Computation and Language (cs.CL)

Abstract: This study constructed a Japanese chat dataset for tuning large language models (LLMs), which consist of about 8.4 million records. Recently, LLMs have been developed and gaining popularity. However, high-performing LLMs are usually mainly for English. There are two ways to support languages other than English by those LLMs: constructing LLMs from scratch or tuning existing models. However, in both ways, datasets are necessary parts. In this study, we focused on supporting Japanese in those LLMs and making a dataset for training or tuning LLMs in Japanese. The dataset we constructed consisted of various tasks, such as translation and knowledge tasks. In our experiment, we tuned an existing LLM using our dataset and evaluated the performance qualitatively. The results suggest that our dataset is possibly beneficial for LLMs. However, we also revealed some difficulties in constructing LLMs in languages other than English.

Abstract: Large language models (LLMs) demonstrate remarkable medical expertise, but data privacy concerns impede their direct use in healthcare environments. Although offering improved data privacy protection, domain-specific small language models (SLMs) often underperform LLMs, emphasizing the need for methods that reduce this performance gap while alleviating privacy concerns. In this paper, we present a simple yet effective method that harnesses LLMs' medical proficiency to boost SLM performance in medical tasks under privacy-restricted scenarios. Specifically, we mitigate patient privacy issues by extracting keywords from medical data and prompting the LLM to generate a medical knowledge-intensive context by simulating clinicians' thought processes. This context serves as additional input for SLMs, augmenting their decision-making capabilities. Our method significantly enhances performance in both few-shot and full training settings across three medical knowledge-intensive tasks, achieving up to a 22.57% increase in absolute accuracy compared to SLM fine-tuning without context, and sets new state-of-the-art results in two medical tasks within privacy-restricted scenarios. Further out-of-domain testing and experiments in two general domain datasets showcase its generalizability and broad applicability.

Abstract: Modeling multi-party conversations (MPCs) with graph neural networks has been proven effective at capturing complicated and graphical information flows. However, existing methods rely heavily on the necessary addressee labels and can only be applied to an ideal setting where each utterance must be tagged with an addressee label. To study the scarcity of addressee labels which is a common issue in MPCs, we propose MADNet that maximizes addressee deduction expectation in heterogeneous graph neural networks for MPC generation. Given an MPC with a few addressee labels missing, existing methods fail to build a consecutively connected conversation graph, but only a few separate conversation fragments instead. To ensure message passing between these conversation fragments, four additional types of latent edges are designed to complete a fully-connected graph. Besides, to optimize the edge-type-dependent message passing for those utterances without addressee labels, an Expectation-Maximization-based method that iteratively generates silver addressee labels (E step), and optimizes the quality of generated responses (M step), is designed. Experimental results on two Ubuntu IRC channel benchmarks show that MADNet outperforms various baseline models on the task of MPC generation, especially under the more common and challenging setting where part of addressee labels are missing.

Abstract: Multilingual semantic parsing aims to leverage the knowledge from the high-resource languages to improve low-resource semantic parsing, yet commonly suffers from the data imbalance problem. Prior works propose to utilize the translations by either humans or machines to alleviate such issues. However, human translations are expensive, while machine translations are cheap but prone to error and bias. In this work, we propose an active learning approach that exploits the strengths of both human and machine translations by iteratively adding small batches of human translations into the machine-translated training set. Besides, we propose novel aggregated acquisition criteria that help our active learning method select utterances to be manually translated. Our experiments demonstrate that an ideal utterance selection can significantly reduce the error and bias in the translated data, resulting in higher parser accuracies than the parsers merely trained on the machine-translated data.

Abstract: Previous studies have shown that large language models (LLMs) like GPTs store massive factual knowledge in their parameters. However, the stored knowledge could be false or out-dated. Traditional knowledge editing methods refine LLMs via fine-tuning on texts containing specific knowledge. However, with the increasing scales of LLMs, these gradient-based approaches bring large computation costs. The trend of model-as-a-service also makes it impossible to modify knowledge in black-box LMs. Inspired by in-context learning (ICL), a new paradigm based on demonstration contexts without parameter updating, we explore whether ICL can edit factual knowledge. To answer this question, we give a comprehensive empirical study of ICL strategies. Experiments show that in-context knowledge editing (IKE), without any gradient and parameter updating, achieves a competitive success rate compared to gradient-based methods on GPT-J (6B) but with much fewer side effects, including less over-editing on similar but unrelated facts and less knowledge forgetting on previously stored knowledge. We also apply the method to larger LMs with tens or hundreds of parameters like OPT-175B, which shows the scalability of our method. The code is available at https://github.com/Zce1112zslx/IKE.

Abstract: Fact-checking real-world claims often requires collecting multiple pieces of evidence and applying complex multi-step reasoning. In this paper, we present Program-Guided Fact-Checking (ProgramFC), a novel fact-checking model that decomposes complex claims into simpler sub-tasks that can be solved using a shared library of specialized functions. We first leverage the in-context learning ability of large language models to generate reasoning programs to guide the verification process. Afterward, we execute the program by delegating each sub-task to the corresponding sub-task handler. This process makes our model both explanatory and data-efficient, providing clear explanations of its reasoning process and requiring minimal training data. We evaluate ProgramFC on two challenging fact-checking datasets and show that it outperforms seven fact-checking baselines across different settings of evidence availability, with explicit output programs that benefit human debugging. Our codes and data are publicly available at https://github.com/mbzuai-nlp/ProgramFC.

Abstract: Etremely Weakly Supervised Text Classification (XWS-TC) refers to text classification based on minimal high-level human guidance, such as a few label-indicative seed words or classification instructions. There are two mainstream approaches for XWS-TC, however, never being rigorously compared: (1) training classifiers based on pseudo-labels generated by (softly) matching seed words (SEED) and (2) prompting (and calibrating) language models using classification instruction (and raw texts) to decode label words (PROMPT). This paper presents the first XWS-TC benchmark to compare the two approaches on fair grounds, where the datasets, supervisions, and hyperparameter choices are standardized across methods. Our benchmarking results suggest that (1) Both SEED and PROMPT approaches are competitive and there is no clear winner; (2) SEED is empirically more tolerant than PROMPT to human guidance (e.g., seed words, classification instructions, and label words) changes; (3) SEED is empirically more selective than PROMPT to the pre-trained language models; (4) Recent SEED and PROMPT methods have close connections and a clustering post-processing step based on raw in-domain texts is a strong performance booster to both. We hope this benchmark serves as a guideline in selecting XWS-TC methods in different scenarios and stimulate interest in developing guidance- and model-robust XWS-TC methods. We release the repo at https://github.com/ZihanWangKi/x-TC.

Abstract: Query-focused meeting summarization(QFMS) aims to generate a specific summary for the given query according to the meeting transcripts. Due to the conflict between long meetings and limited input size, previous works mainly adopt extract-then-summarize methods, which use extractors to simulate binary labels or ROUGE scores to extract utterances related to the query and then generate a summary. However, the previous approach fails to fully use the comparison between utterances. To the extractor, comparison orders are more important than specific scores. In this paper, we propose a Ranker-Generator framework. It learns to rank the utterances by comparing them in pairs and learning from the global orders, then uses top utterances as the generator's input. We show that learning to rank utterances helps to select utterances related to the query effectively, and the summarizer can benefit from it. Experimental results on QMSum show that the proposed model outperforms all existing multi-stage models with fewer parameters.

Abstract: Bias research in NLP seeks to analyse models for social biases, thus helping NLP practitioners uncover, measure, and mitigate social harms. We analyse the body of work that uses prompts and templates to assess bias in language models. We draw on a measurement modelling framework to create a taxonomy of attributes that capture what a bias test aims to measure and how that measurement is carried out. By applying this taxonomy to 90 bias tests, we illustrate qualitatively and quantitatively that core aspects of bias test conceptualisations and operationalisations are frequently unstated or ambiguous, carry implicit assumptions, or be mismatched. Our analysis illuminates the scope of possible bias types the field is able to measure, and reveals types that are as yet under-researched. We offer guidance to enable the community to explore a wider section of the possible bias space, and to better close the gap between desired outcomes and experimental design, both for bias and for evaluating language models more broadly.

Abstract: Recent studies in natural language processing (NLP) have focused on modern languages and achieved state-of-the-art results in many tasks. Meanwhile, little attention has been paid to ancient texts and related tasks. Classical Chinese first came to Japan approximately 2,000 years ago. It was gradually adapted to a Japanese form called Kanbun-Kundoku (Kanbun) in Japanese reading and translating methods, which has significantly impacted Japanese literature. However, compared to the rich resources for ancient texts in mainland China, Kanbun resources remain scarce in Japan. To solve this problem, we construct the first Classical-Chinese-to-Kanbun dataset in the world. Furthermore, we introduce two tasks, character reordering and machine translation, both of which play a significant role in Kanbun comprehension. We also test the current language models on these tasks and discuss the best evaluation method by comparing the results with human scores. We release our code and dataset on GitHub.

Abstract: Cross-lingual natural language inference is a fundamental problem in cross-lingual language understanding. Many recent works have used prompt learning to address the lack of annotated parallel corpora in XNLI. However, these methods adopt discrete prompting by simply translating the templates to the target language and need external expert knowledge to design the templates. Besides, discrete prompts of human-designed template words are not trainable vectors and can not be migrated to target languages in the inference stage flexibly. In this paper, we propose a novel Soft prompt learning framework with the Multilingual Verbalizer (SoftMV) for XNLI. SoftMV first constructs cloze-style question with soft prompts for the input sample. Then we leverage bilingual dictionaries to generate an augmented multilingual question for the original question. SoftMV adopts a multilingual verbalizer to align the representations of original and augmented multilingual questions into the same semantic space with consistency regularization. Experimental results on XNLI demonstrate that SoftMV can achieve state-of-the-art performance and significantly outperform the previous methods under the few-shot and full-shot cross-lingual transfer settings.

Abstract: Large language models (LLMs) have initiated a paradigm shift in transfer learning. In contrast to the classic pretraining-then-finetuning procedure, in order to use LLMs for downstream prediction tasks, one only needs to provide a few demonstrations, known as in-context examples, without adding more or updating existing model parameters. This in-context learning (ICL) capabilities of LLMs is intriguing, and it is not yet fully understood how pretrained LLMs acquire such capabilities. In this paper, we investigate the reason why a transformer-based language model can accomplish in-context learning after pre-training on a general language corpus by proposing one hypothesis that LLMs can simulate kernel regression algorithms when faced with in-context examples. More concretely, we first prove that Bayesian inference on in-context prompts can be asymptotically understood as kernel regression $\hat y = \frac{\sum_i y_i K(x, x_i)}{\sum_i K(x, x_i)}$ as the number of in-context demonstrations grows. Then, we empirically investigate the in-context behaviors of language models. We find that during ICL, the attentions and hidden features in LLMs match the behaviors of a kernel regression. Finally, our theory provides insights on multiple phenomena observed in ICL field: why retrieving demonstrative samples similar to test sample can help, why ICL performance is sensitive to the output formats, and why ICL accuracy benefits from selecting in-distribution and representative samples. We will make our code available to the research community following publication.

Abstract: Simultaneous machine translation (SiMT) starts to output translation while reading the source sentence and needs a precise policy to decide when to output the generated translation. Therefore, the policy determines the number of source tokens read during the translation of each target token. However, it is difficult to learn a precise translation policy to achieve good latency-quality trade-offs, because there is no golden policy corresponding to parallel sentences as explicit supervision. In this paper, we present a new method for constructing the optimal policy online via binary search. By employing explicit supervision, our approach enables the SiMT model to learn the optimal policy, which can guide the model in completing the translation during inference. Experiments on four translation tasks show that our method can exceed strong baselines across all latency scenarios.

Abstract: Evaluating grounded neural language model performance with respect to pragmatic qualities like the trade off between truthfulness, contrastivity and overinformativity of generated utterances remains a challenge in absence of data collected from humans. To enable such evaluation, we present a novel open source image-text dataset "Annotated 3D Shapes" (A3DS) comprising over nine million exhaustive natural language annotations and over 12 million variable-granularity captions for the 480,000 images provided by Burges & Kim (2018). We showcase the evaluation of pragmatic abilities developed by a task-neutral image captioner fine-tuned in a multi-agent communication setting to produce contrastive captions. The evaluation is enabled by the dataset because the exhaustive annotations allow to quantify the presence of contrastive features in the model's generations. We show that the model develops human-like patterns (informativity, brevity, over-informativity for specific features (e.g., shape, color biases)).

Abstract: Generating persona consistent dialogue response is important for developing an intelligent conversational agent. Recent works typically fine-tune large-scale pre-trained models on this task by concatenating persona texts and dialogue history as a single input sequence to generate the target response. While simple and effective, our analysis shows that this popular practice is seriously affected by order sensitivity where different input orders of persona sentences significantly impact the quality and consistency of generated response, resulting in severe performance fluctuations (i.e., 29.4% on GPT2 and 83.2% on BART). To mitigate the order sensitivity problem, we propose a model-agnostic framework, ORder Insensitive Generation (ORIG), which enables dialogue models to learn robust representation under different persona orders and improve the consistency of response generation. Experiments on the Persona-Chat dataset justify the effectiveness and superiority of our method with two dominant pre-trained models (GPT2 and BART).

Abstract: Multi-aspect controllable text generation aims to generate fluent sentences that possess multiple desired attributes simultaneously. Traditional methods either combine many operators in the decoding stage, often with costly iteration or search in the discrete text space, or train separate controllers for each aspect, resulting in a degeneration of text quality due to the discrepancy between different aspects. To address these limitations, we introduce a novel approach for multi-aspect control, namely MacLaSa, that estimates compact latent space for multiple aspects and performs efficient sampling with a robust sampler based on ordinary differential equations (ODEs). To eliminate the domain gaps between different aspects, we utilize a Variational Autoencoder (VAE) network to map text sequences from varying data sources into close latent representations. The estimated latent space enables the formulation of joint energy-based models (EBMs) and the plugging in of arbitrary attribute discriminators to achieve multi-aspect control. Afterwards, we draw latent vector samples with an ODE-based sampler and feed sampled examples to the VAE decoder to produce target text sequences. Experimental results demonstrate that MacLaSa outperforms several strong baselines on attribute relevance and textual quality while maintaining a high inference speed.

Abstract: Despite advances in multilingual neural machine translation (MNMT), we argue that there are still two major challenges in this area: data imbalance and representation degeneration. The data imbalance problem refers to the imbalance in the amount of parallel corpora for all language pairs, especially for long-tail languages (i.e., very low-resource languages). The representation degeneration problem refers to the problem of encoded tokens tending to appear only in a small subspace of the full space available to the MNMT model. To solve these two issues, we propose Bi-ACL, a framework that uses only target-side monolingual data and a bilingual dictionary to improve the performance of the MNMT model. We define two modules, named bidirectional autoencoder and bidirectional contrastive learning, which we combine with an online constrained beam search and a curriculum learning sampling strategy. Extensive experiments show that our proposed method is more effective both in long-tail languages and in high-resource languages. We also demonstrate that our approach is capable of transferring knowledge between domains and languages in zero-shot scenarios.

Abstract: Event Causality Identification (ECI) aims to identify causal relations between events in unstructured texts. This is a very challenging task, because causal relations are usually expressed by implicit associations between events. Existing methods usually capture such associations by directly modeling the texts with pre-trained language models, which underestimate two kinds of semantic structures vital to the ECI task, namely, event-centric structure and event-associated structure. The former includes important semantic elements related to the events to describe them more precisely, while the latter contains semantic paths between two events to provide possible supports for ECI. In this paper, we study the implicit associations between events by modeling the above explicit semantic structures, and propose a Semantic Structure Integration model (SemSIn). It utilizes a GNN-based event aggregator to integrate the event-centric structure information, and employs an LSTM-based path aggregator to capture the event-associated structure information between two events. Experimental results on three widely used datasets show that SemSIn achieves significant improvements over baseline methods.

Abstract: In recent years, large language models (LMs) have achieved remarkable progress across various natural language processing tasks. As pre-training and fine-tuning are costly and might negatively impact model performance, it is desired to efficiently adapt an existing model to different conditions such as styles, sentiments or narratives, when facing different audiences or scenarios. However, efficient adaptation of a language model to diverse conditions remains an open challenge. This work is inspired by the observation that text conditions are often associated with selection of certain words in a context. Therefore we introduce LM-Switch, a theoretically grounded, lightweight and simple method for generative language model conditioning. We begin by investigating the effect of conditions in Hidden Markov Models (HMMs), and establish a theoretical connection with language model. Our finding suggests that condition shifts in HMMs are associated with linear transformations in word embeddings. LM-Switch is then designed to deploy a learnable linear factor in the word embedding space for language model conditioning. We show that LM-Switch can model diverse tasks, and achieves comparable or better performance compared with state-of-the-art baselines in LM detoxification and generation control, despite requiring no more than 1% of parameters compared with baselines and little extra time overhead compared with base LMs. It is also able to learn from as few as a few sentences or one document. Moreover, a learned LM-Switch can be transferred to other LMs of different sizes, achieving a detoxification performance similar to the best baseline. We will make our code available to the research community following publication.

Abstract: Ultra-fine entity typing (UFET) is the task of inferring the semantic types, from a large set of fine-grained candidates, that apply to a given entity mention. This task is especially challenging because we only have a small number of training examples for many of the types, even with distant supervision strategies. State-of-the-art models, therefore, have to rely on prior knowledge about the type labels in some way. In this paper, we show that the performance of existing methods can be improved using a simple technique: we use pre-trained label embeddings to cluster the labels into semantic domains and then treat these domains as additional types. We show that this strategy consistently leads to improved results, as long as high-quality label embeddings are used. We furthermore use the label clusters as part of a simple post-processing technique, which results in further performance gains. Both strategies treat the UFET model as a black box and can thus straightforwardly be used to improve a wide range of existing models.

Abstract: Agency, the capacity to proactively shape events, is crucial to how humans interact and collaborate with other humans. In this paper, we investigate Agency as a potentially desirable function of dialogue agents, and how it can be measured and controlled. We build upon the social-cognitive theory of Bandura (2001) to develop a framework of features through which Agency is expressed in dialogue -- indicating what you intend to do (Intentionality), motivating your intentions (Motivation), having self-belief in intentions (Self-Efficacy), and being able to self-adjust (Self-Regulation). We collect and release a new dataset of 83 human-human collaborative interior design conversations containing 908 conversational snippets annotated for Agency features. Using this dataset, we explore methods for measuring and controlling Agency in dialogue systems. Automatic and human evaluation show that although a baseline GPT-3 model can express Intentionality, models that explicitly manifest features associated with high Motivation, Self-Efficacy, and Self-Regulation are better perceived as being highly agentive. This work has implications for the development of dialogue systems with varying degrees of Agency in collaborative tasks.

Abstract: Pretrained language models have achieved remarkable success in various natural language processing tasks. However, pretraining has recently shifted toward larger models and larger data, and this has resulted in significant computational and energy costs. In this paper, we propose Influence Subset Selection (ISS) for language model, which explicitly utilizes end-task knowledge to select a tiny subset of the pretraining corpus. Specifically, the ISS selects the samples that will provide the most positive influence on the performance of the end-task. Furthermore, we design a gradient matching based influence estimation method, which can drastically reduce the computation time of influence. With only 0.45% of the data and a three-orders-of-magnitude lower computational cost, ISS outperformed pretrained models (e.g., RoBERTa) on eight datasets covering four domains.

Abstract: Colexification in comparative linguistics refers to the phenomenon of a lexical form conveying two or more distinct meanings. In this paper, we propose simple and effective methods to build multilingual graphs from colexification patterns: ColexNet and ColexNet+. ColexNet's nodes are concepts and its edges are colexifications. In ColexNet+, concept nodes are in addition linked through intermediate nodes, each representing an ngram in one of 1,334 languages. We use ColexNet+ to train high-quality multilingual embeddings $\overrightarrow{\mbox{ColexNet+}}$ that are well-suited for transfer learning scenarios. Existing work on colexification patterns relies on annotated word lists. This limits scalability and usefulness in NLP. In contrast, we identify colexification patterns of more than 2,000 concepts across 1,335 languages directly from an unannotated parallel corpus. In our experiments, we first show that ColexNet has a high recall on CLICS, a dataset of crosslingual colexifications. We then evaluate $\overrightarrow{\mbox{ColexNet+}}$ on roundtrip translation, verse retrieval and verse classification and show that our embeddings surpass several baselines in a transfer learning setting. This demonstrates the benefits of colexification for multilingual NLP.

Abstract: Recent advances in tabular question answering (QA) with large language models are constrained in their coverage and only answer questions over a single table. However, real-world queries are complex in nature, often over multiple tables in a relational database or web page. Single table questions do not involve common table operations such as set operations, Cartesian products (joins), or nested queries. Furthermore, multi-table operations often result in a tabular output, which necessitates table generation capabilities of tabular QA models. To fill this gap, we propose a new task of answering questions over multiple tables. Our model, MultiTabQA, not only answers questions over multiple tables, but also generalizes to generate tabular answers. To enable effective training, we build a pre-training dataset comprising of 132,645 SQL queries and tabular answers. Further, we evaluate the generated tables by introducing table-specific metrics of varying strictness assessing various levels of granularity of the table structure. MultiTabQA outperforms state-of-the-art single table QA models adapted to a multi-table QA setting by finetuning on three datasets: Spider, Atis and GeoQuery.

Abstract: In this paper, we provide a holistic analysis of the different sources of bias, Upstream, Sample and Overampflication biases, in NLP models. We investigate how they impact the fairness of the task of text classification. We also investigate the impact of removing these biases using different debiasing techniques on the fairness of text classification. We found that overamplification bias is the most impactful bias on the fairness of text classification. And that removing overamplification bias by fine-tuning the LM models on a dataset with balanced representations of the different identity groups leads to fairer text classification models. Finally, we build on our findings and introduce practical guidelines on how to have a fairer text classification model.

Abstract: Researchers have proposed various information extraction (IE) techniques to convert news articles into structured knowledge for news understanding. However, none of the existing methods have explicitly addressed the issue of framing bias that is inherent in news articles. We argue that studying and identifying framing bias is a crucial step towards trustworthy event understanding. We propose a novel task, neutral event graph induction, to address this problem. An event graph is a network of events and their temporal relations. Our task aims to induce such structural knowledge with minimal framing bias in an open domain. We propose a three-step framework to induce a neutral event graph from multiple input sources. The process starts by inducing an event graph from each input source, then merging them into one merged event graph, and lastly using a Graph Convolutional Network to remove event nodes with biased connotations. We demonstrate the effectiveness of our framework through the use of graph prediction metrics and bias-focused metrics.

Abstract: Recent years have seen remarkable progress in automatic speech recognition (ASR). However, traditional token-level ASR models have struggled with accurately transcribing entities due to the problem of homophonic and near-homophonic tokens. This paper introduces a novel approach called CopyNE, which uses a span-level copying mechanism to improve ASR in transcribing entities. CopyNE can copy all tokens of an entity at once, effectively avoiding errors caused by homophonic or near-homophonic tokens that occur when predicting multiple tokens separately. Experiments on Aishell and ST-cmds datasets demonstrate that CopyNE achieves significant reductions in character error rate (CER) and named entity CER (NE-CER), especially in entity-rich scenarios. Furthermore, even when compared to the strong Whisper baseline, CopyNE still achieves notable reductions in CER and NE-CER. Qualitative comparisons with previous approaches demonstrate that CopyNE can better handle entities, effectively improving the accuracy of ASR.

Abstract: This study proposes a multitask learning architecture for extractive summarization with coherence boosting. The architecture contains an extractive summarizer and coherent discriminator module. The coherent discriminator is trained online on the sentence vectors of the augmented textual input, thus improving its general ability of judging whether the input sentences are coherent. Meanwhile, we maximize the coherent scores from the coherent discriminator by updating the parameters of the summarizer. To make the extractive sentences trainable in a differentiable manner, we introduce two strategies, including pre-trained converting model (model-based) and converting matrix (MAT-based) that merge sentence representations. Experiments show that our proposed method significantly improves the proportion of consecutive sentences in the extracted summaries based on their positions in the original article (i.e., automatic sentence-level coherence metric), while the goodness in terms of other automatic metrics (i.e., Rouge scores and BertScores) are preserved. Human evaluation also evidences the improvement of coherence and consistency of the extracted summaries given by our method.

Abstract: The practice of transferring knowledge from a sophisticated, closed-source large language model (LLM) to a compact, open-source LLM has garnered considerable attention. Previous works have focused on a unidirectional knowledge distillation way by aligning the responses of the student model with those of the teacher model to a set of instructions. Nevertheless, they overlooked the possibility of incorporating any reciprocal "feedback"--identifying challenging instructions where the student model's performance falls short--to boost the student model's proficiency iteratively. To this end, we propose a novel adversarial distillation framework for a more efficient knowledge transfer. Leveraging the versatile role adaptability of LLMs, we prompt the closed-source model to identify "hard" instructions and generate new "hard" instructions for the student model, creating a three-stage adversarial loop of imitation, discrimination, and generation. By applying this adversarial framework, we successfully transfer knowledge from ChatGPT to a 7B student model (named Lion), achieving nearly 95% capability approximation using a mere 70k training data. We aspire that this proposed model may serve as the baseline to reflect the performance of ChatGPT, especially the open-source instruction-following language model baseline for our community.

Abstract: Non-autoregressive translation (NAT) models have been extensively investigated within the context of sentence-level machine translation (MT) tasks, demonstrating comparable quality and superior translation speed when contrasted with autoregressive translation (AT) models. However, the challenges associated with multi-modality and alignment issues within NAT models become more prominent when increasing input and output length, leading to unexpected complications in document-level MT. In this paper, we conduct a comprehensive examination of typical NAT models in the context of document-level MT tasks. Experiments reveal that, although NAT models significantly accelerate text generation on documents, they do not perform as effectively as on sentences. To bridge this performance gap, we introduce a novel design that underscores the importance of sentence-level alignment for non-autoregressive document-level machine translation (NA-DMT). This innovation substantially reduces the performance discrepancy. However, it is worth noting that NA-DMT models are still far from perfect and may necessitate additional research to fully optimize their performance. We delve into the related opportunities and challenges and provide our code at https://github.com/baoguangsheng/nat-on-doc to stimulate further research in this field.

Abstract: Machine Learning models have many potentially beneficial applications in education settings, but a key barrier to their development is securing enough data to train these models. Labelling educational data has traditionally relied on highly skilled raters using complex, multi-class rubrics, making the process expensive and difficult to scale. An alternative, more scalable approach could be to use non-expert crowdworkers to evaluate student work, however, maintaining sufficiently high levels of accuracy and inter-rater reliability when using non-expert workers is challenging. This paper reports on two experiments investigating using non-expert crowdworkers and comparative judgement to evaluate complex student data. Crowdworkers were hired to evaluate student responses to open-ended reading comprehension questions. Crowdworkers were randomly assigned to one of two conditions: the control, where they were asked to decide whether answers were correct or incorrect (i.e., a categorical judgement), or the treatment, where they were shown the same question and answers, but were instead asked to decide which of two candidate answers was more correct (i.e., a comparative/preference-based judgement). We found that using comparative judgement substantially improved inter-rater reliability on both tasks. These results are in-line with well-established literature on the benefits of comparative judgement in the field of educational assessment, as well as with recent trends in artificial intelligence research, where comparative judgement is becoming the preferred method for providing human feedback on model outputs when working with non-expert crowdworkers. However, to our knowledge, these results are novel and important in demonstrating the beneficial effects of using the combination of comparative judgement and crowdworkers to evaluate educational data.

Abstract: There have been many recent investigations into prompt-based training of transformer language models for new text genres in low-resource settings. The prompt-based training approach has been found to be effective in generalizing pre-trained or fine-tuned models for transfer to resource-scarce settings. This work, for the first time, reports results on adopting prompt-based training of transformers for \textit{scholarly knowledge graph object prediction}. The work is unique in the following two main aspects. 1) It deviates from the other works proposing entity and relation extraction pipelines for predicting objects of a scholarly knowledge graph. 2) While other works have tested the method on text genera relatively close to the general knowledge domain, we test the method for a significantly different domain, i.e. scholarly knowledge, in turn testing the linguistic, probabilistic, and factual generalizability of these large-scale transformer models. We find that (i) per expectations, transformer models when tested out-of-the-box underperform on a new domain of data, (ii) prompt-based training of the models achieve performance boosts of up to 40\% in a relaxed evaluation setting, and (iii) testing the models on a starkly different domain even with a clever training objective in a low resource setting makes evident the domain knowledge capture gap offering an empirically-verified incentive for investing more attention and resources to the scholarly domain in the context of transformer models.

Abstract: Large language models have shown tremendous performance in a variety of tasks. In-context learning -- the ability to improve at a task after being provided with a number of demonstrations -- is seen as one of the main contributors to their success. In the present paper, we demonstrate that the in-context learning abilities of large language models can be recursively improved via in-context learning itself. We coin this phenomenon meta-in-context learning. Looking at two idealized domains, a one-dimensional regression task and a two-armed bandit task, we show that meta-in-context learning adaptively reshapes a large language model's priors over expected tasks. Furthermore, we find that meta-in-context learning modifies the in-context learning strategies of such models. Finally, we extend our approach to a benchmark of real-world regression problems where we observe competitive performance to traditional learning algorithms. Taken together, our work improves our understanding of in-context learning and paves the way toward adapting large language models to the environment they are applied purely through meta-in-context learning rather than traditional finetuning.

Abstract: Automatic text simplification systems help to reduce textual information barriers on the internet. However, for languages other than English, only few parallel data to train these systems exists. We propose a two-step approach to overcome this data scarcity issue. First, we fine-tuned language models on a corpus of German Easy Language, a specific style of German. Then, we used these models as decoders in a sequence-to-sequence simplification task. We show that the language models adapt to the style characteristics of Easy Language and output more accessible texts. Moreover, with the style-specific pre-training, we reduced the number of trainable parameters in text simplification models. Hence, less parallel data is sufficient for training. Our results indicate that pre-training on unaligned data can reduce the required parallel data while improving the performance on downstream tasks.

Abstract: We introduce Parallel Paraphrasing ($\text{Para}_\text{both}$), an augmentation method for translation metrics making use of automatic paraphrasing of both the reference and hypothesis. This method counteracts the typically misleading results of speech translation metrics such as WER, CER, and BLEU if only a single reference is available. We introduce two new datasets explicitly created to measure the quality of metrics intended to be applied to Swiss German speech-to-text systems. Based on these datasets, we show that we are able to significantly improve the correlation with human quality perception if our method is applied to commonly used metrics.

Abstract: Understanding the fundamental concepts and trends in a scientific field is crucial for keeping abreast of its ongoing development. In this study, we propose a systematic framework for analyzing the evolution of research topics in a scientific field using causal discovery and inference techniques. By conducting extensive experiments on the ACL Anthology corpus, we demonstrate that our framework effectively uncovers evolutionary trends and the underlying causes for a wide range of natural language processing (NLP) research topics.

Abstract: Entity typing is the task of assigning semantic types to the entities that are mentioned in a text. Since obtaining sufficient amounts of manual annotations is expensive, current state-of-the-art methods are typically trained on automatically labelled datasets, e.g. by exploiting links between Wikipedia pages. In this paper, we propose to use coreference chains as an additional supervision signal. Specifically, we pre-train an entity encoder using a contrastive loss, such that entity embeddings of coreferring entities are more similar to each other than to the embeddings of other entities. Since this strategy is not tied to Wikipedia, we can pre-train our entity encoder on other genres than encyclopedic text and on larger amounts of data. Our experimental results show that the proposed pre-training strategy allows us to improve the state-of-the-art in fine-grained entity typing, provided that only high-quality coreference links are exploited.

Abstract: Speaker diarization(SD) is a classic task in speech processing and is crucial in multi-party scenarios such as meetings and conversations. Current mainstream speaker diarization approaches consider acoustic information only, which result in performance degradation when encountering adverse acoustic conditions. In this paper, we propose methods to extract speaker-related information from semantic content in multi-party meetings, which, as we will show, can further benefit speaker diarization. We introduce two sub-tasks, Dialogue Detection and Speaker-Turn Detection, in which we effectively extract speaker information from conversational semantics. We also propose a simple yet effective algorithm to jointly model acoustic and semantic information and obtain speaker-identified texts. Experiments on both AISHELL-4 and AliMeeting datasets show that our method achieves consistent improvements over acoustic-only speaker diarization systems.

Abstract: One of the essential human skills is the ability to seamlessly build an inner representation of the world. By exploiting this representation, humans are capable of easily finding consensus between visual, auditory and linguistic perspectives. In this work, we set out to understand and emulate this ability through an explicit representation for both vision and language - Graphs of Events in Space and Time (GEST). GEST alows us to measure the similarity between texts and videos in a semantic and fully explainable way, through graph matching. It also allows us to generate text and videos from a common representation that provides a well understood content. In this work we show that the graph matching similarity metrics based on GEST outperform classical text generation metrics and can also boost the performance of state of art, heavily trained metrics.

Abstract: Compositionality is a hallmark of human language that not only enables linguistic generalization, but also potentially facilitates acquisition. When simulating language emergence with neural networks, compositionality has been shown to improve communication performance; however, its impact on imitation learning has yet to be investigated. Our work explores the link between compositionality and imitation in a Lewis game played by deep neural agents. Our contributions are twofold: first, we show that the learning algorithm used to imitate is crucial: supervised learning tends to produce more average languages, while reinforcement learning introduces a selection pressure toward more compositional languages. Second, our study reveals that compositional languages are easier to imitate, which may induce the pressure toward compositional languages in RL imitation settings.

Abstract: As ChatGPT and GPT-4 spearhead the development of Large Language Models (LLMs), more researchers are investigating their performance across various tasks. But more research needs to be done on the interpretability capabilities of LLMs, that is, the ability to generate reasons after an answer has been given. Existing explanation datasets are mostly English-language general knowledge questions, which leads to insufficient thematic and linguistic diversity. To address the language bias and lack of medical resources in generating rationales QA datasets, we present ExplainCPE (over 7k instances), a challenging medical benchmark in Simplified Chinese. We analyzed the errors of ChatGPT and GPT-4, pointing out the limitations of current LLMs in understanding text and computational reasoning. During the experiment, we also found that different LLMs have different preferences for in-context learning. ExplainCPE presents a significant challenge, but its potential for further investigation is promising, and it can be used to evaluate the ability of a model to generate explanations. AI safety and trustworthiness need more attention, and this work makes the first step to explore the medical interpretability of LLMs.The dataset is available at https://github.com/HITsz-TMG/ExplainCPE.

Abstract: The emergence of generative large language models (LLMs) raises the question: what will be its impact on crowdsourcing. Traditionally, crowdsourcing has been used for acquiring solutions to a wide variety of human-intelligence tasks, including ones involving text generation, manipulation or evaluation. For some of these tasks, models like ChatGPT can potentially substitute human workers. In this study, we investigate, whether this is the case for the task of paraphrase generation for intent classification. We quasi-replicated the data collection methodology of an existing crowdsourcing study (similar scale, prompts and seed data) using ChatGPT. We show that ChatGPT-created paraphrases are more diverse and lead to more robust models.

Abstract: In behavioural testing, system functionalities underrepresented in the standard evaluation setting (with a held-out test set) are validated through controlled input-output pairs. Optimising performance on the behavioural tests during training (behavioural learning) would improve coverage of phenomena not sufficiently represented in the i.i.d. data and could lead to seemingly more robust models. However, there is the risk that the model narrowly captures spurious correlations from the behavioural test suite, leading to overestimation and misrepresentation of model performance -- one of the original pitfalls of traditional evaluation. In this work, we introduce BeLUGA, an analysis method for evaluating behavioural learning considering generalisation across dimensions of different granularity levels. We optimise behaviour-specific loss functions and evaluate models on several partitions of the behavioural test suite controlled to leave out specific phenomena. An aggregate score measures generalisation to unseen functionalities (or overfitting). We use BeLUGA to examine three representative NLP tasks (sentiment analysis, paraphrase identification and reading comprehension) and compare the impact of a diverse set of regularisation and domain generalisation methods on generalisation performance.

Abstract: Assessing student answers and providing valuable feedback is crucial for effective learning, but it can be a time-consuming task. Traditional methods of automating student answer assessment through text classification often suffer from issues such as lack of trustworthiness, transparency, and the ability to provide a rationale for the automated assessment process. These limitations hinder their usefulness in practice. In this paper, we explore using ChatGPT, a cutting-edge large language model, for the concurrent tasks of student answer scoring and rationale generation under both the zero-shot and few-shot settings. We introduce a critic module which automatically filters incorrect outputs from ChatGPT and utilizes the remaining ChtaGPT outputs as noisy labelled data to fine-tune a smaller language model, enabling it to perform student answer scoring and rationale generation. Moreover, by drawing multiple samples from ChatGPT outputs, we are able to compute predictive confidence scores, which in turn can be used to identify corrupted data and human label errors in the training set. Our experimental results demonstrate that despite being a few orders of magnitude smaller than ChatGPT, the fine-tuned language model achieves better performance in student answer scoring. Furthermore, it generates more detailed and comprehensible assessments than traditional text classification methods. Our approach provides a viable solution to achieve explainable automated assessment in education.

Abstract: This paper details the process of developing the first native large generative language model for the Nordic languages, GPT-SW3. We cover all parts of the development process, from data collection and processing, training configuration and instruction finetuning, to evaluation and considerations for release strategies. We hope that this paper can serve as a guide and reference for other researchers that undertake the development of large generative models for smaller languages.

Abstract: Recent progress in sentence embedding, which represents the meaning of a sentence as a point in a vector space, has achieved high performance on tasks such as a semantic textual similarity (STS) task. However, sentence representations as a point in a vector space can express only a part of the diverse information that sentences have, such as asymmetrical relationships between sentences. This paper proposes GaussCSE, a Gaussian distribution-based contrastive learning framework for sentence embedding that can handle asymmetric relationships between sentences, along with a similarity measure for identifying inclusion relations. Our experiments show that GaussCSE achieves the same performance as previous methods in natural language inference tasks, and is able to estimate the direction of entailment relations, which is difficult with point representations.

Abstract: Understanding the internal reasoning behind the predictions of machine learning systems is increasingly vital, given their rising adoption and acceptance. While previous approaches, such as LIME, generate algorithmic explanations by attributing importance to input features for individual examples, recent research indicates that practitioners prefer examining language explanations that explain sub-groups of examples. In this paper, we introduce MaNtLE, a model-agnostic natural language explainer that analyzes multiple classifier predictions and generates faithful natural language explanations of classifier rationale for structured classification tasks. MaNtLE uses multi-task training on thousands of synthetic classification tasks to generate faithful explanations. Simulated user studies indicate that, on average, MaNtLE-generated explanations are at least 11% more faithful compared to LIME and Anchors explanations across three tasks. Human evaluations demonstrate that users can better predict model behavior using explanations from MaNtLE compared to other techniques

Abstract: Pre-trained seq2seq models have achieved state-of-the-art results in the grammatical error correction task. However, these models still suffer from a prediction bias due to their unidirectional decoding. Thus, we propose a bidirectional Transformer reranker (BTR), that re-estimates the probability of each candidate sentence generated by the pre-trained seq2seq model. The BTR preserves the seq2seq-style Transformer architecture but utilizes a BERT-style self-attention mechanism in the decoder to compute the probability of each target token by using masked language modeling to capture bidirectional representations from the target context. For guiding the reranking, the BTR adopts negative sampling in the objective function to minimize the unlikelihood. During inference, the BTR gives final results after comparing the reranked top-1 results with the original ones by an acceptance threshold. Experimental results show that, in reranking candidates from a pre-trained seq2seq model, T5-base, the BTR on top of T5-base could yield 65.47 and 71.27 F0.5 scores on the CoNLL-14 and BEA test sets, respectively, and yield 59.52 GLEU score on the JFLEG corpus, with improvements of 0.36, 0.76 and 0.48 points compared with the original T5-base. Furthermore, when reranking candidates from T5-large, the BTR on top of T5-base improved the original T5-large by 0.26 points on the BEA test set.

Abstract: Semi-supervised learning (SSL) is a popular setting aiming to effectively utilize unlabelled data to improve model performance in downstream natural language processing (NLP) tasks. Currently, there are two popular approaches to make use of unlabelled data: Self-training (ST) and Task-adaptive pre-training (TAPT). ST uses a teacher model to assign pseudo-labels to the unlabelled data, while TAPT continues pre-training on the unlabelled data before fine-tuning. To the best of our knowledge, the effectiveness of TAPT in SSL tasks has not been systematically studied, and no previous work has directly compared TAPT and ST in terms of their ability to utilize the pool of unlabelled data. In this paper, we provide an extensive empirical study comparing five state-of-the-art ST approaches and TAPT across various NLP tasks and data sizes, including in- and out-of-domain settings. Surprisingly, we find that TAPT is a strong and more robust SSL learner, even when using just a few hundred unlabelled samples or in the presence of domain shifts, compared to more sophisticated ST approaches, and tends to bring greater improvements in SSL than in fully-supervised settings. Our further analysis demonstrates the risks of using ST approaches when the size of labelled or unlabelled data is small or when domain shifts exist. We offer a fresh perspective for future SSL research, suggesting the use of unsupervised pre-training objectives over dependency on pseudo labels.

Abstract: Speech language models (SpeechLMs) process and generate acoustic data only, without textual supervision. In this work, we propose TWIST, a method for training SpeechLMs using a warm-start from a pretrained textual language models. We show using both automatic and human evaluations that TWIST outperforms a cold-start SpeechLM across the board. We empirically analyze the effect of different model design choices such as the speech tokenizer, the pretrained textual model, and the dataset size. We find that model and dataset scale both play an important role in constructing better-performing SpeechLMs. Based on our observations, we present the largest (to the best of our knowledge) SpeechLM both in terms of number of parameters and training data. We additionally introduce two spoken versions of the StoryCloze textual benchmark to further improve model evaluation and advance future research in the field. Speech samples can be found on our website: https://pages.cs.huji.ac.il/adiyoss-lab/twist/ .

Abstract: Large Language Models (LLMs) have emerged as powerful generative Artificial Intelligence solutions which can be applied to several fields and areas of work. This paper presents results and reflection of an experiment done to use the model GPT 3.5-Turbo to emulate some aspects of an inductive Thematic Analysis. Previous research on this subject has largely worked on conducting deductive analysis. Thematic Analysis is a qualitative method for analysis commonly used in social sciences and it is based on interpretations made by the human analyst(s) and the identification of explicit and latent meanings in qualitative data. Attempting an analysis based on human interpretation with an LLM clearly is a provocation but also a way to learn something about how these systems can or cannot be used in qualitative research. The paper presents the motivations for attempting this emulation, it reflects on how the six steps to a Thematic Analysis proposed by Braun and Clarke can at least partially be reproduced with the LLM and it also reflects on what are the outputs produced by the model. The paper used two existing datasets of open access semi-structured interviews, previously analysed with Thematic Analysis by other researchers. It used the previously produced analysis (and the related themes) to compare with the results produced by the LLM. The results show that the model can infer at least partially some of the main Themes. The objective of the paper is not to replace human analysts in qualitative analysis but to learn if some elements of LLM data manipulation can to an extent be of support for qualitative research.

Abstract: The main objective of Knowledge Graph (KG) embeddings is to learn low-dimensional representations of entities and relations, enabling the prediction of missing facts. A significant challenge in achieving better KG embeddings lies in capturing relation patterns, including symmetry, antisymmetry, inversion, commutative composition, non-commutative composition, hierarchy, and multiplicity. This study introduces a novel model called 3H-TH (3D Rotation and Translation in Hyperbolic space) that captures these relation patterns simultaneously. In contrast, previous attempts have not achieved satisfactory performance across all the mentioned properties at the same time. The experimental results demonstrate that the new model outperforms existing state-of-the-art models in terms of accuracy, hierarchy property, and other relation patterns in low-dimensional space, meanwhile performing similarly in high-dimensional space.

Abstract: Large language models (LLMs) have exhibited an emergent in-context learning (ICL) ability. However, the ICL models that can solve ordinary cases are hardly extended to solve more complex tasks by processing the demonstration examples once. This single-turn ICL is incoordinate with the decision making process of humans by learning from analogy. In this paper, we propose an effective and efficient two-stage framework to boost ICL in LLMs by exploiting a dual form between Transformer attention and gradient descent-based optimization. Concretely, we divide the ICL process into "Deep-Thinking" and inference stages. The "Deep-Thinking" stage performs iterative forward optimization of demonstrations, which is expected to boost the reasoning abilities of LLMs at test time by "thinking" demonstrations multiple times. It produces accumulated meta-gradients by manipulating the Key-Value matrices in the self-attention modules of the Transformer. Then, the inference stage only takes the test query as input without concatenating demonstrations and applies the learned meta-gradients through attention for output prediction. In this way, demonstrations are not required during the inference stage since they are already learned and stored in the definitive meta-gradients. LLMs can be effectively and efficiently adapted to downstream tasks. Extensive experiments on ten classification and multiple-choice datasets show that our method achieves substantially better performance than standard ICL in terms of both accuracy and efficiency.

Abstract: We introduce the Dutch Model Benchmark: DUMB. The benchmark includes a diverse set of datasets for low-, medium- and high-resource tasks. The total set of eight tasks include three tasks that were previously not available in Dutch. Instead of relying on a mean score across tasks, we propose Relative Error Reduction (RER), which compares the DUMB performance of models to a strong baseline which can be referred to in the future even when assessing different sets of models. Through a comparison of 14 pre-trained models (mono- and multi-lingual, of varying sizes), we assess the internal consistency of the benchmark tasks, as well as the factors that likely enable high performance. Our results indicate that current Dutch monolingual models under-perform and suggest training larger Dutch models with other architectures and pre-training objectives. At present, the highest performance is achieved by DeBERTaV3 (large), XLM-R (large) and mDeBERTaV3 (base). In addition to highlighting best strategies for training larger Dutch models, DUMB will foster further research on Dutch. A public leaderboard is available at https://dumbench.nl.

Abstract: Nearest Neighbor Machine Translation ($k$NN-MT) has achieved great success on domain adaptation tasks by integrating pre-trained Neural Machine Translation (NMT) models with domain-specific token-level retrieval. However, the reasons underlying its success have not been thoroughly investigated. In this paper, we provide a comprehensive analysis of $k$NN-MT through theoretical and empirical studies. Initially, we offer a theoretical interpretation of the working mechanism of $k$NN-MT as an efficient technique to implicitly execute gradient descent on the output projection layer of NMT, indicating that it is a specific case of model fine-tuning. Subsequently, we conduct multi-domain experiments and word-level analysis to examine the differences in performance between $k$NN-MT and entire-model fine-tuning. Our findings suggest that: (1) Incorporating $k$NN-MT with adapters yields comparable translation performance to fine-tuning on in-domain test sets, while achieving better performance on out-of-domain test sets; (2) Fine-tuning significantly outperforms $k$NN-MT on the recall of low-frequency domain-specific words, but this gap could be bridged by optimizing the context representations with additional adapter layers.

Abstract: Task-oriented dialogue (TOD) models have great progress in the past few years. However, these studies primarily focus on datasets written by annotators, which has resulted in a gap between academic research and more realistic spoken conversation scenarios. While a few small-scale spoken TOD datasets are proposed to address robustness issues, e.g., ASR errors, they fail to identify the unique challenges in spoken conversation. To tackle the limitations, we introduce SpokenWOZ, a large-scale speech-text dataset for spoken TOD, which consists of 8 domains, 203k turns, 5.7k dialogues and 249 hours of audios from human-to-human spoken conversations. SpokenWOZ incorporates common spoken characteristics such as word-by-word processing and commonsense reasoning. We also present cross-turn slot and reasoning slot detection as new challenges based on the spoken linguistic phenomena. We conduct comprehensive experiments on various models, including text-modal baselines, newly proposed dual-modal baselines and LLMs. The results show the current models still has substantial areas for improvement in spoken conversation, including fine-tuned models and LLMs, i.e., ChatGPT.

Abstract: Automated stance detection and related machine learning methods can provide useful insights for media monitoring and academic research. Many of these approaches require annotated training datasets, which limits their applicability for languages where these may not be readily available. This paper explores the applicability of large language models for automated stance detection in a challenging scenario, involving a morphologically complex, lower-resource language, and a socio-culturally complex topic, immigration. If the approach works in this case, it can be expected to perform as well or better in less demanding scenarios. We annotate a large set of pro and anti-immigration examples, and compare the performance of multiple language models as supervised learners. We also probe the usability of ChatGPT as an instructable zero-shot classifier for the same task. Supervised achieves acceptable performance, and ChatGPT yields similar accuracy. This is promising as a potentially simpler and cheaper alternative for text classification tasks, including in lower-resource languages. We further use the best-performing model to investigate diachronic trends over seven years in two corpora of Estonian mainstream and right-wing populist news sources, demonstrating the applicability of the approach for news analytics and media monitoring settings, and discuss correspondences between stance changes and real-world events.

Abstract: Transformers have revolutionized almost all natural language processing (NLP) tasks but suffer from memory and computational complexity that scales quadratically with sequence length. In contrast, recurrent neural networks (RNNs) exhibit linear scaling in memory and computational requirements but struggle to match the same performance as Transformers due to limitations in parallelization and scalability. We propose a novel model architecture, Receptance Weighted Key Value (RWKV), that combines the efficient parallelizable training of Transformers with the efficient inference of RNNs. Our approach leverages a linear attention mechanism and allows us to formulate the model as either a Transformer or an RNN, which parallelizes computations during training and maintains constant computational and memory complexity during inference, leading to the first non-transformer architecture to be scaled to tens of billions of parameters. Our experiments reveal that RWKV performs on par with similarly sized Transformers, suggesting that future work can leverage this architecture to create more efficient models. This work presents a significant step towards reconciling the trade-offs between computational efficiency and model performance in sequence processing tasks.

Abstract: Multi-label text classification (MLC) is a challenging task in settings of large label sets, where label support follows a Zipfian distribution. In this paper, we address this problem through retrieval augmentation, aiming to improve the sample efficiency of classification models. Our approach closely follows the standard MLC architecture of a Transformer-based encoder paired with a set of classification heads. In our case, however, the input document representation is augmented through cross-attention to similar documents retrieved from the training set and represented in a task-specific manner. We evaluate this approach on four datasets from the legal and biomedical domains, all of which feature highly skewed label distributions. Our experiments show that retrieval augmentation substantially improves model performance on the long tail of infrequent labels especially so for lower-resource training scenarios and more challenging long-document data scenarios.

Abstract: Large language models (LLMs) are becoming attractive as few-shot reasoners to solve NL-related tasks. However, there is still much to be learned about how well LLMs understand structured data, such as tables. While it is true that tables can be used as inputs to LLMs with serialization, there lack comprehensive studies examining whether LLMs can truly comprehend such data. In this paper we try to understand this by designing a benchmark to evaluate structural understanding capabilities (SUC) of LLMs. The benchmark we create includes seven tasks, each with their own unique challenges, e.g,, cell lookup, row retrieval and size detection. We run a series of evaluations on GPT-3 family models (e.g., text-davinci-003). We discover that the performance varied depending on a number of input choices, including table input format, content order, role prompting and partition marks. Drawing from the insights gained through the benchmark evaluations, we then propose self-augmentation for effective structural prompting, e.g., critical value / range identification using LLMs' internal knowledge. When combined with carefully chosen input choices, these structural prompting methods lead to promising improvements in LLM performance on a variety of tabular tasks, e.g., TabFact($\uparrow2.31\%$), HybridQA($\uparrow2.13\%$), SQA($\uparrow2.72\%$), Feverous($\uparrow0.84\%$), and ToTTo($\uparrow5.68\%$). We believe our benchmark and proposed prompting methods can serve as a simple yet generic selection for future research. The code and data are released in https://anonymous.4open.science/r/StructuredLLM-76F3.

Abstract: Biomedical named entity recognition is one of the core tasks in biomedical natural language processing (BioNLP). To tackle this task, numerous supervised/distantly supervised approaches have been proposed. Despite their remarkable success, these approaches inescapably demand laborious human effort. To alleviate the need of human effort, dictionary-based approaches have been proposed to extract named entities simply based on a given dictionary. However, one downside of existing dictionary-based approaches is that they are challenged to identify concept synonyms that are not listed in the given dictionary, which we refer as the synonym generalization problem. In this study, we propose a novel Synonym Generalization (SynGen) framework that recognizes the biomedical concepts contained in the input text using span-based predictions. In particular, SynGen introduces two regularization terms, namely, (1) a synonym distance regularizer; and (2) a noise perturbation regularizer, to minimize the synonym generalization error. To demonstrate the effectiveness of our approach, we provide a theoretical analysis of the bound of synonym generalization error. We extensively evaluate our approach on a wide range of benchmarks and the results verify that SynGen outperforms previous dictionary-based models by notable margins. Lastly, we provide a detailed analysis to further reveal the merits and inner-workings of our approach.

Abstract: Applying knowledge distillation encourages a student model to behave more like a teacher model, largely retaining the performance of the teacher model, even though the student model may have substantially fewer parameters. However, while distillation helps student models behave more like teacher models in-distribution, this is not necessarily the case out-of-distribution. To address this, we use a language model to create task-specific unlabeled data that mimics the data in targeted out-of-distribution domains. We use this generated data for knowledge distillation on the task of Natural Language Inference (NLI), encouraging the student models to behave more like the teacher models for these examples. Our domain-targeted augmentation is highly effective, and outperforms previous robustness methods when evaluating out-of-distribution performance on MNLI. Surprisingly, this method also improves performance on out-of-distribution domains that the data was not generated for. We additionally introduce Distilled Minority Upsampling (DMU), a method for identifying and upsampling minority examples during the distillation. DMU is complementary to the domain-targeted augmentation, and substantially improves performance on SNLI-hard. Finally, we show out-of-distribution improvements on HANS from both of our methods, despite augmenting the training data with fewer than 5k examples.

Abstract: Tools serve as pivotal interfaces that enable humans to understand and reshape the world. With the advent of foundational models, AI systems can utilize tools to expand their capabilities and interact with the world. Existing tool learning methodologies, encompassing supervised fine-tuning and prompt engineering approaches, often induce language models to utilize tools indiscriminately, as complex problems often exceed their own competencies. However, introducing tools for simple tasks, which the models themselves can readily resolve, can inadvertently propagate errors rather than enhance performance. This leads to the research question: can we teach language models when and how to use tools? To meet this need, we propose Tool leaRning wIth exeCution fEedback (TRICE), a two-stage end-to-end framework that enables the model to continually learn through feedback derived from tool execution, thereby learning when and how to use tools effectively. Experimental results, backed by further analysis, show that TRICE can make the language model to selectively use tools by decreasing the model's dependency on tools while enhancing the performance. Code and datasets will be available in https://github.com/zjunlp/trice.

Abstract: There are two types of approaches to solving cross-lingual transfer: multilingual pre-training implicitly aligns the hidden representations of different languages, while the translate-test explicitly translates different languages to an intermediate language, such as English. Translate-test has better interpretability compared to multilingual pre-training. However, the translate-test has lower performance than multilingual pre-training(Conneau and Lample, 2019; Conneau et al, 2020) and can't solve word-level tasks because translation rearranges the word order. Therefore, we propose a new Machine-created Universal Language (MUL) as a new intermediate language. MUL consists of a set of discrete symbols as universal vocabulary and NL-MUL translator for translating from multiple natural languages to MUL. MUL unifies common concepts from different languages into the same universal word for better cross-language transfer. And MUL preserves the language-specific words as well as word order, so the model can be easily applied to word-level tasks. Our experiments show that translating into MUL achieves better performance compared to multilingual pre-training, and our analyses show that MUL has good interpretability.

Abstract: Despite recent progress in text-to-SQL parsing, current semantic parsers are still not accurate enough for practical use. In this paper, we investigate how to build automatic text-to-SQL error correction models. Noticing that token-level edits are out of context and sometimes ambiguous, we propose building clause-level edit models instead. Besides, while most language models of code are not specifically pre-trained for SQL, they know common data structures and their operations in programming languages such as Python. Thus, we propose a novel representation for SQL queries and their edits that adheres more closely to the pre-training corpora of language models of code. Our error correction model improves the exact set match accuracy of different parsers by 2.4-6.5 and obtains up to 4.3 point absolute improvement over two strong baselines. Our code and data are available at https://github.com/OSU-NLP-Group/Auto-SQL-Correction.

Abstract: VoxML is a modeling language used to map natural language expressions into real-time visualizations using commonsense semantic knowledge of objects and events. Its utility has been demonstrated in embodied simulation environments and in agent-object interactions in situated multimodal human-agent collaboration and communication. It introduces the notion of object affordance (both Gibsonian and Telic) from HRI and robotics, as well as the concept of habitat (an object's context of use) for interactions between a rational agent and an object. This paper aims to specify VoxML as an annotation language in general abstract terms. It then shows how it works on annotating linguistic data that express visually perceptible human-object interactions. The annotation structures thus generated will be interpreted against the enriched minimal model created by VoxML as a modeling language while supporting the modeling purposes of VoxML linguistically.

Abstract: We consider the problem of few-shot spoken word classification in a setting where a model is incrementally introduced to new word classes. This would occur in a user-defined keyword system where new words can be added as the system is used. In such a continual learning scenario, a model might start to misclassify earlier words as newer classes are added, i.e. catastrophic forgetting. To address this, we propose an extension to model-agnostic meta-learning (MAML): each inner learning loop, where a model "learns how to learn'' new classes, ends with a single gradient update using stored templates from all the classes that the model has already seen (one template per class). We compare this method to OML (another extension of MAML) in few-shot isolated-word classification experiments on Google Commands and FACC. Our method consistently outperforms OML in experiments where the number of shots and the final number of classes are varied.

Abstract: While large models such as GPT-3 demonstrate exceptional performance in zeroshot and fewshot summarization tasks, their extensive serving and fine-tuning costs hinder their utilization in various applications. Conversely, previous studies have found that although automatic metrics tend to favor smaller fine-tuned models, the quality of the summaries they generate is inferior to that of larger models like GPT-3 when assessed by human evaluators. To address this issue, we propose InheritSumm, a versatile and compact summarization model derived from GPT-3.5 through distillation. InheritSumm not only exhibits comparable zeroshot and fewshot summarization capabilities to GPT-3.5 but is also sufficiently compact for fine-tuning purposes. Experimental results demonstrate that InheritSumm achieves similar or superior performance to GPT-3.5 in zeroshot and fewshot settings. Furthermore, it outperforms the previously established best small models in both prefix-tuning and full-data fine-tuning scenarios.

Abstract: This study investigates machine translation between related languages i.e., languages within the same family that share similar linguistic traits such as word order and lexical similarity. Machine translation through few-shot prompting leverages a small set of translation pair examples to generate translations for test sentences. This requires the model to learn how to generate translations while simultaneously ensuring that token ordering is maintained to produce a fluent and accurate translation. We propose that for related languages, the task of machine translation can be simplified by leveraging the monotonic alignment characteristic of such languages. We introduce a novel approach of few-shot prompting that decomposes the translation process into a sequence of word chunk translations. Through evaluations conducted on multiple related language pairs across various language families, we demonstrate that our novel approach of decomposed prompting surpasses multiple established few-shot baseline models, thereby verifying its effectiveness. For example, our model outperforms the strong few-shot prompting BLOOM model with an average improvement of 4.2 chrF++ scores across the examined languages.

Abstract: Query-focused summarization (QFS) aims to extract or generate a summary of an input document that directly answers or is relevant to a given query. The lack of large-scale datasets in the form of documents, queries, and summaries has hindered model development in this area. In contrast, multiple large-scale high-quality datasets for generic summarization exist. We hypothesize that there is a hidden query for each summary sentence in a generic summarization annotation, and we utilize a large-scale pretrained language model to recover it. In this way, we convert four generic summarization benchmarks into a new QFS benchmark dataset, LMGQS, which consists of over 1 million document-query-summary samples. We thoroughly investigate the properties of our proposed dataset and establish baselines with state-of-the-art summarization models. By fine-tuning a language model on LMGQS, we achieve state-of-the-art zero-shot and supervised performance on multiple existing QFS benchmarks, demonstrating the high quality and diversity of LMGQS.

Abstract: The abundance of annotated data in natural language processing (NLP) poses both opportunities and challenges. While it enables the development of high-performing models for a variety of tasks, it also poses the risk of models learning harmful biases from the data, such as gender stereotypes. In this work, we investigate the role of attention, a widely-used technique in current state-of-the-art NLP models, in the propagation of social biases. Specifically, we study the relationship between the entropy of the attention distribution and the model's performance and fairness. We then propose a novel method for modulating attention weights to improve model fairness after training. Since our method is only applied post-training and pre-inference, it is an intra-processing method and is, therefore, less computationally expensive than existing in-processing and pre-processing approaches. Our results show an increase in fairness and minimal performance loss on different text classification and generation tasks using language models of varying sizes. WARNING: This work uses language that is offensive.

Abstract: Human evaluations are often required for abstractive summary evaluations to give fairer judgments. However, they are often time-consuming, costly, inconsistent, and non-reproducible. To overcome these challenges, we explore the potential of using an out-of-the-box LLM (i.e. "gpt-3.5-turbo") for summarization evaluation without manually selecting demonstrations or complex prompt tuning. We compare different evaluation methods, including 2 methods for Likert-scale scoring and 1 method for head-to-head comparisons, to investigate the performance of the LLM as a zero-shot evaluator. We further propose a meta-correlation metric to measure the stability of the LLM's evaluation capability. With extensive experiments, we show that certain prompt formats can produce better results than others. We also bring attention to the LLM's deteriorating evaluation capability with the rising qualities of summaries. In addition, we find that the LLM's evaluation capability also depends on the evaluated dimensions. We discuss the pros and cons of each method, make recommendations, and suggest some future directions for improvement.

Abstract: Meta-learning and few-shot prompting are viable methods to induce certain types of compositional behaviour. However, these methods can be very sensitive to the choice of support examples used. Choosing good supports from the training data for a given test query is already a difficult problem, but in some cases solving this may not even be enough. We consider a grounded language learning problem (gSCAN) where good support examples for certain test splits might not even exist in the training data, or would be infeasible to search for. We design an agent which instead generates possible supports which are relevant to the test query and current state of the world, then uses these supports via meta-learning to solve the test query. We show substantially improved performance on a previously unsolved compositional behaviour split without a loss of performance on other splits. Further experiments show that in this case, searching for relevant demonstrations even with an oracle function is not sufficient to attain good performance when using meta-learning.

Abstract: The recent success of large language models (LLMs) has shown great potential to develop more powerful conversational recommender systems (CRSs), which rely on natural language conversations to satisfy user needs. In this paper, we embark on an investigation into the utilization of ChatGPT for conversational recommendation, revealing the inadequacy of the existing evaluation protocol. It might over-emphasize the matching with the ground-truth items or utterances generated by human annotators, while neglecting the interactive nature of being a capable CRS. To overcome the limitation, we further propose an interactive Evaluation approach based on LLMs named iEvaLM that harnesses LLM-based user simulators. Our evaluation approach can simulate various interaction scenarios between users and systems. Through the experiments on two publicly available CRS datasets, we demonstrate notable improvements compared to the prevailing evaluation protocol. Furthermore, we emphasize the evaluation of explainability, and ChatGPT showcases persuasive explanation generation for its recommendations. Our study contributes to a deeper comprehension of the untapped potential of LLMs for CRSs and provides a more flexible and easy-to-use evaluation framework for future research endeavors. The codes and data are publicly available at https://github.com/RUCAIBox/iEvaLM-CRS.

Abstract: Existing datasets for automated fact-checking have substantial limitations, such as relying on artificial claims, lacking annotations for evidence and intermediate reasoning, or including evidence published after the claim. In this paper we introduce AVeriTeC, a new dataset of 4,568 real-world claims covering fact-checks by 50 different organizations. Each claim is annotated with question-answer pairs supported by evidence available online, as well as textual justifications explaining how the evidence combines to produce a verdict. Through a multi-round annotation process, we avoid common pitfalls including context dependence, evidence insufficiency, and temporal leakage, and reach a substantial inter-annotator agreement of $\kappa=0.619$ on verdicts. We develop a baseline as well as an evaluation scheme for verifying claims through several question-answering steps against the open web.

Abstract: Word sense disambiguation (WSD), which aims to determine an appropriate sense for a target word given its context, is crucial for natural language understanding. Existing supervised methods treat WSD as a classification task and have achieved remarkable performance. However, they ignore uncertainty estimation (UE) in the real-world setting, where the data is always noisy and out of distribution. This paper extensively studies UE on the benchmark designed for WSD. Specifically, we first compare four uncertainty scores for a state-of-the-art WSD model and verify that the conventional predictive probabilities obtained at the end of the model are inadequate to quantify uncertainty. Then, we examine the capability of capturing data and model uncertainties by the model with the selected UE score on well-designed test scenarios and discover that the model reflects data uncertainty satisfactorily but underestimates model uncertainty. Furthermore, we explore numerous lexical properties that intrinsically affect data uncertainty and provide a detailed analysis of four critical aspects: the syntactic category, morphology, sense granularity, and semantic relations.

Abstract: Motivation: Named Entity Recognition (NER) is a key task to support biomedical research. In Biomedical Named Entity Recognition (BioNER), obtaining high-quality expert annotated data is laborious and expensive, leading to the development of automatic approaches such as distant supervision. However, manually and automatically generated data often suffer from the unlabeled entity problem, whereby many entity annotations are missing, degrading the performance of full annotation NER models. Results: To address this problem, we systematically study the effectiveness of partial annotation learning methods for biomedical entity recognition over different simulated scenarios of missing entity annotations. Furthermore, we propose a TS-PubMedBERT-Partial-CRF partial annotation learning model. We harmonize 15 biomedical NER corpora encompassing five entity types to serve as a gold standard and compare against two commonly used partial annotation learning models, BiLSTM-Partial-CRF and EER-PubMedBERT, and the state-of-the-art full annotation learning BioNER model PubMedBERT tagger. Results show that partial annotation learning-based methods can effectively learn from biomedical corpora with missing entity annotations. Our proposed model outperforms alternatives and, specifically, the PubMedBERT tagger by 38% in F1-score under high missing entity rates. The recall of entity mentions in our model is also competitive with the upper bound on the fully annotated dataset.

Abstract: The increasing adoption of text-to-speech technologies has led to a growing demand for natural and emotive voices that adapt to a conversation's context and emotional tone. This need is particularly relevant for interactive narrative-driven systems such as video games, TV shows, and graphic novels. To address this need, we present the Emotive Narrative Storytelling (EMNS) corpus, a dataset of high-quality British English speech with labelled utterances designed to enhance interactive experiences with dynamic and expressive language. We provide high-quality clean audio recordings and natural language description pairs with transcripts and user-reviewed and self-reported labels for features such as word emphasis, expressiveness, and emotion labels. EMNS improves on existing datasets by providing higher quality and clean recording to aid more natural and expressive speech synthesis techniques for interactive narrative-driven experiences. Additionally, we release our remote and scalable data collection system to the research community.

Abstract: Multilingual understanding models (or encoder-based), pre-trained via masked language modeling, have achieved promising results on many language understanding tasks (e.g., mBERT). However, these non-autoregressive (NAR) models still struggle to generate high-quality texts compared with autoregressive (AR) models. Considering that encoder-based models have the advantage of efficient generation and self-correction abilities, this paper explores methods to empower multilingual understanding models the generation abilities to get a unified model. Specifically, we start from a multilingual encoder (XLM-R) and propose a \textbf{S}emantic-\textbf{G}uided \textbf{A}lignment-then-Denoising (SGA) approach to adapt an encoder to a multilingual generator with a small number of new parameters. Experiments show that the proposed approach is an effective adaption method, outperforming widely-used initialization-based methods with gains of 9.4 BLEU on machine translation, 8.1 Rouge-L on question generation, and 5.5 METEOR on story generation on XLM-R$_{large}$. On the other hand, we observe that XLM-R is still inferior to mBART in supervised settings despite better results on zero-shot settings, indicating that more exploration is required to make understanding models strong generators.

Abstract: Distantly supervised named entity recognition (DS-NER) has been proposed to exploit the automatically labeled training data instead of human annotations. The distantly annotated datasets are often noisy and contain a considerable number of false negatives. The recent approach uses a weighted sampling approach to select a subset of negative samples for training. However, it requires a good classifier to assign weights to the negative samples. In this paper, we propose a simple and straightforward approach for selecting the top negative samples that have high similarities with all the positive samples for training. Our method achieves consistent performance improvements on four distantly supervised NER datasets. Our analysis also shows that it is critical to differentiate the true negatives from the false negatives.

Abstract: Large language models (LLMs) have revolutionized the field of AI, demonstrating unprecedented capacity across various tasks. However, the inference process for LLMs comes with significant computational costs. In this paper, we propose an efficient LLM inference pipeline that harnesses the power of LLMs. Our approach begins by tapping into the potential of LLMs to accurately perceive and predict the response length with minimal overhead. By leveraging this information, we introduce an efficient sequence scheduling technique that groups queries with similar response lengths into micro-batches. We evaluate our approach on real-world instruction datasets using the LLaMA-based model, and our results demonstrate an impressive 86% improvement in inference throughput without compromising effectiveness. Notably, our method is orthogonal to other inference acceleration techniques, making it a valuable addition to many existing toolkits (e.g., FlashAttention, Quantization) for LLM inference.

Abstract: We explore testing the reasoning ability of large language models (LLMs), such as ChatGPT, by engaging with them in a debate-like conversation that probes deeper into their understanding of the subject. Specifically, we formulate a new task where given a question, the LLM can generate a correct solution while the user believes in a wrong solution in the beginning, and they need to discuss to make the correct decision through dialogue. Such a setting requires the LLM to not only achieve the correct answer on its own (which could be done by shallow memorization), but also be able to defend the truth instead of blindly believing or getting misled by the user's (invalid) arguments and critiques, thus testing in greater depth whether the LLM grasps the essence of the reasoning required to solve the problem. To automate this evaluation framework and save human labor, we simulate the user using another LLM conditioned on a synthesized wrong solution. Across a range of complex reasoning benchmarks spanning math, commonsense, logic and tasks from BIG-Bench, we find that despite being able to generate correct step-by-step solutions in the beginning, ChatGPT cannot maintain its belief in truth for a significant portion of examples when challenged by often-time absurdly invalid arguments. Our work reveals LLMs' weaknesses not captured by conventional benchmarking, and also points to danger zones of aligning models with human feedback.

Abstract: This paper presents an exhaustive quantitative and qualitative evaluation of Large Language Models (LLMs) for Knowledge Graph (KG) construction and reasoning. We employ eight distinct datasets that encompass aspects including entity, relation and event extraction, link prediction, and question answering. Empirically, our findings suggest that GPT-4 outperforms ChatGPT in the majority of tasks and even surpasses fine-tuned models in certain reasoning and question-answering datasets. Moreover, our investigation extends to the potential generalization ability of LLMs for information extraction, which culminates in the presentation of the Virtual Knowledge Extraction task and the development of the VINE dataset. Drawing on these empirical findings, we further propose AutoKG, a multi-agent-based approach employing LLMs for KG construction and reasoning, which aims to chart the future of this field and offer exciting opportunities for advancement. We anticipate that our research can provide invaluable insights for future undertakings of KG\footnote{Code and datasets will be available in https://github.com/zjunlp/AutoKG.

Abstract: Pretraining is the preliminary and fundamental step in developing capable language models (LM). Despite this, pretraining data design is critically under-documented and often guided by empirically unsupported intuitions. To address this, we pretrain 28 1.5B parameter decoder-only models, training on data curated (1) at different times, (2) with varying toxicity and quality filters, and (3) with different domain compositions. First, we quantify the effect of pretraining data age. A temporal shift between evaluation data and pretraining data leads to performance degradation, which is not overcome by finetuning. Second, we explore the effect of quality and toxicity filters, showing a trade-off between performance on standard benchmarks and risk of toxic generations. Our findings indicate there does not exist a one-size-fits-all solution to filtering training data. We also find that the effects of different types of filtering are not predictable from text domain characteristics. Lastly, we empirically validate that the inclusion of heterogeneous data sources, like books and web, is broadly beneficial and warrants greater prioritization. These findings constitute the largest set of experiments to validate, quantify, and expose many undocumented intuitions about text pretraining, which we hope will help support more informed data-centric decisions in LM development.

Abstract: Recent advancements in deep learning have precipitated the emergence of large language models (LLMs) which exhibit an impressive aptitude for understanding and producing text akin to human language. Despite the ability to train highly capable LLMs, the methodology for maintaining their relevancy and rectifying errors remains elusive. To that end, the past few years have witnessed a surge in techniques for editing LLMs, the objective of which is to alter the behavior of LLMs within a specific domain without negatively impacting performance across other inputs. This paper embarks on a deep exploration of the problems, methods, and opportunities relating to model editing for LLMs. In particular, we provide an exhaustive overview of the task definition and challenges associated with model editing, along with an in-depth empirical analysis of the most progressive methods currently at our disposal. We also build a new benchmark dataset to facilitate a more robust evaluation and pinpoint enduring issues intrinsic to existing techniques. Our objective is to provide valuable insights into the effectiveness and feasibility of each model editing technique, thereby assisting the research community in making informed decisions when choosing the most appropriate method for a specific task or context. Code and datasets will be available at https://github.com/zjunlp/EasyEdit.

Abstract: This study employs Independent Component Analysis (ICA) to uncover universal properties of embeddings of words or images. Our approach extracts independent semantic components of embeddings, enabling each embedding to be represented as a composition of intrinsic interpretable axes. We demonstrate that embeddings can be expressed as a combination of a few axes and that these semantic axes are consistent across different languages, modalities, and embedding algorithms. This discovery of universal properties in embeddings contributes to model interpretability, potentially facilitating the development of highly interpretable models and the compression of large-scale models.

Abstract: Recent research on transformer-based language models investigates their reasoning ability over logical rules expressed in natural language text. However, their logic is not yet well-understood as we cannot explain the abstractions made by the models that help them in reasoning. These models are criticized for merely memorizing complex patterns in the data, which often creates issues for their generalizability in unobserved situations. In this work, we analyze the use of probabilistic logical rules in transformer-based language models. In particular, we propose a new approach, Probabilistic Constraint Training (PCT), that explicitly models probabilistic logical reasoning by imposing the rules of reasoning as constraints during training. We create a new QA benchmark for evaluating probabilistic reasoning over uncertain textual rules, which creates instance-specific rules, unlike the only existing relevant benchmark. Experimental results show that our proposed technique improves the base language models' accuracy and explainability when probabilistic logical reasoning is required for question answering. Moreover, we show that the learned probabilistic reasoning abilities are transferable to novel situations.

Abstract: Scientific fact-checking is crucial for ensuring the accuracy, reliability, and trustworthiness of scientific claims. However, existing benchmarks are limited in terms of their claim diversity, reliance on text-based evidence, and oversimplification of scientific reasoning. To address these gaps, we introduce SCITAB, a novel dataset comprising 1,225 challenging scientific claims requiring compositional reasoning with scientific tables. The claims in SCITAB are derived from the actual scientific statements, and the evidence is presented as tables, closely mirroring real-world fact-checking scenarios. We establish benchmarks on SCITAB using state-of-the-art models, revealing its inherent difficulty and highlighting limitations in existing prompting methods. Our error analysis identifies unique challenges, including ambiguous expressions and irrelevant claims, suggesting future research directions. The code and the data are publicly available at https://github.com/XinyuanLu00/SciTab.

Abstract: Training a Named Entity Recognition (NER) model often involves fixing a taxonomy of entity types. However, requirements evolve and we might need the NER model to recognize additional entity types. A simple approach is to re-annotate entire dataset with both existing and additional entity types and then train the model on the re-annotated dataset. However, this is an extremely laborious task. To remedy this, we propose a novel approach called Partial Label Model (PLM) that uses only partially annotated datasets. We experiment with 6 diverse datasets and show that PLM consistently performs better than most other approaches (0.5 - 2.5 F1), including in novel settings for taxonomy expansion not considered in prior work. The gap between PLM and all other approaches is especially large in settings where there is limited data available for the additional entity types (as much as 11 F1), thus suggesting a more cost effective approaches to taxonomy expansion.

Abstract: This paper aims to improve contrastive learning for sentence embeddings from two perspectives: handling dropout noise and addressing feature corruption. Specifically, for the first perspective, we identify that the dropout noise from negative pairs affects the model's performance. Therefore, we propose a simple yet effective method to deal with such type of noise. Secondly, we pinpoint the rank bottleneck of current solutions to feature corruption and propose a dimension-wise contrastive learning objective to address this issue. Both proposed methods are generic and can be applied to any contrastive learning based models for sentence embeddings. Experimental results on standard benchmarks demonstrate that combining both proposed methods leads to a gain of 1.8 points compared to the strong baseline SimCSE configured with BERT base. Furthermore, applying the proposed method to DiffCSE, another strong contrastive learning based baseline, results in a gain of 1.4 points.

Abstract: Reliable automatic evaluation of summarization systems is challenging due to the multifaceted and subjective nature of the task. This is especially the case for languages other than English, where human evaluations are scarce. In this work, we introduce SEAHORSE, a dataset for multilingual, multifaceted summarization evaluation. SEAHORSE consists of 96K summaries with human ratings along 6 quality dimensions: comprehensibility, repetition, grammar, attribution, main ideas, and conciseness, covering 6 languages, 9 systems and 4 datasets. As a result of its size and scope, SEAHORSE can serve both as a benchmark to evaluate learnt metrics, as well as a large-scale resource for training such metrics. We show that metrics trained with SEAHORSE achieve strong performance on the out-of-domain meta-evaluation benchmarks TRUE (Honovich et al., 2022) and mFACE (Aharoni et al., 2022). We make SEAHORSE publicly available for future research on multilingual and multifaceted summarization evaluation.

Abstract: We introduce a multilingual extension of the HOLISTICBIAS dataset, the largest English template-based taxonomy of textual people references: MULTILINGUALHOLISTICBIAS. This extension consists of 20,459 sentences in 50 languages distributed across all 13 demographic axes. Source sentences are built from combinations of 118 demographic descriptors and three patterns, excluding nonsensical combinations. Multilingual translations include alternatives for gendered languages that cover gendered translations when there is ambiguity in English. Our benchmark is intended to uncover demographic imbalances and be the tool to quantify mitigations towards them. Our initial findings show that translation quality for EN-to-XX translations is an average of 8 spBLEU better when evaluating with the masculine human reference compared to feminine. In the opposite direction, XX-to-EN, we compare the robustness of the model when the source input only differs in gender (masculine or feminine) and masculine translations are an average of almost 4 spBLEU better than feminine. When embedding sentences to a joint multilingual sentence representations space, we find that for most languages masculine translations are significantly closer to the English neutral sentences when embedded.

Abstract: Most existing task-oriented dialog (TOD) systems track dialog states in terms of slots and values and use them to query a database to get relevant knowledge to generate responses. In real-life applications, user utterances are noisier, and thus it is more difficult to accurately track dialog states and correctly secure relevant knowledge. Recently, a progress in question answering and document-grounded dialog systems is retrieval-augmented methods with a knowledge retriever. Inspired by such progress, we propose a retrieval-based method to enhance knowledge selection in TOD systems, which significantly outperforms the traditional database query method for real-life dialogs. Further, we develop latent variable model based semi-supervised learning, which can work with the knowledge retriever to leverage both labeled and unlabeled dialog data. Joint Stochastic Approximation (JSA) algorithm is employed for semi-supervised model training, and the whole system is referred to as that JSA-KRTOD. Experiments are conducted on a real-life dataset from China Mobile Custom-Service, called MobileCS, and show that JSA-KRTOD achieves superior performances in both labeled-only and semi-supervised settings.

Abstract: To translate speech for automatic dubbing, machine translation needs to be isochronous, i.e. translated speech needs to be aligned with the source in terms of speech durations. We introduce target factors in a transformer model to predict durations jointly with target language phoneme sequences. We also introduce auxiliary counters to help the decoder to keep track of the timing information while generating target phonemes. We show that our model improves translation quality and isochrony compared to previous work where the translation model is instead trained to predict interleaved sequences of phonemes and durations.

Abstract: State-of-the-art neural models can now reach human performance levels across various natural language understanding tasks. However, despite this impressive performance, models are known to learn from annotation artefacts at the expense of the underlying task. While interpretability methods can identify influential features for each prediction, there are no guarantees that these features are responsible for the model decisions. Instead, we introduce a model-agnostic logical framework to determine the specific information in an input responsible for each model decision. This method creates interpretable Natural Language Inference (NLI) models that maintain their predictive power. We achieve this by generating facts that decompose complex NLI observations into individual logical atoms. Our model makes predictions for each atom and uses logical rules to decide the class of the observation based on the predictions for each atom. We apply our method to the highly challenging ANLI dataset, where our framework improves the performance of both a DeBERTa-base and BERT baseline. Our method performs best on the most challenging examples, achieving a new state-of-the-art for the ANLI round 3 test set. We outperform every baseline in a reduced-data setting, and despite using no annotations for the generated facts, our model predictions for individual facts align with human expectations.

Abstract: ChatGPT and GPT-4 have attracted substantial interest from both academic and industrial circles, owing to their remarkable few-shot (or even zero-shot) ability to handle various tasks. Recent work shows that, after being fine-tuned with a few sets of instruction-driven data, the recently proposed LLM, LLaMa, exhibits an impressive capability to address a broad range of tasks. However, the zero-shot performance of LLMs does not consistently outperform that of models fined-tuned for specific scenarios. To explore whether the capabilities of LLMs can be further enhanced for specific scenarios, we choose the writing-assistance scenario as the testbed, including seven writing tasks. We collect training data for these tasks, reframe them in an instruction-following format, and subsequently refine LLaMa via instruction tuning. Experimental results show that continually fine-tuning LLaMa on writing instruction data significantly improves its ability on writing tasks. We also conduct more experiments and analyses to offer insights for future work on effectively fine-tuning LLaMa for specific scenarios.

Abstract: Explaining the decisions of neural models is crucial for ensuring their trustworthiness at deployment time. Using Natural Language Explanations (NLEs) to justify a model's predictions has recently gained increasing interest. However, this approach usually demands large datasets of human-written NLEs for the ground-truth answers, which are expensive and potentially infeasible for some applications. For models to generate high-quality NLEs when only a few NLEs are available, the fine-tuning of Pre-trained Language Models (PLMs) in conjunction with prompt-based learning recently emerged. However, PLMs typically have billions of parameters, making fine-tuning expensive. We propose SparseFit, a sparse few-shot fine-tuning strategy that leverages discrete prompts to jointly generate predictions and NLEs. We experiment with SparseFit on the T5 model and four datasets and compare it against state-of-the-art parameter-efficient fine-tuning techniques. We perform automatic and human evaluations to assess the quality of the model-generated NLEs, finding that fine-tuning only 6.8% of the model parameters leads to competitive results for both the task performance and the quality of the NLEs.

Abstract: Recent advances in large language models have enabled them to reach a level of text generation comparable to that of humans. These models show powerful capabilities across a wide range of content, including news article writing, story generation, and scientific writing. Such capability further narrows the gap between human-authored and machine-generated texts, highlighting the importance of deepfake text detection to avoid potential risks such as fake news propagation and plagiarism. However, previous work has been limited in that they testify methods on testbed of specific domains or certain language models. In practical scenarios, the detector faces texts from various domains or LLMs without knowing their sources. To this end, we build a wild testbed by gathering texts from various human writings and deepfake texts generated by different LLMs. Human annotators are only slightly better than random guessing at identifying machine-generated texts. Empirical results on automatic detection methods further showcase the challenges of deepfake text detection in a wild testbed. In addition, out-of-distribution poses a greater challenge for a detector to be employed in realistic application scenarios. We release our resources at https://github.com/yafuly/DeepfakeTextDetect.

Abstract: Multi-query attention (MQA), which only uses a single key-value head, drastically speeds up decoder inference. However, MQA can lead to quality degradation, and moreover it may not be desirable to train a separate model just for faster inference. We (1) propose a recipe for uptraining existing multi-head language model checkpoints into models with MQA using 5% of original pre-training compute, and (2) introduce grouped-query attention (GQA), a generalization of multi-query attention which uses an intermediate (more than one, less than number of query heads) number of key-value heads. We show that uptrained GQA achieves quality close to multi-head attention with comparable speed to MQA.

Abstract: Interactive Natural Language Processing (iNLP) has emerged as a novel paradigm within the field of NLP, aimed at addressing limitations in existing frameworks while aligning with the ultimate goals of artificial intelligence. This paradigm considers language models as agents capable of observing, acting, and receiving feedback iteratively from external entities. Specifically, language models in this context can: (1) interact with humans for better understanding and addressing user needs, personalizing responses, aligning with human values, and improving the overall user experience; (2) interact with knowledge bases for enriching language representations with factual knowledge, enhancing the contextual relevance of responses, and dynamically leveraging external information to generate more accurate and informed responses; (3) interact with models and tools for effectively decomposing and addressing complex tasks, leveraging specialized expertise for specific subtasks, and fostering the simulation of social behaviors; and (4) interact with environments for learning grounded representations of language, and effectively tackling embodied tasks such as reasoning, planning, and decision-making in response to environmental observations. This paper offers a comprehensive survey of iNLP, starting by proposing a unified definition and framework of the concept. We then provide a systematic classification of iNLP, dissecting its various components, including interactive objects, interaction interfaces, and interaction methods. We proceed to delve into the evaluation methodologies used in the field, explore its diverse applications, scrutinize its ethical and safety issues, and discuss prospective research directions. This survey serves as an entry point for researchers who are interested in this rapidly evolving area and offers a broad view of the current landscape and future trajectory of iNLP.

Abstract: Large Language Models (LLMs) may hallucinate and generate fake information, despite pre-training on factual data. Inspired by the journalistic device of "according to sources", we propose according-to prompting: directing LLMs to ground responses against previously observed text. To quantify this grounding, we propose a novel evaluation metric (QUIP-Score) that measures the extent to which model-produced answers are directly found in underlying text corpora. We illustrate with experiments on Wikipedia that these prompts improve grounding under our metrics, with the additional benefit of often improving end-task performance. Furthermore, prompts that ask the model to decrease grounding (or to ground to other corpora) decrease grounding, indicating the ability of language models to increase or decrease grounded generations on request.

Abstract: Recent work in NLP has shown promising results in training models on large amounts of tasks to achieve better generalization. However, it is not well-understood how tasks are related, and how helpful training tasks can be chosen for a new task. In this work, we investigate whether knowing task relationships via pairwise task transfer improves choosing one or more source tasks that help to learn a new target task. We provide TaskWeb, a large-scale benchmark of pairwise task transfers for 22 NLP tasks using three different model types, sizes, and adaptation methods, spanning about 25,000 experiments. Then, we design a new method TaskShop based on our analysis of TaskWeb. TaskShop uses TaskWeb to estimate the benefit of using a source task for learning a new target, and to choose a subset of helpful training tasks for multi-task learning. Our method improves overall rankings and top-k precision of source tasks by 12% and 29%, respectively. We also use TaskShop to build smaller multi-task training sets that improve zero-shot performances across 11 different target tasks by at least 4.3%.

Abstract: Prompting is now a dominant method for evaluating the linguistic knowledge of large language models (LLMs). While other methods directly read out models' probability distributions over strings, prompting requires models to access this internal information by processing linguistic input, thereby implicitly testing a new type of emergent ability: metalinguistic judgment. In this study, we compare metalinguistic prompting and direct probability measurements as ways of measuring models' knowledge of English. Broadly, we find that LLMs' metalinguistic judgments are inferior to quantities directly derived from representations. Furthermore, consistency gets worse as the prompt diverges from direct measurements of next-word probabilities. Our findings suggest that negative results relying on metalinguistic prompts cannot be taken as conclusive evidence that an LLM lacks a particular linguistic competence. Our results also highlight the lost value with the move to closed APIs where access to probability distributions is limited.

Abstract: Pre-trained visual language models (VLM) have shown excellent performance in image caption tasks. However, it sometimes shows insufficient reasoning ability. In contrast, large language models (LLMs) emerge with powerful reasoning capabilities. Therefore, we propose a method called TReE, which transfers the reasoning ability of a large language model to a visual language model in zero-shot scenarios. TReE contains three stages: observation, thinking, and re-thinking. Observation stage indicates that VLM obtains the overall information of the relative image. Thinking stage combines the image information and task description as the prompt of the LLM, inference with the rationals. Re-Thinking stage learns from rationale and then inference the final result through VLM.

Abstract: We introduce Chain of Knowledge (CoK), a framework that augments large language models with structured knowledge bases to improve factual correctness and reduce hallucination. Compared to previous works which only retrieve unstructured texts, CoK leverages structured knowledge bases which support complex queries and offer more direct factual statements. To assist large language models to effectively query knowledge bases, we propose a query generator model with contrastive instruction-tuning. As the query generator is separate from the frozen large language model, our framework is modular and thus easily adapted to various knowledge sources and models. Experiments show that our framework significantly enhances the factual correctness of large language models on knowledge-intensive tasks.

Abstract: We present a design framework called Conversational Learning with Analytical Step-by-Step Strategies (CLASS) for developing high-performance Intelligent Tutoring Systems (ITS). The CLASS framework aims to empower ITS with with two critical capabilities: imparting tutor-like step-by-step guidance and enabling tutor-like conversations in natural language to effectively engage learners. To empower ITS with the aforementioned capabilities, the CLASS framework employs two carefully curated synthetic datasets. The first scaffolding dataset encompasses a variety of elements, including problems, their corresponding subproblems, hints, incorrect solutions, and tailored feedback. This dataset provides ITS with essential problem-solving strategies necessary for guiding students through each step of the conversation. The second conversational dataset contains simulated student-tutor conversations that involve the application of problem-solving strategies learned from the first dataset. In the second dataset, the tutoring system adheres to a pre-defined response template, which helps to maintain consistency and structure in ITS's responses during its interactions. This structured methodology facilitates seamless integration of user feedback and yields valuable insights into ITS's internal decision-making process, allowing for continuous refinement and improvement of the system. We also present a proof-of-concept ITS, referred to as SPOCK, trained using the CLASS framework with a focus on college level introductory biology content. A carefully constructed protocol was developed for SPOCK's preliminary evaluation, examining aspects such as the factual accuracy and relevance of its responses. Experts in the field of biology offered favorable remarks, particularly highlighting SPOCK's capability to break down questions into manageable subproblems and provide step-by-step guidance to students.

Abstract: Hate speech is a severe issue that affects many online platforms. So far, several studies have been performed to develop robust hate speech detection systems. Large language models like ChatGPT have recently shown great potential in performing several tasks, including hate speech detection. However, it is crucial to comprehend the limitations of these models to build more robust hate speech detection systems. Thus to bridge the gap, our study aims to evaluate the weaknesses of the ChatGPT model in detecting hate speech at a granular level across 11 languages. In addition, we investigate the influence of complex emotions, such as the use of emojis in hate speech, on the performance of the ChatGPT model. Through our analysis, we examine and investigate the errors made by the model, shedding light on its shortcomings in detecting certain types of hate speech and highlighting the need for further research and improvements in hate speech detection.

Abstract: A prominent weakness of modern language models (LMs) is their tendency to generate factually incorrect text, which hinders their usability. A natural question is whether such factual errors can be detected automatically. Inspired by truth-seeking mechanisms in law, we propose a factuality evaluation framework for LMs that is based on cross-examination. Our key idea is that an incorrect claim is likely to result in inconsistency with other claims that the model generates. To discover such inconsistencies, we facilitate a multi-turn interaction between the LM that generated the claim and another LM (acting as an examiner) which introduces questions to discover inconsistencies. We empirically evaluate our method on factual claims made by multiple recent LMs on four benchmarks, finding that it outperforms existing methods and baselines, often by a large gap. Our results demonstrate the potential of using interacting LMs for capturing factual errors.

Abstract: Out-of-distribution (OOD) detection is a critical task for reliable predictions over text. Fine-tuning with pre-trained language models has been a de facto procedure to derive OOD detectors with respect to in-distribution (ID) data. Despite its common use, the understanding of the role of fine-tuning and its necessity for OOD detection is largely unexplored. In this paper, we raise the question: is fine-tuning necessary for OOD detection? We present a study investigating the efficacy of directly leveraging pre-trained language models for OOD detection, without any model fine-tuning on the ID data. We compare the approach with several competitive fine-tuning objectives, and offer new insights under various types of distributional shifts. Extensive evaluations on 8 diverse ID-OOD dataset pairs demonstrate near-perfect OOD detection performance (with 0% FPR95 in many cases), strongly outperforming its fine-tuned counterparts. We show that using distance-based detection methods, pre-trained language models are near-perfect OOD detectors when the distribution shift involves a domain change. Furthermore, we study the effect of fine-tuning on OOD detection and identify how to balance ID accuracy with OOD detection performance. Our code is publically available at https://github.com/Uppaal/lm-ood.

Abstract: Multilingual large language models (MLLMs) are jointly trained on data from many different languages such that representation of individual languages can benefit from other languages' data. Impressive performance on zero-shot cross-lingual transfer shows that these models are capable of exploiting data from other languages. Yet, it remains unclear to what extent, and under which conditions, languages rely on each other's data. In this study, we use TracIn (Pruthi et al., 2020), a training data attribution (TDA) method, to retrieve the most influential training samples seen during multilingual fine-tuning for a particular test language. This allows us to analyse cross-lingual sharing mechanisms of MLLMs from a new perspective. While previous work studied cross-lingual sharing at the level of model parameters, we present the first approach to study cross-lingual sharing at the data level. We find that MLLMs rely on data from multiple languages from the early stages of fine-tuning and that this reliance gradually increases as fine-tuning progresses. We further study how different fine-tuning languages influence model performance on a given test language and find that they can both reinforce and complement the knowledge acquired from data of the test language itself.

Abstract: In this paper, we introduce Parallel Attention and Feed-Forward Net Design (PAF) for transformer models. Transformer models are indisputably the backbone of all Natural Language Processing applications. Therefore, any efforts aimed at improving their efficiency are guaranteed to have an enormous impact. Transformer models consist of many layers and each layer has an attention block followed by a feed-forward network (FFN) that processes the input based on the attention block's output. We refer to this standard design as Series Attention and Feed-Forward Net Design (SAF). For each layer in our proposed PAF design for transformer models, we make FFN block's computations independent of the output of the attention block. This decoupling allows FFN block of each layer to run in parallel to the attention block of that layer. We evaluate PAF design by training two large language models (RoBERTa-large and bert-large-uncased) and comparing them to their SAF counterparts on six tasks of the General Language Understanding (GLUE) benchmark which test a multitude of semantic attributes. PAF models achieves nearly identical performance as their SAF counterparts on all the six tasks. We also compare time complexities of attention blocks with FFN blocks and find that running both blocks in parallel can theoretically and in practice achieve upto 1.5x to 2x gains in speed. We leave the development of fast and efficient libraries for implementation of PAF design for future work.

Abstract: In this paper, we propose DiffusionNER, which formulates the named entity recognition task as a boundary-denoising diffusion process and thus generates named entities from noisy spans. During training, DiffusionNER gradually adds noises to the golden entity boundaries by a fixed forward diffusion process and learns a reverse diffusion process to recover the entity boundaries. In inference, DiffusionNER first randomly samples some noisy spans from a standard Gaussian distribution and then generates the named entities by denoising them with the learned reverse diffusion process. The proposed boundary-denoising diffusion process allows progressive refinement and dynamic sampling of entities, empowering DiffusionNER with efficient and flexible entity generation capability. Experiments on multiple flat and nested NER datasets demonstrate that DiffusionNER achieves comparable or even better performance than previous state-of-the-art models.

Abstract: In-context learning (ICL) is an important paradigm for adapting large language models (LLMs) to new tasks, but the generalization behavior of ICL remains poorly understood. We investigate the inductive biases of ICL from the perspective of feature bias: which feature ICL is more likely to use given a set of underspecified demonstrations in which two features are equally predictive of the labels. First, we characterize the feature biases of GPT-3 models by constructing underspecified demonstrations from a range of NLP datasets and feature combinations. We find that LLMs exhibit clear feature biases - for example, demonstrating a strong bias to predict labels according to sentiment rather than shallow lexical features, like punctuation. Second, we evaluate the effect of different interventions that are designed to impose an inductive bias in favor of a particular feature, such as adding a natural language instruction or using semantically relevant label words. We find that, while many interventions can influence the learner to prefer a particular feature, it can be difficult to overcome strong prior biases. Overall, our results provide a broader picture of the types of features that ICL may be more likely to exploit and how to impose inductive biases that are better aligned with the intended task.

Abstract: By providing external information to large language models (LLMs), tool augmentation (including retrieval augmentation) has emerged as a promising solution for addressing the limitations of LLMs' static parametric memory. However, how receptive are LLMs to such external evidence, especially when the evidence conflicts with their parametric memory? We present the first comprehensive and controlled investigation into the behavior of LLMs when encountering knowledge conflicts. We propose a systematic framework to elicit high-quality parametric memory from LLMs and construct the corresponding counter-memory, which enables us to conduct a series of controlled experiments. Our investigation reveals seemingly contradicting behaviors of LLMs. On the one hand, different from prior wisdom, we find that LLMs can be highly receptive to external evidence even when that conflicts with their parametric memory, given that the external evidence is coherent and convincing. On the other hand, LLMs also demonstrate a strong confirmation bias when the external evidence contains some information that is consistent with their parametric memory, despite being presented with conflicting evidence at the same time. These results pose important implications that are worth careful consideration for the further development and deployment of tool- and retrieval-augmented LLMs.

Abstract: Pretrained language models (PLMs) are key components in NLP, but they contain strong social biases. Quantifying these biases is challenging because current methods focusing on fill-the-mask objectives are sensitive to slight changes in input. To address this, we propose LABDet, a robust language-agnostic method for evaluating bias in PLMs. For nationality as a case study, we show that LABDet "surfaces" nationality bias by training a classifier on top of a frozen PLM on non-nationality sentiment detection. Collaborating with political scientists, we find consistent patterns of nationality bias across monolingual PLMs in six languages that align with historical and political context. We also show for English BERT that bias surfaced by LABDet correlates well with bias in the pretraining data; thus, our work is one of the few studies that directly links pretraining data to PLM behavior. Finally, we verify LABDet's reliability and applicability to different templates and languages through an extensive set of robustness checks.

Abstract: Automatically highlighting words that cause semantic differences between two documents could be useful for a wide range of applications. We formulate recognizing semantic differences (RSD) as a token-level regression task and study three unsupervised approaches that rely on a masked language model. To assess the approaches, we begin with basic English sentences and gradually move to more complex, cross-lingual document pairs. Our results show that an approach based on word alignment and sentence-level contrastive learning has a robust correlation to gold labels. However, all unsupervised approaches still leave a large margin of improvement. Code to reproduce our experiments is available at https://github.com/ZurichNLP/recognizing-semantic-differences

Abstract: The fixed-size context of Transformer makes GPT models incapable of generating arbitrarily long text. In this paper, we introduce RecurrentGPT, a language-based simulacrum of the recurrence mechanism in RNNs. RecurrentGPT is built upon a large language model (LLM) such as ChatGPT and uses natural language to simulate the Long Short-Term Memory mechanism in an LSTM. At each timestep, RecurrentGPT generates a paragraph of text and updates its language-based long-short term memory stored on the hard drive and the prompt, respectively. This recurrence mechanism enables RecurrentGPT to generate texts of arbitrary length without forgetting. Since human users can easily observe and edit the natural language memories, RecurrentGPT is interpretable and enables interactive generation of long text. RecurrentGPT is an initial step towards next-generation computer-assisted writing systems beyond local editing suggestions. In addition to producing AI-generated content (AIGC), we also demonstrate the possibility of using RecurrentGPT as an interactive fiction that directly interacts with consumers. We call this usage of generative models by ``AI As Contents'' (AIAC), which we believe is the next form of conventional AIGC. We further demonstrate the possibility of using RecurrentGPT to create personalized interactive fiction that directly interacts with readers instead of interacting with writers. More broadly, RecurrentGPT demonstrates the utility of borrowing ideas from popular model designs in cognitive science and deep learning for prompting LLMs. Our code is available at https://github.com/aiwaves-cn/RecurrentGPT and an online demo is available at https://www.aiwaves.org/recurrentgpt.

Abstract: Automated evaluation of text generation systems has recently seen increasing attention, particularly checking whether generated text stays truthful to input sources. Existing methods frequently rely on an evaluation using task-specific language models, which in turn allows for little interpretability of generated scores. We introduce SRLScore, a reference-free evaluation metric designed with text summarization in mind. Our approach generates fact tuples constructed from Semantic Role Labels, applied to both input and summary texts. A final factuality score is computed by an adjustable scoring mechanism, which allows for easy adaption of the method across domains. Correlation with human judgments on English summarization datasets shows that SRLScore is competitive with state-of-the-art methods and exhibits stable generalization across datasets without requiring further training or hyperparameter tuning. We experiment with an optional co-reference resolution step, but find that the performance boost is mostly outweighed by the additional compute required. Our metric is available online at https://github.com/heyjing/SRLScore.