Computation and Language (cs.CL)

Abstract: Large language models (LLMs) are increasingly adopted for knowledge-intensive tasks and contexts. Existing approaches improve the knowledge capabilities of general-purpose LLMs through retrieval or generated knowledge prompting, but they fall short of reflecting two key properties of knowledge-rich models: knowledge should be modular, ever-growing, sourced from diverse domains; knowledge acquisition and production should be a collaborative process, where diverse stakeholders contribute new information. To this end, we propose CooK, a novel framework to empower general-purpose large language models with modular and collaboratively sourced knowledge. We first introduce specialized language models, autoregressive models trained on corpora from a wide range of domains and sources. These specialized LMs serve as parametric knowledge repositories that are later prompted to generate background knowledge for general-purpose LLMs. We then propose three knowledge filters to dynamically select and retain information in generated documents by controlling for relevance, brevity, and factuality. Finally, we propose bottom-up and top-down knowledge integration approaches to augment general-purpose LLMs with the curated (relevant, factual) knowledge from community-driven specialized LMs that enable multi-domain knowledge synthesis and on-demand knowledge requests. Through extensive experiments, we demonstrate that CooK achieves state-of-the-art performance on six benchmark datasets. Our results highlight the potential of enriching general-purpose LLMs with evolving and modular knowledge -- relevant knowledge that can be continuously updated through the collective efforts of the research community.

Abstract: Enhancing word usage is a desired feature for writing assistance. To further advance research in this area, this paper introduces "Smart Word Suggestions" (SWS) task and benchmark. Unlike other works, SWS emphasizes end-to-end evaluation and presents a more realistic writing assistance scenario. This task involves identifying words or phrases that require improvement and providing substitution suggestions. The benchmark includes human-labeled data for testing, a large distantly supervised dataset for training, and the framework for evaluation. The test data includes 1,000 sentences written by English learners, accompanied by over 16,000 substitution suggestions annotated by 10 native speakers. The training dataset comprises over 3.7 million sentences and 12.7 million suggestions generated through rules. Our experiments with seven baselines demonstrate that SWS is a challenging task. Based on experimental analysis, we suggest potential directions for future research on SWS. The dataset and related codes is available at https://github.com/microsoft/SmartWordSuggestions.

Abstract: Textual response generation is an essential task for multimodal task-oriented dialog systems.Although existing studies have achieved fruitful progress, they still suffer from two critical limitations: 1) focusing on the attribute knowledge but ignoring the relation knowledge that can reveal the correlations between different entities and hence promote the response generation}, and 2) only conducting the cross-entropy loss based output-level supervision but lacking the representation-level regularization. To address these limitations, we devise a novel multimodal task-oriented dialog system (named MDS-S2). Specifically, MDS-S2 first simultaneously acquires the context related attribute and relation knowledge from the knowledge base, whereby the non-intuitive relation knowledge is extracted by the n-hop graph walk. Thereafter, considering that the attribute knowledge and relation knowledge can benefit the responding to different levels of questions, we design a multi-level knowledge composition module in MDS-S2 to obtain the latent composed response representation. Moreover, we devise a set of latent query variables to distill the semantic information from the composed response representation and the ground truth response representation, respectively, and thus conduct the representation-level semantic regularization. Extensive experiments on a public dataset have verified the superiority of our proposed MDS-S2. We have released the codes and parameters to facilitate the research community.

Abstract: In this paper, we introduce self-distillation and online clustering for self-supervised speech representation learning (DinoSR) which combines masked language modeling, self-distillation, and online clustering. We show that these concepts complement each other and result in a strong representation learning model for speech. DinoSR first extracts contextualized embeddings from the input audio with a teacher network, then runs an online clustering system on the embeddings to yield a machine-discovered phone inventory, and finally uses the discretized tokens to guide a student network. We show that DinoSR surpasses previous state-of-the-art performance in several downstream tasks, and provide a detailed analysis of the model and the learned discrete units. The source code will be made available after the anonymity period.

Abstract: Knowledge distillation has attracted a great deal of interest recently to compress pre-trained language models. However, existing knowledge distillation methods suffer from two limitations. First, the student model simply imitates the teacher's behavior while ignoring the underlying reasoning. Second, these methods usually focus on the transfer of sophisticated model-specific knowledge but overlook data-specific knowledge. In this paper, we present a novel attribution-driven knowledge distillation approach, which explores the token-level rationale behind the teacher model based on Integrated Gradients (IG) and transfers attribution knowledge to the student model. To enhance the knowledge transfer of model reasoning and generalization, we further explore multi-view attribution distillation on all potential decisions of the teacher. Comprehensive experiments are conducted with BERT on the GLUE benchmark. The experimental results demonstrate the superior performance of our approach to several state-of-the-art methods.

Abstract: Large pre-trained language models (PLMs) have garnered significant attention for their versatility and potential for solving a wide spectrum of natural language processing (NLP) tasks. However, the cost of running these PLMs may be prohibitive. Furthermore, PLMs may not be open-sourced due to commercial considerations and potential risks of misuse, such as GPT-3. The parameters and gradients of PLMs are unavailable in this scenario. To solve the issue, black-box tuning has been proposed, which utilizes derivative-free optimization (DFO), instead of gradient descent, for training task-specific continuous prompts. However, these gradient-free methods still exhibit a significant gap compared to gradient-based methods. In this paper, we introduce gradient descent into black-box tuning scenario through knowledge distillation. Furthermore, we propose a novel method GDFO, which integrates gradient descent and derivative-free optimization to optimize task-specific continuous prompts in a harmonized manner. Experimental results show that GDFO can achieve significant performance gains over previous state-of-the-art methods.

Abstract: Large language models (LLMs) have demonstrated powerful capabilities in both text understanding and generation. Companies have begun to offer Embedding as a Service (EaaS) based on these LLMs, which can benefit various natural language processing (NLP) tasks for customers. However, previous studies have shown that EaaS is vulnerable to model extraction attacks, which can cause significant losses for the owners of LLMs, as training these models is extremely expensive. To protect the copyright of LLMs for EaaS, we propose an Embedding Watermark method called EmbMarker that implants backdoors on embeddings. Our method selects a group of moderate-frequency words from a general text corpus to form a trigger set, then selects a target embedding as the watermark, and inserts it into the embeddings of texts containing trigger words as the backdoor. The weight of insertion is proportional to the number of trigger words included in the text. This allows the watermark backdoor to be effectively transferred to EaaS-stealer's model for copyright verification while minimizing the adverse impact on the original embeddings' utility. Our extensive experiments on various datasets show that our method can effectively protect the copyright of EaaS models without compromising service quality.

Abstract: Large language models (LLMs) are increasingly adopted for a variety of tasks with implicit graphical structures, such as planning in robotics, multi-hop question answering or knowledge probing, structured commonsense reasoning, and more. While LLMs have advanced the state-of-the-art on these tasks with structure implications, whether LLMs could explicitly process textual descriptions of graphs and structures, map them to grounded conceptual spaces, and perform structured operations remains underexplored. To this end, we propose NLGraph (Natural Language Graph), a comprehensive benchmark of graph-based problem solving designed in natural language. NLGraph contains 29,370 problems, covering eight graph reasoning tasks with varying complexity from simple tasks such as connectivity and shortest path up to complex problems such as maximum flow and simulating graph neural networks. We evaluate LLMs (GPT-3/4) with various prompting approaches on the NLGraph benchmark and find that 1) language models do demonstrate preliminary graph reasoning abilities, 2) the benefit of advanced prompting and in-context learning diminishes on more complex graph problems, while 3) LLMs are also (un)surprisingly brittle in the face of spurious correlations in graph and problem settings. We then propose Build-a-Graph Prompting and Algorithmic Prompting, two instruction-based approaches to enhance LLMs in solving natural language graph problems. Build-a-Graph and Algorithmic prompting improve the performance of LLMs on NLGraph by 3.07% to 16.85% across multiple tasks and settings, while how to solve the most complicated graph reasoning tasks in our setup with language models remains an open research question. The NLGraph benchmark and evaluation code are available at https://github.com/Arthur-Heng/NLGraph.

Abstract: In most Vision-Language models (VL), the understanding of the image structure is enabled by injecting the position information (PI) about objects in the image. In our case study of LXMERT, a state-of-the-art VL model, we probe the use of the PI in the representation and study its effect on Visual Question Answering. We show that the model is not capable of leveraging the PI for the image-text matching task on a challenge set where only position differs. Yet, our experiments with probing confirm that the PI is indeed present in the representation. We introduce two strategies to tackle this: (i) Positional Information Pre-training and (ii) Contrastive Learning on PI using Cross-Modality Matching. Doing so, the model can correctly classify if images with detailed PI statements match. Additionally to the 2D information from bounding boxes, we introduce the object's depth as new feature for a better object localization in the space. Even though we were able to improve the model properties as defined by our probes, it only has a negligible effect on the downstream performance. Our results thus highlight an important issue of multimodal modeling: the mere presence of information detectable by a probing classifier is not a guarantee that the information is available in a cross-modal setup.

Abstract: A recent trend in the domain of open-domain conversational agents is enabling them to converse empathetically to emotional prompts. Current approaches either follow an end-to-end approach or condition the responses on similar emotion labels to generate empathetic responses. But empathy is a broad concept that refers to the cognitive and emotional reactions of an individual to the observed experiences of another and it is more complex than mere mimicry of emotion. Hence, it requires identifying complex human conversational strategies and dynamics in addition to generic emotions to control and interpret empathetic responding capabilities of chatbots. In this work, we make use of a taxonomy of eight empathetic response intents in addition to generic emotion categories in building a dialogue response generation model capable of generating empathetic responses in a controllable and interpretable manner. It consists of two modules: 1) a response emotion/intent prediction module; and 2) a response generation module. We propose several rule-based and neural approaches to predict the next response's emotion/intent and generate responses conditioned on these predicted emotions/intents. Automatic and human evaluation results emphasize the importance of the use of the taxonomy of empathetic response intents in producing more diverse and empathetically more appropriate responses than end-to-end models.

Abstract: Konkani is a highly nasalised language which makes it unique among Indo-Aryan languages. This work investigates the acoustic-phonetic properties of Konkani oral and nasal vowels. For this study, speech samples from six speakers (3 male and 3 female) were collected. A total of 74 unique sentences were used as a part of the recording script, 37 each for oral and nasal vowels, respectively. The final data set consisted of 1135 vowel phonemes. A comparative F1-F2 plot of Konkani oral and nasal vowels is presented with an experimental result and formant analysis. The average F1, F2 and F3 values are also reported for the first time through experimentation for all nasal and oral vowels. This study can be helpful for the linguistic research on vowels and speech synthesis systems specific to the Konkani language.

Abstract: Automatic extraction of party (dis)similarities from texts such as party election manifestos or parliamentary speeches plays an increasing role in computational political science. However, existing approaches are fundamentally limited to targeting only global party (dis)-similarity: they condense the relationship between a pair of parties into a single figure, their similarity. In aggregating over all policy domains (e.g., health or foreign policy), they do not provide any qualitative insights into which domains parties agree or disagree on. This paper proposes a workflow for estimating policy domain aware party similarity that overcomes this limitation. The workflow covers (a) definition of suitable policy domains; (b) automatic labeling of domains, if no manual labels are available; (c) computation of domain-level similarities and aggregation at a global level; (d) extraction of interpretable party positions on major policy axes via multidimensional scaling. We evaluate our workflow on manifestos from the German federal elections. We find that our method (a) yields high correlation when predicting party similarity at a global level and (b) provides accurate party-specific positions, even with automatically labelled policy domains.

Abstract: We study whether multiple large language models (LLMs) can autonomously improve each other in a negotiation game by playing, reflecting, and criticizing. We are interested in this question because if LLMs were able to improve each other, it would imply the possibility of creating strong AI agents with minimal human intervention. We ask two LLMs to negotiate with each other, playing the roles of a buyer and a seller, respectively. They aim to reach a deal with the buyer targeting a lower price and the seller a higher one. A third language model, playing the critic, provides feedback to a player to improve the player's negotiation strategies. We let the two agents play multiple rounds, using previous negotiation history and AI feedback as in-context demonstrations to improve the model's negotiation strategy iteratively. We use different LLMs (GPT and Claude) for different roles and use the deal price as the evaluation metric. Our experiments reveal multiple intriguing findings: (1) Only a subset of the language models we consider can self-play and improve the deal price from AI feedback, weaker models either do not understand the game's rules or cannot incorporate AI feedback for further improvement. (2) Models' abilities to learn from the feedback differ when playing different roles. For example, it is harder for Claude-instant to improve as the buyer than as the seller. (3) When unrolling the game to multiple rounds, stronger agents can consistently improve their performance by meaningfully using previous experiences and iterative AI feedback, yet have a higher risk of breaking the deal. We hope our work provides insightful initial explorations of having models autonomously improve each other with game playing and AI feedback.

Abstract: Retrieving proper domain knowledge from an external database lies at the heart of end-to-end task-oriented dialog systems to generate informative responses. Most existing systems blend knowledge retrieval with response generation and optimize them with direct supervision from reference responses, leading to suboptimal retrieval performance when the knowledge base becomes large-scale. To address this, we propose to decouple knowledge retrieval from response generation and introduce a multi-grained knowledge retriever (MAKER) that includes an entity selector to search for relevant entities and an attribute selector to filter out irrelevant attributes. To train the retriever, we propose a novel distillation objective that derives supervision signals from the response generator. Experiments conducted on three standard benchmarks with both small and large-scale knowledge bases demonstrate that our retriever performs knowledge retrieval more effectively than existing methods. Our code has been made publicly available.\footnote{https://github.com/18907305772/MAKER}

Abstract: It has been argued that semantic systems reflect pressure for efficiency, and a current debate concerns the cultural evolutionary process that produces this pattern. We consider efficiency as instantiated in the Information Bottleneck (IB) principle, and a model of cultural evolution that combines iterated learning and communication. We show that this model, instantiated in neural networks, converges to color naming systems that are efficient in the IB sense and similar to human color naming systems. We also show that iterated learning alone, and communication alone, do not yield the same outcome as clearly.

Abstract: Comprehending characters' personalities is a crucial aspect of story reading. As readers engage with a story, their understanding of a character evolves based on new events and information; and multiple fine-grained aspects of personalities can be perceived. This leads to a natural problem of situated and fine-grained personality understanding. The problem has not been studied in the NLP field, primarily due to the lack of appropriate datasets mimicking the process of book reading. We present the first labeled dataset PersoNet for this problem. Our novel annotation strategy involves annotating user notes from online reading apps as a proxy for the original books. Experiments and human studies indicate that our dataset construction is both efficient and accurate; and our task heavily relies on long-term context to achieve accurate predictions for both machines and humans. The dataset is available at https://github.com/Gorov/personet_acl23.

Abstract: Data contamination has become especially prevalent and challenging with the rise of models pretrained on very large, automatically-crawled corpora. For closed models, the training data becomes a trade secret, and even for open models, it is not trivial to ascertain whether a particular test instance has been compromised. Strategies such as live leaderboards with hidden answers, or using test data which is guaranteed to be unseen, are expensive and become fragile with time. Assuming that all relevant actors value clean test data and will cooperate to mitigate data contamination, what can be done? We propose three strategies that can make a difference: (1) Test data made public should be encrypted with a public key and licensed to disallow derivative distribution; (2) demand training exclusion controls from closed API holders, and protect your test data by refusing to evaluate until demands are met; (3) in case of test data based on internet text, avoid data which appears with its solution on the internet, and release the context of internet-derived data along with the data. These strategies are practical and can be effective in preventing data contamination and allowing trustworthy evaluation of models' capabilities.

Abstract: Generative Pre-Training (GPT) models like ChatGPT have demonstrated exceptional performance in various Natural Language Processing (NLP) tasks. Although ChatGPT has been integrated into the overall workflow to boost efficiency in many domains, the lack of flexibility in the finetuning process hinders its applications in areas that demand extensive domain expertise and semantic knowledge, such as healthcare. In this paper, we evaluate ChatGPT on the China National Medical Licensing Examination (CNMLE) and propose a novel approach to improve ChatGPT from two perspectives: integrating medical domain knowledge and enabling few-shot learning. By using a simple but effective retrieval method, medical background knowledge is extracted as semantic instructions to guide the inference of ChatGPT. Similarly, relevant medical questions are identified and fed as demonstrations to ChatGPT. Experimental results show that directly applying ChatGPT fails to qualify the CNMLE at a score of 51 (i.e., only 51\% of questions are answered correctly). While our knowledge-enhanced model achieves a high score of 70 on CNMLE-2022 which not only passes the qualification but also surpasses the average score of humans (61). This research demonstrates the potential of knowledge-enhanced ChatGPT to serve as versatile medical assistants, capable of analyzing real-world medical problems in a more accessible, user-friendly, and adaptable manner.

Abstract: Unlike empathetic dialogues, the system in emotional support conversations (ESC) is expected to not only convey empathy for comforting the help-seeker, but also proactively assist in exploring and addressing their problems during the conversation. In this work, we study the problem of mixed-initiative ESC where the user and system can both take the initiative in leading the conversation. Specifically, we conduct a novel analysis on mixed-initiative ESC systems with a tailor-designed schema that divides utterances into different types with speaker roles and initiative types. Four emotional support metrics are proposed to evaluate the mixed-initiative interactions. The analysis reveals the necessity and challenges of building mixed-initiative ESC systems. In the light of this, we propose a knowledge-enhanced mixed-initiative framework (KEMI) for ESC, which retrieves actual case knowledge from a large-scale mental health knowledge graph for generating mixed-initiative responses. Experimental results on two ESC datasets show the superiority of KEMI in both content-preserving evaluation and mixed initiative related analyses.

Abstract: The language-independency of encoded representations within multilingual neural machine translation (MNMT) models is crucial for their generalization ability on zero-shot translation. Neural interlingua representations have been shown as an effective method for achieving this. However, fixed-length neural interlingua representations introduced in previous work can limit its flexibility and representation ability. In this study, we introduce a novel method to enhance neural interlingua representations by making their length variable, thereby overcoming the constraint of fixed-length neural interlingua representations. Our empirical results on zero-shot translation on OPUS, IWSLT, and Europarl datasets demonstrate stable model convergence and superior zero-shot translation results compared to fixed-length neural interlingua representations. However, our analysis reveals the suboptimal efficacy of our approach in translating from certain source languages, wherein we pinpoint the defective model component in our proposed method.

Abstract: AI-driven chatbots have become an emerging solution to address psychological distress. Due to the lack of psychotherapeutic data, researchers use dialogues scraped from online peer support forums to train them. But since the responses in such platforms are not given by professionals, they contain both conforming and non-conforming responses. In this work, we attempt to recognize these conforming and non-conforming response types present in online distress-support dialogues using labels adapted from a well-established behavioral coding scheme named Motivational Interviewing Treatment Integrity (MITI) code and show how some response types could be rephrased into a more MI adherent form that can, in turn, enable chatbot responses to be more compliant with the MI strategy. As a proof of concept, we build several rephrasers by fine-tuning Blender and GPT3 to rephrase MI non-adherent "Advise without permission" responses into "Advise with permission". We show how this can be achieved with the construction of pseudo-parallel corpora avoiding costs for human labor. Through automatic and human evaluation we show that in the presence of less training data, techniques such as prompting and data augmentation can be used to produce substantially good rephrasings that reflect the intended style and preserve the content of the original text.

Abstract: Zero pronouns (ZPs) are frequently omitted in pro-drop languages (e.g. Chinese, Hungarian, and Hindi), but should be recalled in non-pro-drop languages (e.g. English). This phenomenon has been studied extensively in machine translation (MT), as it poses a significant challenge for MT systems due to the difficulty in determining the correct antecedent for the pronoun. This survey paper highlights the major works that have been undertaken in zero pronoun translation (ZPT) after the neural revolution, so that researchers can recognise the current state and future directions of this field. We provide an organisation of the literature based on evolution, dataset, method and evaluation. In addition, we compare and analyze competing models and evaluation metrics on different benchmarks. We uncover a number of insightful findings such as: 1) ZPT is in line with the development trend of large language model; 2) data limitation causes learning bias in languages and domains; 3) performance improvements are often reported on single benchmarks, but advanced methods are still far from real-world use; 4) general-purpose metrics are not reliable on nuances and complexities of ZPT, emphasizing the necessity of targeted metrics; 5) apart from commonly-cited errors, ZPs will cause risks of gender bias.

Abstract: Natural language processing models tend to learn and encode social biases present in the data. One popular approach for addressing such biases is to eliminate encoded information from the model's representations. However, current methods are restricted to removing only linearly encoded information. In this work, we propose Iterative Gradient-Based Projection (IGBP), a novel method for removing non-linear encoded concepts from neural representations. Our method consists of iteratively training neural classifiers to predict a particular attribute we seek to eliminate, followed by a projection of the representation on a hypersurface, such that the classifiers become oblivious to the target attribute. We evaluate the effectiveness of our method on the task of removing gender and race information as sensitive attributes. Our results demonstrate that IGBP is effective in mitigating bias through intrinsic and extrinsic evaluations, with minimal impact on downstream task accuracy.

Abstract: Spoken Language Understanding (SLU) is one of the core components of a task-oriented dialogue system, which aims to extract the semantic meaning of user queries (e.g., intents and slots). In this work, we introduce OpenSLU, an open-source toolkit to provide a unified, modularized, and extensible toolkit for spoken language understanding. Specifically, OpenSLU unifies 10 SLU models for both single-intent and multi-intent scenarios, which support both non-pretrained and pretrained models simultaneously. Additionally, OpenSLU is highly modularized and extensible by decomposing the model architecture, inference, and learning process into reusable modules, which allows researchers to quickly set up SLU experiments with highly flexible configurations. OpenSLU is implemented based on PyTorch, and released at \url{https://github.com/LightChen233/OpenSLU}.

Abstract: Recent years witnessed an increase in the amount of research on the task of Question Difficulty Estimation from Text QDET with Natural Language Processing (NLP) techniques, with the goal of targeting the limitations of traditional approaches to question calibration. However, almost the entirety of previous research focused on single silos, without performing quantitative comparisons between different models or across datasets from different educational domains. In this work, we aim at filling this gap, by quantitatively analyzing several approaches proposed in previous research, and comparing their performance on three publicly available real world datasets containing questions of different types from different educational domains. Specifically, we consider reading comprehension Multiple Choice Questions (MCQs), science MCQs, and math questions. We find that Transformer based models are the best performing across different educational domains, with DistilBERT performing almost as well as BERT, and that they outperform other approaches even on smaller datasets. As for the other models, the hybrid ones often outperform the ones based on a single type of features, the ones based on linguistic features perform well on reading comprehension questions, while frequency based features (TF-IDF) and word embeddings (word2vec) perform better in domain knowledge assessment.

Abstract: Revolutionary advancements in Large Language Models have drastically reshaped our interactions with artificial intelligence systems. Despite this, a notable hindrance remains-the deficiency of a long-term memory mechanism within these models. This shortfall becomes increasingly evident in situations demanding sustained interaction, such as personal companion systems and psychological counseling. Therefore, we propose MemoryBank, a novel memory mechanism tailored for LLMs. MemoryBank enables the models to summon relevant memories, continually evolve through continuous memory updates, comprehend, and adapt to a user personality by synthesizing information from past interactions. To mimic anthropomorphic behaviors and selectively preserve memory, MemoryBank incorporates a memory updating mechanism, inspired by the Ebbinghaus Forgetting Curve theory, which permits the AI to forget and reinforce memory based on time elapsed and the relative significance of the memory, thereby offering a human-like memory mechanism. MemoryBank is versatile in accommodating both closed-source models like ChatGPT and open-source models like ChatGLM. We exemplify application of MemoryBank through the creation of an LLM-based chatbot named SiliconFriend in a long-term AI Companion scenario. Further tuned with psychological dialogs, SiliconFriend displays heightened empathy in its interactions. Experiment involves both qualitative analysis with real-world user dialogs and quantitative analysis with simulated dialogs. In the latter, ChatGPT acts as users with diverse characteristics and generates long-term dialog contexts covering a wide array of topics. The results of our analysis reveal that SiliconFriend, equipped with MemoryBank, exhibits a strong capability for long-term companionship as it can provide emphatic response, recall relevant memories and understand user personality.

Abstract: Large language models have recently made tremendous progress in a variety of aspects, e.g., cross-task generalization, instruction following. Comprehensively evaluating the capability of large language models in multiple tasks is of great importance. In this paper, we propose M3KE, a Massive Multi-Level Multi-Subject Knowledge Evaluation benchmark, which is developed to measure knowledge acquired by Chinese large language models by testing their multitask accuracy in zero- and few-shot settings. We have collected 20,477 questions from 71 tasks. Our selection covers all major levels of Chinese education system, ranging from the primary school to college, as well as a wide variety of subjects, including humanities, history, politics, law, education, psychology, science, technology, art and religion. All questions are multiple-choice questions with four options, hence guaranteeing a standardized and unified assessment process. We've assessed a number of state-of-the-art open-source Chinese large language models on the proposed benchmark. The size of these models varies from 335M to 130B parameters. Experiment results demonstrate that they perform significantly worse than GPT-3.5 that reaches an accuracy of ~ 48% on M3KE. The dataset is available at https://github.com/tjunlp-lab/M3KE.

Abstract: Large, multilingual language models exhibit surprisingly good zero- or few-shot machine translation capabilities, despite having never seen the intentionally-included translation examples provided to typical neural translation systems. We investigate the role of incidental bilingualism -- the unintentional consumption of bilingual signals, including translation examples -- in explaining the translation capabilities of large language models, taking the Pathways Language Model (PaLM) as a case study. We introduce a mixed-method approach to measure and understand incidental bilingualism at scale. We show that PaLM is exposed to over 30 million translation pairs across at least 44 languages. Furthermore, the amount of incidental bilingual content is highly correlated with the amount of monolingual in-language content for non-English languages. We relate incidental bilingual content to zero-shot prompts and show that it can be used to mine new prompts to improve PaLM's out-of-English zero-shot translation quality. Finally, in a series of small-scale ablations, we show that its presence has a substantial impact on translation capabilities, although this impact diminishes with model scale.

Abstract: We introduce the problem of phone classification in the context of speech recognition, and explore several sets of local spectro-temporal features that can be used for phone classification. In particular, we present some preliminary results for phone classification using two sets of features that are commonly used for object detection: Haar features and SVM-classified Histograms of Gradients (HoG)

Abstract: In this paper, we take the initiative to investigate the performance of LLMs on complex planning tasks that require LLMs to understand a virtual spatial environment simulated via natural language and act correspondingly in text. We propose a benchmark named Natural Language Planning (NLP) composed of a set of novel tasks: Brick World, NLVR-based Manipulations, and Natural Language Navigation. We found that current popular LLMs such as ChatGPT still lack abilities in complex planning. This arises a question -- do the LLMs have a good understanding of the environments described in natural language, or maybe other alternatives such as symbolic representations are neater and hence better to be understood by LLMs? To this end, we propose a novel method called CoS (Chain-of-Symbol Prompting) that represents the complex environments with condensed symbolic spatial representations during the chained intermediate thinking steps. CoS is easy to use and does not need additional training on LLMs. Extensive experiments indicate that CoS clearly surpasses the performance of the Chain-of-Thought (CoT) Prompting in all three planning tasks with even fewer tokens used in the inputs compared with CoT on ChatGPT and InstructGPT. The performance gain is strong, by up to 60.8% accuracy (from 31.8% to 92.6%) on Brick World for ChatGPT. CoS also reduces the number of tokens in the prompt obviously, by up to 65.8% of the tokens (from 407 to 139) for the intermediate steps from demonstrations on Brick World.

Abstract: The evaluation of natural language processing (NLP) systems is crucial for advancing the field, but current benchmarking approaches often assume that all systems have scores available for all tasks, which is not always practical. In reality, several factors such as the cost of running baseline, private systems, computational limitations, or incomplete data may prevent some systems from being evaluated on entire tasks. This paper formalize an existing problem in NLP research: benchmarking when some systems scores are missing on the task, and proposes a novel approach to address it. Our method utilizes a compatible partial ranking approach to impute missing data, which is then aggregated using the Borda count method. It includes two refinements designed specifically for scenarios where either task-level or instance-level scores are available. We also introduce an extended benchmark, which contains over 131 million scores, an order of magnitude larger than existing benchmarks. We validate our methods and demonstrate their effectiveness in addressing the challenge of missing system evaluation on an entire task. This work highlights the need for more comprehensive benchmarking approaches that can handle real-world scenarios where not all systems are evaluated on the entire task.

Abstract: We propose a new paradigm for universal information extraction (IE) that is compatible with any schema format and applicable to a list of IE tasks, such as named entity recognition, relation extraction, event extraction and sentiment analysis. Our approach converts the text-based IE tasks as the token-pair problem, which uniformly disassembles all extraction targets into joint span detection, classification and association problems with a unified extractive framework, namely UniEX. UniEX can synchronously encode schema-based prompt and textual information, and collaboratively learn the generalized knowledge from pre-defined information using the auto-encoder language models. We develop a traffine attention mechanism to integrate heterogeneous factors including tasks, labels and inside tokens, and obtain the extraction target via a scoring matrix. Experiment results show that UniEX can outperform generative universal IE models in terms of performance and inference-speed on $14$ benchmarks IE datasets with the supervised setting. The state-of-the-art performance in low-resource scenarios also verifies the transferability and effectiveness of UniEX.

Abstract: Measuring the distance between machine-produced and human language is acritical open problem. Inspired by empirical findings from psycholinguistics on theperiodicity of entropy in language, we propose FACE, a set of metrics based onFourier Analysis of the estimated Cross-Entropy of language, for measuring thesimilarity between model-generated and human-written languages. Based on anopen-ended generation task and the experimental data from previous studies, weind that FACE can effectively identify the human-model gap, scales with modelsize, reflects the outcomes of different sampling methods for decoding, correlateswell with other evaluation metrics and with human judgment scores. FACE iscomputationally efficient and provides intuitive interpretations.

Abstract: Finetuning pre-trained language models (LMs) enhances the models' capabilities. Prior techniques fine-tune a pre-trained LM on input-output pairs (e.g., instruction fine-tuning), or with numerical rewards that gauge the quality of its outputs (e.g., reinforcement learning from human feedback). We explore LMs' potential to learn from textual interactions (LeTI) that not only check their correctness with binary labels, but also pinpoint and explain errors in their outputs through textual feedback. Our investigation focuses on the code generation task, where the model produces code pieces in response to natural language instructions. This setting invites a natural and scalable way to acquire the textual feedback: the error messages and stack traces from code execution using a Python interpreter. LeTI iteratively fine-tunes the model, using the LM objective, on a concatenation of natural language instructions, LM-generated programs, and textual feedback, which is only provided when the generated program fails to solve the task. Prepended to this fine-tuning text, a binary reward token is used to differentiate correct and buggy solutions. On MBPP, a code generation dataset, LeTI substantially improves the performance of two base LMs of different scales. LeTI requires no ground-truth outputs for training and even outperforms a fine-tuned baseline that does. LeTI's strong performance generalizes to other datasets. Trained on MBPP, it achieves comparable or better performance than the base LMs on unseen problems in HumanEval. Furthermore, compared to binary feedback, we observe that textual feedback leads to improved generation quality and sample efficiency, achieving the same performance with fewer than half of the gradient steps. LeTI is equally applicable in natural language tasks when they can be formulated as code generation, which we empirically verified on event argument extraction.

Abstract: Appropriate prosody is critical for successful spoken communication. Contextual word embeddings are proven to be helpful in predicting prosody but do not allow for choosing between plausible prosodic renditions. Reference-based TTS models attempt to address this by conditioning speech generation on a reference speech sample. These models can generate expressive speech but this requires finding an appropriate reference. Sufficiently large generative language models have been used to solve various language-related tasks. We explore whether such models can be used to suggest appropriate prosody for expressive TTS. We train a TTS model on a non-expressive corpus and then prompt the language model to suggest changes to pitch, energy and duration. The prompt can be designed for any task and we prompt the model to make suggestions based on target speaking style and dialogue context. The proposed method is rated most appropriate in 49.9\% of cases compared to 31.0\% for a baseline model.

Abstract: Labeling data is essential for training text classifiers but is often difficult to accomplish accurately, especially for complex and abstract concepts. Seeking an improved method, this paper employs a novel approach using a generative language model (GPT-4) to produce labels and rationales for large-scale text analysis. We apply this approach to the task of discovering public value expressions in US AI patents. We collect a database comprising 154,934 patent documents using an advanced Boolean query submitted to InnovationQ+. The results are merged with full patent text from the USPTO, resulting in 5.4 million sentences. We design a framework for identifying and labeling public value expressions in these AI patent sentences. A prompt for GPT-4 is developed which includes definitions, guidelines, examples, and rationales for text classification. We evaluate the quality of the labels and rationales produced by GPT-4 using BLEU scores and topic modeling and find that they are accurate, diverse, and faithful. These rationales also serve as a chain-of-thought for the model, a transparent mechanism for human verification, and support for human annotators to overcome cognitive limitations. We conclude that GPT-4 achieved a high-level of recognition of public value theory from our framework, which it also uses to discover unseen public value expressions. We use the labels produced by GPT-4 to train BERT-based classifiers and predict sentences on the entire database, achieving high F1 scores for the 3-class (0.85) and 2-class classification (0.91) tasks. We discuss the implications of our approach for conducting large-scale text analyses with complex and abstract concepts and suggest that, with careful framework design and interactive human oversight, generative language models can offer significant advantages in quality and in reduced time and costs for producing labels and rationales.

Abstract: Automated text simplification, a technique useful for making text more accessible to people such as children and emergent bilinguals, is often thought of as a monolingual translation task from complex sentences to simplified sentences using encoder-decoder models. This view fails to account for elaborative simplification, where new information is added into the simplified text. This paper proposes to view elaborative simplification through the lens of the Question Under Discussion (QUD) framework, providing a robust way to investigate what writers elaborate upon, how they elaborate, and how elaborations fit into the discourse context by viewing elaborations as explicit answers to implicit questions. We introduce ElabQUD, consisting of 1.3K elaborations accompanied with implicit QUDs, to study these phenomena. We show that explicitly modeling QUD (via question generation) not only provides essential understanding of elaborative simplification and how the elaborations connect with the rest of the discourse, but also substantially improves the quality of elaboration generation.

Abstract: Automatically determining whether a text and a corresponding image are semantically aligned is a significant challenge for vision-language models, with applications in generative text-to-image and image-to-text tasks. In this work, we study methods for automatic text-image alignment evaluation. We first introduce SeeTRUE: a comprehensive evaluation set, spanning multiple datasets from both text-to-image and image-to-text generation tasks, with human judgements for whether a given text-image pair is semantically aligned. We then describe two automatic methods to determine alignment: the first involving a pipeline based on question generation and visual question answering models, and the second employing an end-to-end classification approach by finetuning multimodal pretrained models. Both methods surpass prior approaches in various text-image alignment tasks, with significant improvements in challenging cases that involve complex composition or unnatural images. Finally, we demonstrate how our approaches can localize specific misalignments between an image and a given text, and how they can be used to automatically re-rank candidates in text-to-image generation.

Abstract: We introduce PaLM 2, a new state-of-the-art language model that has better multilingual and reasoning capabilities and is more compute-efficient than its predecessor PaLM. PaLM 2 is a Transformer-based model trained using a mixture of objectives. Through extensive evaluations on English and multilingual language, and reasoning tasks, we demonstrate that PaLM 2 has significantly improved quality on downstream tasks across different model sizes, while simultaneously exhibiting faster and more efficient inference compared to PaLM. This improved efficiency enables broader deployment while also allowing the model to respond faster, for a more natural pace of interaction. PaLM 2 demonstrates robust reasoning capabilities exemplified by large improvements over PaLM on BIG-Bench and other reasoning tasks. PaLM 2 exhibits stable performance on a suite of responsible AI evaluations, and enables inference-time control over toxicity without additional overhead or impact on other capabilities. Overall, PaLM 2 achieves state-of-the-art performance across a diverse set of tasks and capabilities. When discussing the PaLM 2 family, it is important to distinguish between pre-trained models (of various sizes), fine-tuned variants of these models, and the user-facing products that use these models. In particular, user-facing products typically include additional pre- and post-processing steps. Additionally, the underlying models may evolve over time. Therefore, one should not expect the performance of user-facing products to exactly match the results reported in this report.

Abstract: Application Tracking Systems (ATS) have allowed talent managers, recruiters, and college admissions committees to process large volumes of potential candidate applications efficiently. Traditionally, this screening process was conducted manually, creating major bottlenecks due to the quantity of applications and introducing many instances of human bias. The advent of large language models (LLMs) such as ChatGPT and the potential of adopting methods to current automated application screening raises additional bias and fairness issues that must be addressed. In this project, we wish to identify and quantify the instances of social bias in ChatGPT and other OpenAI LLMs in the context of candidate screening in order to demonstrate how the use of these models could perpetuate existing biases and inequalities in the hiring process.

Abstract: Learning from human feedback has been shown to be effective at aligning language models with human preferences. Past work has often relied on Reinforcement Learning from Human Feedback (RLHF), which optimizes the language model using reward scores assigned from a reward model trained on human preference data. In this work we show how the recently introduced Sequence Likelihood Calibration (SLiC), can also be used to effectively learn from human preferences (SLiC-HF). Furthermore, we demonstrate this can be done with human feedback data collected for a different model, similar to off-policy, offline RL data. Automatic and human evaluation experiments on the TL;DR summarization task show that SLiC-HF significantly improves supervised fine-tuning baselines. Furthermore, SLiC-HF presents a competitive alternative to the PPO RLHF implementation used in past work while being much simpler to implement, easier to tune and more computationally efficient in practice.

Abstract: Autoregressive decoding limits the efficiency of transformers for Machine Translation (MT). The community proposed specific network architectures and learning-based methods to solve this issue, which are expensive and require changes to the MT model, trading inference speed at the cost of the translation quality. In this paper, we propose to address the problem from the point of view of decoding algorithms, as a less explored but rather compelling direction. We propose to reframe the standard greedy autoregressive decoding of MT with a parallel formulation leveraging Jacobi and Gauss-Seidel fixed-point iteration methods for fast inference. This formulation allows to speed up existing models without training or modifications while retaining translation quality. We present three parallel decoding algorithms and test them on different languages and models showing how the parallelization introduces a speedup up to 38% w.r.t. the standard autoregressive decoding and nearly 2x when scaling the method on parallel resources. Finally, we introduce a decoding dependency graph visualizer (DDGviz) that let us see how the model has learned the conditional dependence between tokens and inspect the decoding procedure.

Abstract: The mixture proportions of pretraining data domains (e.g., Wikipedia, books, web text) greatly affect language model (LM) performance. In this paper, we propose Domain Reweighting with Minimax Optimization (DoReMi), which first trains a small proxy model using group distributionally robust optimization (Group DRO) over domains to produce domain weights (mixture proportions) without knowledge of downstream tasks. We then resample a dataset with these domain weights and train a larger, full-sized model. In our experiments, we use DoReMi on a 280M-parameter proxy model to find domain weights for training an 8B-parameter model (30x larger) more efficiently. On The Pile, DoReMi improves perplexity across all domains, even when it downweights a domain. DoReMi improves average few-shot downstream accuracy by 6.5% over a baseline model trained using The Pile's default domain weights and reaches the baseline accuracy with 2.6x fewer training steps. On the GLaM dataset, DoReMi, which has no knowledge of downstream tasks, even matches the performance of using domain weights tuned on downstream tasks.