Computation and Language (cs.CL)

Abstract: Sentence Boundary Detection (SBD) is one of the foundational building blocks of Natural Language Processing (NLP), with incorrectly split sentences heavily influencing the output quality of downstream tasks. It is a challenging task for algorithms, especially in the legal domain, considering the complex and different sentence structures used. In this work, we curated a diverse multilingual legal dataset consisting of over 130'000 annotated sentences in 6 languages. Our experimental results indicate that the performance of existing SBD models is subpar on multilingual legal data. We trained and tested monolingual and multilingual models based on CRF, BiLSTM-CRF, and transformers, demonstrating state-of-the-art performance. We also show that our multilingual models outperform all baselines in the zero-shot setting on a Portuguese test set. To encourage further research and development by the community, we have made our dataset, models, and code publicly available.

Abstract: The prompt-based learning paradigm, which bridges the gap between pre-training and fine-tuning, achieves state-of-the-art performance on several NLP tasks, particularly in few-shot settings. Despite being widely applied, prompt-based learning is vulnerable to backdoor attacks. Textual backdoor attacks are designed to introduce targeted vulnerabilities into models by poisoning a subset of training samples through trigger injection and label modification. However, they suffer from flaws such as abnormal natural language expressions resulting from the trigger and incorrect labeling of poisoned samples. In this study, we propose {\bf ProAttack}, a novel and efficient method for performing clean-label backdoor attacks based on the prompt, which uses the prompt itself as a trigger. Our method does not require external triggers and ensures correct labeling of poisoned samples, improving the stealthy nature of the backdoor attack. With extensive experiments on rich-resource and few-shot text classification tasks, we empirically validate ProAttack's competitive performance in textual backdoor attacks. Notably, in the rich-resource setting, ProAttack achieves state-of-the-art attack success rates in the clean-label backdoor attack benchmark without external triggers. All data and code used in our models are publically available\footnote{\url{https://github.com/shuaizhao95/Prompt_attack}}.

Abstract: Generative agents that simulate human society show tremendous potential for further research and practical applications. Specifically, the generative agent architecture comprising several meticulously designed modules constitutes the most critical component. To facilitate progress in this research, this report presents our integrated perspective on comprehending generative agents through summarization, since we believe summarization is the most fundamental and indispensable capacity of generative agents manifested across diverse scenarios. We hope this report can provide insight into understanding the importance of summarization capacity in generative agents and motivate future research.

Abstract: Flowchart-grounded troubleshooting dialogue (FTD) systems, which follow the instructions of a flowchart to diagnose users' problems in specific domains (eg., vehicle, laptop), have been gaining research interest in recent years. However, collecting sufficient dialogues that are naturally grounded on flowcharts is costly, thus FTD systems are impeded by scarce training data. To mitigate the data sparsity issue, we propose a plan-based data augmentation (PlanDA) approach that generates diverse synthetic dialog data at scale by transforming concise flowchart into dialogues. Specifically, its generative model employs a variational-base framework with a hierarchical planning strategy that includes global and local latent planning variables. Experiments on the FloDial dataset show that synthetic dialogue produced by PlanDA improves the performance of downstream tasks, including flowchart path retrieval and response generation, in particular on the Out-of-Flowchart settings. In addition, further analysis demonstrate the quality of synthetic data generated by PlanDA in paths that are covered by current sample dialogues and paths that are not covered.

Abstract: Influence functions (IFs) are a powerful tool for detecting anomalous examples in large scale datasets. However, they are unstable when applied to deep networks. In this paper, we provide an explanation for the instability of IFs and develop a solution to this problem. We show that IFs are unreliable when the two data points belong to two different classes. Our solution leverages class information to improve the stability of IFs. Extensive experiments show that our modification significantly improves the performance and stability of IFs while incurring no additional computational cost.

Abstract: The focus is on critical problems in NLP related to linguistic diversity and variation across the African continent, specifically with regards to African local dialects and Arabic dialects that have received little attention. We evaluated our various approaches, demonstrating their effectiveness while highlighting the potential impact of the proposed approach on businesses seeking to improve customer experience and product development in African local dialects. The idea of using the model as a teaching tool for product-based instruction is interesting, as it could potentially stimulate interest in learners and trigger techno entrepreneurship. Overall, our modified approach offers a promising analysis of the challenges of dealing with African local dialects. Particularly Arabic dialects, which could have a significant impact on businesses seeking to improve customer experience and product development.

Abstract: With the advent of deep learning methods, Neural Machine Translation (NMT) systems have become increasingly powerful. However, deep learning based systems are susceptible to adversarial attacks, where imperceptible changes to the input can cause undesirable changes at the output of the system. To date there has been little work investigating adversarial attacks on sequence-to-sequence systems, such as NMT models. Previous work in NMT has examined attacks with the aim of introducing target phrases in the output sequence. In this work, adversarial attacks for NMT systems are explored from an output perception perspective. Thus the aim of an attack is to change the perception of the output sequence, without altering the perception of the input sequence. For example, an adversary may distort the sentiment of translated reviews to have an exaggerated positive sentiment. In practice it is challenging to run extensive human perception experiments, so a proxy deep-learning classifier applied to the NMT output is used to measure perception changes. Experiments demonstrate that the sentiment perception of NMT systems' output sequences can be changed significantly.

Abstract: Most existing multi-document summarization (MDS) datasets lack human-generated and genuine (i.e., not synthetic) summaries or source documents with explicit inter-document relationships that a summary must capture. To enhance the capabilities of MDS systems we present PeerSum, a novel dataset for generating meta-reviews of scientific papers, where the meta-reviews are highly abstractive and genuine summaries of reviews and corresponding discussions. These source documents have rich inter-document relationships of an explicit hierarchical structure with cross-references and often feature conflicts. As there is a scarcity of research that incorporates hierarchical relationships into MDS systems through attention manipulation on pre-trained language models, we additionally present Rammer (Relationship-aware Multi-task Meta-review Generator), a meta-review generation model that uses sparse attention based on the hierarchical relationships and a multi-task objective that predicts several metadata features in addition to the standard text generation objective. Our experimental results show that PeerSum is a challenging dataset, and Rammer outperforms other strong baseline MDS models under various evaluation metrics.

Abstract: In this paper, we release a largest ever medical Question Answering (QA) dataset with 26 million QA pairs. We benchmark many existing approaches in our dataset in terms of both retrieval and generation. Experimental results show that the existing models perform far lower than expected and the released dataset is still challenging in the pre-trained language model era. Moreover, we also experimentally show the benefit of the proposed dataset in many aspects: (i) trained models for other QA datasets in a zero-shot fashion; and (ii) as external knowledge for retrieval-augmented generation (RAG); and (iii) improving existing pre-trained language models by using the QA pairs as a pre-training corpus in continued training manner. We believe that this dataset will not only contribute to medical research but also facilitate both the patients and clinical doctors. See \url{https://github.com/FreedomIntelligence/Huatuo-26M}.

Abstract: Dungeons & Dragons (D&D) is a tabletop roleplaying game with complex natural language interactions between players and hidden state information. Recent work has shown that large language models (LLMs) that have access to state information can generate higher quality game turns than LLMs that use dialog history alone. However, previous work used game state information that was heuristically created and was not a true gold standard game state. We present FIREBALL, a large dataset containing nearly 25,000 unique sessions from real D\&D gameplay on Discord with true game state info. We recorded game play sessions of players who used the Avrae bot, which was developed to aid people in playing D&D online, capturing language, game commands and underlying game state information. We demonstrate that FIREBALL can improve natural language generation (NLG) by using Avrae state information, improving both automated metrics and human judgments of quality. Additionally, we show that LLMs can generate executable Avrae commands, particularly after finetuning.

Abstract: Large Language models (LLMs) are trained on large amounts of data, which can include sensitive information that may compromise personal privacy. LLMs showed to memorize parts of the training data and emit those data verbatim when an adversary prompts appropriately. Previous research has primarily focused on data preprocessing and differential privacy techniques to address memorization or prevent verbatim memorization exclusively, which can give a false sense of privacy. However, these methods rely on explicit and implicit assumptions about the structure of the data to be protected, which often results in an incomplete solution to the problem. To address this, we propose a novel framework that utilizes a reinforcement learning approach (PPO) to fine-tune LLMs to mitigate approximate memorization. Our approach utilizes a negative similarity score, such as BERTScore or SacreBLEU, as a reward signal to learn a dissimilarity policy. Our results demonstrate that this framework effectively mitigates approximate memorization while maintaining high levels of coherence and fluency in the generated samples. Furthermore, our framework is robust in mitigating approximate memorization across various circumstances, including longer context, which is known to increase memorization in LLMs.

Abstract: Scaling language models have revolutionized widespread NLP tasks, yet little comprehensively explored few-shot relation extraction with large language models. In this paper, we investigate principal methodologies, in-context learning and data generation, for few-shot relation extraction via GPT-3.5 through exhaustive experiments. To enhance few-shot performance, we further propose task-related instructions and schema-constrained data generation. We observe that in-context learning can achieve performance on par with previous prompt learning approaches, and data generation with the large language model can boost previous solutions to obtain new state-of-the-art few-shot results on four widely-studied relation extraction datasets. We hope our work can inspire future research for the capabilities of large language models in few-shot relation extraction. Code is available in \url{https://github.com/zjunlp/DeepKE/tree/main/example/llm.

Abstract: Entity alignment(EA) is a crucial task for integrating cross-lingual and cross-domain knowledge graphs(KGs), which aims to discover entities referring to the same real-world object from different KGs. Most existing methods generate aligning entity representation by mining the relevance of triple elements via embedding-based methods, paying little attention to triple indivisibility and entity role diversity. In this paper, a novel framework named TTEA -- Type-enhanced Ensemble Triple Representation via Triple-aware Attention for Cross-lingual Entity Alignment is proposed to overcome the above issues considering ensemble triple specificity and entity role features. Specifically, the ensemble triple representation is derived by regarding relation as information carrier between semantic space and type space, and hence the noise influence during spatial transformation and information propagation can be smoothly controlled via specificity-aware triple attention. Moreover, our framework uses triple-ware entity enhancement to model the role diversity of triple elements. Extensive experiments on three real-world cross-lingual datasets demonstrate that our framework outperforms state-of-the-art methods.

Abstract: Cross-lingual and cross-domain knowledge alignment without sufficient external resources is a fundamental and crucial task for fusing irregular data. As the element-wise fusion process aiming to discover equivalent objects from different knowledge graphs (KGs), entity alignment (EA) has been attracting great interest from industry and academic research recent years. Most of existing EA methods usually explore the correlation between entities and relations through neighbor nodes, structural information and external resources. However, the complex intrinsic interactions among triple elements and role information are rarely modeled in these methods, which may lead to the inadequate illustration for triple. In addition, external resources are usually unavailable in some scenarios especially cross-lingual and cross-domain applications, which reflects the little scalability of these methods. To tackle the above insufficiency, a novel universal EA framework (OTIEA) based on ontology pair and role enhancement mechanism via triple-aware attention is proposed in this paper without introducing external resources. Specifically, an ontology-enhanced triple encoder is designed via mining intrinsic correlations and ontology pair information instead of independent elements. In addition, the EA-oriented representations can be obtained in triple-aware entity decoder by fusing role diversity. Finally, a bidirectional iterative alignment strategy is deployed to expand seed entity pairs. The experimental results on three real-world datasets show that our framework achieves a competitive performance compared with baselines.

Abstract: Most existing retrieval-augmented language models (LMs) for question answering assume all retrieved information is factually correct. In this work, we study a more realistic scenario in which retrieved documents may contain misinformation, causing conflicts among them. We observe that the existing models are highly brittle to such information in both fine-tuning and in-context few-shot learning settings. We propose approaches to make retrieval-augmented LMs robust to misinformation by explicitly fine-tuning a discriminator or prompting to elicit discrimination capability in GPT-3. Our empirical results on open-domain question answering show that these approaches significantly improve LMs' robustness to knowledge conflicts. We also provide our findings on interleaving the fine-tuned model's decision with the in-context learning process, paving a new path to leverage the best of both worlds.

Abstract: Pre-trained language models (PLMs) have achieved remarkable success in NLP tasks. Despite the great success, mainstream solutions largely follow the pre-training then finetuning paradigm, which brings in both high deployment costs and low training efficiency. Nevertheless, fine-tuning on a specific task is essential because PLMs are only pre-trained with language signal from large raw data. In this paper, we propose a novel fine-tuning-free strategy for language models, to consider both language signal and teacher signal. Teacher signal is an abstraction of a battery of downstream tasks, provided in a unified proposition format. Trained with both language and strong task-aware teacher signals in an interactive manner, our FreeLM model demonstrates strong generalization and robustness. FreeLM outperforms large models e.g., GPT-3 and InstructGPT, on a range of language understanding tasks in experiments. FreeLM is much smaller with 0.3B parameters, compared to 175B in these models.

Abstract: Recently there have been efforts to introduce new benchmark tasks for spoken language understanding (SLU), like semantic parsing. In this paper, we describe our proposed spoken semantic parsing system for the quality track (Track 1) in Spoken Language Understanding Grand Challenge which is part of ICASSP Signal Processing Grand Challenge 2023. We experiment with both end-to-end and pipeline systems for this task. Strong automatic speech recognition (ASR) models like Whisper and pretrained Language models (LM) like BART are utilized inside our SLU framework to boost performance. We also investigate the output level combination of various models to get an exact match accuracy of 80.8, which won the 1st place at the challenge.

Abstract: Recent research demonstrates that external knowledge injection can advance pre-trained language models (PLMs) in a variety of downstream NLP tasks. However, existing knowledge injection methods are either applicable to structured knowledge or unstructured knowledge, lacking a unified usage. In this paper, we propose a UNified knowledge inTERface, UNTER, to provide a unified perspective to exploit both structured knowledge and unstructured knowledge. In UNTER, we adopt the decoder as a unified knowledge interface, aligning span representations obtained from the encoder with their corresponding knowledge. This approach enables the encoder to uniformly invoke span-related knowledge from its parameters for downstream applications. Experimental results show that, with both forms of knowledge injected, UNTER gains continuous improvements on a series of knowledge-driven NLP tasks, including entity typing, named entity recognition and relation extraction, especially in low-resource scenarios.

Abstract: Transformer-based models typically have a predefined bound to their input length, because of their need to potentially attend to every token in the input. In this work, we propose Unlimiformer: a general approach that can wrap any existing pretrained encoder-decoder transformer, and offload the attention computation across all layers to a single $k$-nearest-neighbor index; this index can be kept on either the GPU or CPU memory and queried in sub-linear time. This way, we can index extremely long input sequences, while every attention head in every decoder layer retrieves its top-$k$ keys, instead of attending to every key. We demonstrate Unlimiformers's efficacy on several long-document and multi-document summarization benchmarks, showing that it can summarize even 350k token-long inputs from the BookSum dataset, without any input truncation at test time. Unlimiformer improves pretrained models such as BART and Longformer by extending them to unlimited inputs without additional learned weights and without modifying their code. We make our code and models publicly available at https://github.com/abertsch72/unlimiformer .

Abstract: Computational models of syntax are predominantly text-based. Here we propose that basic syntax can be modeled directly from raw speech in a fully unsupervised way. We focus on one of the most ubiquitous and basic properties of syntax -- concatenation. We introduce spontaneous concatenation: a phenomenon where convolutional neural networks (CNNs) trained on acoustic recordings of individual words start generating outputs with two or even three words concatenated without ever accessing data with multiple words in the input. Additionally, networks trained on two words learn to embed words into novel unobserved word combinations. To our knowledge, this is a previously unreported property of CNNs trained on raw speech in the Generative Adversarial Network setting and has implications both for our understanding of how these architectures learn as well as for modeling syntax and its evolution from raw acoustic inputs.

Abstract: Applying language models to natural language processing tasks typically relies on the representations in the final model layer, as intermediate hidden layer representations are presumed to be less informative. In this work, we argue that due to the gradual improvement across model layers, additional information can be gleaned from the contrast between higher and lower layers during inference. Specifically, in choosing between the probable next token predictions of a generative model, the predictions of lower layers can be used to highlight which candidates are best avoided. We propose a novel approach that utilizes the contrast between layers to improve text generation outputs, and show that it mitigates degenerative behaviors of the model in open-ended generation, significantly improving the quality of generated texts. Furthermore, our results indicate that contrasting between model layers at inference time can yield substantial benefits to certain aspects of general language model capabilities, more effectively extracting knowledge during inference from a given set of model parameters.

Abstract: We report our efforts in identifying a set of previous human evaluations in NLP that would be suitable for a coordinated study examining what makes human evaluations in NLP more/less reproducible. We present our results and findings, which include that just 13\% of papers had (i) sufficiently low barriers to reproduction, and (ii) enough obtainable information, to be considered for reproduction, and that all but one of the experiments we selected for reproduction was discovered to have flaws that made the meaningfulness of conducting a reproduction questionable. As a result, we had to change our coordinated study design from a reproduce approach to a standardise-then-reproduce-twice approach. Our overall (negative) finding that the great majority of human evaluations in NLP is not repeatable and/or not reproducible and/or too flawed to justify reproduction, paints a dire picture, but presents an opportunity for a rethink about how to design and report human evaluations in NLP.

Abstract: Fine-tuning large models is highly effective, however, inference using these models can be expensive and produces carbon emissions. Knowledge distillation has been shown to be a practical solution to reduce inference costs, but the distillation process itself requires significant computational resources. Rather than buying or renting GPUs to fine-tune, then distill a large model, an NLP practitioner who needs a compact model might also choose to simply allocate an available budget to hire annotators and manually label additional fine-tuning data. In this paper, we investigate how to most efficiently use a fixed budget to build a compact model. Through our extensive experiments on six diverse NLP tasks, we find that distilling from T5-XXL (11B) to T5-Small (60M) leads to almost always a cost-efficient option compared to annotating more data to directly train a compact model (T5-Small (60M)). We further demonstrate that the optimal amount of distillation that maximizes utility varies across different budgetary scenarios.

Abstract: Pre-trained language models (LMs) are used for knowledge intensive tasks like question answering, but their knowledge gets continuously outdated as the world changes. Prior work has studied targeted updates to LMs, injecting individual facts and evaluating whether the model learns these facts while not changing predictions on other contexts. We take a step forward and study LMs' abilities to make inferences based on injected facts (or propagate those facts): for example, after learning that something is a TV show, does an LM predict that you can watch it? We study this with two cloze-style tasks: an existing dataset of real-world sentences about novel entities (ECBD) as well as a new controlled benchmark with manually designed templates requiring varying levels of inference about injected knowledge. Surprisingly, we find that existing methods for updating knowledge (gradient-based fine-tuning and modifications of this approach) show little propagation of injected knowledge. These methods improve performance on cloze instances only when there is lexical overlap between injected facts and target inferences. Yet, prepending entity definitions in an LM's context improves performance across all settings, suggesting that there is substantial headroom for parameter-updating approaches for knowledge injection.