NAACL 2024 论文列表

Evaluating the veracity of everyday claims is time consuming and in some cases requires domain expertise. We empirically demonstrate that the commonly used fact checking pipeline, known as the retriever-reader, suffers from performance deterioration when it is trained on the labeled data from one domain and used in another domain. Afterwards, we delve into each component of the pipeline and propose novel algorithms to address this problem. We propose an adversarial algorithm to make the retriever component robust against distribution shift. Our core idea is to initially train a bi-encoder on the labeled source data, and then, to adversarially train two separate document and claim encoders using unlabeled target data. We then focus on the reader component and propose to train it such that it is insensitive towards the order of claims and evidence documents. Our empirical evaluations support the hypothesis that such a reader shows a higher robustness against distribution shift. To our knowledge, there is no publicly available multi-topic fact checking dataset. Thus, we propose a simple automatic method to re-purpose two well-known fact checking datasets. We then construct eight fact checking scenarios from these datasets, and compare our model to a set of strong baseline models, including recent domain adaptation models that use GPT4 for generating synthetic data.

Standard fine-tuning of language models typically performs well on in-distribution data, but suffers with generalization to distribution shifts. In this work, we aim to improve the generalization of adapter-based cross-lingual task transfer where such cross-language distribution shifts are imminent. We investigate scheduled unfreezing algorithms –originally proposed to mitigate catastrophic forgetting in transfer learning – for fine-tuning task adapters. Our experiments show that scheduled unfreezing methods close the gap to full fine-tuning and achieve stronger cross-lingual transfer performance, suggesting that these methods can go beyond just mitigating catastrophic forgetting. Next, aiming to understand these empirical findings, we investigate the learning dynamics of scheduled unfreezing using Fisher Information. Our experiments reveal that scheduled unfreezing induces different learning dynamics compared to standard fine-tuning, and provide evidence that the dynamics of Fisher Information during training correlate with cross-lingual generalization performance. We additionally propose a general scheduled unfreezing algorithm that achieves an average of 2 points improvement over four datasets compared to standard fine-tuning and provides empirical evidence for a theory-based justification of the heuristic unfreezing schedule for task adapter training.

Table Question Answering (TQA) aims at composing an answer to a question based on tabular data. While prior research has shown that TQA models lack robustness, understanding the underlying cause and nature of this issue remains predominantly unclear, posing a significant obstacle to the development of robust TQA systems. In this paper, we formalize three major desiderata for a fine-grained evaluation of robustness of TQA systems. They should (i) answer questions regardless of alterations in table structure, (ii) base their responses on the content of relevant cells rather than on biases, and (iii) demonstrate robust numerical reasoning capabilities. To investigate these aspects, we create and publish a novel TQA evaluation benchmark in English. Our extensive experimental analysis reveals that none of the examined state-of-the-art TQA systems consistently excels in these three aspects. Our benchmark is a crucial instrument for monitoring the behavior of TQA systems and paves the way for the development of robust TQA systems. We release our benchmark publicly.

Prompt-based methods have achieved promising results in most few-shot text classification tasks. However, for readability assessment tasks, traditional prompt methods lack crucial linguistic knowledge, which has already been proven to be essential.Moreover, previous studies on utilizing linguistic features have shown non-robust performance in few-shot settings and may even impair model performance.To address these issues, we propose a novel prompt-based tuning framework that incorporates rich linguistic knowledge, called Feature Prompt Tuning (FPT). Specifically, we extract linguistic features from the text and embed them into trainable soft prompts. Further, we devise a new loss function to calibrate the similarity ranking order between categories. Experimental results demonstrate that our proposed method FTPnot only exhibits a significant performance improvement over the prior best prompt-based tuning approaches, but also surpasses the previous leading methods that incorporate linguistic features. Also, our proposed model significantly outperforms the large language model gpt-3.5-turbo-16k in most cases. Our proposed method establishes a new architecture for prompt tuning that sheds light on how linguistic features can be easily adapted to linguistic-related tasks.

Planning is a crucial task for agents in task oriented dialogs (TODs). Human agents typically resolve user issues by following predefined workflows, decomposing workflow steps into actionable items, and performing actions by executing APIs in order; all of which require reasoning and planning. With the recent advances in LLMs, there have been increasing attempts to use them for task planning and API usage. However, the faithfulness of the plans to predefined workflows and API dependencies, is not guaranteed with LLMs. Moreover, workflows in real life are often custom-defined and prone to changes; hence, adaptation is desirable. To study this, we propose the problem of faithful planning in TODs that needs to resolve user intents by following predefined flows and preserving API dependencies. To solve this problem, we propose FLAP, a Flow-Adhering Planning algorithm based on constrained decoding with lookahead heuristic for LLMs. Our algorithm alleviates the need for finetuning LLMs using domain specific (plan/dependency) data, enables quick adaptation to predefined flows, and outperforms other decoding and prompting-based baselines. Further, our algorithm empowers smaller LLMs (≈7B) to perform at par larger LLMs (≈30B-40B).

This paper introduces FIRE (**FI**nancial **R**elation **E**xtraction), a sentence-level dataset of named entities and relations within the financial sector. Comprising 3,025 instances, the dataset encapsulates 13 named entity types along with 18 relation types. Sourced from public financial reports and financial news articles, FIRE captures a wide array of financial information about a business including, but not limited to, corporate structure, business model, revenue streams, and market activities such as acquisitions. The full dataset was labeled by a single annotator to minimize labeling noise. The labeling time for each sentence was recorded during the labeling process. We show how this feature, along with curriculum learning techniques, can be used to improved a model’s performance. The FIRE dataset is designed to serve as a valuable resource for training and evaluating machine learning algorithms in the domain of financial information extraction. The dataset and the code to reproduce our experimental results are available at https://github.com/hmhamad/FIRE. The repository for the labeling tool can be found at https://github.com/abhinav-kumar-thakur/relation-extraction-annotator.

Understanding event descriptions is a central aspect of language processing, but current approaches focus overwhelmingly on single sentences or documents. Aggregating information about an event across documents can offer a much richer understanding. To this end, we present FAMuS, a new corpus of Wikipedia passages that report on some event, paired with underlying, genre-diverse (non-Wikipedia) source articles for the same event. Events and (cross-sentence) arguments in both report and source are annotated against FrameNet, providing broad coverage of different event types. We present results on two key event understanding tasks enabled by FAMuS: source validation—determining whether a document is a valid source for a target report event—and cross-document argument extraction—full-document argument extraction for a target event from both its report and the correct source article.

In the evolving landscape of Neural Machine Translation (NMT), the pretrain-then-finetune paradigm has yielded impressive results. However, the persistent challenge of Catastrophic Forgetting (CF) remains a hurdle. While previous work has introduced Continual Learning (CL) methods to address CF, these approaches grapple with the delicate balance between avoiding forgetting and maintaining system extensibility. To address this, we propose a CL method, named F-MALLOC (Feed-forward Memory ALLOCation). F-MALLOC is inspired by recent insights highlighting that feed-forward layers emulate neural memories and encapsulate crucial translation knowledge. It decomposes feed-forward layers into discrete memory cells and allocates these memories to different tasks. By learning to allocate and safeguard these memories, our method effectively alleviates CF while ensuring robust extendability. Besides, we propose a comprehensive assessment protocol for multi-stage CL of NMT systems. Experiments conducted following this new protocol showcase the superior performance of F-MALLOC, evidenced by higher BLEU scores and almost zero forgetting.

Existing dense retrieval systems utilize the same model architecture for encoding both the passages and the queries, even though queries are much shorter and simpler than passages. This leads to high latency of the query encoding, which is performed online and therefore might impact user experience. We show that combining a standard large passage encoder with a small efficient query encoder can provide significant latency drops with only a small decrease in quality. We offer a pretraining and training solution for multiple small query encoder architectures. Using a small transformer architecture we are able to decrease latency by up to ∼12×, while MRR@10 on the MS MARCO dev set only decreases from 38.2 to 36.2. If this solution does not reach the desired latency requirements, we propose an efficient RNN as the query encoder, which processes the query prefix incrementally and only infers the last word after the query is issued. This shortens latency by ∼38× with only a minor drop in quality, reaching 35.5 MRR@10 score.

Extremely weakly-supervised text classification aims to classify texts without any labeled data, but only relying on class names as supervision. Existing works include prompt-based and seed-based methods. Prompt-based methods prompt language model with instructions, while seed-based methods generate pseudo-labels with word matching. Both of them have significant flaws, including zero-shot instability and context-dependent ambiguities. This paper introduces SetSync, which follows a new paradigm, i.e. wordset-based, which can avoid the above problems. In SetSync, a class is represented with wordsets, and pseudo-labels are generated with wordsets matching. To facilitate this, we propose to use information bottleneck to identify class-relevant wordsets. Moreover, we regard the classifier training as a hybrid learning of semi-supervised and noisy-labels, and propose a new training strategy, termed sync-denoising. Extensive experiments on 11 datasets show that SetSync outperforms all existing prompt and seed methods, exceeding SOTA by an impressive average of 8 points.

Standard extractive systems suffer from the lack of gold training signals since existing corpora solely provide document and human-written summary pairs while disregarding extractive labels. As a result, existing methods resort to imperfect pseudo-labels that are both biased and error-prone, thereby hindering the learning process of extractive models. In contrast, text generators which are commonly employed in abstractive summarization can effortlessly overcome this predicament on account of flexible sequence-to-sequence architectures. Motivated to bypass this inherent limitation, we investigate the possibility of conducting extractive summarization with text generators. Through extensive experiments covering six summarization benchmarks, we show that high-quality extractive summaries can be assembled via approximating the outputs (abstractive summaries) of these generators. Moreover, we find that the approximate summaries correlate positively with the auxiliary summaries (i.e. a better generator enables the production of better extractive summaries). Our results signify a new paradigm for training extractive summarizers i.e. learning with generation (abstractive) objectives rather than extractive schemes.

Identifying linguistic differences between dialects of a language often requires expert knowledge and meticulous human analysis. This is largely due to the complexity and nuance involved in studying various dialects. We present a novel approach to extract distinguishing lexical features of dialects by utilizing interpretable dialect classifiers, even in the absence of human experts. We explore both post-hoc and intrinsic approaches to interpretability, conduct experiments on Mandarin, Italian, and Low Saxon, and experimentally demonstrate that our method successfully identifies key language-specific lexical features that contribute to dialectal variations.

Effectively training language models on longinputs poses many technical challenges. As acost consideration, languages models are pre-trained on a fixed sequence length before beingadapted to longer sequences. We explore var-ious methods for adapting models to longerinputs by training on segmented sequences andan interpolation-based method for extendingabsolute positional embeddings. We developa training procedure to extend the input con-text size of pretrained models with no architec-tural changes and no additional memory coststhan training on the original input lengths. Bysub-sampling segments from long inputs whilemaintaining their original position the model isable to learn new positional interactions. Ourmethod benefits both models trained with abso-lute positional embeddings, by extending theirinput contexts, as well as popular relative posi-tional embedding methods showing a reducedperplexity on sequences longer than they weretrained on. We demonstrate our method canextend input contexts by a factor of 4× whileimproving perplexity.

Referring Image Segmentation (RIS) is a cross-modal task that aims to segment an instance described by a natural language expression. Recent methods leverage large-scale pretrained unimodal models as backbones along with fusion techniques for joint reasoning across modalities. However, the inherent cross-modal nature of RIS raises questions about the effectiveness of unimodal backbones. We propose RISCLIP, a novel framework that effectively leverages the cross-modal nature of CLIP for RIS. Observing CLIP’s inherent alignment between image and text features, we capitalize on this starting point and introduce simple but strong modules that enhance unimodal feature extraction and leverage rich alignment knowledge in CLIP’s image-text shared-embedding space. RISCLIP exhibits outstanding results on all three major RIS benchmarks and also outperforms previous CLIP-based methods, demonstrating the efficacy of our strategy in extending CLIP’s image-text alignment to RIS.

Saliency post-hoc explainability methods are important tools for understanding increasingly complex NLP models. While these methods can reflect the model’s reasoning, they may not align with human intuition, making the explanations not plausible. In this work, we present a methodology for incorporating rationales, which are text annotations explaining human decisions, into text classification models. This incorporation enhances the plausibility of post-hoc explanations while preserving their faithfulness. Our approach is agnostic to model architectures and explainability methods. We introduce the rationales during model training by augmenting the standard cross-entropy loss with a novel loss function inspired by contrastive learning. By leveraging a multi-objective optimization algorithm, we explore the trade-off between the two loss functions and generate a Pareto-optimal frontier of models that balance performance and plausibility. Through extensive experiments involving diverse models, datasets, and explainability methods, we demonstrate that our approach significantly enhances the quality of model explanations without causing substantial (sometimes negligible) degradation in the original model’s performance.

Augmenting Large Language Models (LLMs) for Question Answering (QA) with domain specific data has attracted wide attention. However, domain data often exists in a hybrid format, including text and semi-structured tables, posing challenges for the seamless integration of information. Table-to-Text Generation is a promising solution by facilitating the transformation of hybrid data into a uniformly text-formatted corpus. Although this technique has been widely studied by the NLP community, there is currently no comparative analysis on how corpora generated by different table-to-text methods affect the performance of QA systems.In this paper, we address this research gap in two steps. First, we innovatively integrate table-to-text generation into the framework of enhancing LLM-based QA systems with domain hybrid data. Then, we utilize this framework in real-world industrial data to conduct extensive experiments on two types of QA systems (DSFT and RAG frameworks) with four representative methods: Markdown format, Template serialization, TPLM-based method, and LLM-based method. Based on the experimental results, we draw some empirical findings and explore the underlying reasons behind the success of some methods. We hope the findings of this work will provide a valuable reference for the academic and industrial communities in developing robust QA systems.

LLMs (Large Language Models) usually interact with users in the form of dialogue and generate responses following their instructions, which naturally require dialogue comprehension abilities. However, dialogue comprehension is a general language ability which is hard to be evaluated directly. In this work, we propose to perform the evaluation focusing on the factual consistency issue with the help of the dialogue summarization task. Besides evaluating and analyzing the dialogue summarization performance (DIAC-Sum) of different LLMs, we also derive factual questions from the generated summaries and use them as a more flexible measurement of dialogue comprehension (DIAC-FactQA). Our evaluation shows that, on average, 26.8% of the summaries generated by LLMs contain factual inconsistency. Even ChatGPT, the strongest model evaluated, has such errors in 16% of its summaries. For answering the factual questions, which is more challenging, the average error rate of all evaluated LLMs is 36.1%. Both results indicate serious deficiencies. Detailed analysis shows that the understanding of subject/object of the conversation is still challenging for LLMs. Furthermore, to stimulate and enhance the dialogue comprehension ability of LLMs, we propose a fine-tuning paradigm with auto-constructed multi-task data, which achieved a relative error rate reduction of 11% on DIAC-FactQA.

Traditional attempts to enhance the logical reasoning abilities of language models often rely on supervised fine-tuning, limiting their generalization to new tasks or domains. Large Language Models (LLMs), with their capacity to condense vast knowledge, can effectively tackle many tasks. Yet, our experiments reveal a gap in their performance on logical reasoning benchmarks when compared to state-of-the-art fine-tuning based models. To bridge this gap, we present LogicLLM, a first-of-its-kind, fully self-supervised framework for integrating logical reasoning capabilities into LLMs, and activating them via in-context learning. We apply this to two LLM series, FLAN-T5 and LLaMA, with parameter sizes from 3 billion to 33 billion. LogicLLM demonstrates its effectiveness through successful improvements on two logical reasoning benchmarks (ReClor and LogiQA-v2). Additionally, LogicLLM based on FLAN-T5-11B attains comparable results to ChatGPT, and evaluations with LLaMA-based models on three language understanding benchmarks (RACE, MMLU and Big-Bench-Hard) confirm that the improvements come without compromising the model’s general language understanding capabilities.

Auxiliary function is a helpful component to improve language model’s code generation ability. However, a systematic exploration of how they affect has yet to be done. In this work, we comprehensively evaluate the ability to utilize auxiliary functions encoded in recent code-pretrained language models. First, we construct a human-crafted evaluation set, called HumanExtension, which contains examples of two functions where one function assists the other.With HumanExtension, we design several experiments to examine their ability in a multifaceted way. Our evaluation processes enable a comprehensive understanding of including auxiliary functions in the prompt in terms of effectiveness and robustness. An additional implementation style analysis captures the models’ various implementation patterns when they access the auxiliary function. Through this analysis, we discover the models’ promising ability to utilize auxiliary functions including their self-improving behavior by implementing the two functions step-by-step. However, our analysis also reveals the model’s underutilized behavior to call the auxiliary function, suggesting the future direction to enhance their implementation by eliciting the auxiliary function call ability encoded in the models. We release our code and dataset to facilitate this research direction.