ACL 2024 论文列表

Generative retrieval is a promising new paradigm in text retrieval that generates identifier strings of relevant passages as the retrieval target. This paradigm leverages powerful generative language models, distinct from traditional sparse or dense retrieval methods. In this work, we identify a viable direction to further enhance generative retrieval via distillation and propose a feasible framework, named DGR. DGR utilizes sophisticated ranking models, such as the cross-encoder, in a teacher role to supply a passage rank list, which captures the varying relevance degrees of passages instead of binary hard labels; subsequently, DGR employs a specially designed distilled RankNet loss to optimize the generative retrieval model, considering the passage rank order provided by the teacher model as labels. This framework only requires an additional distillation step to enhance current generative retrieval systems and does not add any burden to the inference stage. We conduct experiments on four public datasets, and the results indicate that DGR achieves state-of-the-art performance among the generative retrieval methods. Additionally, DGR demonstrates exceptional robustness and generalizability with various teacher models and distillation losses.

Among the recently emerged knowledge editing methods, in-context knowledge editing (IKE) has shown respectable abilities on knowledge editing in terms of generalization and specificity. Noting the promising advantages but unexplored issues of IKE, we propose **DistillMIKE** as a novel extension of IKE, i.e., editing **distill**ation of "**M**assive” **I**n-context **K**nowledge **E**diting in large language models (LLMs), mainly consisting of two expansions; 1) *Massive in-context knowledge editing (MIKE)*, which extends IKE to a massive editing task, aiming to inject not a single edit but a set of massive edits to LLMs; To preserve specificity, our key novel extension is a “selective” retrieval augmentation, where the retrieval-augmented IKE is only applied to “in-scope” examples, whereas the unedited model without IKE is employed for “out-of-scope” ones. 2) *Editing distillation* of MIKE using low-rank adaptation (LoRA), which distills editing abilities of MIKE to parameters of LLMs in a manner of eliminating the need of lengthy in-context demonstrations, thus removing the computational overhead encountered at the inference time. Experimental results on the zsRE and CounterFact datasets demonstrate that MIKE shows the state-of-the-art perfomrances and DistilMIKE show comparable performances with MIKE. Our code is available at https://github.com/JoveReCode/DistillMIKE.git.

Joint entity and relation extraction is a process that identifies entity pairs and their relations using a single model. We focus on the problem of joint extraction in distantly-labeled data, whose labels are generated by aligning entity mentions with the corresponding entity and relation tags using a knowledge base (KB). One key challenge is the presence of noisy labels arising from both incorrect entity and relation annotations, which significantly impairs the quality of supervised learning. Existing approaches, either considering only one source of noise or making decisions using external knowledge, cannot well-utilize significant information in the training data. We propose DENRL, a generalizable framework that 1) incorporates a lightweight transformer backbone into a sequence labeling scheme for joint tagging, and 2) employs a noise-robust framework that regularizes the tagging model with significant relation patterns and entity-relation dependencies, then iteratively self-adapts to instances with less noise from both sources. Surprisingly, experiments on two benchmark datasets show that DENRL, using merely its own parametric distribution and simple data-driven heuristics, outperforms strong baselines by a large margin with better interpretability.

As a relative quality comparison of model responses, human and Large Language Model (LLM) preferences serve as common alignment goals in model fine-tuning and criteria in evaluation. Yet, these preferences merely reflect broad tendencies, resulting in less explainable and controllable models with potential safety risks. In this work, we dissect the preferences of human and 32 different LLMs to understand their quantitative composition, using annotations from real-world user-model conversations for a fine-grained, scenario-wise analysis. We find that humans are less sensitive to errors, favor responses that support their stances, and show clear dislike when models admit their limits. On the contrary, advanced LLMs like GPT-4-Turbo emphasize correctness, clarity, and harmlessness more. Additionally, LLMs of similar sizes tend to exhibit similar preferences, regardless of their training methods, and fine-tuning for alignment does not significantly alter the preferences of pretrained-only LLMs. Finally, we show that preference-based evaluation can be intentionally manipulated. In both training-free and training-based settings, aligning a model with the preferences of judges boosts scores, while injecting the least preferred properties lowers them. This results in notable score shifts: up to 0.59 on MT-Bench (1-10 scale) and 31.94 on AlpacaEval 2.0 (0-100 scale), highlighting the significant impact of this strategic adaptation. We have made all resources of this project publicly available.

Supervised fine-tuning (SFT) on instruction-following corpus is a crucial approach toward the alignment of large language models (LLMs). However, the performance of LLMs on standard knowledge and reasoning benchmarks tends to suffer from deterioration at the latter stage of the SFT process, echoing the phenomenon of alignment tax. Through our pilot study, we put a hypothesis that the data biases are probably one cause behind the phenomenon. To address the issue, we introduce a simple disperse-then-merge framework. To be concrete, we disperse the instruction-following data into portions and then train multiple sub-models using different data portions. Lastly, we merge multiple models into a single one via model merging techniques. Despite its simplicity, our framework outperforms various sophisticated methods such as data curation and training regularization on a series of standard knowledge and reasoning benchmarks.

Automated disinformation generation is often listed as one of the risks of large language models (LLMs). The theoretical ability to flood the information space with disinformation content might have dramatic consequences for democratic societies around the world. This paper presents a comprehensive study of the disinformation capabilities of the current generation of LLMs to generate false news articles in English language. In our study, we evaluated the capabilities of 10 LLMs using 20 disinformation narratives. We evaluated several aspects of the LLMs: how well they are at generating news articles, how strongly they tend to agree or disagree with the disinformation narratives, how often they generate safety warnings, etc. We also evaluated the abilities of detection models to detect these articles as LLM-generated. We conclude that LLMs are able to generate convincing news articles that agree with dangerous disinformation narratives.

Reinforcement Learning from Human Feedback (RLHF) has been a crucial component in the recent success of Large Language Models. However, RLHF is know to exploit biases in human preferences, such as verbosity. A well-formatted and eloquent answer is often more highly rated by users, even when it is less helpful and objective. A number of approaches have been developed to control those biases in the classical RLHF literature, but the problem remains relatively under-explored for Direct Alignment Algorithms such as Direct Preference Optimization (DPO). Unlike classical RLHF, DPO does not train a separate reward model or use reinforcement learning directly, so previous approaches developed to control verbosity cannot be directly applied to this setting. Our work makes several contributions. For the first time, we study the length problem in the DPO setting, showing significant exploitation in DPO and linking it to out-of-distribution bootstrapping. We then develop a principled but simple regularization strategy that prevents length exploitation, while still maintaining improvements in model quality. We demonstrate these affects across datasets on summarization and dialogue, where we achieve up to 20% improvement in win rates when controlling for length, despite the GPT4 judge’s well-known verbosity bias.

Dialects introduce syntactic and lexical variations in language that occur in regional or social groups. Most NLP methods are not sensitive to such variations. This may lead to unfair behavior of the methods, conveying negative bias towards dialect speakers. While previous work has studied dialect-related fairness for aspects like hate speech, other aspects of biased language, such as lewdness, remain fully unexplored. To fill this gap, we investigate performance disparities between dialects in the detection of five aspects of biased language and how to mitigate them. To alleviate bias, we present a multitask learning approach that models dialect language as an auxiliary task to incorporate syntactic and lexical variations. In our experiments with African-American English dialect, we provide empirical evidence that complementing common learning approaches with dialect modeling improves their fairness. Furthermore, the results suggest that multitask learning achieves state-of-the-art performance and helps to detect properties of biased language more reliably.

Text style transfer (TST) is an important task in natural language generation, which aims to transfer the text style (e.g., sentiment) while keeping its semantic information. Due to the absence of parallel datasets for supervision, most existing studies have been conducted in an unsupervised manner, where the generated sentences often suffer from high semantic divergence and thus low semantic preservation. In this paper, we propose a novel disentanglement-based framework for TST named DisenTrans, where disentanglement means that we separate the attribute and content components in the natural language corpus and consider this task from these two perspectives. Concretely, we first create a disentangled Chain-of-Thought prompting procedure to synthesize parallel data and corresponding attribute components for supervision. Then we develop a disentanglement learning method with synthetic data, where two losses are designed to enhance the focus on attribute properties and constrain the semantic space, thereby benefiting style control and semantic preservation respectively. Instructed by the disentanglement concept, our framework creates valuable supervised information and utilizes it effectively in TST tasks. Extensive experiments on mainstream datasets present that our framework achieves significant performance with great sample efficiency.

While large language models have significantly enhanced the effectiveness of discourse relation classifications, it remains unclear whether their comprehension is faithful and reliable. We provide DiSQ, a new method for evaluating the faithfulness of understanding discourse based on question answering. We first employ in-context learning to annotate the reasoning for discourse comprehension, based on the connections among key events within the discourse. Following this, DiSQ interrogates the model with a sequence of questions to assess its grasp of core event relations, its resilience to counterfactual queries, as well as its consistency to its previous responses. then evaluate language models with different architectural designs using DiSQ, finding: (1) DiSQ presents a significant challenge for all models, with the top-performing GPT model attaining only 41% of the ideal performance in PDTB; (2) DiSQ is robust to domain shifts and paraphrase variations; (3) Open-source models generally lag behind their closed-source GPT counterparts, with notable exceptions being those enhanced with chat and code/math features; (4) Our analysis validates the effectiveness of explicitly signalled discourse connectives, the role of contextual information, and the benefits of using historical QA data.

Experimental methods for estimating the impacts of text on human evaluation have been widely used in the social sciences. However, researchers in experimental settings are usually limited to testing a small number of pre-specified text treatments. While efforts to mine unstructured texts for features that causally affect outcomes have been ongoing in recent years, these models have primarily focused on the topics or specific words of text, which may not always be the mechanism of the effect. We connect these efforts with NLP interpretability techniques and present a method for flexibly discovering clusters of similar text phrases that are predictive of human reactions to texts using convolutional neural networks. When used in an experimental setting, this method can identify text treatments and their effects under certain assumptions. We apply the method to two data sets. The first enables direct validation of the model’s ability to detect phrases known to cause the outcome. The second demonstrates its ability to flexibly discover text treatments with varying textual structures. In both cases, the model learns a greater variety of text treatments compared to benchmark methods, and these text features quantitatively meet or exceed the ability of benchmark methods to predict the outcome.

End-to-end argumentation mining (AM) aims to extract the argumentation structure including argumentation components and their argumentation relations from text. Recent developments in end-to-end AM models have demonstrated significant progress by redefining the AM task as a sequence generation task, exhibiting simplicity and competitive performance. Nevertheless, these models overlook the integration of supplementary discourse structure information, a crucial factor for comprehending argumentation structures, resulting in suboptimal outcomes. In this study, we propose the DENIM framework, which generates discourse structure-aware prefixes for each layer of the generation model. These prefixes imbue the generation-based AM model with discourse structures, thereby augmenting the overall generation process. Moreover, we introduce a multi-task prompt coupled with a three-step decoding strategy, aiming to optimize the efficiency and effectiveness of argumentation structure decoding. Extensive experiments and analyses on two benchmark datasets show that DENIM achieves state-of-the-art performances on two AM benchmarks.

Large language models (LLMs) internalize enormous parametric knowledge during pre-training. Concurrently, realistic applications necessitate external contextual knowledge to aid models on the underlying tasks. This raises a crucial dilemma known as knowledge conflicts, where the contextual knowledge clashes with the parametric knowledge. However, existing decoding works are specialized in resolving knowledge conflicts and could inadvertently deteriorate performance in absence of conflicts. In this paper, we propose an adaptive decoding method, termed as contextual information-entropy constraint decoding (COIECD), to discern whether the knowledge conflicts occur and resolve them. It can improve the model’s faithfulness to conflicting context, and simultaneously maintain high performance among non-conflicting context. Our experiments show that COIECD exhibits strong performance and robustness over knowledge conflicts in realistic datasets.

In parataxis languages like Chinese, word meanings are highly correlated with morphological knowledge, which can help to disambiguate word senses. However, in-depth exploration of morphological knowledge in previous word sense disambiguation (WSD) methods is still lacking due to the absence of publicly available resources. In this paper, we are motivated to enhance Chinese WSD with full morphological knowledge, including both word-formations and morphemes. We first construct the largest and releasable Chinese WSD resources, including the lexico-semantic inventories MorInv and WrdInv, a Chinese WSD dataset MiCLS, and an out-of-volcabulary (OOV) test set. Then, we propose a model, MorBERT, to fully leverage this morphology-informed knowledge for Chinese WSD and achieve a SOTA F1 of 92.18% in the task. Finally, we demonstrated the model’s robustness in low-resource settings and generalizability to OOV senses. These resources and methods may bring new insights into and solutions for various downstream tasks in both computational and humanistic fields.

Direct preference optimization (DPO) is a successful fine-tuning strategy for aligning large language models with human preferences without the need to train a reward model or employ reinforcement learning. DPO, as originally formulated, relies on binary preference data and fine-tunes a language model to increase the likelihood of a preferred response over a dispreferred response. However, not all preference pairs are equal. Sometimes, the preferred response is only slightly better than the dispreferred one. In other cases, the preference is much stronger. For instance, if a response contains harmful or toxic content, the annotator will have a strong preference for that response. In this paper, we propose a generalization of DPO, termed DPO with an offset (ODPO), that does not treat every preference pair equally during fine-tuning. Intuitively, ODPO requires the difference between the likelihood of the preferred and dispreferred response to be greater than an offset value. The offset is determined based on the extent to which one response is preferred over another. Our experiments on various tasks suggest that ODPO significantly outperforms DPO in aligning language models, especially when the number of preference pairs is limited.

While image captioning is an essential field of vision language models (VLM), a lack of continuity between the learning objective and final performance metrics of VLMs complicates their training and optimization. Reinforcement learning (RL) can directly optimize such metrics, but it is accompanied by a significant computational cost, making it difficult to apply to recent large-scale VLMs. In this paper, we propose Direct Metric Optimization (DMO), which is a lightweight final-metric-optimizing training method. We replace the computationally expensive exploration process in RL with an offline, diverse text data augmentation and show that self-supervised training on reward-weighted augmented data leads to direct and stable metric optimization. Our experiments demonstrate that DMO achieves performance comparable to those of the state-of-the-art RL method while saving hundreds of times more model forwarding iterations and greater amounts of computation time. This suggests that DMO constitutes a promising alternative for metric optimization in the era of large-scale VLMs.

Aligning large language models (LLMs) with human expectations without human-annotated preference data is an important problem. In this paper, we propose a method to evaluate the response preference by using the output probabilities of response pairs under contrastive prompt pairs, which could achieve better performance on LLaMA2-7B and LLaMA2-13B compared to RLAIF. Based on this, we propose an automatic alignment method, Direct Large Model Alignment (DLMA). First, we use contrastive prompt pairs to automatically generate preference data. Then, we continue to evaluate the generated preference data using contrastive prompt pairs and calculate a self-rewarding score. Finally, we use the DPO algorithm to effectively align LLMs by combining this self-rewarding score. In the experimental stage, our DLMA method could surpass the RLHF method without relying on human-annotated preference data.

Large language models (LLMs) have demonstrated strong reasoning abilities when prompted to generate chain-of-thought (CoT) explanations alongside answers. However, previous research on evaluating LLMs has solely focused on answer accuracy, neglecting the correctness of the generated CoT. In this paper, we delve deeper into the CoT reasoning capabilities of LLMs in multi-hop question answering by utilizing knowledge graphs (KGs). We propose a novel discriminative and generative CoT evaluation paradigm to assess LLMs’ knowledge of reasoning and the accuracy of the generated CoT. Through experiments conducted on 5 different families of LLMs across 2 multi-hop question-answering datasets, we find that LLMs possess sufficient knowledge to perform reasoning. However, there exists a significant disparity between answer accuracy and faithfulness of the CoT generated by LLMs, indicating that they often arrive at correct answers through incorrect reasoning.

While LLMs can provide reasoned explanations along with their answers, the nature and quality of those explanations are still poorly understood. In response, our goal is to define a detailed way of characterizing the explanation capabilities of modern models and to create a nuanced, interpretable explanation evaluation tool that can generate such characterizations automatically, without relying on expensive API calls or human annotations. Our approach is to (a) define the new task of explanation critiquing - identifying and categorizing any main flaw in an explanation and providing suggestions to address the flaw, (b) create a sizeable, human-verified dataset for this task, and (c) train an open-source, automatic critique model (called Digital Socrates) using this data. Through quantitative and qualitative analysis, we demonstrate how Digital Socrates is useful for revealing insights about student models by examining their reasoning chains, and how it can provide high-quality, nuanced, automatic evaluation of those model explanations for the first time. Digital Socrates thus fills an important gap in evaluation tools for understanding and improving the explanation behavior of models.