EMNLP 2024 论文列表

Journalists engage in multiple steps in the news writing process that depend on human creativity, like exploring different “angles” (i.e. the specific perspectives a reporter takes). These can potentially be aided by large language models (LLMs). By affecting planning decisions, such interventions can have an outsize impact on creative output. We advocate a careful approach to evaluating these interventions to ensure alignment with human values.In a case study of journalistic coverage of press releases, we assemble a large dataset of 250k press releases and 650k articles covering them. We develop methods to identify news articles that _challenge and contextualize_ press releases. Finally, we evaluate suggestions made by LLMs for these articles and compare these with decisions made by human journalists. Our findings are three-fold: (1) Human-written news articles that challenge and contextualize press releases more take more creative angles and use more informational sources. (2) LLMs align better with humans when recommending angles, compared with informational sources. (3) Both the angles and sources LLMs suggest are significantly less creative than humans.

Large Language Models (LLMs) have shown remarkable capabilities in various natural language processing tasks. However, LLMs may rely on dataset biases as shortcuts for prediction, which can significantly impair their robustness and generalization capabilities. This paper presents Shortcut Suite, a comprehensive test suite designed to evaluate the impact of shortcuts on LLMs’ performance, incorporating six shortcut types, five evaluation metrics, and four prompting strategies. Our extensive experiments yield several key findings: 1) LLMs demonstrate varying reliance on shortcuts for downstream tasks, which significantly impairs their performance. 2) Larger LLMs are more likely to utilize shortcuts under zero-shot and few-shot in-context learning prompts. 3) Chain-of-thought prompting notably reduces shortcut reliance and outperforms other prompting strategies, while few-shot prompts generally underperform compared to zero-shot prompts. 4) LLMs often exhibit overconfidence in their predictions, especially when dealing with datasets that contain shortcuts. 5) LLMs generally have a lower explanation quality in shortcut-laden datasets, with errors falling into three types: distraction, disguised comprehension, and logical fallacy. Our findings offer new insights for evaluating robustness and generalization in LLMs and suggest potential directions for mitigating the reliance on shortcuts.

Prior study shows that LLMs sometimes generate content that violates copyright. In this paper, we study another important yet underexplored problem, i.e., will LLMs respect copyright information in user input, and behave accordingly? The research problem is critical, as a negative answer would imply that LLMs will become the primary facilitator and accelerator of copyright infringement behavior. We conducted a series of experiments using a diverse set of language models, user prompts, and copyrighted materials, including books, news articles, API documentation, and movie scripts. Our study offers a conservative evaluation of the extent to which language models may infringe upon copyrights when processing user input containing protected material. This research emphasizes the need for further investigation and the importance of ensuring LLMs respect copyright regulations when handling user input to prevent unauthorized use or reproduction of protected content. We also release a benchmark dataset serving as a test bed for evaluating infringement behaviors by LLMs and stress the need for future alignment.

The variations between in-group and out-group speech (intergroup bias) are subtle and could underlie many social phenomena like stereotype perpetuation and implicit bias. In this paper, we model intergroup bias as a tagging task on English sports comments from forums dedicated to fandom for NFL teams. We curate a dataset of over 6 million game-time comments from opposing perspectives (the teams in the game), each comment grounded in a non-linguistic description of the events that precipitated these comments (live win probabilities for each team). Expert and crowd annotations justify modeling the bias through tagging of implicit and explicit referring expressions and reveal the rich, contextual understanding of language and the world required for this task. For large-scale analysis of intergroup variation, we use LLMs for automated tagging, and discover that LLMs occasionally perform better when prompted with linguistic descriptions of the win probability at the time of the comment, rather than numerical probability. Further, large-scale tagging of comments using LLMs uncovers linear variations in the form of referent across win probabilities that distinguish in-group and out-group utterances.

Emotions play important epistemological and cognitive roles in our lives, revealing our values and guiding our actions. Previous work has shown that LLMs display biases in emotion attribution along gender lines. However, unlike gender, which says little about our values, religion, as a socio-cultural system, prescribes a set of beliefs and values for its followers. Religions, therefore, cultivate certain emotions. Moreover, these rules are explicitly laid out and interpreted by religious leaders. Using emotion attribution, we explore how different religions are represented in LLMs. We find that:Major religions in the US and European countries are represented with more nuance, displaying a more shaded model of their beliefs.Eastern religions like Hinduism and Buddhism are strongly stereotyped.Judaism and Islam are stigmatized – the models’ refusal skyrocket. We ascribe these to cultural bias in LLMs and the scarcity of NLP literature on religion. In the rare instances where religion is discussed, it is often in the context of toxic language, perpetuating the perception of these religions as inherently toxic. This finding underscores the urgent need to address and rectify these biases. Our research emphasizes the crucial role emotions play in shaping our lives and how our values influence them.

Recent methods have demonstrated that Large Language Models (LLMs) can solve reasoning tasks better when they are encouraged to solve subtasks of the main task first. In this paper we devise a similar strategy that breaks down reasoning tasks into a problem decomposition phase and a problem solving phase and show that the strategy is able to outperform a single stage solution. Further, we hypothesize that the decomposition should be easier to distill into a smaller model compared to the problem solving because the latter requires large amounts of domain knowledge while the former only requires learning general problem solving strategies. We propose methods to distill these two capabilities and evaluate their impact on reasoning outcomes and inference cost. We find that we can distill the problem decomposition phase and at the same time achieve good generalization across tasks, datasets, and models. However, it is harder to distill the problem solving capability without losing performance and the resulting distilled model struggles with generalization. These results indicate that by using smaller, distilled problem decomposition models in combination with problem solving LLMs we can achieve reasoning with cost-efficient inference and local adaptation.

Evaluating the quality and consistency of text generated by Large Language Models (LLMs) poses a significant, yet unresolved challenge for industry research. We propose , an automated framework for evaluating and improving the consistency of LLM-generated texts using a divide-conquer-reasoning approach. Unlike existing LLM-based evaluators operating at the paragraph level, our method employs a divide-and-conquer evaluator () that breaks down the paragraph-to-paragraph comparison into sentence-to-paragraph comparisons. To facilitate this approach, we also introduce an automatic metric converter () that translates the output from into an interpretable numeric score. Beyond the consistency evaluation, we further present a reason-assisted improver () that mitigates inconsistencies by leveraging the analytical reasons identified by . Through comprehensive and systematic empirical analysis, we show that our approach outperforms state-of-the-art methods by a large margin (e.g., +16.8% and +32.5% on the SummEval dataset) in consistency evaluation across multiple benchmarks. Our approach also substantially reduces nearly 90% output inconsistencies in one iteration, showing promise for effective hallucination mitigation in real-world industrial applications.

The Criminal Court View Generation task aims to produce explanations that inform judicial decisions. This necessitates a nuanced understanding of diverse legal concepts, such as Recidivism, Confess, and Robbery, which often coexist within cases, complicating holistic analysis. However, existing methods mainly rely on the generation capability of language models, without paying enough attention to the important legal concepts.To enhance the precision and depth of such explanations, we introduce Legal Concept-guided Criminal Court Views Generation (LeGen), a three-stage approach designed for iterative reasoning tailored to individual legal constructs.Specifically, in the first stage, we design a decomposer to divide the court views into focused sub-views, each anchored around a distinct legal concept. Next, a concept reasoning module generates targeted rationales by intertwining the deconstructed facts with their corresponding legal frameworks, ensuring contextually relevant interpretations.Finally, a verifier and a generator are employed to align the rationale with the case fact and obtain synthesized comprehensive and legally sound final court views, respectively.We evaluate LeGen by conducting extensive experiments on a real-world dataset and experimental results validate the effectiveness of our proposed model. Our codes are available at https://anonymous.4open.science/r/LeGen-5625.

The automation of radiology report generation (RRG) holds immense potential to alleviate radiologists’ workloads and improve diagnostic accuracy. Despite advancements in image captioning and vision-language pretraining, RRG remains challenging due to the lengthy and complex nature of radiology reports. In this work, we proposes the Divide and Conquer Radiology Report Generation (DCRRG) model, which breaks down full-text radiology reports into concise observation descriptions. This approach enables the model to capture fine-grained representations from each observation through a two-stage process: an encoding stage focusing on observation prediction tasks to learn fine-grained representations, and a decoding stage for integrating these descriptions into cohesive and comprehensive radiology reports. Experimental results on two benchmark datasets demonstrate that DCRRG achieves significant improvements across all evaluated metrics, underscoring its capability to generate semantically coherent and clinically accurate radiology reports.

Topic-Dependent Argument Mining (TDAM), that is extracting and classifying argument components for a specific topic from large document sources, is an inherently difficult task for machine learning models and humans alike, as large TDAM datasets are rare and recognition of argument components requires expert knowledge. The task becomes even more difficult if it also involves stance detection of retrieved arguments. In this work, we investigate the effect of TDAM dataset composition in few- and zero-shot settings. Our findings show that, while fine-tuning is mandatory to achieve acceptable model performance, using carefully composed training samples and reducing the training sample size by up to almost 90% can still yield 95% of the maximum performance. This gain is consistent across three TDAM tasks on three different datasets. We also publish a new dataset and code for future benchmarking.

Knowledge-based Visual Qustion-answering (K-VQA) often requires the use of background knowledge beyond the image. However, we discover that a single knowledge generation strategy is often insuffcient for all K-VQA questions. To this end, we propose Diversifcation, Evidence Truncation, and Combination for Knowledge-based Elucidation (DietCoke), which utilizes a bundle of complementary question-answering tactics and aggregates their answers using textual rationales. DietCoke comprises of three stages: diversifcation, rationalization, and ensemble. The diversification stage generates three distinctive decision contexts, each leading to its own answer candidate. The rationalization stage generates two rationales, the automatic rationale and the mechanistic rationale, for each answer candidate using decorrelated techniques. Finally, in the ensemble stage, an LLM informed by the rationales selects one answer from the three candidates. Experiments show that DietCoke significantly outperforms state-of-the-art LLM-based baselines by 2.8% on OK-VOA and 4.7% on A-OKVOA and that the strategies in the ensembles are highly complementary.

We demonstrate substantial performance gains in zero-shot dialogue state tracking (DST) by enhancing training data diversity through synthetic data generation.Existing DST datasets are severely limited in the number of application domains and slot types they cover due to the high costs of data collection, restricting their adaptability to new domains.This work addresses this challenge with a novel, fully automatic data generation approach that creates synthetic zero-shot DST datasets.Distinguished from previous methods, our approach can generate dialogues across a massive range of application domains, complete with silver-standard dialogue state annotations and slot descriptions.This technique is used to create the D0T dataset for training zero-shot DST models, encompassing an unprecedented 1,000+ domains. Experiments on the MultiWOZ benchmark show that training models on diverse synthetic data improves Joint Goal Accuracy by 6.7%, achieving results competitive with models 13.5 times larger than ours.

Subword regularization, used widely in NLP, improves model performance by reducing the dependency on exact tokenizations, augmenting the training corpus, and exposing the model to more unique contexts during training. BPE and MaxMatch, two popular subword tokenization schemes, have stochastic dropout regularization variants. However, there has not been an analysis of the distributions formed by them.We show that these stochastic variants are heavily biased towards a small set of tokenizations per word. If the benefits of subword regularization are as mentioned, we hypothesize that biasedness artificially limits the effectiveness of these schemes. Thus, we propose an algorithm to uniformly sample tokenizations that we use as a drop-in replacement for the stochastic aspects of existing tokenizers, and find that it improves machine translation quality.

The distractor generation task focuses on generating incorrect but plausible options for objective questions such as fill-in-the-blank and multiple-choice questions. This task is widely utilized in educational settings across various domains and subjects. The effectiveness of these questions in assessments relies on the quality of the distractors, as they challenge examinees to select the correct answer from a set of misleading options. The evolution of artificial intelligence (AI) has transitioned the task from traditional methods to the use of neural networks and pre-trained language models. This shift has established new benchmarks and expanded the use of advanced deep learning methods in generating distractors. This survey explores distractor generation tasks, datasets, methods, and current evaluation metrics for English objective questions, covering both text-based and multi-modal domains. It also evaluates existing AI models and benchmarks and discusses potential future research directions.

Extensive efforts have been made before the public release of Large language models (LLMs) to align their behaviors with human values. However, even meticulously aligned LLMs remain vulnerable to malicious manipulations such as jailbreaking, leading to unintended behaviors. In this work, we propose a novel black-box jailbreak framework for automated red teaming of LLMs. We designed malicious content concealing and memory reframing with an iterative optimization algorithm to jailbreak LLMs, motivated by the research about the distractibility and over-confidence phenomenon of LLMs. Extensive experiments of jailbreaking both open-source and proprietary LLMs demonstrate the superiority of our framework in terms of effectiveness, scalability and transferability. We also evaluate the effectiveness of existing jailbreak defense methods against our attack and highlight the crucial need to develop more effective and practical defense strategies.

Image-text contrastive models like CLIP have wide applications in zero-shot classification, image-text retrieval, and transfer learning. However, they often struggle on compositional visio-linguistic tasks (e.g., attribute-binding or object-relationships) where their performance is no better than random chance. To address this, we introduce SDS-CLIP, a lightweight and sample-efficient distillation method to enhance CLIP’s compositional visio-linguistic reasoning. Our approach fine-tunes CLIP using a distillation objective borrowed from large text-to-image generative models like Stable-Diffusion, which are known for their strong visio-linguistic reasoning abilities. On the challenging Winoground benchmark, SDS-CLIP improves the visio-linguistic performance of various CLIP models by up to 7%, while on the ARO dataset, it boosts performance by up to 3%. This work underscores the potential of well-designed distillation objectives from generative models to enhance contrastive image-text models with improved visio-linguistic reasoning capabilities.

Instruction tuning aims to align large language models (LLMs) with open-domain instructions and human-preferred responses. While several studies have explored autonomous approaches to distilling and annotating instructions from powerful proprietary LLMs, such as ChatGPT, they often neglect the impact of the distributions and characteristics of tasks, together with the varying difficulty of instructions in training sets. This oversight can lead to imbalanced knowledge capabilities and poor generalization powers of student LLMs. To address these challenges, we introduce Task-Aware Curriculum Planning for Instruction Refinement (TAPIR), a multi-round distillation framework that utilizes an oracle LLM to select instructions that are difficult for a student LLM to follow. To balance the student’s capabilities, task distributions in training sets are adjusted with responses automatically refined according to their corresponding tasks. In addition, by incorporating curriculum planning, our approach systematically escalates the difficulty levels of tasks, progressively enhancing the student LLM’s capabilities. We rigorously evaluate TAPIR using several widely recognized benchmarks (such as AlpacaEval 2.0, MT-Bench, etc.) and multiple student LLMs. Empirical results demonstrate that student LLMs, trained with our method and less training data, outperform larger instruction-tuned models and strong distillation baselines.

Language Models for text classification often produce overconfident predictions for both in-distribution and out-of-distribution samples, i.e., the model’s output probabilities do not match their accuracy. Prior work showed that simple post-hoc approaches are effective for mitigating this issue, but are not robust in noisy settings, e.g., when the distribution shift is caused by spelling mistakes. In this work, we propose Distance Aware Calibration (DAC), a post-hoc approach that changes the confidence scores of a Language Model leveraging the distance between new samples been evaluated and the in-domain training set. We show that using DAC on top of a Language Model can improve in-domain calibration, robustness to different kind of distribution shift and also the model’s ability to detect out-of-distribution samples. We provide an extensive evaluation on common text classification benchmark for both calibration and out-of-distribution detection tasks.

Fine-tuning-based unlearning methods prevail for erasing targeted harmful, sensitive, or copyrighted information within large language models while preserving overall capabilities. However, the true effectiveness of the methods is unclear. In this paper, we delve into the limitations of fine-tuning-based unlearning through activation patching and parameter restoration experiments. Our findings reveal that these methods alter the model’s knowledge retrieval process, rather than genuinely erasing the problematic knowledge embedded in the model parameters. Furthermore, behavioral tests demonstrate that the unlearning mechanisms inevitably impact the global behavior of the models, affecting unrelated knowledge or capabilities. Our work advocates the development of more resilient unlearning techniques for truly erasing knowledge.