ACL 2024 论文列表

Large language models (LLMs) suffer from catastrophic forgetting during continual learning. Conventional rehearsal-based methods rely on previous training data to retain the model’s ability, which may not be feasible in real-world applications. When conducting continual learning based on a publicly-released LLM checkpoint, the availability of the original training data may be non-existent. To address this challenge, we propose a framework called Self-Synthesized Rehearsal (SSR) that uses the LLM to generate synthetic instances for rehearsal. Concretely, we first employ the base LLM for in-context learning to generate synthetic instances. Subsequently, we utilize the latest LLM to refine the instance outputs based on the synthetic inputs, preserving its acquired ability. Finally, we select diverse high-quality synthetic instances for rehearsal in future stages. Experimental results demonstrate that SSR achieves superior or comparable performance compared to conventional rehearsal-based approaches while being more data-efficient. Besides, SSR effectively preserves the generalization capabilities of LLMs in general domains.

Text classification is a crucial task encountered frequently in practical scenarios, yet it is still under-explored in the era of large language models (LLMs). This study shows that LLMs are vulnerable to changes in the number and arrangement of options in text classification. Our extensive empirical analyses reveal that the key bottleneck arises from ambiguous decision boundaries and inherent biases towards specific tokens and positions.To mitigate these issues, we make the first attempt and propose a novel two-stage classification framework for LLMs. Our approach is grounded in the empirical observation that pairwise comparisons can effectively alleviate boundary ambiguity and inherent bias. Specifically, we begin with a self-reduction technique to efficiently narrow down numerous options, which contributes to reduced decision space and a faster comparison process. Subsequently, pairwise contrastive comparisons are employed in a chain-of-thought manner to draw out nuances and distinguish confusable options, thus refining the ambiguous decision boundary.Extensive experiments on four datasets (Banking77, HWU64, LIU54, and Clinic150) verify the effectiveness of our framework. Furthermore, benefitting from our framework, various LLMs can achieve consistent improvements. Our code and data are available in https://github.com/Chuge0335/PC-CoT.

Instruction-following language models often show undesirable biases. These undesirable biases may be accelerated in the real-world usage of language models, where a wide range of instructions is used through zero-shot example prompting. To solve this problem, we first define the bias neuron, which significantly affects biased outputs, and prove its existence empirically. Furthermore, we propose a novel and practical bias mitigation method, CRISPR, to eliminate bias neurons of language models in instruction-following settings. CRISPR automatically determines biased outputs and categorizes neurons that affect the biased outputs as bias neurons using an explainability method. Experimental results demonstrate the effectiveness of our method in mitigating biases under zero-shot instruction-following settings without losing the model’s task performance and existing knowledge. The experimental results reveal the generalizability of our method as it shows robustness under various instructions and datasets. Surprisingly, our method can mitigate the bias in language models by eliminating only a few neurons (at least three).

Lifelong prompt tuning has significantly advanced parameter-efficient lifelong learning with its efficiency and minimal storage demands on various tasks.Our empirical studies, however, highlights certain transferability constraints in the current methodologies: a universal algorithm that guarantees consistent positive transfer across all tasks is currently unattainable, especially when dealing dissimilar tasks that may engender negative transfer.Identifying the misalignment between algorithm selection and task specificity as the primary cause of negative transfer, we present the Similarity Heuristic Lifelong Prompt Tuning (SHLPT) framework. This innovative strategy partitions tasks into two distinct subsets by harnessing a learnable similarity metric, thereby facilitating fruitful transfer from tasks regardless of their similarity or dissimilarity. Additionally, SHLPT incorporates a parameter pool to combat catastrophic forgetting effectively. Our experiments shows that SHLPT outperforms state-of-the-art techniques in lifelong learning benchmarks and demonstrates robustness against negative transfer in diverse task sequences.

Chomsky and others have very directly claimed that large language models (LLMs) are equally capable of learning languages that are possible and impossible for humans to learn. However, there is very little published experimental evidence to support such a claim. Here, we develop a set of synthetic impossible languages of differing complexity, each designed by systematically altering English data with unnatural word orders and grammar rules. These languages lie on an impossibility continuum: at one end are languages that are inherently impossible, such as random and irreversible shuffles of English words, and on the other, languages that may not be intuitively impossible but are often considered so in linguistics, particularly those with rules based on counting word positions. We report on a wide range of evaluations to assess the capacity of GPT-2 small models to learn these uncontroversially impossible languages, and crucially, we perform these assessments at various stages throughout training to compare the learning process for each language. Our core finding is that GPT-2 struggles to learn impossible languages when compared to English as a control, challenging the core claim. More importantly, we hope our approach opens up a productive line of inquiry in which different LLM architectures are tested on a variety of impossible languages in an effort to learn more about how LLMs can be used as tools for these cognitive and typological investigations.

Health-related misinformation on social networks can lead to poor decision-making and real-world dangers. Such misinformation often misrepresents scientific publications and cites them as “proof” to gain perceived credibility. To effectively counter such claims automatically, a system must explain how the claim was falsely derived from the cited publication. Current methods for automated fact-checking or fallacy detection neglect to assess the (mis)used evidence in relation to misinformation claims, which is required to detect the mismatch between them. To address this gap, we introduce Missci, a novel argumentation theoretical model for fallacious reasoning together with a new dataset for real-world misinformation detection that misrepresents biomedical publications. Unlike previous fallacy detection datasets, Missci (i) focuses on implicit fallacies between the relevant content of the cited publication and the inaccurate claim, and (ii) requires models to verbalize the fallacious reasoning in addition to classifying it. We present Missci as a dataset to test the critical reasoning abilities of large language models (LLMs), that are required to reconstruct real-world fallacious arguments, in a zero-shot setting. We evaluate two representative LLMs and the impact of different levels of detail about the fallacy classes provided to the LLM via prompts. Our experiments and human evaluation show promising results for GPT 4, while also demonstrating the difficulty of this task.

While Large language models (LLMs) have the capability to iteratively reflect on their own outputs, recent studies have observed their struggles with knowledge-rich problems without access to external resources. In addition to the inefficiency of LLMs in self-assessment, we also observe that LLMs struggle to revisit their predictions despite receiving explicit negative feedback. Therefore, We propose Mirror, a Multiple-perspective self-reflection method for knowledge-rich reasoning, to avoid getting stuck at a particular reflection iteration. Mirror enables LLMs to reflect from multiple-perspective clues, achieved through a heuristic interaction between a Navigator and a Reasoner. It guides agents toward diverse yet plausibly reliable reasoning trajectory without access to ground truth by encouraging (1) diversity of directions generated by Navigator and (2) agreement among strategically induced perturbations in responses generated by the Reasoner. The experiments on five reasoning datasets demonstrate that Mirror’s superiority over several contemporary self-reflection approaches. Additionally, the ablation study studies clearly indicate that our strategies alleviate the aforementioned challenges.

Large language models (LLMs) have achieved remarkable performance in natural language understanding and generation tasks. However, they often suffer from limitations such as difficulty in incorporating new knowledge, generating hallucinations, and explaining their reasoning process. To address these challenges, we propose a novel prompting pipeline, named MindMap, that leverages knowledge graphs (KGs) to enhance LLMs’ inference and transparency. Our method enables LLMs to comprehend KG inputs and infer with a combination of implicit and external knowledge. Moreover, our method elicits the mind map of LLMs, which reveals their reasoning pathways based on the ontology of knowledge. We evaluate our method on diverse question & answering tasks, especially in medical domains, and show significant improvements over baselines. We also introduce a new hallucination evaluation benchmark and analyze the effects of different components of our method. Our results demonstrate the effectiveness and robustness of our method in merging knowledge from LLMs and KGs for combined inference.

Recent advances in few-shot question answering (QA) mostly rely on the power of pre-trained large language models (LLMs) and fine-tuning in specific settings. Although the pre-training stage has already equipped LLMs with powerful reasoning capabilities, LLMs still need to be fine-tuned to adapt to specific domains to achieve the best results. In this paper, we propose to select the most informative data for fine-tuning, thereby improving the efficiency of the fine-tuning process with comparative or even better accuracy on the open-domain QA task. We present MinPrompt, a minimal data augmentation framework for open-domain QA based on an approximate graph algorithm and unsupervised question generation. We transform the raw text into a graph structure to build connections between different factual sentences, then apply graph algorithms to identify the minimal set of sentences needed to cover the most information in the raw text. We then generate QA pairs based on the identified sentence subset and train the model on the selected sentences to obtain the final model. Empirical results on several benchmark datasets and theoretical analysis show that MinPrompt is able to achieve comparable or better results than baselines with a high degree of efficiency, bringing consistent improvements in F-1 scores.

Metaphors in natural language are a reflection of fundamental cognitive processes such as analogical reasoning and categorisation, and are deeply rooted in everyday communication. Metaphor understanding is therefore an essential task for large language models (LLMs). We release the Metaphor Understanding Challenge Dataset (MUNCH), designed to evaluate the metaphor understanding capabilities of LLMs. The dataset provides over 10k paraphrases for sentences containing metaphor use, as well as 1.5k instances containing inapt paraphrases. The inapt paraphrases were carefully selected to serve as control to determine whether the model indeed performs full metaphor interpretation or rather resorts to lexical similarity. All apt and inapt paraphrases were manually annotated. The metaphorical sentences cover natural metaphor uses across 4 genres (academic, news, fiction, and conversation), and they exhibit different levels of novelty. Experiments with LLaMA and GPT-3.5 demonstrate that MUNCH presents a challenging task for LLMs. The dataset is freely accessible at https://github.com/xiaoyuisrain/metaphor-understanding-challenge.

Fact-checking real-world claims often requires reviewing multiple multimodal documents in order to assess the claim’s truthfulness, a highly laborious and time-consuming task. In this paper, we present a summarization model crafted to generate claim-specific summaries useful for fact-checking from multimodal multi-document datasets. The model takes inputs in the form of documents, images, and a claim, with the objective of assisting in fact-checking tasks. We introduce a dynamic perceiver-based model that is able to handle inputs from multiple modalities of arbitrary lengths. To train our model, we leverage a novel reinforcement learning-based entailment objective in order to generate summaries that provide evidence distinguishing between different truthfulness labels. To assess the efficacy of our approach, we conduct experiments on both an existing benchmark as well as a new dataset of multi-document claims which we contribute. Our approach outperforms the SOTA approach by 4.6% in the claim verification task on the MOCHEG dataset and demonstrates strong performance on our new Multi-News-Fact-Checking dataset.

Metaphor interpretation is a difficult task in natural language understanding. The development of relevant techniques in this domain is slow, mostly because of the lack of large annotated datasets and effective pre-trained language models (PLMs) for metaphor learning. Thus, we propose a large annotated dataset and a PLM for the metaphor interpretation task. Our foundation model is based on a novel anomalous language modeling (ALM) method, which we benchmark with comparable PLM baselines on the new dataset, finding that it largely improves model performance on metaphor identification and interpretation.

Language style is often used by writers to convey their intentions, identities, and mastery of language. In this paper, we show that current large language models struggle to capture some language styles without fine-tuning. To address this challenge, we investigate whether LLMs can be meta-trained based on representative lexicons to recognize new styles they have not been fine-tuned on. Experiments on 13 established style classification tasks, as well as 63 novel tasks generated using LLMs, demonstrate that meta-training with style lexicons consistently improves zero-shot transfer across styles. We release the code and data at https://github.com/octaviaguo/Style-LLM.

We introduce a new unsupervised text embedding method, Meta-Task Prompting with Explicit One-Word Limitation (MetaEOL), for generating high-quality sentence embeddings from Large Language Models (LLMs) without the need for model fine-tuning. Leveraging meta-task prompting, MetaEOL guides LLMs to produce embeddings through a series of carefully designed prompts that address multiple representational aspects. Our comprehensive experiments demonstrate that embeddings averaged from various meta-tasks are versatile embeddings that yield competitive performance on Semantic Textual Similarity (STS) benchmarks and excel in downstream tasks, surpassing contrastive-trained models. Our findings suggest a new scaling law, offering a versatile and resource-efficient approach for embedding generation across diverse scenarios.

Neural-symbolic methods have demonstrated efficiency in enhancing the reasoning abilities of large language models (LLMs). However, existing methods mainly rely on syntactically mapping natural languages to complete formal languages like Python and SQL. Those methods require that reasoning tasks be convertible into programs, which cater to the computer execution mindset and deviate from human reasoning habits. To broaden symbolic methods’ applicability and adaptability in the real world, we propose Meta-Reasoning from a linguistic perspective. This method empowers LLMs to deconstruct reasoning-independent semantic information into generic symbolic representations, thereby efficiently capturing more generalized reasoning knowledge. We conduct extensive experiments on more than ten datasets encompassing conventional reasoning tasks like arithmetic, symbolic, and logical reasoning, and the more complex interactive reasoning tasks like theory-of-mind reasoning. Experimental results demonstrate that Meta-Reasoning significantly enhances in-context reasoning accuracy, learning efficiency, out-of-domain generalization, and output stability compared to the Chain-of-Thought technique.

Long-form generations from large language models (LLMs) contain a mix of factual and non-factual claims, making evaluating factuality difficult.Prior works evaluate the factuality of a long paragraph by decomposing it into multiple facts, verifying those facts independently, and aggregating the results.Such methods assume that combining factual claims forms a factual paragraph.The above assumption can be violated: we show that strong open-source models like Llama-chat can generate paragraphs that contain verifiable facts, but the facts are combined into a non-factual paragraph due to entity ambiguity.We further reveal that existing factuality metrics, including FActScore and citation recall, cannot properly evaluate these non-factual paragraphs and overestimate their factuality.To address this, we introduce an enhanced metric, **D-FActScore**, specifically designed for content with ambiguous entities.We evaluate the D-FActScores of people biographies generated by retrieval-augmented LLMs.We show that D-FActScore can better assess the factuality of paragraphs with entity ambiguity than FActScore.We also find that four widely used open-source LLMs tend to mix information of distinct entities to form non-factual paragraphs, making their D-FActScore much lower than FActScore by over 10%.

Mental manipulation, a significant form of abuse in interpersonal conversations, presents a challenge to identify due to its context-dependent and often subtle nature. The detection of manipulative language is essential for protecting potential victims, yet the field of Natural Language Processing (NLP) currently faces a scarcity of resources and research on this topic. Our study addresses this gap by introducing a new dataset, named MentalManip, which consists of 4,000 annotated fictional dialogues. This dataset enables a comprehensive analysis of mental manipulation, pinpointing both the techniques utilized for manipulation and the vulnerabilities targeted in victims. Our research further explores the effectiveness of leading-edge models in recognizing manipulative dialogue and its components through a series of experiments with various configurations. The results demonstrate that these models inadequately identify and categorize manipulative content. Attempts to improve their performance by fine-tuning with existing datasets on mental health and toxicity have not overcome these limitations. We anticipate that MentalManip will stimulate further research, leading to progress in both understanding and mitigating the impact of mental manipulation in conversations.

Multimodal Large Language Models (MLLMs) have demonstrated proficiency in handling a variety of visual-language tasks. However, current MLLM benchmarks are predominantly designed to evaluate reasoning based on static information about a single image, and the ability of modern MLLMs to extrapolate from image sequences, which is essential for understanding our ever-changing world, has been less investigated. To address this challenge, this paper introduces Mementos, a new benchmark designed to assess MLLMs’ sequential image reasoning abilities. Mementos features 4,761 diverse image sequences with varying lengths. We also employ a GPT-4 assisted method to evaluate MLLM reasoning performance. Through a careful evaluation of nine recent MLLMs on Mementos, including GPT-4V and Gemini, we find that they struggle to accurately describe dynamic information about given image sequences, often leading to hallucinations/misrepresentations of objects and their corresponding behaviors. Our quantitative analysis and case studies identify three key factors impacting MLLMs’ sequential image reasoning: the correlation between object and behavioral hallucinations, the influence of co-occurring behaviors, and the compounding impact of behavioral hallucinations.

Memes have evolved as a prevalent medium for diverse communication, ranging from humour to propaganda. With the rising popularity of image-focused content, there is a growing need to explore its potential harm from different aspects. Previous studies have analyzed memes in closed settings - detecting harm, applying semantic labels, and offering natural language explanations. To extend this research, we introduce MemeMQA, a multimodal question-answering framework aiming to solicit accurate responses to structured questions while providing coherent explanations. We curate MemeMQACorpus, a new dataset featuring 1,880 questions related to 1,122 memes with corresponding answer-explanation pairs. We further propose ARSENAL, a novel two-stage multimodal framework that leverages the reasoning capabilities of LLMs to address MemeMQA. We benchmark MemeMQA using competitive baselines and demonstrate its superiority - ~18% enhanced answer prediction accuracy and distinct text generation lead across various metrics measuring lexical and semantic alignment over the best baseline. We analyze ARSENAL’s robustness through diversification of question-set, confounder-based evaluation regarding MemeMQA’s generalizability, and modality-specific assessment, enhancing our understanding of meme interpretation in the multimodal communication landscape.