ACL 2024 论文列表

Large Language Models (LLMs) have shown their impressive capabilities, while also raising concerns about the data contamination problems due to privacy issues and leakage of benchmark datasets in the pre-training phase. Therefore, it is vital to detect the contamination by checking whether an LLM has been pre-trained on the target texts. Recent studies focus on the generated texts and compute perplexities, which are superficial features and not reliable. In this study, we propose to utilize the probing technique for pre-training data detection by examining the model’s internal activations. Our method is simple and effective and leads to more trustworthy pre-training data detection. Additionally, we propose ArxivMIA, a new challenging benchmark comprising arxiv abstracts from Computer Science and Mathematics categories. Our experiments demonstrate that our method outperforms all baselines, and achieves state-of-the-art performance on both WikiMIA and ArxivMIA, with additional experiments confirming its efficacy.

Lexical Substitution discovers appropriate substitutes for a given target word in a context sentence. However, the task fails to consider substitutes that are of equal or higher proficiency than the target, an aspect that could be beneficial for language learners looking to improve their writing. To bridge this gap, we propose a new task — language proficiency-oriented lexical substitution. We also introduce ProLex, a novel benchmark designed to assess systems’ ability to generate not only appropriate substitutes but also substitutes that demonstrate better language proficiency. Besides the benchmark, we propose models that can automatically perform the new task. We show that our best model, a Llama2-13B model fine-tuned with task-specific synthetic data, outperforms ChatGPT by an average of 3.2% in F-score and achieves comparable results with GPT-4 on ProLex.

Gender bias has been widely observed in NLP models, which has the potential to perpetuate harmful stereotypes and discrimination. In this paper, we construct a dataset GenderStance of 36k samples to measure gender bias in stance detection, determining whether models consistently predict the same stance for a particular gender group. We find that all models are gender-biased and prone to classify sentences that contain male nouns as Against and those with female nouns as Favor. Moreover, extensive experiments indicate that sources of gender bias stem from the fine-tuning data and the foundation model itself. We will publicly release our code and dataset.

The rapid development of language models (LMs) brings unprecedented accessibility and usage for both models and users. On the one hand, powerful LMs achieve state-of-the-art performance over numerous downstream NLP tasks. On the other hand, more and more attention is paid to unrestricted model accesses that may bring malicious privacy risks of data leakage. To address these issues, many recent works propose privacy-preserving language models (PPLMs) with differential privacy (DP). Unfortunately, different DP implementations make it challenging for a fair comparison among existing PPLMs. In this paper, we present PrivLM-Bench, a multi-perspective privacy evaluation benchmark to empirically and intuitively quantify the privacy leakage of LMs. Instead of only reporting DP parameters, PrivLM-Bench sheds light on the neglected inference data privacy during actual usage. PrivLM-Bench first clearly defines multi-faceted privacy objectives. Then, PrivLM-Bench constructs a unified pipeline to perform private fine-tuning. Lastly, PrivLM-Bench performs existing privacy attacks on LMs with pre-defined privacy objectives as the empirical evaluation results. The empirical attack results are used to fairly and intuitively evaluate the privacy leakage of various PPLMs. We conduct extensive experiments on three datasets of GLUE for mainstream LMs.

Recent studies show that large language models (LLMs) improve their performance through self-feedback on certain tasks while degrade on others. We discovered that such a contrary is due to LLM’s bias in evaluating their own output. In this paper, we formally define LLM’s self-bias – the tendency to favor its own generation – using two statistics. We analyze six LLMs (GPT-4, GPT-3.5, Gemini, LLaMA2, Mixtral and DeepSeek) on translation, constrained text generation, and mathematical reasoning tasks. We find that self-bias is prevalent in all examined LLMs across multiple languages and tasks. Our analysis reveals that while the self-refine pipeline improves the fluency and understandability of model outputs, it further amplifies self-bias. To mitigate such biases, we discover that larger model size and external feedback with accurate assessment can significantly reduce bias in the self-refine pipeline, leading to actual performance improvement in downstream tasks. The code and data are released at https://github.com/xu1998hz/llm_self_bias.

While supervised fine-tuning (SFT) has been a straightforward approach for tailoring the output of foundation large language model (LLM) to specific preferences, concerns have been raised about the depth of this alignment, with some critiques suggesting it is merely “superficial”. We critically examine this hypothesis within the scope of cross-lingual generation tasks, proposing that the effectiveness of SFT may be constrained by its reliance on prior tokens to guide cross-lingual generation. Based on this crucial insight, and in response to the challenges posed by the costly and limited availability of non-English data for SFT, we introduce a novel training-free alignment method named PreTTY, which employs minimal task-related prior tokens to bridge the foundation LLM and the SFT LLM, achieving comparable performance without training. Experiments on machine translation and part-of-speech tagging across seven languages demonstrate the efficacy of PreTTY in cross-lingual settings. Remarkably, by initiating the decoding process with only one or two prior tokens, foundation LLMs can attain up to 98% of the performance metrics of their SFT counterparts. This method presents a cost-effective alternative to traditional SFT and advances the democratization of multilingual LLMs.

Large language models (LLMs) showcase impressive reasoning capabilities when coupled with Chain-of-Thought (CoT) prompting. However, the robustness of this approach warrants further investigation. In this paper, we introduce a novel scenario termed preemptive answers, where the LLM obtains an answer before engaging in reasoning. This situation can arise inadvertently or induced by malicious users by prompt injection attacks. Experiments reveal that preemptive answers significantly impair the model’s reasoning capability across various CoT methods and a broad spectrum of datasets. To bolster the robustness of reasoning, we propose two measures aimed at mitigating this issue to some extent.

Temporal Knowledge Graph (TKG) reasoning seeks to predict future incomplete facts leveraging historical data. While existing approaches have shown effectiveness in addressing the task through various perspectives, such as graph learning and logic rules, they are limited in capturing the indeterminacy in future events, particularly in the case of rare/unseen facts. To tackle the highlighted issues, we introduce a novel approach by conceptualizing TKG reasoning as a sequence denoising process for future facts, namely DiffuTKG. Concretely, we first encodes the historical events as the conditional sequence. Then we gradually introduce Gaussian noise to corrupt target facts during the forward process and then employ a transformer-based conditional denoiser to restore them in the reverse phase. Moreover, we introduce an uncertainty regularization loss to mitigate the risk of prediction biases by favoring frequent scenarios over rare/unseen facts. Empirical results on four real-world datasets show that DiffuTKG outperforms state-of-the-art methods across multiple evaluation metrics.

Comparative reasoning plays a crucial role in predicting text preferences; however, large language models (LLMs) often demonstrate inconsistencies in their reasoning, leading to incorrect preference predictions. While approaches like Chain-of-Thought improve accuracy in many settings, they struggle to consistently distinguish the similarities and differences of complex texts. We introduce SC2, a model that prompts LLMs to predict text preferences by generating structured intermediate comparisons. SC2 begins by proposing aspects for comparison, followed by generating textual comparisons under each aspect. We select consistent comparisons with a pairwise comparator that ensures each comparison of a given aspect clearly distinguishes differences between texts, significantly reducing hallucination and improving consistency. Our empirical studies across various NLP tasks, including summarization, retrieval, and automatic rating, demonstrate that SC2‘s enhanced performance in text preference prediction is significant.

Social media advertising offers a platform for fossil fuel value chain companies and their agents to reinforce their narratives, often emphasizing economic, labor market, and energy security benefits to promote oil and gas policy and products. Whether such narratives can be detected automatically and the extent to which the cost of human annotation can be reduced is our research question. We introduce a task of classifying narratives into seven categories, based on existing definitions and data.Experiments showed that RoBERTa-large outperforms other methods, while GPT-4 Turbo can serve as a viable annotator for the task, thereby reducing human annotation costs. Our findings and insights provide guidance to automate climate-related ad analysis and lead to more scalable ad scrutiny.

Large Multimodal Models (LMMs) excel in natural language and visual understanding but are challenged by exacting tasks such as Knowledge-based Visual Question Answering (KB-VQA) which involve the retrieval of relevant information from document collections to use in shaping answers to questions. We present an extensive training and evaluation framework, M2KR, for KB-VQA. M2KR contains a collection of vision and language tasks which we have incorporated into a single suite of benchmark tasks for training and evaluating general-purpose multi-modal retrievers. We use M2KR to develop PreFLMR, a pre-trained version of the recently developed Fine-grained Late-interaction Multi-modal Retriever (FLMR) approach to KB-VQA, and we report new state-of-the-art results across a range of tasks. We also present investigations into the scaling behaviors of PreFLMR intended to be useful in future developments in general-purpose multi-modal retrievers.

Empathy is a social mechanism used to support and strengthen emotional connection with others, including in online communities. However, little is currently known about the nature of these online expressions, nor the particular factors that may lead to their improved detection. In this work, we study the role of a specific and complex subcategory of linguistic phenomena, figurative language, in online expressions of empathy. Our extensive experiments reveal that incorporating features regarding the use of metaphor, idiom, and hyperbole into empathy detection models improves their performance, resulting in impressive maximum F1 scores of 0.942 and 0.809 for identifying posts without and with empathy, respectively.

The guidance from capability evaluations has greatly propelled the progress of human society and the development of Artificial Intelligence. However, as LLMs evolve, it becomes challenging to construct evaluation benchmark with accurate labels for SuperLLMs whose capabilities approach or even surpass those of humans. To credibly conduct poor-supervised evaluation without accurate labels, we first prove that the consistency between the model under evaluation and the reference model, when their prediction distributions are independent and the sample size is infinite, can equivalently assess the true capabilities of the model to be evaluated. However, using either humans or LLMs as the reference model cannot sufficiently meet the conditions, for which we propose the PEEM algorithm. By treating all models under evaluation as reference models, PEEM alternately optimizes model weights and filters reference models based on EM algorithm to maximally alleviate the insufficiency of the conditions. Comprehensive experiments across 3 types of tasks with 16 mainstream LLMs validate the efficiency, universality, and effectiveness of PEEM. More generally, PEEM has advanced the evaluation paradigm evolution from human-centric to human&model-centric, alleviating the limitations of human capabilities for evaluating SuperLLMs.

Much recent work seeks to evaluate values and opinions in large language models (LLMs) using multiple-choice surveys and questionnaires. Most of this work is motivated by concerns around real-world LLM applications. For example, politically-biased LLMs may subtly influence society when they are used by millions of people. Such real-world concerns, however, stand in stark contrast to the artificiality of current evaluations: real users do not typically ask LLMs survey questions. Motivated by this discrepancy, we challenge the prevailing *constrained* evaluation paradigm for values and opinions in LLMs and explore more realistic *unconstrained* evaluations. As a case study, we focus on the popular Political Compass Test (PCT). In a systematic review, we find that most prior work using the PCT *forces models to comply with the PCT’s multiple-choice format. We show that models give substantively different answers when not forced; that answers change depending on how models are forced; and that answers lack paraphrase robustness. Then, we demonstrate that models give different answers yet again in a more realistic open-ended answer setting. We distill these findings into recommendations and open challenges in evaluating values and opinions in LLMs.

Word sense disambiguation (WSD) can be viewed as two subtasks: textual word sense disambiguation (Textual-WSD) and visual word sense disambiguation (Visual-WSD). They aim to identify the most semantically relevant senses or images to a given context containing ambiguous target words. However, existing WSD models seldom address these two subtasks jointly due to lack of images in Textual-WSD datasets or lack of senses in Visual-WSD datasets. To bridge this gap, we propose PolCLIP, a unified image-text WSD model. By employing an image-text complementarity strategy, it not only simulates stable diffusion models to generate implicit visual representations for word senses but also simulates image captioning models to provide implicit textual representations for images. Additionally, a disambiguation-oriented image-sense dataset is constructed for the training objective of learning multimodal polysemy representations. To the best of our knowledge, PolCLIP is the first model that can cope with both Textual-WSD and Visual-WSD. Extensive experimental results on benchmarks demonstrate the effectiveness of our method, achieving a 2.53% F1-score increase over the state-of-the-art models on Textual-WSD and a 2.22% HR@1 improvement on Visual-WSD.

Multi-hop question answering (MQA) is one of the challenging tasks to evaluate machine’s comprehension and reasoning abilities, where large language models (LLMs) have widely achieved the human-comparable performance. Due to the dynamics of knowledge facts in real world, knowledge editing has been explored to update model with the up-to-date facts while avoiding expensive re-training or fine-tuning. Starting from the edited fact, the updated model needs to provide cascading changes in the chain of MQA. The previous art simply adopts a mix-up prompt to instruct LLMs conducting multiple reasoning tasks sequentially, including question decomposition, answer generation, and conflict checking via comparing with edited facts. However, the coupling of these functionally-diverse reasoning tasks inhibits LLMs’ advantages in comprehending and answering questions while disturbing them with the unskilled task of conflict checking. We thus propose a framework, Programmable knowledge editing for Multi-hop Question Answering (PokeMQA), to decouple the jobs. Specifically, we prompt LLMs to decompose knowledge-augmented multi-hop question, while interacting with a detached trainable scope detector to modulate LLMs behavior depending on external conflict signal. The experiments on three LLM backbones and two benchmark datasets validate our superiority in knowledge editing of MQA, outperforming all competitors by a large margin in almost all settings and consistently producing reliable reasoning process.

Since the emergence of large language models, prompt learning has become a popular method for optimizing and customizing these models. Special prompts, such as Chain-of-Thought, have even revealed previously unknown reasoning capabilities within these models. However, the progress of discovering effective prompts has been slow, driving a desire for general prompt optimization methods. Unfortunately, few existing prompt learning methods satisfy the criteria of being truly “general”, i.e., automatic, discrete, black-box, gradient-free, and interpretable all at once. In this paper, we introduce metaheuristics, a branch of discrete non-convex optimization methods with over 100 options, as a promising approach to prompt learning. Within our paradigm, we test six typical methods: hill climbing, simulated annealing, genetic algorithms with/without crossover, tabu search, and harmony search, demonstrating their effectiveness in white-box and black-box prompt learning. Furthermore, we show that these methods can be used to discover more human-understandable prompts that were previously unknown in both reasoning and image generation tasks, opening the door to a cornucopia of possibilities in prompt optimization.

With the development of LLMs, the security threats of LLMs are getting more and more attention. Numerous jailbreak attacks have been proposed to assess the security defense of LLMs. Current jailbreak attacks primarily utilize scenario camouflage techniques. However their explicitly mention of malicious intent will be easily recognized and defended by LLMs. In this paper, we propose an indirect jailbreak attack approach, Puzzler, which can bypass the LLM’s defensive strategies and obtain malicious response by implicitly providing LLMs with some clues about the original malicious query. In addition, inspired by the wisdom of “When unable to attack, defend” from Sun Tzu’s Art of War, we adopt a defensive stance to gather clues about the original malicious query through LLMs. The experimental results indicate that the Query Success Rate of the Puzzler is 14.0%-82.7% higher than baselines on the most prominent LLMs. Furthermore, when tested against the state-of-the-art jailbreak detection approaches, Puzzler proves to be more effective at evading detection compared to baselines.

The increasing use of complex and opaque black box models requires the adoption of interpretable measures, one such option is extractive rationalizing models, which serve as a more interpretable alternative. These models, also known as Explain-Then-Predict models, employ an explainer model to extract rationales and subsequently condition the predictor with the extracted information. Their primary objective is to provide precise and faithful explanations, represented by the extracted rationales. In this paper, we take a semi-supervised approach to optimize for the plausibility of extracted rationales. We adopt a pre-trained natural language inference (NLI) model and further fine-tune it on a small set of supervised rationales (10%). The NLI predictor is leveraged as a source of supervisory signals to the explainer via entailment alignment. We show that, by enforcing the alignment agreement between the explanation and answer in a question-answering task, the performance can be improved without access to ground truth labels. We evaluate our approach on the ERASER dataset and show that our approach achieves comparable results with supervised extractive models and outperforms unsupervised approaches by > 100%.