EMNLP 2024 论文列表

Recently, leveraging reinforcement learning (RL) to fine-tune language models (LMs), known as reinforcement learning from human feedback (RLHF), has become an important research topic. However, there is still a lack of theoretical understanding of how RLHF works, the conditions under which it succeeds or fails, and whether it guarantees optimization of both likelihood 𝛽(⋅) and reward R(⋅) objectives. To address these issues, we consider RLHF as a bi-objective problem that has the nature of a Pareto optimization, present a Pareto improvement condition that is necessary to obtain Pareto-efficient policies, and propose a simple yet powerful method named reward dropout that guarantees a Pareto improvement. To demonstrate the performance of reward dropout, two benchmark datasets commonly used in text style transfer tasks were utilized in our study: sentiment and topic datasets sourced from Yelp and AG_News, respectively. Our experiments highlight that paying attention to a few samples with higher rewards leads to greater Pareto improvements regardless of model size. We also demonstrate that the effect of reward dropout is generalizable and most effective with non-pretrained target models, saving the effort of pretraining.

Research on continuous sign language recognition (CSLR) is essential to bridge the communication gap between deaf and hearing individuals. Numerous previous studies have trained their models using the connectionist temporal classification (CTC) loss. During inference, these CTC-based models generally require the entire sign video as input to make predictions, a process known as offline recognition, which suffers from high latency and substantial memory usage. In this work, we take the first step towards online CSLR. Our approach consists of three phases: 1) developing a sign dictionary; 2) training an isolated sign language recognition model on the dictionary; and 3) employing a sliding window approach on the input sign sequence, feeding each sign clip to the optimized model for online recognition. Additionally, our online recognition model can be extended to support online translation by integrating a gloss-to-text network and can enhance the performance of any offline model. With these extensions, our online approach achieves new state-of-the-art performance on three popular benchmarks across various task settings. Code and models are available at https://github.com/FangyunWei/SLRT.

Most existing Visual Question Answering (VQA) systems are constrained to support domain-specific questions, i.e., to train different models separately for different VQA tasks, thus generalizing poorly to others. For example, models trained on the reasoning-focused dataset GQA struggle to effectively handle samples from the knowledge-emphasizing dataset OKVQA. Meanwhile, in real-world scenarios, it is user-unfriendly to restrict the domain of questions. Therefore, this paper proposes a necessary task: One-to-Many Visual Question Answering, of which the ultimate goal is to enable a single model to answer as many different domains of questions as possible by the effective integration of available VQA resources. To this end, we first investigate into ten common VQA datasets, and break the task of VQA down into the integration of three key abilities.Then, considering assorted questions rely on different VQA abilities, this paper proposes a novel dynamic Mixture of LoRAs (MoL) strategy. MoL mixes three individually trained LoRA adapters (corresponding to each VQA ability) dynamically for different samples demanding various VQA abilities. The proposed MoL strategy is verified to be highly effective by experiments, establishing SOTAs on four datasets. In addition, MoL generalizes well to three extra zero-shot datasets.Data and codes will be released.

The evaluation of natural language processing (NLP) systems is crucial for advancing the field, but current benchmarking approaches often assume that all systems have scores available for all tasks, which is not always practical. In reality, several factors such as the cost of running baseline, private systems, computational limitations, or incomplete data may prevent some systems from being evaluated on entire tasks. This paper formalize an existing problem in NLP research: benchmarking when some systems scores are missing on the task, and proposes a novel approach to address it. Our method utilizes a compatible partial ranking approach to impute missing data, which is then aggregated using the Borda count method. It includes two refinements designed specifically for scenarios where either task-level or instance-level scores are available. We also introduce an extended benchmark, which contains over 131 million scores, an order of magnitude larger than existing benchmarks. We validate our methods and demonstrate their effectiveness in addressing the challenge of missing system evaluation on an entire task. This work highlights the need for more comprehensive benchmarking approaches that can handle real-world scenarios where not all systems are evaluated on the entire task.

We present a survey of more than 90 recent papers that aim to study cultural representation and inclusion in large language models (LLMs). We observe that none of the studies explicitly define “culture, which is a complex, multifaceted concept; instead, they probe the models on some specially designed datasets which represent certain aspects of “culture”. We call these aspects the proxies of culture, and organize them across two dimensions of demographic and semantic proxies. We also categorize the probing methods employed. Our analysis indicates that only certain aspects of “culture,” such as values and objectives, have been studied, leaving several other interesting and important facets, especially the multitude of semantic domains (Thompson et al., 2020) and aboutness (Hershcovich et al., 2022), unexplored. Two other crucial gaps are the lack of robustness of probing techniques and situated studies on the impact of cultural mis- and under-representation in LLM-based applications.

The proliferation of Internet memes in the age of social media necessitates effective identification of harmful ones. Due to the dynamic nature of memes, existing data-driven models may struggle in low-resource scenarios where only a few labeled examples are available. In this paper, we propose an agency-driven framework for low-resource harmful meme detection, employing both outward and inward analysis with few-shot annotated samples. Inspired by the powerful capacity of Large Multimodal Models (LMMs) on multimodal reasoning, we first retrieve relative memes with annotations to leverage label information as auxiliary signals for the LMM agent. Then, we elicit knowledge-revising behavior within the LMM agent to derive well-generalized insights into meme harmfulness. By combining these strategies, our approach enables dialectical reasoning over intricate and implicit harm-indicative patterns. Extensive experiments conducted on three meme datasets demonstrate that our proposed approach achieves superior performance than state-of-the-art methods on the low-resource harmful meme detection task.

While transformer models exhibit strong capabilities on linguistic tasks, their complex architectures make them difficult to interpret. Recent work has aimed to reverse engineer transformer models into human-readable representations called circuits that implement algorithmic functions. We extend this research by analyzing and comparing circuits for similar sequence continuation tasks, which include increasing sequences of Arabic numerals, number words, and months. By applying circuit interpretability analysis, we identify a key sub-circuit in both GPT-2 Small and Llama-2-7B responsible for detecting sequence members and for predicting the next member in a sequence. Our analysis reveals that semantically related sequences rely on shared circuit subgraphs with analogous roles. Additionally, we show that this sub-circuit has effects on various math-related prompts, such as on intervaled circuits, Spanish number word and months continuation, and natural language word problems. Overall, documenting shared computational structures enables better model behavior predictions, identification of errors, and safer editing procedures. This mechanistic understanding of transformers is a critical step towards building more robust, aligned, and interpretable language models.

The rapid development of multimodal large language models (MLLMs), such as GPT-4V, has led to significant advancements. However, these models still face challenges in medical multimodal capabilities due to limitations in the quantity and quality of medical vision-text data, stemming from data privacy concerns and high annotation costs. While pioneering approaches utilize PubMed’s large-scale, de-identified medical image-text pairs to address these limitations, they often fall short due to inherent data noise. To tackle this, we refined medical image-text pairs from PubMed and employed MLLMs (GPT-4V) in an ‘unblinded’ capacity to denoise and reformat the data, resulting in the creation of the **PubMedVision** dataset with 1.3 million medical VQA samples. Our validation demonstrates that: (1) PubMedVision can significantly enhance the medical multimodal capabilities of MLLMs, showing significant improvement in benchmarks including the MMMU Health & Medicine track; (2) manual checks by medical experts and empirical results validate the superior data quality of our dataset compared to other data construction methods. Using PubMedVision, we train a 34B medical MLLM **HuatuoGPT-Vision**, which shows superior performance in medical multimodal scenarios among open-source MLLMs. Our code and data are available at https://github.com/FreedomIntelligence/HuatuoGPT-Vision.

As Large Language Models (LLMs) continue to evolve, they are increasingly being employed in numerous studies to simulate societies and execute diverse social tasks. However, LLMs are susceptible to societal biases due to their exposure to human-generated data. Given that LLMs are being used to gain insights into various societal aspects, it is essential to mitigate these biases. To that end, our study investigates the presence of implicit gender biases in multi-agent LLM interactions and proposes two strategies to mitigate these biases. We begin by creating a dataset of scenarios where implicit gender biases might arise, and subsequently develop a metric to assess the presence of biases. Our empirical analysis reveals that LLMs generate outputs characterized by strong implicit bias associations (≥ ≈ 50% of the time). Furthermore, these biases tend to escalate following multi-agent interactions. To mitigate them, we propose two strategies: self-reflection with in-context examples (ICE); and supervised fine-tuning. Our research demonstrates that both methods effectively mitigate implicit biases, with the ensemble of fine-tuning and self-reflection proving to be the most successful.

Large language models (LLMs) have revolutionized natural language processing and broadened their applicability across diverse commercial applications. However, the deployment of these models is constrained by high inference time in multilingual settings. To mitigate this challenge, this paper explores a training recipe of an assistant model in speculative decoding, which are leveraged to draft and-then its future tokens are verified by the target LLM. We show that language-specific draft models, optimized through a targeted pretrain-and-finetune strategy, substantially brings a speedup of inference time compared to the previous methods. We validate these models across various languages in inference time, out-of-domain speedup, and GPT-4o evaluation.

The fusion of language models (LMs) and knowledge graphs (KGs) is widely used in commonsense question answering, but generating faithful explanations remains challenging. Current methods often overlook path decoding faithfulness, leading to divergence between graph encoder outputs and model predictions. We identify confounding effects and LM-KG misalignment as key factors causing spurious explanations. To address this, we introduce the LM-KG Fidelity metric to assess KG representation reliability and propose the LM-KG Distribution-aware Alignment (LKDA) algorithm to improve explanation faithfulness. Without ground truth, we evaluate KG explanations using the proposed Fidelity-Sparsity Trade-off Curve. Experiments on CommonsenseQA and OpenBookQA show that LKDA significantly enhances explanation fidelity and model performance, highlighting the need to address distributional misalignment for reliable commonsense reasoning.

As intelligent education evolves, it will provide students with multiple personalized learning services based on their individual abilities. Computerized adaptive testing (CAT) is designed to accurately measure a student’s ability using the least questions, providing an efficient and personalized testing method. However, existing methods mainly focus on minimizing the number of questions required to assess ability, often lacking clear and reliable explanations for the question selection process. Educators and students can hardly trust and accept CAT systems without an understanding of the rationale behind the question selection process. To address this issue, we introduce LLM-Agent-Based CAT (LACAT), a novel agent powered by large language models to enhance CAT with human-like interpretability and explanation capabilities. LACAT consists of three key modules: the Summarizer, which generates interpretable student profiles; the Reasoner, which personalizes questions and provides human-readable explanations; and the Critic, which learns from past choices to optimize future question selection. We conducted extensive experiments on three real-world educational datasets. The results demonstrate that LACAT can perform comparably or superior to traditional CAT methods in accuracy and significantly improve the transparency and acceptability of the testing process. Human evaluations further confirm that LACAT can generate high-quality, understandable explanations, thereby enhancing student trust and satisfaction.

Grammatical Error Correction (GEC) is a crucial technique in Automated Essay Scoring (AES) for evaluating the fluency of essays. However, in Chinese, existing GEC datasets often fail to consider the importance of specific grammatical error types within compositional scenarios, lack research on data collected from native Chinese speakers, and largely overlook cross-sentence grammatical errors. Furthermore, the measurement of the overall fluency of an essay is often overlooked. To address these issues, we present CEFA (Chinese Essay Fluency Assessment), an extensive corpus that is derived from essays authored by native Chinese-speaking primary and secondary students and encapsulates essay fluency scores along with both coarse and fine-grained grammatical error types and corrections. Experiments employing various benchmark models on CEFA substantiate the challenge of our dataset. Our findings further highlight the significance of fine-grained annotations in fluency assessment and the mutually beneficial relationship between error types and corrections

Extractive summarization plays a pivotal role in natural language processing due to its wide-range applications in summarizing diverse content efficiently, while also being faithful to the original content. Despite significant advancement achieved in extractive summarization by Large Language Models (LLMs), these summaries frequently exhibit incoherence. An important aspect of the coherent summary is its readability for intended users. Although there have been many datasets and benchmarks proposed for creating coherent extractive summaries, none of them currently incorporate user intent to improve coherence in extractive summarization. Motivated by this, we propose a systematically created human-annotated dataset consisting of coherent summaries for five publicly available datasets and natural language user feedback, offering valuable insights into how to improve coherence in extractive summaries. We utilize this dataset for aligning LLMs through supervised fine-tuning with natural language human feedback to enhance the coherence of their generated summaries. Preliminary experiments with Falcon-40B and Llama-2-13B show significant performance improvements (~10% Rouge-L) in terms of producing coherent summaries. We further utilize human feedback to benchmark results over instruction-tuned models such as FLAN-T5 which resulted in several interesting findings.

Recent advancements in large language models have demonstrated enhanced capabilities in visual reasoning tasks by employing additional encoders for aligning different modalities. While the Q-Former has been widely used as a general encoder for aligning several modalities including image, video, audio, and 3D with large language models, previous works on its efficient training and the analysis of its individual components have been limited. In this work, we investigate the effectiveness of parameter efficient fine-tuning (PEFT) the Q-Former using InstructBLIP with visual reasoning benchmarks ScienceQA and IconQA. We observe that applying PEFT to the Q-Former achieves comparable performance to full fine-tuning using under 2% of the trainable parameters. Additionally, we employ AdaLoRA for dynamic parameter budget reallocation to examine the relative importance of the Q-Former’s sublayers with 4 different benchmarks. Our findings reveal that the self-attention layers are noticeably more important in perceptual visual-language reasoning tasks, and relative importance of FFN layers depends on the complexity of visual-language patterns involved in tasks. The code is available at https://github.com/AttentionX/InstructBLIP_PEFT.

Trolling in online communities typically involves disruptive behaviors such as provoking anger and manipulating discussions, leading to a polarized atmosphere and emotional distress. Robust moderation is essential for mitigating these negative impacts and maintaining a healthy and constructive community atmosphere. However, effectively addressing trolls is difficult because their behaviors vary widely and require different response strategies (RSs) to counter them. This diversity makes it challenging to choose an appropriate RS for each specific situation.To address this challenge, our research investigates whether humans have preferred strategies tailored to different types of trolling behaviors.Our findings reveal a correlation between the types of trolling encountered and the preferred RS. In this paper, we introduce a methodology for generating counter-responses to trolls by recommending appropriate RSs, supported by a dataset aligning these strategies with human preferences across various troll contexts. The experimental results demonstrate that our proposed approach guides constructive discussion and reduces the negative effects of trolls, thereby enhancing the online community environment.

Vision-language models (VLMs) like CLIP have demonstrated remarkable applicability across a variety of downstream tasks, including zero-shot image classification. Recently, the use of prompts or adapters for efficient transfer learning (ETL) has gained significant attention for effectively adapting to downstream tasks. However, previous studies have overlooked the challenge of varying transfer difficulty of downstream tasks. In this paper, we empirically analyze how each ETL method behaves with respect to transfer difficulty. Our observations indicate that utilizing vision prompts and text adapters is crucial for adaptability and generalizability in domains with high difficulty. Also, by applying an adaptive ensemble approach that integrates task-adapted VLMs with pre-trained VLMs and strategically leverages more general knowledge in low-difficulty and less in high-difficulty domains, we consistently enhance performance across both types of domains. Based on these observations, we propose an adaptive ensemble method that combines visual prompts and text adapters with pre-trained VLMs, tailored by transfer difficulty, to achieve optimal performance for any target domain. Upon experimenting with extensive benchmarks, our method consistently outperforms all baselines, particularly on unseen tasks, demonstrating its effectiveness.

Translating text that contains entity names is a challenging task, as cultural-related references can vary significantly across languages. These variations may also be caused by transcreation, an adaptation process that entails more than transliteration and word-for-word translation. In this paper, we address the problem of cross-cultural translation on two fronts: (i) we introduce XC-Translate, the first large-scale, manually-created benchmark for machine translation that focuses on text that contains potentially culturally-nuanced entity names, and (ii) we propose KG-MT, a novel end-to-end method to integrate information from a multilingual knowledge graph into a neural machine translation model by leveraging a dense retrieval mechanism. Our experiments and analyses show that current machine translation systems and large language models still struggle to translate texts containing entity names, whereas KG-MT outperforms state-of-the-art approaches by a large margin, obtaining a 129% and 62% relative improvement compared to NLLB-200 and GPT-4, respectively.

Situational awareness refers to the capacity to perceive and comprehend the present context and anticipate forthcoming events, which plays a critical role in aiding decision-making, anticipating potential issues, and adapting to dynamic circumstances. Nevertheless, the situational awareness capabilities of large language models have not yet been comprehensively assessed. To address this, we propose SA-Bench, a comprehensive benchmark that covers three tiers of situational awareness capabilities, covering environment perception, situation comprehension and future projection. SA-Bench provides a comprehensive evaluation to explore the situational awareness capabilities of LLMs. We conduct extensive experiments on advanced LLMs, including GPT-4, LLaMA3, Qwen1.5, among others. Our experimental results indicate that even SOTA LLMs still exhibit substantial capability gaps compared to humans. In addition, we thoroughly analysis and examine the challenges encountered by LLMs across various tasks, as well as emphasize the deficiencies they confront. We hope SA-Bench will foster research within the field of situational awareness.