EMNLP 2024 论文列表

While language model (LM)-powered chatbots and generative search engines excel at answering concrete queries, discovering information in the terrain of unknown unknowns remains challenging for users. To emulate the common educational scenario where children/students learn by listening to and participating in conversations of their parents/teachers, we create Collaborative STORM (Co-STORM). Unlike QA systems that require users to ask all the questions, Co-STORM lets users observe and occasionally steer the discourse among several LM agents. The agents ask questions on the user’s behalf, allowing the user to discover unknown unknowns serendipitously. To facilitate user interaction, Co-STORM assists users in tracking the discourse by organizing the uncovered information into a dynamic mind map, ultimately generating a comprehensive report as takeaways. For automatic evaluation, we construct the WildSeek dataset by collecting real information-seeking records with user goals. Co-STORM outperforms baseline methods on both discourse trace and report quality. In a further human evaluation, 70% of participants prefer Co-STORM over a search engine, and 78% favor it over a RAG chatbot.

Recently, pre-trained language models (PLMs) have been increasingly adopted in spoken language understanding (SLU). However, automatic speech recognition (ASR) systems frequently produce inaccurate transcriptions, leading to noisy inputs for SLU models, which can significantly degrade their performance. To address this, our objective is to train SLU models to withstand ASR errors by exposing them to noises commonly observed in ASR systems, referred to as ASR-plausible noises. Speech noise injection (SNI) methods have pursued this objective by introducing ASR-plausible noises, but we argue that these methods are inherently biased towards specific ASR systems, or ASR-specific noises. In this work, we propose a novel and less biased augmentation method of introducing the noises that are plausible to any ASR system, by cutting off the non-causal effect of noises. Experimental results and analyses demonstrate the effectiveness of our proposed methods in enhancing the robustness and generalizability of SLU models against unseen ASR systems by introducing more diverse and plausible ASR noises in advance.

Understanding the inner workings of large language models (LLMs) is crucial for advancing their theoretical foundations and real-world applications. While the attention mechanism and multi-layer perceptrons (MLPs) have been studied independently, their interactions remain largely unexplored. This study investigates how attention heads and next-token neurons interact in LLMs to predict new words. We propose a methodology to identify next-token neurons, find prompts that highly activate them, and determine the upstream attention heads responsible. We then generate and evaluate explanations for the activity of these attention heads in an automated manner. Our findings reveal that some attention heads recognize specific contexts relevant to predicting a token and activate a downstream token-predicting neuron accordingly. This mechanism provides a deeper understanding of how attention heads work with MLP neurons to perform next-token prediction. Our approach offers a foundation for further research into the intricate workings of LLMs and their impact on text generation and understanding.

We find arithmetic ability resides within a limited number of attention heads, with each head specializing in distinct operations. To delve into the reason, we introduce the Comparative Neuron Analysis (CNA) method, which identifies an internal logic chain consisting of four distinct stages from input to prediction: feature enhancing with shallow FFN neurons, feature transferring by shallow attention layers, feature predicting by arithmetic heads, and prediction enhancing among deep FFN neurons. Moreover, we identify the human-interpretable FFN neurons within both feature-enhancing and feature-predicting stages. These findings lead us to investigate the mechanism of LoRA, revealing that it enhances prediction probabilities by amplifying the coefficient scores of FFN neurons related to predictions. Finally, we apply our method in model pruning for arithmetic tasks and model editing for reducing gender bias. Code is on https://github.com/zepingyu0512/arithmetic-mechanism.

Reinforcement learning from human feedback (RLHF) has emerged as the primary method for aligning large language models (LLMs) with human preferences. The RLHF process typically starts by training a reward model (RM) using human preference data. Conventional RMs are trained on pairwise responses to the same user request, with relative ratings indicating which response humans prefer. The trained RM serves as a proxy for human preferences. However, due to the black-box nature of RMs, their outputs lack interpretability, as humans cannot intuitively understand why an RM thinks a response is good or not. As RMs act as human preference proxies, it is desirable for them to be human-interpretable to ensure that their internal decision processes are consistent with human preferences and to prevent reward hacking in LLM alignment. To build RMs with interpretable preferences, we propose a two-stage approach: i) train an Absolute-Rating Multi-Objective Reward Model (ArmoRM) with multi-dimensional absolute-rating data, each dimension corresponding to a human-interpretable objective (e.g., honesty, verbosity, safety); ii) employ a Mixture-of-Experts (MoE) strategy with a gating network that automatically selects the most suitable reward objectives based on the context. We efficiently trained an ArmoRM with Llama-3 8B and a gating network consisting of a shallow MLP on top of the ArmoRM. Our trained model, ArmoRM-Llama3-8B, obtains state-of-the-art performance on RewardBench, a benchmark evaluating RMs for language modeling. Notably, the performance of our model surpasses the LLM-as-a-judge method with GPT-4 judges by a margin, and approaches the performance of the much larger Nemotron-4 340B reward model.

Contrastively trained vision-language models such as CLIP have achieved remarkable progress in vision and language representation learning. Despite the promising progress, their proficiency in compositional reasoning over attributes and relations (e.g., distinguishing between “the car is underneath the person” and “the person is underneath the car”) remains notably inadequate. We investigate the cause for this deficient behavior is the composition attribution issue, where the attribution scores (e.g., attention scores or GradCAM scores) for relations (e.g., underneath) or attributes (e.g., red) in the text are substantially lower than those for object terms. In this work, we show such issue is mitigated via a novel framework called CAE (Composition Attribution Enhancement). This generic framework incorporates various interpretable attribution methods to encourage the model to pay greater attention to composition words denoting relationships and attributes within the text. Detailed analysis shows that our approach enables the models to adjust and rectify the attribution of the texts. Extensive experiments across seven benchmarks reveal that our framework significantly enhances the ability to discern intricate details and construct more sophisticated interpretations of combined visual and linguistic elements.

Language models recognized as a new form of knowledge bases, face challenges of outdated, erroneous, and privacy-sensitive information, necessitating knowledge editing to rectify errors without costly retraining. Existing methods, spanning model’s parameters modification, external knowledge integration, and in-context learning, lack in-depth analysis from a model interpretability perspective. Our work explores the instability in in-context learning outcomes, providing insights into its reasons and distinctions from other methods. Leveraging findings on the critical role of feed-forward MLPs in decoder-only models, we propose a tailored knowledge editing method, TailoredKE, that considers the unique information flow of each sample. Model interpretability reveals diverse attribute recall across transformer layers, guiding edits to specific features at different depths and mitigating over-editing issues.

Despite their vast capabilities, Large Language Models (LLMs) often struggle with generating reliable outputs, frequently producing high-confidence inaccuracies known as hallucinations. Addressing this challenge, our research introduces InternalInspector, a novel framework designed to enhance confidence estimation in LLMs by leveraging contrastive learning on internal states including attention states, feed-forward states, and activation states of all layers. Unlike existing methods that primarily focus on the final activation state, InternalInspector conducts a comprehensive analysis across all internal states of every layer to accurately identify both correct and incorrect prediction processes. By benchmarking InternalInspector against existing confidence estimation methods across various natural language understanding and generation tasks, including factual question answering, commonsense reasoning, and reading comprehension, InternalInspector achieves significantly higher accuracy in aligning the estimated confidence scores with the correctness of the LLM’s predictions and lower calibration error. Furthermore, InternalInspector excels at HaluEval, a hallucination detection benchmark, outperforming other internal-based confidence estimation methods in this task.

Large language models (LLMs) have demonstrated the potential to mimic human social intelligence. However, most studies focus on simplistic and static self-report or performance-based tests, which limits the depth and validity of the analysis. In this paper, we developed a novel framework, InterIntent, to assess LLMs’ social intelligence by mapping their ability to understand and manage intentions in a game setting. We focus on four dimensions of social intelligence: situational awareness, self-regulation, self-awareness, and theory of mind. Each dimension is linked to a specific game task: intention selection, intention following, intention summarization, and intention guessing. Our findings indicate that while LLMs exhibit high proficiency in selecting intentions, achieving an accuracy of 88%, their ability to infer the intentions of others is significantly weaker, trailing human performance by 20%. Additionally, game performance correlates with intention understanding, highlighting the importance of the four components towards success in this game. These findings underline the crucial role of intention understanding in evaluating LLMs’ social intelligence and highlight the potential of using social deduction games as a complex testbed to enhance LLM evaluation. InterIntent contributes a structured approach to bridging the evaluation gap in social intelligence within multiplayer LLM-based games.

Enhancing Language Models’ (LMs) ability to understand purchase intentions in E-commerce scenarios is crucial for their effective assistance in various downstream tasks. However, previous approaches that distill intentions from LMs often fail to generate meaningful and human-centric intentions applicable in real-world E-commerce contexts. This raises concerns about the true comprehension and utilization of purchase intentions by LMs. In this paper, we present IntentionQA, a double-task multiple-choice question answering benchmark to evaluate LMs’ comprehension of purchase intentions in E-commerce. Specifically, LMs are tasked to infer intentions based on purchased products and utilize them to predict additional purchases. IntentionQA consists of 4,360 carefully curated problems across three difficulty levels, constructed using an automated pipeline to ensure scalability on large E-commerce platforms. Human evaluations demonstrate the high quality and low false-negative rate of our benchmark. Extensive experiments across 19 language models show that they still struggle with certain scenarios, such as understanding products and intentions accurately, jointly reasoning with products and intentions, and more, in which they fall far behind human performances.

Intent detection is a critical component of task-oriented dialogue systems (TODS) which enables the identification of suitable actions to address user utterances at each dialog turn. Traditional approaches relied on computationally efficient supervised sentence transformer encoder models, which require substantial training data and struggle with out-of-scope (OOS) detection. The emergence of generative large language models (LLMs) with intrinsic world knowledge presents new opportunities to address these challenges.In this work, we adapt SOTA LLMs using adaptive in-context learning and chain-of-thought prompting for intent detection, and compare their performance with contrastively fine-tuned sentence transformer (SetFit) models to highlight prediction quality and latency tradeoff. We propose a hybrid system using uncertainty based routing strategy to combine the two approaches that along with negative data augmentation results in achieving the best of both worlds ( i.e. within 2% of native LLM accuracy with 50% less latency). To better understand LLM OOS detection capabilities, we perform controlled experiments revealing that this capability is significantly influenced by the scope of intent labels and the size of the label space. We also introduce a two-step approach utilizing internal LLM representations, demonstrating empirical gains in OOS detection accuracy and F1-score by >5% for the Mistral-7B model.

In this study, we investigate the potential of language models (LMs) in aiding patients experiencing anomia, a difficulty identifying the names of items. Identifying the intended target item from patient’s circumlocution involves the two challenges of term failure and error. (1) The terms relevant to identifying the item remain unseen. (2) What makes the challenge unique is inherent perturbed terms by semantic paraphasia, which are not exactly related to the target item, hindering the identification process. To address each, we propose robustifying the model from semantically paraphasic errors and enhancing the model with unseen terms with gradient-based selective augmentation (GradSelect). Specifically, the gradient value controls augmented data quality amid semantic errors, while the gradient variance guides the inclusion of unseen but relevant terms. Due to limited domain-specific datasets, we evaluate the model on the Tip of the Tongue dataset as an intermediary task and then apply our findings to real patient data from AphasiaBank. Our results demonstrate strong performance against baselines, aiding anomia patients by addressing the outlined challenges.

Information Extraction (IE), aiming to extract structured information from unstructured natural language texts, can significantly benefit from pre-trained language models. However, existing pre-training methods solely focus on exploiting the textual knowledge, relying extensively on annotated large-scale datasets, which is labor-intensive and thus limits the scalability and versatility of the resulting models. To address these issues, we propose SKIE, a novel pre-training framework tailored for IE that integrates structural semantic knowledge via contrastive learning, effectively alleviating the annotation burden. Specifically, SKIE utilizes Abstract Meaning Representation (AMR) as a low-cost supervision source to boost model performance without human intervention. By enhancing the topology of AMR graphs, SKIE derives high-quality cohesive subgraphs as additional training samples, providing diverse multi-level structural semantic knowledge. Furthermore, SKIE refines the graph encoder to better capture cohesive information and edge relation information, thereby improving the pre-training efficacy. Extensive experimental results demonstrate that SKIE outperforms state-of-the-art baselines across multiple IE tasks and showcases exceptional performance in few-shot and zero-shot settings.

Emotion significantly influences human behavior and decision-making processes. We propose a labeling methodology grounded in Plutchik’s Wheel of Emotions theory for emotion classification. Furthermore, we employ a Mixture of Experts (MoE) architecture to evaluate the efficacy of this labeling approach, by identifying the specific emotions that each expert learns to classify. Experimental results reveal that our methodology improves the performance of emotion classification.

Annotation tools are increasingly only steps in a larger process into which they need to be integrated, for instance by calling out to web services for labeling support or importing documents from external sources. This requires certain capabilities that annotation tools need to support in order to keep up. Here, we define the respective requirements and how popular annotation tools support them. As a demonstration for how these can be implemented, we adapted INCEpTION, a semantic annotation platform offering intelligent assistance and knowledge management. For instance, support for a range of APIs has been added to INCEpTION through which it can be controlled and which allow it to interact with external services such as authorization services, crowdsourcing platforms, terminology services or machine learning services. Additionally, we introduce new capabilities that allow custom rendering of XML documents and even the ability to add new JavaScript-based editor plugins, thereby making INCEpTION usable in an even wider range of annotation tasks.

The proliferation of online misinformation presents a significant challenge, requiring scalable strategies for effective mitigation. While detection methods exist, current reactive approaches, like content flagging and banning, are short-term and insufficient. Additionally, advancements like large language models (LLMs) exacerbate the issue by enabling large-scale creation and dissemination of misinformation. Thus, sustainable, scalable solutions that encourage behavior change and broaden perspectives by persuading misinformants against their viewpoints or broadening their perspectives are needed. To this end, we propose persuasive LLM-based dialogue systems to tackle misinformation. However, challenges arise due to the lack of suitable datasets and formal frameworks for generating persuasive responses. Inspired by existing methods for countering online hate speech, we explore adapting counter-hate response strategies for misinformation. Since misinformation and hate speech often coexist despite differing intentions, we develop classifiers to identify and annotate response strategies from hate-speech counter-responses for use in misinformation scenarios. Human evaluations show a 91% agreement on the applicability of these strategies to misinformation. Next, as a scalable counter-misinformation solution, we create an LLM-based argument graph framework that generates persuasive responses, using the strategies as control codes to adjust the style and content. Human evaluations and case studies demonstrate that our framework generates expert-like responses and is 14% more engaging, 21% more natural, and 18% more factual than the best available alternatives.

Image-text contrastive models such as CLIP learn transferable and robust representations for zero-shot transfer to a variety of downstream tasks. However, to obtain strong downstream performances, prompts need to be carefully curated, which can be a tedious engineering task. To address the issue of manual prompt engineering, prompt-tuning is used where a set of contextual vectors are learned by leveraging information from the training data. Despite their effectiveness, existing prompt-tuning frameworks often lack interpretability, thus limiting their ability to understand the compositional nature of images. In this work, we first identify that incorporating compositional attributes (e.g., a “green” tree frog) in the design of manual prompts can significantly enhance image-text alignment scores. Building upon this observation, we propose a novel and interpretable prompt-tuning method named IntCoOp, which learns to jointly align attribute-level inductive biases and class embeddings during prompt-tuning. To assess the effectiveness of our approach, we evaluate IntCoOp across two representative tasks in a few-shot learning setup: generalization to novel classes, and unseen domain shifts. Through extensive experiments across 10 downstream datasets on CLIP, we find that introducing attribute-level inductive biases leads to superior performance against state-of-art prompt tuning frameworks. Notably, in a 16-shot setup, IntCoOp improves CoOp by 7.35% in average performance across 10 diverse datasets.

The Instruction-Driven Game Engine (IDGE) project aims to democratize game development by enabling a large language model (LLM) to follow free-form game descriptions and generate game-play processes. The IDGE allows users to create games simply by natural language instructions, which significantly lowers the barrier for game development. We approach the learning process for IDGEs as a Next State Prediction task, wherein the model autoregressively predicts the game states given player actions. The computation of game states must be precise; otherwise, slight errors could corrupt the game-play experience. This is challenging because of the gap between stability and diversity. To address this, we train the IDGE in a curriculum manner that progressively increases its exposure to complex scenarios.Our initial progress lies in developing an IDGE for Poker, which not only supports a wide range of poker variants but also allows for highly individualized new poker games through natural language inputs. This work lays the groundwork for future advancements in transforming how games are created and played.

Unsupervised multitask pre-training has been the critical method behind the recent success of language models (LMs). However, supervised multitask learning still holds significant promise, as scaling it in the post-training stage trends towards better generalization. In this paper, we explore supervised multitask pre-training by proposing Instruction Pre-training, a framework that scalably augments massive raw corpora with instruction-response pairs to pre-train LMs. The instruction-response pairs are generated by an efficient instruction synthesizer built on open-source models. In our experiments, we synthesize 200M instruction response pairs covering 40+ task categories to verify the effectiveness of Instruction Pre-training. In pre-training from scratch, Instruction Pre-training not only consistently enhances pre-trained base models but also benefits more from further instruction tuning. In continual pre-training, Instruction Pre-training enables Llama3-8B to be comparable to or even outperform Llama3-70B. Our model, code, and data are available at https://github.com/microsoft/LMOps.