EMNLP 2024 Papers

Inspired by the legal doctrine of stare decisis, which leverages precedents (prior cases) for informed decision-making, we explore methods to integrate them into LJP models. To facilitate precedent retrieval, we train a retriever with a fine-grained relevance signal based on the overlap ratio of alleged articles between cases. We investigate two strategies to integrate precedents: direct incorporation at inference via label interpolation based on case proximity and during training via a precedent fusion module using a stacked-cross attention model. We employ joint training of the retriever and LJP models to address latent space divergence between them. Our experiments on LJP tasks from the ECHR jurisdiction reveal that integrating precedents during training coupled with joint training of the retriever and LJP model, outperforms models without precedents or with precedents incorporated only at inference, particularly benefiting sparser articles.

Although existing fashionable generation methods on Incomplete Utterance Rewriting (IUR) can generate coherent utterances, they often result in the inclusion of irrelevant and redundant tokens in rewritten utterances due to their inability to focus on critical tokens in dialogue context. Furthermore, the limited size of the training datasets also contributes to the insufficient training of the IUR model. To address the first issue, we propose a multi-task learning framework EO-IUR (Editing Operation-guided Incomplete Utterance Rewriting) that introduces the editing operation labels generated by sequence labeling module to guide generation model to focus on critical tokens. Furthermore, we introduce a token-level heterogeneous graph to represent dialogues. To address the second issue, we propose a two-dimensional utterance augmentation strategy, namely editing operation-based incomplete utterance augmentation and LLM-based historical utterance augmentation. The experimental results on three datasets demonstrate that our EO-IUR outperforms previous state-of-the-art (SOTA) baselines in both open-domain and task-oriented dialogue.

Despite advancements, fine-tuning Large Language Models (LLMs) remains costly due to the extensive parameter count and substantial data requirements for model generalization. Accessibility to computing resources remains a barrier for the open-source community. To address this challenge, we propose the In2Core algorithm, which selects a coreset by analyzing the correlation between training and evaluation samples with a trained model. Notably, we assess the model’s internal gradients to estimate this relationship, aiming to rank the contribution of each training point. To enhance efficiency, we propose an optimization to compute influence functions with a reduced number of layers while achieving similar accuracy. By applying our algorithm to instruction fine-tuning data of LLMs, we can achieve similar performance with just 50% of the training data. Meantime, using influence functions to analyze model coverage to certain testing samples could provide a reliable and interpretable signal on the training set’s coverage of those test points.

Event Causality Identification (ECI) aims at determining the existence of a causal relation between two events. Although recent prompt learning-based approaches have shown promising improvements on the ECI task, their performance are often subject to the delicate design of multiple prompts and the positive correlations between the main task and derivate tasks. The in-context learning paradigm provides explicit guidance for label prediction in the prompt learning paradigm, alleviating its reliance on complex prompts and derivative tasks. However, it does not distinguish between positive and negative demonstrations for analogy learning. Motivated from such considerations, this paper proposes an **I**n-**C**ontext **C**ontrastive **L**earning (ICCL) model that utilizes contrastive learning to enhance the effectiveness of both positive and negative demonstrations. Additionally, we apply contrastive learning to event pairs to better facilitate event causality identification. Our ICCL is evaluated on the widely used corpora, including the EventStoryLine and Causal-TimeBank, and results show significant performance improvements over the state-of-the-art algorithms.

Spurred by advancements in scale, large language models (LLMs) have demonstrated strong few-shot learning ability via in-context learning (ICL). However, the performance of ICL has been shown to be highly sensitive to the selection of few-shot demonstrations. Selecting the most suitable examples as context remains an ongoing challenge and an open problem. Existing literature has highlighted the importance of selecting examples that are diverse or semantically similar to the test sample while ignoring the fact that the optimal selection dimension, i.e., diversity or similarity, is task-specific. Based on how the test sample is answered, we propose Iterative Demonstration Selection (IDS) to leverage the merits of both dimensions. Using zero-shot chain-of-thought reasoning (Zero-shot-CoT), IDS iteratively selects examples that are diverse but still strongly correlated with the test sample as ICL demonstrations. Specifically, IDS applies Zero-shot-CoT to the test sample before demonstration selection. The output reasoning path is then used to choose demonstrations that are prepended to the test sample for inference. The generated answer is followed by its corresponding reasoning path for extracting a new set of demonstrations in the next iteration. After several iterations, IDS adopts majority voting to obtain the final result. Through extensive experiments on tasks including reasoning, question answering, and topic classification, we demonstrate that IDS can consistently outperform existing ICL demonstration selection methods.

Recent advancements in artificial intelligence have led to the creation of highly capable large language models (LLMs) that can perform tasks in a human-like manner. However, LLMs exhibit only infant-level cognitive abilities in certain areas. One such area is the A-Not-B error, a phenomenon seen in infants where they repeat a previously rewarded behavior despite well-observed changed conditions. This highlights their lack of inhibitory control – the ability to stop a habitual or impulsive response. In our work, we design a text-based multi-choice QA scenario similar to the A-Not-B experimental settings to systematically test the inhibitory control abilities of LLMs. We found that state-of-the-art LLMs (like Llama3-8b) perform consistently well with in-context learning (ICL) but make errors and show a significant drop of as many as 83.3% in reasoning tasks when the context changes trivially. This suggests that LLMs only have inhibitory control abilities on par with human infants in this regard, often failing to suppress the previously established response pattern during ICL.

With the rising popularity of Transformer-based large language models (LLMs), reducing their high inference costs has become a significant research focus. One effective approach to mitigate these costs is compressing the long input contexts. Existing methods typically leverage the self-attention mechanism of the large model itself for context compression. While these methods have achieved notable results, the compression process still entails quadratic complexity. To mitigate this limitation, we propose the In-Context Former (IC-Former). This method does not rely on the target large model but instead utilizes cross-attention mechanisms to extract and condense information from the contextual embeddings. The computational overhead of our method grows linearly with the compression range. Experimental results indicate that our method requires only 1/32 of the floating-point operations of the baseline during compression and improves processing speed by 68 to 112 times while achieving 90% of the baseline performance on evaluation metrics. Additionally, IC-Former demonstrates strong regularity in its interactions with the context, enhancing its interpretability. Overall, IC-Former significantly reduces compression costs, making real-time compression scenarios feasible.

Recent work has revealed that in-context learning for large language models exhibits compositional generalization capacity, which can be enhanced by selecting in-context demonstrations similar to test cases to provide contextual information. However, how to exhibit in-context compositional generalization (ICCG) of large vision-language models (LVLMs) is non-trival. Due to the inherent asymmetry between visual and linguistic modalities, ICCG in LVLMs faces an inevitable challenge—redundant information on the visual modality. The redundant information affects in-context learning from two aspects: (1) Similarity calculation may be dominated by redundant information, resulting in sub-optimal demonstration selection. (2) Redundant information in in-context demonstrations brings misleading contextual information to in-context learning. To alleviate these problems, we propose a demonstration selection method to achieve ICCG for LVLMs, by considering two key factors of demonstrations: content and structure, from a multimodal perspective. Specifically, we design a diversity-coverage-based matching score to select demonstrations with maximum coverage, and avoid selecting demonstrations with redundant information via their content redundancy and structural complexity. We build a GQA-ICCG dataset to simulate the ICCG setting, and conduct experiments on GQA-ICCG and the VQA v2 dataset. Experimental results demonstrate the effectiveness of our method.

To effectively use large language models (LLMs) for real-world queries, it is imperative that they generalize to the long-tail distribution, i.e. rare examples where models exhibit low confidence. In this work, we take the first step towards evaluating LLMs in the long-tail distribution of inferential knowledge. We exemplify long-tail evaluation on the Natural Language Inference task. First, we introduce Logic-Induced-Knowledge-Search (LINK), a systematic long-tail data generation framework, to obtain factually-correct yet long-tail inferential statements. LINK uses variable-wise prompting grounded on symbolic rules to seek low-confidence statements while ensuring factual correctness. We then use LINK to curate Logic-Induced-Long-Tail (LINT), a large-scale long-tail inferential knowledge dataset that contains 108K statements spanning four domains. We evaluate popular LLMs on LINT; we find that state-of-the-art LLMs show significant performance drop (21% relative drop for GPT4) on long-tail data as compared to on head distribution data, and smaller models show even more generalization weakness. These results further underscore the necessity of long-tail evaluation in developing generalizable LLMs.

Adapting large language models (LLMs) to specific tasks remains challenging due to the extensive retraining required, prompting the need for efficient adapter techniques. Despite this, the concurrent serving of multiple adapters, each with unique matrix shapes, poses significant system-level challenges. To address these issues, we identify an opportunity in structurally sparse adapters, which, unlike low-rank adapters, maintain consistent matrix shapes while varying in sparsity patterns. Leveraging this characteristic, we introduce SpartanServe, a system designed for efficient concurrent serving of LLMs using multiple structurally sparse adapters. SpartanServe employs a unified matrix multiplication operation and a novel memory management technique to enable effective batching. Furthermore, the incorporation of Triton kernels enhances the acceleration of matrix multiplication in the serving process. Experimental results demonstrate that SpartanServe achieves 2.12× speedup over S-LoRA when serving 96 adapters using a single NVIDIA A100 GPU (40GB), showcasing its efficacy in concurrent LLM serving.

Recent progress in Spoken Language Modeling has shown that learning language directly from speech is feasible. Generating speech through a pipeline that operates at the text level typically loses nuances, intonations, and non-verbal vocalizations. Modeling directly from speech opens up the path to more natural and expressive systems. On the other hand, speech-only systems require up to three orders of magnitude more data to catch up to their text-based counterparts in terms of their semantic abilities. We show that fine-tuning speech representation models on phoneme classification leads to more context-invariant representations, and language models trained on these units achieve comparable lexical comprehension to ones trained on hundred times more data.

Accurately retrieving relevant bid keywords for user queries is critical in Sponsored Search but remains challenging, particularly for short, ambiguous queries. Existing dense and generative retrieval models often fail to capture the nuanced user intent in these cases. To address this, we propose an approach to enhance query understanding by augmenting queries with rich contextual signals derived from web search results and large language models, stored in an online cache. Specifically, we use web search titles and snippets to ground queries in real-world information, and utilize GPT-4 to generate query rewrites and explanations that clarify user intent. These signals are efficiently integrated through a Fusion-in-Decoder based Unity architecture, enabling both dense and generative retrieval with serving costs on par with traditional context-free models. To address scenarios where context is unavailable in the cache, we introduce context glancing, a curriculum learning strategy that improves model robustness and performance even without contextual signals during inference. Extensive offline experiments demonstrate that our context-aware approach substantially outperforms context-free models. Furthermore, online A/B testing on a prominent search engine across 160+ countries shows significant improvements in user engagement and revenue.

Existing auto-regressive large language models (LLMs) are primarily trained using documents from general domains. In the biomedical domain, continual pre-training is a prevalent method for domain adaptation to inject professional knowledge into powerful LLMs that have been pre-trained in general domains. Previous studies typically conduct standard pre-training by randomly packing multiple documents into a long pre-training sequence. Recently, some existing works suggest that enhancing the relatedness of documents within the same pre-training sequence may be advantageous. However, these studies primarily focus on general domains, which cannot be readily applied in the biomedical domain where the distinction of fine-grained topics is harder. Is it possible to further improve the pre-training for biomedical language models (LMs) using exactly the same corpus? In this paper, we explore an improved approach to continual pre-training, which is a prevalent method for domain adaptation, by utilizing information from the citation network in this challenging scenario. Empirical studies demonstrate that our proposed LinkLM data improves both the intra-sample and inter-sample referring abilities of auto-regressive LMs in the biomedical domain, encouraging more profound consideration of task-specific pre-training sequence design for continual pre-training.

Humans naturally attribute utterances of direct speech to their speaker in literary works.When attributing quotes, we process contextual information but also access mental representations of characters that we build and revise throughout the narrative. Recent methods to automatically attribute such utterances have explored simulating human logic with deterministic rules or learning new implicit rules with neural networks when processing contextual information.However, these systems inherently lack character representations, which often leads to errors in more challenging examples of attribution: anaphoric and implicit quotes.In this work, we propose to augment a popular quotation attribution system, BookNLP, with character embeddings that encode global stylistic information of characters derived from an off-the-shelf stylometric model, Universal Authorship Representation (UAR).We create DramaCV, a corpus of English drama plays from the 15th to 20th century that we automatically annotate for Authorship Verification of fictional characters utterances, and release two versions of UAR trained on DramaCV, that are tailored for literary characters analysis.Then, through an extensive evaluation on 28 novels, we show that combining BookNLP’s contextual information with our proposed global character embeddings improves the identification of speakers for anaphoric and implicit quotes, reaching state-of-the-art performance.Code and data can be found at https://github.com/deezer/character_embeddings_qa.

Advancements in Large Language Models (LLMs) have significantly enhanced instruction-following capabilities. However, most Instruction Fine-Tuning (IFT) datasets are predominantly in English, limiting model performance in other languages. Traditional methods for creating multilingual IFT datasets—such as translating existing English IFT datasets or converting existing NLP datasets into IFT datasets by templating—struggle to capture linguistic nuances and ensure prompt (instruction) diversity. To address this issue, we propose a novel method for collecting multilingual IFT datasets that preserves linguistic naturalness and ensures prompt diversity. This approach leverages English-focused LLMs, monolingual corpora, and a scoring function to create high-quality, diversified IFT datasets in multiple languages. Experiments demonstrate that LLMs finetuned using these IFT datasets show notable improvements in both generative and discriminative tasks, indicating enhanced language comprehension by LLMs in non-English contexts. Specifically, on the multilingual summarization task, LLMs using our IFT dataset achieved 17.57% and 15.23% improvements over LLMs fine-tuned with translation-based and template-based datasets, respectively.

Previous studies on multi-party dialogue generation predominantly concentrated on modeling the reply-to structure of dialogue histories, always overlooking the coherence between generated responses and target utterances. To address this issue, we propose a Reinforcement Learning approach emphasizing both Topic and Rhetorical Coherence (RL-TRC). In particular, the topic- and rhetorical-coherence tasks are designed to enhance the model’s perception of coherence with the target utterance. Subsequently, an agent is employed to learn a coherence policy, which guides the generation of responses that are topically and rhetorically aligned with the target utterance. Furthermore, three discourse-aware rewards are developed to assess the coherence between the generated response and the target utterance, with the objective of optimizing the policy. The experimental results and in-depth analyses on two popular datasets demonstrate that our RL-TRC significantly outperforms the state-of-the-art baselines, particularly in generating responses that are more coherent with the target utterances.

Multi-agent debate has proven effective in improving large language models quality for reasoning and factuality tasks. While various role-playing strategies in multi-agent debates have been explored, in terms of the communication among agents, existing approaches adopt a brute force algorithm – each agent can communicate with all other agents. In this paper, we systematically investigate the effect of communication connectivity in multi-agent systems. Our experiments on GPT and Mistral models reveal that multi-agent debates leveraging sparse communication topology can achieve comparable or superior performance while significantly reducing computational costs. Furthermore, we extend the multi-agent debate framework to multi-modal reasoning and alignment labeling tasks, showcasing its broad applicability and effectiveness. Our findings underscore the importance of communication connectivity on enhancing the efficiency and effectiveness of the “society of minds” approach.

While instruction fine-tuned LLMs are effective text generators, sensitivity to prompt construction makes performance unstable and sub-optimal in practice. Relying on a single ‘best’ prompt cannot capture all differing approaches to a generation problem. Using this observation, we propose multi-prompt decoding, where many candidate generations are decoded from a prompt bank at inference-time. To ensemble candidates, we use Minimum Bayes Risk (MBR) decoding, which selects a final output using a trained value metric. We show multi-prompt improves MBR across a comprehensive set of conditional generation tasks, and show this is a result of estimating a more diverse and higher quality candidate space than that of a single prompt. Our experiments confirm multi-prompt improves generation across tasks, models and metrics.

We present Randomized Path-Integration (RPI) - a path-integration method for explaining language models via randomization of the integration path over the attention information in the model. RPI employs integration on internal attention scores and their gradients along a randomized path, which is dynamically established between a baseline representation and the attention scores of the model. The inherent randomness in the integration path originates from modeling the baseline representation as a randomly drawn tensor from a Gaussian diffusion process. As a consequence, RPI generates diverse baselines, yielding a set of candidate attribution maps. This set facilitates the selection of the most effective attribution map based on the specific metric at hand. We present an extensive evaluation, encompassing 11 explanation methods and 5 language models, including the Llama2 and Mistral models. Our results demonstrate that RPI outperforms latest state-of-the-art methods across 4 datasets and 5 evaluation metrics.