EMNLP 2024 Papers

Modular programming, which aims to construct the final program by integrating smaller, independent building blocks, has been regarded as a desirable practice in software development. However, with the rise of recent code generation agents built upon large language models (LLMs), a question emerges: is this traditional practice equally effective for these new tools? In this work, we assess the impact of modularity in code generation by introducing a novel metric for its quantitative measurement. Surprisingly, unlike conventional wisdom on the topic, we find that modularity is not a core factor for improving the performance of code generation models. We also explore potential explanations for why LLMs do not exhibit a preference for modular code compared to non-modular code.

This paper investigates Who’s Harry Potter (WHP), a pioneering yet insufficiently understood method for LLM unlearning. We explore it in two steps. First, we introduce a new task of LLM targeted unlearning, where given an unlearning target (e.g., a person) and some unlearning documents, we aim to unlearn only the information about the target, rather than everything in the unlearning documents. We further argue that a successful unlearning should satisfy criteria such as not outputting gibberish, not fabricating facts about the unlearning target, and not releasing factual information under jailbreak attacks. Second, we construct a causal intervention framework for targeted unlearning, where the knowledge of the unlearning target is modeled as a confounder between LLM input and output, and the unlearning process as a deconfounding process. This framework justifies and extends WHP, deriving a simple unlearning algorithm that includes WHP as a special case. Experiments on existing and new datasets show that our approach, without explicitly optimizing for the aforementioned criteria, achieves competitive performance in all of them.

Microblog content (e.g., Tweets) is noisy due to its informal use of language and its lack of contextual information within each post. To tackle these challenges, state-of-the-art microblog classification models rely on pre-training language models (LMs). However, pre-training dedicated LMs is resource-intensive and not suitable for small labs. Supervised contrastive learning (SCL) has shown its effectiveness with small, available resources. In this work, we examine the effectiveness of fine-tuning transformer-based language models, regularized with a SCL loss for English microblog classification. Despite its simplicity, the evaluation on two English microblog classification benchmarks (TweetEval and Tweet Topic Classification) shows an improvement over baseline models. The result shows that, across all subtasks, our proposed method has a performance gain of up to 11.9 percentage points. All our models are open source.

We present new supertaggers trained on English HPSG-based treebanks and test the effects of the best tagger on parsing speed and accuracy. HPSG treebanks are produced automatically by large manually built grammars and feature high-quality annotation based on a well-developed linguistic theory. The English Resource Grammar treebanks include diverse and challenging test datasets, beyond the usual WSJ section 23 and Wikipedia data. HPSG supertagging has previously relied on MaxEnt-based models. We use SVM and neural CRF- and BERT-based methods and show that both SVM and neural supertaggers achieve considerably higher accuracy compared to the baseline and lead to an increase not only in the parsing speed but also the parser accuracy with respect to gold dependency structures. Our fine-tuned BERT-based tagger achieves 97.26% accuracy on 950 sentences from WSJ23 and 93.88% on the out-of-domain technical essay The Cathedral and the Bazaar. We present experiments with integrating the best supertagger into an HPSG parser and observe a speedup of a factor of 3 with respect to the system which uses no tagging at all, as well as large recall gains and an overall precision gain. We also compare our system to an existing integrated tagger and show that although the well-integrated tagger remains the fastest, our experimental system can be more accurate. Finally, we hope that the diverse and difficult datasets we used for evaluation will gain more popularity in the field: we show that results can differ depending on the dataset, even if it is an in-domain one. We contribute the complete datasets reformatted for Huggingface token classification.

The most commonly used transformers for retrieval at present, BERT and T5, have been shown not to be robust to query variations such as typos or paraphrases. Although this is an important prerequisite for their practicality, this problem has hardly been investigated. More recent large language models (LLMs), including instruction-tuned LLMs, have not been analyzed yet, and only one study looks beyond typos. We close this gap by reproducing this study and extending it with a systematic analysis of more recent models, including Sentence-BERT, CharacterBERT, E5-Mistral, AnglE, and Ada v2. We further investigate if instruct-LLMs can be prompted for robustness. Our results are mixed in that the previously observed robustness issues for cross-encoders also apply to bi-encoders that use much larger LLMs, albeit to a lesser extent. While further LLM scaling may improve their embeddings, their cost-effective use for all but large deployments is limited. Training data that includes query variations allows LLMs to be fine-tuned for more robustness, but focusing on a single category of query variation may even degrade the effectiveness on others. Our code, results, and artifacts can be found at https://github.com/webis-de/EMNLP-24

Catastrophic Forgetting (CF) means models forgetting previously acquired knowledge when learning new data. It compromises the effectiveness of large language models (LLMs) during fine-tuning, yet the underlying causes have not been thoroughly investigated. This paper takes the first step to reveal the direct link between the flatness of the model loss landscape and the extent of CF in the field of LLMs. Based on this, we introduce the sharpness-aware minimization to mitigate CF by flattening the loss landscape. Experiments on three widely-used fine-tuning datasets, spanning different model scales, demonstrate the effectiveness of our method in alleviating CF. Analyses show that we nicely complement the existing anti-forgetting strategies, further enhancing the resistance of LLMs to CF.

In the era of data-driven decision-making, Long-Form Table Question Answering (LFTQA) is essential for integrating structured data with complex reasoning. Despite recent advancements in Large Language Models (LLMs) for LFTQA, evaluating their effectiveness remains a significant challenge. We introduce LFTQA-Eval, a meta-evaluation dataset comprising 2,988 human-annotated examples, to rigorously assess the efficacy of current automated metrics in assessing LLM-based LFTQA systems, with a focus on faithfulness and comprehensiveness. Our findings reveal that existing automatic metrics poorly correlate with human judgments and fail to consistently differentiate between factually accurate responses and those that are coherent but factually incorrect. Additionally, our in-depth examination of the limitations associated with automated evaluation methods provides essential insights for the improvement of LFTQA automated evaluation.

Numerous previous studies have sought to determine to what extent language models, pretrained on natural language text, can serve as useful models of human cognition.In this paper, we are interested in the opposite question: whether we can directly optimize a language model to be a useful cognitive model by aligning it to human psychometric data.To achieve this, we introduce a novel alignment technique in which we fine-tune a language model to implicitly optimize the parameters of a linear regressor that directly predicts humans’ reading times of in-context linguistic units, e.g., phonemes, morphemes, or words, using surprisal estimates derived from the language model. Using words as a test case, we evaluate our technique across multiple model sizes and datasets and find that it improves language models’ psychometric predictive power.However, we find an inverse relationship between psychometric power and a model’s performance on downstream NLP tasks as well as its perplexity on held-out test data.While this latter trend has been observed before (Oh et al., 2022; Shain et al., 2024), we are the first to induce it by manipulating a model’s alignment to psychometric data.

Large language models (LLMs) can handle multilingual and cross-lingual text within a single input; however, previous works leveraging multilingualism in LLMs primarily focus on using English as the pivot language to enhance language understanding and reasoning. Given that multiple languages are a compensation for the losses caused by a single language’s limitations, it’s a natural next step to enrich the model’s learning context through the integration of the original input with its multiple translations. In this paper, we start by revealing that LLMs learn from parallel multilingual input (PMI). Our comprehensive evaluation shows that PMI enhances the model’s comprehension of the input, achieving superior performance than conventional in-context learning (ICL). Furthermore, to explore how multilingual processing affects prediction, we examine the activated neurons in LLMs. Surprisingly, involving more languages in the input activates fewer neurons, leading to more focused and effective neural activation patterns. Also, this neural reaction coincidently mirrors the neuroscience insight about synaptic pruning, highlighting a similarity between artificial and biological ‘brains’.

Personality recognition aims to identify the personality traits implied in user data such as dialogues and social media posts. Current research predominantly treats personality recognition as a classification task, failing to reveal the supporting evidence for the recognized personality. In this paper, we propose a novel task named Explainable Personality Recognition, aiming to reveal the reasoning process as supporting evidence of the personality trait. Inspired by personality theories, personality traits are made up of stable patterns of personality state, where the states are short-term characteristic patterns of thoughts, feelings, and behaviors in a concrete situation at a specific moment in time. We propose an explainable personality recognition framework called Chain-of-Personality-Evidence (CoPE), which involves a reasoning process from specific contexts to short-term personality states to long-term personality traits. Furthermore, based on the CoPE framework, we construct an explainable personality recognition dataset from dialogues, PersonalityEvd. We introduce two explainable personality state recognition and explainable personality trait recognition tasks, which require models to recognize the personality state and trait labels and their corresponding support evidence. Our extensive experiments based on Large Language Models on the two tasks show that revealing personality traits is very challenging and we present some insights for future research. We will release our dataset and source code to facilitate further studies in this direction.

Social media is recognized as an important source for deriving insights into public opinion dynamics and social impacts due to the vast textual data generated daily and the ‘unconstrained’ behavior of people interacting on these platforms. However, such analyses prove challenging due to the semantic shift phenomenon, where word meanings evolve over time. This paper proposes an unsupervised dynamic word embedding method to capture longitudinal semantic shifts in social media data without predefined anchor words. The method leverages word co-occurrence statistics and dynamic updating to adapt embeddings over time, addressing the challenges of data sparseness, imbalanced distributions, and synergistic semantic effects. Evaluated on a large COVID-19 Twitter dataset, the method reveals semantic evolution patterns of vaccine- and symptom-related entities across different pandemic stages, and their potential correlations with real-world statistics. Our key contributions include the dynamic embedding technique, empirical analysis of COVID-19 semantic shifts, and discussions on enhancing semantic shift modeling for computational social science research. This study enables capturing longitudinal semantic dynamics on social media to understand public discourse and collective phenomena.

Large language models (LLMs) have brought a great breakthrough to the natural language processing (NLP) community, while leading the challenge of handling concurrent customer queries due to their high throughput demands. Data multiplexing addresses this by merging multiple inputs into a single composite input, allowing more efficient inference through a shared forward pass. However, as distinguishing individuals from a composite input is challenging, conventional methods typically require training the entire backbone, yet still suffer from performance degradation. In this paper, we introduce RevMUX, a parameter-efficient data multiplexing framework that incorporates a reversible design in the multiplexer, which can be reused by the demultiplexer to perform reverse operations and restore individual samples for classification. Extensive experiments on four datasets and three types of LLM backbones demonstrate the effectiveness of RevMUX for enhancing LLM inference efficiency while retaining a satisfactory classification performance.

Text style transfer (TST) is crucial in natural language processing, aiming to endow text with a new style without altering its meaning. In real-world scenarios, not all styles have abundant resources. This work introduces TWIST (reusing Transferable Weight Increments for Style Text generation), a novel framework to mitigate data scarcity by utilizing style features in weight increments to transfer low-resource styles effectively. During target style learning, we derive knowledge via a specially designed weight pool and initialize the parameters for the unseen style. To enhance the effectiveness of merging, the target style weight increments are often merged from multiple source style weight increments through singular vectors. Considering the diversity of styles, we also designed a multi-key memory network that simultaneously focuses on task- and instance-level information to derive the most relevant weight increments. Results from multiple style transfer datasets show that TWIST demonstrates remarkable performance across different backbones, achieving particularly effective results in low-resource scenarios.

Aligning language models (LMs) based on human-annotated preference data is a crucial step in obtaining practical and performant LM-based systems. However, multilingual human preference data are difficult to obtain at scale, making it challenging to extend this framework to diverse languages. In this work, we evaluate a simple approach for zero-shot cross-lingual alignment, where a reward model is trained on preference data in one source language and directly applied to other target languages. On summarization and open-ended dialog generation, we show that this method is consistently successful under comprehensive evaluation settings, including human evaluation: cross-lingually aligned models are preferred by humans over unaligned models on up to >70% of evaluation instances. We moreover find that a different-language reward model sometimes yields better aligned models than a same-language reward model. We also identify best practices when there is no language-specific data for even supervised finetuning, another component in alignment.

Large language models (LLMs) possess extensive knowledge and commonsense reasoning capabilities, making them valuable for creating powerful agents. However, existing LLM agent frameworks have not fully utilized past experiences for improvement. This work introduces a new LLM-based agent framework called Retrospex, which addresses this challenge by analyzing past experiences in depth. Unlike previous approaches, Retrospex does not directly integrate experiences into the LLM’s context. Instead, it combines the LLM’s action likelihood with action values estimated by a Reinforcement Learning (RL) Critic, which is trained on past experiences through an offline “retrospection” process. Additionally, Retrospex employs a dynamic action rescoring mechanism that increases the importance of experience-based values for tasks that require more interaction with the environment. We evaluate Retrospex in ScienceWorld, ALFWorld and Webshop environments, demonstrating its advantages over strong baselines.

Recent Retrieval Augmented Generation (RAG) aims to enhance Large Language Models (LLMs) by incorporating extensive knowledge retrieved from external sources. However, such approach encounters some challenges: Firstly, the original queries may not be suitable for precise retrieval, resulting in erroneous contextual knowledge; Secondly, the language model can easily generate inconsistent answer with external references due to their knowledge boundary limitation. To address these issues, we propose the chain-of-verification (CoV-RAG) to enhance the external retrieval correctness and internal generation consistency. Specifically, we integrate the verification module into the RAG, engaging in scoring, judgment, and rewriting. To correct external retrieval errors, CoV-RAG retrieves new knowledge using a revised query. To correct internal generation errors, we unify QA and verification tasks with a Chain-of-Thought (CoT) reasoning during training. Our comprehensive experiments across various LLMs demonstrate the effectiveness and adaptability compared with other strong baselines. Especially, our CoV-RAG can significantly surpass the state-of-the-art baselines using different LLM backbones.

Long-form question answering (LFQA) aims at generating in-depth answers to end-user questions, providing relevant information beyond the direct answer. However, existing retrievers are typically optimized towards information that directly targets the question, missing out on such contextual information. Furthermore, there is a lack of training data for relevant context. To this end, we propose and compare different weak supervision techniques to optimize retrieval for contextual information. Experiments demonstrate improvements on the end-to-end QA performance on ASQA, a dataset for long-form question answering. Importantly, as more contextual information is retrieved, we improve the relevant page recall for LFQA by 14.7% and the groundedness of generated long-form answers by 12.5%. Finally, we show that long-form answers often anticipate likely follow-up questions, via experiments on a conversational QA dataset.

Protein Language Models traditionally depend on Multiple Sequence Alignments (MSA) to incorporate evolutionary knowledge. However, MSA-based approaches suffer from substantial computational overhead and generally underperform in generalizing to de novo proteins. This study reevaluates the role of MSA, proposing it as a retrieval augmentation method and questioning the necessity of sequence alignment. We show that a simple alternative, Retrieved Sequence Augmentation (RSA), can enhance protein representation learning without the need for alignment and cumbersome preprocessing. RSA surpasses MSA Transformer by an average of 5% in both structural and property prediction tasks while being 373 times faster. Additionally, RSA demonstrates enhanced transferability for predicting de novo proteins. This methodology addresses a critical need for efficiency in protein prediction and can be rapidly employed to identify homologous sequences, improve representation learning, and enhance the capacity of Large Language Models to interpret protein structures.

In-context learning (ICL) has been instrumental in adapting large language models (LLMs) to downstream tasks using correct input-output examples. Recent advances have attempted to improve model performance through principles derived from mistakes, yet these approaches suffer from lack of customization and inadequate error coverage. To address these limitations, we propose Retrieved In-Context Principles (RICP), a novel teacher-student framework. In RICP, the teacher model analyzes mistakes from the student model to generate reasons and insights for preventing similar mistakes. These mistakes are clustered based on their underlying reasons for developing task-level principles, enhancing the error coverage of principles. During inference, the most relevant mistakes for each question are retrieved to create question-level principles, improving the customization of the provided guidance. RICP is orthogonal to existing prompting methods and does not require intervention from the teacher model during inference. Experimental results across seven reasoning benchmarks reveal that RICP effectively enhances performance when applied to various prompting strategies.