EMNLP 2024 Papers

Recent large language models (LLMs) have demonstrated remarkable generalization abilities in mathematics and logical reasoning tasks.Prior research indicates that LLMs pre-trained with programming language data exhibit high mathematical and reasoning abilities; however, this causal relationship has not been rigorously tested. Our research aims to verify which programming languages and features during pre-training affect logical inference performance. Specifically, we pre-trained decoder-based language models from scratch using datasets from ten programming languages (e.g., Python, C, Java) and three natural language datasets (Wikipedia, Fineweb, C4) under identical conditions. Thereafter, we evaluated the trained models in a few-shot in-context learning setting on logical reasoning tasks: FLD and bAbi, which do not require commonsense or world knowledge. The results demonstrate that nearly all models trained with programming languages consistently outperform those trained with natural languages, indicating that programming languages contain factors that elicit logic inference performance. In addition, we found that models trained with programming languages exhibit a better ability to follow instructions compared to those trained with natural languages. Further analysis reveals that the depth of Abstract Syntax Trees representing parsed results of programs also affects logical reasoning performance. These findings will offer insights into the essential elements of pre-training for acquiring the foundational abilities of LLMs.

Large Language Models (LLMs) are typically shipped with tokenizers that *deterministically* encode text into so-called *canonical* token sequences, to which the LLMs assign probability values.One common assumption is that the probability of a piece of text is the probability of its canonical token sequence.However, the tokenization of a string is not unique: e.g., the Llama2 tokenizer encodes ‘Tokens‘ as ‘[Tok,ens]‘, but ‘[Tok,en,s]‘ also represents the same text.In this paper, we study non-canonical tokenizations.We prove that, given a string, it is computationally hard to find the most likely tokenization for an autoregressive LLM, as well as to compute the marginal probability over all possible tokenizations.We then show how the marginal is, in most cases, indistinguishable from the canonical probability.Surprisingly, we then empirically demonstrate the existence of a significant amount of signal hidden within tokenization space.Notably, by simply aggregating the probabilities of non-canonical tokenizations, we achieve improvements across a range of LLM evaluation benchmarks for a variety of architectures, including transformers and state space models.

As the utilization of Large Language Models (LLMs) becomes more widespread, there is a growing demand for their ability to handle more complex and longer external knowledge across various use cases. Most existing evaluations of the open-ended question answering (ODQA) task, which necessitates the use of external knowledge, focus solely on whether the model provides the correct answer. However, even when LLMs answer correctly, they often fail to provide an obvious source for their responses. Therefore, it is necessary to jointly evaluate and verify the correctness of the answers and the appropriateness of grounded evidence in complex external contexts. To address this issue, we examine the phenomenon of discrepancies in abilities across two distinct tasks—QA and evidence selection—when performed simultaneously, from the perspective of task alignment. To verify LLMs’ task alignment, we introduce a verification framework and resources considering both semantic relevancy and structural diversity of the given long context knowledge. Through extensive experiments and detailed analysis, we provide insights into the task misalignment between QA and evidence selection. Our code and resources will be available upon acceptance.

In visual speech processing, context modeling capability is one of the most important requirements due to the ambiguous nature of lip movements. For example, homophenes, words that share identical lip movements but produce different sounds, can be distinguished by considering the context. In this paper, we propose a novel framework, namely Visual Speech Processing incorporated with LLMs (VSP-LLM), to maximize the context modeling ability by bringing the overwhelming power of LLMs. Specifically, VSP-LLM is designed to perform multi-tasks of visual speech recognition and translation, where the given instructions control the type of task. The input video is mapped to the input latent space of an LLM by employing a self-supervised visual speech model. Focused on the fact that there is redundant information in input frames, we propose a novel deduplication method that reduces the embedded visual features by employing visual speech units. Through the proposed deduplication and low rank adaptation, VSP-LLM can be trained in a computationally efficient manner. In the translation dataset, the MuAViC benchmark, we demonstrate that VSP-LLM trained on just 30 hours of labeled data can more effectively translate compared to the recent model trained with 433 hours of data.

Generative models, particularly large language models (LLMs), have achieved remarkable success in producing natural and high-quality content. However, their widespread adoption raises concerns regarding copyright infringement, privacy violations, and security risks associated with AI-generated content. To address these concerns, we propose a novel digital forensics framework for LLMs, enabling the tracing of AI-generated content back to its source. This framework embeds a secret watermark directly into the generated output, eliminating the need for model retraining. To enhance traceability, especially for short outputs, we introduce a “depth watermark” that strengthens the link between content and generator. Our approach ensures accurate tracing while maintaining the quality of the generated content. Extensive experiments across various settings and datasets validate the effectiveness and robustness of our proposed framework.

Prompting serves as the major way humans interact with Large Language Models (LLM). Commercial AI systems commonly define the role of the LLM in system prompts. For example, ChatGPT uses ”You are a helpful assistant” as part of its default system prompt. Despite current practices of adding personas to system prompts, it remains unclear how different personas affect a model’s performance on objective tasks. In this study, we present a systematic evaluation of personas in system prompts. We curate a list of 162 roles covering 6 types of interpersonal relationships and 8 domains of expertise. Through extensive analysis of 4 popular families of LLMs and 2,410 factual questions, we demonstrate that adding personas in system prompts does not improve model performance across a range of questions compared to the control setting where no persona is added. Nevertheless, further analysis suggests that the gender, type, and domain of the persona can all influence the resulting prediction accuracies. We further experimented with a list of persona search strategies and found that, while aggregating results from the best persona for each question significantly improves prediction accuracy, automatically identifying the best persona is challenging, with predictions often performing no better than random selection. Overall, our findings suggest that while adding a persona may lead to performance gains in certain settings, the effect of each persona can be largely random. %Our results can help inform the design of system prompts for AI systems. Code and data are available at https://github.com/Jiaxin-Pei/Prompting-with-Social-Roles.

This study investigates the predicate-argument structure in Korean language processing. Despite the importance of distinguishing mandatory arguments and optional modifiers in sentences, research in this area has been limited. We introduce a dataset with token-level annotations which labels mandatory and optional elements as complements and adjuncts, respectively. Particularly, we reclassify certain Korean phrases, previously misidentified as adverbial phrases, as complements, addressing misuses of the term adjunct in existing Korean treebanks. Utilizing a Korean dependency treebank, we develop an automatic labeling technique for complements and adjuncts. Experiments using the proposed dataset yield satisfying results, demonstrating that the dataset is trainable and reliable.

We introduce a neural architecture finetuned for the task of scenario context generation: The relevant location and time of an event or entity mentioned in text. Contextualizing information extraction helps to scope the validity of automated finings when aggregating them as knowledge graphs. Our approach uses a high-quality curated dataset of time and location annotations in a corpus of epidemiology papers to train an encoder-decoder architecture. We also explored the use of data augmentation techniques during training. Our findings suggest that a relatively small fine-tuned encoder-decoder model performs better than out-of-the-box LLMs and semantic role labeling parsers to accurate predict the relevant scenario information of a particular entity or event.

Reasoning is most powerful when an LLM accurately aggregates relevant information. We examine the critical role of information aggregation in reasoning by requiring the LLM to analyze sports narratives. To succeed at this task, an LLM must infer points from actions, identify related entities, attribute points accurately to players and teams, and compile key statistics to draw conclusions. We conduct comprehensive experiments with real NBA basketball data and present SportsGen, a new method to synthesize game narratives. By synthesizing data, we can rigorously evaluate LLMs’ reasoning capabilities under complex scenarios with varying narrative lengths and density of information. Our findings show that most models, including GPT-4o, often fail to accurately aggregate basketball scores due to frequent scoring patterns. Open-source models like Llama-3 further suffer from significant score hallucinations. Finally, the effectiveness of reasoning is influenced by narrative complexity, information density, and domain-specific terms, highlighting the challenges in analytical reasoning tasks.

The introduction of Large Language Models (LLMs) has advanced data representation and analysis, bringing significant progress in their use for medical questions and answering. Despite these advancements, integrating tabular data, especially numerical data pivotal in clinical contexts, into LLM paradigms has not been thoroughly explored. In this study, we examine the effectiveness of vector representations from last hidden states of LLMs for medical diagnostics and prognostics using electronic health record (EHR) data. We compare the performance of these embeddings with that of raw numerical EHR data when used as feature inputs to traditional machine learning (ML) algorithms that excel at tabular data learning, such as eXtreme Gradient Boosting. We focus on instruction-tuned LLMs in a zero-shot setting to represent abnormal physiological data and evaluating their utilities as feature extractors to enhance ML classifiers for predicting diagnoses, length of stay, and mortality. Furthermore, we examine prompt engineering techniques on zero-shot and few-shot LLM embeddings to measure their impact comprehensively. Although findings suggest the raw data features still prevail in medical ML tasks, zero-shot LLM embeddings demonstrate competitive results, suggesting a promising avenue for future research in medical applications.

This paper studies in-context learning by decomposing the output of large language models into the individual contributions of attention heads and MLPs (components). We observe curious components: good-performing ones that individually do well on a classification task, even when the model performs poorly; bad-performing ones that do much worse than chance; and label-biased components that always predict the same label. We find that component accuracies are well-correlated across different demonstration sets and perturbations of prompt templates. Based on our findings, we propose component reweighting, which learns to linearly re-scale the component activations from a few labeled examples. Given 24 labeled examples, our method improves by an average of 6.0% accuracy points over 24-shot ICL across 8 tasks on Llama-2-7B. Overall, this paper both enriches our understanding of ICL and provides a practical method for improvement by examining model internals.

Depression is a critical concern in global mental health, prompting extensive research into AI-based detection methods. Among various AI technologies, Large Language Models (LLMs) stand out for their versatility in healthcare applications. However, the application of LLMs in the identification and analysis of depressive states remains relatively unexplored, presenting an intriguing avenue for future research. In this paper, we present an innovative approach to employ an LLM in the realm of depression detection, integrating acoustic speech information into the LLM framework for this specific application. We investigate an efficient method for automatic depression detection by integrating speech signals into LLMs utilizing Acoustic Landmarks. This approach is not only valuable for the detection of depression but also represents a new perspective in enhancing the ability of LLMs to comprehend and process speech signals. By incorporating acoustic landmarks, which are specific to the pronunciation of spoken words, our method adds critical dimensions to text transcripts. This integration also provides insights into the unique speech patterns of individuals, revealing the potential mental states of individuals. By encoding acoustic landmarks information into LLMs, evaluations of the proposed approach on the DAIC-WOZ dataset reveal state-of-the-art results when compared with existing Audio-Text baselines.

Multilingual language models are widely used to extend NLP systems to low-resource languages. However, concrete evidence for the effects of multilinguality on language modeling performance in individual languages remains scarce. Here, we pre-train over 10,000 monolingual and multilingual language models for over 250 languages, including multiple language families that are under-studied in NLP. We assess how language modeling performance in each language varies as a function of (1) monolingual dataset size, (2) added multilingual dataset size, (3) linguistic similarity of the added languages, and (4) model size (up to 45M parameters). We find that in moderation, adding multilingual data improves low-resource language modeling performance, similar to increasing low-resource dataset sizes by up to 33%. Improvements depend on the syntactic similarity of the added multilingual data, with marginal additional effects of vocabulary overlap. However, high-resource languages consistently perform worse in multilingual pre-training scenarios. As dataset sizes increase, adding multilingual data begins to hurt performance for both low-resource and high-resource languages, likely due to limited model capacity (the “curse of multilinguality”). These results suggest that massively multilingual pre-training may not be optimal for any languages involved, but that more targeted models can significantly improve performance.

The current framework for sequence labeling encompasses a feature extractor and a sequence tagger. This study introduces a unified framework named SLGAN, which harnesses the capabilities of Generative Adversarial Networks to address the challenges associated with Sequence Labeling tasks. SLGAN not only mitigates the limitation of GANs in backpropagating loss to discrete data but also exhibits strong adaptability to various sequence labeling tasks. Unlike traditional GANs, the discriminator within SLGAN does not discriminate whether data originates from the discriminator or the generator; instead, it focuses on predicting the correctness of each tag within the tag sequence. We conducted evaluations on six different tasks spanning four languages, including Chinese, Japanese, and Korean Word Segmentation, Chinese and English Named Entity Recognition, and Chinese Part-of-Speech Tagging. Our experimental results illustrate that SLGAN represents a versatile and highly effective solution, consistently achieving state-of-the-art or competitive performance results, irrespective of the specific task or language under consideration.

Large language models (LLMs) have demonstrated remarkable progress in leveraging diverse knowledge sources. This study investigates how nine widely used LLMs allocate knowledge between local context and global parameters when answering open-ended questions in knowledge-consistent scenarios. We introduce a novel dataset, WikiAtomic, and systematically vary context sizes to analyze how LLMs prioritize and utilize the provided information and their parametric knowledge in knowledge-consistent scenarios. Additionally, we also study their tendency to hallucinate under varying context sizes. Our findings reveal consistent patterns across models, including a consistent reliance on both contextual (around 70%) and parametric (around 30%) knowledge, and a decrease in hallucinations with increasing context. These insights highlight the importance of more effective context organization and developing models that use input more deterministically for robust performance.

Large language models (LLMs) exhibit excellent performance in various tasks. However, the memory requirements of LLMs present a great challenge when deploying on memory-limited devices, even for quantized LLMs. This paper introduces a framework to compress LLM after quantization further, achieving about 2.2x compression ratio. A compression-aware quantization is first proposed to enhance model weight compressibility by re-scaling the model parameters before quantization, followed by a pruning method to improve further. Upon this, we notice that decompression can be a bottleneck during practical scenarios. We then give a detailed analysis of the trade-off between memory usage and latency brought by the proposed method. A speed-adaptive method is proposed to overcome it. The experimental results show inference with the compressed model can achieve a 40% reduction in memory size with negligible loss in accuracy and inference speed.

Meeting summarization has become a critical task considering the increase in online interactions. Despite new techniques being proposed regularly, the evaluation of meeting summarization techniques relies on metrics not tailored to capture meeting-specific errors, leading to ineffective assessment. This paper explores what established automatic metrics capture and the errors they mask by correlating metric scores with human evaluations across a comprehensive error taxonomy. We start by reviewing the literature on English meeting summarization to identify key challenges, such as speaker dynamics and contextual turn-taking, and error types, including missing information and linguistic inaccuracy, concepts previously loosely defined in the field. We then examine the relationship between these challenges and errors using human annotated transcripts and summaries from encoder-decoder-based and autoregressive Transformer models on the QMSum dataset. Experiments reveal that different model architectures respond variably to the challenges, resulting in distinct links between challenges and errors. Current established metrics struggle to capture the observable errors, showing weak to moderate correlations, with a third of the correlations indicating error masking. Only a subset of metrics accurately reacts to specific errors, while most correlations show either unresponsiveness or failure to reflect the error’s impact on summary quality.

Best practices for high conflict conversations like counseling or customer support almost always include recommendations to paraphrase the previous speaker. Although paraphrase classification has received widespread attention in NLP, paraphrases are usually considered independent from context, and common models and datasets are not applicable to dialog settings. In this work, we investigate paraphrases across turns in dialog (e.g., Speaker 1: “That book is mine.” becomes Speaker 2: “That book is yours.”). We provide an operationalization of context-dependent paraphrases, and develop a training for crowd-workers to classify paraphrases in dialog. We introduce ContextDeP, a dataset with utterance pairs from NPR and CNN news interviews annotated for context-dependent paraphrases. To enable analyses on label variation, the dataset contains 5,581 annotations on 600 utterance pairs. We present promising results with in-context learning and with token classification models for automatic paraphrase detection in dialog.

Gender bias in machine translation (MT) is recognized as an issue that can harm people and society. And yet, advancements in the field rarely involve people, the final MT users, or inform how they might be impacted by biased technologies. Current evaluations are often restricted to automatic methods, which offer an opaque estimate of what the downstream impact of gender disparities might be. We conduct an extensive human-centered study to examine if and to what extent bias in MT brings harms with tangible costs, such as quality of service gaps across women and men. To this aim, we collect behavioral data from ~90 participants, who post-edited MT outputs to ensure correct gender translation. Across multiple datasets, languages, and types of users, our study shows that feminine post-editing demands significantly more technical and temporal effort, also corresponding to higher financial costs. Existing bias measurements, however, fail to reflect the found disparities. Our findings advocate for human-centered approaches that can inform the societal impact of bias.