EMNLP 2024 Papers

Hate speech (HS) on social media exacerbates misinformation and baseless prejudices. Evidence-supported counterspeech (CS) is crucial for correcting misinformation and reducing prejudices through facts. Existing methods for generating evidence-supported CS often lack clear guidance with a core claim for organizing evidence and do not adequately address factuality and faithfulness hallucinations in CS within anti-hate contexts. In this paper, to mitigate the aforementioned, we propose F2RL, a Factuality and Faithfulness Reinforcement Learning framework for generating claim-guided and evidence-supported CS. Firstly, we generate counter-claims based on hate speech and design a self-evaluation mechanism to select the most appropriate one. Secondly, we propose a coarse-to-fine evidence retrieval method. This method initially generates broad queries to ensure the diversity of evidence, followed by carefully reranking the retrieved evidence to ensure its relevance to the claim. Finally, we design a reinforcement learning method with a triplet-based factuality reward model and a multi-aspect faithfulness reward model. The method rewards the generator to encourage greater factuality, more accurate refutation of hate speech, consistency with the claim, and better utilization of evidence. Extensive experiments on three benchmark datasets demonstrate that the proposed framework achieves excellent performance in CS generation, with strong factuality and faithfulness.

Hate speech is a complex and subjective phenomenon. In this paper, we present a dataset (GAZE4HATE) that provides gaze data collected in a hate speech annotation experiment. We study whether the gaze of an annotator provides predictors of their subjective hatefulness rating, and how gaze features can improve Hate Speech Detection (HSD). We conduct experiments on statistical modeling of subjective hate ratings and gaze and analyze to what extent rationales derived from hate speech models correspond to human gaze and explanations in our data. Finally, we introduce MEANION, a first gaze-integrated HSD model. Our experiments show that particular gaze features like dwell time or fixation counts systematically correlate with annotators’ subjective hate ratings and improve predictions of text-only hate speech models.

Productive interactions between diverse users and language technologies require outputs from the latter to be culturally relevant and sensitive. Prior works have evaluated models’ knowledge of cultural norms, values, and artefacts, without considering how this knowledge manifests in downstream applications. In this work, we focus on extrinsic evaluation of cultural competence in two text generation tasks, open-ended question answering and story generation. We quantitatively and qualitatively evaluate model outputs when an explicit cue of culture, specifically nationality, is perturbed in the prompts. Although we find that model outputs do vary when varying nationalities and feature culturally relevant words, we also find weak correlations between text similarity of outputs for different countries and the cultural values of these countries. Finally, we discuss important considerations in designing comprehensive evaluation of cultural competence in user-facing tasks.

Medical entity disambiguation (MED) aims to ground medical mentions in text with ontological entities in knowledge bases (KBs). A notable challenge of MED is the long medical text usually contains multiple entities’ mentions with intricate correlations. However, limited by computation overhead, many existing methods consider only a single candidate entity mention during the disambiguation process. As such, they focus only on local MED optimal while ignoring the sole-mention disambiguation possibly boosted by richer context from other mentions’ disambiguating processes – missing global optimal on entity combination in the text. Motivated by this, we propose a new approach called Extractive Medical Entity Disambiguation with Memory Mechanism and Memorized Entity Information (M3E). Specifically, we reformulate MED as a text extraction task, which simultaneously accepts the context of medical mentions, all possible candidate entities, and entity definitions, and it is then trained to extract the text span corresponding to the correct entity. Upon our new formulation, 1) to alleviate the computation overhead from the enriched context, we devise a memory mechanism module that performs memory caching, retrieval, fusion and cross-network residual; and 2) to utilize the disambiguation clues from other mentions, we design an auxiliary disambiguation module that employs a gating mechanism to assist the disambiguation of remaining mentions. Extensive experiments on two benchmark datasets demonstrate the superiority of M3E over the state-of-the-art MED methods on all metrics.

We consider the problem of language model inversion: given outputs of a language model, we seek to extract the prompt that generated these outputs. We develop a new black-box method, output2prompt, that extracts prompts without access to the model’s logits and without adversarial or jailbreaking queries. Unlike previous methods, output2prompt only needs outputs of normal user queries. To improve memory efficiency, output2prompt employs a new sparse encoding techique. We measure the efficacy of output2prompt on a variety of user and system prompts and demonstrate zero-shot transferability across different LLMs.

In this work, we are interested in automated methods for knowledge graph creation (KGC) from input text. Progress on large language models (LLMs) has prompted a series of recent works applying them to KGC, e.g., via zero/few-shot prompting. Despite successes on small domain-specific datasets, these models face difficulties scaling up to text common in many real-world applications. A principal issue is that, in prior methods, the KG schema has to be included in the LLM prompt to generate valid triplets; larger and more complex schemas easily exceed the LLMs’ context window length. Furthermore, there are scenarios where a fixed pre-defined schema is not available and we would like the method to construct a high-quality KG with a succinct self-generated schema. To address these problems, we propose a three-phase framework named Extract-Define-Canonicalize (EDC): open information extraction followed by schema definition and post-hoc canonicalization. EDC is flexible in that it can be applied to settings where a pre-defined target schema is available and when it is not; in the latter case, it constructs a schema automatically and applies self-canonicalization. To further improve performance, we introduce a trained component that retrieves schema elements relevant to the input text; this improves the LLMs’ extraction performance in a retrieval-augmented generation-like manner. We demonstrate on three KGC benchmarks that EDC is able to extract high-quality triplets without any parameter tuning and with significantly larger schemas compared to prior works. Code for EDC is available at https://github.com/clear-nus/edc.

Argument mining (AM) involves the identification of argument relations (AR) between Argumentative Discourse Units (ADUs). The essence of ARs among ADUs is context-dependent and lies in maintaining a coherent flow of ideas, often centered around the relations between discussed entities, topics, themes or concepts. However, these relations are not always explicitly stated; rather, inferred from implicit chains of reasoning connecting the concepts addressed in the ADUs. While humans can infer such background knowledge, machines face challenges when the contextual cues are not explicitly provided. This paper leverages external resources, including WordNet, ConceptNet, and Wikipedia to identify semantic paths (knowledge paths) connecting the concepts discussed in the ADUs to obtain the implicit chains of reasoning. To effectively leverage these paths for AR prediction, we propose attention-based Multi-Network architectures. Various architecture are evaluated on the external resources, and the Wikipedia based configuration attains F-scores of 0.85, 0.84, 0.70, and 0.87, respectively, on four diverse datasets, showing strong performance over the baselines.

Scaling the rotary position embedding (RoPE) has become a common method for extending the context window of RoPE-based large language models (LLMs). However, existing scaling methods often rely on empirical approaches and lack a profound understanding of the internal distribution within RoPE, resulting in suboptimal performance in extending the context window length. In this paper, we propose to optimize the context window extending task from the view of rotary angle distribution. Specifically, we first estimate the distribution of the rotary angles within the model and analyze the extent to which length extension perturbs this distribution. Then, we present a novel extension strategy that minimizes the disturbance between rotary angle distributions to maintain consistency with the pre-training phase, enhancing the model’s capability to generalize to longer sequences. Experimental results compared to the strong baseline methods demonstrate that our approach reduces by up to 72% of the distributional disturbance when extending LLaMA2’s context window to 8k, and reduces by up to 32% when extending to 16k. On the LongBench-E benchmark, our method achieves an average improvement of up to 4.33% over existing state-of-the-art methods. Furthermore, Our method maintains the model’s performance on the Hugging Face Open LLM benchmark after context window extension, with only an average performance fluctuation ranging from -0.12 to +0.22.

When performing complex multi-step reasoning tasks, the ability of Large Language Models (LLMs) to derive structured intermediate proof steps is important for ensuring that the models truly perform the desired reasoning and for improving models’ explainability. This paper is centred around a focused study: whether the current state-of-the-art generalist LLMs can leverage the structures in a few examples to better construct the proof structures with in-context learning. Our study specifically focuses on structure-aware demonstration and structure-aware pruning. We demonstrate that they both help improve performance. A detailed analysis is provided to help understand the results.

In-Context Learning (ICL) and Instruction Tuning (IT) are two primary paradigms of adopting Large Language Models (LLMs) to downstream applications. However, they are significantly different. In ICL, a set of demonstrations is provided at the inference time, but the LLM’s parameters are not updated. In IT, a set of demonstrations is used to adjust the parameters of the LLM during training, but no demonstrations are provided at the inference time. Although a growing body of literature has explored ICL and IT, studies on these topics have largely been conducted in isolation, leading to a disconnect between these two paradigms. In this work, we explore the relationship between ICL and IT by examining how the hidden states of LLMs change in these two paradigms. Through carefully designed experiments conducted with LLaMA-2 and LLaMA-2-Chat (7B and 13B), we find that ICL and IT converge in LLM hidden states despite their apparent differences in implementation. Specifically, ICL changes an LLM’s hidden states as if its accompanying demonstrations were used to instructionally tune the model. Furthermore, the convergence between ICL and IT is largely contingent upon several factors related to the demonstration. Overall, this work offers a unique perspective to explore the connection between ICL and IT and sheds light on understanding the behaviors of LLMs.

Benchmarking the performance of information retrieval (IR) is mostly conducted with a fixed set of documents (static corpora). However, in realistic scenarios, this is rarely the case and the documents to be retrieved are constantly updated and added. In this paper, we focus on Generative Retrievals (GR), which apply autoregressive language models to IR problems, and explore their adaptability and robustness in dynamic scenarios. We also conduct an extensive evaluation of computational and memory efficiency, crucial factors for real-world deployment of IR systems handling vast and ever-changing document collections. Our results on the StreamingQA benchmark demonstrate that GR is more adaptable to evolving knowledge (4–11%), robust in learning knowledge with temporal information, and efficient in terms of inference FLOPs (x2), indexing time (x6), and storage footprint (x4) compared to Dual Encoders (DE), which are commonly used in retrieval systems. Our paper highlights the potential of GR for future use in practical IR systems within dynamic environments.

Recent advancements in multimodal large language models (MLLMs) have made significant progress in integrating information across various modalities, yet real-world applications in educational and scientific domains remain challenging. This paper introduces the Multimodal Scientific ASR (MS-ASR) task, which focuses on transcribing scientific conference videos by leveraging visual information from slides to enhance the accuracy of technical terminologies. Realized that traditional metrics like WER fall short in assessing performance accurately, prompting the proposal of severity-aware WER (SWER) that considers the content type and severity of ASR errors. We propose the Scientific Vision Augmented ASR (SciVASR) framework as a baseline method, enabling MLLMs to improve transcript quality through post-editing. Evaluations of state-of-the-art MLLMs, including GPT-4o, show a 45% improvement over speech-only baselines, highlighting the importance of multimodal information integration.

Language models (LMs) can hallucinate when performing complex mathematical reasoning. Physics provides a rich domain for assessing their mathematical capabilities, where physical context requires that any symbolic manipulation satisfies complex semantics (e.g., units, tensorial order). In this work, we systematically remove crucial context from prompts to force instances where model inference may be algebraically coherent, yet unphysical. We assess LM capabilities in this domain using a curated dataset encompassing multiple notations and Physics subdomains. Further, we improve zero-shot scores using synthetic in-context examples, and demonstrate non-linear degradation of derivation quality with perturbation strength via the progressive omission of supporting premises. We find that the models’ mathematical reasoning is not physics-informed in this setting, where physical context is predominantly ignored in favour of reverse-engineering solutions.

Learning a model of a stochastic setting often involves learning both general structure rules and specific properties of the instance. This paper investigates the interplay between learning the general and the specific in various learning methods, with emphasis on sample efficiency. We design a framework called LEVERWORLDS, which allows the generation of simple physics-inspired worlds that follow a similar generative process with different distributions, and their instances can be expressed in natural language. These worlds allow for controlled experiments to assess the sample complexity of different learning methods. We experiment with classic learning algorithms as well as Transformer language models, both with fine-tuning and In-Context Learning (ICL). Our general finding is that (1) Transformers generally succeed in the task; but (2) they are considerably less sample efficient than classic methods that make stronger assumptions about the structure, such as Maximum Likelihood Estimation and Logistic Regression. This finding is in tension with the recent tendency to use Transformers as general-purpose estimators. We propose an approach that leverages the ICL capabilities of contemporary language models to apply simple algorithms for this type of data. Our experiments show that models currently struggle with the task but show promising potential.

Human cognition exhibits systematic compositionality, the algebraic ability to generate infinite novel combinations from finite learned components, which is the key to understanding and reasoning about complex logic. In this work, we investigate the compositionality of large language models (LLMs) in mathematical reasoning. Specifically, we construct a new dataset MathTrap by introducing carefully designed logical traps into the problem descriptions of MATH and GSM8K. Since problems with logical flaws are quite rare in the real world, these represent “unseen” cases to LLMs. Solving these requires the models to systematically compose (1) the mathematical knowledge involved in the original problems with (2) knowledge related to the introduced traps. Our experiments show that while LLMs possess both components of requisite knowledge, they do not spontaneously combine them to handle these novel cases. We explore several methods to mitigate this deficiency, such as natural language prompts, few-shot demonstrations, and fine-tuning. We find that LLMs’ performance can be improved through the above external intervention. Overall, systematic compositionality remains an open challenge for large language models.

Yonkoma Manga, characterized by its four-panel structure, presents unique challenges due to its rich contextual information and strong sequential features. To address the limitations of current multimodal large language models (MLLMs) in understanding this type of data, we create a novel dataset named YManga from the Internet. After filtering out low-quality content, we collect a dataset of 1,015 yonkoma strips, containing 10,150 human annotations. We then define three challenging tasks for this dataset: panel sequence detection, generation of the author’s creative intention, and description generation for masked panels. These tasks progressively introduce the complexity of understanding and utilizing such image-text data. To the best of our knowledge, YManga is the first dataset specifically designed for yonkoma manga strips understanding. Extensive experiments conducted on this dataset reveal significant challenges faced by current multimodal large language models. Our results show a substantial performance gap between models and humans across all three tasks.

Large Language Models (LLMs) are foundational in language technologies, particularly in information retrieval (IR). In this paper, we thoroughly explore the best practice of leveraging LLMs for query expansion. To this end, we introduce a training-free, straightforward yet effective framework called Multi-Text Generation Integration (MuGI). This approach leverages LLMs to generate multiple pseudo-references, which are then integrated with the original queries to enhance both sparse and dense retrieval methods. Additionally, we introduce a retrieval pipeline based on MuGI, which combines the strengths of sparse and dense retrievers to achieve superior performance without the need for costly pre-indexing. Our empirical findings reveal that: (1) Increasing the number of samples from LLMs benefits IR systems; (2) A balance between the query and pseudo-documents, and an effective integration strategy, is critical for high performance; (3) Contextual information from LLMs is essential, even boost a 23M model to outperform a 7B baseline model; (4) Pseudo relevance feedback can further calibrate queries for improved performance; and (5) Query expansion is widely applicable and versatile, consistently enhancing models ranging from 23M to 7B parameters. Our code and all generated references are made available at https://github.com/lezhang7/Retrieval_MuGI.

Multiple-choice visual question answering (VQA) is to automatically choose a correct answer from a set of choices after reading an image. Existing efforts have been devoted to a separate generation of an image-related question, a correct answer, or challenge distractors. By contrast, we turn to a holistic generation and optimization of questions, answers, and distractors (QADs) in this study. This integrated generation strategy eliminates the need for human curation and guarantees information consistency. Furthermore, we first propose to put the spotlight on different image regions to diversify QADs. Accordingly, a novel framework ReBo is formulated in this paper. ReBo cyclically generates each QAD based on a recurrent multimodal encoder, and each generation is focusing on a different area of the image compared to those already concerned by the previously generated QADs. In addition to traditional VQA comparisons with state-of-the-art approaches, we also validate the capability of ReBo in generating augmented data to benefit VQA models.

Extreme multi-label classification (XMC) aims to identify relevant subsets from numerous labels. Among the various approaches for XMC, tree-based linear models are effective due to their superior efficiency and simplicity. However, the space complexity of tree-based methods is not well-studied. Many past works assume that storing the model is not affordable and apply techniques such as pruning to save space, which may lead to performance loss. In this work, we conduct both theoretical and empirical analyses on the space to store a tree model under the assumption of sparse data, a condition frequently met in text data. We found that, some features may be unused when training binary classifiers in a tree method, resulting in zero values in the weight vectors. Hence, storing only non-zero elements can greatly save space. Our experimental results indicate that tree models can require less than 10% of the size of the standard one-vs-rest method for multi-label text classification. Our research provides a simple procedure to estimate the size of a tree model before training any classifier in the tree nodes. Then, if the model size is already acceptable, this approach can help avoid modifying the model through weight pruning or other techniques.