EMNLP 2024 Papers

Despite tremendous advancements, current state-of-the-art Vision-Language Models (VLMs) are still far from perfect. They tend to hallucinate and may generate biased responses. In such circumstances, having a way to assess the reliability of a given response generated by a VLM is quite useful. Existing methods, such as estimating uncertainty using answer likelihoods or prompt-based confidence generation, often suffer from overconfidence. Other methods use self-consistency comparison but are affected by confirmation biases. To alleviate these, we propose Decompose and Compare Consistency (DeCC) for reliability measurement. By comparing the consistency between the direct answer generated using the VLM’s internal reasoning process, and the indirect answers obtained by decomposing the question into sub-questions and reasoning over the sub-answers produced by the VLM, DeCC measures the reliability of VLM’s direct answer. Experiments across six vision-language tasks with three VLMs show DeCC’s reliability estimation achieves better correlation with task accuracy compared to the existing methods.

Susceptibility to misinformation describes the degree of belief in unverifiable claims, a latent aspect of individuals’ mental processes that is not observable. Existing susceptibility studies heavily rely on self-reported beliefs, which can be subject to bias, expensive to collect, and challenging to scale for downstream applications. To address these limitations, in this work, we propose a computational approach to efficiently model users’ latent susceptibility levels. As shown in previous work, susceptibility is influenced by various factors (e.g., demographic factors, political ideology), and directly influences people’s reposting behavior on social media. To represent the underlying mental process, our susceptibility modeling incorporates these factors as inputs, guided by the supervision of people’s sharing behavior. Using COVID-19 as a testbed, our experiments demonstrate a significant alignment between the susceptibility scores estimated by our computational modeling and human judgments, confirming the effectiveness of this latent modeling approach. Furthermore, we apply our model to annotate susceptibility scores on a large-scale dataset and analyze the relationships between susceptibility with various factors. Our analysis reveals that political leanings and other psychological factors exhibit varying degrees of association with susceptibility to COVID-19 misinformation, and shows that susceptibility is unevenly distributed across different professional and geographical backgrounds.

Adapting Large Language Models (LLMs) for recommendation requires careful consideration of the decoding process, given the inherent differences between generating items and natural language. Existing approaches often directly apply LLMs’ original decoding methods. However, we find these methods encounter significant challenges: 1) amplification bias—where standard length normalization inflates scores for items containing tokens with generation probabilities close to 1 (termed ghost tokens), and 2) homogeneity issue—generating multiple similar or repetitive items for a user. To tackle these challenges, we introduce a new decoding approach named Debiasing-Diversifying Decoding (D3). D3 disables length normalization for ghost tokens to alleviate amplification bias, and it incorporates a text-free assistant model to encourage tokens less frequently generated by LLMs for counteracting recommendation homogeneity. Extensive experiments on real-world datasets demonstrate the method’s effectiveness in enhancing accuracy and diversity.

Generative language models often struggle with specialized or less-discussed knowledge. A potential solution is found in Retrieval-Augmented Generation (RAG) models which act like retrieving information before generating responses. In this study, we explore how the Atlas approach, a RAG model, decides between what it already knows (parametric) and what it retrieves (non-parametric). We use causal mediation analysis and controlled experiments to examine how internal representations influence information processing. Our findings disentangle the effects of parametric knowledge and the retrieved context. They indicate that in cases where the model can choose between both types of information (parametric and non-parametric), it relies more on the context than the parametric knowledge. Furthermore, the analysis investigates the computations involved in how the model uses the information from the context. We find that multiple mechanisms are active within the model and can be detected with mediation analysis: first, the decision of whether the context is relevant, and second, how the encoder computes output representations to support copying when relevant.

Chain-of-Thought (CoT) prompting has been shown to enhance the multi-step reasoning capabilities of Large Language Models (LLMs). However, debates persist about whether LLMs exhibit *abstract generalization* or rely on *shallow heuristics* when given CoT prompts. To understand the factors influencing CoT reasoning we provide a detailed case study of the symbolic reasoning task of decoding shift ciphers, where letters are shifted forward some number of steps in the alphabet. We analyze the pattern of results produced by three LLMs—GPT-4, Claude 3, and Llama 3.1—performing this task using CoT prompting. By focusing on a single relatively simple task, we are able to identify three factors that systematically affect CoT performance: the probability of the task’s expected output (probability), what the model has implicitly learned during pre-training (memorization), and the number of intermediate operations involved in reasoning (noisy reasoning). We show that these factors can drastically influence task accuracy across all three LLMs; e.g., when tested with GPT-4, varying the output’s probability of occurrence shifts accuracy from 26% to 70%. Overall, we conclude that CoT prompting performance reflects both memorization and a probabilistic version of genuine reasoning.

Social networks are rife with noise and misleading information, presenting multifaceted challenges for rumor detection. In this paper, from the perspective of human cognitive subjectivity, we introduce the mining of individual latent intentions and propose a novel multi-task learning framework, the Intent-Aware Rumor Detection Network (IRDNet). IRDNet is designed to discern multi-level rumor semantic features and latent user intentions, addressing the challenges of robustness and key feature mining and alignment that plague existing models. In IRDNet, the multi-level semantic extraction module captures sequential and hierarchical features to generate robust semantic representations. The hierarchical contrastive learning module incorporates two complementary strategies, event-level and intent-level, to establish cognitive anchors that uncover the latent intentions of information disseminators. Event-level contrastive learning employs high-quality data augmentation and adversarial perturbations to enhance model robustness. Intent-level contrastive learning leverages the intent encoder to capture latent intent features and optimize consistency within the same intent while ensuring heterogeneity between different intents to clearly distinguish key features from irrelevant elements. Experimental results demonstrate that IRDNet significantly improves the effectiveness of rumor detection and effectively addresses the challenges present in the field of rumor detection.

Cognitive distortion research holds increasing significance as it sheds light on pervasive errors in thinking patterns, providing crucial insights into mental health challenges and fostering the development of targeted interventions and therapies. This paper delves into the complex domain of cognitive distortions which are prevalent distortions in cognitive processes often associated with mental health issues. Focusing on patient-doctor dialogues, we introduce a pioneering method for detecting and reasoning about cognitive distortions utilizing Large Language Models (LLMs). Operating within a multimodal context encompassing audio, video, and textual data, our approach underscores the critical importance of integrating diverse modalities for a comprehensive understanding of cognitive distortions. By leveraging multimodal information, including audio, video, and textual data, our method offers a nuanced perspective that enhances the accuracy and depth of cognitive distortion detection and reasoning in a zero-shot manner. Our proposed hierarchical framework adeptly tackles both detection and reasoning tasks, showcasing significant performance enhancements compared to current methodologies. Through comprehensive analysis, we elucidate the efficacy of our approach, offering promising insights into the diagnosis and understanding of cognitive distortions in multimodal settings.The code and dataset can be found here: https://github.com/clang1234/ZS-CoDR.git

Computational notebooks became indispensable tools for research-related development, offering unprecedented interactivity and flexibility in the development process. However, these benefits come at the cost of reproducibility and an increased potential for bugs.With the rise of code-fluent Large Language Models empowered with agentic techniques, smart bug-fixing tools with a high level of autonomy have emerged.However, those tools are tuned for classical script programming and still struggle with non-linear computational notebooks.In this paper, we present an AI agent designed specifically for error resolution in a computational notebook. We have developed an agentic system capable of exploring a notebook environment by interacting with it—similar to how a user would—and integrated the system into the JetBrains service for collaborative data science called Datalore.We evaluate our approach against the pre-existing single-action solution by comparing costs and conducting a user study. Users rate the error resolution capabilities of the agentic system higher but experience difficulties with UI. We share the results of the study and consider them valuable for further improving user-agent collaboration.

Increasing use of large language models (LLMs) demand performant guardrails to ensure the safety of inputs and outputs of LLMs. When these safeguards are trained on imbalanced data, they can learn the societal biases. We present a light-weight, post-processing method for mitigating counterfactual fairness in closed-source text safety classifiers. Our approach involves building an ensemble that not only outperforms the input classifiers and policy-aligns them, but also acts as a debiasing regularizer. We introduce two threshold-agnostic metrics to assess the counterfactual fairness of a model, and demonstrate how combining these metrics with Fair Data Reweighting (FDW) helps mitigate biases. We create an expanded Open AI dataset, and a new templated LLM-generated dataset based on user-prompts, both of which are counterfactually balanced across identity groups and cover four key areas of safety; we will work towards publicly releasing these datasets. Our results show that our approach improves counterfactual fairness with minimal impact on model performance.

We propose a multi-agent debate as optimization (DAO) system for event extraction, where the primary objective is to iteratively refine the large language models (LLMs) outputs through debating without parameter tuning. In DAO, we introduce two novel modules: the Diverse-RAG (DRAG) module and the Adaptive Conformal Prediction (AdaCP) module. DRAG systematically retrieves supporting information that best fits the debate discussion, while AdaCP enhances the accuracy and reliability of event extraction by effectively rejecting less promising answers. Experimental results demonstrate a significant reduction in the performance gap between supervised approaches and tuning-free LLM-based methods by 18.1% and 17.8% on ACE05 and 17.9% and 15.2% on CASIE for event detection and argument extraction respectively.

There is a mismatch between psychological and computational studies on emotions. Psychological research aims at explaining and documenting internal mechanisms of these phenomena, while computational work often simplifies them into labels. Many emotion fundamentals remain under-explored in natural language processing, particularly how emotions develop and how people cope with them. To help reduce this gap, we follow theories on coping, and treat emotions as strategies to cope with salient situations (i.e., how people deal with emotion-eliciting events). This approach allows us to investigate the link between emotions and behavior, which also emerges in language. We introduce the task of coping identification, together with a corpus to do so, constructed via role-playing. We find that coping strategies realize in text even though they are challenging to recognize, both for humans and automatic systems trained and prompted on the same task. We thus open up a promising research direction to enhance the capability of models to better capture emotion mechanisms from text.

The IIT-CDIP document collection is the source of several widely used and publicly accessible document understanding datasets. In this paper, manual inspection of 5 datasets derived from IIT-CDIP uncovers the presence of thousands of instances of sensitive personal data, including US Social Security Numbers (SSNs), birth places and dates, and home addresses of individuals. The presence of such sensitive personal data in commonly-used and publicly available datasets is startling and has ethical and potentially legal implications; we believe such sensitive data ought to be removed from the internet. Thus, in this paper, we develop a modular data de-identification pipeline that replaces sensitive data with synthetic, but realistic, data. Via experiments, we demonstrate that this de-identification method preserves the utility of the de-identified documents so that they can continue be used in various document understanding applications. We will release redacted versions of these datasets publicly.

Answer Sentence Selection (AS2) is a critical task for designing effective retrieval-based Question Answering (QA) systems. Most advancements in AS2 focus on English due to the scarcity of annotated datasets for other languages. This lack of resources prevents the training of effective AS2 models in different languages, creating a performance gap between QA systems in English and other locales. In this paper, we introduce new high-quality datasets for AS2 in five European languages (French, German, Italian, Portuguese, and Spanish), obtained through supervised Automatic Machine Translation (AMT) of existing English AS2 datasets such as ASNQ, WikiQA, and TREC-QA using a Large Language Model (LLM). We evaluated our approach and the quality of the translated datasets through multiple experiments with different Transformer architectures. The results indicate that our datasets are pivotal in producing robust and powerful multilingual AS2 models, significantly contributing to closing the performance gap between English and other languages.

We introduce DataTales, a novel benchmark designed to assess the proficiency of language models in data narration, a task crucial for transforming complex tabular data into accessible narratives. Existing benchmarks often fall short in capturing the requisite analytical complexity for practical applications. DataTales addresses this gap by offering 4.9k financial reports paired with corresponding market data, showcasing the demand for models to create clear narratives and analyze large datasets while understanding specialized terminology in the field. Our findings highlights the significant challenge that language models face in achieving the necessary precision and analytical depth for proficient data narration, suggesting promising avenues for future model development and evaluation methodologies.

Data-driven storytelling is a powerful method for conveying insights by combining narrative techniques with visualizations and text. These stories integrate visual aids, such as highlighted bars and lines in charts, along with textual annotations explaining insights. However, creating such stories requires a deep understanding of the data and meticulous narrative planning, often necessitating human intervention, which can be time-consuming and mentally taxing. While Large Language Models (LLMs) excel in various NLP tasks, their ability to generate coherent and comprehensive data stories remains underexplored. In this work, we introduce a novel task for data story generation and a benchmark containing 1,449 stories from diverse sources. To address the challenges of crafting coherent data stories, we propose a multi-agent framework employing two LLM agents designed to replicate the human storytelling process: one for understanding and describing the data (Reflection), generating the outline, and narration, and another for verification at each intermediary step. While our agentic framework generally outperforms non-agentic counterparts in both model-based and human evaluations, the results also reveal unique challenges in data story generation.

While coreference resolution is traditionally used as a component in individual document understanding, in this work we take a more global view and explore what can we learn about a domain from the set of all document-level coreference relations that are present in a large corpus. We derive coreference chains from a corpus of 30 million biomedical abstracts and construct a graph based on the string phrases within these chains, establishing connections between phrases if they co-occur within the same coreference chain. We then use the graph structure and the betweeness centrality measure to distinguish between edges denoting hierarchy, identity and noise, assign directionality to edges denoting hierarchy, and split nodes (strings) that correspond to multiple distinct concepts. The result is a rich, data-driven ontology over concepts in the biomedical domain, parts of which overlaps significantly with human-authored ontologies. We release the coreference chains and resulting ontology under a creative-commons license.

The impressive capabilities of recent language models can be largely attributed to the multi-trillion token pretraining datasets that they are trained on. However, model developers fail to disclose their construction methodology which has lead to a lack of open information on how to develop effective pretraining sets. To address this issue, we perform the first systematic study across the entire pipeline of pretraining set construction. First, we run ablations on existing techniques for pretraining set development to identify which methods translate to the largest gains in model accuracy on downstream evaluations. Then, we categorize the most widely used data source, web crawl snapshots, across the attributes of toxicity, quality, type of speech, and domain. Finally, we show how such attribute information can be used to further refine and improve the quality of a pretraining set. These findings constitute an actionable set of steps that practitioners can use to develop high quality pretraining sets.

Recent works have shown that by curating high quality and diverse instruction tuning datasets, we can significantly improve instruction-following capabilities. However, creating such datasets is difficult and most works rely on manual curation or proprietary language models. Automatic data curation is difficult as it is still not clear how we can define diversity for instruction tuning, how diversity and quality depend on one other, and how we can optimize dataset quality and diversity. To resolve these issue, we propose a new algorithm, Quality-Diversity Instruction Tuning (QDIT). QDIT provides a simple method to simultaneously control dataset diversity and quality, allowing us to conduct an in-depth study on the effect of diversity and quality on instruction tuning performance. From this study we draw two key insights (1) there is a natural tradeoff between data diversity and quality and (2) increasing data diversity significantly improves the worst case instruction following performance, therefore improving robustness. We validate the performance of QDIT on several large scale instruction tuning datasets, where we find it can substantially improve worst and average case performance compared to quality-driven data selection.

The opacity in developing large language models (LLMs) is raising growing concerns about the potential contamination of public benchmarks in the pre-training data. Existing contamination detection methods are typically based on the text overlap between training and evaluation data, which can be too superficial to reflect deeper forms of contamination. In this paper, we first present a cross-lingual form of contamination that inflates LLMs’ performance while evading current detection methods, deliberately injected by overfitting LLMs on the translated versions of benchmark test sets. Then, we propose generalization-based approaches to unmask such deeply concealed contamination. Specifically, we examine the LLM’s performance change after modifying the original benchmark by replacing the false answer choices with correct ones from other questions. Contaminated models can hardly generalize to such easier situations, where the false choices can be not even wrong, as all choices are correct in their memorization. Experimental results demonstrate that cross-lingual contamination can easily fool existing detection methods, but not ours. In addition, we discuss the potential utilization of cross-lingual contamination in interpreting LLMs’ working mechanisms and in post-training LLMs for enhanced multilingual capabilities.