EMNLP 2024 Papers

This paper studies the problem of information retrieval, to adapt to unseen tasks. Existing work generates synthetic queries from domain-specific documents to jointly train the retriever. However, the conventional query generator assumes the query as a question, thus failing to accommodate general search intents. A more lenient approach incorporates task-adaptive elements, such as few-shot learning with an 137B LLM. In this paper, we challenge a trend equating query and question, and instead conceptualize query generation task as a “compilation” of high-level intent into task-adaptive query. Specifically, we propose EGG, a query generator that better adapts to wide search intents expressed in the BeIR benchmark. Our method outperforms baselines and existing models on four tasks with underexplored intents, while utilizing a query generator 47 times smaller than the previous state-of-the-art. Our findings reveal that instructing the LM with explicit search intent is a key aspect of modeling an effective query generator.

Pretrained language models (LMs) encode implicit representations of knowledge in their parameters. However, localizing these representations and disentangling them from each other remains an open problem. In this work, we investigate whether pretrained language models contain various *knowledge-critical* subnetworks: particular sparse computational subgraphs that can, if removed, precisely suppress specific knowledge the model has memorized. We propose a multi-objective differentiable masking scheme that can be applied to both weights and neurons to discover such subnetworks and show that we can use them to precisely remove specific knowledge from models while minimizing adverse effects on the behavior of the original model. We demonstrate our method on multiple GPT2 variants, uncovering highly sparse subnetworks (98%+ sparsity) that are critical for expressing specific collections of relational knowledge. When these subnetworks are removed, the remaining network maintains most of its initial abilities but struggles to represent the suppressed knowledge.

Most of the efforts in interpreting neural relevance models have been on local explanations, which explain the relevance of a document to a query. However, local explanations are not effective in predicting the model’s behavior on unseen texts. We aim at explaining a neural relevance model by providing lexical explanations that can be globally generalized. Specifically, we construct a relevance thesaurus containing semantically relevant query term and document term pairs, which can augment BM25 scoring functions to better approximate the neural model’s predictions. We propose a novel method to build a relevance thesaurus construction. Our method involves training a neural relevance model which can score the relevance for partial segments of query and documents. The trained model is used to identify relevant terms over the vocabulary space. The resulting thesaurus explanation is evaluated based on ranking effectiveness and fidelity to the targeted neural ranking model. Finally, our thesaurus reveals the existence of brand name bias in ranking models, which further supports the utility of our explanation method.

Recent advancements in Natural Language Processing (NLP) have impacted numerous sub-fields such as natural language generation, natural language inference, question answering, and more. However, in the field of question generation, the creation of distractors for multiple-choice questions (MCQ) remains a challenging task. In this work, we present a simple, generic framework for distractor generation using readily available Pre-trained Language Models (PLMs). Unlike previous methods, our framework relies solely on pre-trained language models and does not require additional training on specific datasets. Building upon previous research, we introduce a two-stage framework consisting of candidate generation and candidate selection. Our proposed distractor generation framework outperforms previous methods without the need for training or fine-tuning. Human evaluations confirm that our approach produces more effective and engaging distractors. The related codebase is publicly available at https://github.com/obss/disgem.

Adapting Large Language Models (LLMs) for agent tasks is critical in developing language agents. Direct Preference Optimization (DPO) is a promising technique for this adaptation with the alleviation of compounding errors, offering a means to directly optimize Reinforcement Learning (RL) objectives. However, applying DPO to multi-turn tasks presents challenges due to the inability to cancel the partition function. Overcoming this obstacle involves making the partition function independent of the current state and addressing length disparities between preferred and dis-preferred trajectories. In this light, we replace the policy constraint with the state-action occupancy measure constraint in the RL objective and add length normalization to the Bradley-Terry model, yielding a novel loss function named DMPO for multi-turn agent tasks with theoretical explanations. Extensive experiments on three multi-turn agent task datasets confirm the effectiveness and superiority of the DMPO loss.

Multimodal Large Language Models (MLLMs) demonstrate a strong understanding of the real world and can even handle complex tasks. However, they still fail on some straightforward visual question-answering (VQA) problems. This paper dives deeper into this issue, revealing that models tend to err when answering easy questions (e.g., Yes/No questions) about an image, even though they can correctly describe it.We refer to this model behavior discrepancy between difficult and simple questions as model laziness.To systematically investigate model laziness, we manually construct LazyBench, a benchmark that includes Yes/No, multiple choice, short answer questions, and image description tasks that are related to the same subjects in the images.Based on LazyBench. we observe that laziness widely exists in current advanced MLLMs (e.g., GPT-4o, Gemini-1.5-pro, Claude 3, LLaVA-1.5, LLaVA-1.6, and QWen-VL). We also analyzed the failure cases of LLaVA-1.5-13B on the VQA-v2 benchmark and discovered that about half of these failures are due to the model’s laziness. This further highlights the importance of ensuring that the model fully utilizes its capability.To this end, we conduct a preliminary exploration of how to mitigate laziness and find that chain of thought can effectively avoid this issue. The data can be accessed at https://github.com/Akutagawa1998/LazyBench.

Domain-specific neural machine translation (NMT) systems (, in educational applications) are socially significant with the potential to help make information accessible to a diverse set of users in multilingual societies. Such NMT systems should be lexically constrained and draw from domain-specific dictionaries. Dictionaries could present multiple candidate translations for a source word/phrase due to the polysemous nature of words. The onus is then on the NMT model to choose the contextually most appropriate candidate. Prior work has largely ignored this problem and focused on the single candidate constraint setting wherein the target word or phrase is replaced by a single constraint. In this work, we present DictDis, a lexically constrained NMT system that disambiguates between multiple candidate translations derived from dictionaries. We achieve this by augmenting training data with multiple dictionary candidates to actively encourage disambiguation during training by implicitly aligning multiple candidate constraints. We demonstrate the utility of DictDis via extensive experiments on English-Hindi, English-German, and English-French datasets across a variety of domains including regulatory, finance, engineering, health and standard benchmark test datasets. In comparison with existing approaches for lexically constrained and unconstrained NMT, we demonstrate superior performance for the copy constraint and disambiguation-related measures on all domains, while also obtaining improved fluency of up to 2-3 BLEU points on some domains. We also release our test set consisting of 4K English-Hindi sentences in multiple domains.

Efficiently deriving structured workflows from unannotated dialogs remains an underexplored and formidable challenge in computational linguistics. Automating this process could significantly accelerate the manual design of workflows in new domains and enable the grounding of large language models in domain-specific flowcharts, enhancing transparency and controllability.In this paper, we introduce Dialog2Flow (D2F) embeddings, which differ from conventional sentence embeddings by mapping utterances to a latent space where they are grouped according to their communicative and informative functions (i.e., the actions they represent). D2F allows for modeling dialogs as continuous trajectories in a latent space with distinct action-related regions. By clustering D2F embeddings, the latent space is quantized, and dialogs can be converted into sequences of region/action IDs, facilitating the extraction of the underlying workflow.To pre-train D2F, we build a comprehensive dataset by unifying twenty task-oriented dialog datasets with normalized per-turn action annotations. We also introduce a novel soft contrastive loss that leverages the semantic information of these actions to guide the representation learning process, showing superior performance compared to standard supervised contrastive loss.Evaluation against various sentence embeddings, including dialog-specific ones, demonstrates that D2F yields superior qualitative and quantitative results across diverse domains.

Factual faithfulness is a crucial requirement in information-seeking dialogue: the system should respond to the user queries so that the responses are meaningful and aligned with the knowledge provided to the system. However, most modern large language models (LLMs) suffer from hallucinations, that is, they generate responses not supported by or even contradicting the knowledge source. To mitigate the issue and increase faithfulness of information-seeking dialogue systems supported by the LLMs, we introduce BeInfo, a simple yet effective method that applies ‘behavioural tuning’ on the LLMs to aid information-seeking dialogue. Relying on three standard information seeking dialogue datasets, we show that models tuned with BeInfo become considerably more faithful to the knowledge source both for datasets and domains seen during BeInfo-tuning, as well as on unseen domains, when applied in a zero-shot manner. In addition, we present a ‘real-life’ case study on conversations with real users, showcasing that the models with 3B parameters (e.g., Flan-T5) tuned with BeInfo demonstrate strong performance on data from real ‘production’ conversations: when tuned on a limited amount of such realistic in-domain dialogues, they surpass much larger LLMs used ‘off-the-shelf’, both on automatic and human evaluation metrics.

Though large language models (LLMs) achieve significant success in recent years, the hallucination issue remains a challenge, and numerous benchmarks are proposed for hallucination detection. Nevertheless, some of these benchmarks are not naturally generated by LLMs but are intentionally induced. Also, many merely focus on the factuality hallucination while ignoring the faithfulness hallucination. Additionally, although dialogue pattern is more widely utilized in the era of LLMs, current benchmarks only concentrate on sentence-level and passage-level hallucination. In this study, we propose DiaHalu, the first dedicated dialogue-level hallucination evaluation benchmark for LLMs to our knowledge. Initially, we integrate the collected topics into system prompts and facilitate a dialogue between two LLMs. Subsequently, we manually modify the contents that do not adhere to human language conventions and then have LLMs re-generate, simulating authentic human-machine interaction scenarios. Finally, professional scholars annotate all the samples in the dataset. DiaHalu covers four common multi-turn dialogue domains and five hallucination subtypes, extended from factuality and faithfulness hallucination. Experiments through some well-known LLMs and detection methods on the dataset show that DiaHalu is a challenging benchmark, holding significant value for further research.

High-quality distractors are crucial to both the assessment and pedagogical value of multiple-choice questions (MCQs), where manually crafting ones that anticipate knowledge deficiencies or misconceptions among real students is difficult. Meanwhile, automated distractor generation, even with the help of large language models (LLMs), remains challenging for subjects like math. It is crucial to not only identify plausible distractors but also understand the error behind them. In this paper, we introduce DiVERT (Distractor Generation with Variational Errors Represented as Text), a novel variational approach that learns an interpretable representation of errors behind distractors in math MCQs. Through experiments on a real-world math MCQ dataset with 1,434 questions used by hundreds of thousands of students, we show that DiVERT, despite using a base open-source LLM with 7B parameters, outperforms state-of-the-art approaches using GPT-4o on downstream distractor generation. We also conduct a human evaluation with math educators and find that DiVERT leads to error labels that are of comparable quality to human-authored ones.

Query Auto-Complete (QAC) is an essential search feature that suggests users with a list of potential search keyword completions as they type, enabling them to complete their queries faster. While the QAC systems in eCommerce stores generally use the Learning to Rank (LTR) approach optimized based on customer feedback, it struggles to provide diverse suggestions, leading to repetitive queries and limited navigational suggestions related to product categories, attributes, and brands. This paper proposes a novel DiAL framework that explicitly optimizes for diversity alongside customer feedback signals. It achieves this by leveraging a smooth approximation of the diversity-based metric (𝛼NDCG) as a listwise loss function and modifying it to balance relevance and diversity. The proposed approach yielded an improvement of 8.5% in mean reciprocal rank (MRR) and 22.8% in 𝛼NDCG compared to the pairwise ranking approach on an eCommerce dataset, while meeting the ultra-low latency constraints of QAC systems. In an online experiment, the diversity-aware listwise QAC model resulted in a 0.48% lift in revenue. Furthermore, we replicated the proposed approach on a publicly available search log, demonstrating improvements in both diversity and relevance of the suggested queries.

In this work, we introduce a novel approach that equips LLM agents with introspection, enhancing consistency and adaptability in solving complex tasks. Our approach prompts LLM agents to decompose a given task into manageable subtasks (i.e., to make a plan), and to continuously introspect upon the suitability and results of their actions. We implement a three-fold introspective intervention: 1) anticipatory reflection on potential failures and alternative remedy before action execution, 2) post-action alignment with subtask objectives and backtracking with remedy to ensure utmost effort in plan execution, and 3) comprehensive review upon plan completion for future strategy refinement. By deploying and experimenting with this methodology—a zero-shot approach—within WebArena for practical tasks in web environments, our agent demonstrates superior performance with a success rate of 23.5% over existing zero-shot methods by 3.5%. The experimental results suggest that our introspection-driven approach not only enhances the agent’s ability to navigate unanticipated challenges through a robust mechanism of plan execution, but also improves efficiency by reducing the number of trials and plan revisions by 45% needed to achieve a task.

Recent studies show evidence for emergent cognitive abilities in Large Pre-trained Language Models (PLMs). The increasing cognitive alignment of these models has made them candidates for cognitive science theories. Prior research into the emergent cognitive abilities of PLMs has been path-independent to model training, i.e. has only looked at the final model weights and not the intermediate steps. However, building plausible models of human cognition using PLMs also requires aligning their performance during training to the developmental trajectories of children’s thinking. Guided by psychometric tests of human intelligence, we choose four task categories to investigate the alignment of ten popular families of PLMs and evaluate each of their available intermediate and final training steps: Numerical ability, Linguistic abilities, Conceptual understanding, and Fluid reasoning. We find a striking regularity: regardless of model size, the developmental trajectories of PLMs consistently exhibit a window of maximal alignment to human cognitive development. Before that window, training appears to endow models with the requisite structure to be poised to rapidly learn from experience. After that window, training appears to serve the engineering goal of reducing loss but not the scientific goal of increasing alignment with human cognition.

Large language models (LLMs) have demonstrated remarkable performance in the legal domain, with GPT-4 even passing the Uniform Bar Exam in the U.S. However their efficacy remains limited for non-standardized tasks and tasks in languages other than English. This underscores the need for careful evaluation of LLMs within each legal system before application.Here, we introduce KBL, a benchmark for assessing the Korean legal language understanding of LLMs, consisting of (1) 7 legal knowledge tasks (510 examples), (2) 4 legal reasoning tasks (288 examples), and (3) the Korean bar exam (4 domains, 53 tasks, 2,510 examples). First two datasets were developed in close collaboration with lawyers to evaluate LLMs in practical scenarios in a certified manner. Furthermore, considering legal practitioners’ frequent use of extensive legal documents for research, we assess LLMs in both a closed book setting, where they rely solely on internal knowledge, and a retrieval-augmented generation (RAG) setting, using a corpus of Korean statutes and precedents. The results indicate substantial room and opportunities for improvement.

Prior works on detoxification are scattered in the sense that they do not cover all aspects of detoxification needed in a real-world scenario. Notably, prior works restrict the task of developing detoxification models to only a seen subset of platforms, leaving the question of how the models would perform on unseen platforms unexplored. Additionally, these works do not address non-detoxifiability, a phenomenon whereby the toxic text cannot be detoxified without altering the meaning. We propose DetoxLLM, the first comprehensive end-to-end detoxification framework, which attempts to alleviate the aforementioned limitations. We first introduce a cross-platform pseudo-parallel corpus applying multi-step data processing and generation strategies leveraging ChatGPT. We then train a suite of detoxification models with our cross-platform corpus. We show that our detoxification models outperform the SoTA model trained with human-annotated parallel corpus. We further introduce explanation to promote transparency and trustworthiness. DetoxLLM additionally offers a unique paraphrase detector especially dedicated for the detoxification task to tackle the non-detoxifiable cases. Through experimental analysis, we demonstrate the effectiveness of our cross-platform corpus and the robustness of DetoxLLM against adversarial toxicity.

Emotion Recognition in conversations (ERC) involves an internal cognitive process that interprets emotional cues by using a collection of past emotional experiences. However, many existing methods struggle to decipher emotional cues in dialogues since they are insufficient in understanding the rich historical emotional context. In this work, we introduce an innovative Detective Network (DetectiveNN), a novel model that is grounded in the cognitive theory of emotion and utilizes a “recall-detect-predict” framework to imitate human emotional reasoning. This process begins by ‘recalling’ past interactions of a specific speaker to collect emotional cues. It then ‘detects’ relevant emotional patterns by interpreting these cues in the context of the ongoing conversation. Finally, it ‘predicts’ the speaker’s current emotional state. Tested on three benchmark datasets, our approach significantly outperforms existing methods. This highlights the advantages of incorporating cognitive factors into deep learning for ERC, enhancing task efficacy and prediction accuracy.

The diversity of text can be measured beyond word-level features, however existing diversity evaluation focuses primarily on word-level features. Here we propose a method for evaluating diversity over syntactic features to characterize general repetition in models, beyond frequent n-grams. Specifically, we define syntactic templates (e.g., strings comprising parts-of-speech) and show that models tend to produce templated text in downstream tasks at a higher rate than what is found in human-reference textsWe find that most (76%) templates in model-generated text can be found in pre-training data (compared to only 35% of human-authored text), and are not overwritten during fine-tuning or alignment processes such as RLHF. The connection between templates in generated text and the pre-training data allows us to analyze syntactic templates in models where we do not have the pre-training data.We also find that templates as features are able to differentiate between models, tasks, and domains, and are useful for qualitatively evaluating common model constructions.Finally, we demonstrate the use of templates as a useful tool for analyzing style memorization of training data in LLMs.

Detecting and answering ambiguous questions has been a challenging task in open-domain question answering. Ambiguous questions have different answers depending on their interpretation and can take diverse forms. Temporally ambiguous questions are one of the most common types of such questions. In this paper, we introduce TEMPAMBIQA, a manually annotated temporally ambiguous QA dataset consisting of 8,162 open-domain questions derived from existing datasets. Our annotations focus on capturing temporal ambiguity to study the task of detecting temporally ambiguous questions. We propose a novel approach by using diverse search strategies based on disambiguate versions of the questions. We also introduce and test non-search, competitive baselines for detecting temporal ambiguity using zero-shot and few-shot approaches.