NAACL 2024 论文列表

Retrieval Augmented Generation (RAG) is a prominent approach in real-word applications for grounding large language model (LLM) generations in up to date and domain-specific knowledge. However, there is a lack of systematic investigations of the impact of each component (retrieval quality, prompts, generation models) on the generation quality of a RAG pipeline in real world scenarios. In this study, we benchmark 6 LLMs in 15 retrieval scenarios exploring 9 prompts over 2 real world financial domain datasets. We thoroughly discuss the impact of each component in RAG pipeline on answer generation quality and formulate specific recommendations for the design of RAG systems.

Instruction-tuned Large Language Models (LLMs) have recently showcased remarkable ability to generate fitting responses to natural language instructions. However, an open research question concerns the inherent biases of trained models and their responses. For instance, if the data used to tune an LLM is dominantly written by persons with a specific political bias, we might expect generated answers to share this bias. Current research work seeks to de-bias such models, or suppress potentially biased answers.With this demonstration, we take a different view on biases in instruction-tuning: Rather than aiming to suppress them, we aim to make them explicit and transparent. To this end, we present OpinionGPT, a web demo in which users can ask questions and select all biases they wish to investigate. The demo will answer this question using a model fine-tuned on text representing each of the selected biases, allowing side-by-side comparison. To train the underlying model, we identified 11 different biases (political, geographic, gender, age) and derived an instruction-tuning corpus in which each answer was written by members of one of these demographics. This paper presents OpinionGPT, illustrates how we trained the bias-aware model and showcases the web application (available at https://opiniongpt.informatik.hu-berlin.de).

Object navigation (ObjectNav) requires an agent to navigate through unseen environments to find queried objects. Many previous methods attempted to solve this task by relying on supervised or reinforcement learning, where they are trained on limited household datasets with close-set objects. However, two key challenges are unsolved: understanding free-form natural language instructions that demand open-set objects, and generalizing to new environments in a zero-shot manner. Aiming to solve the two challenges, in this paper, we propose **OpenFMNav**, an **Open**-set **F**oundation **M**odel based framework for zero-shot object **Nav**igation. We first unleash the reasoning abilities of large language models (LLMs) to extract proposed objects from natural language instructions that meet the user’s demand. We then leverage the generalizability of large vision language models (VLMs) to actively discover and detect candidate objects from the scene, building a *Versatile Semantic Score Map (VSSM)*. Then, by conducting common sense reasoning on *VSSM*, our method can perform effective language-guided exploration and exploitation of the scene and finally reach the goal. By leveraging the reasoning and generalizing abilities of foundation models, our method can understand free-form human instructions and perform effective open-set zero-shot navigation in diverse environments. Extensive experiments on the HM3D ObjectNav benchmark show that our method surpasses all the strong baselines on all metrics, proving our method’s effectiveness. Furthermore, we perform real robot demonstrations to validate our method’s open-set-ness and generalizability to real-world environments.

Existing federated learning (FL) studies usuallyassume the training label space and test labelspace are identical. However, in real-world applications, this assumption is too ideal to betrue. A new user could come up with queriesthat involve data from unseen classes, and suchopen-vocabulary queries would directly defectsuch FL systems. Therefore, in this work, weexplicitly focus on the under-explored openvocabulary challenge in FL. That is, for a newuser, the global server shall understand her/hisquery that involves arbitrary unknown classes.To address this problem, we leverage the pretrained vision-language models (VLMs). Inparticular, we present a novel adaptation framework tailored for VLMs in the context of FL,named as Federated Multimodal Prototyping(Fed-MP). Fed-MP adaptively aggregates thelocal model weights based on light-weightclient residuals, and makes predictions basedon a novel multimodal prototyping mechanism.Fed-MP exploits the knowledge learned fromthe seen classes, and robustifies the adaptedVLM to unseen categories. Our empirical evaluation on various datasets validates the effectiveness of Fed-MP.

Despite their general capabilities, LLMs still struggle on biomedicalNER tasks, which are difficult due to the presence of specialized terminology and lack of training data. In this work we set out to improve LLM performance on biomedical NER in limited data settings via a new knowledge augmentation approach which incorporates definitions of relevant concepts on-the-fly. During this process, to provide a test bed for knowledge augmentation, we perform a comprehensive exploration of prompting strategies. Our experiments show that definition augmentation is useful for both open source and closed LLMs.For example, it leads to a relative improvement of 15% (on average) in GPT-4 performance (F1) across all (six) of our test datasets. We conduct extensive ablations and analyses to demonstrate that our performance improvements stem from adding relevant definitional knowledge. We find that careful prompting strategies also improve LLM performance, allowing them to outperform fine-tuned language models in few-shot settings. To facilitate future research in this direction, we release our code at https://github.com/allenai/beacon.

We propose on-the-fly ensembling of a neural machine translation (NMT) model with a large language model (LLM), prompted on the same task and input. Through experiments on 4 language directions with varying data amounts, we find that a slightly weaker-at-translation LLM can improve translations of a NMT model, and such an ensemble can produce better translations than ensembling two stronger NMT models.We demonstrate that our ensemble method can be combined with various techniques from LLM prompting, such as in context learning and translation context.

Politeness is a multifaceted concept influenced by individual perceptions of what is considered polite or impolite. With this objective, we introduce a novel task - Politeness Cause Elicitation and Intensity Tagging (PCEIT). This task focuses on conversations and aims to identify the underlying reasons behind the use of politeness and gauge the degree of politeness conveyed. To address this objective, we create HING-POEM, a new conversational dataset in Hinglish (a blend of Hindi and English) for mental health and legal counseling of crime victims. The rationale for the domain selection lies in the paramount importance of politeness in mental health and legal counseling of crime victims to ensure a compassionate and cordial atmosphere for them. We enrich the HING-POEM dataset by annotating it with politeness labels, politeness causal spans, and intensity values at the level of individual utterances. In the context of the introduced PCEIT task, we present PAANTH (Politeness CAuse ElicitAion and INtensity Tagging in Hinglish), a comprehensive framework based on Contextual Enhanced Attentive Convolution Transformer. We conduct extensive quantitative and qualitative evaluations to establish the effectiveness of our proposed approach using the newly constructed dataset. Our approach is compared against state-of-the-art baselines, and these analyses help demonstrate the superiority of our method.

Minimum Bayes-risk (MBR) decoding has recently gained renewed attention in text generation.MBR decoding considers texts sampled from a model as pseudo-references and selects the text with the highest similarity to the others.Therefore, sampling is one of the key elements of MBR decoding, and previous studies reported that the performance varies by sampling methods.From a theoretical standpoint, this performance variation is likely tied to how closely the samples approximate the true distribution of references.However, this approximation has not been the subject of in-depth study.In this study, we propose using anomaly detection to measure the degree of approximation.We first closely examine the performance variation and then show that previous hypotheses about samples do not correlate well with the variation, but our introduced anomaly scores do.The results are the first to empirically support the link between the performance and the core assumption of MBR decoding.

At the heart of the Pyramid evaluation method for text summarization lie human written summary content units (SCUs). These SCUs areconcise sentences that decompose a summary into small facts. Such SCUs can be used to judge the quality of a candidate summary, possibly partially automated via natural language inference (NLI) systems. Interestingly, with the aim to fully automate the Pyramid evaluation, Zhang and Bansal (2021) show that SCUs can be approximated by automatically generated semantic role triplets (STUs). However, several questions currently lack answers, in particular: i) Are there other ways of approximating SCUs that can offer advantages?ii) Under which conditions are SCUs (or their approximations) offering the most value? In this work, we examine two novel strategiesto approximate SCUs: generating SCU approximations from AMR meaning representations (SMUs) and from large language models (SGUs), respectively. We find that while STUs and SMUs are competitive, the best approximation quality is achieved by SGUs. We also show through a simple sentence-decomposition baseline (SSUs) that SCUs (and their approximations) offer the most value when rankingshort summaries, but may not help as much when ranking systems or longer summaries.

Current decoder-based pre-trained language models (PLMs) successfully demonstrate multilingual capabilities. However, it is unclear how these models handle multilingualism.We analyze the neuron-level internal behavior of multilingual decoder-based PLMs, Specifically examining the existence of neurons that fire “uniquely for each language” within decoder-only multilingual PLMs.We analyze six languages: English, German, French, Spanish, Chinese, and Japanese, and show that language-specific neurons are unique, with a slight overlap (< 5%) between languages. These neurons are mainly distributed in the models’ first and last few layers. This trend remains consistent across languages and models.Additionally, we tamper with less than 1% of the total neurons in each model during inference and demonstrate that tampering with a few language-specific neurons drastically changes the probability of target language occurrence in text generation.

Recent work on automated approaches to counterspeech have mostly focused on synthetic data but seldom look into how the public deals with abuse. While these systems identifying and generating counterspeech have the potential for abuse mitigation, it remains unclear how robust a model is against adversarial attacks across multiple domains and how models trained on synthetic data can handle unseen user-generated abusive content in the real world. To tackle these issues, this paper first explores the dynamics of abuse and replies using our novel dataset of 6,955 labelled tweets targeted at footballers for studying public figure abuse. We then curate DynaCounter, a new English dataset of 1,911 pairs of abuse and replies addressing nine minority identity groups, collected in an adversarial human-in-the-loop process over four rounds. Our analysis shows that adversarial attacks do not necessarily result in better generalisation. We further present a study of multi-domain counterspeech generation, comparing Flan-T5 and T5 models. We observe that handling certain abuse targets is particularly challenging.

Augmenting a language model (LM) with k-nearest neighbors (kNN) retrieval on its training data alone can decrease its perplexity, though the underlying reasons for this remain elusive. In this work, we rule out one previously posited possibility — the “softmax bottleneck.” We then create a new dataset to evaluate LM generalization ability in the setting where training data contains additional information that is not causally relevant. This task is challenging even for GPT-3.5 Turbo. We show that, for both GPT-2 and Mistral 7B, kNN retrieval augmentation consistently improves per formance in this setting. Finally, to make kNN retrieval more accessible, we propose using amulti-layer perceptron model that maps datastore keys to values as a drop-in replacement for traditional retrieval. This reduces storage costsby over 25x.

Narrative Question Answering is an important task for evaluating and improving reading comprehension abilities in both humans and machines. However, there is a lack of consensus on the skill taxonomy that would enable systematic and comprehensive assessment and learning of the various aspects of Narrative Question Answering. Existing task-level skill views oversimplify the multidimensional nature of tasks, while question-level taxonomies face issues in evaluation and methodology. To address these challenges, we introduce a more inclusive skill taxonomy that synthesizes and redefines narrative understanding skills from previous taxonomies and includes a generation skill dimension from the answering perspective.

Structured data, prevalent in tables, databases, and knowledge graphs, poses a significant challenge in its representation. With the advent of large language models (LLMs), there has been a shift towards linearization-based methods, which process structured data as sequential token streams, diverging from approaches that explicitly model structure, often as a graph. Crucially, there remains a gap in our understanding of how these linearization-based methods handle structured data, which is inherently non-linear.This work investigates the linear handling of structured data in encoder-decoder language models, specifically T5. Our findings reveal the model’s ability to mimic human-designed processes such as schema linking and syntax prediction, indicating a deep, meaningful learning of structure beyond simple token sequencing. We also uncover insights into the model’s internal mechanisms, including the ego-centric nature of structure node encodings and the potential for model compression due to modality fusion redundancy. Overall, this work sheds light on the inner workings of linearization-based methods and could potentially provide guidance for future research.

Recent studies have found that summaries generated by large language models (LLMs) are favored by human annotators over the original reference summaries in commonly used summarization datasets. Therefore, we study an LLM-as-reference learning setting for smaller text summarization models to investigate whether their performance can be substantially improved. To this end, we use LLMs as both oracle summary generators for standard supervised fine-tuning and oracle summary evaluators for efficient contrastive learning that leverages the LLMs’ supervision signals. We conduct comprehensive experiments with source news articles and find that (1) summarization models trained under the LLM-as-reference setting achieve significant performance improvement in both LLM and human evaluations; (2) contrastive learning outperforms standard supervised fine-tuning under both low and high resource settings. Our experimental results also enable a meta-analysis of LLMs’ summary evaluation capacities under a challenging setting, showing that LLMs are not well-aligned with human evaluators. Particularly, our expert human evaluation reveals remaining nuanced performance gaps between LLMs and our fine-tuned models, which LLMs fail to capture. Thus, we call for further studies into both the potential and challenges of using LLMs in summarization model development.

Large language models are successful in answering factoid questions but are also prone to hallucination.We investigate the phenomenon of LLMs possessing correct answer knowledge yet still hallucinating from the perspective of inference dynamics, an area not previously covered in studies on hallucinations.We are able to conduct this analysis via two key ideas.First, we identify the factual questions that query the same triplet knowledge but result in different answers. The difference between the model behaviors on the correct and incorrect outputs hence suggests the patterns when hallucinations happen.Second, to measure the pattern, we utilize mappings from the residual streams to vocabulary space.We reveal the different dynamics of the output token probabilities along the depths of layers between the correct and hallucinated cases. In hallucinated cases, the output token’s information rarely demonstrates abrupt increases and consistent superiority in the later stages of the model.Leveraging the dynamic curve as a feature, we build a classifier capable of accurately detecting hallucinatory predictions with an 88% success rate. Our study shed light on understanding the reasons for LLMs’ hallucinations on their known facts, and more importantly, on accurately predicting when they are hallucinating.

This study introduces a new long-form database question answering dataset designed to evaluate how Large Language Models (LLMs) interact with a SQL interpreter. The task necessitates LLMs to strategically generate multiple SQL queries to retrieve sufficient data from a database, to reason with the acquired context, and to synthesize them into a comprehensive analytical narrative. Our findings highlight that this task poses great challenges even for the state-of-the-art **GPT-4** model. We propose and evaluate two interaction strategies, and provide a fine-grained analysis of the individual stages within the interaction. A key discovery is the identification of two primary bottlenecks hindering effective interaction: the capacity for planning and the ability to generate multiple SQL queries. To address the challenge of accurately assessing answer quality, we introduce a multi-agent evaluation framework that simulates the academic peer-review process, enhancing the precision and reliability of our evaluations. This framework allows for a more nuanced understanding of the strengths and limitations of current LLMs in complex retrieval and reasoning tasks.

In NLP, Event Coreference Resolution (ECR) is the task of connecting event clusters that refer to the same underlying real-life event, usually via neural systems. In this work, we investigate using abductive free-text rationales (FTRs) generated by modern autoregressive LLMs as distant supervision of smaller student models for cross-document coreference (CDCR) of events. We implement novel rationale-oriented event clustering and knowledge distillation methods for event coreference scoring that leverage enriched information from the FTRs for improved CDCR without additional annotation or expensive document clustering. Our model using coreference-specific knowledge distillation achieves SOTA B3 F1 on the ECB+ and GVC corpora and we establish a new baseline on the AIDA Phase 1 corpus. Our code can be found at https://github.com/csu-signal/llama_cdcr.

The capability of intelligent models to extrapolate and comprehend changes in object states is a crucial yet demanding aspect of AI research, particularly through the lens of human interaction in real-world settings. This task involves describing complex visual environments, identifying active objects, and interpreting their changes as conveyed through language. Traditional methods, which isolate object captioning and state change detection, offer a limited view of dynamic environments. Moreover, relying on a small set of symbolic words to represent changes has restricted the expressiveness of language. To address these challenges, in this paper, we introduce the Object State Captioning and State Change Representation (OSCaR) dataset and benchmark. OSCaR consists of 14,084 annotated video segments with nearly 1,000 unique objects from various egocentric video collections. It sets a new testbed for evaluating Multimodal Large Language Models (MLLMs). Our experiments demonstrate that while MLLMs show some skill, they lack a full understanding of object state changes. The benchmark includes a fine-tuned model that, despite initial capabilities, requires significant improvements in accuracy and generalization ability for effective understanding of these changes. Our code and dataset are available at https://github.com/nguyennm1024/OSCaR.