NAACL 2024 论文列表

Existing open-source helpfulness preference datasets do not specify what makes some responses more helpful and others less so. Models trained on these datasets can incidentally learn to model dataset artifacts (e.g. preferring longer but unhelpful responses only due to their length). To alleviate this problem, we collect HelpSteer, a multi-attribute helpfulness dataset annotated for the various aspects that make responses helpful. Specifically, our 37k-sample dataset has annotations for correctness, coherence, complexity, and verbosity in addition to overall helpfulness of responses. Training Llama 2 70B using the HelpSteer dataset with SteerLM technique produces a model that scores 7.54 on MT Bench, which is currently the highest score for open models that do not require training data from more powerful models (e.g. GPT-4). We release this dataset with CC-BY-4.0 license at https://huggingface.co/datasets/nvidia/HelpSteer

Since the recent prosperity of Large Language Models (LLMs), there have been interleaved discussions regarding how to reduce hallucinations from LLM responses, how to increase the factuality of LLMs, and whether Knowledge Graphs (KGs), which store the world knowledge in a symbolic form, will be replaced with LLMs. In this paper, we try to answer these questions from a new angle: How knowledgeable are LLMs?To answer this question, we constructed Head-to-Tail, a benchmark that consists of 18K question-answer (QA) pairs regarding head, torso, and tail facts in terms of popularity. We designed an automated evaluation method and a set of metrics that closely approximate the knowledge an LLM confidently internalizes. Through a comprehensive evaluation of 16 publicly available LLMs, we show that existing LLMs are still far from being perfect in terms of their grasp of factual knowledge, especially for facts of torso-to-tail entities.

To protect users from massive hateful content, existing works studied automated hate speech detection. Despite the existing efforts, one question remains: Do automated hate speech detectors conform to social media content policies? A platform’s content policies are a checklist of content moderated by the social media platform. Because content moderation rules are often uniquely defined, existing hate speech datasets cannot directly answer this question. This work seeks to answer this question by creating HateModerate, a dataset for testing the behaviors of automated content moderators against content policies. First, we engage 28 annotators and GPT in a six-step annotation process, resulting in a list of hateful and non-hateful test suites matching each of Facebook’s 41 hate speech policies. Second, we test the performance of state-of-the-art hate speech detectors against HateModerate, revealing substantial failures these models have in their conformity to the policies. Third, using HateModerate, we augment the training data of a top-downloaded hate detector on HuggingFace. We observe significant improvement in the models’ conformity to content policies while having comparable scores on the original test data. Our dataset and code can be found on https://github.com/stevens-textmining/HateModerate.

Large Language Models (LLMs) have shown propensity to generate hallucinated outputs, i.e., texts that are factually incorrect or unsupported. Existing methods for alleviating hallucinations typically require costly human annotations to identify and correct hallucinations in LLM outputs. Moreover, most of these methods focus on a specific type of hallucination, e.g., entity or token errors, which limits their effectiveness in addressing various types of hallucinations exhibited in LLM outputs. To our best knowledge, in this paper we propose the first active learning framework to alleviate LLM hallucinations, reducing costly human annotations of hallucination needed. By measuring fine-grained hallucinations from errors in semantic frame, discourse and content verifiability in text summarization, we propose HAllucination Diversity-Aware Sampling (HADAS) to select diverse hallucinations for annotations in active learning for LLM finetuning. Extensive experiments on three datasets and different backbone models demonstrate advantages of our method in effectively and efficiently mitigating LLM hallucinations.

Temporal knowledge graphs (TKGs) serve as powerful tools for storing and modeling dynamic facts, holding immense potential in anticipating future facts. Since future facts are inherently unknowable, effectively modeling the intricate temporal structure of historical facts becomes paramount for accurate prediction. However, current models often rely heavily on fact recurrence or periodicity, leading to information loss due to prolonged evolutionary processes. Notably, the occurrence of one fact always influences the likelihood of another. To this end, we propose HTCCN, a novel Hawkes process-based temporal causal convolutional network designed for temporal reasoning under extrapolation settings. HTCCN employs a temporal causal convolutional network to model the historical interdependence of facts and leverages Hawkes to model link formation processes inductively in TKGs. Importantly, HTCCN introduces dual-level dynamics to comprehensively capture the temporal evolution of facts. Rigorous experimentation on four real-world datasets underscores the superior performance of HTCCN.

Micro-enterprises and individual developers emerge analysis demands for long sequence with powerful Large Language Models (LLMs). They try to deploy the LLMs at local, but only possess various commodity devices and the unreliable interconnection between devices. Existing parallel techniques do not lead to the same effectiveness in limited environment. The heterogeneity of devices, coupled with their limited capacity and expensive communication, brings challenges to private deployment for maximized utilization of available devices while masking latency. Hence, we introduce HPipe, a pipeline inference framework that successfully mitigates LLMs from high-performance clusters to heterogeneous commodity devices. By ensuring a balanced distribution of workloads, HPipe facilitates the parallel execution of LLMs through pipelining the sequences on the token dimension. The evaluation conducted on LLaMA-7B and GPT3-2B demonstrates that HPipe holds the potential for context analysis on LLM with heterogeneity devices, achieving an impressive speedup in latency and throughput up to 2.28 times.

Existing self-supervised methods in natural language processing (NLP), especially hierarchical text classification (HTC), mainly focus on self-supervised contrastive learning, extremely relying on human-designed augmentation rules to generate contrastive samples, which can potentially corrupt or distort the original information. In this paper, we tend to investigate the feasibility of a contrastive learning scheme in which the semantic and syntactic information inherent in the input sample is adequately reserved in the contrastive samples and fused during the learning process. Specifically, we propose an information lossless contrastive learning strategy for HTC, namely Hierarchy-aware Information Lossless contrastive Learning (HILL), which consists of a text encoder representing the input document, and a structure encoder directly generating the positive sample. The structure encoder takes the document embedding as input, extracts the essential syntactic information inherent in the label hierarchy with the principle of structural entropy minimization, and injects the syntactic information into the text representation via hierarchical representation learning. Experiments on three common datasets are conducted to verify the superiority of HILL.

Advancements in dense retrieval models have brought ColBERT to prominence in Information Retrieval (IR) with its advanced interaction techniques.However, ColBERT is reported to frequently underperform in zero-shot scenarios, where traditional techniques such as BM25 still exceed it.Addressing this, we propose to balance representation isotropy and anisotropy for zero-shot model performance, based on our observations that isotropy can enhance cosine similarity computations and anisotropy may aid in generalizing to unseen data.Striking a balance between these isotropic and anisotropic qualities stands as a critical objective to refine model efficacy.Based on this, we present ours, a Hybrid Isotropy Learning (HIL) architecture that integrates isotropic and anisotropic representations.Our experiments with the BEIR benchmark show that our model significantly outperforms the baseline ColBERT model, highlighting the importance of harmonized isotropy in improving zero-shot retrieval performance.

While major languages often enjoy substantial attention and resources, the linguistic diversity across the globe encompasses a multitude of smaller, indigenous, and regional languages that lack the same level of computational support. One such region is the Caribbean. While commonly labeled as “English speaking”, the ex-British Caribbean region consists of a myriad of Creole languages thriving alongside English. In this paper, we present Guylingo: a comprehensive corpus designed for advancing NLP research in the domain of Creolese (Guyanese English-lexicon Creole), the most widely spoken language in the culturally rich nation of Guyana. We first outline our framework for gathering and digitizing this diverse corpus, inclusive of colloquial expressions, idioms, and regional variations in a low-resource language. We then demonstrate the challenges of training and evaluating NLP models for machine translation for Creolese. Lastly, we discuss the unique opportunities presented by recent NLP advancements for accelerating the formal adoption of Creole languages as official languages in the Caribbean.

Machine Translation (MT) remains one of the last NLP tasks where large language models (LLMs) have not yet replaced dedicated supervised systems. This work exploits the complementary strengths of LLMs and supervised MT by guiding LLMs to automatically post-edit MT with external feedback on its quality, derived from Multidimensional Quality Metric (MQM) annotations. Working with LLaMA-2 models, we consider prompting strategies varying the nature of feedback provided and then fine-tune the LLM to improve its ability to exploit the provided guidance. Through experiments on Chinese-English, English-German, and English-Russian MQM data, we demonstrate that prompting LLMs to post-edit MT improves TER, BLEU and COMET scores, although the benefits of fine-grained feedback are not clear. Fine-tuning helps integrate fine-grained feedback more effectively and further improves translation quality based on both automatic and human evaluation.

We evaluate the performance disparity of the Whisper and MMS families of ASR models across the VoxPopuli and Common Voice multilingual datasets, with an eye toward intersectionality. Our two most important findings are that model size, surprisingly, correlates logarithmically with worst-case performance disparities, meaning that larger (and better) models are less fair. We also observe the importance of intersectionality. In particular, models often exhibit significant performance disparity across binary gender for adolescents.

Effective conversation requires common ground: a shared understanding between the participants. Common ground, however, does not emerge spontaneously in conversation. Speakers and listeners work together to both identify and construct a shared basis while avoiding misunderstanding. To accomplish grounding, humans rely on a range of dialogue acts, like clarification (What do you mean?) and acknowledgment (I understand.). However, it is unclear whether large language models (LLMs) generate text that reflects human grounding. To this end, we curate a set of grounding acts and propose corresponding metrics that quantify attempted grounding. We study whether LLM generations contain grounding acts, simulating turn-taking from several dialogue datasets and comparing results to humans. We find that—compared to humans—LLMs generate language with less conversational grounding, instead generating text that appears to simply presume common ground. To understand the roots of the identified grounding gap, we examine the role of instruction tuning and preference optimization, finding that training on contemporary preference data leads to a reduction in generated grounding acts. Altogether, we highlight the need for more research investigating conversational grounding in human-AI interaction.

We present an empirical study of groundedness in long-form question answering (LFQA) by retrieval-augmented large language models (LLMs).In particular, we evaluate whether every generated sentence is grounded in the retrieved documents or the model’s pre-training data.Across 3 datasets and 4 model families, our findings reveal that a significant fraction of generated sentences are consistently ungrounded, even when those sentences contain correct ground-truth answers.Additionally, we examine the impacts of factors such as model size, decoding strategy, and instruction tuning on groundedness. Our results show that while larger models tend to ground their outputs more effectively, a significant portion of correct answers remains compromised by hallucinations. This study provides novel insights into the groundedness challenges in LFQA and underscores the necessity for more robust mechanisms in LLMs to mitigate the generation of ungrounded content.

The injection of syntactic information in Variational AutoEncoders (VAEs) can result in an overall improvement of performances and generalisation. An effective strategy to achieve such a goal is to separate the encoding of distributional semantic features and syntactic structures into heterogeneous latent spaces via multi-task learning or dual encoder architectures. However, existing works employing such techniques are limited to LSTM-based VAEs. This work investigates latent space separation methods for structural syntactic injection in Transformer-based VAE architectures (i.e., Optimus) through the integration of graph-based models. Our empirical evaluation reveals that the proposed end-to-end VAE architecture can improve theoverall organisation of the latent space, alleviating the information loss occurring in standard VAE setups, and resulting in enhanced performances on language modelling and downstream generation tasks.

Current Conversational AI systems employ different machine learning pipelines, as well as external knowledge sources and business logic to predict the next action. Maintaining various components in dialogue managers’ pipeline adds complexity in expansion and updates, increases processing time, and causes additive noise through the pipeline that can lead to incorrect next action prediction. This paper investigates graph integration into language transformers to improve understanding the relationships between humans’ utterances, previous, and next actions without the dependency on external sources or components. Experimental analyses on real calls indicate that the proposed Graph Integrated Language Transformer models can achieve higher performance compared to other production level conversational AI systems in driving interactive calls with human users in real-world settings.

One of the main problems low-resource languages face in NLP can be pictured as a vicious circle: data is needed to build and test tools, but the available text is scarce and there are not powerful tools to collect it.In order to break this circle for Guarani, we explore if text automatically generated from a grammar can work as a Data Augmentation technique to boost the performance of Guarani-Spanish Machine Translation (MT) systems.After building a grammar-based system that generates Spanish text and syntactically transfers it to Guarani, we perform several experiments by pretraining models using this synthetic text.We find that the MT systems that are pretrained with synthetic text perform better, even outperforming previous baselines.

Pretrained Language Models (PLMs) benefit from external knowledge stored in graph structures for various downstream tasks. However, bridging the modality gap between graph structures and text remains a significant challenge. Traditional methods like linearizing graphs for PLMs lose vital graph connectivity, whereas Graph Neural Networks (GNNs) require cumbersome processes for integration into PLMs. In this work, we propose a novel graph-guided self-attention mechanism, GraSAME. GraSAME seamlessly incorporates token-level structural information into PLMs without necessitating additional alignment or concatenation efforts. As an end-to-end, lightweight multimodal module, GraSAME follows a multi-task learning strategy and effectively bridges the gap between graph and textual modalities, facilitating dynamic interactions between GNNs and PLMs. Our experiments on the graph-to-text generation task demonstrate that GraSAME outperforms baseline models and achieves results comparable to state-of-the-art (SOTA) models on WebNLG datasets. Furthermore, compared to SOTA models, GraSAME eliminates the need for extra pre-training tasks to adjust graph inputs and reduces the number of trainable parameters by over 100 million.

With the widespread use of language models (LMs) in NLP tasks, researchers have discovered the potential of Chain-of-thought (CoT) to assist LMs in accomplishing complex reasoning tasks by generating intermediate steps. However, human thought processes are often non-linear, rather than simply sequential chains of thoughts. Therefore, we propose Graph-of-Thought (GoT) reasoning, which models human thought processes not only as a chain but also as a graph. By representing thought units as nodes and connections between them as edges, our approach captures the non-sequential nature of human thinking and allows for a more realistic modeling of thought processes. GoT adopts a two-stage framework with an additional GoT encoder for thought graph representation and fuses the graph representation with the original input representation through a gated fusion mechanism. We evaluate GoT’s performance on a text-only reasoning task (AQUA-RAT) and a multimodal reasoning task (ScienceQA). Our model achieves significant improvement over the strong CoT baseline on the AQUA-RAT test set and boosts accuracy from 85.19% to 87.59% using the T5-base model over the state-of-the-art Multimodal-CoT on the ScienceQA test set. Our code is publicly available at https://github.com/Zoeyyao27/Graph-of-Thought

Exploring the intersection of language and culture in Large Language Models (LLMs), this study critically examines their capability to encapsulate cultural nuances across diverse linguistic landscapes. Central to our investigation are three research questions: the efficacy of language-specific instruction tuning, the impact of pretraining on dominant language data, and the identification of optimal approaches to elicit accurate cultural knowledge from LLMs. Utilizing the GeoMLaMA benchmark for multilingual commonsense knowledge and an adapted CAMeL dataset (English-only) for evaluation of nuanced cultural aspects, our experiments span six different languages and cultural contexts, revealing the extent of LLMs’ cultural awareness. Our findings highlight a nuanced landscape: while language-specific tuning and bilingual pretraining enhance cultural understanding in certain contexts, they also uncover inconsistencies and biases, particularly in non-Western cultures. This work expands our understanding of LLMs’ cultural competence and emphasizes the importance of integrating diverse cultural perspectives in their development, aiming for a more globally representative and equitable approach in language modeling.