NAACL 2024 论文列表

Most research about natural language generation (NLG) relies on evaluation benchmarks with limited references for a sample, which may result in poor correlations with human judgements. The underlying reason is that one semantic meaning can actually be expressed in different forms, and the evaluation with a single or few references may not accurately reflect the quality of the model’s hypotheses. To address this issue, this paper presents a simple and effective method, named **Div-Ref**, to enhance existing evaluation benchmarks by enriching the number of references. We leverage large language models (LLMs) to diversify the expression of a single reference into multiple high-quality ones to cover the semantic space of the reference sentence as much as possible. We conduct comprehensive experiments to empirically demonstrate that diversifying the expression of reference can significantly enhance the correlation between automatic evaluation and human evaluation. This idea is compatible with recent LLM-based evaluation which can similarly derive advantages from incorporating multiple references. *We strongly encourage future generation benchmarks to include more references, even if they are generated by LLMs, which is once for all.* We release all the code and data at https://github.com/RUCAIBox/Div-Ref to facilitate research.

Conservation of historical documents benefits from computational methods by alleviating the manual labor related to digitization and modernization of textual content. Languages usually evolve over time and keeping historical wordforms is crucial for diachronic studies and digital humanities. However, spelling conventions did not necessarily exist when texts were originally written and orthographic variations are commonly observed depending on scribes and time periods. In this study, we propose to automatically normalize orthographic wordforms found in historical archives written in Middle French during the 16th century without fully modernizing textual content. We leverage pre-trained models in a low resource setting based on a manually curated parallel corpus and produce additional resources with artificial data generation approaches. Results show that causal language models and knowledge distillation improve over a strong baseline, thus validating the proposed methods.

Sample contrastive methods, typically referred to simply as contrastive are the foundation of most unsupervised methods to learn text and sentence embeddings. On the other hand, a different class of self-supervised non-contrastive loss functions and methods have been considered in the computer vision community and referred to as dimension contrastive. In this paper, we thoroughly compare this class of methods with the standard baseline for contrastive sentence embeddings, SimCSE. We find that self-supervised embeddings trained using dimension contrastive objectives can outperform SimCSE on downstream tasks without needing auxiliary loss functions.

In noisy label learning, instance selection based on small-loss criteria has been proven to be highly effective. However, in the case of noisy multi-label text classification (NMLTC), the presence of noise is not limited to the instance-level but extends to the (instance-label) pair-level.This gives rise to two main challenges.(1) The loss information at the pair-level fails to capture the variations between instances. (2) There are two types of noise at the pair-level: false positives and false negatives. Identifying false negatives from a large pool of negative pairs presents an exceedingly difficult task. To tackle these issues, we propose a novel approach called instance-label pair correction (iLaCo), which aims to address the problem of noisy pair selection and correction in NMLTC tasks.Specifically, we first introduce a holistic selection metric that identifies noisy pairs by simultaneously considering global loss information and instance-specific ranking information.Secondly, we employ a filter guided by label correlation to focus exclusively on negative pairs with label relevance. This filter significantly reduces the difficulty of identifying false negatives.Experimental analysis indicates that our framework effectively corrects noisy pairs in NMLTC datasets, leading to a significant improvement in model performance.

Reasoning in the presence of idiomatic expressions (IEs) remains a challenging frontier in natural language understanding (NLU). Unlike standard text, the non-compositional nature of an IE makes it difficult for model comprehension, as their figurative or non-literal mean- ing usually cannot be inferred from the constituent words alone. It stands to reason that in these challenging circumstances, pre-trained language models (PTLMs) should make use of the surrounding context to infer additional in- formation about the IE. In this paper, we investigate the utilization of said context for idiomatic reasoning tasks, which is under-explored relative to arithmetic or commonsense reason- ing (Liu et al., 2022; Yu et al., 2023). Preliminary findings point to a surprising observation: general purpose PTLMs are actually negatively affected by the context, as performance almost always increases with its removal. In these scenarios, models may see gains of up to 3.89%. As a result, we argue that only IE-aware models remain suitable for idiomatic reasoning tasks, given the unexpected and unexplainable manner in which general purpose PTLMs reason over IEs. Additionally, we conduct studies to examine how models utilize the context in various situations, as well as an in-depth analysis on dataset formation and quality. Finally, we provide some explanations and insights into the reasoning process itself based on our results.

To establish sophisticated monitoring of newspaper articles for detecting crisis-related signals, natural language processing has to cope with unstructured data, media, and cultural bias as well as multiple languages. So far, research on detecting signals in newspaper articles is focusing on structured data, restricted language settings, and isolated application domains. When considering complex crisis-related signals, a high number of diverse newspaper articles in terms of language and culture reduces potential biases. We demonstrate MENDEL – a model for multi-lingual and open-domain newspaper signaling for detecting crisis-related indicators in newspaper articles. The model works with unstructured news data and combines multiple transformer-based models for pre-processing (STANZA) and content filtering (RoBERTa, GPT-3.5). Embedded in a Question-Answering (QA) setting, MENDEL supports multiple languages (>66) and can detect early newspaper signals for open crisis domains in real-time.

This paper introduces Neurocache, an approach to extend the effective context size of large language models (LLMs) using an external vector cache to store its past states. Like recent vector retrieval approaches, Neurocache uses an efficient k-nearest-neighbor (kNN) algorithm to retrieve relevant past states and incorporate them into the attention process. Neurocache improves upon previous methods by (1) storing compressed states, which reduces cache size; (2) performing a single retrieval operation per token which increases inference speed; and (3) extending the retrieval window to neighboring states, which improves both language modeling and downstream task accuracy. Our experiments show the effectiveness of Neurocache both for models trained from scratch and for pre-trained models such as Llama2-7B and Mistral-7B when enhanced with the cache mechanism. We also compare Neurocache with text retrieval methods and show improvements in single-document question-answering and few-shot learning tasks. We made the source code available under: https://github.com/alisafaya/neurocache

Comparative knowledge (e.g., steel is stronger and heavier than styrofoam) is an essential component of our world knowledge, yet understudied in prior literature. In this paper, we harvest the dramatic improvements in knowledge capabilities of language models into a large-scale comparative knowledge base. While the ease of acquisition of such comparative knowledge is much higher from extreme-scale models like GPT-4, compared to their considerably smaller and weaker counterparts such as GPT-2, not even the most powerful models are exempt from making errors. We thus ask: to what extent are models at different scales able to generate valid and diverse comparative knowledge?We introduce NeuroComparatives, a novel framework for comparative knowledge distillation overgenerated from language models such as GPT-variants and LLaMA, followed by stringent filtering of the generated knowledge. Our framework acquires comparative knowledge between everyday objects, producing a corpus of up to 8.8M comparisons over 1.74M entity pairs - 10X larger and 30% more diverse than existing resources. Moreover, human evaluations show that NeuroComparatives outperform existing resources in terms of validity (up to 32% absolute improvement). Our acquired NeuroComparatives leads to performance improvements on five downstream tasks.We find that neuro-symbolic manipulation of smaller models offers complementary benefits to the currently dominant practice of prompting extreme-scale language models for knowledge distillation.

Federated embodied agent learning protects the data privacy of individual visual environments by keeping data locally at each client (the individual environment) during training. However, since the local data is inaccessible to the server under federated learning, attackers may easily poison the training data of the local client to build a backdoor in the agent without notice. Deploying such an agent raises the risk of potential harm to humans, as the attackers may easily navigate and control the agent as they wish via the backdoor. Towards Byzantine-robust federated embodied agent learning, in this paper, we study the attack and defense for the task of vision-and-language navigation (VLN), where the agent is required to follow natural language instructions to navigate indoor environments. First, we introduce a simple but effective attack strategy, Navigation as Wish (NAW), in which the malicious client manipulates local trajectory data to implant a backdoor into the global model. Results on two VLN datasets (R2R and RxR) show that NAW can easily navigate the deployed VLN agent regardless of the language instruction, without affecting its performance on normal test sets. Then, we propose a new Prompt-Based Aggregation (PBA) to defend against the NAW attack in federated VLN, which provides the server with a ”prompt” of the vision-and-language alignment variance between the benign and malicious clients so that they can be distinguished during training. We validate the effectiveness of the PBA method on protecting the global model from the NAW attack, which outperforms other state-of-the-art defense methods by a large margin in the defense metrics on R2R and RxR.

This paper investigates the potential of using natural language descriptions as an alternative to direct image-based observations for learning policies in reinforcement learning. Due to the inherent challenges in managing image-based observations, which include abundant information and irrelevant features, we propose a method that compresses images into a natural language form for state representation. This approach allows better interpretability and leverages the processing capabilities of large-language models. We conducted several experiments involving tasks that required image-based observation. The results demonstrated that policies trained using natural language descriptions of images yield better generalization than those trained directly from images, emphasizing the potential of this approach in practical settings.

How can we perform computations over natural language representations to solve tasks that require symbolic and numeric reasoning? We propose natural language embedded programs (NLEP) as a unifying framework for addressing math/symbolic reasoning, natural language understanding, and instruction following tasks. Our approach prompts a language model to generate full Python programs that define functions over data structures which contain natural language representations of structured knowledge. A Python interpreter then executes the generated code and prints the output. Despite using a task-general prompt, we find that this approach can improve upon strong baselines across a range of different tasks including math and symbolic reasoning, text classification, question answering, and instruction following. We found that the generated programs are interpretable since they outline the exact reasoning process followed by the program interpreter.

We present the first comprehensive survey of Native Language Identification (NLI) applied to texts. NLI is the task of automatically identifying an author’s native language (L1) based on their second language (L2) production. NLI is an important task with practical applications in second language teaching and NLP. The task has been widely studied for both text and speech, particularly for L2 English due to the availability of suitable corpora. Speech-based NLI relies heavily on accent modeled by pronunciation patterns and prosodic cues while text-based NLI relies primarily on modeling spelling errors and grammatical patterns that reveal properties of an individuals’ L1 influencing L2 production. We survey over one hundred papers on the topic including the papers associated with the NLI and INLI shared tasks. We describe several text representations and computational techniques used in text-based NLI. Finally, we present a comprehensive account of publicly available datasets used for the task thus far.

Large language models trained primarily in a monolingual setting have demonstrated their ability to generalize to machine translation using zero- and few-shot examples with in-context learning. However, even though zero-shot translations are relatively good, there remains a discernible gap comparing their performance with the few-shot setting. In this paper, we investigate the factors contributing to this gap and find that this gap can largely be closed (for about 70%) by matching the writing styles of the target corpus. Additionally, we explore potential approaches to enhance zero-shot baselines without the need for parallel demonstration examples, providing valuable insights into how these methods contribute to improving translation metrics.

Named entity recognition is a key component of Information Extraction (IE), particularly in scientific domains such as biomedicine and chemistry, where large language models (LLMs), e.g., ChatGPT, fall short. We investigate the applicability of transfer learning for enhancing a named entity recognition model trained in the biomedical domain (the source domain) to be used in the chemical domain (the target domain). A common practice for training such a model in a few-shot learning setting is to pretrain the model on the labeled source data, and then, to finetune it on a hand-full of labeled target examples. In our experiments, we observed that such a model is prone to mislabeling the source entities, which can often appear in the text, as the target entities. To alleviate this problem, we propose a model to transfer the knowledge from the source domain to the target domain, but, at the same time, to project the source entities and target entities into separate regions of the feature space. This diminishes the risk of mislabeling the source entities as the target entities. Our model consists of two stages: 1) entity grouping in the source domain, which incorporates knowledge from annotated events to establish relations between entities, and 2) entity discrimination in the target domain, which relies on pseudo labeling and contrastive learning to enhance discrimination between the entities in the two domains. We conduct our extensive experiments across three source and three target datasets, demonstrating that our method outperforms the baselines by up to 5% absolute value. Code, data, and resources are publicly available for research purposes: https://github.com/Lhtie/Bio-Domain-Transfer .

Large Language Models (LLMs) have shown promising in-context learning abilities. However, conventional In-Context Learning (ICL) approaches are often impeded by length limitations of transformer architecture, which pose challenges when attempting to effectively integrate supervision from a substantial number of demonstration examples. In this paper, we introduce a novel framework, called Naive Bayes-based Context Extension (NBCE), to enable existing LLMs to perform ICL with an increased number of demonstrations by significantly expanding their context size. Importantly, this expansion does not require fine-tuning or dependence on particular model architectures, all the while preserving linear efficiency. NBCE initially splits the context into equal-sized windows fitting the target LLM’s maximum length. Then, it introduces a voting mechanism to select the most relevant window, regarded as the posterior context. Finally, it employs Bayes’ theorem to generate the test task. Our experimental results demonstrate that NBCE substantially enhances performance, particularly as the number of demonstration examples increases, consistently outperforming alternative methods. The NBCE code will be made publicly accessible. The code NBCE is available at: https://github.com/amurtadha/NBCE-master

In recent years, counterspeech has emerged as one of the most promising strategies to fight online hate. These non-escalatory responses tackle online abuse while preserving the freedom of speech of the users, and can have a tangible impact in reducing online and offline violence. Recently, there has been growing interest from the Natural Language Processing (NLP) community in addressing the challenges of analysing, collecting, classifying, and automatically generating counterspeech, to reduce the huge burden of manually producing it. In particular, researchers have taken different directions in addressing these challenges, thus providing a variety of related tasks and resources. In this paper, we provide a guide for doing research on counterspeech, by describing - with detailed examples - the steps to undertake, and providing best practices that can be learnt from the NLP studies on this topic. Finally, we discuss open challenges and future directions of counterspeech research in NLP.

The use of words to convey speaker’s intent is traditionally distinguished from the ‘mention’ of words for quoting what someone said, or pointing out properties of a word. Here we show that computationally modeling this use-mention distinction is crucial for dealing with counterspeech online. Counterspeech that refutes problematic content often mentions harmful language but is not harmful itself (e.g., calling a vaccine dangerous is not the same as expressing disapproval of someone for calling vaccines dangerous). We show that even recent language models fail at distinguishing use from mention, and that this failure propagates to two key downstream tasks: misinformation and hate speech detection, resulting in censorship of counterspeech. We introduce prompting mitigations that teach the use-mention distinction, and show they reduce these errors. Our work highlights the importance of the use-mention distinction for NLP and CSS and offers ways to address it.

The paper focuses on the marginalization of indigenous language communities in the face of rapid technological advancements. We highlight the cultural richness of these languages and the risk they face of being overlooked in the realm of Natural Language Processing (NLP). We aim to bridge the gap between these communities and researchers, emphasizing the need for inclusive technological advancements that respect indigenous community perspectives. We show the NLP progress of indigenous Latin American languages and the survey that covers the status of indigenous languages in Latin America, their representation in NLP, and the challenges and innovations required for their preservation and development. The paper contributes to the current literature in understanding the need and progress of NLP for indigenous communities of Latin America, specifically low-resource and indigenous communities in general.

Humans frequently experience emotions. When emotions arise, they affect not only our mental state but can also change our physical state. For example, we often open our eyes wide when we are surprised, or clap our hands when we feel excited. Physical manifestations of emotions are referred to as embodied emotion in the psychology literature. From an NLP perspective, recognizing descriptions of physical movements or physiological responses associated with emotions is a type of implicit emotion recognition. Our work introduces a new task of recognizing expressions of embodied emotion in natural language. We create a dataset of sentences that contains 7,300 body part mentions with human annotations for embodied emotion. We develop a classification model for this task and present two methods to acquire weakly labeled instances of embodied emotion by extracting emotional manner expressions and by prompting a language model. Our experiments show that the weakly labeled data can train an effective classification model without gold data, and can also improve performance when combined with gold data. Our dataset is publicly available at https://github.com/yyzhuang1991/Embodied-Emotions.