NAACL 2024 Papers

Multilingual pretraining has been a successful solution to the challenges posed by the lack of resources for languages. These models can transfer knowledge to target languages with minimal or no examples. Recent research suggests that monolingual models also have a similar capability, but the mechanisms behind this transfer remain unclear. Some studies have explored factors like language contamination and syntactic similarity. An emerging line of research suggests that the representations learned by language models contain two components: a language-specific and a language-agnostic component. The latter is responsible for transferring a more universal knowledge. However, there is a lack of comprehensive exploration of these properties across diverse target languages. To investigate this hypothesis, we conducted an experiment inspired by the work on the Scaling Laws for Transfer. We measured the amount of data transferred from a source language to a target language and found that models initialized from diverse languages perform similarly to a target language in a cross-lingual setting. This was surprising because the amount of data transferred to 10 diverse target languages, such as Spanish, Korean, and Finnish, was quite similar. We also found evidence that this transfer is not related to language contamination or language proximity, which strengthens the hypothesis that the model also relies on language-agnostic knowledge. Our experiments have opened up new possibilities for measuring how much data represents the language-agnostic representations learned during pretraining.

Measuring semantic similarity between texts is a crucial task in natural language processing. While existing semantic text matching focuses on pairs of similar-length sequences, matching texts with non-comparable lengths has broader applications in specific domains, such as comparing professional document summaries and content. Current approaches struggle with text pairs of non-comparable lengths due to truncation issues. To address this, we split texts into natural sentences and decouple sentence representations using supervised contrastive learning (SCL). Meanwhile, we adopt the embedded topic model (ETM) for specific domain data. Our experiments demonstrate the effectiveness of our model, based on decoupled and topic-informed sentence embeddings, in matching texts of significantly different lengths across three well-studied datasets.

We investigate massive end-to-end automatic speech recognition (ASR) models with efficiency improvements achieved by time reduction. The encoders of our models use the neural architecture of Google’s universal speech model (USM), with additional funnel pooling layers to significantly reduce the frame rate and speed up training and inference. We also explore a few practical methods to mitigate potential accuracy loss due to time reduction, while enjoying most efficiency gain. Our methods are demonstrated to work with both Connectionist Temporal Classification (CTC) and RNN-Transducer (RNN-T), with up to 2B model parameters, and over two domains. For a large-scale voice search recognition task, we perform extensive studies on vocabulary size, time reduction strategy, and its generalization performance on long-form test sets, and show that a 900M RNN-T is very tolerant to severe time reduction, with as low encoder output frame rate as 640ms. We also provide ablation studies on the Librispeech benchmark for important training hyperparameters and architecture designs, in training 600M RNN-T models at the frame rate of 160ms.

Social media is a valuable data source for exploring mental health issues. However, previous studies have predominantly focused on the semantic content of these posts, overlooking the importance of their temporal attributes, as well as the evolving nature of mental disorders and symptoms.In this paper, we study the causality between psychiatric symptoms and life events, as well as among different symptoms from social media posts, which leads to better understanding of the underlying mechanisms of mental disorders. By applying these extracted causality features to tasks such as diagnosis point detection and early risk detection of depression, we notice considerable performance enhancement. This indicates that causality information extracted from social media data can boost the efficacy of mental disorder diagnosis and treatment planning.

Decoding continuous language from brain activity is a formidable yet promising field of research. It is particularly significant for aiding people with speech disabilities to communicate through brain signals. This field addresses the complex task of mapping brain signals to text. The previous best attempt reverse-engineered this process in an indirect way: it began by learning to encode brain activity from text and then guided text generation by aligning with predicted brain responses. In contrast, we propose a simple yet effective method that guides text reconstruction by directly comparing them with the predicted text embeddings mapped from brain activities. Comprehensive experiments reveal that our method significantly outperforms the current state-of-the-art model, showing average improvements of 77% and 54% on BLEU and METEOR scores. We further validate the proposed modules through detailed ablation studies and case analyses and highlight a critical correlation: the more precisely we map brain activities to text embeddings, the better the text reconstruction results. Such insight can simplify the task of reconstructing language from brain activities for future work, emphasizing the importance of improving brain-to-text-embedding mapping techniques.

Tools serve as pivotal interfaces that enable humans to understand and reshape the environment. With the advent of foundation models, AI systems can utilize tools to expand their capabilities and interact with the real world. Existing tool learning methodologies, encompassing supervised fine-tuning and prompt engineering approaches, often induce large language models to utilize tools indiscriminately, as complex tasks often exceed their own competencies. However, introducing tools for simple tasks, which the models themselves can readily resolve, can inadvertently propagate errors rather than enhance performance. This leads to the research question: can we teach language models when and how to use tools? To meet this need, we propose Tool leaRning wIth exeCution fEedback (TRICE), a two-stage end-to-end framework that enables the model to continually learn through feedback derived from tool execution, thereby learning when and how to use tools effectively. Experimental results, backed by further analysis, show that TRICE can make the large language model selectively use tools by improving the accuracy of tool usage while enhancing insufficient tool learning and mitigating excessive reliance on tools.

We explore the creative problem-solving capabilities of modern LLMs in a novel constrained setting. To this end, we create MACGYVER, an automatically generated dataset consisting of over 1,600 real-world problems deliberately designed to trigger innovative usage of objects and necessitate out-of-the-box thinking. We then present our collection to both LLMs and humans to compare and contrast their problem-solving abilities. MACGYVER is challenging for both groups, but in unique and complementary ways. For instance, humans excel in tasks they are familiar with but struggle with domain-specific knowledge, leading to a higher variance. In contrast, LLMs, exposed to a variety of specialized knowledge, attempt broader problems but fail by proposing physically-infeasible actions. Finally, we provide a detailed error analysis of LLMs, and demonstrate the potential of enhancing their problem-solving ability with novel prompting techniques such as iterative step-wise reflection and divergent-convergent thinking.This work (1) introduces a fresh arena for intelligent agents focusing on intricate aspects of physical reasoning, planning, and unconventional thinking, which supplements the existing spectrum of machine intelligence; and (2) provides insight into the constrained problem-solving capabilities of both humans and AI.

Pre-trained language models (PLMs) that rely solely on textual data may exhibit limitations in multimodal semantics comprehension. Existing solutions attempt to alleviate this issue by incorporating explicit image retrieval or generation techniques.However, these methods: (1) focus exclusively on the static image modality; (2) inevitably encounter modality gaps and noise; (3) indiscriminately treat all modalities.In this paper, we propose a novel multimodal-augmented framework termed MaCSC, which can infuse multimodal semantics into PLMs and facilitate a self-balancing calibration of information allocation.Specifically, MaCSC obtains modal-specific conceptual prototypes from contrastive pre-training models (e.g., CLIP),and aggregates the intra- and inter-modal semantics of the conceptual prototype to enhance PLMs.In addition, we utilize a novel self-balancing contrastive loss to achieve multi-scale self-balancing calibration of multimodal information during fine-tuning PLMs.Experimental results show that MaCSC consistently improves the performance of PLMs across various architectures and scales, and outperforms competitive baselines on multiple NLP tasks.

Large Language Models (LLM) have demonstrated their strong ability in the field of machine translation, yet they suffer from high computational cost and latency. Therefore, transferring translation knowledge from giant LLMs to medium-sized machine translation models is a promising research direction. However, traditional knowledge distillation methods ignore the capability of student and teacher models, therefore repeatedly teaching student models on the knowledge they have learned, and failing to extend to novel contexts and knowledge. In this paper, we propose a framework called MT-Patcher, which transfers knowledge from LLMs to existing MT models in a selective, comprehensive and proactive manner. Considering the current translation ability of student MT models, we only identify and correct their translation errors, instead of distilling the whole translation from the teacher. Leveraging the strong language abilities of LLMs, we instruct LLM teachers to synthesize diverse contexts and anticipate more potential errors for the student. Experiment results on translating both specific language phenomena and general MT benchmarks demonstrate that finetuning the MT model on about 10% examples can achieve comparable results to the traditional knowledge distillation method, and synthesized potential errors and diverse contexts further improve MT performances on unseen contexts and words.

The task of scientific Natural Language Inference (NLI) involves predicting the semantic relation between two sentences extracted from research articles. This task was recently proposed along with a new dataset called SciNLI derived from papers published in the computational linguistics domain. In this paper, we aim to introduce diversity in the scientific NLI task and present MSciNLI, a dataset containing 132,320 sentence pairs extracted from five new scientific domains. The availability of multiple domains makes it possible to study domain shift for scientific NLI. We establish strong baselines on MSciNLI by fine-tuning Pre-trained Language Models (PLMs) and prompting Large Language Models (LLMs). The highest Macro F1 scores of PLM and LLM baselines are 77.21% and 51.77%, respectively, illustrating that MSciNLI is challenging for both types of models. Furthermore, we show that domain shift degrades the performance of scientific NLI models which demonstrates the diverse characteristics of different domains in our dataset. Finally, we use both scientific NLI datasets in an intermediate task transfer learning setting and show that they can improve the performance of downstream tasks in the scientific domain. We make our dataset and code available on Github.

Several Natural Language Understanding (NLU) tasks focus on linking text to explicit knowledge, including Word Sense Disambiguation, Semantic Role Labeling, Semantic Parsing, and Relation Extraction. In addition to the importance of connecting raw text with explicit knowledge bases, the integration of such carefully curated knowledge into deep learning models has been shown to be beneficial across a diverse range of applications, including Language Modeling and Machine Translation. Nevertheless, the scarcity of semantically-annotated corpora across various tasks and languages limits the potential advantages significantly. To address this issue, we put forward MOSAICo, the first endeavor aimed at equipping the research community with the key ingredients to model explicit semantic knowledge at a large scale, providing hundreds of millions of silver yet high-quality annotations for four NLU tasks across five languages. We describe the creation process of MOSAICo, demonstrate its quality and variety, and analyze the interplay between different types of semantic information. MOSAICo, available at https://github.com/SapienzaNLP/mosaico, aims to drop the requirement of closed, licensed datasets and represents a step towards a level playing field across languages and tasks in NLU.

News media often strive to minimize explicit moral language in news articles, yet most articles are dense with moral values as expressed through the reported events themselves. However, values that are reflected in the intricate dynamics among *participating entities* and *moral events* are far more challenging for most NLP systems to detect, including LLMs. To study this phenomenon, we annotate a new dataset, **MORAL EVENTS**, consisting of 5,494 structured event annotations on 474 news articles by diverse US media across the political spectrum. We further propose **MOKA**, a moral event extraction framework with **MO**ral **K**nowledge **A**ugmentation, which leverages knowledge derived from moral words and moral scenarios to produce structural representations of morality-bearing events. Experiments show that **MOKA** outperforms competitive baselines across three moral event understanding tasks. Further analysis shows even ostensibly nonpartisan media engage in the selective reporting of moral events.

With the rapid development of large language models (LLMs) and their integration into large multimodal models (LMMs), there has beenimpressive progress in zero-shot completion of user-oriented vision-language tasks. However, a gap remains in the domain of chartimage understanding due to the distinct abstract components in charts. To address this, we introduce a large-scale MultiModal ChartInstruction (MMC-Instruction) dataset comprising 600k instances supporting diverse tasks and chart types. Leveraging this data, we de-velop MultiModal Chart Assistant (MMCA), an LMM that achieves state-of-the-art performance on existing chart QA benchmarks. Recognizing the need for a comprehensive evaluation of LMM chart understanding, we also propose a MultiModal Chart Benchmark (MMC-Benchmark), a comprehensive human-annotated benchmark with nine distinct tasks evaluating reasoning capabilities over charts.Extensive experiments on MMC-Benchmark reveal the limitations of existing LMMs on correctly interpreting charts, even for the mostrecent GPT-4V model. Our work provides an instruction-tuning methodology and benchmark to advance multimodal understanding ofcharts. Code and data are available at https://github.com/FuxiaoLiu/MMC.

Query expansion, pivotal in search engines, enhances the representation of user information needs with additional terms. While existing methods expand queries using retrieved or generated contextual documents, each approach has notable limitations. Retrieval-based methods often fail to accurately capture search intent, particularly with brief or ambiguous queries. Generation-based methods, utilizing large language models (LLMs), generally lack corpus-specific knowledge and entail high fine-tuning costs. To address these gaps, we propose a novel zero-shot query expansion framework utilizing LLMs for mutual verification. Specifically, we first design a query-query-document generation method, leveraging LLMs’ zero-shot reasoning ability to produce diverse sub-queries and corresponding documents. Then, a mutual verification process synergizes generated and retrieved documents for optimal expansion. Our proposed method is fully zero-shot, and extensive experiments on three public benchmark datasets are conducted to demonstrate its effectiveness over existing methods. Our code is available online at https://github.com/Applied-Machine-Learning-Lab/MILL to ease reproduction.

Large Language Models (LLMs) are powerful tools which have been both dominant and commonplace in the field of Artificial Intelligence. Yet, LLMs have a tendency to devolve into toxic degeneration, wherein otherwise safe and unproblematic models begin generating toxic content. For the sake of social responsibility and inspired by the biological mechanisms of inhibition control, we introduce the paradigm of Education for Societal Norms (ESN). By collecting and labeling examples as acceptable and unacceptable (in this case toxic and non-toxic), and including a corresponding acceptable rewrite with every unacceptable example, we introduce a new mechanism for LLM detoxification. We annotate a dataset of 2,850 entries and use it to fine-tune a model, which we call a Model with Inhibition Control (MICo). Evaluating this model on toxicity detection capability, rewrite detoxification, meaning preservation, and overall toxicity reduction, we discover significant improvements over the baseline model. In our experiments we show that overall toxicity of this model is more than 60% reduced, with over 75% reduction in severe toxicity.

With the rising human-like precision of Large Language Models (LLMs) in numerous tasks, their utilization in a variety of real-world applications is becoming more prevalent. Several studies have shown that LLMs excel on many standard NLP benchmarks. However, it is challenging to evaluate LLMs due to test dataset contamination and the limitations of traditional metrics. Since human evaluations are difficult to collect, there is a growing interest in the community to use LLMs themselves as reference-free evaluators for subjective metrics. However, past work has shown that LLM-based evaluators can exhibit bias and have poor alignment with human judgments. In this study, we propose a framework for an end-to-end assessment of LLMs as evaluators in multilingual scenarios. We create a carefully curated dataset, covering 10 languages containing native speaker judgments for the task of summarization. This dataset is created specifically to evaluate LLM-based evaluators, which we refer to as meta-evaluation (METAL). We compare the performance of LLM-based evaluators created using GPT-3.5-Turbo, GPT-4, and PaLM2. Our results indicate that LLM-based evaluators based on GPT-4 perform the best across languages, while GPT-3.5-Turbo performs poorly. Additionally, we perform an analysis of the reasoning provided by LLM-based evaluators and find that it often does not match the reasoning provided by human judges.

Retrieval augmentation is a powerful but expensive method to make language models more knowledgeable about the world. Memory-based methods like LUMEN (de Jong et al., 2023a) pre-compute token representations for retrieved passages to drastically speed up inference. However, memory also leads to much greater storage requirements from storing pre-computed representations. We propose MEMORY-VQ, a new method to reduce storage requirements of memory-augmented models without sacrificing performance. Our method uses a vector quantization variational autoencoder (VQ-VAE) to compress token representations. We apply MEMORY-VQ to the LUMEN model to obtain LUMEN-VQ, a memory model that achieves a 16x compression rate with comparable performance on the KILT benchmark. LUMEN-VQ enables practical retrieval augmentation even for extremely large retrieval corpora.

There has been a surge in LLM evaluation research to understand LLM capabilities and limitations. However, much of this research has been confined to English, leaving LLM building and evaluation for non-English languages relatively unexplored. Several new LLMs have been introduced recently, necessitating their evaluation on non-English languages. This study aims to perform a thorough evaluation of the non-English capabilities of SoTA LLMs (GPT-3.5-Turbo, GPT-4, PaLM2, Gemini-Pro, Mistral, Llama2, and Gemma) by comparing them on the same set of multilingual datasets. Our benchmark comprises 22 datasets covering 83 languages, including low-resource African languages. We also include two multimodal datasets in the benchmark and compare the performance of LLaVA models, GPT-4-Vision and Gemini-Pro-Vision. Our experiments show that larger models such as GPT-4, Gemini-Pro and PaLM2 outperform smaller models on various tasks, notably on low-resource languages, with GPT-4 outperforming PaLM2 and Gemini-Pro on more datasets. We also perform a study on data contamination and find that several models are likely to be contaminated with multilingual evaluation benchmarks, necessitating approaches to detect and handle contamination while assessing the multilingual performance of LLMs.

Recently, retrieval-based in-context learning (ICL) methods for selecting demonstrations have been widely investigated. Existing methods train a dense retriever to retrieve the most appropriate demonstrations for a given test query, which improves ICL performance. However, we find that distinct LLMs exhibit different biases for “what is a good demonstration” since they possess differences in training data, model architectures and training methods. As a result, a demonstration suitable for one LLM may not be appropriate for others.Previous approaches ignore the model bias and fail to retrieve the most appropriate demonstrations for different inference LLMs, resulting in a degradation of ICL performance.To address this problem, we propose a simple yet effective metric to evaluate the appropriateness of demonstrations for a specific inference LLM. Furthermore, we introduce a Model-specific Demonstration Retrieval (MDR) method for ICL at inference time, which considers the biases of different LLMs. We test MDR on seen and unseen tasks with multi-scale inference LLMs, such as GPT-Neo-2.7B, LLaMA-7B and Vicuna-13B. Experiments on 23 datasets across 11 data domains highlight the remarkable effectiveness of MDR, showcasing improvements of up to 41.2% in comparison to methods that neglect model biases.