EMNLP 2023 论文列表

Knowledge of norms is needed to understand and reason about acceptable behavior in human communication and interactions across sociocultural scenarios. Most computational research on norms has focused on a single culture, and manually built datasets, from non-conversational settings. We address these limitations by proposing a new framework, NormSage, to automatically extract culture-specific norms from multi-lingual conversations. NormSage uses GPT-3 prompting to 1) extract candidate norms directly from conversations and 2) provide explainable self-verification to ensure correctness and relevance. Comprehensive empirical results show the promise of our approach to extract high-quality culture-aware norms from multi-lingual conversations (English and Chinese), across several quality metrics. Further, our relevance verification can be extended to assess the adherence and violation of any norm with respect to a conversation on-the-fly, along with textual explanation. NormSage achieves an AUC of 94.6\% in this grounding setup, with generated explanations matching human-written quality.

Modern embedding-based metrics for evaluation of generated text generally fall into one of two paradigms: discriminative metrics that are trained to directly predict which outputs are of higher quality according to supervised human annotations, and generative metrics that are trained to evaluate text based on the probabilities of a generative model. Both have their advantages; discriminative metrics are able to directly optimize for the problem of distinguishing between good and bad outputs, while generative metrics can be trained using abundant raw text. In this paper, we present a framework that combines the best of both worlds, using both supervised and unsupervised signals from whatever data we have available. We operationalize this idea by training T5Score, a metric that uses these training signals with mT5 as backbone. We perform an extensive empirical comparison with other existing metrics on 5 datasets, 19 languages and 280 systems, demonstrating the utility of our method. Experimental results show that: T5Score achieves the best performance on all datasets against existing top-scoring metrics at the segment level.

In the absence of readily available labeled data for a given sequence labeling task and language, annotation projection has been proposed as one of the possible strategies to automatically generate annotated data. Annotation projection has often been formulated as the task of transporting, on parallel corpora, the labels pertaining to a given span in the source language into its corresponding span in the target language. In this paper we present T-Projection, a novel approach for annotation projection that leverages large pretrained text2text language models and state-of-the-art machine translation technology. T-Projection decomposes the label projection task into two subtasks: (i) A candidate generation step, in which a set of projection candidates using a multilingual T5 model is generated and, (ii) a candidate selection step, in which the generated candidates are ranked based on translation probabilities. We conducted experiments on intrinsic and extrinsic tasks in 5 Indo-European and 8 low-resource African languages. We demostrate that T-projection outperforms previous annotation projection methods by a wide margin. We believe that T-Projection can help to automatically alleviate the lack of high-quality training data for sequence labeling tasks. Code and data are publicly available.

The facts and time in the document are intricately intertwined, making temporal reasoning over documents challenging. Previous work models time implicitly, making it difficult to handle such complex relationships. To address this issue, we propose MTGER, a novel Multi-view Temporal Graph Enhanced Reasoning framework for temporal reasoning over time-involved documents. Concretely, MTGER explicitly models the temporal relationships among facts by multi-view temporal graphs. On the one hand, the heterogeneous temporal graphs explicitly model the temporal and discourse relationships among facts; on the other hand, the multi-view mechanism captures both time-focused and fact-focused information, allowing the two views to complement each other through adaptive fusion. To further improve the implicit reasoning capability of the model, we design a self-supervised time-comparing objective. Extensive experimental results demonstrate the effectiveness of our method on the TimeQA and SituatedQA datasets. Furthermore, MTGER gives more consistent answers under question perturbations.

The impressive success of recent deep neural network (DNN)-based systems is significantly influenced by the high-quality datasets used in training. However, the effects of the datasets, especially how they interact with each other, remain underexplored. We propose a state-vector framework to enable rigorous studies in this direction. This framework uses idealized probing test results as the bases of a vector space. This framework allows us to quantify the effects of both standalone and interacting datasets. We show that the significant effects of some commonly-used language understanding datasets are characteristic and are concentrated on a few linguistic dimensions. Additionally, we observe some ``spill-over'' effects: the datasets could impact the models along dimensions that may seem unrelated to the intended tasks. Our state-vector framework paves the way for a systematic understanding of the dataset effects, a crucial component in responsible and robust model development.

Transformer models have achieved impressive success in various natural language processing tasks. But it is also limited used in some areas and the heavy computation complexity is one of the main limitations. Many model structures have been proposed to reduce the computation complexity and some are really effective. The previous research can be divided into two categories. One is to use more effective training and inference strategies and the other is focused on how to replace the standard self-attention mechanism with linear attention method. Differently, we revisit the design in Transformer and find that the feed forward network (FFN) is also computationally expensive, especially when the hidden dimension is large. In this paper, we propose a new FFN structure, named MSCFFN, which splits the large matrix space to several small space to reduce the computation complexity and uses the Multi-Space Cross method to ensure the accurate result. To the best of our knowledge, this is the first time to redesign FFN to accelerate Transformers. We experimentally validate the effectiveness of the proposed method on the Long-Range Arena benchmark. And the results show MSCFFN can achieve a faster speed with a similar or even better accuracy.

Context-aware neural machine translation, a paradigm that involves leveraging information beyond sentence-level context to resolve inter-sentential discourse dependencies and improve document-level translation quality, has given rise to a number of recent techniques. However, despite well-reasoned intuitions, most context-aware translation models show only modest improvements over sentence-level systems. In this work, we investigate and present several core challenges that impede progress within the field, relating to discourse phenomena, context usage, model architectures, and document-level evaluation. To address these problems, we propose a more realistic setting for document-level translation, called paragraph-to-paragraph (PARA2PARA) translation, and collect a new dataset of Chinese-English novels to promote future research.

We propose task-adaptive tokenization\footnote{Our work will be publicly available upon acceptance.} as a way to adapt the generation pipeline to the specifics of a downstream task and enhance long-form generation in mental health. Inspired by insights from cognitive science, our task-adaptive tokenizer samples variable segmentations from multiple outcomes, with sampling probabilities optimized based on task-specific data. We introduce a strategy for building a specialized vocabulary and introduce a vocabulary merging protocol that allows for the integration of task-specific tokens into the pre-trained model's tokenization step. Through extensive experiments on psychological question-answering tasks in both Chinese and English, we find that our task-adaptive tokenization approach brings a significant improvement in generation performance while using up to 60\% fewer tokens. Preliminary experiments point to promising results when using our tokenization approach with very large language models.

Combating disinformation is one of the burning societal crises - about 67% of the American population believes that disinformation produces a lot of uncertainty, and 10% of them knowingly propagate disinformation. Evidence shows that disinformation can manipulate democratic processes and public opinion, causing disruption in the share market, panic and anxiety in society, and even death during crises. Therefore, disinformation should be identified promptly and, if possible, mitigated. With approximately 3.2 billion images and 720,000 hours of video shared online daily on social media platforms, scalable detection of multimodal disinformation requires efficient fact verification. Despite progress in automatic text-based fact verification (e.g., FEVER, LIAR), the research community lacks substantial effort in multimodal fact verification. To address this gap, we introduce FACTIFY 3M, a dataset of 3 million samples that pushes the boundaries of the domain of fact verification via a multimodal fake news dataset, in addition to offering explainability through the concept of 5W question-answering. Salient features of the dataset include: (i) textual claims, (ii) ChatGPT-generated paraphrased claims, (iii) associated images, (iv) stable diffusion-generated additional images (i.e., visual paraphrases), (v) pixel-level image heatmap to foster image-text explainability of the claim, (vi) 5W QA pairs, and (vii) adversarial fake news stories.

This paper investigates the challenges in building Swiss German speech translation systems, specifically focusing on the impact of dialect diversity and differences between Swiss German and Standard German. Swiss German is a spoken language with no formal writing system, it comprises many diverse dialects and is a low-resource language with only around 5 million speakers. The study is guided by two key research questions: how does the inclusion and exclusion of dialects during the training of speech translation models for Swiss German impact the performance on specific dialects, and how do the differences between Swiss German and Standard German impact the performance of the systems? We show that dialect diversity and linguistic differences pose significant challenges to Swiss German speech translation, which is in line with linguistic hypotheses derived from empirical investigations.

Existing multi-domain dialog state tracking (DST) models are developed based on multi-domain dialogs, which require significant manual effort to define domain relations and collect data. This process can be challenging and expensive, particularly when numerous domains are involved. In this paper, we propose a divide-and-conquer (DAC) DST paradigm and a multi-domain dialog synthesis framework, which makes building multi-domain DST models from single-domain dialogs possible. The DAC paradigm segments a multi-domain dialog into multiple single-domain dialogs for DST, which makes models generalize better on dialogs involving unseen domain combinations. The multi-domain dialog synthesis framework merges several potentially related single-domain dialogs into one multi-domain dialog and modifies the dialog to simulate domain relations. The synthesized dialogs can help DST models capture the value transfer between domains. Experiments with three representative DST models on two datasets demonstrate the effectiveness of our proposed DAC paradigm and data synthesis framework.

Vision-and-language navigation (VLN) agents are trained to navigate in real-world environments based on natural language instructions. A major challenge in VLN is the limited available training data, which hinders the models' ability to generalize effectively. Previous approaches have attempted to alleviate this issue by using external tools to generate pseudo-labeled data or integrating web-scaled image-text pairs during training. However, these methods often rely on automatically-generated or out-of-domain data, leading to challenges such as suboptimal data quality and domain mismatch. In this paper, we introduce a masked path modeling (MPM) objective. MPM pretrains an agent using self-collected data for subsequent navigation tasks, eliminating the need for external tools. Specifically, our method allows the agent to explore navigation environments and record the paths it traverses alongside the corresponding agent actions. Subsequently, we train the agent on this collected data to reconstruct the original action sequence when given a randomly masked subsequence of the original path. This approach enables the agent to accumulate a diverse and substantial dataset, facilitating the connection between visual observations of paths and the agent's actions, which is the foundation of the VLN task. Importantly, the collected data are in-domain, and the training process avoids synthetic data with uncertain quality, addressing previous issues. We conduct experiments on various VLN datasets and demonstrate the applications of MPM across different levels of instruction complexity. Our results exhibit significant improvements in success rates, with enhancements of 1.3\%, 1.1\%, and 1.2\% on the val-unseen split of the Room-to-Room, Room-for-Room, and Room-across-Room datasets, respectively. Additionally, we underscore the adaptability of MPM as well as the potential for additional improvements when the agent is allowed to explore unseen environments prior to testing.

The potential of large language models (LLMs) to simultaneously perform a wide range of natural language processing (NLP) tasks has been the subject of extensive research. Although instruction tuning has proven to be a data-efficient method for transforming LLMs into such generalist models, their performance still lags behind specialist models trained exclusively for specific tasks. In this paper, we investigate whether incorporating broadcoverage generalist instruction tuning can contribute to building a specialist model. We hypothesize that its efficacy depends on task specificity and skill requirements. Our experiments assess four target tasks with distinct coverage levels, revealing that integrating generalist instruction tuning consistently enhances model performance when the task coverage is broad. The effect is particularly pronounced when the amount of task-specific training data is limited. Further investigation into three target tasks focusing on different capabilities demonstrates that generalist instruction tuning improves understanding and reasoning abilities. However, for tasks requiring factual knowledge, generalist data containing hallucinatory information may negatively affect the model’s performance. Overall, our work provides a systematic guide for developing specialist models with general instruction tuning.

Style representation learning builds content-independent representations of author style in text. To date, no large dataset of texts with stylometric annotations on a wide range of style dimensions has been compiled, perhaps because the linguistic expertise to perform such annotation would be prohibitively expensive. Therefore, current style representation approaches make use of unsupervised neural methods to disentangle style from content to create style vectors. These approaches, however, result in uninterpretable representations, complicating their usage in downstream applications like authorship attribution where auditing and explainability is critical. In this work, we use prompting to perform stylometry on a large number of texts to generate a synthetic stylometry dataset. We use this synthetic data to then train human-interpretable style representations we call LISA embeddings. We release our synthetic dataset (StyleGenome) and our interpretable style embedding model (LISA) as resources.

Although large language models (LLMs) have advanced the state-of-the-art in NLP significantly, deploying them for downstream applications is still challenging due to cost, responsiveness, control, or concerns around privacy and security. As such, trainable models are still the preferred option in some cases. However, these models still require human-labeled data for optimal performance, which is expensive and time-consuming to obtain. In order to address this issue, several techniques to reduce human effort involve labeling or generating data using LLMs. Although these methods are effective for certain applications, in practice they encounter difficulties in real-world scenarios. Labeling data requires careful data selection, while generating data necessitates task-specific prompt engineering. In this paper, we propose a unified data creation pipeline that requires only a single formatting example, and which is applicable to a broad range of tasks, including traditionally problematic ones with semantically devoid label spaces. In our experiments we demonstrate that instruction-following LLMs are highly cost-effective data creators, and that models trained with these data exhibit performance better than those trained with human-labeled data (by up to 17.5\%) on out-of-distribution evaluation, while maintaining comparable performance on in-distribution tasks. These results have important implications for the robustness of NLP systems deployed in the real-world.

Previous studies on machine translation evaluation mostly focused on the quality of individual sentences, while overlooking the important role of contextual information. Although WMT Metrics Shared Tasks have introduced context content into the human annotations of translation evaluation since 2019, the relevant metrics and methods still did not take advantage of the corresponding context. In this paper, we propose a context-aware machine translation evaluation metric called Cont-COMET, built upon the effective COMET framework. Our approach simultaneously considers the preceding and subsequent contexts of the sentence to be evaluated and trains our metric to be aligned with the setting during human annotation. We also introduce a content selection method to extract and utilize the most relevant information. The experiments and evaluation of Cont-COMET on the official test framework from WMT show improvements in both system-level and segment-level assessments.

In the context of multi-step reasoning, e.g., with chain-of-thought, language models (LMs) can easily assign a high likelihood to incorrect steps. As a result, decoding strategies that optimize for solution likelihood often yield incorrect solutions. To address this issue, we propose Guiding chain-of-thought ReAsoning with a CorrectnEss Discriminator (GRACE), a stepwise decoding approach that steers the decoding process towards producing correct reasoning steps. GRACE employs a discriminator trained with a contrastive loss over correct and incorrect steps, which is used during decoding to score next-step candidates based on their correctness. Importantly, GRACE only requires sampling from the LM, without the need for LM training or fine-tuning. Using models from FLAN-T5 and LLaMA families, we evaluate GRACE over four math and two symbolic reasoning tasks, where it exhibits substantial performance gains compared to greedy decoding, verifiers, and self-consistency in most settings. When further combined with self-consistency, GRACE outperforms all the baselines by sizeable margins. Human and LLM evaluations over GSM8K show that GRACE not only improves the final answer accuracy but also the correctness of the intermediate reasoning.

Zero-shot commonsense Question-Answering (QA) requires models to reason about general situations beyond specific benchmarks. State-of-the-art approaches fine-tune language models on QA pairs constructed from CommonSense Knowledge Bases (CSKBs) to equip the models with more commonsense knowledge in a QA context. However, current QA synthesis protocols may introduce noise from the CSKBs and generate ungrammatical questions and false negative options, which impede the model’s ability to generalize. To address these issues, we propose QADYNAMICS, a training dynamics-driven framework for QA diagnostics and refinement. Our approach analyzes the training dynamics of each QA pair at both the question level and option level, discarding machine-detectable artifacts by removing uninformative QA pairs and mislabeled or false-negative options. Extensive experiments demonstrate the effectiveness of our approach, which outperforms all baselines while using only 33% of the synthetic data, even including LLMs such as ChatGPT. Moreover, expert evaluations confirm that our framework significantly improves the quality of QA synthesis. Our code and model checkpoints are available at https://github.com/HKUST-KnowComp/QaDynamics.

Large Language Models (LLMs) have made remarkable advancements in the field of natural language generation. However, the propensity of LLMs to generate inaccurate or non-factual content, termed "hallucinations", remains a significant challenge. Current hallucination detection methods often necessitate the retrieval of great numbers of relevant evidence, thereby increasing response times. We introduce a unique framework that leverages statistical decision theory and Bayesian sequential analysis to optimize the trade-off between costs and benefits during the hallucination detection process. This approach does not require a predetermined number of observations. Instead, the analysis proceeds in a sequential manner, enabling an expeditious decision towards "belief" or "disbelief" through a stop-or-continue strategy. Extensive experiments reveal that this novel framework surpasses existing methods in both efficiency and precision of hallucination detection. Furthermore, it requires fewer retrieval steps on average, thus decreasing response times.