COLING 2024 论文列表

Synthetic data generation has the potential to impact applications and domains with scarce data. However, before such data is used for sensitive tasks such as mental health, we need an understanding of how different demographics are represented in it. In our paper, we analyze the potential of producing synthetic data using GPT-3 by exploring the various stressors it attributes to different race and gender combinations, to provide insight for future researchers looking into using LLMs for data generation. Using GPT-3, we develop HeadRoom, a synthetic dataset of 3,120 posts about depression-triggering stressors, by controlling for race, gender, and time frame (before and after COVID-19). Using this dataset, we conduct semantic and lexical analyses to (1) identify the predominant stressors for each demographic group; and (2) compare our synthetic data to a human-generated dataset. We present the procedures to generate queries to develop depression data using GPT-3, and conduct analyzes to uncover the types of stressors it assigns to demographic groups, which could be used to test the limitations of LLMs for synthetic data generation for depression data. Our findings show that synthetic data mimics some of the human-generated data distribution for the predominant depression stressors across diverse demographics.

Topic models aim to reveal latent structures within a corpus of text, typically through the use of term-frequency statistics over bag-of-words representations from documents. In recent years, conceptual entities — interpretable, language-independent features linked to external knowledge resources — have been used in place of word-level tokens, as words typically require extensive language processing with a minimal assurance of interpretability. However, current literature is limited when it comes to exploring purely entity-driven neural topic modeling. For instance, despite the advantages of using entities for eliciting thematic structure, it is unclear whether current techniques are compatible with these sparsely organised, information-dense conceptual units. In this work, we explore entity-based neural topic modeling and propose a novel topic clustering approach using bimodal vector representations of entities. Concretely, we extract these latent representations from large language models and graph neural networks trained on a knowledge base of symbolic relations, in order to derive the most salient aspects of these conceptual units. Analysis of coherency metrics confirms that our approach is better suited to working with entities in comparison to state-of-the-art models, particularly when using graph-based embeddings trained on a knowledge base.

Multimodal Conversational Emotion (MCE) detection, generally spanning across the acoustic, vision and language modalities, has attracted increasing interest in the multimedia community. Previous studies predominantly focus on learning contextual information in conversations with only a few considering the topic information in single language modality, while always neglecting the acoustic and vision topic information. On this basis, we propose a model-agnostic Topic-enriched Diffusion (TopicDiff) approach for capturing multimodal topic information in MCE tasks. Particularly, we integrate the diffusion model into neural topic model to alleviate the diversity deficiency problem of neural topic model in capturing topic information. Detailed evaluations demonstrate the significant improvements of TopicDiff over the state-of-the-art MCE baselines, justifying the importance of multimodal topic information to MCE and the effectiveness of TopicDiff in capturing such information. Furthermore, we observe an interesting finding that the topic information in acoustic and vision is more discriminative and robust compared to the language.

Topic-controllable summarization is an emerging research area with a wide range of potential applications. However, existing approaches suffer from significant limitations. For example, the majority of existing methods built upon recurrent architectures, which can significantly limit their performance compared to more recent Transformer-based architectures, while they also require modifications to the model’s architecture for controlling the topic. At the same time, there is currently no established evaluation metric designed specifically for topic-controllable summarization. This work proposes a new topic-oriented evaluation measure to automatically evaluate the generated summaries based on the topic affinity between the generated summary and the desired topic. The reliability of the proposed measure is demonstrated through appropriately designed human evaluation. In addition, we adapt topic embeddings to work with powerful Transformer architectures and propose a novel and efficient approach for guiding the summary generation through control tokens. Experimental results reveal that control tokens can achieve better performance compared to more complicated embedding-based approaches while also being significantly faster.

The time at which a message is communicated is a vital piece of metadata in many real-world natural language processing tasks such as Topic Detection and Tracking (TDT). TDT systems aim to cluster a corpus of news articles by event, and in that context, stories that describe the same event are likely to have been written at around the same time. Prior work on time modeling for TDT takes this into account, but does not well capture how time interacts with the semantic nature of the event. For example, stories about a tropical storm are likely to be written within a short time interval, while stories about a movie release may appear over weeks or months. In our work, we design a neural method that fuses temporal and textual information into a single representation of news documents for event detection. We fine-tune these time-aware document embeddings with a triplet loss architecture, integrate the model into downstream TDT systems, and evaluate the systems on two benchmark TDT data sets in English. In the retrospective setting, we apply clustering algorithms to the time-aware embeddings and show substantial improvements over baselines on the News2013 data set. In the online streaming setting, we add our document encoder to an existing state-of-the-art TDT pipeline and demonstrate that it can benefit the overall performance. We conduct ablation studies on the time representation and fusion algorithm strategies, showing that our proposed model outperforms alternative strategies. Finally, we probe the model to examine how it handles recurring events more effectively than previous TDT systems.

The development of NLP tools for low-resource languages is impeded by the lack of data. While recent unsupervised pre-training approaches ease this requirement, the need for labelled data is crucial to progress the development of such tools. Moreover, publicly available datasets for such languages typically cover low-level syntactic tasks. In this work, we introduce new semantic datasets for Maltese generated automatically using associated metadata from a corpus in the news domain. The datasets are a news tag multi-label classification and a news abstractive summarisation task by generating its title. We also present an evaluation using publicly available models as baselines. Our results show that current models are lacking the semantic knowledge required to solve such tasks, shedding light on the need to use better modelling approaches for Maltese.

Tool learning aims to extend the capabilities of large language models (LLMs) with external tools. A major challenge in tool learning is how to support a large number of tools, including unseen tools. To address this challenge, previous studies have proposed retrieving suitable tools for the LLM based on the user query. However, previously proposed methods do not consider the differences between seen and unseen tools, nor do they take the hierarchy of the tool library into account, which may lead to suboptimal performance for tool retrieval. Therefore, to address the aforementioned issues, we propose ToolRerank, an adaptive and hierarchy-aware reranking method for tool retrieval to further refine the retrieval results. Specifically, our proposed ToolRerank includes Adaptive Truncation, which truncates the retrieval results related to seen and unseen tools at different positions, and Hierarchy-Aware Reranking, which makes retrieval results more concentrated for single-tool queries and more diverse for multi-tool queries. Experimental results show that ToolRerank can improve the quality of the retrieval results, leading to better execution results generated by the LLM.

Minimal-pair paradigm datasets have been used as benchmarks to evaluate the linguistic knowledge of models and provide an unsupervised method of acceptability judgment. The model performances are evaluated based on the percentage of minimal pairs in the MPP dataset where the model assigns a higher sentence log-likelihood to an acceptable sentence than to an unacceptable sentence. Each minimal pair in MPP datasets is controlled to align the number of words per sentence because the sentence length affects the sentence log-likelihood. However, aligning the number of words may be insufficient because recent language models tokenize sentences with subwords. Tokenization may cause a token length difference in minimal pairs, introducing token-length bias that skews the evaluation results. This study demonstrates that MPP datasets suffer from token-length bias and fail to evaluate the linguistic knowledge of a language model correctly. The results proved that sentences with a shorter token length would likely be assigned a higher log-likelihood regardless of their acceptability, which becomes problematic when comparing models with different tokenizers. To address this issue, we propose a debiased minimal pair generation method, allowing MPP datasets to measure language ability correctly and provide comparable results for all models.

In recent years, the fine-tuned generative models have been proven more powerful than the previous tagging-based or span-based models on named entity recognition (NER) task. It has also been found that the information related to entities, such as entity types, can prompt a model to achieve NER better. However, it is not easy to determine the entity types indeed existing in the given sentence in advance, and inputting too many potential entity types would distract the model inevitably. To exploit entity types’ merit on promoting NER task, in this paper we propose a novel NER framework, namely ToNER based on a generative model. In ToNER, a type matching model is proposed at first to identify the entity types most likely to appear in the sentence. Then, we append a multiple binary classification task to fine-tune the generative model’s encoder, so as to generate the refined representation of the input sentence. Moreover, we add an auxiliary task for the model to discover the entity types which further fine-tunes the model to output more accurate results. Our extensive experiments on some NER benchmarks verify the effectiveness of our proposed strategies in ToNER that are oriented towards entity types’ exploitation.

Although the NLP community has adopted central differential privacy as a go-to framework for privacy-preserving model training or data sharing, the choice and interpretation of the key parameter, privacy budget 𝜀 that governs the strength of privacy protection, remains largely arbitrary. We argue that determining the 𝜀 value should not be solely in the hands of researchers or system developers, but must also take into account the actual people who share their potentially sensitive data. In other words: Would you share your instant messages for 𝜀 of 10? We address this research gap by designing, implementing, and conducting a behavioral experiment (311 lay participants) to study the behavior of people in uncertain decision-making situations with respect to privacy-threatening situations. Framing the risk perception in terms of two realistic NLP scenarios and using a vignette behavioral study help us determine what 𝜀 thresholds would lead lay people to be willing to share sensitive textual data – to our knowledge, the first study of its kind.

State-of-the-art language models perform well on a variety of language tasks, but they continue to struggle with understanding negation cues in tasks like natural language inference (NLI). Inspired by Hossain et al. (2020), who show under-representation of negation in language model pretraining datasets, we experiment with additional pretraining with negation data for which we introduce two new datasets. We also introduce a new learning strategy for negation building on ELECTRA’s (Clark et al., 2020) replaced token detection objective. We find that continuing to pretrain ELECTRA-Small’s discriminator leads to substantial gains on a variant of RTE (Recognizing Textual Entailment) with additional negation. On SNLI (Stanford NLI) (Bowman et al., 2015), there are no gains due to the extreme under-representation of negation in the data. Finally, on MNLI (Multi-NLI) (Williams et al., 2018), we find that performance on negation cues is primarily stymied by neutral-labeled examples.

Unpredictability, especially unpredictability with unknown error characteristics, is a highly undesirable trait, particularly in medical patient care applications. Although large pre-trained language models (LLM) have been applied to a variety of unseen tasks with highly competitive and successful results, their sensitivity to language inputs and resulting performance variability is not well-studied. In this work, we test state-of-the-art pre-trained language models from a variety of families to characterize their error generation and reliability in medical assessment ability. Particularly, we experiment with general medical assessment multiple choice tests, as well as their open-ended and true-false alternatives. We also profile model consistency, error agreements with each other and to humans; and finally, quantify their ability to recover and explain errors. The findings in this work can be used to give further information about medical models so that modelers can make better-informed decisions rather than relying on standalone performance metrics alone.

Speakers sometimes omit certain arguments of a predicate in a sentence; such omission is especially frequent in pro-drop languages. This study addresses a question about ellipsis—what can explain the native speakers’ ellipsis decisions?—motivated by the interest in human discourse processing and writing assistance for this choice. To this end, we first collect large-scale human annotations of whether and why a particular argument should be omitted across over 2,000 data points in the balanced corpus of Japanese, a prototypical pro-drop language. The data indicate that native speakers overall share common criteria for such judgments and further clarify their quantitative characteristics, e.g., the distribution of related linguistic factors in the balanced corpus. Furthermore, the performance of the language model–based argument ellipsis judgment model is examined, and the gap between the systems’ prediction and human judgments in specific linguistic aspects is revealed. We hope our fundamental resource encourages further studies on natural human ellipsis judgment.

Existing studies on extractive summarization have primarily focused on scoring and selecting summary sentences independently. However, these models are limited to sentence-level extraction and tend to select highly generalized sentences while overlooking the overall content of a document. To effectively consider the semantics of a document, in this study, we introduce a novel machine reading comprehension (MRC) framework for extractive summarization (MRCSum) by setting a query as the title. Our framework enables MRCSum to consider the semantic coherence and relevance of summary sentences in relation to the overall content. In particular, when a title is not available, we generate a title-like query, which is expected to achieve the same effect as a title. Our title-like query consists of the topic and keywords to serve as information on the main topic or theme of the document. We conduct experiments in both Korean and English languages, evaluating the performance of MRCSum on datasets comprising both long and short summaries. Our results demonstrate the effectiveness of MRCSum in extractive summarization, showcasing its ability to generate concise and informative summaries with or without explicit titles. Furthermore, our MRCSum outperforms existing models by capturing the essence of the document content and producing more coherent summaries.

To better handle commonsense knowledge, which is difficult to acquire in ordinary training of language models, commonsense knowledge graphs and commonsense knowledge models have been constructed. The former manually and symbolically represents commonsense, and the latter stores these graphs’ knowledge in the models’ parameters. However, the existing commonsense knowledge models that deal with events do not consider granularity or time axes. In this paper, we propose a time-aware commonsense knowledge model, TaCOMET. The construction of TaCOMET consists of two steps. First, we create TimeATOMIC using ChatGPT, which is a commonsense knowledge graph with time. Second, TaCOMET is built by continually finetuning an existing commonsense knowledge model on TimeATOMIC. TimeATOMIC and continual finetuning let the model make more time-aware generations with rich commonsense than the existing commonsense models. We also verify the applicability of TaCOMET on a robotic decision-making task. TaCOMET outperformed the existing commonsense knowledge model when proper times are input. Our dataset and models will be made publicly available.

Parliamentary debates constitute a substantial and somewhat underutilized reservoir of publicly available written content. Despite their potential, the Italian parliamentary documents remain largely unexplored and most importantly inaccessible in their original paper-based form. In this paper we attempt to transform these valuable historical documents into IPSA, a digitally readable structured corpus containing speeches, reports of the Standing Committees, and law proposals spanning 175 years of Italian history, from the issuing of the Statuto Albertino in 1848, up to the present day. At first, the PDF documents, available on the official websites of Senato della Repubblica and Camera dei Deputati, the two chambers that form the Italian Parliament, are digitized using Optical Character Recognition (OCR) techniques. Then, the speeches are tagged with the corresponding speakers. The final dataset is released both in textual and structured format.

This work explores the intrinsic limitations of the popular one-hot encoding method in classification of intents when detection of out-of-scope (OOS) inputs is required. Although recent work has shown that there can be significant improvements in OOS detection when the intent classes are represented as dense-vectors based on domain-specific knowledge, we argue in this paper that such gains are more likely due to advantages of the much richer topologies that can be created with dense vectors compared to the equidistant class representation assumed by one-hot encodings. We start by demonstrating how dense-vector encodings are able to create OOS spaces with much richer topologies. Then, we show empirically, using four standard intent classification datasets, that knowledge-free, randomly generated dense-vector encodings of intent classes can yield over 20% gains over one-hot encodings, producing better systems for open-world classification tasks, mostly from improvements in OOS detection.

While human values play a crucial role in making arguments persuasive, we currently lack the necessary extensive datasets to develop methods for analyzing the values underlying these arguments on a large scale. To address this gap, we present the Touché23-ValueEval dataset, an expansion of the Webis-ArgValues-22 dataset. We collected and annotated an additional 4780 new arguments, doubling the dataset’s size to 9324 arguments. These arguments were sourced from six diverse sources, covering religious texts, community discussions, free-text arguments, newspaper editorials, and political debates. Each argument is annotated by three crowdworkers for 54 human values, following the methodology established in the original dataset. The Touché23-ValueEval dataset was utilized in the SemEval 2023 Task 4. ValueEval: Identification of Human Values behind Arguments, where an ensemble of transformer models demonstrated state-of-the-art performance. Furthermore, our experiments show that a fine-tuned large language model, Llama-2-7B, achieves comparable results.

We present a preview of the Syntactic Acceptability Dataset, a resource being designed for both syntax and computational linguistics research. In its current form, the dataset comprises 1,000 English sequences from the syntactic discourse: Half from textbooks and half from the journal Linguistic Inquiry, the latter to ensure a representation of the contemporary discourse. Each entry is labeled with its grammatical status (“well-formedness” according to syntactic formalisms) extracted from the literature, as well as its acceptability status (“intuitive goodness” as determined by native speakers) obtained through crowdsourcing, with highest experimental standards. Even in its preliminary form, this dataset stands as the largest of its kind that is publicly accessible. We also offer preliminary analyses addressing three debates in linguistics and computational linguistics: We observe that grammaticality and acceptability judgments converge in about 83% of the cases and that “in-betweenness” occurs frequently. This corroborates existing research. We also find that while machine learning models struggle with predicting grammaticality, they perform considerably better in predicting acceptability. This is a novel finding. Future work will focus on expanding the dataset.