COLING 2024 Papers

Many systems rely on the ability to effectively search through databases of personal and organization entity names in multiple writing scripts. Despite this, there is a relative lack of research studying this problem in isolation. In this work, we discuss this problem in detail and support future research by publishing what we believe is the first comprehensive dataset designed for this task. Additionally, we present a number of baselines against which future work can be compared; among which, we describe a neural solution based on ByT5 (Xue et al. 2022) which demonstrates up to a 12% performance gain over preexisting baselines, indicating that there remains much room for improvement in this space.

Dialogue datasets are crucial for deep learning-based task-oriented dialogue system research. While numerous English language multi-domain task-oriented dialogue datasets have been developed and contributed to significant advancements in task-oriented dialogue systems, such a dataset does not exist in Japanese, and research in this area is limited compared to that in English. In this study, towards the advancement of research and development of task-oriented dialogue systems in Japanese, we constructed JMultiWOZ, the first Japanese language large-scale multi-domain task-oriented dialogue dataset. Using JMultiWOZ, we evaluated the dialogue state tracking and response generation capabilities of the state-of-the-art methods on the existing major English benchmark dataset MultiWOZ2.2 and the latest large language model (LLM)-based methods. Our evaluation results demonstrated that JMultiWOZ provides a benchmark that is on par with MultiWOZ2.2. In addition, through evaluation experiments of interactive dialogues with the models and human participants, we identified limitations in the task completion capabilities of LLMs in Japanese.

Models, such as BERT, have made a significant breakthrough in the Natural Language Processing (NLP) domain solving 11+ tasks. This is achieved by training on a large scale of unlabelled text resources and leveraging Transformers architecture making it the “Jack of all NLP trades”. However, one of the popular and challenging tasks in Sequence Classification is Short Text Classification (STC). Short Texts face the problem of being short, equivocal, and non-standard. In this paper, we address two major problems: 1. Improving STC tasks performance in Japanese language which consists of many varieties and dialects. 2. Building a light-weight Japanese BERT model with cross-domain functionality and comparable accuracy with State of the Art (SOTA) BERT models. To solve this, we propose a novel cross-domain scalable model called JLBert, which is pre-trained on a rich, diverse and less explored Japanese e-commerce corpus. We present results from extensive experiments to show that JLBert is outperforming SOTA Multilingual and Japanese specialized BERT models on three Short Text datasets by approx 1.5% across various domain.

The detection of hate speech is a subject extensively explored by researchers, and machine learning algorithms play a crucial role in this domain. The existing resources mostly focus on text sequence classification for the task of hate speech detection. However, the target of hateful content is another dimension that has not been studied in details due to the lack of data resources. In this study, we address this gap by introducing a novel tweet dataset for the task of joint learning of hate speech detection and target detection, called JL-Hate, for the tasks of sequential text classification and token classification, respectively. The JL-Hate dataset consists of 1,530 tweets divided equally in English and Turkish languages. Leveraging this dataset, we conduct a series of benchmark experiments. We utilize a joint learning model to concurrently perform sequence and token classification tasks on our data. Our experimental results demonstrate consistent performance with the prevalent studies, both in sequence and token classification tasks.

Large language models (LLMs) have proficiently solved a broad range of tasks with their rich knowledge but often struggle with logical reasoning. To foster the research on logical reasoning, many benchmarks have been proposed so far. However, most of these benchmarks are limited to English, hindering the evaluation of LLMs specialized for each language. To address this, we propose **JFLD** (**J**apanese **F**ormal **L**ogic **D**eduction), a deductive reasoning benchmark for Japanese. JFLD assess whether LLMs can generate logical steps to (dis-)prove a given hypothesis based on a given set of facts. Its key features are assessing pure logical reasoning abilities isolated from knowledge and assessing various reasoning rules. We evaluate various Japanese LLMs and see that they are still poor at logical reasoning, thus highlighting a substantial need for future research.

We present JEMHopQA, a multi-hop QA dataset for the development of explainable QA systems. The dataset consists not only of question-answer pairs, but also of supporting evidence in the form of derivation triples, which contributes to making the QA task more realistic and difficult. It is created based on Japanese Wikipedia using both crowd-sourced human annotation as well as prompting a large language model (LLM), and contains a diverse set of question, answer and topic categories as compared with similar datasets released previously. We describe the details of how we built the dataset as well as the evaluation of the QA task presented by this dataset using GPT-4, and show that the dataset is sufficiently challenging for the state-of-the-art LLM while showing promise for combining such a model with existing knowledge resources to achieve better performance.

Document question answering is a task of question answering on given documents such as reports, slides, pamphlets, and websites, and it is a truly demanding task as paper and electronic forms of documents are so common in our society. This is known as a quite challenging task because it requires not only text understanding but also understanding of figures and tables, and hence visual question answering (VQA) methods are often examined in addition to textual approaches. We introduce Japanese Document Question Answering (JDocQA), a large-scale document-based QA dataset, essentially requiring both visual and textual information to answer questions, which comprises 5,504 documents in PDF format and annotated 11,600 question-and-answer instances in Japanese. Each QA instance includes references to the document pages and bounding boxes for the answer clues. We incorporate multiple categories of questions and unanswerable questions from the document for realistic question-answering applications. We empirically evaluate the effectiveness of our dataset with text-based large language models (LLMs) and multimodal models. Incorporating unanswerable questions in finetuning may contribute to harnessing the so-called hallucination generation.

Neural language models have exhibited outstanding performance in a range of downstream tasks. However, there is limited understanding regarding the extent to which these models internalize syntactic knowledge, so that various datasets have recently been constructed to facilitate syntactic evaluation of language models across languages. In this paper, we introduce JCoLA (Japanese Corpus of Linguistic Acceptability), which consists of 10,020 sentences annotated with binary acceptability judgments. Specifically, those sentences are manually extracted from linguistics textbooks, handbooks and journal articles, and split into in-domain data (86 %; relatively simple acceptability judgments extracted from textbooks and handbooks) and out-of-domain data (14 %; theoretically significant acceptability judgments extracted from journal articles), the latter of which is categorized by 12 linguistic phenomena. We then evaluate the syntactic knowledge of 9 different types of Japanese and multilingual language models on JCoLA. The results demonstrated that several models could surpass human performance for the in-domain data, while no models were able to exceed human performance for the out-of-domain data. Error analyses by linguistic phenomena further revealed that although neural language models are adept at handling local syntactic dependencies like argument structure, their performance wanes when confronted with long-distance syntactic dependencies like verbal agreement and NPI licensing.

Many languages use adpositions (prepositions or postpositions) to mark a variety of semantic relations, with different languages exhibiting both commonalities and idiosyncrasies in the relations grouped under the same lexeme. We present the first Japanese extension of the SNACS framework (Schneider et al., 2018), which has served as the basis for annotating adpositions in corpora from several languages. After establishing which of the set of particles (joshi) in Japanese qualify as case markers and adpositions as defined in SNACS, we annotate 10 chapters (≈10k tokens) of the Japanese translation of Le Petit Prince (The Little Prince), achieving high inter-annotator agreement. We find that, while a majority of the particles and their uses are captured by the existing and extended SNACS annotation guidelines from the previous work, some unique cases were observed. We also conduct experiments investigating the cross-lingual similarity of adposition and case marker supersenses, showing that the language-agnostic SNACS framework captures similarities not clearly observed in multilingual embedding space.

Understanding expressions that refer to the physical world is crucial for such human-assisting systems in the real world, as robots that must perform actions that are expected by users. In real-world reference resolution, a system must ground the verbal information that appears in user interactions to the visual information observed in egocentric views. To this end, we propose a multimodal reference resolution task and construct a Japanese Conversation dataset for Real-world Reference Resolution (J-CRe3). Our dataset contains egocentric video and dialogue audio of real-world conversations between two people acting as a master and an assistant robot at home. The dataset is annotated with crossmodal tags between phrases in the utterances and the object bounding boxes in the video frames. These tags include indirect reference relations, such as predicate-argument structures and bridging references as well as direct reference relations. We also constructed an experimental model and clarified the challenges in multimodal reference resolution tasks.

Generic commercial language-based assistants have become ubiquitously available, originally in the form of smart speakers and mobile apps, and more recently in the form of systems based on generative AI. At first glance, their capabilities seem remarkable. Speech recognition works well, NLU mostly works, and access to back-end information sources is usually quite good. However, there is still a lot of work to be done. In the area of NLU in particular, focused probes into the capabilities of language-based assistants easily reveal significant areas of brittleness that demonstrate large gaps in their coverage. For example, the straightforward disjunctive query is this monday or tuesday elicited the nonsensical response it’s 2:50 p.m. many consider it to be the afternoon. These gaps are difficult to identify if the development process relies on training the system with an ongoing supply of natural user data, because this natural data can become distorted by a self-reinforcing feedback loop where the system ‘trains’ the user to produce data that works. This paper describes a process for collecting specific kinds of data to uncover these gaps and an annotation scheme for system responses, and includes examples of simple utterances that nonetheless fail to be correctly processed. The systems tested include both Conventional assistants, such as Amazon Alexa and Google Assistant, as well as GenAI systems, including ChatGPT and Bard/Gemini. We claim that these failures are due to a lack of attention to the full spectrum of input possibilities, and argue that systems would benefit from the inclusion of focused manual assessment to directly target likely gaps.

Neural word embeddings have proven valuable in the development of medical applications. However, for the Italian language, there are no publicly available corpora, embeddings, or evaluation resources tailored to this domain. In this paper, we introduce an Italian corpus for the medical domain, that includes texts from Wikipedia, medical journals, drug leaflets, and specialized websites. Using this corpus, we generate neural word embeddings from scratch. These embeddings are then evaluated using standard evaluation resources, that we translated into Italian exploiting the concept graph in the UMLS Metathesaurus. Despite the relatively small size of the corpus, our experimental results indicate that the new embeddings correlate well with human judgments regarding the similarity and the relatedness of medical concepts. Moreover, these medical-specific embeddings outperform a baseline model trained on the full Wikipedia corpus, which includes the medical pages we used. We believe that our embeddings and the newly introduced textual resources will foster further advancements in the field of Italian medical Natural Language Processing.

We introduce IsraParlTweet, a new linked corpus of Hebrew-language parliamentary discussions from the Knesset (Israeli Parliament) between the years 1992-2023 and Twitter posts made by Members of the Knesset between the years 2008-2023, containing a total of 294.5 million Hebrew tokens. In addition to raw text, the corpus contains comprehensive metadata on speakers and Knesset sessions as well as several linguistic annotations. As a result, IsraParlTweet can be used to conduct a wide variety of quantitative and qualitative analyses and provide valuable insights into political discourse in Israel.

Research on automated text summarization typically uses human and automatic evaluation methods. While most recent studies focus on intrinsic evaluation, which assesses the general quality of summaries, e.g. coherence and informativeness, we concentrate on task-based extrinsic evaluation to determine the usefulness of summaries. We incorporate three downstream tasks, namely question answering, text classification, and text similarity assessment, and measure the usefulness of summaries for these tasks by several metrics. Our findings reveal that summaries are generally useful in tasks that require a comprehensive grasp of the text but are less useful in tasks requiring a more specific understanding of the text. We also analyze the usefulness and inherent properties of summaries from different models, and find that fine-tuned models consistently produce more useful summaries across all three tasks. In contrast, zero-shot models tend to lean towards text classification and similarity assessment, providing more general and less detailed summaries. Additionally, we assess the correlation between 14 intrinsic automatic metrics and human judgments. Intrinsic metrics perform well in evaluating summaries for question answering but are less effective in the other two tasks. This highlights the limitations of relying solely on intrinsic metrics for assessing summary performance and usefulness.

Even though various speech data sets are available in Hungarian, there is a lack of a general overview about their types and sizes. To fill in this gap, we provide a survey of available data sets in spoken Hungarian in five categories (e.g., monolingual, Hungarian part of multilingual, pathological, child-related and dialectal collections). In total, the estimated size of available data is about 2800 hours (across 7500 speakers) and it represents a rich spoken language diversity. However, the distribution of the data and its alignment to real-life (e.g. speech recognition) tasks is far from optimal indicating the need for additional larger-scale natural language speech data sets. Our survey presents an overview of available data sets for Hungarian explaining their strengths and weaknesses which is useful for researchers working on Hungarian across disciplines. In addition, our survey serves as a starting point towards a unified foundational speech model specific to Hungarian.

The rise of Modular Deep Learning showcases its potential in various Natural Language Processing applications. Parameter-efficient fine-tuning (PEFT) modularity has been shown to work for various use cases, from domain adaptation to multilingual setups. However, all this work covers the case where the modular components are trained and deployed within one single Pre-trained Language Model (PLM). This model-specific setup is a substantial limitation on the very modularity that modular architectures are trying to achieve. We ask whether current modular approaches are transferable between models and whether we can transfer the modules from more robust and larger PLMs to smaller ones. In this work, we aim to fill this gap via a lens of Knowledge Distillation, commonly used for model compression, and present an extremely straightforward approach to transferring pre-trained, task-specific PEFT modules between same-family PLMs. Moreover, we propose a method that allows the transfer of modules between incompatible PLMs without any change in the inference complexity. The experiments on Named Entity Recognition, Natural Language Inference, and Paraphrase Identification tasks over multiple languages and PEFT methods showcase the initial potential of transferable modularity.

The use of large language models (LLM), especially ChatGPT, to help with research has come into practice. Researchers use it for timely advice and hope to obtain in-depth feedback. However, can LLM be a qualified and reliable reviewer? Although there already exist several review-related datasets, few works have carefully and thoroughly inspected model’s capability as a reviewer, especially the correctness of generated reviews. In this paper, we first evaluate GPT-3.5 and GPT-4 (the current top-performing LLM) on 2 types of tasks under different settings: the score prediction task and the review generation task. In addition, we propose a dataset containing 197 review-revision multiple-choice questions (RR-MCQ) with detailed labels from the review-rebuttal forum in ICLR-2023. By asking questions from technical details to the overall presentation and quality, our RR-MCQ data provides a more complete model ability assessment. The results show that LLM is generally helpful, but great caution is needed as it always makes mistakes. Although it can give passable decisions (> 60% accuracy) on single options, completely correct answers are still rare (about 20%); models are still weak on long paper processing, zero-shot scoring, and giving critical feedback like human reviewers.

Text simplification is a common task where the text is adapted to make it easier to understand. Similarly, text elaboration can make a passage more sophisticated, offering a method to control the complexity of reading comprehension tests. However, text simplification and elaboration tasks are limited to only relatively alter the readability of texts. It is useful to directly modify the readability of any text to an absolute target readability level to cater to a diverse audience. Ideally, the readability of readability-controlled generated text should be independent of the source text. Therefore, we propose a novel readability-controlled text modification task. The task requires the generation of 8 versions at various target readability levels for each input text. We introduce novel readability-controlled text modification metrics. The baselines for this task use ChatGPT and Llama-2, with an extension approach introducing a two-step process (generating paraphrases by passing through the language model twice). The zero-shot approaches are able to push the readability of the paraphrases in the desired direction but the final readability remains correlated with the original text’s readability. We also find greater drops in semantic and lexical similarity between the source and target texts with greater shifts in the readability.

We investigate how gender authorship influences polar, i.e. positive and negative gender reference. Given German-language newspaper texts where the full name of the authors are known and their gender can be inferred from the first names. And given that nouns in the text have gender reference, i.e. are labeled by a gender classifier as female or male denoting nouns. If these nouns carry a polar load, they count towards the gender-specific statistics we are interested in. A polar load is given either via phrase-level sentiment composition, or by a verb-based analysis of the polar role a noun (phrase) plays: is it framed by the verb as a positive or negative actor, or as receiving a positive or negative effect? Also, reported gender-gender relations (in favor, against) might be gender-specific. Statistical hypothesis testing is carried out in order to find out whether significant gender-wise correlations exist. We found that, in fact, gender reference is gender-specific: each gender significantly more often focuses on their own gender than the other one and e.g. positive actorship supremacy is claimed (intra-) gender-wise.