COLING 2024 Papers

Sign language is the primary communication medium for people who are deaf or have hearing loss. However, given the divergent range of sensory abilities of these individuals, there is a communication gap that needs to be addressed. In this paper, we present action-concentrated embedding (ACE), which is a novel sign token embedding framework. Additionally, to provide a more structured foundation for sign language analysis, we introduce a dedicated notation system tailored for sign language that endeavors to encapsulate the nuanced gestures and movements that are integral with sign communication. The proposed ACE approach tracks a signer’s actions based on human posture estimation. Tokenizing these actions and capturing the token embedding using a short-time Fourier transform encapsulates the time-based behavioral changes. Hence, ACE offers input embedding to translate sign language into natural language sentences. When tested against a disaster sign language dataset using automated machine translation measures, ACE notably surpasses prior research in terms of translation capabilities, improving the performance by up to 5.79% for BLEU-4 and 5.46% for ROUGE-L metric.

Sarcasm primarily involves saying something but “meaning the opposite” or “meaning something completely different” in order to convey a particular tone or mood. In both the above cases, the “meaning” is reflected by the communicative intention of the speaker, known as dialogue acts. In this paper, we seek to investigate a novel phenomenon of analyzing sarcasm in the context of dialogue acts with the hypothesis that the latter helps to understand the former better. Toward this aim, we extend the multi-modal MUStARD dataset to enclose dialogue acts for each dialogue. To demonstrate the utility of our hypothesis, we develop a dialogue act-aided multi-modal transformer network for sarcasm identification (MM-SARDAC), leveraging interrelation between these tasks. In addition, we introduce an order-infused, multi-modal infusion mechanism into our proposed model, which allows for a more intuitive combined modality representation by selectively focusing on relevant modalities in an ordered manner. Extensive empirical results indicate that dialogue act-aided sarcasm identification achieved better performance compared to performing sarcasm identification alone. The dataset and code are available at https://github.com/mohit2b/MM-SARDAC.

Empathy is critical for effective communication and mental health support, and in many online health communities people anonymously engage in conversations to seek and provide empathetic support. The ability to automatically recognize and detect empathy contributes to the understanding of human emotions expressed in text, therefore advancing natural language understanding across various domains. Existing empathy and mental health-related corpora focus on broader contexts and lack domain specificity, but similarly to other tasks (e.g., learning distinct patterns associated with COVID-19 versus skin allergies in clinical notes), observing empathy within different domains is crucial to providing tailored support. To address this need, we introduce AcnEmpathize, a dataset that captures empathy expressed in acne-related discussions from forum posts focused on its emotional and psychological effects. We find that transformer-based models trained on our dataset demonstrate excellent performance at empathy classification. Our dataset is publicly released to facilitate analysis of domain-specific empathy in online conversations and advance research in this challenging and intriguing domain.

This paper introduces the licence-based access control framework developed by the Language Data Commons of Australia (LDaCA) for a range of language collections, with examples given of implementation for significant Indigenous and Australian English collections. Language collections may be curated for many reasons, such as documentation for language revival, for research, security or commercial purposes. Some language collections are created with the intention of being “Open Access”; publicly available with no restriction. Other collections require that access be limited to individuals or groups of people, either at the collection level or at the level of individual items, such as a recording. To facilitate access, while respecting the intended access conditions for a collection, or collection items, some form of user identification and authorisation process is typically required. The access control framework described in this paper is based upon descriptions of access conditions in easy-to-read licences which are stored alongside data files in the collections; and is implemented using identity-based authentication and authorisation systems where required. The framework accommodates accessibility needs from unrestricted to extremely limited access, is dynamic, and able to be modified in response to changes in access needs. Storing licences with the data is a significant development in separating language data and access requirements from access infrastructure.

This paper introduces a new task, abstractive multi-video captioning, which focuses on abstracting multiple videos with natural language. Unlike conventional video captioning tasks generating a specific caption for a video, our task generates an abstract caption of the shared content in a video group containing multiple videos. To address our task, models must learn to understand each video in detail and have strong abstraction abilities to find commonalities among videos. We construct a benchmark dataset for abstractive multi-video captioning named AbstrActs. AbstrActs contains 13.5k video groups and corresponding abstract captions. As abstractive multi-video captioning models, we explore two approaches: end-to-end and cascade. For evaluation, we proposed a new metric, CocoA, which can evaluate the model performance based on the abstractness of the generated captions. In experiments, we report the impact of the way of combining multiple video features, the overall model architecture, and the number of input videos.

Pre-trained Language Models have been shown to be able to emulate deductive reasoning in natural language. However, PLMs are easily affected by irrelevant information (e.g., entity) in instance-level proofs when learning deductive reasoning. To address this limitation, we propose an Abstract-level Deductive Reasoner (ADR). ADR is trained to predict the abstract reasoning proof of each sample, which guides PLMs to learn general reasoning patterns rather than instance-level knowledge. Experimental results demonstrate that ADR significantly reduces the impact of PLMs learning instance-level knowledge (over 70%).

Effective feature representations play a critical role in enhancing the performance of text generation models that rely on deep neural networks. However, current approaches suffer from several drawbacks, such as the inability to capture the deep semantics of language and sensitivity to minor input variations, resulting in significant changes in the generated text. In this paper, we present a novel solution to these challenges by employing a mixture of experts, multiple encoders, to offer distinct perspectives on the emotional state of the user’s utterance while simultaneously enhancing performance. We propose an end-to-end model architecture called ASEM that performs emotion analysis on top of sentiment analysis for open-domain chatbots, enabling the generation of empathetic responses that are fluent and relevant. In contrast to traditional attention mechanisms, the proposed model employs a specialized attention strategy that uniquely zeroes in on sentiment and emotion nuances within the user’s utterance. This ensures the generation of context-rich representations tailored to the underlying emotional tone and sentiment intricacies of the text. Our approach outperforms existing methods for generating empathetic embeddings, providing empathetic and diverse responses. The performance of our proposed model significantly exceeds that of existing models, enhancing emotion detection accuracy by 6.2% and lexical diversity by 1.4%. ASEM code is released at https://github.com/MIRAH-Official/Empathetic-Chatbot-ASEM.git

We introduce the Alternating Reading Task (ART) Corpus, a collection of dyadic sentence reading for studying the entrainment and imitation behaviour in speech communication. The ART corpus features three experimental conditions - solo reading, alternating reading, and deliberate imitation - as well as three subcorpora encompassing French-, Italian-, and Slovak-accented English. This design allows systematic investigation of speech entrainment in a controlled and less spontaneous setting. Alongside detailed transcriptions, it includes English proficiency scores, demographics, and in-experiment questionnaires for probing linguistic, personal and interpersonal influences on entrainment. Our presentation covers its design, collection, annotation processes, initial analysis, and future research prospects.

ARBRES is an ongoing project of open science implemented as a platform (“wikigrammar”) documenting both the Breton language itself and the state of research and engineering work in linguistics and NLP. Along its nearly 15 years of operation, it has aggregated a wealth of linguistic data in the form of interlinear glosses with translations illustrating lexical items, grammatical features, dialectal variations... While these glosses were primarily meant for human consumption, their volume and the regular format imposed by the wiki engine used for the website also make them suitable for machine processing. ARBRES Kenstur is a new parallel corpus derived from the glosses in ARBRES, including about 5k phrases and sentences in Breton along with translations in standard French. The nature of the original data — sourced from field linguistic inquiries meant to document the structure of Breton — leads to a resource that is mechanically more concerned with the internal variations of the language and rare phenomena than typical parallel corpora. Preliminaries experiments in using this corpus show that it can help improve machine translation for Breton, demonstrating that sourcing data from field linguistic documentation can be a way to help provide NLP tools for minority and low-resource languages.

Long-form numerical reasoning aims to generate a reasoning program to calculate the answer for a given question. Previous work followed a retriever-generator framework, where the retriever selects key facts from a long-form document, and the generator generates a reasoning program based on the retrieved facts. However, they treated all facts equally without considering the different contributions of facts with and without numerical information. Furthermore, they ignored program consistency, leading to the wrong punishment of programs that differed from the ground truth. In order to address these issues, we proposed APOLLO (An optimized training aPproach fOr Long-form numericaL reasOning), to improve long-form numerical reasoning. APOLLO includes a number-aware negative sampling strategy for the retriever to discriminate key numerical facts, and a consistency-based reinforcement learning with target program augmentation for the generator to ultimately increase the execution accuracy. Experimental results on the FinQA and ConvFinQA leaderboards verify the effectiveness of our proposed methods, achieving the new state-of-the-art.

Contextualised Language Models (LM) improve on traditional word embeddings by encoding the meaning of words in context. However, such models have also made it possible to learn high-quality decontextualised concept embeddings. Three main strategies for learning such embeddings have thus far been considered: (i) fine-tuning the LM to directly predict concept embeddings from the name of the concept itself, (ii) averaging contextualised representations of mentions of the concept in a corpus, and (iii) encoding definitions of the concept. As these strategies have complementary strengths and weaknesses, we propose to learn a unified embedding space in which all three types of representations can be integrated. We show that this allows us to outperform existing approaches in tasks such as ontology completion, which heavily depends on access to high-quality concept embeddings. We furthermore find that mentions and definitions are well-aligned in the resulting space, enabling tasks such as target sense verification, even without the need for any fine-tuning.

This paper presents ALLIES, a meta corpus which gathers and extends existing French corpora collected from radio and TV shows. The corpus contains 1048 audio files for about 500 hours of speech. Agglomeration of data is always a difficult issue, as the guidelines used to collect, annotate and transcribe speech are generally different from one corpus to another. ALLIES intends to homogenize and correct speaker labels among the different files by integrated human feedback within a speaker verification system. The main contribution of this article is the design of a protocol in order to evaluate properly speech segmentation (including music and overlap detection), speaker diarization, speech transcription and speaker change detection. As part of it, a test partition has been carefully manually 1) segmented and annotated according to speech, music, noise, speaker labels with specific guidelines for overlap speech, 2) orthographically transcribed. This article also provides as a second contribution baseline results for several speech processing tasks.

Introducing ADEA: a German dataset that captures online dialogues and focuses on ethical issues related to future AI applications. This dataset, which includes over 2800 labeled user utterances on four different topics, is specifically designed for the training of chatbots that can navigate the complexities of real-world ethical AI conversations. The creation of these dialogues is the result of two carefully conducted studies in which university students interacted with an argumentative dialogue system. A fundamental part of our methodology is the use of German argument graphs. These graphs not only form the knowledge base of the dialogue system but also serve as an effective annotation scheme for the dialogues. Apart from the introduction of the dataset and the argument graphs, we provide a preliminary benchmark using GPT-4 via the OpenAI API. This provides researchers with a concrete reference point while demonstrating the potential of our dataset. We make our dataset and argument graphs available at https://github.com/HaupChris/ADEA-Dialogue-Dataset.

Natural Language Processing (NLP) models tend to inherit and amplify stereotypical biases present in their training data, leading to harmful societal consequences. Current efforts to rectify these biases typically revolve around making models oblivious to bias, which is at odds with the idea that humans require increased awareness to tackle these biases better. This prompts a fundamental research question: are bias-oblivious models the only viable solution to combat stereotypical biases? This paper answers this question by proposing the Agency-BeLiefs Embedding (ABLE) model, a novel approach that actively encodes stereotypical biases into the embedding space. ABLE draws upon social psychological theory to acquire and represent stereotypical biases in the form of agency and belief scores rather than directly representing stereotyped groups. Our experimental results showcase ABLE’s effectiveness in learning agency and belief stereotypes while preserving the language model’s proficiency. Furthermore, we underscore the practical significance of incorporating stereotypes within the ABLE model by demonstrating its utility in various downstream tasks. Our approach exemplifies the potential benefits of addressing bias through awareness, as opposed to the prevailing approach of mitigating bias through obliviousness.

The recent emergence of Large Language Models (LLMs) has enabled significant advances in the field of Natural Language Processing (NLP). While these new models have demonstrated superior performance on various tasks, their application and potential are still underexplored, both in terms of the diversity of tasks they can handle and their domain of application. In this context, we evaluate four state-of-the-art instruction-tuned LLMs (ChatGPT, Flan-T5 UL2, Tk-Instruct, and Alpaca) on a set of 13 real-world clinical and biomedical NLP tasks in English, including named-entity recognition (NER), question-answering (QA), relation extraction (RE), and more. Our overall results show that these evaluated LLMs approach the performance of state-of-the-art models in zero- and few-shot scenarios for most tasks, particularly excelling in the QA task, even though they have never encountered examples from these tasks before. However, we also observe that the classification and RE tasks fall short of the performance achievable with specifically trained models designed for the medical field, such as PubMedBERT. Finally, we note that no single LLM outperforms all others across all studied tasks, with some models proving more suitable for certain tasks than others.

We describe ongoing work for developing a workflow for the applied use case of classifying diachronic and regional language variation in Pre-Modern Slavic texts. The data were obtained via handwritten text recognition (HTR) on medieval manuscripts and printings and partly by manual transcription. Our goal is to develop a workflow for such historical language data, covering HTR-postprocessing, annotating and classifying the digitized texts. We test and adapt existing language resources to fit the pipeline with low-barrier tooling, accessible for Humanists with limited experience in research data infrastructures, computational analysis or advanced methods of natural language processing (NLP). The workflow starts by addressing ground truth (GT) data creation for diagnosing and correcting HTR errors via string metrics and data-driven methods. On GT and on HTR data, we subsequently show classification results using transfer learning on sentence-level text snippets. Next, we report on our token-level data labeling efforts. Each step of the workflow is complemented with describing current limitations and our corresponding work in progress.

This paper presents a new web portal with information about the state of the art of natural language processing tasks in Spanish. It provides information about forums, competitions, tasks and datasets in Spanish, that would otherwise be spread in multiple articles and web sites. The portal consists of overview pages where information can be searched for and filtered by several criteria and individual pages with detailed information and hyperlinks to facilitate navigation. Information has been manually curated from publications that describe competitions and NLP tasks from 2013 until 2023 and will be updated as new tasks appear. A total of 185 tasks and 128 datasets from 94 competitions have been introduced.

In this work we present two datasets for the development of virtual patients and the first evaluation results. We firstly introduce a Spanish corpus of medical dialogue questions annotated with intents, built upon prior research in French. We also propose a second dataset of dialogues using a novel annotation approach that involves doctor questions, patient answers, and corresponding clinical records, organized as triples of the form (clinical report, question, patient answer). This way, the doctor-patient conversation is modeled as a question-answering system that tries to find responses to questions taking a clinical record as input. This approach can help to eliminate the need for manually structured patient records, as commonly used in previous studies, thereby expanding the pool of diverse virtual patients available. Leveraging these annotated corpora, we develop and assess an automatic system designed to answer medical dialogue questions posed by medical students to simulated patients in medical exams. Our approach demonstrates robust generalization, relying solely on medical records to generate new patient cases. The two datasets and the code will be freely available for the research community.

This paper discusses the challenges non-prescriptive language uses in chatbot communication create for Semantic Parsing (SP). To help SP developers improve their systems, we propose a flexible error typology based on an analysis of a sample of non-prescriptive language uses mined from a domain-specific chatbot logs. This typology is not tied to any specific language model. We also present a framework for automatically mapping these errors to the typology. Finally, we show how our framework can help evaluate SP systems from a linguistic robustness perspective. Our framework can be expanded to include new classes of errors across different domains and user demographics.