COLING 2024 论文列表

This paper presents the Slovak Autistic and Non-Autistic Child Speech Corpus, which consists of audio-recordings and transcripts of collaborative, task-oriented conversations between children (with or without autism spectrum disorder, ASD) and a non-autistic adult experimenter. The task used to elicit this corpus was the Maps task. This corpus was primarily recorded to investigate lexical alignment, but can also be used to study other conversation coordination strategies and behaviours. Scores on various standardised psychometric tests, such as those measuring IQ, executive functioning, and theory of mind, are included for each participant. In total, the corpus contains over 15 hours of speech. This relatively large database contains a non-Germanic language and can be shared with any qualified researcher, making it a valuable resource for replication of existing findings regarding communication and ASD as well as future research into communication between individuals with and without ASD.

We present the SAMER Corpus, the first manually annotated Arabic parallel corpus for text simplification targeting school-aged learners. Our corpus comprises texts of 159K words selected from 15 publicly available Arabic fiction novels most of which were published between 1865 and 1955. Our corpus includes readability level annotations at both the document and word levels, as well as two simplified parallel versions for each text targeting learners at two different readability levels. We describe the corpus selection process, and outline the guidelines we followed to create the annotations and ensure their quality. Our corpus is publicly available to support and encourage research on Arabic text simplification, Arabic automatic readability assessment, and the development of Arabic pedagogical language technologies.

Post-hoc explanation methods are an important tool for increasing model transparency for users. Unfortunately, the currently used methods for attributing token importance often yield diverging patterns. In this work, we study potential sources of disagreement across methods from a linguistic perspective. We find that different methods systematically select different classes of words and that methods that agree most with other methods and with humans display similar linguistic preferences. Token-level differences between methods are smoothed out if we compare them on the syntactic span level. We also find higher agreement across methods by estimating the most important spans dynamically instead of relying on a fixed subset of size k. We systematically investigate the interaction between k and spans and propose an improved configuration for selecting important tokens.

Recent advancements in spontaneous text-to-speech (TTS) have enabled the realistic synthesis of creaky voice, a voice quality known for its diverse pragmatic and paralinguistic functions. In this study, we used synthesized creaky voice in perceptual tests, to explore how listeners without formal training perceive two distinct types of creaky voice. We annotated a spontaneous speech corpus using creaky voice detection tools and modified a neural TTS engine with a creaky phonation embedding to control the presence of creaky phonation in the synthesized speech. We performed an objective analysis using a creak detection tool which revealed significant differences in creaky phonation levels between the two creaky voice types and modal voice. Two subjective listening experiments were performed to investigate the effect of creaky voice on perceived certainty, valence, sarcasm, and turn finality. Participants rated non-positional creak as less certain, less positive, and more indicative of turn finality, while positional creak was rated significantly more turn final compared to modal phonation.

The dearth of literature combining hypothesis-making and collaborative problem solving presents a problem in the investigation into how hypotheses are generated in group environments. A new dataset, the Resolving Investigative hyPotheses (RIP) corpus, is introduced to address this issue. The corpus uses the fictionalised environment of a murder investigation game. An artificial environment restricts the number of possible hypotheses compared to real-world situations, allowing a deeper dive into the data. In three groups of three, participants collaborated to solve the mystery: two groups came to the wrong conclusion in different ways, and one succeeded in solving the game. RIP is a 49k-word dialogical corpus, consisting of three sub-corpora, annotated for argumentation and discourse structure on the basis of Inference Anchoring Theory. The corpus shows the emergent roles individuals took on and the strategies the groups employed, showing what can be gained through a deeper exploration of this domain. The corpus bridges the gap between these two areas – hypothesis generation and collaborative problem solving – by using an environment rich with potential for hypothesising within a highly collaborative space.

The paper presents a new training dataset of sentences in 7 languages, manually annotated for sentiment, which are used in a series of experiments focused on training a robust sentiment identifier for parliamentary proceedings. The paper additionally introduces the first domain-specific multilingual transformer language model for political science applications, which was additionally pre-trained on 1.72 billion words from parliamentary proceedings of 27 European parliaments. We present experiments demonstrating how the additional pre-training on parliamentary data can significantly improve the model downstream performance, in our case, sentiment identification in parliamentary proceedings. We further show that our multilingual model performs very well on languages not seen during fine-tuning, and that additional fine-tuning data from other languages significantly improves the target parliament’s results. The paper makes an important contribution to multiple disciplines inside the social sciences, and bridges them with computer science and computational linguistics. Lastly, the resulting fine-tuned language model sets up a more robust approach to sentiment analysis of political texts across languages, which allows scholars to study political sentiment from a comparative perspective using standardized tools and techniques.

Parallel corpora are still scarce for most of the world’s language pairs. The situation is by no means different for regional languages of France. In addition, adequate web interfaces facilitate and encourage the use of parallel corpora by target users, such as language learners and teachers, as well as linguists. In this paper, we describe ParCoLab, a parallel corpus and a web platform for querying the corpus. From its onset, ParCoLab has been geared towards lower-resource languages, with an initial corpus in Serbian, along with French and English (later Spanish). We focus here on the extension of ParCoLab with a parallel corpus for four regional languages of France: Alsatian, Corsican, Occitan and Poitevin-Saintongeais. In particular, we detail criteria for choosing texts and issues related to their collection. The new parallel corpus contains more than 20k tokens per regional language.

Most existing methods of Out-of-Domain (OOD) intent classification rely on extensive auxiliary OOD corpora or specific training paradigms. However, they are underdeveloped in the underlying principle that the models should have differentiated confidence in In- and Out-of-domain intent. In this work, we shed light on the fundamental cause of model overconfidence on OOD and demonstrate that calibrated subnetworks can be uncovered by pruning the overparameterized model. Calibrated confidence provided by the subnetwork can better distinguish In- and Out-of-domain, which can be a benefit for almost all post hoc methods. In addition to bringing fundamental insights, we also extend the Lottery Ticket Hypothesis to open-world scenarios. We conduct extensive experiments on four real-world datasets to demonstrate our approach can establish consistent improvements compared with a suite of competitive baselines.

The paper reports on the first results of the design and implementation of a new digital tool for romance philology: the digital Onomastic Repertoire for the medieval French romance (12th-15th centuries). This tool, projected with a modular and integrable architecture, was implemented from a selection of romances, the corpus of the Medieval French Roman d’Alexandre. After introducing the peculiarities of the onomastic system in the Middle Ages (and, more generally, the peculiarities of medieval literary texts), the paper describes 1) the methodological challenges faced in the preparatory work, illustrates and comments on the first results achieved and 2) the design and implementation of the first integrated system for the interactive creation of the Onomastic Repertoire of the romaN d’AlexandRE (ORNARE), and 3) the current research output in terms of both a digital edition and the digital onomastic index of the corpus.

We present an extension of the Low Saxon Universal Dependencies dataset and discuss a few annotation-related challenges. Low Saxon is a West-Germanic low-resource language that lacks a common standard and therefore poses challenges for NLP. The 1,000 sentences in our dataset cover the last 200 years and 8 of the 9 major dialects. They are presented both in original and in normalised spelling and two lemmata are provided: A Modern Low Saxon lemma and a Middle Low Saxon lemma. Several annotation-related issues result from dialectal variation in morphological categories, and we explain differences in the pronoun, gender, case, and mood system. Furthermore, we take up three syntactic constructions that do not occur in Standard Dutch or Standard German: the possessive dative, pro-drop in pronominal adverbs, and complementiser doubling in subordinate interrogative clauses. These constructions are also rare in the other Germanic UD datasets and have not always been annotated consistently.

Pose estimation keypoints are widely used in sign language recognition (SLR) as a means of generalising to unseen signers. Despite the advantages of keypoints, SLR models struggle to achieve high recognition accuracy for many signed languages due to the large degree of variability between occurrences of the same signs, the lack of large datasets and the imbalanced nature of the data therein. In this paper we seek to provide a deeper analysis into the ways that these keypoints are used by models in order to determine which are most informative to SLR, identify potentially redundant ones and investigate whether keypoints that are central to differentiating signs in practice are being effectively used as expected by models.

Alzheimer’s disease (AD) represents a major problem for society and a heavy burden for those affected. The study of changes in speech offers a potential means for large-scale AD screening that is non-invasive and inexpensive. Automatic Speech Recognition (ASR) is necessary for a fully automated system. We compare different ASR systems in terms of Word Error Rate (WER) using a publicly available benchmark dataset of speech recordings of AD patients and controls. Furthermore, this study is the first to quantify how popular linguistic features change when replacing manual transcriptions with ASR output. This contributes to the understanding of linguistic features in the context of AD detection. Moreover, we investigate how ASR affects AD classification performance by implementing two popular approaches: A fine-tuned BERT model, and Random Forest on popular linguistic features. Our results show best classification performance when using manual transcripts, but the degradation when using ASR is not dramatic. Performance stays strong, achieving an AUROC of 0.87. Our BERT-based approach is affected more strongly by ASR transcription errors than the simpler and more explainable approach based on linguistic features.

Stance as an expression of an author’s standpoint and as a means of communication has long been studied by computational linguists. Automatically identifying the stance of a subject toward an object is an active area of research in natural language processing. Significant work has employed topics and claims as the object of stance, with frames of communication becoming more recently considered as alternative objects of stance. However, little attention has been paid to finding what are the benefits and what are the drawbacks when inferring the stance of a text towards different possible stance objects. In this paper we seek to answer this question by analyzing the implied knowledge and the judgments required when deciding the stance of a text towards each stance object type. Our analysis informed experiments with models capable of inferring the stance of a text towards any of the stance object types considered, namely topics, claims, and frames of communication. Experiments clearly indicate that it is best to infer the stance of a text towards a frame of communication, rather than a claim or a topic. It is also better to infer the stance of a text towards a claim rather than a topic. Therefore we advocate that rather than continuing efforts to annotate the stance of texts towards topics, it is better to use those efforts to produce annotations towards frames of communication. These efforts will allow us to better capture the stance towards claims and topics as well.

The Igbo language is facing a risk of becoming endangered, as indicated by a 2025 UNESCO study. This highlights the need to develop language technologies for Igbo to foster communication, learning and preservation. To create robust, impactful, and widely adopted language technologies for Igbo, it is essential to incorporate the multi-dialectal nature of the language. The primary obstacle in achieving dialectal-aware language technologies is the lack of comprehensive dialectal datasets. In response, we present the IgboAPI dataset, a multi-dialectal Igbo-English dictionary dataset, developed with the aim of enhancing the representation of Igbo dialects. Furthermore, we illustrate the practicality of the IgboAPI dataset through two distinct studies: one focusing on Igbo semantic lexicon and the other on machine translation. In the semantic lexicon project, we successfully establish an initial Igbo semantic lexicon for the Igbo semantic tagger, while in the machine translation study, we demonstrate that by finetuning existing machine translation systems using the IgboAPI dataset, we significantly improve their ability to handle dialectal variations in sentences.

Creating language technology based on language data has become very popular with the recent advances of large language models and neural network technologies. This makes language resources very valuable, and especially in case of indigenous languages, the scarce resources are even more precious. Given the good results of simply fetching everything you can from the internet and feeding it to neural networks in English, there has been more work on doing the same for all languages. However, indigenous language resources as they are on the web are not comparable in that they would encode the most recent normativised language in all its aspects. This problematic is further due to not understanding the texts input to models or output by models by the people who work on them. Corpora also have intelligent property rights and copyrights that are not respected. Furthermore, the web is filled with the result of language model -generated texts. In this article we describe an ethical and sustainable way to work with indigenous languages.

Massively multilingual models can process text in several languages relying on a shared set of parameters; however, little is known about the encoding of multilingual information in single network units. In this work, we study how two semantic variables, namely valence and arousal, are processed in the latent dimensions of mBERT and XLM-R across 13 languages. We report a significant cross-lingual overlap in the individual neurons processing affective information, which is more pronounced when considering XLM-R vis-à-vis mBERT. Furthermore, we uncover a positive relationship between cross-lingual alignment and performance, where the languages that rely more heavily on a shared cross-lingual neural substrate achieve higher performance scores in semantic encoding.

We propose an approach that improves the performance of VMT (Video-guided Machine Translation) models, which integrate text and video modalities. We experiment with the MAD (Movie Audio Descriptions) dataset, a new dataset which contains transcribed audio descriptions of movies. We find that the MAD dataset is more lexically rich than the VATEX dataset (the current VMT baseline), and we experiment with MAD pretraining to improve performance on the VATEX dataset. We experiment with two different video encoder architectures: a Conformer (Convolution-augmented Transformer) and a Transformer. Additionally, we conduct experiments by masking the source sentences to assess the degree to which the performance of both architectures improves due to pretraining on additional video data. Finally, we conduct an analysis of the transfer learning potential of a video dataset and compare it to pretraining on a text-only dataset. Our findings demonstrate that pretraining with a lexically rich dataset leads to significant improvements in model performance when models use both text and video modalities.

Can emojis be composed to convey intricate meanings like English phrases? As a pioneering study, we present the Emoji-Lexical Composition (ELCo) dataset, a new resource that offers parallel annotations of emoji sequences corresponding to English phrases. Our dataset contains 1,655 instances, spanning 209 diverse concepts from tangible ones like “right man” (✔️👨) to abstract ones such as “full attention” (🧐✍️, illustrating a metaphoric composition of a focusing face and writing hand). ELCo enables the analysis of the patterns shared between emoji and lexical composition. Through a corpus study, we discovered that simple strategies like direct representation and reduplication are sufficient for conveying certain concepts, but a richer, metaphorical strategy is essential for expressing more abstract ideas. We further introduce an evaluative task, Emoji-based Textual Entailment (EmoTE), to assess the proficiency of NLP models in comprehending emoji compositions. Our findings reveals the challenge of understanding emoji composition in a zero-shot setting for current models, including ChatGPT. Our analysis indicates that the intricacy of metaphorical compositions contributes to this challenge. Encouragingly, models show marked improvement when fine-tuned on the ELCo dataset, with larger models excelling in deciphering nuanced metaphorical compositions.

In the realm of human communication, feedback plays a pivotal role in shaping the dynamics of conversations. This study delves into the multifaceted relationship between listener feedback, narration quality and distraction effects. We present an analysis conducted on the SMYLE corpus, specifically enriched for this study, where 30 dyads of participants engaged in 1) face-to-face storytelling (8.2 hours) followed by 2) a free conversation (7.8 hours). The storytelling task unfolds in two conditions, where a storyteller engages with either a “normal” or a “distracted” listener. Examining the feedback impact on storytellers, we discover a positive correlation between the frequency of specific feedback and the narration quality in normal conditions, providing an encouraging conclusion regarding the enhancement of interaction through specific feedback in distraction-free settings. In contrast, in distracted settings, a negative correlation emerges, suggesting that increased specific feedback may disrupt narration quality, underscoring the complexity of feedback dynamics in human communication. The contribution of this paper is twofold: first presenting a new and highly enriched resource for the analysis of discourse phenomena in controlled and normal conditions; second providing new results on feedback production, its form and its consequence on the discourse quality (with direct applications in human-machine interaction).