COLING 2024 Papers

Classical bias detection methods used in Machine Learning are themselves biased because of the different confounding variables implied in the assessment of the initial biases. First they are using templates that are syntactically simple and distant from the target data on which the model will deployed. Second, current methods are assessing biases in pre-trained language models or in dataset, but not directly on the fine-tuned classifier that can actually produce harms. We propose a simple method to detect the biases of a specific fine-tuned classifier on any type of unlabeled data. The idea is to study the classifier behavior by creating counterfactual examples directly on the target data distribution and quantify the amount of changes. In this work, we focus on named entity perturbations by applying a Named Entity Recognition on target-domain data and modifying them accordingly to most common names or location of a target group (gender and country), and this for several morphosynctactically different languages spoken in relation with the countries of the target groups. We used our method on two models available open-source that are likely to be deployed by industry, and on two tasks and domains. We first assess the bias of a multilingual sentiment analysis model trained over multiple-languages tweets and available open-source, and then a multilingual stance recognition model trained over several languages and assessed over English language. Finally we propose to link the perplexity of each example with the bias of the model, by looking at the change in label distribution with respect to the language of the target group. Our work offers a fine-grained analysis of the interactions between names and languages, revealing significant biases in multilingual models.

Recent studies on LLMs do not pay enough attention to linguistic and lexical semantic tasks, such as taxonomy learning. In this paper, we explore the capacities of Large Language Models featuring LLaMA-2 and Mistral for several Taxonomy-related tasks. We introduce a new methodology and algorithm for data collection via stochastic graph traversal leading to controllable data collection. Collected cases provide the ability to form nearly any type of graph operation. We test the collected dataset for learning taxonomy structure based on English WordNet and compare different input templates for fine-tuning LLMs. Moreover, we apply the fine-tuned models on such datasets on the downstream tasks achieving state-of-the-art results on the TexEval-2 dataset.

Distinguished from traditional knowledge graphs (KGs), temporal knowledge graphs (TKGs) must explore and reason over temporally evolving facts adequately. However, existing TKG approaches still face two main challenges, i.e., the limited capability to model arbitrary timestamps continuously and the lack of rich inference patterns under temporal constraints. In this paper, we propose an innovative TKGE method (PTBox) via polynomial decomposition-based temporal representation and box embedding-based entity representation to tackle the above-mentioned problems. Specifically, we decompose time information by polynomials and then enhance the model’s capability to represent arbitrary timestamps flexibly by incorporating the learnable temporal basis tensor. In addition, we model every entity as a hyperrectangle box and define each relation as a transformation on the head and tail entity boxes. The entity boxes can capture complex geometric structures and learn robust representations, improving the model’s inductive capability for rich inference patterns. Theoretically, our PTBox can encode arbitrary time information or even unseen timestamps while capturing rich inference patterns and higher-arity relations of the knowledge base. Extensive experiments on real-world datasets demonstrate the effectiveness of our method.

Arabic diacritic recovery i.e. diacritization is necessary for proper vocalization and an enabler for downstream applications such as language learning and text to speech. Diacritics come in two varieties, namely: core-word diacritics and case endings. In this paper we introduce a highly effective morphologically informed character-level model that can recover both types of diacritics simultaneously. The model uses a Recurrent Neural Network (RNN) based architecture that takes in text as a sequence of characters, with markers for morphological segmentation, and outputs a sequence of diacritics. We also introduce a character-based morphological segmentation model that we train for Modern Standard Arabic (MSA) and dialectal Arabic. We demonstrate the efficacy of our diacritization model on Classical Arabic, MSA, and two dialectal (Moroccan and Tunisian) texts. We achieve the lowest reported word-level diacritization error rate for MSA (3.4%), match the best results for Classical Arabic (5.4%), and report competitive results for dialectal Arabic.

Due to the huge scale of literary works, automatic text analysis technologies are urgently needed for literary studies such as Digital Humanities. However, the domain-generality of existing NLP technologies limits their effectiveness on in-depth literary studies. It is valuable to explore how to adapt NLP technologies to the literary-specific tasks. Fictional characters are the most essential elements of a novel, and thus crucial to understanding the content of novels. The prerequisite of collecting a character’s information is to resolve its different representations. It is a specific problem of anaphora resolution which is a classical and open-domain NLP task. We adapt a state-of-the-art anaphora resolution model to resolve character representations in Chinese novels by making some modifications, and train a widely used BERT fine-tuned model for speaker extraction as assistance. We also analyze the challenges and potential solutions for character-resolution in Chinese novels according to the resolution results on a specific Chinese novel.

Empathy is essential in healthcare communication. We introduce an annotation approach that draws on well-established frameworks for clinical empathy and breaking bad news (BBN) conversations for considering the interactive dynamics of discourse relations. We construct Empathy in BBNs, a span-relation task dataset of simulated BBN conversations in German, using our annotation scheme, in collaboration with a large medical school to support research on educational tools for medical didactics. The annotation is based on 1) Pounds (2011)’s appraisal framework for clinical empathy, which is grounded in systemic functional linguistics, and 2) the SPIKES protocol for breaking bad news (Baile et al., 2000), commonly taught in medical didactics training. This approach presents novel opportunities to study clinical empathic behavior and enables the training of models to detect causal relations involving empathy, a highly desirable feature of systems that can provide feedback to medical professionals in training. We present illustrative examples, discuss applications of the annotation scheme, and insights we can draw from the framework.

This paper presents results in transfer-learning metaphor recognition in German. Starting from an English language corpus annotated for metaphor at the sentence level, and its machine-translation to German, we annotate 1000 sentences of the German part to use it as a Gold standard for two different metaphor prediction setups: i) a sequence labeling set-up (on the token-level), and ii) a classification (based on sentences) setup. We test two transfer leaning approaches: i) a group of transformer models, and ii) a technique that utilizes bilingual embeddings together with an RNN classifier. We find out that the transformer models do moderately in a zero-shot scenario (up to 61% F1 for classification) and the embeddings approaches do not even beat the guessing baseline (36% F1 for classification). We use our Gold data to fine-tune the classification tasks on target-language data achieving up to 90% F1 with both, the multilingual BERT and the bilingual embeddings. We also publish the annotated bilingual corpus.

Argument mining aims to detect all possible argumentative components and identify their relationships automatically. As a thriving task in natural language processing, there has been a large amount of corpus for academic study and application development in this field. However, the research in this area is still constrained by the inherent limitations of existing datasets. Specifically, all the publicly available datasets are relatively small in scale, and few of them provide information from other modalities to facilitate the learning process. Moreover, the statements and expressions in these corpora are usually in a compact form, which restricts the generalization ability of models. To this end, we collect a novel dataset AntCritic to serve as a helpful complement to this area, which consists of about 10k free-form and visually-rich financial comments and supports both argument component detection and argument relation prediction tasks. Besides, to cope with the challenges brought by scenario expansion, we thoroughly explore the fine-grained relation prediction and structure reconstruction scheme and discuss the encoding mechanism for visual styles and layouts. On this basis, we design two simple but effective model architectures and conduct various experiments on this dataset to provide benchmark performances as a reference and verify the practicability of our proposed architecture. We release our data and code in this link, and this dataset follows CC BY-NC-ND 4.0 license.

We present the Prague Discourse Treebank 3.0 – a new version of the annotation of discourse relations marked by primary and secondary discourse connectives in the data of the Prague Dependency Treebank. Compared to the previous version (PDiT 2.0), the version 3.0 comes with three types of major updates: (i) it brings a largely revised annotation of discourse relations: pragmatic relations have been thoroughly reworked, many inconsistencies across all discourse types have been fixed and previously unclear cases marked in annotators’ comments have been resolved, (ii) it achieves consistency with a Lexicon of Czech Discourse Connectives (CzeDLex), and (iii) it provides the data not only in its native format (Prague Markup Language, discourse relations annotated at the top of the dependency trees), but also in the Penn Discourse Treebank 3.0 format (plain text plus a stand-off discourse annotation) and sense taxonomy. PDiT 3.0 contains 21,662 discourse relations (plus 445 list relations) in 49 thousand sentences.

Current foundation models have shown impressive performance across various tasks. However, several studies have revealed that these models are not effective for everyone due to the imbalanced geographical and economic representation of the data used in the training process. Most of this data comes from Western countries, leading to poor results for underrepresented countries. To address this issue, more data needs to be collected from these countries, but the cost of annotation can be a significant bottleneck. In this paper, we propose methods to identify the data to be annotated to balance model performance and annotation costs. Our approach first involves finding the countries with images of topics (objects and actions) most visually distinct from those already in the training datasets used by current large vision-language foundation models. Next, we identify countries with higher visual similarity for these topics and show that using data from these countries to supplement the training data improves model performance and reduces annotation costs. The resulting lists of countries and corresponding topics are made available at https://github.com/MichiganNLP/visual_diversity_budget.

This research builds upon the Latvian Twitter Eater Corpus (LTEC), which is focused on the narrow domain of tweets related to food, drinks, eating and drinking. LTEC has been collected for more than 12 years and reaching almost 3 million tweets with the basic information as well as extended automatically and manually annotated metadata. In this paper we supplement the LTEC with manually annotated subsets of evaluation data for machine translation, named entity recognition, timeline-balanced sentiment analysis, and text-image relation classification. We experiment with each of the data sets using baseline models and highlight future challenges for various modelling approaches.

Few speech resources describe interruption phenomena, especially for TV and media content. The description of these phenomena may vary across authors: it thus leaves room for improved annotation protocols. We present an annotation of Transition-Relevance Places (TRP) and Floor-Taking event types on an existing French TV and Radio broadcast corpus to facilitate studies of interruptions and turn-taking. Each speaker change is annotated with the presence or absence of a TRP, and a classification of the next-speaker floor-taking as Smooth, Backchannel or different types of turn violations (cooperative or competitive, successful or attempted interruption). An inter-rater agreement analysis shows such annotations’ moderate to substantial reliability. The inter-annotator agreement for TRP annotation reaches κ=0.75, κ=0.56 for Backchannel and κ=0.5 for the Interruption/non-interruption distinction. More precise differences linked to cooperative or competitive behaviors lead to lower agreements. These results underline the importance of low-level features like TRP to derive a classification of turn changes that would be less subject to interpretation. The analysis of the presence of overlapping speech highlights the existence of interruptions without overlaps and smooth transitions with overlaps. These annotations are available at https://lium.univ-lemans.fr/corpus-allies/.

In this study, we focus on the inference presupposed in the concessive discourse relation and present the discourse relation annotation for the Japanese connectives ‘nagara’ and ‘tsutsu’, both of which have two usages: Synchronous and Concession, just like English while. We also present the annotation for ‘tokorode’, which is ambiguous in three ways: Temporal, Location, and Concession. While corpora containing concessive discourse relations already exist, the distinctive feature of our study is that it aims to identify the concessive inferential relations by writing out the implicit presupposed inferences. In this paper, we report on the annotation methodology and its results, as well as the characteristics of concession that became apparent during annotation.

In this paper, we propose a new annotation scheme to classify different types of clauses in Terms-and-Conditions contracts with the ultimate goal of supporting legal experts to quickly identify and assess problematic issues in this type of legal documents. To this end, we built a small corpus of Terms-and-Conditions contracts and finalized an annotation scheme of 14 categories, eventually reaching an inter-annotator agreement of 0.92. Then, for 11 of them, we experimented with binary classification tasks using few-shot prompting with a multilingual T5 and two fine-tuned versions of two BERT-based LLMs for Italian. Our experiments showed the feasibility of automatic classification of our categories by reaching accuracies ranging from .79 to .95 on validation tasks.

Customer-oriented behaviour (COB) plays an important role in call centre interactions, particularly in the context of successful sales negotiation. However, the evaluation of COB in customer-agent conversations often lacks clarity in its definition and robust computational assessment methods. This paper addresses these challenges by presenting a comprehensive conceptual and empirical framework. We conducted multidimensional dialogue act annotations on authentic call centre interactions using the ISO 24617-2 taxonomy, capturing the multifaceted nature of these interactions. This process led to the identification of relevant dialogue act categories, proposed extensions concerning relationship-building aspects, and derived corpus statistics. The findings highlight specific facets of COB that positively impact on Customer Satisfaction (CS), as determined through correlation analysis. Additionally, we delved into the dependencies between COB and feedback acts, leveraging the hierarchical structure of the DIT++ model. This framework improves our understanding of the dynamics shaping sales strategies in call centres and holds promise for practical applications in optimising customer-agent interactions.

In this paper, we present our exploration of annotating Chinese word senses using English WordNet synsets, with examples extracted from OntoNotes Chinese sense inventories. Given a target word along with the example that contains it, the annotators select a WordNet synset that best describes the meaning of the target word in the context. The result demonstrates an inter-annotator agreement of 38% between two annotators. We delve into the instances of disagreement by comparing the two annotated synsets, including their positions within the WordNet hierarchy. The examination reveals intriguing patterns among closely related synsets, shedding light on similar concepts represented within the WordNet structure. The data offers as an indirect linking of Chinese word senses defined in OntoNotes Chinese sense inventories to WordNet sysnets, and thus promotes the value of the OntoNotes corpus. Compared to a direct linking of Chinese word senses to WordNet synsets, the example-based annotation has the merit of not being affected by inaccurate sense definitions and thus offers a new way of mapping WordNets of different languages. At the same time, the annotated data also serves as a valuable linguistic resource for exploring potential lexical differences between English and Chinese, with potential contributions to the broader understanding of cross-linguistic semantic mapping

The scarcity of public datasets for the summarization of medical conversations has been a limiting factor for advancing NLP research in the healthcare domain, and the structure of the existing data is largely limited to the simple format of conversation-summary pairs. We therefore propose a novel Incremental Note Generation (ING) annotation framework capable of greatly enriching summarization datasets in the healthcare domain and beyond. Our framework is designed to capture the human summarization process via an annotation task by instructing the annotators to first incrementally create a draft note as they accumulate information through a conversation transcript (Generation) and then polish the draft note into a reference note (Rewriting). The annotation results include both the reference note and a comprehensive editing history of the draft note in tabular format. Our pilot study on the task of SOAP note generation showed reasonable consistency between four expert annotators, established a solid baseline for quantitative targets of inter-rater agreement, and demonstrated the ING framework as an improvement over the traditional annotation process for future modeling of summarization.

Aspect is a valuable tool for determining the perspective from which an event is observed, allowing for viewing both at the situation and viewpoint level. Uniform Meaning Representation (UMR) seeks to provide a standard, typologically-informed representation of aspects across languages. It employs an aspectual lattice to adapt to different languages and design values that encompass both viewpoint aspect and situation aspects. In the context of annotating the Chinese version of The Little Prince, we paid particular attention to the interactions between aspect values and aspect markers and we also want to know the annotation effectiveness and challenges under the UMR aspectual lattice. During our annotation process, we identified the relationships between aspectual markers and labels. We further categorized and analyzed complex examples that led to low inter-annotator agreement. The factors contributing to disagreement among annotators included the interpretations of lexical semantics, implications, and the influence of aspectual markers, which is related to the inclination of the situation aspect and the exclusivity between the two aspects’ perspectives. Overall, our work sheds light on the challenges of aspect annotation in Chinese and highlights the need for more comprehensive guidelines.

More than 7,000 known languages are spoken around the world. However, due to the lack of annotated resources, only a small fraction of them are currently covered by speech technologies. Albeit self-supervised speech representations, recent massive speech corpora collections, as well as the organization of challenges, have alleviated this inequality, most studies are mainly benchmarked on English. This situation is aggravated when tasks involving both acoustic and visual speech modalities are addressed. In order to promote research on low-resource languages for audio-visual speech technologies, we present AnnoTheia, a semi-automatic annotation toolkit that detects when a person speaks on the scene and the corresponding transcription. In addition, to show the complete process of preparing AnnoTheia for a language of interest, we also describe the adaptation of a pre-trained model for active speaker detection to Spanish, using a database not initially conceived for this type of task. Prior evaluations show that the toolkit is able to speed up to four times the annotation process. The AnnoTheia toolkit, tutorials, and pre-trained models are available at https://github.com/joactr/AnnoTheia/.