Language Resources and Evaluation最新文献

We present a comprehensive and valuable resource in the form of an aligned parallel corpus comprising translations of the Bible in Spanish. Our collection encompasses a total of eleven Bibles, originating from diverse centuries (XVI, XIX, XX), various religious denominations (Protestant, Catholic), and geographical regions (Spain, Latin America). The process of aligning the verses across these translations has been meticulously carried out, ensuring that the content is organized in a coherent manner. As a result, this corpus serves as a useful convenient resource for various linguistic analyses, including paraphrase detection, semantic clustering, and the exploration of biases present within the texts. To illustrate the utility of this resource, we provide several examples that demonstrate how it can be effectively employed in these applications.

The widespread of offensive content online, such as hate speech and cyber-bullying, is a global phenomenon. This has sparked interest in the artificial intelligence (AI) and natural language processing (NLP) communities, motivating the development of various systems trained to detect potentially harmful content automatically. These systems require annotated datasets to train the machine learning (ML) models. However, with a few notable exceptions, most datasets on this topic have dealt with English and a few other high-resource languages. As a result, the research in offensive language identification has been limited to these languages. This paper addresses this gap by tackling offensive language identification in Sinhala, a low-resource Indo-Aryan language spoken by over 17 million people in Sri Lanka. We introduce the Sinhala Offensive Language Dataset (SOLD) and present multiple experiments on this dataset. SOLD is a manually annotated dataset containing 10,000 posts from Twitter annotated as offensive and not offensive at both sentence-level and token-level, improving the explainability of the ML models. SOLD is the first large publicly available offensive language dataset compiled for Sinhala. We also introduce SemiSOLD, a larger dataset containing more than 145,000 Sinhala tweets, annotated following a semi-supervised approach.

Event-based surveillance (EBS) requires the analysis of an ever-increasing volume of documents, requiring automated processing to support human analysts. Few annotated corpora are available for the evaluation of information extraction tools in the EBS domain. The main objective of this work was to build a corpus containing documents which are representative of those collected in the current EBS information systems, and to annotate them with events and their novelty. We proposed new definitions of infectious events and their novelty suited to the background work of analysts working in the EBS domain, and we compiled a corpus of 305 documents describing 283 infectious events. There were 36 included documents in French, representing a total of 11 events, with the remainder in English. We annotated novelty for the 110 most recent documents in the corpus, resulting in 101 events. The inter-annotator agreement was 0.74 for event identification (F1-Score) and 0.69 [95% CI: 0.51; 0.88] (Kappa) for novelty annotation. The overall agreement for entity annotation was lower, with a significant variation according to the type of entities considered (range 0.30–0.68). This corpus is a useful tool for creating and evaluating algorithms and methods submitted by EBS research teams for event detection and annotation of their novelties, aiming at the operational improvement of document flow processing. The small size of this corpus makes it less suitable for training natural language processing models, although this limitation tends to fade away when using few-shots learning methods.

Abstract

In this paper, we approach the problem of semantic search by introducing a task of paraphrase span detection, i.e. given a segment of text as a query phrase, the task is to identify its paraphrase in a given document, the same modelling setup as typically used in extractive question answering. While current work in paraphrasing has almost uniquely focused on sentence-level approaches, the novel span detection approach gives a possibility to retrieve a segment of arbitrary length. On the Turku Paraphrase Corpus of 100,000 manually extracted Finnish paraphrase pairs including their original document context, we find that by achieving an exact match of 88.73 our paraphrase span detection approach outperforms widely adopted sentence-level retrieval baselines (lexical similarity as well as BERT and SBERT sentence embeddings) by more than 20pp in terms of exact match, and 11pp in terms of token-level F-score. This demonstrates a strong advantage of modelling the paraphrase retrieval in terms of span extraction rather than commonly used sentence similarity, the sentence-level approaches being clearly suboptimal for applications where the retrieval targets are not guaranteed to be full sentences. Even when limiting the evaluation to sentence-level retrieval targets only, the span detection model still outperforms the sentence-level baselines by more than 4 pp in terms of exact match, and almost 6pp F-score. Additionally, we introduce a method for creating artificial paraphrase data through back-translation, suitable for languages where manually annotated paraphrase resources for training the span detection model are not available.

Concept maps are two-dimensional visual tools that describe the relationships between concepts belonging to a particular subject. The manual creation of these maps entails problems such as requiring expertise in the relevant field, minimizing visual complexity, and integrating maps, especially in terms of text-intensive documents. In order to overcome these problems, automatic creation of concept maps is required. On the other hand, the production of a fully automated and human-hand quality concept map from a document has not yet been achieved satisfactorily. Motivated by this observation, this study aims to develop a new methodology for automatic creation of the concept maps from Turkish text documents for the first time in the literature. In this respect, within the scope of this study, a new heuristic algorithm has been developed using the Turkish Natural Language Processing software chain and the Graphviz tool to automatically extract concept maps from Turkish texts. The proposed algorithm works with the principle of obtaining concepts based on the dependencies of Turkish words in sentences. The algorithm also determines the sentences to be added to the concept map with a new sentence scoring mechanism. The developed algorithm has been applied on a total of 20 data sets in the fields of Turkish Literature, Geography, Science, and Computer Sciences. The effectiveness of the algorithm has been analyzed with three different performance evaluation criteria, namely precision, recall and F-score. The findings have revealed that the proposed algorithm is quite effective in Turkish texts containing concepts. It has also been observed that the sentence selection algorithm produces results close to the average value in terms of the performance criteria being evaluated. According to the findings, the concept maps automatically obtained by the proposed algorithm are quite similar to the concept maps extracted manually. On the other hand, there is a limitation of the developed algorithm since it is dependent on a natural language processing tool and therefore requires manual intervention in some cases.

Abstract

With globalization and cultural exchange around the globe, most of the population gained knowledge of at least two languages. The bilingual user base on the Social Media Platform (SMP) has significantly contributed to the popularity of code-mixing. However, apart from multiple vital uses, SMP also suffer with abusive text content. Identifying abusive instances for a single language is a challenging task, and even more challenging for code-mix. The abusive posts detection problem is more complicated than it seems due to its unseemly, noisy data and uncertain context. To analyze these contents, the research community needs an appropriate dataset. A small dataset is not a suitable sample for the research work. In this paper, we have analyzed the dimensions of Devanagari-Roman code-mix in short noisy text. We have also discussed the challenges of abusive instances. We have proposed a cost-effective methodology with 20.38% relevancy score to collect and annotate the code-mix abusive text instances. Our dataset is eight times to the related state-of-the-art dataset. Our dataset ensures the balance with 55.81% instances in the abusive class and 44.19% in the non-abusive class. We have also conducted experiments to verify the usefulness of the dataset. We have performed experiments with traditional machine learning techniques, traditional neural network architecture, recurrent neural network architectures, and pre-trained Large Language Model (LLM). From our experiments, we have observed the suitability of the dataset for further scientific work.

In this paper we present a database management and annotation tool for running an enriched FrameNet database, the FrameNet Brasil WebTool. We demonstrate how the entity-based model of such a tool allows for the addition of two types of data-structure to FrameNet Brasil, both of which aimed at refining the granularity of the semantic representations: the frame element-to-frame and the ternary qualia relations. We report on three proof-of-concept applications of such an enriched database: a domain-specific structured lexicon, a recommendation system for tourists and a post-editing system for domain adaptation in machine translation.

In this article, we present the NewsCom-TOX corpus, a new corpus manually annotated for toxicity in Spanish. NewsCom-TOX consists of 4359 comments in Spanish posted in response to 21 news articles on social media related to immigration, in order to analyse and identify messages with racial and xenophobic content. This corpus is multi-level annotated with different binary linguistic categories -stance, target, stereotype, sarcasm, mockery, insult, improper language, aggressiveness and intolerance- taking into account not only the information conveyed in each comment, but also the whole discourse thread in which the comment occurs, as well as the information conveyed in the news article, including their images. These categories allow us to identify the presence of toxicity and its intensity, that is, the level of toxicity of each comment. All this information is available for research purposes upon request. Here we describe the NewsCom-TOX corpus, the annotation tagset used, the criteria applied and the annotation process carried out, including the inter-annotator agreement tests conducted. A quantitative analysis of the results obtained is also provided. NewsCom-TOX is a linguistic resource that will be valuable for both linguistic and computational research in Spanish in NLP tasks for the detection of toxic information.

Detecting toxic comments and rationale for the offensiveness of a social media post promotes moderation of social media content. For this purpose, we propose a Co-Attentive Multi-task Learning (CA-MTL) model through transfer learning for low-resource Hindi-English (commonly known as Hinglish) toxic texts. Together, the cooperative tasks of rationale/span detection and toxic comment classification create a strong multi-task learning objective. A task collaboration module is designed to leverage the bi-directional attention between the classification and span prediction tasks. The combined loss function of the model is constructed using the individual loss functions of these two tasks. Although an English toxic span detection dataset exists, one for Hinglish code-mixed text does not exist as of today. Hence, we developed a dataset with toxic span annotations for Hinglish code-mixed text. The proposed CA-MTL model is compared against single-task and multi-task learning models that lack the co-attention mechanism, using multilingual and Hinglish BERT variants. The F1 scores of the proposed CA-MTL model with HingRoBERTa encoder for both tasks are significantly higher than the baseline models. Caution: This paper may contain words disturbing to some readers.

Abstract

This paper stems from the project A World of Possibilities. Modal pathways over an extra-long period of time: the diachrony of modality in the Latin language (WoPoss) which involves a corpus-based approach to the study of modality in the history of the Latin language. Linguistic annotation and, in particular, the semantic annotation of modality is a keystone of the project. Besides the difficulties intrinsic to any annotation task dealing with semantics, our annotation scheme involves multiple layers of annotation that are interconnected, adding complexity to the task. Considering the intricacies of our fine-grained semantic annotation, we needed to develop well-documented schemas in order to control the consistency of the annotation, but also to enable an efficient reuse of our annotated corpus. This paper presents the different elements involved in the annotation task, and how the description and the relations between the different linguistic components were formalised and documented, combining schema languages with XML documentation.