The CRECIL Corpus: a New Dataset for Extraction of Relations between Characters in Chinese Multi-party Dialogues

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 2337–2344 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 2337 The CRECIL Corpus: a New Dataset for Extraction of Relations between Characters in Chinese Multi-party Dialogues Yuru Jiang∗, Yang Xu∗, Yuhang Zhan∗, Weikai He∗, Yilin Wang∗ , Zixuan Xi∗, Meiyun Wang∗, Xinyu Li∗, Yu Li∗, Yanchao Yu† ∗ Computer School, Beijing Information Science and Technology University jiangyuru@bistu.edu.cn † School of Computing, Edinburgh Napier University y.yu@napier.ac.uk Abstract We describe a new freely available Chinese multi-party dialogue data set for automatic extraction of dialogue- based character relationships. The data has been extracted from the original TV scripts of a Chinese sitcom called “I Love My Family” with complex family-based human daily spoken conversations in Chinese. First, we introduced human annotation scheme for both global Character relationship map and character reference relationship. And then we generated the dialogue-based character relationship triples. The corpus annotates relationships between 140 entities in total. We also carried out a data exploration experiment by deploying a BERT-based model to extract character relationships on the CRECIL corpus and another existing relation extraction corpus (DialogRE (Yu et al., 2020)).The results demonstrate that extracting character relationships is more challenging in CRECIL than in DialogRE. Keywords: Multi-Party Dialogue; Character Relation; Corpus; Annotation Scheme; Relation Extraction 1. Introduction While bringing intelligent systems/robots out of the laboratory into the physical world, especially into the daily environment, they must become ca- pable of understanding natural daily conversations between two or more human speakers. Among other capabilities, this involves the ability to iden- tify/extract the character1/speaker relationships by analyzing the semantic meanings of conversa- tions between different users – this is widely known as the Relation Extraction (RE) problem. Solving such a problem has become a critical task in Natu- ral Language Processing, and it plays an essential role in downstream tasks. In the past decades, the RE problem in language processing has received considerable attention in computational linguistics. On the one hand, there is work that only addresses the relation extraction problem on the document level. Some datasets fo- cus on the RE problem between two entities in one sentence, such as the SemEval-2010 Task 8 dataset(Hendrickx et al., 2010) and the TACRED dataset(Zhang et al., 2017). And some datasets fo- cus on the cross-sentence relationships, i.e. two en- tities whose relation needs to be determined are not in the same sentence or even the same paragraph, such as BC5CDR (Li et al., 2016) and DocRED 1This paper is about social relationships between lit- erary characters or protagonists in a story, rather than relations between orthographic Chinese characters. dataset (Yao et al., 2019). Furthermore, different to the formal document, some RE datasets focus on annotating dialogues. There are much more complicated connections between different speak- ers in a dialogue, especially with more than two participants. The relationship in the dialogue is more character-related, and a complete dialogue exists between different speakers. In addition to many cross-sentence relationships, …

Content truncated to 500 words