Comparison of Multilingual Entity Linking Approaches

Proceedings of Recent Advances in Natural Language Processing, pages 224–233 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_025 224 Comparison of Multilingual Entity Linking Approaches Ivelina Bozhinova Solutions Unit, Ontotext AD 79 Nikola Gabrovski St Sofia, Bulgaria ivelina.bozhinova@ontotext.com Andrey Tagarev Research Unit, Ontotext AD, Sofia, Bulgaria and Computer Science, University of Sheffield andrey.tagarev@ontotext.com Abstract Despite rapid developments in the field of Nat- ural Language Processing (NLP) in the past few years, the task of Multilingual Entity Link- ing (MEL) and especially its end-to-end formu- lation remains challenging. In this paper we aim to evaluate solutions for general end-to- end multilingual entity linking by conducting experiments using both existing complete ap- proaches and novel combinations of pipelines for solving the task. The results identify the best performing current solutions and suggest some directions for further research. 1 Introduction Entity linking (EL) (Hoffart et al., 2011), (Cucerzan, 2007) is the task of mapping mentions in unstructured text to entities in an existing Knowl- edge Base (KB). It has drawn the attention of many researchers in the past few years due to its appli- cation in different areas of NLP, including Ques- tion Answering (De Cao et al., 2019), (Yin et al., 2016), (Wang et al., 2021), Relation Extraction (Baldini Soares et al., 2019), Dialogue (Chen et al., 2017a), (Bordes et al., 2017), (Wen et al., 2017) and Biomedical systems (Bhowmik et al., 2021), (Zheng et al., 2015). Even though there has been a significant improvement in the field recently (Cao et al., 2021), (Wu et al., 2020), (Ayoola et al., 2022), the task of EL and especially in the cross-lingual (Ji et al., 2015), (McNamee et al., 2011), and MEL setups remain challenging. Different approaches have been proposed for solving this task, some of which are based on more traditional methods (Brank et al., 2017), (Delpeuch, 2020) and others exploit the recent discoveries in the field of natural language processing (Cao et al., 2021), (Wu et al., 2020), (Ayoola et al., 2022), (Botha et al., 2020). This paper will present experiments comparing the performance of various methods for a MEL task. 2 Multilingual Entity Recognition and Disambiguation Methods In EL, also known as named-entity recognition and disambiguation (NERD) words of interest in an unstructured text are mapped to corresponding unique entities in an existing target KB. Formally it can be defined as the task of linking a given entity mention m in a given context c to the corresponding entity e in a KB. For the multilingual definition a set of languages L is added and the context is defined as language specific (context c of language l). It also requires a multilingual KB. As the name NERD suggests, the task consists of two subtasks, namely named-entity recognition (NER) and entity disambiguation (ED). Two general groups of EL methods exist. One focuses on performing entity disambiguation but requires correctly annotated entities or at least entity spans in its input. The second takes plain text input and performs both recognition and disambiguation in one or more …

Content truncated to 500 words