RACAI at SemEval-2022 Task 11: Complex named entity recognition using a lateral inhibition mechanism

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1562 - 1569 July 14-15, 2022 ©2022 Association for Computational Linguistics RACAI at SemEval-2022 Task 11: Complex named entity recognition using a lateral inhibition mechanism Vasile P˘ais, Research Institute for Artificial Intelligence ”Mihai Dr˘ag˘anescu”, Romanian Academy Bucharest, Romania vasile@racai.ro Abstract This paper presents RACAI’s system used for the shared task of "Multilingual Complex Named Entity Recognition (MultiCoNER)", organized as part of "The 16th International Workshop on Semantic Evaluation (SemEval 2022)". The system employs a novel layer in- spired by the biological mechanism of lateral inhibition. This allowed the system to achieve good results without any additional resources apart from the provided training data. In ad- dition to the system’s architecture, results are provided as well as observations regarding the provided dataset. 1 Introduction Named entity recognition (NER) is a well known task in natural language processing. It aims to detect spans of text associated with known enti- ties. Initially, much work focused on detecting per- sons, organizations, and locations (Grishman and Sundheim, 1996; Tjong Kim Sang and De Meulder, 2003). However, this limited approach is not suit- able for every domain, thus leading to research in domain-specific NER. For example, in the biomed- ical domain, a number of works have addressed entities such as genes, proteins, diseases (Hu and Verberne, 2020), cell types (Settles, 2004), chemi- cals (Gonzalez-Agirre et al., 2019; Ion et al., 2019). Similarly, in the legal domain additional classes are employed such as money value (Glaser et al., 2018), legal reference (Landthaler et al., 2016; P˘ais, et al., 2021), judge, and lawyer (Leitner et al., 2019). In the context of "The 16th International Work- shop on Semantic Evaluation (SemEval 2022)"1, the task number 11 "Multilingual Complex Named Entity Recognition (MultiCoNER)"2 (Malmasi et al., 2022b) required participants to build a NER 1https://semeval.github.io/ SemEval2022/ 2https://multiconer.github.io/ system able to recognize complex entities in 11 lan- guages: Bangla, Chinese, Dutch, English, Farsi, German, Hindi, Korean, Russian, Spanish, and Turkish. In addition, a multilingual track and a code-mixed track were available. The task focused on 6 entity types: person, location, group, corpora- tion, product and creative work. As noted by Ashwini and Choi (2014), non- traditional entities can pose a challenge for NER systems. This happens because datasets are harder to build and certain entities (such as creative works) are updated more frequently than traditional ones (persons, locations). Furthermore, traditional enti- ties tend to occur as noun phrases, while the newly proposed entities (for the purposes of the task) may be linguistically complex (complex noun phrases, gerunds, infinitives or full clauses). An interesting result was provided by Aguilar et al. (2017), where the top system from WNUT 2017 achieved only 8% recall when dealing with creative works. This paper describes a system for complex NER in a multilingual context, developed at the Research Institute for Artificial Intelligence of the Romanian Academy (RACAI), that participated in the Multi- CoNER task. The system employs a new artificial neural network layer trying to mimic …

Content truncated to 500 words