SAB at SemEval-2023 Task 2: Does Linguistic Information Aid in Named Entity Recognition?

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1131–1137 July 13-14, 2023 ©2023 Association for Computational Linguistics SAB at SemEval-2023 Task 2: Does Linguistic Information Aid in Named Entity Recognition? Siena Biales University of Tübingen sienab1@gmail.com Abstract This paper describes the submission to SemEval-2023 Task 2: Multilingual Complex Named Entity Recognition (MultiCoNER II) by team SAB. This task aims to encourage growth in the field of Named Entity Recog- nition (NER) by focusing on complex and dif- ficult categories of entities, in 12 different lan- guage tracks. The task of NER has historically shown the best results when a model incorpo- rates an external knowledge base or gazetteer, however, less research has been applied to ex- amining the effects of incorporating linguistic information into the model. In this task, we ex- plored combining NER, part-of-speech (POS), and dependency relation labels into a multi-task model and report on the findings. We deter- mine that the addition of POS and dependency relation information in this manner does not improve results. 1 Introduction Named Entity Recognition (NER, Grishman and Sundheim, 1996) is a well established Natural Lan- guage Processing (NLP) task which entails ex- tracting named entities (NEs), or specific types of proper nouns, from a text and classifying the type of entity. In the original task from 1996, entities were considered either a person, organization, or geographic location, but since then the classes have expanded to include more complex entity types such as creative works (book or film titles, etc.), groups, and products. While it can be considered a task on its own, NER has many applications in- cluding Information Extraction (IE) and Question Answering (QA) (Sun et al., 2018). Complex named entities (such as titles of cre- ative works, products, or groups) are especially dif- ficult for NER systems, because unlike traditional NEs which are generally noun phrases, complex named entities may take various other forms (Ash- wini and Choi, 2014). In addition, the domain of complex named entities is constantly expanding as new movies, books, and products are released. As a result, NER models must be adaptable in order to successfully identify and classify such entities. SemEval-2023 Task 2: Multilingual Complex Named Entity Recognition II (MultiCoNER II, Fe- tahu et al., 2023b) is a shared task with the aim of investigating methods of identifying complex named entities in 11 individual language tracks, as well as a multilingual track. This task is a sequel to SemEval-2022 Task 11 (Malmasi et al., 2022b), but with different language tracks and a much more in-depth set of entity classes. Apart from having 33 fine-grained entity classes, the task also included simulated errors added to some of the test set lan- guage tracks to make the task more realistic and challenging. A large portion of NLP research focuses on En- glish and neglects lower-resource languages (Kle- men et al., 2022). The language tracks defined in the task are English (EN), Spanish (ES), Hindi (HI), Bangla (BN), Chinese (ZH), Swedish (SV), Farsi …

Content truncated to 500 words