VBD_NLP at SemEval-2023 Task 2: Named Entity Recognition Systems Enhanced by BabelNet and Wikipedia

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1833–1843 July 13-14, 2023 ©2023 Association for Computational Linguistics VBD_NLP at SemEval-2023 Task 2: Named Entity Recognition Systems Enhanced by BabelNet and Wikipedia Phu Gia Hoang, Thanh Tuan Le, and Long Hai Trieu {v.phuhg1, v.thanhlt41, v.longth12}@vinbigdata.com VinBigData JSC, Hanoi, Vietnam Abstract We describe our systems participated in the SemEval-2023 shared task for Named Entity Recognition (NER) in English and Bangla. In order to address the challenges of the task, where a large number of fine-grained named entity types need to be detected with only a small amount of training data, we use a method to augment the training data based on Babel- Net concepts and Wikipedia redirections to automatically annotate named entities from Wikipedia articles. We build our NER systems based on the powerful mDeBERTa pretrained language model and trained on the augmented data. Our approach significantly enhances the performance of the fine-grained NER task in both English and Bangla subtracks, outper- forming the baseline models. Specifically, our augmented systems achieve macro-f1 scores of 52.64% and 64.31%, representing improve- ments of 2.38% and 11.33% over the English and Bangla baselines, respectively. 1 Introduction The goal of the SemEval-2023 Task 2: Multi- CoNER II Multilingual Complex Named Entity Recognition (Fetahu et al., 2023b) is to focus on extracting semantically ambiguous complex named entities based on the MULTICONER v2 (Fetahu et al., 2023a). The first version of MULTICONER (Malmasi et al., 2022b), which used the same tax- onomy schema as introduced in WNUT 2017 (Der- czynski et al., 2017), was utilized for SemEval- 2022 Task 11: MultiCoNER (Malmasi et al., 2022b), that had 13 subtracks in multilingual, code- mixing, and 11 monolingual languages, namely: English, Spanish, Dutch, Russian, Turkish, Korean, Farsi, German, Chinese, Hindi, and Bangla. In the second version, the the MULTICONER v2 taxon- omy was changed to challenge the previous sys- tems and evaluate their performance on noisy and fine-grained data. The second version provides 13 subtracks in multilingual and 12 monolingual lan- guages, including English, Spanish, Farsi, German, Chinese, Hindi, Bangla, Swedish, French, Italian, Portuguese, and Ukrainian. This task is derived from the MultiCoNER shared task last year (Malmasi et al., 2022b), which was a challenging task due to scarcity training in- stances, with approximately ten thousand training instances and testing data that was ten times larger than the training data. Furthermore, the shared task this year has become more challenging when the number of entity types increases from 6 to 35 fine- grained entity types. This increase in the number of entity types has caused issues for models that train on few training instances and detect entities in a large number of classes. However, this task presents an interesting opportunity to investigate techniques for extending training data to deal with these issues. As a step towards addressing the gap, we devel- oped and evaluated our systems for the SemEval- 2023 Task 2 in English and Bangla by augmenting the training data from a knowledge base. While …

Content truncated to 500 words