LSJSP at SemEval-2023 Task 2: FTBC: A FastText based framework with pre-trained BERT for NER

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1254–1259 July 13-14, 2023 ©2023 Association for Computational Linguistics LSJSP at SemEval-2023 Task 2: FTBC: A FastText based framework with pre-trained BERT for NER Shilpa Chatterjee* IIT Kanpur shilpa@cse.iitk.ac.in Leo Evenss* IIT Kanpur leoevenss@cse.iitk.ac.in Pramit Bhattacharyya* IIT Kanpur pramitb@cse.iitk.ac.in Joydeep Mondal IIT Kanpur joydeep20@iitk.ac.in Abstract This study introduces the system submitted to the SemEval 2022 Task 2: MultiCoNER II (Mul- tilingual Complex Named Entity Recognition) by the LSJSP team. We propose FTBC, a FastText- based framework with pre-trained Bert for NER tasks with complex entities and over a noisy dataset. Our system achieves an average of 58.27% F1 score (fine-grained) and 75.79% F1 score (coarse- grained) across all languages. FTBC outperforms the baseline BERT-CRF model on all 12 monolin- gual tracks. 1 Introduction NER (Named Entity Recognition) is a sequence la- beling task that finds spans of text and tags them to a named entity (persons, locations, organizations, and others.). NER finds its applications in various Natural Language Processing (NLP) tasks, includ- ing Information Retrieval, Machine Translation, Question Answering, and Automatic Summariza- tion (Jurafsky and Martin, 2021). It also finds ap- plication across various domains, including social media, news, e-commerce, and healthcare. Words having multiple meanings increase the complexity of the NER task. In English, the word “Sunny” can refer to both a person and the weather. The prob- lem of ambiguous entities become all the more pro- nounced for Indian languages. In the Bangla sen- tence “cañcala caudhur¯ı cañcala bodha karachena" the first “cañcala” refers to a person named cañ- cala whereas the second “cañcala” (restless) is an adjective. Pre-trained contextual embeddings such as BERT (Devlin et al., 2019), XLM-R (Conneau et al., 2019), and other transformer-based (Vaswani et al., 2017) models have improved the perfor- mance of NER. The embeddings are trained on *These authors contributed equally to this work large corpora, improving the named entities’ con- textual representations. However, for low-resource languages like Bengali, Hindi, Swedish, and Farsi, the performance of these pre-trained models lag compared to their English counterpart. MultiCoNER II shared task (Fetahu et al., 2023b) aims at building Named Entity Recognition (NER) systems for 12 languages, including Bangla (BN), German (DE), English (EN), Spanish (ES), Farsi (FA), French (FR), Hindi (HI), Italian (IT), Portuguese (PT), Swedish (SV), Ukrainian (UK), Chinese (ZH). The task has two kinds of tracks, in- cluding one multilingual track and 12 monolingual tracks for all the languages. The monolingual track requires the development of NER models for that particular language, whereas the multilingual track requires the development of a single multilingual model that can handle all 12 languages. The se- quence of tokens in MultiCoNER II is classified into 30+ classes. Further, including simulated er- rors, like spelling mistakes, in the test set makes the task more realistic and difficult. This paper proposes FTBC, a framework com- prising FastText (Bojanowski et al., 2017) and BERT for the MultiCoNER II task without us- ing any other external knowledge bases. …

Content truncated to 500 words