Janko at SemEval-2023 Task 2: Bidirectional LSTM Model Based on Pre-training for Chinese Named Entity Recognition

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 958–962 July 13-14, 2023 ©2023 Association for Computational Linguistics Janko at SemEval-2023 Task 2: Bidirectional LSTM Model Based on Pre-training for Chinese Named Entity Recognition Jiankuo Li Yunnan University janko0329@163.com Zhengyi Guan Yunnan University 1941028528@qq.com Haiyan Ding∗ Yunnan University teidhy@163.com Abstract This paper describes the method we submitted as the Janko team in the SemEval-2023 Task 2, Multilingual Complex Named Entity Recogni- tion (MultiCoNER 2). We only participated in the Chinese track. In this paper, we implement the BERT-BiLSTM-RDrop model. We use the fine-tuned BERT models, take the output of BERT as the input of the BiLSTM network, and finally use R-Drop technology to optimize the loss function. Our submission achieved a macro-averaged F1 score of 0.579 on the test set. 1 Introduction Named entity recognition (NER) is to identify en- tities with specific meanings from a given text. Named entities generally refer to entities with a specific meaning or strong referential meaning in the text, usually including personal names, place names, organizational structure names, etc. NER is a key task in natural language processing, and it is very useful in NLP tasks such as information extraction (Toda and Kataoka, 2005), question an- swering system (Mollá et al., 2006), and machine translation (Babych and Hartley, 2003). However, due to the nature of the language, NER is also a challenging problem. First, a word often has mul- tiple meanings, leading to different categories of entities in different sentences. For example, Wash- ington can mean either a place name or a person’s name. Secondly, the boundary of the named entity is vague, and its forms are various. It can be a simple noun, a gerund phrase, or even a complete sentence. MultiCoNER 2 (Fetahu et al., 2023b) aims to identify named entities with finer precision, and the organizers divided the whole competition into 13 tracks based on language. Each data set (Fe- tahu et al., 2023a) is composed of a training set, a validation set, and a test set. The training and validation set provides labels for named entities, and the test set provides text only. Compared with MultiCoNER (Malmasi et al., 2022b), the number of entity tags has been increased from 6 (Malmasi et al., 2022a) to more than 30, while the semantics are vaguer and the text length is shorter. In addition, mock errors are added to the test set to make the task more realistic and difficult. Similar to other NER tasks, MultiCoNER 2 uses B-I-O tags to label the tokens, where “B” indicates the first token of a concept, “I” indicates tokens inside of a concept, and “O” indicates tokens that don’t belong to any entity. This paper explains our submission to the Chi- nese track in MultiCoNER 2. We implemented a simple neural network system based on a pre- trained language model. Our submission ranked 14th among 21 teams with 0.579 F1 score in the test set. This paper is organized as follows: In Section …

Content truncated to 500 words