PAI at SemEval-2022 Task 11: Name Entity Recognition with Contextualized Entity Representations and Robust Loss Functions

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1665 - 1670 July 14-15, 2022 ©2022 Association for Computational Linguistics PAI at SemEval-2022 Task 11: Name Entity Recognition with Contextualized Entity Representations and Robust Loss Functions Long Ma∗, Xiaorong Jian∗,Xuan Li Pingan Life Insurance of China, Ltd {MALONG633,JIANXIAORONG974,LIXUAN208}@pingan.com.cn Abstract This paper describes our system used in the SemEval-2022 Task 11 Multilingual Complex Named Entity Recognition, achieving 3rd for track 1 on the leaderboard. We propose Dictionary-fused BERT, a flexible approach for entity dictionaries integration. The main ideas of our systems are: 1) integrating exter- nal knowledge (an entity dictionary) into pre- trained models to obtain contextualized word and entity representations 2) designing a ro- bust loss function leveraging a logit matrix 3) adding an auxiliary task, which is an on-top binary classification to decide whether the to- ken is a mention word or not, makes the main task easier to learn. It is worth noting that our system achieves an F1 of 0.914 in the post- evaluation stage by updating the entity dictio- nary to the one of Meng et al. (2021), which is higher than the score of 1st on the leaderboard of the evaluation stage. 1 Introduction Name entity recognition (NER) is a fundamental task in natural language processing. Processing complex and ambiguous Named Entities (NEs) and in low-context situations is a challenging NLP task in practical and open-domain settings, which was recently outlined by Meng et al. (2021). Other work has extended this to multilingual and code- mixed settings (Fetahu et al., 2021) since code- mixed queries, with entities in a different language than the rest of the query, pose a particular chal- lenge in domains like e-commerce (e.g. queries containing movie or product names). SemEval-2022 task 11 Multilingual Complex Named Entity Recognition (Malmasi et al., 2022b) focuses on dealing with the challenges above: de- tecting complex entities in short, low-context, and code-mixed settings. * The first two authors contributed equally. For this task, we propose Dictionary-fused BERT to integrate external dictionaries into the NER model, and it is compatible with emerging entities and user-defined entities, without retraining the model. 2 Related Work Named Entity Recognition (NER) is a core task in knowledge extraction and is important to various downstream applications such as question answer- ing and dialogue systems. NER is the task of detect- ing mentions of real-world entities from text and recognizing their types (e.g., locations, persons). However, the NER task is facing many chal- lenges outlined by Meng et al. (2021), such as short texts like search queries (Wang et al., 2014), emerging entities (Craswell et al., 2020), long-tail entities, complex entities, and entities in code-mixed queries (Fetahu et al., 2021). For example, complex NEs, like the titles of creative works (movie/book/song/software names) are not simple nouns and are harder to recognize. The neural models driven by memorization perform well on “easy” entities (person names) but fail to recognize complex/unseen entities when entities overlap less between train/test sets. A lot of works …

Content truncated to 500 words