CodeNLP at SemEval-2023 Task 2: Data Augmentation for Named Entity Recognition by Combination of Sequence Generation Strategies

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1798–1804 July 13-14, 2023 ©2023 Association for Computational Linguistics CodeNLP at SemEval-2023 Task 2: Data Augmentation for Named Entity Recognition by Combination of Sequence Generation Strategies Michał Marci´nczuk and Wiktor Walentynowicz Wrocław University of Science and Technology {michal.marcinczuk,wiktor.walentynowicz}@pwr.edu.pl Abstract In the article, we present the CodeNLP sub- mission to the SemEval-2023 Task 2: Multi- CoNER II Multilingual Complex Named En- tity Recognition. Our approach is based on data augmentation by combining various strategies of sequence generation for training. We show that the extended procedure of fine-tuning a pre-trained language model can bring improve- ments compared to any single strategy. On the development subsets the improvements were 1.7 pp and 3.1 pp of F-measure, for English and multilingual datasets, respectively. On the test subsets our models achieved 63.51% and 73.22% of Macro F1, respectively. 1 Introduction In this study, we address the topic of named entity recognition from the field of natural language pro- cessing. The task is to identify sequences of words in a text that refer to some categories of entities — people, locations, organizations, objects, events, etc. There is no one firm definition of what a named entity is. The definition and the range of categories may vary from one application to another. Named entity recognition is a challenging task due to several factors. One is that many named en- tities are proper names, and proper names are rigid designators (Kripke, 1980). Proper names refer to entities. They do not describe the entities by defini- tion. This implies that you should know that a given term is a named entity to identify a named entity. The set of named entities is unlimited. To make a model recognize a term as a named entity, you can train on recognition terms from the annotated dataset, provide some form of a common-sense knowledge base about the world, or infer some characteristics from training data. For instance, in the sample sentence "Mark works in Xax.", the term Xax is probably a named entity because it is capitalized. However, semantic categorization might require some additional information about the term, as it might refer to a location (city or country) or a company name. For example, the fol- lowing sentence, "He loves this city." might clarify that Xax is a name of a city. The less information we have from the input text, the more information we need in the training dataset or external sources. In the paper, we deal with the named entity recognition task called MultiCoNER II (Fetahu et al., 2023b; Malmasi et al., 2022b). The main challenges of this task are: a) fine-grained cate- gories of named entities – 6 main categories and 36 subcategories; b) all texts are lowercase, and c)the dataset consists of short sentences. 2 Related Work Current state-of-the-art methods in the field of named entity recognition are based on either LSTM networks (Yu et al., 2020; Xu et al., 2021), or pre- trained language models …

Content truncated to 500 words