KDDIE at SemEval-2023 Task 2: External Knowledge Injection for Named Entity Recognition

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1498–1501 July 13-14, 2023 ©2023 Association for Computational Linguistics KDDIE at SemEval-2023 Task 2: External Knowledge Injection for Named Entity Recognition Caleb Martin, Huichen Yang, and William Hsu Computer Science of Kansas State University Manhattan, Kansas 66502 {calebjm288, huichen, bhsu}@ksu.edu Abstract This paper introduces our system for the Se- mEval 2023 Task 2: Multilingual Complex Named Entity Recognition (MultiCoNER II) competition. Our team (KDDIE) focused on the sub-task of Named Entity Recognition (NER) for the language of English in the com- petition and reported our results. To achieve our goal, we utilized transfer learning by fine- tuning pre-trained language models (PLMs) on the competition dataset. Our approach involved combining a BERT-based PLM with external knowledge to provide additional context to the model. In this report, we present our findings and results. 1 Introduction There is a large and ever-increasing supply of un- structured data present in today’s world. Much of this data is in the form of free text. Named Entity Recognition (NER) is the task of labeling named entities within the text. This allows us to gather structured data from the free text. SemEval 2023 Task 2 is a competition in which groups compete to build the best NER system for the provided data (Fetahu et al. (2023b)). This task contains 13 differ- ent tracks: one for each of 12 different languages and a multilingual track, which combines all the other languages (Fetahu et al. (2023b)). In this pa- per, we will focus on Track 1 which is monolingual English NER. The task requires NER systems to identify 36 different labels. The six coarse labels to be used were person, location, group, medical, product, and creative work (Fetahu et al. (2023b)). Each of these coarse-level labels is then split into several fine-grain labels. For example, location is split into facility, human settlement, station, and other locations. Figure 1 shows some examples of words being labeled with their corresponding tags. In this paper, we used external knowledge in- jection from outside data sources to provide addi- tional context to our PLMs, as described in (Wang christoph haberland designed a new marble pulpit for the church which was built in italy in 1793. OtherPER OtherPROD HumanSettlement produced by ekta kapoor under the production company name balaji telefilms. Artist ORG Figure 1: An example of labeling some entities with the SemEval 2023 Task 2 tagging scheme. These two sentences are taken from the training set of SemEval 2023 Task 2 (Fetahu et al. (2023a)). et al. (2022)). Our approach involved training mul- tiple transformer models from the HuggingFace library (Wolf et al. (2020)) on the provided train- ing data that was enriched with additional context. We then combined the predictions from several of these models to further improve their performance on the dataset. We fine-tuned several training pa- rameters, including the number of epochs, learning rate, batch size, and other parameters for all the models to achieve the best possible …

Content truncated to 500 words