Infrrd.ai at SemEval-2022 Task 11: A system for named entity recognition using data augmentation, transformer-based sequence labeling model, and EnsembleCRF

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1501 - 1510 July 14-15, 2022 ©2022 Association for Computational Linguistics Infrrd.ai at SemEval-2022 Task 11: A system for named entity recognition using data augmentation, transformer-based sequence labeling model, and EnsembleCRF JiangLong He, Akshay Uppal, Mamatha N, Shiv Vignesh, Deepak Kumar, Aditya Kumar Sarda Infrrd.ai {jianglong,akshayuppal,mamathan}@infrrd.ai {shivvignesh,deepakumar,adityasarda}@infrrd.ai Abstract In low-resource languages, the amount of train- ing data is limited. Hence, the model has to perform well in unseen sentences and syntax on which the model has not trained. We propose a method that addresses the problem through an encoder and an ensemble of language models. A language-specific language model performed poorly when compared to a multilingual lan- guage model. So, the multilingual language model checkpoint is fine-tuned to a specific lan- guage. A novel approach of one hot encoder is introduced between the model outputs and the CRF to combine the results in an ensemble format. Our team, Infrrd.ai, competed in the MultiCoNER competition. The results are en- couraging where the team is positioned within the top 10 positions. There is less than a 4% percent difference from the third position in most of the tracks that we participated in. The proposed method shows that the ensemble of models with a multilingual language model as the base with the help of an encoder performs better than a single language-specific model. 1 Introduction In conll-2003 (Sang and De Meulder, 2003), a shared task was conducted to identify the named entities such as person (PER), location (LOC), orga- nization (ORG), and miscellaneous (MISC). Over a period of time, there was an improvement in the developed systems (Marrero et al., 2009) which resulted in an improved performance that resulted in an increased number of entities. The named en- tities which needed to be identified and extracted were now six. They are person (PER), location (LOC), organization (ORG), group (GRP), prod- uct (PROD), and creative work (CW). Some of these named entities are created by coining a new word that may be non-existent or a combination of existing words in the entire corpus of words in a language. People are most likely interested in coining a new term from the list of words to have an identity tag of a location, an event, or similar concepts. The newly coined words become novel or emerging entities in the list of entities (Derczyn- ski et al., 2017). These words form a most part of ambiguous words, where the word belongs to a particular entity or not, and is difficult to judge. For example, Microsoft is an organization, Windows is a product, and Microsoft Windows is a software product. Apart from this, creating a non-existent word as an entity is an expression of the creativity of the creator which belongs to the creative word entities. A competition was conducted in WNUT 2017 to identify the novel and emerging entities that look for unseen entities as described earlier (Derczynski et al., 2017). These types of entities are …

Content truncated to 500 words