RIGA at SemEval-2023 Task 2: NER Enhanced with GPT-3

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 331–339 July 13-14, 2023 ©2023 Association for Computational Linguistics RIGA at SemEval-2023 Task 2: NER enhanced with GPT-3 Eduards Mukans mukans.work@gmail.com University of Latvia, Faculty of Computing Guntis Barzdins guntis.barzdins@lumii.lv University of Latvia, IMCS Abstract The following is a description of the RIGA team’s submissions for the English track of the SemEval-2023 Task 2: Multilingual Complex Named Entity Recognition (MultiCoNER) II. Our approach achieves 17% boost in results by utilizing pre-existing Large-scale Language Models (LLMs), such as GPT-3, to gather addi- tional contexts. We then fine-tune a pre-trained neural network utilizing these contexts. The final step of our approach involves meticulous model and compute resource scaling, which results in improved performance. Our results placed us 12th out of 34 teams in terms of over- all ranking and 7th in terms of the noisy subset ranking. The code for our method is available on GitHub1. 1 Introduction The SemEval 2023 Task 2: MultiCoNER II (Multi- lingual Complex Named Entity Recognition) (Fe- tahu et al., 2023b) presented participants with a challenge to classify words into 36 named and nominal entities. Unlike traditional named entity recognition (NER) tasks that deal with relatively simple classes such as persons, locations, and orga- nizations, the MultiCoNER task involves entities with any linguistic constituents. These include en- tries such as creative works, food, and medical pro- cedures, among others. Furthermore, rather than simply classifying a word as part of a high-level group, the task requires participants to choose the specific type within the group to which the word belongs. For instance, if the word pertains to a per- son, it should be classified as an artist, an athlete, a politician, and so on. The datasets provided for the MultiCoNER II Task included 12 languages, with participants hav- ing the flexibility to submit their results for any language of their choosing. In our submission, we 1https://github.com/emukans/multiconer2-riga focused solely on the most competitive English track and did not submit to other languages. The task at hand is considerably more complex when compared to traditional NER tasks, as all input texts are lower-cased. Thus, our submission primarily relied on the utilization of pre-existing pre-trained large language models (LLMs). The history of machine learning compute can be divided into three eras, namely the Pre-Deep Learn- ing Era, the Deep Learning Era, and the Large- Scale Era (Sevilla et al., 2022). The Large-Scale Era was initiated by the release of AlphaGo (Sil- ver et al., 2016) in late 2015. This trend is likely to have emerged due to large corporations identi- fying new opportunities in the field and making significant investments in it, thereby resulting in considerable increases in model training budgets. This, in turn, marked the beginning of a new era of large-scale language models. Concurrently, it also prompted small institutes with limited budgets to explore novel methods for parameter efficient fine-tuning (PEFT) of LLMs developed by large corporations. Several approaches have been proposed to ad- dress the issue …

Content truncated to 500 words