CAIR-NLP at SemEval-2023 Task 2: A Multi-Objective Joint Learning System for Named Entity Recognition

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1926–1935 July 13-14, 2023 ©2023 Association for Computational Linguistics CAIR-NLP at SemEval-2023 Task 2: A Multi-Objective Joint Learning System for Named Entity Recognition Sangeeth N1,2, Biswajit Paul2, Chandramani Chaudhary1 1 National Insititute of Technology, Tiruchirappalli, India 2 Centre for Artificial Intelligence and Robotics, CV Raman Nagar, Bangalore, India 406120005@nitt.edu, biswajit.cair@gov.in, chandramani@nitt.edu Abstract This paper describes the NER system designed by the CAIR-NLP team for submission to Mul- tilingual Complex Named Entity Recognition (MultiCoNER II) shared task, which presented a novel challenge of recognizing complex, ambiguous, and fine-grained entities in low- context, multi-lingual, multi-domain dataset and further evaluation on the noisy subset. We propose a Multi-Objective Joint Learning Sys- tem (MOJLS) for NER, which aims to enhance the representation of entities and improve la- bel predictions through joint implementation of a set of learning objectives. Our official sub- mission MOJLS implements four objectives. These include, representation of the named en- tities should be close to its entity type definition, low-context inputs should have representation close to their augmented context, and also min- imization of two label prediction errors, one based on CRF and another biaffine based pre- dictions, where both are producing distribu- tions over the output labels. The official results ranked our system 2nd in five tracks (Multilin- gual, Spanish, Swedish, Ukrainian, and Farsi) and 3rd in three tracks (French, Italian, and Portuguese) out of 13 tracks. Also evaluation on the noisy subset, our model achieved rela- tively better ranks. Official ranks indicate the effectiveness of the proposed MOJLS in deal- ing with the contemporary challenges of NER. 1 Introduction Named Entity Recognition (NER) is an estab- lished natural language processing task that aims to locate entity mentions by marking entity-span boundaries and further classify them into corre- sponding entity types from a set of pre-defined entity categories in unstructured text (Wang et al., 2022b). NER is a critical building block in In- formation Extraction, and Knowledge-base con- struction pipeline (Lample et al., 2016). The Se- mEval 2023 Task 2: Multilingual Complex Named Entity Recognition (MultiCoNER-II) presented a novel challenge of building a NER system cover- ing 33 fine-grained categories over 12 languages, namely English, French, Spanish, German, Italian, Portuguese, Ukrainian, Swedish, Chinese, Farsi, Hindi, Bangla and also in multilingual setting in 13 tracks (Fetahu et al., 2023b). MultiCoNER-II task presents a host of real-world contemporary challenges in terms of (1) Low-context sentences (short and uncased text), (2) Syntactically complex and ambiguous fine-grained entities, (3) Entities having a large long-tail distribution and evolving nature, (4) Multi-linguality, (5) Evaluation on a noisy subset, where inputs are corrupted with noise either on context tokens or entity tokens, and (6) Limited training examples per language. In this paper, we propose a Multi-Objective Joint Learning System (MOJLS) for NER, which aims to learn an enhanced representation of low-context, fine-grained entities and thereby improves recog- nition. We considered four training objectives. Minimization of (1) Representation gaps between the entities to the corresponding entity …

Content truncated to 500 words