Sakura at SemEval-2023 Task 2: Data Augmentation via Translation

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1718–1722 July 13-14, 2023 ©2023 Association for Computational Linguistics Sakura at SemEval-2023 Task 2: Data Augmentation via Translation Alberto Poncelas Maksim Tkachenko Rakuten Institute of TechnologyOhnmar Htun Rakuten Group, Inc. {first.last}@rakuten.com, Abstract We demonstrate a simple yet effective approach to augmenting training data for multilingual named entity recognition using machine trans- lation. The named entity spans from the origi- nal sentences are transferred to the translations via word alignment and then filtered with the baseline recognizer to retain high quality an- notations. The proposed data augmentation approach improves the baseline performance of XLM-Roberta on the multilingual dataset. 1 Introduction Named Entity Recognition (NER) is an essential task in natural language processing (NLP) that in- volves identifying and classifying entities such as people, organizations, locations, and more recently, complex entities such as events, creative works, diseases. Multilingual Complex Named Entity Recognition (MultiCoNER) shared task (Malmasi et al., 2022a,b) is organized to promote research in multilingual NER for complex entities. Multi- CoNER II (Fetahu et al., 2023b) features one mul- tilingual track and twelve monolingual tracks for the following languages: English, Spanish, Hindi, Bangla, Chinese, Swedish, Farsi, French, Italian, Portuguese, Ukrainian, and German. The datasets contain fine-grained tagsets that mainly cover three domains (Wiki sentences, questions, and search queries) and include six top-level tags: Location (LOC), Creative Work (CW), Group (GRP), Person (PER), Product (PROD), and Medical (MED). Complex named entities are highly ambiguous and can take the form of any linguistic constituency: “Sir” (Creative-Work/Visual-Work), “Catch Me If You Can” (Creative-Work/Visual-Work), “Jeremy Corbyn” (Person/Politician). Thereby, making a granular distinction between entity classes (e.g., Politician vs. Scientist) and identifying rare enti- ties solely from the context are challenging prob- lems. To alleviate this challenge, we propose to augment multilingual training data with transla- tions, which we believe can help the NER model to disambiguate more entities by learning from diverse and noisy contexts. Table 1 shows a few ex- amples of such translations with the aligned entity annotations. The proposed augmentation strategy involves three steps: translating training example, transfer- ring entity annotations from the source sentences to the target, and filtering out low-quality annotations. To produce translations, we use publicly-available mBART-50 (Tang et al., 2020). To transfer en- tity annotations, we compute the token alignment between the source and its translation using SimA- lign (Jalili Sabet et al., 2020) and then build an entity alignments on top of the token alignments. To ensure high-quality annotations, we use XLM- RoBERTA (XLM-R) (Conneau et al., 2020) fine- tuned on the original training data to filter out low-quality annotations from the translated data. We find that XLM-R fine-tuned on the augmented dataset outperforms the same model fine-tuned on the original training dataset. 2 Related Work NER is one of the most popular NLP task and is widely explored across domains: news (Tjong Kim Sang and De Meulder, 2003), biomedical (Settles, 2004), social media (Ritter et al., 2011; Ashwini and Choi, 2014). With the recent …

Content truncated to 500 words