L3i at SemEval-2022 Task 11: Straightforward Additional Context for Multilingual Named Entity Recognition

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1630 - 1638 July 14-15, 2022 ©2022 Association for Computational Linguistics L3i at SemEval-2022 Task 11: Straightforward Additional Context for Multilingual Named Entity Recognition Emanuela Boros University of La Rochelle, L3i La Rochelle, France emanuela.boros@univ-lr.fr Carlos-Emiliano González-Gallardo University of La Rochelle, L3i La Rochelle, France carlos.gonzalez_gallardo@univ-lr.fr Jose G. Moreno University of Toulouse, IRIT Toulouse, France jose.moreno@irit.fr Antoine Doucet University of La Rochelle, L3i La Rochelle, France antoine.doucet@univ-lr.fr Abstract This paper summarizes the participation of the L3i laboratory of the University of La Rochelle in the SemEval-2022 Task 11, Mul- tilingual Complex Named Entity Recognition (MultiCoNER). The task focuses on detect- ing semantically ambiguous and complex enti- ties in short and low-context monolingual and multilingual settings. We argue that using a language-specific and a multilingual language model could improve the performance of mul- tilingual and mixed NER. Also, we consider that using additional contexts from the training set could improve the performance of a NER on short texts. Thus, we propose a straightfor- ward technique for generating additional con- texts with and without the presence of entities. Our findings suggest that, in our internal ex- perimental setup, this approach is promising. However, we ranked above average for the high- resource languages and lower than average for low-resource and multilingual models. 1 Introduction Named entity recognition (NER) is the task of de- tecting entities and recognizing their type (e.g., per- son, location, organization) (Grishman and Sund- heim, 1996). Standard benchmarks (e.g., CoNLL- 2003 (Tjong Kim Sang and De Meulder, 2003)) for NER are either focused on a high-resource lan- guage or on a single domain (e.g., news). However, NER is required in different fields, bringing thus several new challenges regarding: the size of the textual content (e.g., short textual snippets like Web queries (Fetahu et al., 2021) or social media posts (Rajoria, 2021)), the characteristics of the entities to be extracted (i.e., capitalization/punctuation fea- tures (Mayhew et al., 2019)), unseen and emerging entities (e.g., entities that have a fast growth rate such as new songs and movies are released weekly (Bernier-Colborne and Langlais, 2020)), complex entities (i.e., complex noun phrases such as particu- larly long nested person names and dates in histori- cal documents (Boros et al., 2020a,c)). Moreover, multilingual documents in which entities of differ- ent languages than the rest of the text are present (i.e., code-mixed documents (Winata et al., 2021; Fetahu et al., 2021)) add another level of complex- ity to the task. Transformer-based (Vaswani et al., 2017) architectures for NER became popular since the release of the BERT model and they hold the state of the art in NER (Akbik et al., 2018; Nie et al., 2020; Yamada et al., 2020; Wang et al., 2020, 2021). However, while most NER systems have been developed to generally address contemporary news data in high resource languages (e.g. English) (Yamada et al., 2020; Wang et al., 2020, 2021), many challenges remain in NER from short texts with complex entities in …

Content truncated to 500 words