SU-NLP at SemEval-2022 Task 11: Complex Named Entity Recognition with Entity Linking

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1648 - 1653 July 14-15, 2022 ©2022 Association for Computational Linguistics SU-NLP at SemEval-2022 Task 11: Complex Named Entity Recognition with Entity Linking Buse Çarık∗, Fatih Beyhan∗, Reyyan Yeniterzi Sabancı University, ˙Istanbul, Turkey {busecarik,fatihbeyhan,reyyan}@sabanciuniv.edu Abstract This paper describes the system proposed by Sabancı University Natural Language Process- ing Group in the SemEval-2022 MultiCoNER task. We developed an unsupervised entity link- ing pipeline that detects potential entity men- tions with the help of Wikipedia and also uses the corresponding Wikipedia context to help the classifier in finding the named entity type of that mention. The proposed pipeline signif- icantly improved the performance, especially for complex entities in low-context settings. 1 Introduction Named Entity Recognition (NER) is a widely stud- ied task of Natural Language Processing. Recent leading architectures achieved impressive perfor- mances, especially in formal contexts (like news articles, etc.) and for commonly worked named entity types (like Person, Organization, etc.). In Multilingual Complex Named Entity Recognition task (Malmasi et al., 2022b), organizers introduced a dataset which is a large collection of short texts that includes complex entities. As they stated, com- plex entities can be noun phrases, gerunds, or full clauses (Malmasi et al., 2022a). The participants were challenged to develop NER systems for iden- tifying these complex entities in short and low- context settings. Recent contextual neural architectures like trans- formers work quite well on a variety of NLP tasks and datasets. However, they are strongly dependent on the context of the given text. In case of lack of context, these models alone may not gather enough evidence to make a correct prediction. When that is the case, getting help from knowledge bases can be useful. The dataset provided with this task has sim- ilar low-context characteristics together with some complex named entity types to be resolved. There- fore, in this work, we investigated whether perform- ing entity linking to provide additional context for *These authors contributed equally to this work. a transformer model can be useful for identifying complex named entities. We initially used an un- supervised entity linking approach to identify the possible entity mentions and their corresponding Wikipedia pages. Later on, we used the additional context retrieved from Wikipedia to predict the types of the identified entities. Performing entity linking prior to the classifier seems to be beneficial in the case of complex entities in short and low- context settings. Our code is publicly available*. 2 Related Work NER is an important and also well-studied task in NLP. Similar to other NLP tasks, recent neural ar- chitectures provided significant improvements in NER as well. However, since these models are heavily dependent on context information, apply- ing them to short, noisy texts such as very short social media posts or web queries results in signifi- cant performance degradation (Meng et al., 2021). To overcome the lack of context in these short texts, as well as the code-mixed terms, it has been prominent to use external knowledge …

Content truncated to 500 words