USTC-NELSLIP at SemEval-2023 Task 2: Statistical Construction and Dual Adaptation of Gazetteer for Multilingual Complex NER

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 64–76 July 13-14, 2023 ©2023 Association for Computational Linguistics Brooke-English at SemEval-2023 Task 5: Clickbait Spoiling Shirui Tang Pingan / 2040179500@qq.com Abstract The task of clickbait spoiling is: generating a short text that satisfies the curiosity induced by a clickbait post. Clickbait links to a web page and advertises its contents by arousing curiosity instead of providing an informative summary. Previous studies on clickbait spoil- ing has shown the approach that classifing the type of spoilers is needed, then generating the appropriate spoilers is more effective on the Webis Clickbait Spoiling Corpus 2022 dataset. Our contribution focused on study of the three classes (phrase, passage and multi) and finding appropriate models to generate spoilers fore- ach class. Results were analysed in each type of spoilers, revealed some reasons of having di- versed results in different spoiler types. "pas- sage" type spoiler was identified as the most difficult and the most valuable type of spoiler. 1 Introduction According to the clickbait spoiling generate re- search paper (Hagen et al., 2022) (Fröbe et al., 2023a) (Fröbe et al., 2023b), clickbait spoiling gen- eration is similar to the task of machine reading comprehension (MRC), where context is the full article/advertises with its title waiting for a reader. Question of MRC corresponds to the clickbait, which is a short text that arouses curiosity with- out providing much informative summary, aiming at attracting people to click into the link towards the context article. And the answer is a spoiling text that could satisfy people’s curiosity raised by the clickbait, therefore people could save the time of reading. Detailed research about the task and its dataset has been conducted in the referring paper. In order to spoil its content, we are listing a couple of key points already verified by the paper and its dataset, and they would be considered as important priors in our experiments. • All the answers can be located as a span in the context, using extraction models could produce the correct answers. • Classifying the spoiler into tree types "phrase", "passage" and "multi" is helpful to the span prediction, under the condition of having a good classifier, otherwise using an unclassified model is a better option. • DeBERTa (He et al., 2020) model showed a general good performance for most of the spoiler-type and score-metric combinations, it is also a good spoiler-type classifier. Although "multi" type was not evaluated, DeBERTa QA model can be regarded as a good initial model for the task fine tune. During data processing some flaw on target span positions were found, target fixing procedure was applied, details would be described in Section3.1. Due to this target fix procedure, direct result com- parison between this paper and the previous re- search paper may not be appropriate. On the other hand, "multi" type spoilers are included in our ex- periment. Within a multi example, multiple text spans that are often locate in different paragraph lines of …

Content truncated to 500 words