Hitachi at SemEval-2023 Task 3: Exploring Cross-lingual Multi-task Strategies for Genre and Framing Detection in Online News

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1702–1711 July 13-14, 2023 ©2023 Association for Computational Linguistics Hitachi at SemEval-2023 Task 3: Exploring Cross-lingual Multi-task Strategies for Genre and Framing Detection in Online News Yuta Koreeda∗, Ken-ichi Yokote∗, Hiroaki Ozaki, Atsuki Yamaguchi, Masaya Tsunokake and Yasuhiro Sogawa Research and Development Group, Hitachi, Ltd. Kokubunji, Tokyo, Japan {yuta.koreeda.pb, kenichi.yokote.fb, hiroaki.ozaki.yu, atsuki.yamaguchi.xn, masaya.tsunokake.qu, yasuhiro.sogawa.tp}@hitachi.com Abstract This paper explains the participation of team Hitachi to SemEval-2023 Task 3 “Detecting the genre, the framing, and the persuasion tech- niques in online news in a multi-lingual setup.” Based on the multilingual, multi-task nature of the task and the low-resource setting, we inves- tigated different cross-lingual and multi-task strategies for training the pretrained language models. Through extensive experiments, we found that (a) cross-lingual/multi-task training, and (b) collecting an external balanced dataset, can benefit the genre and framing detection. We constructed ensemble models from the re- sults and achieved the highest macro-averaged F1 scores in Italian and Russian genre catego- rization subtasks. 1 Introduction As we pay more and more attention to the so- cially influencing problems like COVID-19 and the Russo-Ukrainian war, there has been an increasing concern about infodemic of false and misleading information (Piskorski et al., 2023). In particular, cross-lingual understanding of such information is becoming more important due to polarization of political stances, economical decoupling and echo chamber effect in social media. To that end, Pisko- rski et al. (2023) put together SemEval-2023 Task 3 “Detecting the genre, the framing, and the persua- sion techniques in online news in a multi-lingual setup.” The shared task aims to analyze several as- pects of what makes a text persuasive and to foster development of building blocks for multilingual media analysis. Creating annotated data for media analysis is time consuming thus we cannot assume that we can obtain training data of enough quality and quan- tity. To tackle the problem, we investigated and compared strategies for multilingual media anal- ysis under a low-resource setting. Through ex- ∗ Equal contribution tensive experiments, we found that (a) cross-lin- gual/multi-task training, and (b) collecting an ex- ternal balanced dataset, can benefit the genre and framing detection. We constructed ensemble mod- els from the results and participated in genre catego- rization (subtask 1) and framing detection (subtask 2) in six languages, where we achieved the highest macro-averaged F1 scores in Italian and Russian subtask 1. 2 Task Definition and our Strategy SemEval-2023 Task 3 aims to analyze several as- pects of what makes a text persuasive. It offers three subtasks on news articles in six languages: German (de), English (en), French (fr), Italian (it), Polish (pl) and Russian (ru). There are three ad- ditional languages (Georgian, Greek and Spanish) without training datasets (i.e., participants need to perform zero-shot language transfer). Subtask 1: News genre categorization Given a news article, a system has to determine whether it is an opinion piece, it aims at objective news reporting, or it is a satire piece. This is multi-class document …

Content truncated to 500 words