HHU at SemEval-2023 Task 3: An Adapter-based Approach for News Genre Classification

Proceedings of the The 17th International Workshop on Semantic Evaluation (SemEval-2023), pages 1166–1171 July 13-14, 2023 ©2023 Association for Computational Linguistics HHU at SemEval-2023 Task 3: An Adapter-based Approach for News Genre Classification Fabian Billert and Stefan Conrad Heinrich-Heine University of Düsseldorf {fabian.billert, stefan.conrad}@hhu.de Abstract This paper describes our approach for Subtask 1 of Task 3 at SemEval-2023. In this subtask, task participants were asked to classify multi- lingual news articles for one of three classes: Reporting, Opinion Piece or Satire. By training an AdapterFusion layer composing the task- adapters from different languages, we success- fully combine the language-exclusive knowl- edge and show that this improves the results in nearly all cases, including in zero-shot scenar- ios. 1 Introduction In SemEval-2023 Task-3 (Piskorski et al., 2023), the authors intend to find out what makes a news article persuasive. To this end, they designed three subtasks: News Genre Categorisation, Framing De- tection and Persuasion Techniques Detection. In this paper, we present our results for the first sub- task, which aims to classify the genre of a news article into one of three classes; reporting, opinion piece or satire. The news articles are available in different languages, meaning this is a multilingual problem. Furthermore there are a few languages without training data in order to test the perfor- mance of the models in a zero-shot scenario. The results for the different languages were evaluated separately while taking the F1-macro score as the evaluation metric. In the following, we present our approach for En- glish, Italian, German and Russian, as well as the zero-shot languages Spanish and Greek. We use an adapter-based setup (Pfeiffer et al., 2020b), which enhances BERT-type architectures (Devlin et al., 2018) with modular weights. We train a task-adapter for each of the languages before using AdapterFusion (Pfeiffer et al., 2020a) to combine what the language-specific task adapters learned. We show that using AdapterFusion largely im- proves the performances of the single-language task-adapters, which is a behavior that has already been observed for low-resource tasks. With an F1- macro score of 0.5942, we achieved 4th place out of 27 teams for the English dataset. In addition, for the zero-shot language Greek, we came in third place out of 15 teams with an F1-macro score of 0.7504. For the other languages, we ranked near 10th place, meaning that our approach still shows some instabilities. 2 Background The trainig data for subtask-1 consists of whole texts in six different languages: English, Italian, German, French, Polish and Russian. As we noted above, we only submitted results for English, Ital- ian, German and Russian. We will elaborate on the reason for this in subsection 3.4. However, we still used the French and Polish datasets to augment our data, as we explain in the next section. The main challenge of this subtask is presented by the limited availability, as well as the imbal- ance of the training data. Here, the opinion label is over-represented (76% of the labels) while the satire label is the clear …

Content truncated to 500 words