GateNLP-UShef at SemEval-2022 Task 8: Entity-Enriched Siamese Transformer for Multilingual News Article Similarity

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1121 - 1128 July 14-15, 2022 ©2022 Association for Computational Linguistics GateNLP-UShef at SemEval-2022 Task 8: Entity-Enriched Siamese Transformer for Multilingual News Article Similarity Iknoor Singh, Yue Li, Melissa Thong and Carolina Scarton Department of Computer Science, University of Sheffield (UK) {i.singh, yli381, mlthong1, c.scarton}@sheffield.ac.uk Abstract This paper describes the second-placed system on the leaderboard of SemEval-2022 Task 8: Multilingual News Article Similarity. We pro- pose an entity-enriched Siamese Transformer which computes news article similarity based on different sub-dimensions, such as the shared narrative, entities, location and time of the event discussed in the news article. Our system exploits a Siamese network architecture using a Transformer encoder to learn document-level representations for the purpose of capturing the narrative together with the auxiliary entity- based features extracted from the news articles. The intuition behind using all these features together is to capture the similarity between news articles at different granularity levels and to assess the extent to which different news outlets write about “the same events”. Our ex- perimental results and detailed ablation study demonstrate the effectiveness and the validity of our proposed method. 1 Introduction News article similarity measures could facilitate various important tasks such as the clustering of news (Montalvo et al., 2007; Azzopardi and Staff, 2012), duplicate news detection (Alonso et al., 2013; Gibson et al., 2008; Singh et al., 2021a; Theobald et al., 2008), fact-checking (Hassan et al., 2017; Jiang et al., 2021) and tracking of the spread of news (Zhai and Shah, 2005). SemEval-2022 task 8 (Chen et al., 2022) assesses the similarity be- tween news articles in terms of the real world hap- penings. Therefore, it mainly considers the loca- tion, time, entities and narratives, instead of writing style, political spin, or tone of the article. The train- ing set comprises of monolingual and cross-lingual news articles pairs in 10 different languages. The main contributions in this paper are: 1) We propose an entity-enriched Siamese Transformer whose general idea is to enable the model to ex- plicitly learn from entity features (geolocation, or- ganization, date and quantity) that are crucial to determine the similarity between news events but difficult to extract directly from the language mod- els during fine-tuning. 2) We explore different data augmentation approaches through semi-supervised learning in order to tackle the imbalanced data prob- lem. 3) We compare our proposed model with strong baselines and conduct an ablation study to analyse the contribution of each component in our model. We also present an error analysis at the end of this paper. Our best system which exploits Lan- guage Agnostic BERT Sentence Representations (Feng et al., 2020) ranks 2nd in the competition. 2 Background The goal of the task is to predict similarity scores ranging from 1 (most similar) to 4 (least similar) for a give news article pair. The training data consists of 4,964 article pairs in seven distinct languages (English, German, Spanish, Turkish, Polish, Ara- bic and French), with seven …

Content truncated to 500 words