IIIT-MLNS at SemEval-2022 Task 8: Siamese Architecture for Modeling Multilingual News Similarity

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1145 - 1150 July 14-15, 2022 ©2022 Association for Computational Linguistics IIIT-MLNS at SemEval-2022 Task 8: Siamese Architecture for Modeling Multilingual News Similarity Sagar Joshi∗, Dhaval Taunk∗, Vasudeva Varma IIIT Hyderabad, India {sagar.joshi, dhaval.taunk}@research.iiit.ac.in vv@iiit.ac.in Abstract The task of multilingual news article similarity en- tails determining the degree of similarity of a given pair of news articles in a language-agnostic setting. This task aims to determine the extent to which the articles deal with the entities and events in ques- tion without much consideration of the subjective aspects of the discourse. Considering the supe- rior representations being given by these models as validated on other tasks in NLP across an array of high and low-resource languages and this task not having any restricted set of languages to focus on, we adopted using the encoder representations from these models as our choice throughout our experiments. For modeling the similarity task by using the representations given by these models, a Siamese architecture was used as the underlying architecture. In experimentation, we investigated on several fronts including features passed to the encoder model, data augmentation and ensembling among our major experiments. We found data aug- mentation to be the most effective working strategy among our experiments. 1 Introduction News articles from the web covering the same event tend to differ on regional and political biases, style of writing, conciseness and preciseness of cover- age and the intended audience of the news outlet. Misleading and confusing articles might lead to un- necessary confusion, chaos and tensions potentially exacerbating or even creating non-existent issues. Often, news articles covering the same event in dif- ferent languages from varied regional sources are readily available. This can apply for happenings of local, regional as well as international signifi- cance. In such a scenario, it is pertinent to able to identify or cluster together news articles covering *These authors contributed equally to this work. the same story in different languages and different view points while also being able to distinguish be- tween articles similar in style but covering different happenings. The significance of the problem of multilingual news article similarity (Chen et al., 2022) lies to- wards the applicability in this area by scoring a pair of news articles based on their similarity in coverage of the event without focusing on the sub- jectivity of the content. In our approach, we model the task as a regres- sion problem by making use of a Siamese neural network (Bromley et al., 1993) as the base architec- ture. We perform experiments in feature engineer- ing by making use of metadata such as title, descrip- tion, keywords and tags in addition to the news text. We also experimented with artificially augmenting the data by document permutation. Also, we tried with ensembling of two models - one trained us- ing only textual features, the other using metadata information. While data augmentation strategy did give a de- cent improvement in performance, considering the …

Content truncated to 500 words