Countering Misinformation Through Semantic-Aware Multilingual Models

ABSTRACT

The presence of misinformation and harmful content on social networks is an emerging problem that endangers public health. One of the most successful approaches for detecting, assessing, and providing prompt responses to this misinformation problem is Natural Language Processing (NLP) techniques based on semantic similarity. However, language constitutes one of the most significant barriers to address, denoting the need to develop multilingual tools for an effective fight against misinformation. This paper presents an approach for countering misinformation through a semantic-aware multilingual architecture. Due to the specificity of the task addressed, which involves assessing the level of similarity between a pair of texts in a multilingual scenario, we built an extension of the well-known Semantic Textual Similarity Benchmark (STSb) to 15 languages. This new dataset allows to fine-tune and evaluate multilingual models based on Transformers with a siamese network topology on monolingual and cross-lingual Semantic Textual Similarity (STS) tasks, achieving a maximum average Spearman correlation coefficient of 83.60%. We validate our proposal using the Covid-19 MLIA @ Eval Multilingual Semantic Search Task. The results reported demonstrate that semantic-aware multilingual architectures are successful at measuring the degree of similarity between pairs of texts, while broadening our understanding of the multilingual capabilities of this type of models. The results and the new multilingual STS Benchmark data presented and made publicly in this study constitute an initial step towards extending methods proposed in the literature that employ semantic similarity to combat misinformation at a multilingual level. © 2021, Springer Nature Switzerland AG.