Exploiting Siamese Neural Networks on Short Text Similarity Tasks for Multiple Domains and Languages

Semantic textual similarity algorithms are essential to several natural language processing tasks as clustering documents and text summarization. Many shared tasks regarding this subject were performed during the last few years, but generally, focused on a unique domain and/or language. Siamese Neural Network (SNN) is well known for its ability to compute similarity requiring less training data. We proposed a SNN architecture incorporated with language-independent features, aiming to perform short text similarity calculation in multiple languages and domains. We explored three different corpora from shared tasks: ASSIN 1 and ASSIN 2 with Portuguese journalistic texts and N2C2 (English clinical texts). We adapted theSNNproposed by Mueller and Thyagarajan (2016), in twoways: (i) the activation functions were changed to the ReLU, instead of the sigmoid function, and; (ii) we incorporated the architecture to accept three new lexical features and an embedding layer to infer the values of the pre-trained word embeddings. The evaluation was performed by the Pearson correlation (PC) and the Mean Squared Error (MSE) between the models' predicted values and corpora's gold standard. Our approach achieved better results than the baseline in both languages and domains.