Semantic Textual Similarity with Siamese Neural Networks

Proceedings of Recent Advances in Natural Language Processing, pages 1004–1011, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_116 1004 Semantic Textual Similarity with Siamese Neural Networks Tharindu Ranasinghe, Constantin Or˘asan and Ruslan Mitkov Research Group in Computational Linguistics University of Wolverhampton, UK {t.d.ranasinghehettiarachchige, c.orasan, r.mitkov }@wlv.ac.uk Abstract Calculating the Semantic Textual Similar- ity (STS) is an important research area in natural language processing which plays a significant role in many applications such as question answering, document summar- isation, information retrieval and informa- tion extraction. This paper evaluates Sia- mese recurrent architectures, a special type of neural networks, which are used here to measure STS. Several variants of the archi- tecture are compared with existing meth- ods. 1 Introduction Measuring Semantic Textual Similarity (STS) is the task of calculating the similarity between a pair of texts using both direct and indir- ect relationships between them (Rus et al., 2013). Originally, the work on STS largely fo- cused on similarity between short texts such as abstracts and product descriptions (Li et al., 2006; Mihalcea et al., 2006). The introduction of the STS tasks at the International Work- shops on Semantic Evaluation (SemEval) lead to an increase of the interest that the field received from the research community. The SemEval tasks also led to the development of standard datasets like the SICK corpus (Bentivogli et al., 2016) and standardised the similarity score as a numerical value between 1 and 5 (Agirre et al., 2014). Having a good STS metric is very important in many natural language processing (NLP) applications. As an example, for certain types of question answering systems, having an ac- curate STS component is the key to success since the questions with similar meanings can be answered similarly (Majumder et al., 2016). STS is also important in translation memories retrieval and matching (Gupta et al., 2014b). Translation memories help translators by find- ing in the database they maintain previously translated sentences, which are similar to the one to be translated, and retrieving their translations. Hence, accurate STS methods are beneficial for translation memory. Given the growing importance of having a good STS metric and as a result of the Se- mEval workshops, researchers have proposed numerous STS methods. Most of the early approaches were based on traditional machine learning and involved heavy feature engineer- ing (B´echara et al., 2015). With the advances of word embeddings, and as a result of the suc- cess neural networks have achieved in other fields, most of the methods proposed in re- cent years rely on neural architectures (Tai et al., 2015; Shao, 2017). Neural networks are preferred over traditional machine learn- ing models as they generally tend to perform better than traditional machine learning mod- els. They also do not rely on linguistic fea- tures which means they can be easily applied to languages other than English. The ar- chitecture employed in this paper is a spe- cial class of neural networks called Siamese neural networks. These networks contain two or more identical sub-networks. The networks are …

Content truncated to 500 words