Cross-Lingual Semantic Textual Similarity Modeling Using Neural Networks

Cross-lingual semantic textual similarity is to measure the semantic similarity of sentences in different languages. Previous work pay more attention on leveraging traditional NLP features (e.g., alignment features, syntactic features) to evaluate the semantic similarity of sentences. In this paper, we only use word embedding as basic features without any handcrafted features and build a model which is able to capture local and global semantic information of the sentences to evaluate semantic textual similarity. We test our model on SemEval-2017 and STS benchmark datasets. Our experiments show that our model improves the performance of the semantic textual similarity and achieves the best results compared with the baseline neural-network based methods reported on the two datasets.