Non-Linear Similarity Learning for Compositionality

Many NLP applications rely on the existence of similarity measures over text data. Although word vector space models provide good similarity measures between words, phrasal and sentential similarities derived from composition of individual words remain as a difficult problem. In this paper, we propose a new method of of non-linear similarity learning for semantic compositionality. In this method, word representations are learned through the similarity learning of sentences in a high-dimensional space with kernel functions. On the task of predicting the semantic relatedness of two sentences (SemEval 2014, Task 1), our method outperforms linear baselines, feature engineering approaches, recursive neural networks, and achieve competitive results with long short-term memory models.