Y-X-Y Encoding for Identifying Types of Sentence Similarity

Abstract

Determining the semantic similarity of any two arbitrary sentences requires two steps, i.e. sentence encoding and semantic similarity measure. The most important step is to encode a set of sentences into a set of equal-length vectors for similarity measure in forms of classification. Two practical encoding schemes had been proposed, statistical-based direct encoding and pretrained encoding. The first approach lacks considering word correlation and the dimension of encoded vector is very large. For the second approach, it requires an extra training time prior to the classification process. This study compromises the previous approaches by considering shallow neural networks for encoding sentences and classifying entailment relations between two sentences. A set of y-x-y encoder models is proposed where y can be greater or less than x depending on given dataset. Neither encoder models nor their corresponding classifiers are built upon big and complex structure, and hence is suitable for carrying out such task. The encoding scheme is tested with SICK 2014 dataset [1], specially designed for neutral, entailment, and contradiction sentence pairs. Comparison results (neutral 97.1%, entailment 91.1%, contradiction 94.6%) support the possibility of the proposed scheme to sentence similarity measure.