Improving Semantic Textual Similarity with Phrase Entity Alignment

Abstract

Semantic Textual Similarity (STS) measures the degree of semantic equivalence between two segments of text, even though the similar context is expressed using different words. The textual segments are word phrases, sentences, paragraphs or documents. The similarity can be measured using lexical, syntactic and semantic information embedded in the sentences. The STS task in SemEval workshop is viewed as a regression problem, where real-valued output is clipped to the range 0-5 on a sentence pair. In this paper, empirical evaluations are carried using lexical, syntactic and semantic features on STS 2016 dataset. A new syntactic feature, Phrase Entity Alignment (PEA) is proposed. A phrase entity is a conceptual unit in a sentence with a subject or an object and its describing words. PEA aligns phrase entities present in the sentences based on their similarity scores. STS score is measured by combing the similarity scores of all aligned phrase entities. The impact of PEA on semantic textual equivalence is depicted using Pearson correlation between system generated scores and the human annotations. The proposed system attains a mean score of 0.7454 using random forest regression model. The results indicate that the system using the lexical, syntactic and semantic features together with PEA feature perform comparably better than existing systems.