A Spectral Learning Based Model to Evaluate Semantic Textual Similarity

Abstract

Semantic Textual Similarity (STS) is a task in NLP that compares two sentences in a sentence-pair and scores the relationship between them using the degree of semantic equivalence. It has wide applicability in various fields. Consequently, the research around the task is constantly evolving. The demand for new as well as improved methods is endless. Numerous methods have been proposed that largely belong to either unsupervised or supervised learning approaches. The model proposed here is fairly simple and provides a fresh take on this classification problem using spectral learning. The model does not engage a large labeled corpus or lexical database like most STS supervised and unsupervised methods. Although, supervised STS methods achieve an accuracy that outperforms humans in some cases, but are often held back due to a lack of interpretation of the features instrumental in molding the decision-making process. The proposed model on the other hand generates features (latent knowledge) that are easy to ascertain and have a mathematical foundation. Given a sentence pair, the work focuses on finding latent states and variables from each sentence and performs classification by generating a similarity score. The latent variables are a result of projections learned by performing Canonical Correlation Analysis (CCA) amongst the sentence pair. To perform matching and determine the similarity score, Cosine similarity and Word Mover’s Distance (WMD) are employed. The performance of the proposed model does exhibit an improvement over various sophisticated supervised techniques such as LSTM and BiLSTM.