LSTM based paraphrase identification using combined word embedding features

Abstract

Paraphrase identification is the process of analyzing two text entities (sentences) and determining whether the two entities represent the similar sense or not. This is a task of Natural Language Processing (NLP) in which we need to identify the sentences whether it is a paraphrase or not. Here, the chosen approach for this task is a deep Learning model that is Recurrent Neural Network-LSTM with word embedding features. Word embedding is an approach, from where we can extract the semantics of the word in dense vector representation. The word embedding models that are used for the feature extraction in Telugu are Word2Vec, Glove and Fasttext. These extracted feature models are added in the embedding layer of Long Short-Term Memory algorithm in order to classify the Telugu sentence pairs whether they are Paraphrase or not. The corpus for Telugu is generated manually from various Telugu newspapers. The sentences for word embedding model is also gathered from Telugu newspapers. This is the first attempt for paraphrase identification in Telugu using deep learning approach.