A Novel Paraphrase Generation Model Using Semantically and Syntactically Controlled Structures

Natural Language Processing (NLP) is an integral part of machine-to-human conversation. It requires a deep understanding of diverse human utterances. This expectation from machines encouraged the researchers to study the behavior of machines in order to understand NLP. Because humans can understand the semantics of a word during conversation, but making a machine understand it is a difficult task. Paraphrasing is a way to create multiple forms of human utterances. Thus, it is capable enough for the machines to understand the utterances while communicating with humans. Paraphrasing is a comprehensive task that involves the replacement of words or phrases and high-level syntactic rearrangement processes such as passiveness, left-splitting, and topicalization. So, this research aims to develop a novel hybrid approach that incorporates rules with machine learning techniques to generate diverse paraphrases without the use of labeled paraphrased datasets. The model is trained and evaluated using two datasets namely Quora and mobile service provider queries dataset. In this paper, two algorithms are designed; one is used to create a labeled template-based dataset, and another algorithm to produce a high level of diverse syntactic paraphrases. The evaluation shows that the proposed approaches generate more high-quality paraphrases than baselines along with an increase in the diversity of paraphrases. The proposed model attained the ROUGE-1 value of 66 which is higher than the other baseline models. Thus, it proved that the proposed approach can effectively improve its performance in both syntactic and semantic structure during paraphrase generation.