Automated Paraphrase Generation with Over-Generation and Pruning Services

Conversational services are emerging as a new paradigm for accessing information by simply uttering questions in natural language, posing whole set of challenges to the design and engineering systems. Training conversational deal with nuances language often requires collecting high-quality diverse training samples (i.e., paraphrases). Traditional approaches such hiring an expert or crowdsourcing involve data collection processes that costly time-consuming. Automated paraphrase generation is promising cost-effective scalable approach generating samples. Current automatic techniques, however, tend specialise specific types lexical syntactic variations. As result, generated paraphrases may not perform well relevant quality aspects diversity semantic relatedness. In this paper, we follow inspired integration address these issues generate English semantically diverse. We propose extensible reusable pipeline combines paraphrasing techniques two-step process first focus on i) leveraging strengths multiple most (and possibly noisy) paraphrases, then ii) common separate step. Through empirical evaluations show benefits combining more balancing relevance diversity.