Improving Quality and Domain-Relevancy of Paraphrase Generation with Graph-Based Retrieval Augmented Generation
Lasal Jayawardena, Prasan Yapa · 2024
Paraphrase generation is a fundamental area of research in Natural Language Processing (NLP) and Natural Language Generation (NLG), due to its sequence-to-sequence (Seq2Seq) nature. Paraphrasing, spanning across various domains, poses challenges for simpler model architectures due to the extensive knowledge required to generate paraphrases. The added constraint of generating diverse paraphrases further complicates the task for models trained on existing datasets. We present a methodology that leverages Graph-Based Retrieval Augmented Generation (G-RAG), capable of utilizing both entity and phrasal knowledge to address this issue. We demonstrate through experiments that this approach enables both complex models like Large Language models (LLMs) and smaller Seq2Seq models to generate more diverse paraphrases without compromising semantic similarity. Furthermore, this approach’s capacity to integrate domain-specific knowledge makes it particularly effective across different domains, enhancing its applicability in varied contexts. The results are further corroborated by human evaluation and extensive quantitative analysis focusing on semantic similarity, lexical diversity, syntactic diversity, and grammatical correctness to gauge high-quality paraphrases.