[4P3-OS-17c] OS-17

[4P3-OS-17c-04] Data Augmentation for Retriever Using Generative Model

〇Ayaka Yomogida¹, Kazutada Ban¹, Tetsuya Iida¹, Fumihiko Murase¹, Akira Mitani¹, Takanori Takeno¹, Toru Hirano² (1. DENSO CORPORATION, 2. DENSO INTERNATIONAL AMERICA, INC.)

Keywords:Data Augmentation, Retiever, Generative Model

The Retrieval-Augmented-Generation (RAG) method is gaining attention for enhancing language generation ability in a specific domain. Although retriever using dense embedding, which have reported high accuracy, are critical for the overall accuracy, the challenge has been the high burden of preparing supervised data for the target domain. In this study, we considered data augmentation using the GPT-3.5 and clarified that queries that presuppose a document and queries without clear answers in the document . The proposed method confirmed the effect of improving the accuracy of the retriever with less data by removing such data from the generation results.