DSLR: Doc Refinement with Sentence-Degree Re-ranking and Reconstruction to Improve Retrieval-Augmented Technology
Authors: Taeho Hwang, Soyeong Jeong, Sukmin Cho, SeungYoon Han, Jong C. Park
Summary: Latest developments in Giant Language Fashions (LLMs) have considerably improved their efficiency throughout varied Pure Language Processing (NLP) duties. Nevertheless, LLMs nonetheless battle with producing non-factual responses attributable to limitations of their parametric reminiscence. Retrieval-Augmented Technology (RAG) techniques deal with this concern by incorporating exterior information with a retrieval module. Regardless of their successes, nevertheless, present RAG techniques face challenges with retrieval failures and the restricted potential of LLMs to filter out irrelevant data. Subsequently, on this work, we suggest DSLR (Doc Refinement with Sentence-Degree Re-ranking and Reconstruction), an unsupervised framework that decomposes retrieved paperwork into sentences, filters out irrelevant sentences, and reconstructs them once more into coherent passages. We experimentally validate DSLR on a number of open-domain QA datasets and the outcomes exhibit that DSLR considerably enhances the RAG efficiency over typical fixed-size passage. Moreover, our DSLR enhances efficiency in particular, but reasonable situations with out the necessity for extra coaching, offering an efficient and environment friendly answer for refining retrieved paperwork in RAG system
