Search papers, labs, and topics across Lattice.
The paper introduces DTCRS, a dynamic tree construction method for recursive summarization in RAG, which analyzes question type and decomposes questions into sub-questions to guide the clustering of text chunks. By using embeddings of sub-questions as initial cluster centers, DTCRS reduces redundant summary nodes and improves the relevance of summaries to the question. Experiments show DTCRS significantly reduces summary tree construction time and improves performance on three QA tasks, while also providing insights into the applicability of recursive summarization across different question types.
Forget static summarization trees – DTCRS dynamically constructs them based on question type and semantics, slashing construction time and boosting QA accuracy.
Retrieval-Augmented Generation (RAG) mitigates the hallucination problem of Large Language Models (LLMs) by incorporating external knowledge. Recursive summarization constructs a hierarchical summary tree by clustering text chunks, integrating information from multiple parts of a document to provide evidence for abstractive questions involving multi-step reasoning. However, summary trees often contain a large number of redundant summary nodes, which not only increase construction time but may also negatively impact question answering. Moreover, recursive summarization is not suitable for all types of questions. We introduce DTCRS, a method that dynamically generates summary trees based on document structure and query semantics. DTCRS determines whether a summary tree is necessary by analyzing the question type. It then decomposes the question and uses the embeddings of sub-questions as initial cluster centers, reducing redundant summaries while improving the relevance between summaries and the question. Our approach significantly reduces summary tree construction time and achieves substantial improvements across three QA tasks. Additionally, we investigate the applicability of recursive summarization to different question types, providing valuable insights for future research.
