KeyColBERT-PRF: Optimized Text Retrieval via Keyword Augmentation
Min Pan, Yumin Xie, Shuting Zhou, Mengfei Gao, Wenrui Xiong, Mao Wang · 2024
Pseudo-relevance feedback (PRF) techniques are widely employed in information retrieval, but existing PRF methods for dense retrieval frequently encounter issues with semantic information loss. In this paper, we propose the KeyColBERT-PRF, a novel approach designed to improve search accuracy through query expansion using keyword extraction. Specifically, KeyBERT is utilized to identify keywords that are highly relevant to the original query from pseudo-relevant documents, which are then incorporated into the original query to generate an expanded query. To evaluate the effectiveness of KeyColBERT-PRF, we conducted experiments on the TREC 2019 and TREC 2020 datasets, comparing our method with baseline models. Additionally, we explore the influence of various parameters such as the number of pseudo-relevant documents and the weighting of query expansion on system performance. Results demonstrate that the KeyColBERT-PRF method significantly enhances retrieval accuracy by effectively combining neural ranking with keyword-based query expansion.