High precision retrieval using relevance-flow graph
Jangwon Seo, Jiwoon Jeon · 2009
Traditional bag-of-words information retrieval models use aggregated term statistics to measure the relevance of documents, making it difficult to detect non-relevant documents that contain many query terms by chance or in the wrong context. In-depth document analysis is needed to filter out these deceptive documents. In this paper, we hypothesize that truly relevant documents have relevant sentences in predictable patterns. Our experimental results show that we can successfully identify and exploit these patterns to significantly improve retrieval precision at top ranks.