Applying completely-arbitrary passage for pseudo-relevance feedback in language modeling approach

摘要: Different from the traditional document-level feedback, passage-level feedback restricts context of selecting relevant terms to a passage in document, rather than entire document. It can thus avoid selection nonrelevant non-relevant parts The most recent work has been investigated viewpoint fixed-window type passage. However, limitation optimizing since it includes query-independent portion. To minimize query-independence passage, this paper proposes new called completely-arbitrary Based on this, we devise novel two-stage - which consists passage-retrieval and passage-extension as sub-steps, unlike previous single-stage relying only retrieval. Experimental results show that proposed much significantly improves uses