Query-driven document partitioning and collection selection

Abstract

Abstract-We present a novel strategy to partition a document collection onto several servers and to perform effective collection selection. The method is based on the analysis of query logs. We proposed a novel document representation called query-vectors model. Each document is represented as a list recording the queries for which the document itself is a match, along with their ranks. To both partition the collection and build the collection selection function, we co-cluster queries and documents. The document clusters are then assigned to the underlying IR servers, while the query clusters represent queries that return similar results, and are used for collection selection. We show that this document partition strategy greatly boosts the performance of standard collection selection algorithms, including CORI, w.r.t. a round-robin assignment. Secondly, we show that performing collection selection by matching the query to the existing query clusters and successively choosing only one server, we reach an average precision-at-5 up to 1.74 and we constantly improve CORI precision of a factor between 11% and 15%. As a side result we show a way to select rarely asked-for documents. Separating these documents from the rest of the collection allows the indexer to produce a more compact index containing only relevant documents that are likely to be requested in the future. In our tests, around 52% of the documents (3,128,366) are not returned among the first 100 top-ranked results of any query.

Document allocation policies for selective searching of distributed indexes

“…Puppin et al [15] used query logs to organize document collection into multiple shards. The query log covered a period of time when exhaustive search was used for each query.…”

An Evaluation Measure for Distributed Information Retrieval Systems

“…Various approximations of relevance have been studied in P2PIR: assuming documents containing all query keywords to be relevant [2], using "approximate descriptions of relevant material" [1] or comparing results of distributed algorithms to results of a centralised system [6,4,9,8].…”

“…those with score > 0) relevant [6] -resulting in what is sometimes called relative recall (RR) -or just the N most highly ranked documents [4,9,8]. In the latter case, precision at k documents is used as an evaluation measurewe will call it P N @k in the rest of this work, denoting its dependence on N .…”

“…In [8], N is chosen equal to k, in [4,9], values of 50 and 100 are used without further justification. Besides the problem of choosing N , this set-based approach also neglects the ranking of the centralised system within the first N documents.…”

“…via author information [2], built-in categories [1] or domains of web pages [4], or it is established in less natural ways via clustering [6] or even randomly [5]. Since generally these collections lack queries and relevance judgments, queries are either constructed from the documents [4,2,1] or taken from query logs matching the collection [6,8].…”