Recurrent Pseudo Relevance Feedback on Web Collections

Various Relevance Feedback techniques exist in Information Retrieval such as Simulated Relevance Feedback and Pseudo Relevance Feedback. In a Simulated Relevance Feedback technique a new query is reformulated based on the documents selected by the user from the top-ranked documents whereas in a Pseudo Relevance Feedback, the query is reformulated based on the assumption that N top-ranked documents are relevant. A new Relevance Feedback technique has been developed in this paper which provides better retrieval effectiveness than Pseudo Relevance Feedback at high precision levels at low recall levels. The new technique, called Recurrent Pseudo Relevance Feedback (RPRF), goes into a repetitive process by considering the top-ranked documents that were initially retrieved from Pseudo Relevance Feedback to be the new top-ranked documents. RPRF shows greater retrieval effectiveness in the Probabilistic model using the DB2 model in the context of the Divergence from Randomness (DFR) framework. However, it is also shown that retrieval effectiveness increases as the value of N decreases on the second feedback process. The effectiveness of the results is assessed based on the measurement of recall and precision ratios using the Interpolated Mean Average Precision (MAP) and was tested on the Text Retrieval Conference (TREC) Web collections of 10 and 18 Gigabytes WT10G and GOV .