NPRF: A Neural Pseudo Relevance Feedback Framework for Ad-hoc Information Retrieval

Canjia Li,Ben He,Yingfei Sun,Andrew Yates,Le Sun,Jungang Xu,Le Wang,Kai Hui

doi:10.18653/v1/d18-1478

Abstract

Pseudo relevance feedback (PRF) is commonly used to boost the performance of traditional information retrieval (IR) models by using top-ranked documents to identify and weight new query terms, thereby reducing the effect of query-document vocabulary mismatches. While neural retrieval models have recently demonstrated strong results for ad-hoc retrieval, combining them with PRF is not straightforward due to incompatibilities between existing PRF approaches and neural architectures. To bridge this gap, we propose an end-to-end neural PRF framework that can be used with existing neural IR models by embedding different neural models as building blocks. Extensive experiments on two standard test collections confirm the effectiveness of the proposed NPRF framework in improving the performance of two state-of-the-art neural IR models.

Highlights

Recent progress in neural information retrieval models (NIRMs) has highlighted promising performance on the ad-hoc search task
Existing neural IR models do not have a mechanism for treating expansion terms differently from the original query terms, making it non-trivial to combine them with existing pseudo relevance feedback (PRF) approaches
We instantiate the neural framework for pseudo relevance feedback (NPRF) framework using two state-of-the-art neural IR models, and we evaluate their performance on two widely-used TREC benchmark datasets for ad-hoc retrieval

Summary

Introduction

Recent progress in neural information retrieval models (NIRMs) has highlighted promising performance on the ad-hoc search task. State-of-theart NIRMs, such as DRMM (Guo et al, 2016), HiNT (Fan et al, 2018), (Conv)-KNRM (Xiong et al, 2017; Dai et al, 2018), and (Co)PACRR (Hui et al, 2017, 2018), have successfully implemented insights from traditional IR models using neural building blocks. PRF models expand the query with terms selected from top-ranked documents, thereby boosting ranking performance by reducing the problem of vocabulary mismatch between the original query and documents (Rocchio, 1971). Existing neural IR models do not have a mechanism for treating expansion terms differently from the original query terms, making it non-trivial to combine them with existing PRF approaches. Neural IR models differ in their architectures, making the development of a widely-applicable PRF approach a challenging task

Methods

Results

Conclusion