Regularized estimation of mixture models for robust pseudo-relevance feedback

By Tao Tao, Chengxiang Zhai. 2006

237 citations recorded by the source corpus.

Pseudo-relevance feedback has proven to be an effective strategy for improving retrieval accuracy in all retrieval models. However the performance of existing pseudo feedback methods is often affected significantly by some parameters, such as the number of feedback documents to use and the relative weight of original query terms; these parameters generally have to be set by trial-and-error without any guidance. In this paper, we present a more robust method for pseudo feedback based on statistical language models. Our main idea is to integrate the original query with feedback documents in a single probabilistic mixture model and regularize the estimation of the language model parameters in the model so that the information in the feedback documents can be gradually added to the original query. Unlike most existing feedback methods, our new method has no parameter to tune. Experiment results on two representative data sets show that the new method is significantly more robust than a state-of-the-art baseline language modeling approach for feedback with comparable or better retrieval accuracy.

Subfields: Artificial Intelligence, Information Systems

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Information Systems

Citations in the corpus

Relevance-based Word Embedding (2017, methodology): “ distributions are estimated based on the topk retrieved documents (in our experiments, we setk to 10) for each query. „e distributions can be estimated using any PRF model, such as those proposed in [20, 36, 42, 44]. In this paper, we only focus on the relevance model [20], a state-of-the-art PRF m”
IRGAN: A Minimax Game for Unifying Generative and Discriminative Information Retrieval Models (2017, methodology): “It is also worth comparing IRGAN with pseudo relevance feedback [39, 45, 50], where the top retrieved documents are selected to rene the ranking result.”
LambdaMerge: merging the results of query reformulations (2011, methodology): “Traditional query reformulation methods in IR are based on pseudo-relevance (blind) feedback, which expands a given query by first retrieving its top-k documents, and then extracting document terms to add to the query [20, 26].”
Pseudo-Relevance Feedback Based on Matrix Factorization (2016, methodology): “The regularization strategy for PRF models was previously used by Tao and Zhai [35] for the mixture model.”
Learning to rank for information retrieval (2009, unknown): “Some of them [118] take the proximity of the query terms into consideration, and some others consider the relationship between documents in terms of content similarity [117], hyperlink structure [113], website structure [101], and topic diversity [139].”
Topic sentiment mixture: modeling facets and opinions in weblogs (2007, methodology): “In the experiments, we set the initial values of μs reasonably large (>10,000), and use the regularized estimation strategy in [25] to gradually decay the μs.”
Query Expansion with Freebase (2015, methodology): “The predicted probabilities are used as terms’ expansion scores, and those terms with highest scores are selected for query expansion.”
Query dependent pseudo-relevance feedback based on wikipedia (2009, background): “Large amounts of research has been reported on attempts to improve traditional PRF, e.g. using latent concepts [21], query-regularized estimation [29], additional features other than term distributions [3], and a clustered-based re-sampling method for generating pseudo-relevant documents [19].”
Book Review: Statistical Language Models for Information Retrieval by ChengXiang Zhai (2008, methodology): “Translation models, Relevance models, and Feedback-based query models have all been shown to improve performance significantly over the simple LMs (Parameter tuning is necessary in many cases, but see [Tao & Zhai 06] for “parameter-free” pseudo feedback)”
Spoken Content Retrieval—Beyond Cascading Speech Recognition with Text Retrieval (2015, methodology): “model [167] was used to filter out such acoustic patterns for function words [293].”