Model-based feedback in the language modeling approach to information retrieval

By Chengxiang Zhai, John D. Lafferty. 2001

879 citations recorded by the source corpus.

The language modeling approach to retrieval has been shown to perform well empirically. One advantage of this new approach is its statistical foundations. However, feedback, as one important component in a retrieval system, has only been dealt with heuristically in this new retrieval approach: the original query is usually literally expanded by adding additional terms to it. Such expansion-based feedback creates an inconsistent interpretation of the original and the expanded query. In this paper, we present a more principled approach to feedback in the language modeling approach. Specifically, we treat feedback as updating the query language model based on the extra evidence carried by the feedback documents. Such a model-based feedback strategy easily fits into an extension of the language modeling approach. We propose and evaluate two different approaches to updating a query language model based on feedback documents, one based on a generative probabilistic model of feedback documents and one based on minimization of the KL-divergence over feedback documents. Experiment results show that both approaches are effective and outperform the Rocchio feedback approach.

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

Estimating Embedding Vectors for Queries (2016, background): “PRF has been proven to be highly effective in improving retrieval performance [23, 36, 38].”
Retrieval-Enhanced Machine Learning (2022, background): “It has shown successful results in a wide range of retrieval tasks [2, 8, 14, 28, 36, 68, 73], demonstrating the quality of the produced query representations for retrieval.”
IRGAN: A Minimax Game for Unifying Generative and Discriminative Information Retrieval Models (2017, methodology): “In IR, MLE is also widely used as an estimation method for many relevance features or retrieval models [1], such as Term Frequency (TF), Mixture Model (MM) [49], and Probabilistic Latent Semantic Indexing (PLSI) [17].”
Similarity Measures for Short Segments of Text (2007, methodology): “Many techniques have been proposed to overcome the vocabulary mismatch problem, including stemming [5,9], LSI [3], translation mode ls [1], and query expansion [6,14].”
A Survey of Automatic Query Expansion in Information Retrieval (2012, methodology): “2005] Corpus Web table and form mining Attribute-value pairs Association rules Boolean query [Collins-Thompson and Callan 2005] Stemmer + Wordnet + corpus + top-ranked docs Probabilisticterm association network Single words Markov chain Structured query [Song et al. 2006] Top-ranked docs +corpus K”
Spoken Content Retrieval—Beyond Cascading Speech Recognition with Text Retrieval (2015, neutral): “formulated by component mixture models [270].”
Entity query feature expansion using knowledge base links (2014, background): “Query expansion techniques have been well studied for many models [23, 12, 31, 24, 37].”
Learning to rank for information retrieval (2009, neutral): “, the models based on K–L divergence [140]).”
Query Expansion with Freebase (2015, background): “For example, in Freebase the basketball player Michael Jordan is represented by an object that is linked to a brief text description, attributes such as his career statistics, categories such as people and athlete, and related entities such as Chicago Bulls.”
From Neural Re-Ranking to Neural Ranking: Learning a Sparse Representation for Inverted Indexing (2018, positive): “Following previous work on PRF [24, 29, 54, 59], we only keep the top t terms with the highest values in the updated query vector ® q∗.”
Hashtag retrieval in a microblogging environment (2010, methodology): “Given a query q generated by the query model Θ q , we rank hashtags in decreasing order of the negative KL divergence between their models and Θ q [3]: where unless otherwise noted the calculation uses ˆΘ MLq , the maximum likelihood estimator for Θ q .”
Pseudo-Relevance Feedback Based on Matrix Factorization (2016, methodology): “Indeed, similar to [43], the words of each document d are categorized into two types: background words and topical words.”
Answer Interaction in Non-factoid Question Answering Systems (2019, methodology): “Relevance feedback [4, 8, 12, 15, 39] is an important and early interactive method in IR systems.”
Automatic labeling of multinomial topic models (2007, background): “Labeling a cluster of documents is also valuable for many other tasks, such as search result summarization and modelbased feedback [27].”
Query dependent pseudo-relevance feedback based on wikipedia (2009, methodology): “Traditional PRF algorithms such as Okapi [27], Lavrenko and Croft’s relevance model [18], and Zhai and Laﬀerty’s mixture model [34] are based on the assumption that the top-ranked documents from an initial search are relevant to the query.”