Retrieval Models - Zhai

Chapter 6 of "Text Data Management and Analysis A Practical Introduction to Information Retrieval and Text Mining", by Zhai. N.B. I'm not 100% this chapter is from the book! I only received a scan without proper reference.

Two main information retrieval models are discussed:

  1. Vector space models
  2. Query likelihood

Overview

  1. Vector space models are a good example of similarity-based approaches: a document is considered to be more relevant when the document has a higher similarity to the query than another document.

  2. There are also many probabilistic retrieval models:

We assume that queries and documents are all observations from random variables, and we assume there is a binary random variable called R (with a value of either 1 or 0) to indicate whether a document is relevant to a query. We then define the score of a document with respect to a query as the probability that this random variable R is equal to 1 given a particular document and query. (87)

Examples:

  • Classic probabilistic model(Maron and Kuhns 1960)
  • Language modeling (Ponte and Croft 1998)
  • Divergence-from-randomness model (Amati and Van Rijsbergen 2002)

The example of this chapter, the query likelihood retrieval model, is an example of a language modeling approach.

(The classic probabilistic model approach is very dominant, but is closer to vector space models and is therefore not discussed here).

  1. There are also models using probabilistic inference. These try to quantify the probability that "the query follows from the document" (88)

  2. Models using axiomatic thinking. Some models define a set of constraints and then try to find a good ranking function that meets those constraints.

Despite having a different orientation, the resulting ranking functions are very similar to those of other approaches. This approach can be very useful as a diagnostic tool for ranking functions though.

Few models are experimentally proven to be robust and effective. This book covers four models that are considered robust:

  • BM25 (commonly used in search engines)
  • Pivoted length normalization
  • Query likelihood with JM smoothing
  • Query likelihood with Dirichlet prior smoothing

All state-of-the-art retrieval models assume a "bag-of-words representation of text" where the score of a query is computed per word.

Vector Space Retrieval Models

TODO