Application of human associative memory to Information Retrieval
2009 - Present
As a result of information overload, information retrieval (IR) systems play an ever-increasing role in enabling access to relevant information. However, simply retrieving articles containing the search terms are not enough given that users on average look at only the first page of results making ranking a priority. Numerous approaches exist for ranking. For example, ranking by publication date is satisfactory if the information need is finding recently published articles, but suboptimal if the information need is to find older landmark articles. Ranking by relevance to the query is pertinent for information needs such as gathering (find material relevant to new problem) or exploring (browsing material). Early studies found that different approaches resulted in comparable performance; however, the overlap in the documents considered relevant was low. The lack of overlap has resulted in significant research effort in combining relevance judgments from multiple of algorithms, which has led to performance increases; however, these approaches were combined in an ad-hoc fashion and lacked an integrated theoretical framework for integrating multiple sources of evidence.
Anderson's rational analysis of human memory, and the corresponding model of human associative memory found in Anderson’s cognitive model, ACT-R, provide a theoretical framework for combining multiple sources of evidence to rank the relevance of memory items. The ACT-R theory is a Bayesian model, which asserts that the memory system retrieves the item with the highest posterior probability of being needed given an agent’s current contextual cues. Computing posterior probabilities over memory items requires the prior probability of each item and the likelihood of the context given that the memory item is needed. According to the ACT-R theory, the prior probability of a memory item is a function of its past use. The likelihood of the context is calculated in ACT-R by representing memory as a semantic network where the strength of association between memory elements reflect the probability of needing a memory given that other memory items are present as cues. The strength of associations of all cues in the context are summed to approximate the likelihood.
I hypothesize that this theory-driven framework, appropriately adapted to IR, will outperform existing methods because the integration of diverse information as evidence will improve relevance judgments. The proposed framework will adapt the ACT-R theory to develop a concept-based IR model whose results are ranked by calculating the posterior probability for each document given the query.
Relevant Publications
Poster Abstracts
J.C. Goodwin, J. Herskovic, and T.R. Johnson. (2010, Abstract). Application of Associative Memory Theory to Biomedical Document Retrieval. AMIA 2010 Annual Symposium.