Feature Weighting in Finding Feedback Documents for Query Expansion in Biomedical Document Retrieval

Abstract

Finding good feedback documents for query expansion is a well-known problem in the field of information retrieval. This paper describes a novel approach for finding relevant documents for feedback in query expansion for biomedical document retrieval. The proposed approach relies on a small amount of human intervention to find good feedback documents and tries to learn the relation between query and documents in terms of usefullness of document for query expansion. This proposed approach uses an NLP-based feature weighting technique with classification and clustering method on the documents and identifies relevant documents for feedback. The documents are represented using term frequency and inverse document frequency (TF–IDF) features and these features are weighted according to the type of query and type of the terms. The experiments performed on CDS 2014, 2015 and 2016 datasets show that the feature weighting in finding feedback documents for query expansion approach gives good results as compared to the results of pseudo-relevance feedback, relevance feedback and the results of TF–IDF features without weighting.

Feedback Document Discovery: Finding Good Feedback Documents

First Algorithm

The first algorithm is based on classification. If we have human judgements available for some of the feedback documents, then it will serve as a training data for classification. The documents are represented as a collection of bag-of-words, the TF–IDF scores of the words represent features and human relevance scores provides the classes. Using this as a training data, we want to predict the relevance of other top retrieved feedback documents represented by TF–IDF scores of words.

Second Algorithm

The second algorithm is an extension of first algorithm. The analysis of results of first algorithm shows that the feedback document set still contains some non-relevant docs and it is responsible for insignificant improvement. The detailed analysis is given in “Experiments and Results”. This approach further removes non-relevant documents from relevant document class identified by classification approach. The idea is to perform clustering on the relevant identified documents with number of clusters as two: one from actually relevant documents and second from non-relevant documents. This approach relies on the statement that the relevant documents tend to cluster within the space. k-means clustering is used with k = 2. Since, the convergence of k-means clustering depends on the initial choice of cluster centroids, the initial cluster centroids are chosen as the average of relevant documents’ vectors and the average of non-relevant documents’ vectors from training data.