Using Supervised Machine Learning to Automatically Build Relevance Judgments for a Test Collection

Abstract

This paper describes a new approach to building the query based relevance sets (qrels) or relevance judgments for a test collection automatically without using any human intervention. The methods we describe use supervised machine learning algorithms, namely the Naïve Bayes classifier and the Support Vector Machine (SVM). We achieve better Kendall's tau and Spearman correlation results between the TREC system ranking using the newly generated qrels and the ranking obtained from using the human-built qrels than previous baselines. We also apply a variation of these approaches by using the doc2vec representation of the documents rather than using the traditional tf-idf representation.