Learning to Select Good TitleWords: An New Approach based on Reverse Information Retrieval

In this paper, we show how we can learn to select good words for a document title. We view the problem of selecting good title words for a document as a variant of an Information Retrieval problem. Each title word is treated as a “document” and selection of appropriate title words as finding relevant “documents”. Based on our training collection consisting of 40,000 document and title pairs, we learn the “document” representations for all the title words and apply these learned representations to select appropriate title words over 10,000 test documents. Compared to other learning approaches, namely K nearest neighbor approach, a Naïve Bayesian approach and a variant of a machine translation model, we find that our approach is significantly better as indicated by the F1 metric.