To digest tremendous documents efficiently, people often resort to their titles, which normally provide a concise and semantic representation of main text. Some titles however are misleading due to lexical ambiguity or eye-catching intention. The requirement of reference summaries hampers using traditional lexical summarisation evaluation techniques for title evaluation. In this paper we develop semantic title evaluation techniques by comparing a title with other sentences in terms of topic-based similarity with regard to the whole document. We further give a statistical hypothesis test to check whether a title is favourable without any reference summary. As a byproduct, the top similar sentence can be recommended as a candidate for title. Experiments on patents, scientific papers and DUC'04 benchmarks show our Semantic Title Evaluation and Recommendation technique based on a recent Segmented Topic Model (STERSTM), performs substantially better than that based on the canonical model Latent Dirichlet Allocation (STERLDA). It can also recommend titles with quality comparable with the winners of DUC'04 in terms of summarising documents into very short summaries.
展开▼
