Selecting Article Segment Titles Based on Keyphrase Features and Semantic Relatedness

Abstract

Nowadays people can find almost all kinds of information they want from the Internet. However, in most cases, users are not willing to find their target among segment among long paragraphs, by spending much time browsing texts. Existing work on topic labeling works effectively and performs well on document categorization, but inadequate for granularity of detailed contents. Thus we propose a method for selecting titles for segments in long documents. We analyze the characteristics of high quality titles for article segments, from the aspect of semantic relatedness between the target segment and related articles as well as other segments. Then we revise three features proposed before. We improve the phraseness feature, for giving appropriate scores for long titles. Meanwhile, we combine the features SimPF and Embedding-vector to enhance the efficiency and rationality. We use Wikipedia articles for experimental evaluations, in which a large number of article segments are titled manually, and a great number of articles lack detailed segment titles. We evaluate scoring functions by where hidden original segment titles are ranked, through precision@K. Through rigorous evaluations, we show an optimum combination of the features.