Novel Word Features for Keyword Extraction

Abstract

Keyword extraction plays an increasingly crucial role in several texts related researches. Applications that utilize feature word selection include text mining, and information retrieval etc. This paper introduces novel word features for keyword extraction. These new word features are derived according to the background knowledge supplied by patent data. Given a document, to acquire its background knowledge, this paper first generates a query for searching the patent data based on the key facts present in the document. The query is used to find files in patent data that are closely related to the contents of the document. With the patent search result file set, the information of patent inventors, assignees, and citations in each file are used to mining the hidden knowledge and relationship between different patent files. Then the related knowledge is imported to extend the background knowledge base, which would be extracted to derive the novel word features. The newly introduced word features that reflect the document’s background knowledge offer valuable indications on individual words’ importance in the input document and serve as nice complements to the traditional word features derivable from explicit information of a document. The keyword extraction problem can then be regarded as a classification problem and the Support Vector Machine (SVM) is used to extract the keywords. Experiments have been done using two different data sets. The results show our method improves the performance of keyword extraction.