Keyphrase extraction based on semantic relatedness

Keyphrase extraction is a fundamental research task in natural language processing and text mining. A limitation of previous keyphrase extraction methods based on semantic analysis is that the acquisition of the semantic features within phrases is restricted by the constructed thesaurus and language. An approach to the acquisition of the semantic features within phrases from a single document is proposed in this paper, which is used to extract document keyphrases. Semantic relatedness degrees between phrases are computed using word co-occurrence information in the document, and the document is represented as a relatedness graph. Keyphrases are extracted based on the semantic relatedness features acquired from the graph. Our experiments demonstrate that the proposed keyphrase extraction method always outperforms the baseline methods TFIDF and Kea. Furthermore, our approach is not domain-specific and the method generalizes well when it is trained on one domain (journal articles) and tested on another (news web pages).