Abstract
Keyphrases provide semantic metadata producing an overview of the content of a document, they are used in many text-mining applications. This paper proposes a new method that improves automatic keyphrase extraction by using semantic information of candidate keyphrases. Our method is realized in two stages. In selecting candidates stage, after extraction of all phrases from document, a word sense disambiguation method is used to get senses of phrases, then term conflation is performed by using case folding, stemming, and semantic relatedness between candidates. In filtering stage, four features are used to compute for each candidate: the TFxIDF measure describing the specificity of a phrase, first occurrence of a phrase in the document, length of a phrase, and coherence score which measure the semantic relatedness between the phrase and other candidates. A Naive Bayes scheme builds a prediction model training data with known keyphrases, and then uses the model to calculate the overall probability for each candidate. We evaluate semantically improved method against the well known Kea system by using a more effective semantically enhanced evaluation method. The inter-domain experiment shows that quality of keyphrases extraction can be improved significantly when semantic information is exploited. The intra-domain experiment shows our method is competitive with Kea++ algorithm, and not domain-specific.