Ranking Keyphrases from Semantic and Syntactic Features of Textual Terms

Two important lines of research in key phrase extraction from text are methods that use machine learning to discover rules based on statistics of terms, and knowledge-intensive methods that seek to understand the semantics of the text with the help of conceptual bases like Wikipedia. Our argument is that the task of key phrase extraction for different domains requires defining ranking functions that take into account the advantages and shortcomings of each approach to the specific problem. To determine the best ranking function, arrangements of weights are generated, which in turn weight each of the attributes used in both the statistic and semantic functions. The arrangement of weights that presents better average performance sets the weights of the attributes of the ranking function. We show comparative tests conducted with current approaches that use only syntactic or semantic features with a hybrid ranking approach. The later outperformed the state of the art.