Many unsupervised methods for keyphrase extraction typically compute a score for each word in a document based on various measures such as tf-idf or the PageRank score computed from the word graph built from the text document. The final score of a candidate phrase is then calculated by summing up the scores of its constituent words. A potential problem with the sum up scoring scheme is that the length of a phrase highly impacts its score. To reduce this impact and extract keyphrases of varied lengths, we propose a new scheme for scoring phrases which calculates the final score using the average of the scores of individual words weighted by the frequency of the phrase in the document. We show experimentally that the unsupervised approaches that use this new scheme outperform their counterparts that use the sum up scheme to score phrases.
展开▼
