Supervised TextRank

Abstract

In this paper we investigate how to adapt the TextRank method to make it work in a supervised way. TextRank is a graph based method that applies the ideas of the ranking algorithm used in Google (PageRank) to Natural Language Processing (NLP) tasks. This approach has given very good results in many NLP tasks like text summarization, keyword extraction or word sense disambiguation. In all these tasks TextRank operates in an unsupervised way, without using any training corpus. Our main contribution is the definition of a method that allows to apply TextRank to a graph that includes information generated from a training tagged corpus. We have tested our method with the Part of Speech (POS) tagging task, comparing the results with those obtained with tools specialized in this task. The performance of our system is quite near to these tools, improving the results of two of them when the corpus tagset is big and therefore the tagging task more complicated. © Springer-Verlag Berlin Heidelberg 2006.