Single Document Keyphrase Extraction Using Label Information

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1468–1476, Dublin, Ireland, August 23-29 2014. Single Document Keyphrase Extraction Using Label Information Sumit Negi IBM Research Delhi, India sumitneg@in.ibm.com Abstract Keyphrases have found wide ranging application in NLP and IR tasks such as document sum- marization, indexing, labeling, clustering and classification. In this paper we pose the problem of extracting label specific keyphrases from a document which has document level metadata as- sociated with it namely labels or tags (i.e. multi-labeled document). Unlike other, supervised or unsupervised, methods for keyphrase extraction our proposed methods utilizes both the doc- ument’s text and label information for the task of extracting label specific keyphrases. We pro- pose two models for this purpose both of which model the problem of extracting label specific keyphrases as a random walk on the document’s text graph. We evaluate and report the quality of the extracted keyphrases on a popular multi-label text corpus. 1 Introduction The use of graphs to model and solve various problems arising in Natural Language Processing have lately become very popular. Graph theoretical methods or graph based approaches have been success- fully applied for a varied set of NLP tasks such as Word Sense Disambiguation, Text Summarization, Topic detection etc. One of the earliest and most prominent work in this area has been the TextRank (Mi- halcea and Tarau, 2004) method - an unsupervised graph-based ranking model for extracting keyphrases and “key” sentences from natural language text. This unsupervised method extracts prominent terms, phrases and sentences from text. The TextRank models the text as a graph where, depending on the end application, text units of various sizes and characteristics can be added as vertices e.g. open class words, collocations, sentences etc. Similarly, based on the application, connections can be drawn between these vertices e.g. lexical or semantic relation, contextual overlap etc. To identify “central” or “key” text units in this text graph, TextRank runs the PageRank algorithm on this constructed graph. The ranking over vertices (text units), which indicates their centrality and importance, is obtained by finding the stationary distribution of the random walk on the text graph. In this paper, we consider the problem of extracting label specific keyphrases from a document which has document level metadata associated with it namely labels (i.e. multi-labeled document). To elabo- rate, consider a document as shown in Figure 1. This document has been assigned to two categories as indicated by the labels “Air Pollution” and “Plant Physiology”. Running TextRank on this article yields top ranked key-phrases such as “calibrated instrument”, “polluting gases”, “industrial development” etc. These keyphrases, though central to the article, are not specific to any of the labels that have been assigned to the article. For instance, one would associate keyphrases such as “carbon monoxide ”, “air pollutants” to be more relevant to the “Air Pollution” label and keyphrases such as “stomatal movement”, “cell defense” to be more closely associated with the “Plant Physiology” label. The objective of this …

Content truncated to 500 words