Improved Automatic Keyword Extraction Given More Linguistic Knowledge

Improved Automatic Keyword Extraction Given More Linguistic Knowledge Anette Hulth Department of Computer and Systems Sciences Stockholm University Sweden hulth@dsv.su.se Abstract In this paper, experiments on automatic extraction of keywords from abstracts us- ing a supervised machine learning algo- rithm are discussed. The main point of this paper is that by adding linguistic know- ledge to the representation (such as syn- tactic features), rather than relying only on statistics (such as term frequency and n- grams), a better result is obtained as mea- sured by keywords previously assigned by professional indexers. In more detail, ex- tracting NP-chunks gives a better preci- sion than n-grams, and by adding the POS tag(s) assigned to the term as a feature, a dramatic improvement of the results is ob- tained, independent of the term selection approach applied. 1 Introduction Automatic keyword assignment is a research topic that has received less attention than it deserves, con- sidering keywords’ potential usefulness. Keywords may, for example, serve as a dense summary for a document, lead to improved information retrieval, or be the entrance to a document collection. However, relatively few documents have keywords assigned, and therefore finding methods to automate the as- signment is desirable. A related research area is that of terminology ex- traction (see e.g., Bourigault et al. (2001)), where all terms describing a domain are to be extracted. The aim of keyword assignment is to find a small set of terms that describes a specific document, in- dependently of the domain it belongs to. However, the latter may very well benefit from the results of the former, as appropriate keywords often are of a terminological character. In this work, the automatic keyword extraction is treated as a supervised machine learning task, an ap- proach first proposed by Turney (2000). Two im- portant issues are how to define the potential terms, and what features of these terms are considered dis- criminative, i.e., how to represent the data, and con- sequently what is given as input to the learning al- gorithm. In this paper, experiments with three term selection approaches are presented: n-grams; noun phrase (NP) chunks; and terms matching any of a set of part-of-speech (POS) tag sequences. Four dif- ferent features are used: term frequency, collection frequency, relative position of the first occurrence, and the POS tag(s) assigned to the term. 2 Points of Departure Treating the automatic keyword extraction as a su- pervised machine learning task means that a clas- sifier is trained by using documents with known keywords. The trained model is subsequently ap- plied to documents for which no keywords are as- signed: each defined term from these documents is classified either as a keyword or a non-keyword; or—if a probabilistic model is used—the probabil- ity of the defined term being a keyword is given. Turney (2000) presents results for a comparison be- tween an extraction model based on a genetic algo- rithm and an implementation of bagged C4.5 deci- sion trees for the task. The terms are all stemmed uni- …

Content truncated to 500 words