Combining linguistic and machine learning techniques for email summarization

Combining Linguistic and Machine Learning Techniques for Email Summarization Smaranda Muresan Dept. of Computer Science Columbia University 500 West 120 Street New York, NY, 10027 smara@cs.columbia.edu Evelyne Tzoukermann Bell Laboratories Lucent Technologies 700 Mountain Avenue Murray Hill, NJ, 07974 evelyne@lucent.com Judith L. Klavans Columbia University Center for Research on Information Access 535 West 114th Street New York, NY 10027 klavans@cs.columbia.edu Abstract This paper shows that linguistic tech- niques along with machine learning can extract high quality noun phrases for the purpose of providing the gist or summary of email messages. We describe a set of comparative experi- ments using several machine learning algorithms for the task of salient noun phrase extraction. Three main conclu- sions can be drawn from this study: (i) the modifiers of a noun phrase can be semantically as important as the head, for the task of gisting, (ii) linguistic fil- tering improves the performance of ma- chine learning algorithms, (iii) a combi- nation of classifiers improves accuracy. 1 Introduction In this paper we present a comparative study of symbolic machine learning models applied to nat- ural language task of summarizing email mes- sages through topic phrase extraction. Email messages are domain-general text, they are unstructured and not always syntactically well formed. These characteristics raise challenges for automatic text processing, especially for the sum- marization task. Our approach to email summa- rization, implemented in the GIST-IT system, is to identify topic phrases, by first extracting noun phrases as candidate units for representing doc- ument meaning and then using machine learning algorithms to select the most salient ones. The comparative evaluation of several machine learning models in the settings of our experiments indicates that : (i) for the task of gisting the mod- ifiers of the noun phrase are equally as important as the head, (ii) noun phrases are better than n- grams for the phrase-level representation of the document, (iii) linguistic filtering enhances ma- chine learning techniques, (iv) a combination of classifiers improves accuracy. Section 2 of the paper outlines the machine learning aspect of extracting salient noun phrases, emphasizing the features used for classifica- tion and the symbolic machine learning models used in the comparative experiments. Section 3 presents the linguistic filtering steps that im- prove the accuracy of the machine learning algo- rithms. Section 4 discusses in detail our conclu- sions stated above. 2 Machine Learning for Content Extraction Symbolic machine learning has been applied suc- cessfully in conjunction with many NLP applica- tions (syntactic and semantic parsing, POS tag- ging, text categorization, word sense disambigua- tion) as reviewed by Mooney and Cardie (1999). We used machine learning techniques for finding salient noun phrases that can represent the sum- mary of an email message. This section describes the three steps involved in this classification task: 1) what representation is appropriate for the infor- mation to be classified as relevant or non-relevant (candidate phrases), 2) which features should be associated with each candidate, 3) which classifi- cation models should be used. Case 1 CNP: scientific/JJ and/CC …

Content truncated to 500 words