Electronic Thesis and Dissertation Repository


Doctor of Philosophy


Computer Science


Dr. Robert E. Mercer


In this age of information overload, one experiences a rapidly growing over-abundance of written text. To assist with handling this bounty, this plethora of texts is now widely used to develop and optimize statistical natural language processing (NLP) systems. Surprisingly, the use of more fragments of text to train these statistical NLP systems may not necessarily lead to improved performance. We hypothesize that those fragments that help the most with training are those that contain the desired information. Therefore, determining informativeness in text has become a central issue in our view of NLP. Recent developments in this field have spawned a number of solutions to identify informativeness in text. Nevertheless, a shortfall of most of these solutions is their dependency on the genre and domain of the text. In addition, most of them are not efficient regardless of the natural language processing problem areas. Therefore, we attempt to provide a more general solution to this NLP problem.

This thesis takes a different approach to this problem by considering the underlying theme of a linguistic theory known as the Code Quantity Principle. This theory suggests that humans codify information in text so that readers can retrieve this information more efficiently. During the codification process, humans usually change elements of their writing ranging from characters to sentences. Examples of such elements are the use of simple words, complex words, function words, content words, syllables, and so on. This theory suggests that these elements have reasonable discriminating strength and can play a key role in distinguishing informativeness in natural language text. In another vein, Stylometry is a modern method to analyze literary style and deals largely with the aforementioned elements of writing. With this as background, we model text using a set of stylometric attributes to characterize variations in writing style present in it. We explore their effectiveness to determine informativeness in text. To the best of our knowledge, this is the first use of stylometric attributes to determine informativeness in statistical NLP. In doing so, we use texts of different genres, viz., scientific papers, technical reports, emails and newspaper articles, that are selected from assorted domains like agriculture, physics, and biomedical science. The variety of NLP systems that have benefitted from incorporating these stylometric attributes somewhere in their computational realm dealing with this set of multifarious texts suggests that these attributes can be regarded as an effective solution to identify informativeness in text. In addition to the variety of text genres and domains, the potential of stylometric attributes is also explored in some NLP application areas---including biomedical relation mining, automatic keyphrase indexing, spam classification, and text summarization---where performance improvement is both important and challenging. The success of the attributes in all these areas further highlights their usefulness.