Generative Content Models for Structural Analysis of Medical Abstracts

The ability to accurately model the content structure of text is important for many natural language processing applications. This paper describes experiments with generative models for analyzing the discourse structure of medical abstracts, which generally follow the pattern of "introduction", "methods", "results", and "conclusions". We demonstrate that Hidden Markov Models are capable of accurately capturing the structure of such texts, and can achieve classification accuracy comparable to that of discriminative techniques. In addition, generative approaches provide advantages that may make them preferable to discriminative techniques such as Support Vector Machines under certain conditions. Our work makes two contributions: at the application level, we report good performance on an interesting task in an important domain; more generally, our results contribute to an ongoing discussion regarding the tradeoffs between generative and discriminative techniques.