Automatic Discrimination of Genres

The exponential growth in the size of informationresources has motivated research on selectinginformation using non-topical criteria, such asgenre. This work assesses the feasibility ofautomatically classifying documents by genre usingonly adjectives and adverbs as discriminatingfeatures, in particular: (a) trait adjectives, usedby psychologists to assess human personality traits,and (b) speaker-oriented adverbs, studied bylinguists as markers of narrator attitude. It employshuman intelligence to select features a priori basedon insights from linguistics and psychology. Findingsshow that one can build more parsimonious classifierswhose power is equal to or greater than that ofclassifiers which use ad hoc methods.This work (a) presents a model for the systematicstudy of a narrowly-defined set of features, (b)demonstrates the principle of selecting featuresmotivated from other disciplines, and (c) uses a newperformance measure called accuracy gain. This workwill benefit researchersand practitioners ininformation retrieval, computational linguistics,communication, and anyone interested in identifyingstyle and other non-topical characteristics of documents.