Using Grammatical Markov Models for Stylometric Analysis CS224N : Natural Language Processing

Researchers have tackled the problem of authorship attribution in several different ways, using various metrics to identify the author of an anonymous document given a set of writing samples from potential candidates. Common complaints about modern methodologies tend to accuse studies of content bias, which occurs when quantitative models identify similar content rather than similar styles. This artificially increases accuracy by producing good results on test data while failing to identify authors in real-world applications. We examine several quantitative methods that isolate style by using grammar-based features rather than relying on models of word shape and frequency.