Linguistic correlates of style: authorship classification with deep linguistic analysis features

Linguistic correlates of style: authorship classification with deep linguistic analysis features Michael Gamon Microsoft Research Microsoft Corp. One Microsoft Way Redmond, WA 98052 mgamon@microsoft.com Abstract The identification of authorship falls into the category of style classification, an interesting sub-field of text categorization that deals with properties of the form of linguistic expression as opposed to the content of a text. Various fea- ture sets and classification methods have been proposed in the literature, geared towards ab- stracting away from the content of a text, and focusing on its stylistic properties. We demon- strate that in a realistically difficult authorship attribution scenario, deep linguistic analysis features such as context free production fre- quencies and semantic relationship frequencies achieve significant error reduction over more commonly used “shallow” features such as function word frequencies and part of speech trigrams. Modern machine learning techniques like support vector machines allow us to ex- plore large feature vectors, combining these dif- ferent feature sets to achieve high classification accuracy in style-based tasks. 1 Introduction Authorship identification has been a long stand- ing topic in the field of stylometry, the analysis of literary style (Holmes 1998). From a broader per- spective, issues of style, genre, and authorship are an interesting sub-area of text categorization. Typically, text categorization concerns itself with classifying texts according to topics. For that ob- jective, it is crucial to extract information about the content of a text. In contrast, issues of style, genre and authorship are about the “form” of a text. The analysis of style needs to abstract away from the content and focus on content-independent form properties of the linguistic expressions in a text. This makes style analysis a prime candidate for the use of linguistic processing to extract structural features. Viewed from a different angle, the ab- stractness of style assessment features makes them highly domain-independent and reusable, as long as they are used with a classification technique that tolerates large feature vectors. Previously suggested methods of style categori- zation and authorship identification have made use of a number of content independent features: • frequencies of function words (Mosteller et al. 1964) • word length and sentence length statistics (dating back to 1851 according to Holmes 1998) • word tags and tag n-grams (Argamon et al. 1998, Koppel et al. 2003, Santini 2004) • “stability features” (Koppel et al. 2003) cap- turing the extent to which an item can be re- placed by a semantically similar item • rewrite rules in an automatic parse (Baayen et al. 1996) In this paper, we demonstrate that a combination of features based on shallow linguistic analysis (function word frequencies, part of speech tri- grams) and a set of deep linguistic analysis features (context free grammar production frequencies and features derived from semantic graphs) yields very high accuracy in attributing a short random text sample to one of the three Brontë sisters as its au- thor. Through feature ablation experiments we show that both the syntactic information captured in syntactic rewrite rules and the …

Content truncated to 500 words