Ngram and Bayesian Classification of Documents for Topic and Authorship

Abstract

Large, real world, data sets have been investigated in the context of Authorship Attribution of real world documents. Ngram measures can be used to accurately assign authorship for long documents such as novels. A number of 5 (authors) ϫ 5 (movies) arrays of movie reviews were acquired from the Internet Movie Database. Both ngram and naive Bayes classifiers were used to classify along both the authorship and topic (movie) axes. Both approaches yielded similar results, and authorship was as accurately detected, or more accurately detected, than topic. Part of speech tagging and function-word lists were used to investigate the influence of structure on classification tasks on documents with meaning removed but grammatical structure intact.

A survey of modern authorship attribution methods

“…Chaski (2001) described a writing sample database comprising texts of 92 people on 10 common subjects (e.g., a letter of apology to your best friend, a letter to your insurance company, etc.). Clement and Sharp (2003) reported a corpus of movie reviews comprising 5 authors who review the same 5 movies. Another corpus comprising various genres was described by Baayen, van Halteren, Neijt, and Tweedie (2002) and was also used by Baayen (2005) andvan Halteren (2007).…”

“…Several features described in Section 2 are claimed to capture only stylistic information (e.g., function words). However, the application of stylometric features to topic-identification tasks has revealed the potential of these features to indicate content information as well (Clement & Sharp, 2003;Mikros & Argiri, 2007). It seems that low-level features like character n-grams are very successful for representing texts for stylistic purposes Keselj, et al, 2003;Stamatatos, 2006b;Grieve, 2007).…”

“…One of the earliest approaches to author identification that is still used in many modern studies employ the use of probabilistic models (Mosteller & Wallace, 1964;Clement & Sharp, 2003;Peng, et al, 2004;Zhao & Zobel, 2005;Madigan, et al, 2005;Sanderson & Guenter, 2006). Such methods attempt to maximize the probability P(x|a) for a text x to belong to a candidate author a.…”

Seeking Variety: A Dynamic Model of Employee Blog Reading Behavior

“…Comprehensiveness is measured as the length of post (Lu et al 2010). Sentiments expressed in the posts are classified as positive or negative using a Bayesian classifier (Clement and Sharp 2003) that is trained over word n-grams of order 5. The order was selected using cross-validation over a sentiment classification dataset (Pang and Lee 2004).…”

Is 1 noun worth 2 adjectives? Measuring relative feature utility

“…What is the relative degree to which these linguistic and non-linguistic characteristics represent topicality and predict the utility of a document to a system user? Term roots may carry one or more meanings or topics, and the addition of contextual or supporting information, such as suffixes, part-of-speech tags, and larger contexts can contribute to the topicality, therefore improving document ordering (Bossong, 1989;Clement & Sharp, 2003;Losee, 2001) How does one measure how many of one feature or type of feature is equivalent in ordering power to another chosen feature or type of feature? The Relative Feature Utility may be used to empirically analyze the ordering effects of term stemming, the length of natural language phrases, the effect of using different part-of-speech labels, and various information retrieval or filtering assumptions.…”