Network Data Analysis of Word Graphs with Applications to Authorship Attributio

Network data analysis is an emerging area of study that applies quantitative analysis to complex data from a variety of application fields. Methods used in network data analysis enable visualization of relational data in the form of graphs and also yield descriptive characteristics and predictive graph models. This thesis shows that a representation of text as a word graph produces the well documented feature sets used in authorship attribution tasks such as the word frequency model and the part-of-speech (POS) bigram model. This thesis applies nominal assortativity of parts of speech, a network data characteristic of word graphs, to the problem of authorship attribution and shows how these features are produced from a word graph model. Specifically, it is shown that the nominal assortative mixture of parts of speech, a statistic that measures the tendency of words of the same POS in a word network to be connected by an edge, produces a feature set that can be used to predict authorship. These results are compared to the POS bigram model, a highly accurate authorship attribution model, and show that the nominal assortativity model is competitive. Analysis of these models along with word graph characteristics provides insights into the English language. Particularly, analysis of the nominal assortative mixture of parts of speech reveals regular structural properties of English grammar.