Mining Online Diaries for Blogger Identification

摘要

In this paper, we present an investigation of authorship identification on personal blogs or diaries, which are different from other types of text such as essays, emails, or articles based on the text properties. The investigation utilizes couple of intuitive feature sets and studies various parameters that affect the identification performance. Many studies manipulated the problem of authorship identification in manually collected corpora, but only few utilized real data from existing blogs. The complexity of the language model in personal blogs is motivating to identify the correspondent author. The main contribution of this work is at least three folds. Firstly, we utilize the LIWC and MRC feature sets together, which have been developed with Psychology background, for the first time for authorship identification on personal blogs. Secondly, we analyze the effect of various parameters, and feature sets, on the identification performance. This includes the number of authors in the data corpus, the post size or the word count, and the number of posts for each author. Finally, we study applying authorship identification over a limited set of users that have a common personality attributes. This analysis is motivated by the lack of standard or solid recommendations in literature for such task, especially in the domain of personal blogs. The results and evaluation show that the utilized features are compact while their performance is highly comparable with other larger feature sets. The analysis also confirmed the most effective parameters, their ranges in the data corpus, and the usefulness of the common users classifier in improving the performance, for the author identification task.