Applying clustering algorithms to determine authorship of chinese twitter messages

摘要

Author attribution research of character-based languages such as Chinese is still in its early stages. In this paper, we study the effectiveness of two popular clustering algorithms in determining the authorship of Chinese Twitter messages. We create a data-set of ten authors with 100 tweets each from publicly-available Chinese Twitter profiles. We analyze the data using simple k-means (SKM) and Expectation Maximization (EM), two popular clustering algorithms available in the Waikato Environment for Knowledge Analysis (WEKA). Our feature set includes character n-grams and Chinese function words derived from the literature. We achieve accuracy up to 44:53% for three authors, 29:24% for five authors, and 20:52% for ten authors. For our data-sets and the number of authors we compared, SKM returns better accuracy ratings. Lastly, we determine that function words are valuable features in attributing Chinese Tweets, and identify which of these Chinese function words were of most value.