Authorship Identification Using Correlations of Frequent Features Notebook for PAN at CLEF 2013

In this work, we explore how well can we perform the PAN‘13 authorship identification task by using correlation of frequent significant features found in documents written by a known author. We extract features from the context of four types of words: first words of sentences or lines, nouns, verbs, punctuation. We compute the Matthews Correlation Coefficient for all pairs of extracted features and by using principal component analysis, transform them into a form suitable for using simple Euclidian distance as a metric. By hypothesizing that the distances of different documents written by a same author belong to same distribution, we can provide educated guesses for the authorship identification. Our system achieves F1-score of 66.7% for English, 56.7% for Greek and 80% for Spanish texts with an overall F1-score of 67.1% on the PAN‘13 dataset.