Stylometric Studies based on Tone and Word Length Motifs

31st Pacific Asia Conference on Language, Information and Computation (PACLIC 31), pages 56–63 Cebu City, Philippines, November 16-18, 2017 Copyright c⃝2017 Renkui Hou and Chu-Ren Huang Stylometric Studies based on Tone and Word Length Motifs Hou Renkui Huang Chu-Ren The Hong Kong Polytechnic University, Hong Kong The Hong Kong Polytechnic University, Hong Kong; hourk0917@163.com churen.huang@polyu.edu.hk Abstract: We propose a new approach to stylometric analysis combining lexical and textual information, but without annotation or other pre-processing. In particular, our study makes use Chinese tones motifs and word length motifs automatically extracted from unannotated texts. The proposed approach is based on linked data in nature as tone and word-length information is extracted from a lexicon and mapped to the text. Support vector machine and random forest were used to establish the classification models for author differentiation. Based on comparative study of classification results of different models, we conclude that the combination of word-final tones motifs, segment-final motifs and word length motifs provides the best outcome and hence is the best model. Keywords: Stylometric analysis, Tones motifs, Word length motif, Chinese prose 1 Introduction Style refers to linguistic choices made by an author that distinguish his/her writing from those of other authors (Herdan 1966). Stylometric analysis, can distinguish texts written by different authors by measuring some stylistic features in text. It is assumed that quantitative authorship attribution is that the anonymous author of a text can be selected from a set of possible authors by comparing the values of textual measures in that text to their corresponding values in each possible author’s writing samples (Grieve 2007). In fact, textual measurements are assumed to include conscious and unconscious aspects of the author’s style. It would then be an asset to find the features of the unconscious aspect, since they can not be consciously manipulated by the author (García & Martin 2006). Stylometric analysis involves extracting style markers, i.e. stylometric features, and classifying the texts represented by those features according to authors (Stamatatos et al. 2000). These models can be seen as the text classification according to their authors. The most effective features to discriminate between different authors, i.e. style markers, should be determined at first. A great variety of measures, including sentence length, word length, word frequencies, character frequencies and vocabulary richness had been proposed. Savoy (2012) compared the performance obtained when using word types or lemmas as text representations. Koppel et al. (2009) compared the performances of several representative learning methods for authorship attribution and showed that the choice of the learning algorithm is no more important than the choice of the features by which the texts are to be represented. This paper examines whether lexical information, such as tones motifs and word length motifs, can serve as effective stylometric features in authorship attribution. The motivation of such a study is both to find an effective model of stylometric study without annotation and processing, as well as to test the effectiveness of the linked data approach to stylometric studies. 56 1.1 Literature review …

Content truncated to 500 words