Comparing Statistical Similarity Measures for Stylistic Multivariate Analysis

International Conference RANLP 2009 - Borovets, Bulgaria, pages 349–354 Comparing Statistical Similarity Measures for Stylistic Multivariate Analysis Marius Popescu University of Bucharest Department of Computer Science Academiei 14, Bucharest, Romania mpopescu@phobos.cs.unibuc.ro Liviu P. Dinu University of Bucharest Department of Computer Science Academiei 14, Bucharest, Romania ldinu@funinf.cs.unibuc.ro Abstract The goal of this paper is to compare a set of dis- tance/similarity measures, some motivated sta- tistically, others motivated stylistically, regard- ing their ability to reflect stylistic similarity between texts. To assess the ability of these distance/similarity functions to capture stylistic similarity between texts, we have tested them in the two most frequently employed multivariate statistical analysis settings: cluster analysis and (kernel) principal components analysis. Keywords Stylistic Multivariate Analysis, Statistical Similarity Measures, Cluster Analysis, Kernel Principal Components Analysis 1 Introduction Computational stylistics investigates texts from the standpoint of individual style (author identification) or functional style (genres, registers). Because in all computational stylistic studies/approaches, a process of comparison of two or more texts is involved, in a way or another, there was always a need for a dis- tance/similarity function to measure similarity (or dis- similarity) of texts from the stylistic point of view. Usually, the distance/similarity measures are im- plicitly or explicitly used by multivariate statistical analysis techniques typically applied in computational stylistic approaches. In [5], these approaches are char- acterized as: ”[The]...technique essentially picks the N most common words in the corpus under investigation and computes the occurrence rate of these N words in each text or text-unit, thus converting each text into an N-dimensional array of numbers. Multivariate sta- tistical techniques are then applied to the data to look for patterns. The two techniques most frequently em- ployed are principal components analysis and cluster analysis.” The goal of this paper is to compare a set of dis- tance/similarity measures, some motivated statisti- cally, others motivated stylistically, regarding their ability to reflect stylistic similarity between texts. As style markers we have used the function word frequencies. Function words are generally considered good indicators of style because their use is very un- likely to be under the conscious control of the au- thor and because of their psychological and cognitive role [3]. Also function words prove to be very effective in many author attribution studies. The distance/similarity between two texts will be measured as distance/similarity between the func- tion words frequencies corresponding to the respec- tive texts. For this study we selected some similar- ity/distance measures. We started with the most nat- ural distance/similarity measures: euclidean distance and (taking into account the statistical nature of data) Pearson’s correlation coefficient. Since function words frequencies can also be viewed as ordinal variables, we also considered for comparison some specific sim- ilarity measures: Spearman’s rank-order coefficient, Spearman’s footrule, Goodman and Kruskal’s gamma, Kendall’s tau. Finally, we have added a stylistically motivated similarity measure: Burrows’s delta, that has interesting statistic interpretations. To assess the ability of these distance/similarity functions to capture stylistic similarity between texts, we have tested them in the two most frequently em- ployed …

Content truncated to 500 words