Optimum Parameter Selection for K.L.D. Based Authorship Attribution in Gujarati

International Joint Conference on Natural Language Processing, pages 1102–1106, Nagoya, Japan, 14-18 October 2013. Optimum parameter selection for K.L.D. based Authorship Attribution for Gujarati Parth Mehta DA-IICT, Gandhinagar parth.mehta126@gmail.com Prasenjit Majumder DA-IICT, Gandhinagar prasenjit.majumder@gmail.com Abstract We examine several quantitative tech- niques of authorship attribution that have gained importance over the time and com- pare them with the current state of the art Z-score based technique. In this paper we show how comparable the existing tech- niques can be to the Z-score based method, simply by tuning the parameters. We try to find the optimum values for number of terms, smoothing parameter value and the minimum number of texts required for cre- ating an author profile. 1 Introduction Authorship attribution and author profiling have a long standing history dating back to 19th century (Mosteller and Wallace, 1964). While authorship attribution deals with determining whether or not a given author has written the given article, au- thor profiling aims at determining the age, gen- der, education level, etc. of the author form his article (Koppel et al., 2002). In the current work we focus only on authorship attribution. Author- ship attribution techniques can be broadly classi- fied into linguistic techniques and Statistical tech- niques. Until early 90’s majority of the work done was from a linguistic perspective. Only after (Holmes, 1994) did statistical methods gain im- portance. There were many attempts to solve this particular problem using various statistics of texts like mean sentence length, term frequency distri- butions (Zipf, 1932), word lengths, character fre- quencies (Peng et al., 2003), vocabulary richness, etc. Rudman (1998) proposed nearly 1000 differ- ent measures that could be used. A whole new field of study called stylometry evolved from these type of studies. In 2002, (Burrows, 2002) pro- posed a novel technique based on Z-score that cov- ered many of the above features, specially vocabu- lary difference and difference in term distribution into a single measure. Later, Savoy(2010) mod- ified this Z-Score that improved the result drasti- cally and this technique is currently the state of the art for authorship attribution. In this paper we compare three major statistical techniques for au- thorship attribution to the state of the art Z-score based technique. 2 Corpus Details The absence of a replicable and reliable corpora daunts the field of authorship attribution and more so for Indian languages. To the best of our knowl- edge the only work available to date for Indian languages is by (Bagavandas and Manimannan, 2008) where the corpus consisted of 55 Tamil arti- cles, 32 of which were attributed and 23 disputed and there were only three possible authors. Having this fact in mind and following the steps of (Savoy, 2012) the authors developed a corpora consist- ing of 5039 newspaper articles from the newspa- per Gujarat Samachar. These articles consist of 49 different weekly articles, from the supplements Shatdal and Ravi Purti, written by 40 distinct au- thors in the time period of 01-Dec-2011 to 1-Dec- 2012 and is made available on …

Content truncated to 500 words