?
Automation of Forensic Authorship Attribution: Problems and Prospects
The article deals with validation of an integrative attribution algorithm based on the analysis of the author’s idiostyle using methods of interpretative linguistics with ob jectification of the available data with the help of mathematical statistics. The algo rithm addresses the identification problem of the attribution. The choice of parameters describing the individual style of an author assumes that the text is a product of an authentic language personality described by psycholinguistic (Yu.N. Karaulov), socio linguistic and forensic linguistic (S.M. Vul, M. Coulthard, R. Shuy) methods. To validate a hypothesis that the identification problem of attribution is best resolved by the inte grative methodology, we have created the KhoRom application which brings together the aforementioned approaches to the analysis of language personality: http://kho rom-attribution.ru/#/. It can be used to compare two language personality models and determine to what extent they are similar using the following metrics: Pearson correla tion coefficient, linear regression determination coefficient and Student’s t-criterion. Importantly, this application also describes the interpreted model of language person ality to inform the user on the importance of values of each parameter. The system has a wealth of features, with the user able to choose parameters, view parameter implementation in the document and edit the final list of parameter implementations (in case of malfunction, the application performance can be corrected manually). The created application is only a part of the attribution algorithm. The data produced by mathematical statistics need to be analyzed by expert judgment through the use of methodological recommendations developed for the algorithm. The effectiveness of this methodology has been proved by its validation on texts of various length and genres, with a number of documents pertaining to fiction, journalism, official and collo quial styles being analyzed. For texts of all discourses except colloquial, the developed algorithm has demonstrated a high level of accuracy (F-score of 0.8 to 1). For better applicability of the algorithm to colloquial texts, the authors have developed a number of improvements pending implementation.