A comparison of Data Driven models of solving the task of gender identification of author in Russian language texts for cases without and with the gender deception

Abstract

Abstract. In this work we compare several data-driven approaches to the task of author's gender identification for texts with or without gender imitation. The data corpus has been specially gathered with crowdsourcing for this task. The best models are convolutional neural network with input of morphological data (f1-measure: 88%±3) for texts without imitation, and gradient boosting model with vector of character n-grams frequencies as input data (f1-measure: 64% ± 3) for texts with gender imitation. The method to filter the crowdsourced corpus using limited reference sample of texts to increase the accuracy of result is discussed.

Data-driven Approaches to Author’s Profiling Identification for Russian Texts on Base of Complex Machine Learning Models in Combinations with Siamese Networks

“…For such languages as English, Spanish, and Arabic, there are large text corpora to create data-driven models to identify author's profile (see Section 3). A few last years the author's profiling identification task has been rapidly developing for Russian, and it is strongly related to formation of similar Russian corpora, but the sizes of these corpora are currently smaller: in our work we use Gender-Imitation-Crowdsource ("GI cs") corpus [1], , and Gender Imitation corpus [2] Naturally, it makes it more difficult to reach the high precision of task solution.…”

“…The adaptation of complex models, based on convolutional neural networks, gradient boosting methods, LSTM, Siamese networks is described in Section 4.2. We then apply the same methods to the 'GI cs' datasets for the task of gender prediction (Section 7) and compare them to our previous results [1,3], that demonstrated the accuracy of 88% ±3%, which is about 30% more than the baseline.…”

A Gender Identification of Russian Text Author on Base of Multigenre Data-Driven Approach using Machine Learning Models

“…The purpose of this paper is to evaluate the accuracy of solving multigenre profiling task more correctly with cross-validation, using an extended set of text features along with neural net and machine learning models. Models, which are effective and largely independent of genre and external dictionaries, have been proposed in our previous work (Sboev A., 2017).The sets of used features included: morpho-syntactic, linguistic features (LIWC), a generalized dictionary approach of low displacement rank (LDR), along with different variants of vector representation of the text. The adaptations of these models and more complex models, based on convolutional neural networks (CNN), long short-term memories (LSTM), gradient boosting methods, are described in Section Materials and Methods, see Subsection Models.…”