A Feature Selection Technique–Based Approach for Author Profiling Using Word Embedding Techniques

Textual data are tremendously increasing on the World Wide Web through different social networking platforms, such as blogs, review sites, Facebook, Instagram, Twitter, etc. Crimes or frauds are also growing with the development of text in these platforms. Most of the users on these platforms create accounts with false information and send harassing and threatening messages. Knowing the details or basic information of text posted on these platforms is a very important task to control fraud. Author profiling (AP) is one technique used by the research community and social network administrators to know the author’s information related to messages posted on social networking platforms. Author profiling is a method of detecting demographic profiles such as age, gender, education, etc., of authors based on their written texts. Author profiling techniques are used in various applications of text processing, such as marketing, forensic analysis, and security-related fields. Most author profiling techniques use the content information of text to differentiate the writing styles of authors. The dataset contains a huge number of content words and, the identification of relevant words for differentiating the writing style of different authors is one challenging task to the research community. Feature selection algorithms are proposed by the researchers to find relevant features in the dataset. In this article, we developed a feature selection technique–based approach by using word embedding techniques for predicting the gender and age from the selected dataset. The PAN competition author profiling datasets are considered in this article for experimentation. In the proposed approach, word embedding techniques are used for converting words into word vectors. The similarity measure is used for finding the similarity among the word vectors. Similar words are grouped into clusters. Select the most important words from these clusters and consider them for experimentation. Selected words are used for representing the documents as vectors. These vectors are passed to machine learning (ML) algorithms to develop a model for classification. This model predicts the accuracy of the proposed method. The proposed approach attained the best accuracies for gender and age prediction when compared with the accuracies of several well-known approaches.