Deep Ensemble Approach for Classifying Documents Based on Writing Styles

As digital content grows exponentially, efficient methods for automatically classifying texts are becoming increasingly important. Since most scholars and professionals use the web to collect some documents, it is most important to identify the document with a suitable writing style such as formal or informal. As a solution, research suggests a deep learning-based method for classifying texts according to linguistic markers. A dataset was collected from multiple web sources, containing both formal documents (news stories, formal documents) and informal documents (personal letters, personal blogs). After pre-processing, to create the feature vector, we took into account 13 linguistic markers like voice, language type, modal verbs, slang, acronyms, pronouns, phrasal verbs, colloquialisms, contractions, abbreviations, initialism and grammar, and vocabulary. These linguistic features are very important to identify the difference between formal and informal styles. Then, Long Short-Term Memory Networks (LSTM), Convolutional Neural Networks (CNN), and Artificial Neural Networks (ANN) were taken into consideration. Also, the ensemble approach by majority prediction is applied combining the above three algorithm results. The 80%, 90%, 70%, and 60% were used to train and evaluate the classification models through a variety of testing splitting methods. Using performance indicators, f1-score, accuracy, precision, and recall, the models' efficacy was assessed. The ensemble approach did better than the other individual algorithms, with a maximum accuracy of 87.2% and resilience in identifying between formal and informal texts. Subsequent investigations will explore larger datasets, more language features, advanced deep learning models, as well as multilingual and real-time classification systems.