Abstract:
Because of the wide variety of formal and informal writing styles brought about by
the rapid growth of digital communication, the classification of documents based on
it becomes a challenging task. Using a variety of variables, this work seeks to increase
the accuracy of formality classification algorithms. Grammar, vocabulary,
punctuation, and sentence structure are some stylistic components that define various
writing styles, and traditional approaches have trouble distinguishing between them.
Differentiating between formal and informal language is becoming increasingly
important in applications such as research papers, legal documents, informal letters,
NEWS, etc. The objective of this approach is to use linguistic features, examines how
well deep learning algorithms classify documents as formal and informal. The study
collected dataset of 5,000 text samples. The text files contained 2500 formal letters,
news items as formal documents, and remaining are personal blogs, personal letters
as informal documents. Next pre-processed all data using stop word removal,
lemmatization, tokenization and lowercasing. Formal and informal categories which
include pronouns, grammar, vocabulary, slang, acronyms, language and initialisms
seven linguistic features were targeted for this study and those features are extracted.
Then these seven features are combined to generated the feature vector for each
document. The generated feature vector was applied and in order to classify
documents, three deep learning models Artificial Neural Networks (ANN),
Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM)
networks are trained. Here, ANN learns nonlinear patterns in data, CNN identifies
text sections, and LSTM considers word position and those are selected based on the
literature review. The performance of each model is compared using different test
splitting methods and cross-validation techniques. According to experimental data,
the LSTM model outperforms ANN and CNN in terms of precision, recall and f
measure metrics, achieving the highest classification accuracy of 89.4% with an
epoch size of 100 and a batch size of 32 with lowest error rate for Mean Absolute
Error and Root Mean Squared Error. The results highlight how well LSTM can detect
linguistic subtleties and offer suggestions for improving formality recognition in
Natural Language Processing applications, which will help with more context
sensitive text classification.
