Short story genre prediction: a study driven by transformer and sequential models

Abstract

A huge number of genres that represent a variety of themes, styles, and narrativestructures are represented in short stories. Genre detection in short stories is anintegral task for analysis, recommendation systems, content organization, and research. Therefore, the topic we have decided to work on is Short Stories GenreDetection using NLP techniques. The goal is to identify the genre of stories sothat the audience may quickly determine whether or not the story is appropriatefor them. Our approach is based on feeding the model as much big data as possible and maximizing efficiency by emphasizing on non automatically marked-uptext features. Specifically, we experimented with models such as LSTM, Bi-LSTM,and transformer-based architectures like BanglaBERT, DistilBERT, and RoBERTafor genre classification tasks. After research, BanglaBERT showed the best resultswhile the sequential models showed the least improved results. Among the fivemodels evaluated, BanglaBERT achieved the highest performance with an accuracyof 72% , significantly outperforming both sequential and transformer-based models.RoBERTa followed in 2nd place showing strong results especially for certain classes.DistilBERT placed last among the BERT models. Moreover, Bi-LSTM showed better performance than LSTM with an F1-score of 60%, compared to LSTM. Overall,transformer models, particularly BanglaBERT and RoBERTa, performed significantly better compared to the LSTM-based models.