Publications

25TH INTERNATIONAL CONFERENCE ON COMPUTER AND INFORMATION TECHNOLOGY (ICCIT) 2022

DOI : 10.1109/ICCIT57492.2022.10055349

Authors : Md Mezbaur Rahman, Saadman Malik, Mohammed Saidul Islam, Fardin Saad, Md Azam Hossain, Abu Raihan Mostofa Kamal

Abstract:

Automated tag prediction is one of the critical techniques to enhance the retrieval of similar movies in recommendation systems. Additionally, it serves the purpose of drawing a preview of the content the viewers hope to watch. However, the quotidian architectures employed for this task encompass sequence-to-sequence models like Neural Networks (NN), Long Short Term Memory (LSTM), etc., which can not learn the overall context. On the contrary, Transformer-based Language models are tailored toward understanding the broad context since the tokens generated are learned parallelly. Therefore, our study proposes an Encoder-based Transformer Language model to predict movie tags from movie plot synopsis. We use the Movie Plot Synopses with Tags (MPST) dataset and employ lightweight Vanilla Neural Networks (VNN) and RoBERTa language models as our pipeline for leveraging contextual knowledge to predict movie tags from an extensive movie synopses corpus. The MPST dataset has an uneven distribution of tags, and reducing the tag space ensued the model’s efficacy by achieving a 38.34% F1 score and 21.14% Recall, which outperformed previous studies for automated movie tag prediction.

Keywords: Pretrained Language Model, LSTM, Tag Prediction, Vanilla Neural Network