LSTM vs. GRU vs. Bidirectional RNN for script generation
This is an incremental study for scriptwriters and AI researchers, as it applies existing methods to a new domain without achieving state-of-the-art results.
The paper tackled the problem of generating new conversations and scenarios for TV scripts by comparing LSTM, GRU, and Bidirectional RNN models, finding that these models performed similarly with no concrete numerical improvements reported.
Scripts are an important part of any TV series. They narrate movements, actions and expressions of characters. In this paper, a case study is presented on how different sequence to sequence deep learning models perform in the task of generating new conversations between characters as well as new scenarios on the basis of a script (previous conversations). A comprehensive comparison between these models, namely, LSTM, GRU and Bidirectional RNN is presented. All the models are designed to learn the sequence of recurring characters from the input sequence. Each input sequence will contain, say "n" characters, and the corresponding targets will contain the same number of characters, except, they will be shifted one character to the right. In this manner, input and output sequences are generated and used to train the models. A closer analysis of explored models performance and efficiency is delineated with the help of graph plots and generated texts by taking some input string. These graphs describe both, intraneural performance and interneural model performance for each model.