Automatic scene generation using sentiment analysis and bidirectional recurrent neural network with multi-head attention

Text generation is one of the complex tasks associated with natural language processing. For efficient text generation, syntax and semantics of the language have to be considered to assign context to key phrases. The main objective of the proposed work is to perform text generation specifically for movie scripts. The training data consist of a self-annotated corpus of movie scripts depicting scenes, specific to certain genre where the annotation mainly focuses on a specific director’s movie scripts. The scene generation is set forth by word embedding with sentiment classification where the emotionally analyzed words are vectorized using the EmoVec algorithm performing sentiment analysis. Based on the sentiment and location associated with each scene, context for the phrases is identified and proceeded to build a well-defined script. Bidirectional long short-term memory BLSTM with multi-head attention is used to capture the information processed in both forward and backward propagation in order to understand future context. The vocabulary is built using Stanford’s Internet Movie Database IMDB datasets to perform word-based encoding for which requirement of an extensive vocabulary is imminent.