2:20 PM - 2:40 PM
[2E2-04] TULIP: Generation of Screenplay-style Novel (SS) Using Three-step LSTM Trained on Web Novels
Keywords:Novel Generation, Conversation Generation, LSTM, SS, Deep Learning
This paper proposes a new method to generate screenplay-style novels (SS: side story or short story).
The key idea is to incorporate each talker's utterance-style into existing neural conversation model.
Our model, three-step unified LSTM interlocution producer (TULIP), consists of unified three LSTMs: utterance-encoding LSTM, context-updating LSTM, and utterance-decoding LSTM.
We trained our model on screenplay-style novels on the web, and generated new novels from scratch, in other words, not by direct processing of existing novels.
We evaluated novels generated by our model both in quality and quantity, and confirmed that our LSTM based method generated a natural Japanese sentence sequence.
We also indicate the limitation of existing quantitive evaluation methods.
The key idea is to incorporate each talker's utterance-style into existing neural conversation model.
Our model, three-step unified LSTM interlocution producer (TULIP), consists of unified three LSTMs: utterance-encoding LSTM, context-updating LSTM, and utterance-decoding LSTM.
We trained our model on screenplay-style novels on the web, and generated new novels from scratch, in other words, not by direct processing of existing novels.
We evaluated novels generated by our model both in quality and quantity, and confirmed that our LSTM based method generated a natural Japanese sentence sequence.
We also indicate the limitation of existing quantitive evaluation methods.