[1I4-GS-2] Machine learning: Applied machine learning (1)

[1I4-GS-2-01] Evaluating response generation for character by pointer-generator-mechanism

〇Sohei Okui¹, Makoto Nakatsuji¹ (1. NTTresonant)

Keywords:DeepLearning, LSTM, Generation Model, Chatbot, Character

Japanese businesses that earn profits by licensing character content are expanding not only in Japan but also overseas, and their economic effects are significant. In the past, our company has provided many chatbot services that enable users to talk to characters, and we confirmed that there is demand among fans for this opportunity to converse with characters they like. We now proposed a model that generates answers to question documents by extracting word tokens from a number of information sources. This model considers character answers obtained from a model constructed in the past as reference documents, and considers user utterances as question documents. By extracting word tokens from reference documents and using them for answer generation, the model can generate a character response to a user utterance. If this proposal can be realized, we will be able to create various character personalities at low cost. In this paper, we propose a model in which a pointer formation mechanism was applied to character response generation, as well as criteria for its evaluation.