Using Convolutional Neural Network for improving o; inference of interrogative sentences in a dialogue

Developing a dialogue system with high quality of the inference is currently one of the most popular topics in Natural Language Processing. It is important for a dialogue system to make appropriate responses to users' utterances. This short paper proposes a method for automatically recognizing if a sentence from a user's utterance is interrogative or declarative by using Convolutional Neural Network (CNN). Our approach utilizes the pitch of an utterance and Japanese sentence transliterated to Latin alphabet as input. Our proposed method has achieved precision, recall and Fl score 25 points higher than these reported in previous studies. Moreover, this research has compared the evaluation criteria by each component, namely the pitch of an utterance and Japanese sentence transliterated to Latin alphabet. This experiment has confirmed our hypothesis that the pitch component helps the model improve the quality of inference.