Incorporating Human Intuitions into Data Augmentation to Detect Concentration on Conversation

This study proposes a data expansion method to classify the excitement of conversation. This study incorporates human intuitions for conversation excitement into data augmentation. Quantification of the human intuitions would efficiently assign correct labels to the data set generated by data augmentation. The pandemic of the new coronavirus has resulted in a loss of communication opportunities. We have lost opportunities that are important to form good relationships. A deep learning model to discriminate conversation excitement would contribute to increasing such important opportunities. However, training and using models to solve real-world problems requires a lot of data. There are many cases where sufficient data cannot be collected to train a model. In such cases, data augmentation is the most promising solution. We should pay attention to the point that effective data augmentation methods vary depending on the type and characteristics of the data. This study experimentally collects conversational data. It performs data augmentation on the conversational data. It creates datasets by similarity and trains multiple models. Comparing the accuracy of these models verifies the effectiveness of incorporating human intuitions into data augmentation. The paper discusses what kind of data augmentation technique works well to generate realistic conversation data with augmentation.