Machine Feeling by Knowledge Acquisition with Emotion Map

Artificial intelligent systems have been developed by many industries and academic institutes, and there are also many recent studies about emotion or sentiment. The previous studies about the emotion commonly treated the emotion as knowledge trainable from emotion-annotated data, and focused on emotion perception or expression but not emotion generation. In this paper, we design a graph attention network (GAT) for emotion generation based on the assumption that emotion is intrinsic and propagates as more knowledge is acquired. We represent the knowledge as a graph, and formulate that the knowledge acquisition as an expansion of the knowledge graph. The graph gets larger as time-step goes by, and the GAT learns from the time-series graph data. We simulate this knowledge acquisition based on the assumption that emotion propagates to the newly acquired knowledge. Interestingly, the simulation results exhibited behaviors that are consistent with previous findings of psychiatry. We believe our study will contribute to development of human-like emotional agents that have its own unique emotion about what it experienced or learned.