Conversational Analysis of Word Wolf based on the Contextual Transition in Distributed Representation Space

Abstract

In conversation, one cannot express what he/she wants to convey at once since it is hard to encode all the information thoroughly. Therefore, the conversation partners have to interpret what the others want to say from ambiguous utterances. The existing dialogue systems tend to clarify the ambiguity of utterances by posing the questions multiple times, which is stressful for users. To reveal how humans estimate what another person tries to say in such ambiguous utterances, we analyze dialogues in the conversation game "Word Wolf" with employing our previously proposed context estimation method, SCAIN. We developed a method to estimate which keyword the player has using the SCAIN algorithm. In the experiment, we collected each participant's utterances in the game and verified the method by analyzing the transition of the estimations in the conversation.