Inference of the Day Topic Word Using WWW Search Rankings for Computer Conversations

Recently, demand has arisen for robots that can converse naturally with human beings. To accomplish this, such robots should to be able to offer appropriate topics to the person engaging them in conversation. Thus, it would be useful if such robots had the ability to collect news of interest to that person, which could then be used to help conversations proceed more smoothly. This paper proposes a method of extracting Web search keywords of interest by gender, age and day of the week. In this study, the authors compiled Web search engine keywords based on rank using the Biglobe Search Shunkan Ranking system. However, such rankings cannot be used independently because the relevant words for a particular day will not be known until that day is over. Accordingly, their method uses Web search rankings compiled for the day prior to day in question to select keywords for use. Because of this, the method cannot determine keywords that appear on the actual day of the conversation. However, it can extract keywords that appear repeatedly in identifiable patterns, as well as topics of general interest to a certain age and gender. In their experiments, the authors experienced particularly good results by focusing on three concepts： keyword consolidation, forgetting factor and short-term and long-term memory.