Extraction of User Demands Based on Similar Tweets Graph

Abstract

Twitter is used by many users, and posted tweets include user's straightforward real intention. Therefore, we can obtain various opinions on items and events by collecting tweets. However, since the tweets are posted one after another over time and are represented by characters, it is difficult to grasp the overall picture of opinions on items. Therefore, by visualizing opinions on items, it is easier to grasp the whole picture more clearly. In this study, we collect tweets including item names and construct a graph connecting similar tweets. Then, from the connected component, we attempt to extract expressions related to user demands. Also, when constructing a similar tweet graph, it is necessary to appropriately set the similarity threshold. If the threshold is too low, unrelated tweets will be connected and a connected component will consist of different demand expressions. On the other hand, if the threshold value is too high, the demand expression of the same meaning will be divided as other connected components due to some notation fluctuation. In this paper, by focusing on the occurrence probability of the demand expression appearing in each connected component and defining the purity and the cohesiveness, we propose a method of setting the apropriate similarity threshold. In our experimental evaluations using a lot of tweets for two games “Mario tennis ace” and “Dairanto smash brothers SPECIAL”, we confirmed that opinions such as “interesting” or “difficult” can be extracted from similar tweets graph constructed by the appropriate similarity threshold value. We also confirmed that we can overlook the demands related to items.