Ego-state Estimation from Short Texts Based on Sentence Distributed Representation

Human personality multilaterally consists of complex elements. Egogram is a method to classify personalities into patterns according to combinations of five levels of ego-states. With the recent development of Social Networking Services (SNS), a number of studies have attempted to judge personality from statements appearing on various social networking sites. However, there are several problems associated with personality judgment based on the superficial information found in such statements. For example, one's personality is not always reflected in every statement that one makes, and statements are influenced by a personality that tends to change over time. It is also important to collect sufficient amounts of statement data including the results of personality judgments. In this paper, to produce an automatic egogram judgment, we focused on the short texts found on certain SNS sites, especially microblogs. We represented Twitter user comments with a distributed representation (sentence vector) in pre-training and then sought to create a model to estimate the ego-state levels of each Twitter user using a deep neural network. Experimental results showed that our proposed method estimated ego-states with higher accuracy than the baseline method based on bag of words. To investigate changes of personality over time, we analyzed how the match rates of the estimation results changed before/after the egogram judgment. Moreover, we confirmed that the personality pattern classification was improved by adding a feature expressing the degree of formality of the sentence.