Generating Achievement Relationship Graph Between Actions for Alternative Solution Recommendation

Tsukasa Hirano, Yoshiyuki Shoji, Takehiro Yamamoto, Kouzou Ohara

Abstract: This paper proposes a method for generating a graph of the purpose and alternative actions for a given specific action name. For instance, the purpose of “concentrating on the lecture” can be achieved by taking actions such as “preventing sleepiness,” “taking sugar,” or “changing the lecture.” In this case, the actions “concentrating on the lecture” and “preventing sleepiness” have a means-end (i.e., achieved-by) relationship, and “preventing sleepiness” and “taking sugar” have substitutable relationships. Our method presents these achievable and substitutable actions as a graph to give people various choices and help them make decisions. The proposed method extracts descriptions of actions from product review data using a large-scale language model and assigns action names. The relationship between action names is calculated based on graph computation and language patterns in web documents. The results of subject experiments show that the proposed method can find many pairs of action names in an achievement relationship.