Trivia Score and Ranking Estimation Using Support Vector Regression and RankNet

摘要

Dialogue systems have been increasingly important these days. In particular, non-task-oriented dialogue systems are studied because of the success of neural network approaches such as seq2seq models. However, these models tend to generate simple responses such as "yes" and "ok." To construct a dialogue system that holds users' attention continuously, we need to generate utterances that capture the interest of the user. In this paper, we propose a method to extract trivia sentences for the purpose. Trivia information perhaps adds a surprise to users. Therefore, capturing trivia information is beneficial for dialogue systems. We estimate a trivia score of a sentence by using machine learning approaches, Support Vector Regression (SVR) and RankNet. We obtained 0.79 and 0.78 on SVR for the nDCG@5 and RankNet for the nDCG@10, respectively. We focus on the subject word in each sentence. The method with subject information outperformed that without subject information; 0.79 with subject information vs. 0.64 without subject information on the SVR for the nDCG@5.