Response Obligation Estimation That Considers Users' Repetitive Utterances using Knowledge-Guided Random Forest

Abstract

Response obligation is whether a spoken dialogue system should react to an input sound. This paper focuses on the false negative errors in response obligation estimation (ROE) that are displayed as the system's neglect of its users. When the users repeat after the system ignores their speech, ROE will likely fail again because the repeated input is similar to the previous input. Therefore, we propose an improved ROE method that considers users' repetitions. First, we show that a simple concatenation of ROE and repetition features is better than two other integration architectures. Then, we propose a modified random forest algorithm that incorporates human domain knowledge. The effectiveness is demonstrated with simulated repetitions as a 7.6-point gain from the baseline.