The effect of topic-shift characteristics in daily conversation on identification of recognition errors

Abstract

To support senior and reclusive citizens’ smooth conversations, we have developed a conversation support system named “associative board”. it recognizes their conversation and provides several suitable topics for speakers when their conversation progress not so smooth. However. If there are too many recognition errors the system will not be able to present suitable words. The misrecognized words identification function is necessary for our associative board system. In this study, we clarify the problems with conventional misrecognized words identification methods for recognizing daily casual conversation. As results of evaluation, the conventional misrecognized words identification is effective for the conversations with predefined topics, however for casual conversations without predefined topic, the identification is difficult. The distribution of semantic similarity values among words for casual conversation are broader than that with predefined topics. When the semantic similarity values are under 0.3, despite the correct recognition utterances, that semantic similarity values of the recognition results are often lower than that of the misrecognition results. The 21.7% to all topics are that case. That means when the casual conversations in which the topic-shifting occurs frequently, the misrecognized words identification is difficult. The semantic similarity among recognized words should be calculated constantly, and when the semantic similarity values are high continuously or are low rarely, the identification method could be used. When the semantic similarity values are low continuously, the error words extraction and correction process should be stopped.