Applying Topic Recognition to Spoken Language in Human-Robot Interaction Dialogues
Manuel Giuliani, Thomas Marschall, Manfred Tscheligi · 2014
Human-robot interaction systems that work in everyday situations need to be able to talk about different topics, for example when the robot is a bartender that serves drinks to human customers. We applied a topic recognition approach that is based on term frequency-inverse document frequency (TF-IDF) on a test set of spoken language interactions between human customers and bartenders in German bars. We sorted the test set into five topics and evaluated our topic recognition with different topic corpora. Our evaluation shows that recognition accuracy is only as high as 70.2% for certain topics and at 30.0% on average, even for manually created topic corpora. This result suggests that a multimodal approach is needed to automatically recognise topics in spoken language more sufficiently.