Reference: Johnson, K.; Poon, A.; Shiffman, S.; Lin, R.; & Fagan, L. Q-Med: A Spoken-Language System to Conduct Medical Interviews. Knowledge Systems Laboratory, February, 1992.
Abstract: Continuous-speech-recognition technology (CSRT) promises to be a useful modality for human-computer interaction. Unfortunately, usable spoken- language systems have been difficult to build, in part due to problems with misrecognitions when large speaker-independent vocabularies and language models are used. As part of a project to create spoken-language systems that achieve acceptable performance in spite of partially misrecognized input, we have developed Q-MED, a system that creates applications using CSRT in the task of medical interviewing. The system uses questions arranged hierarchically: from open-ended questions that have large language models, to more directed questions that use smaller language models. This hierarchy of questions allows the system to recover from unintended or misinterpreted utterances by asking more directed questions until an adequate answer is recognized. Furthermore, information-retrieval techniques map an utterance to one or more predefined symptoms, even if only some of the words in the utterance are recognized correctly. This paper discusses the rationale behind and implementation of Q-MED, using examples from an application created to interview patients who have abdominal pain.