Spoken language systems are difficult to build in part because they combine many diverse sources of knowledge, each of which needs to fit the needs of a particular domain and application. Learning approaches used in dialogue often (though not always) rely on the existence of a labeled corpus from which models parameters can be estimated; unfortunately this requires human effort to create labels or annotations. This project is developing a new approach, implicit learning, that addresses current shortcomings by leveraging natural patterns that occur in conversation to generate learning instances. The project is approaching the problem from three perspectives: 1) Understanding the leverage provided by implicit learning, specifically how the quality of information, its quantity and the distribution of learning opportunities affects its efficiency. 2) Determining the conditions under which learning opportunities can be elicited, specifically computing the utility of given interventions in terms of cost to the user versus gain in knowledge. 3) Augmenting the set of useful patterns through discovery, specifically through the retrospective analysis of past interactions. The project is using working dialog systems to conduct experiments and gather empirical data for analysis. The techniques and analyses created in this project should be applicable to the design of a wide variety of interactive systems and allow them to incorporate a natural non-obtrusive learning component. Their value lies in the reduction of the need for expert supervision of learning and the corresponding ability to evolve behavior over time. As such they constitute a key element of robust intelligence.
