We propose an entropy minimization dialog management (DM) strategy for goal-driven information retrieval (IR). By associating each goal of an IR task with a set of stochastic attributes, reaching a goal can then be accomplished by filling the “attribute slots” corresponding to the goal. Information access can now be cast as a dialog problem that specific information about the attributes is solicited from a user by the system through multiple dialog turns. For a real-world music search task with 38118 songs and 12 attributes corresponding to each song, we demonstrate the concept by designing a simulation game to order song from the above music database. We show that 8.3 dialog turns are needed on the average if random questions are asked by the system, whereas the entropy minimization DM is a very efficient goal seeking method to order a song with the least amount of 3.3 dialog turns among different strategies. Furthermore, the proposed DM techniques can manage the dialog process in a more efficient and flexible manner.
Bibliographic reference. Wu, Ji / Li, Miao / Lee, Chin-Hui (2015): "An entropy minimization framework for goal-driven dialogue management", In INTERSPEECH-2015, 2027-2031.