A Similarity-Based Approach to Attribute Selection in User-Adaptive Sales Dialogs

Abstract

For dynamic sales dialogs in electronic commerce scenarios, approaches based on an information gain measure used for attribute selection have been suggested. These measures consider the distribution of attribute values in the case base and are focused on the reduction of dialog length. The implicit knowledge contained in the similarity measures is neglected. Another important aspect that has not been investigated either is the quality of the produced dialogs, i.e. if the retrieval result is appropriate to the customer’s demands. Our approach takes the more direct way to the target products by asking the attributes that induce the maximum change of similarity distribution amongst the candidate cases, thereby faster discriminating the case base in similar and dissimilar cases. Evaluations show that this approach produces dialogs that reach the expected retrieval result with fewer questions. In real world scenarios, it is possible that the customer cannot answer a question. To nevertheless reach satisfactory results, one has to balance between a high information gain and the probability that the question will not be answered. We use a Bayesian Network to estimate these probabilities.