Incremental learning of new user formulations in automatic directory assistance

Directory Assistance for business listings is a challenging task: one of its main problems is that customers formulate their requests for the same listing with great variability. Since it is difficult to reliably predict a priori the user formulations, we have proposed a procedure for detecting, from field data, user formulations that were not foreseen by the designers. These formulations can be added, as variants, to the denominations already included in the system to reduce its failures. In this work, we propose an incremental procedure that is able to filter a huge amount of calls routed to the operators, collected every month, and to detect a limited number of phonetic strings that can be included as new formulation variants in the system vocabulary. The results of our experiments, tested on 9 months of calls that the system was unable to serve automatically, show that the incremental procedure, using only additional amount of data collected every month, is able to stay close to the (upper bound) performance of the not incremental one, and offers the possibility of periodically updating the system formulation variants of every city.