Abstract: In embodiments, a reusable and adaptive multi-task orchestration dialogue system orchestrates a set of single-task dialogue systems to provide multi-scenario dialogue processing. In embodiments, for each question propounded by a user, using a deep learning predictive model, a best single-task dialogue system is chosen out of the set. In embodiments, multi-task orchestration is done without the need to change, or even understand, the inner workings or mechanisms of the individual single-task dialogue systems in the set. Moreover, the multi-task orchestration is also unconcerned with what rules are set in each individual single-task dialogue system. In embodiments, prior to selection of the best single-task dialogue system to return the best answer, new intents and entities are discovered and used to update an existing dialogue path. In embodiments, additional data is continually collected, and used to retrain model so as to further improve performance.

Abstract: A method, system, and computer program product processes dialog data. The method includes obtaining dialog data including heterogeneous data items. The method includes generating a heterogeneous network based on the dialog data. The heterogeneous network includes two or more bipartite subnetworks representing the relationship of the data items in the dialog data. The nodes of the two or more bipartite subnetworks correspond to the data items in the dialog data. The method includes determining node representations for the nodes in the heterogeneous network through representation learning.

Abstract: Aspects of the present invention disclose a method for consolidating of a plurality of personal bills from diverse financial sources to reflect the payments, expenses, and balances without duplication. The method includes one or more processors parsing a plurality of bills of a user, the plurality of bills including bills with varying formats. The method further includes identifying a set of bills of the plurality of bills of the user, the set of bills including related bills based at least in part on a prebuilt rule. The method further includes determining a correlation of one or more items of respective bills of the set of bills of the user based at least in part on a machine learning algorithm. The method further includes generating a consolidated bill, from the set of bills of the user, based at least in part on the determined correlation of the one or more items.

Abstract: A computer-implemented method is presented for enabling hierarchical conversational policy learning for sales strategy planning. The method includes enabling a user to have a conversation with a robot via a conversation platform, employing a plan database to store general plans used in the conversation, employing an industry database to store a plurality of candidate plans pertaining to sales promotions, and employing a plan and policy optimizer to allow the robot to select and output an optimal plan from the plurality of candidate plans, the optimal plan determined by hierarchical reinforcement learning via a first learner and a second learner, the first leaner selecting the optimal plan and the second learner selecting an optimal action.

Abstract: In embodiments, a reusable and adaptive multi-task orchestration dialogue system orchestrates a set of single-task dialogue systems to provide multi-scenario dialogue processing. In embodiments, for each question propounded by a user, using a deep learning predictive model, a best single-task dialogue system is chosen out of the set. In embodiments, multi-task orchestration is done without the need to change, or even understand, the inner workings or mechanisms of the individual single-task dialogue systems in the set. Moreover, the multi-task orchestration is also unconcerned with what rules are set in each individual single-task dialogue system. In embodiments, prior to selection of the best single-task dialogue system to return the best answer, new intents and entities are discovered and used to update an existing dialogue path. In embodiments, additional data is continually collected, and used to retrain model so as to further improve performance.

Abstract: A computer-implemented method is presented for enabling hierarchical conversational policy learning for sales strategy planning. The method includes enabling a user to have a conversation with a robot via a conversation platform, employing a plan database to store general plans used in the conversation, employing an industry database to store a plurality of candidate plans pertaining to sales promotions, and employing a plan and policy optimizer to allow the robot to select and output an optimal plan from the plurality of candidate plans, the optimal plan determined by hierarchical reinforcement learning via a first learner and a second learner, the first leaner selecting the optimal plan and the second learner selecting an optimal action.