SYSTEM AND METHOD OF OPERATING A CONTEXT OF HISTORICAL EVENTS FOR THE IMPROVEMENT OF A CHATBOT'S DIALOGUE MANAGER

摘要

The present invention relates to a system comprising a computer providing an interface between a front-end application configured to receive one or more user oral formulations from a smart portable object equipped with a microphone and speaker for transmitting and receiving in a language. natural language and a main system with which it is in communication, the main system comprising one or more processors configured to provide a service to this user, the frontend application consisting in: receiving from a natural language comprehension module (NLU) a input information representative of the text formulated in natural language via the application of the intelligent object; in structuring the information received taking into account the history of the dialogues in the form of several triplets (preferably three) at least two concerning user information, agent information, and a possible third concerning statistical information which provides information through their sub vectors on the type of action taken (question, answer, polite phrase, etc.), the information provided (in the case of an action response for example) and the information requested (in the case of an action question ) and grouping these triplets together in the form of a dialogue observation vector ot; perform at least one calculation on each parameter (Ars 0, Ars 1, ..Ars i) several observation vectors thus constituted by each using a plurality of observation vectors and at least one layer of recurrent neural network of LSTM type per information sub-vectors, to generate a latent representation (denoted h (Ars)) of several observations [o0, ...., ot], one after another in the form of a vector h (t), to apply the output vector to an FC reinforcement algorithm to calculate Q-values to deduce the action to be carried out according to the state ht; communicating the identified action to the Natural Language Generation Module (NLG) to provide the front-end application with natural language text output based on the identified action or perform the requested action. End up storing the observations of new dialogues in a memory (Experience Replay) to be able to reuse them during the next training sessions.