| CPC G06F 16/3329 (2019.01) [G06F 40/284 (2020.01); G06F 40/40 (2020.01)] | 20 Claims |
|
1. A computer-implemented method for improving a predictive response of a machine learning-based virtual dialogue system, the computer-implemented method comprising:
identifying one or more required and unfilled dialogue slots associated with a dialogue intent classification in response to computing the dialogue intent classification for dialogue input data provided by a user;
executing a machine learning-derived context pairing operation using unsupervised learning models, including a transformer-based neural network, to transform the dialogue input data into slot-specific contextualized answers stored in an answer bank,
wherein the answer bank is dynamically generated and includes the slot-specific contextualized answers and slot-specific contextualized questions as key-value pairs for predicting slot values in response to user queries, wherein executing the machine learning-derived context pairing operation includes:
extracting one or more slot values from the dialogue input data based on identifying the one or more required and unfilled dialogue slots, wherein identifying the one or more required and unfilled dialogue slots includes using the dialogue intent classification to perform a slot lookup, within a dialogue reference data structure, of the one or more required and unfilled dialogue slots that are stored in association with the dialogue intent classification, and
(i) using the respective unfilled slot to predict, by a context transformer model, a likely answer to a prospective question about the respective unfilled slot by creating a context transform pairing between the dialogue input data and one dialogue slot of the one or more required and unfilled dialogue slots;
(ii) transforming the context transform pairing, by the context transformer model, to a slot-informed contextualized answer;
(iii) in response to the transformation of the context pairing, storing the slot-informed contextualized answer to the answer bank of a question-answering model;
(iv) using the respective unfilled slot to predict, by a question transformer model, a likely question for discovering a prospective answer containing the respective unfilled slot by creating a question transform pairing between the dialogue intent classification and the one dialogue slot of the one or more required and unfilled dialogue slots;
(v) transforming the question transform pairing, by a question transformer model, to a slot-informed contextualized question;
(vi) using the slot-informed contextualized question as input for questioning the question-answering model about the respective unfilled slot; and
(vii) in response to receiving the slot-informed contextualized question about the respective unfilled slot, predicting by the question-answering model a distinct slot value for the respective unfilled slot based on the answer bank of the question-answering model;
implementing a slot arbiter that generates a signal prompting, via a user interface, a completion response for the respective unfilled slot, wherein the completion response includes a confirmation response to modify, via a user interface, the distinct slot value for the respective unfilled slot; and
automatically executing a set of pre-determined computer-executable instructions for performing a task in response to or responding to the dialogue input data based on the computing the distinct slot value for each of the one or more required and unfilled dialogue slots associated with the dialogue intent classification of the dialogue input data provided by the user.
|