Dialog system with self-learning natural language understanding
AssigneeHitachi, Ltd.
InventorsTakeshi Homma, Masahito Togami
Life of this patent
Estimating the term…
Abstract
Example implementations described herein are directed to a dialog system with self-learning natural language understanding (NLU), involving a client-server configuration. If the NLU results in the client is not confident, the NLU will be done again in the server. In the dialog system, the human user and the system communicate via speech or text information. The examples of such products include robots, interactive voice response system (IVR) for call centers, voice-enabled personal devices, car navigation system, smart phones, and voice input devices in the work environments where the human operator cannot operate the devices by hands.
Claims19
A method for a dialog system comprising a client device configured with a text input method to receive user input, and a first natural language understanding (NLU) model configured to communicate with a server, the user input involving at least one of text input typed by the user or Automatic Speech Recognition (ASR) output obtained from the user speech, and wherein the server manages a second NLU model, wherein the first NLU model and the second NLU model are different, the method comprising:
- determining NLU results of an input dialog at the client device from applying the first NLU model;
- determining a confidence score for the NLU results;
- for the confidence score of the NLU results obtained from applying the first NLU model not meeting a threshold, having the client device access the server to execute an NLU process by using the second NLU model on the input dialog;
- and
- for the confidence score of the NLU results obtained from applying the first NLU model meeting the threshold, having the client device execute actions based on the NLU results obtained from applying the first NLU model;
- wherein the first NLU model and the second NLU model are generated from a training algorithm configured to conduct a sequence labeling method utilizing machine learning;
- wherein the first NLU model and the second NLU model are trained by using training data comprising sentences and associated dialog acts for at least one of each word in the sentences and each sentence.
A method for a dialog system comprising a client device configured with a text input method to receive user input, and a first natural language understanding (NLU) model configured to communicate with a server, the user input involving at least one of text input typed by the user or Automatic Speech Recognition (ASR) output obtained from the user speech, and wherein the server manages a second NLU model, wherein the first NLU model and the second NLU model are different, the method comprising:
- determining NLU results of an input dialog at the client device from applying the first NLU model;
- determining a confidence score for the NLU results;
- for the confidence score of the NLU results obtained from applying the first NLU model not meeting a threshold, having the client device access the server to execute an NLU process by using the second NLU model on the input dialog;
- and
- for the confidence score of the NLU results obtained from applying the first NLU model meeting the threshold, having the client device execute actions based on the NLU results obtained from applying the first NLU model;
- wherein the first NLU model and the second NLU model are generated from a Finite State Transducer (FST)-based algorithm where phrases in the input dialog are mapped on a FST and arcs in the FST are tied with a corresponding dialog act and a weight based on dialog act-chunk pair information;
- wherein the determining NLU results of the input dialog from applying the first NLU model or the second NLU model comprises:
- conducting a shortest path search on the FST;
- and
- outputting dialog acts along one or more paths of the shortest path search for a summation of the weight being below another threshold.
Dependencies are inferred from the claim text.