CROWD SOURCED BASED TRAINING FOR NATURAL LANGUAGE INTERFACE SYSTEMS

• US 20150301795A1
• Filed: 04/14/2015
• Published: 10/22/2015
• Est. Priority Date: 04/16/2014
• Status: Abandoned Application

• Current Assignee: Meta Platforms, Inc. (f/k/a Facebook, Inc.)
• Original Assignee: Meta Platforms, Inc. (f/k/a Facebook, Inc.)
• Inventors: Lebrun, Alexandre
• Application Number:
  US14/686,771
• Publication Number:
  US 20150301795A1
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G06F 40/35 Discourse or dialogue representation
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/18 using natural language modelling

Abstract

A crowdsourcing based community platform includes a natural language configuration system that predicts a user'"'"'s desired function call based on a natural language input (speech or text). The system provides a collaboration platform to configure and optimize quickly natural language systems to leverage the work and data of other developers, thus minimizing the time and data required to improve the quality and accuracy of one single system and providing a network effect to reach quickly critical mass of data. An application developer can provide training data for training a model specific to the developer'"'"'s application. The developer can also obtain training data by forking one or more other applications so that the training data provided for the forked applications is used to train the model for the developer'"'"'s application.

Claims

1. A computer-implemented method comprising:

maintaining a plurality of natural language (NL) instances, each NL instance comprising a set of training data and a prediction model that predicts a user-desired application function based on a NL query;

receiving one or more linking requests to link NL instances to other NL instances of the plurality of NL instances;

linking one or more NL instances of the plurality of NL instances to one or more other NL instances of the plurality of NL instances responsive to the received linking requests; and

training each of the NL instances of the plurality of NL instances based at least in part on;

the training data for the NL instance and the training data for any other NL instances to which the NL instance is linked;

for each of one or more of the plurality of NL instances;

receiving an application request for the NL instance to provide a user-desired application function based on a NL query,

using the prediction model for the NL instance to determine an application function according to the prediction model, and

responding to the application request with the determined application function.

2. The method of claim 1, wherein content indicators are associated with the one or more NL instances.

3. The method of claim 1 further comprising:

ranking each of the linked NL instances of the plurality of NL instances according to the number of links associated with each linked NL instance,

wherein an NL instance that is not linked to any other NL instance is ranked at zero.

4. The method of claim 3, wherein the receiving of the one linking requests to link NL instances is at least in part based on the rank of each NL instance of the plurality of NL instances.

5. The method of claim 1 further comprising:

displaying a list of NL instances of the plurality of NL instance for inclusion in the one or more linking requests.

6. The method of claim 5, wherein the displaying of the list of NL instances is at least in part based on the rank of each NL instance of the plurality of NL instances.

7. The method of claim 1 further comprising:

searching the plurality of NL instances based on user-specified search criteria to determine a set of NL instance for inclusion in the one or more linking requests.

8. The method of claim 7, wherein the user-specified search criteria comprises a similarity measure to a local NL instance of the plurality of NL instances.

9. A computer program product comprising a non-transitory computer readable storage medium for version control for application development, the non-transitory computer readable storage medium storing instructions for:

maintaining a plurality of natural language (NL) instances, each NL instance comprising a set of training data and a prediction model that predicts a user-desired application function based on a natural language input;

receiving one or more requests to link NL instances to other NL instances of the plurality of NL instances;

linking one or more NL instances of the plurality of NL instances to one or more other NL instances of the plurality of NL instances responsive to the received requests; and

training each of the NL instances of the plurality of NL instances based at least in part on;

the training data for the NL instance and the training data for any other NL instances to which the NL instance is linked;

for each of one or more of the plurality of NL instances;

receiving a request for the NL instance to provide a user-desired application function based on a natural language input,

using the prediction model for the NL instance to determine an application function according to the prediction model, and

responding to the request with the determined application function.

10. The computer program product of claim 9 further comprising:

ranking each of the linked NL instances of the plurality of NL instances according to the number of links associated with each linked NL instance,

wherein an NL instance that is not linked to any other NL instance is ranked at zero.

11. The computer program product of claim 10, wherein the receiving of the one linking requests to link NL instances is at least in part based on the rank of each NL instance of the plurality of NL instances.

12. The computer program product of claim 9 further comprising:

displaying a list of NL instances of the plurality of NL instance for inclusion in the one or more linking requests.

13. The computer program product of claim 12, wherein the displaying of the list of NL instances is at least in part based on the rank of each NL instance of the plurality of NL instances.

14. The computer program product of claim 9 further comprising:

searching the plurality of NL instances based on user-specified search criteria to determine a set of NL instance for inclusion in the one or more linking requests.

15. The method of claim 14, wherein the user-specified search criteria comprises a similarity measure to a local NL instance of the plurality of NL instances.