Dataset generation using large language models

Patent metadata

Field	Value
Publication number	US-2024185001-A1
Application number	US-202218075942-A
Country	US
Kind code	A1
Filing date	Dec 6, 2022
Priority date	Dec 6, 2022
Publication date	Jun 6, 2024
Grant date	—

Abstract

Disclosed are systems and techniques that may generate datasets for training task-oriented dialogue systems. The techniques include generating natural language queries by selecting a template query, sampling one or more tokens from a data store of domain-specific tokens, modifying the selected template query using the one or more sampled tokens to generate a query prompt, and using a natural language generative machine-learning model to generate, based on the query prompt, a respective natural language query of the subset of the plurality of natural language queries, and causing the generated plurality of natural language queries to be provided to a machine-learning model training engine configured to train, using the generated plurality of natural language queries, a conversational machine-learning model to perform a domain-specific conversational task.

First claim

What is claimed is:

1 . A method comprising:
performing one or more conversational tasks in a target domain using a machine learning model (MLM), the MLM trained, at least, by:
generating a plurality of natural language (NL) queries, at least one NL query of the plurality of NL queries generated, at least, by:
selecting a template query from a plurality of template queries;
sampling one or more tokens from a plurality of tokens corresponding to the target domain;
modifying the template query using the one or more tokens to generate a query prompt; and
using a large language model (LLM) to generate, based at least on the query prompt, the at least one NL query of the plurality of NL queries; and

updating one or more parameters of the MLM using the plurality of NL queries.

2 . The method of claim 1 , wherein the template query comprises:
an example query corresponding to the target domain, and an NL response to the example query.

3 . The method of claim 2 , wherein the template query further comprises one or more placeholders, and wherein the modifying the template query comprises replacing the one or more placeholders with the one or more tokens.

4 . The method of claim 2 , wherein the example query corresponds to an example conversation between a customer and a service provider in at least one of:
a financial services domain, a food services domain, a health services domain, or a retail services domain.

5 . The method of claim 1 , wherein the sampling the one or more tokens is performed randomly.

6 . The method of claim 1 , wherein at least one template query of the plurality of template queries is associated with a selection weight, and wherein the template query is selected with a probability determined using the selection weight.

7 . The method of claim 1 , further comprising:
obtaining a condition associated with a sub-task of the conversational task; augmenting the query prompt with the obtained condition to generate an augmented query prompt; and applying the augmented query prompt to the LLM to generate the at least one NL query.

8 . The method of claim 7 , wherein the one or more tokens includes a plurality of tokens, and the obtaining the condition comprises:
selecting a subset of at least one of the plurality of tokens; and combining the subset with a condition template associated with the sub-task, wherein the training the MLM includes using the subset.

9 . The method of claim 1 , further comprising:
associating a quality label with one or more NL queries, the quality label being indicative of a quality of a respective NL query of the one or more NL queries; creating a new template query using one or more template queries of the plurality of template queries, the new template query comprising one or more virtual tokens; using another MLM to learn the one or more virtual tokens of the new template query using the one or more NL queries and associated quality labels; and adding the new template query comprising the one or more learned virtual tokens to the plurality of template queries.

10 . A system comprising:
one or more processing units to:
generate a plurality of natural language (NL) queries, at least, by:
selecting a template query of a plurality of template queries;
sampling one or more tokens from a data store of domain-specific tokens;
modifying the selected template query using the one or more sampled tokens to generate a query prompt; and
using a large language model (LLM) to generate, based at least on the query prompt, a respective NL query of the subset of the plurality of NL queries; and

provided the generated plurality of NL queries to an MLM training engine to cause the MLM training engine to train a MLM to perform a domain-specific conversational task.

11 . The system of claim 10 , wherein the selected template query comprises:
a domain-specific example query, and an NL response to the example query.

12 . The system of claim 11 , wherein the selected template query further comprises one or more placeholders, and wherein to modify the selected template query, the one or more processing units replace the one or more placeholders with the one or more sampled tokens.

13 . The system of claim 11 , wherein the domain-specific example query corresponds to an example conversation between a customer and a service provider in at least one of:
a financial services domain, a food services domain, a health services domain, or a retail services domain.

14 . The system of claim 10 , wherein the sampling the one or more tokens is performed randomly.

15 . The system of claim 10 , wherein one or more individual template queries of the plurality of template queries is associated with a selection weight, and wherein the selected template query is selected with a probability determined using the selection weight.

16 . The system of claim 10 , wherein the one or more processing units are further to:
obtain a condition associated with a sub-task of the domain-specific conversational task; augment the generated query prompt with the obtained condition; and apply the LLM to the augmented query prompt to generate the respective NL query.

17 . The system of claim 16 , wherein, to obtain the condition associated with the sub-task of the domain-specific conversational task, the one or more processing units are to:
select a subset of at least one of the one or more sampled tokens; and combine the selected subset with a condition template associated with the sub-task, wherein the one or more processing units are further to provide the selected subset to the MLM training engine.

18 . The system of claim 10 , wherein the one or more processing units are further to:
associate a quality label with one or more generated NL queries, the quality label being indicative of quality of a respective NL query of the one or more generated NL queries; create a new template query using one or more template queries of the plurality of template queries, the new template query comprising one or more virtual tokens; use a token-generating MLM to learn the one or more virtual tokens of the new template query using the one or more generated NL queries and associated quality labels; and add the new template query comprising the one or more learned virtual tokens to the plurality of template queries.

19 . The system of claim 10 , wherein the system is comprised in at least one of:
an in-vehicle infotainment system for an autonomous or semi-autonomous machine; a system for performing simulation operations; a system for performing digital twin operations; a system for performing light transport simulation; a system for performing collaborative content creation for 3D assets; a system for performing deep learning operations; a system implemented using an edge device; a system for generating or presenting at least one of virtual reality content, mixed reality content, or augmented reality content; a system implemented using a robot; a system for performing conversational AI operations; a system for generating synthetic data; a system incorporating one or more virtual machines (VMs); a system implemented at least partially in a data center; or a system implemented at least partially using cloud computing resources.

20 . A processor comprising:
one or more processing units to perform one or more natural language processing tasks using a machine learning model, the machine learning model trained, at leas

Assignees

• Nvidia Corp

Inventors

• Nagaraju Divija
• Parisien Christopher

Classifications

• G06F40/56 Primary

  Natural language generation · CPC title

• G06F40/284 Primary

  Lexical analysis, e.g. tokenisation or collocates · CPC title