US 12,608,181 B2Granted

Generation of synthetic training data using grammar mapping

AssigneeMicrosoft Technology Licensing, LLC

InventorsKonstantin Andreyevich Golobokov, Zeqi Lin, Haizhen Zhang, Yu Hu, Yousef Ahmed Al-Kofahi, Jonathan Richard Malsan, Haiyuan Cao, Daniel Akintola Fatade

Life of this patent

Estimating the term…

patentgrep

Loading patent…

Abstract

The automatic generation of synthetic training data that can be used to train a language model to generate code examples following a code language based on a natural language input. Thus, new language models may be created, or existing language models may be fine-tuned, to adapt to automatically generate code without having to manually generate bulk quantities of training data. Rather, a many-to-many grammar mapping is navigated to generate training data. Specifically, the many-to-many grammar mapping maps code grammar to natural grammar. Then, each training data is generated by navigating the many-to-many grammar mapping definition to generate a mapping of a respective code expression to a respective natural language expression.

Claims20

Dependent claims

A computing system that generates synthetic training data, said computing system comprising:

one or more processors;
and
one or more hardware storage devices that store instructions that are executable by the one or more processors to cause the computing system to:
access a many-to-many grammar mapping definition for mapping code grammar and natural grammar, the code grammar being associated with a code language and defining how to form code strings from an alphabet of the code language in a way that the code strings are valid according to a syntax of the code language, the natural grammar being associated with a natural language and defining how to form natural language strings from an alphabet of the natural language in a way that the natural language strings at least approximate a syntax of the natural language;
generate a plurality of mappings between code expressions and natural language expressions by navigating the many-to-many grammar mapping definition to respectively generate a respective mapping of a respective code expression to a respective natural language expression, wherein the navigation includes a random sampling of the many-to-many grammar mappings definition such that each mapping in the plurality of mappings is different than at least most of the other mappings in the plurality of mappings, and wherein, as a result of the random sampling of the many-to-many grammar mappings definition, a different navigation results in a different intermediate form of an expression mapping, which is associated with the plurality of mappings;
select a first pair from among the plurality of mappings, the first pair comprising a first mapping between a first code expression and a first natural language expression;
syntactically validate the first code expression by executing an argument checker for a specific application development tool against the first code expression;
and
in response to the argument checker syntactically validating the first code expression, include the first pair in a plurality of training data.

6 more elements

A method for generating synthetic training data to train a language model to generate code examples following a code language based on a natural language input, the method comprising:

accessing a many-to-many grammar mapping definition for mapping code grammar and natural grammar, the code grammar being associated with a code language and defining how to form code strings from an alphabet of the code language in a way that the code strings are valid according to a syntax of the code language, the natural grammar being associated with a natural language and defining how to form natural language strings from an alphabet of the natural language in a way that the natural language strings at least approximate a syntax of the natural language;
generating a plurality of mappings between code expressions and natural language expressions by navigating the many-to-many grammar mapping definition to respectively generate a respective mapping of a respective code expression to a respective natural language expression, wherein the navigation includes a random sampling of the many-to-many grammar mappings definition such that each mapping in the plurality of mappings is different than at least most of the other mappings in the plurality of mappings, and wherein, as a result of the random sampling of the many-to-many grammar mappings definition, a different navigation results in a different intermediate form of an expression mapping, which is associated with the plurality of mappings;
selecting a first pair from among the plurality of mappings, the first pair comprising a first mapping between a first code expression and a first natural language expression;
syntactically validating the first code expression by executing an argument checker for a specific application development tool against the first code expression;
and
in response to the argument checker syntactically validating the first code expression, including the first pair in a plurality of training data.

3 more elements

Dependencies are inferred from the claim text.