Natural language query disambiguation
A method, computer program product, and system are provided. A first natural language text is received via a user interface. A generative pretrained transformer machine learning model processes the first natural language text and a context to identify a first intent. The processing is based in part on a syntax determined from a set of natural language completion paradigms. The generative transformer machine learning model maps the first set of parameters to a first query. The mapping is associated with a first confidence. The generative transformer machine learning model processes the first set of parameters and the first query to generate a set of execution steps. The processing is performed when the first confidence satisfies a threshold. The set of execution steps is parsed into a query object that is forwarded to a reporting service.
1. A method comprising:
iteratively until a confidence satisfies a threshold:
receiving a first natural language text via a user interface,
processing, by a generative pretrained transformer machine learning model, the first natural language text and a context to identify a first intent, wherein:
the generative pretrained transformer machine learning model determines a latent space representation of both encoder and decoder sides of the generative pretrained transformer machine learning model,
each layer of the encoder and decoder sides of the generative pretrained transformer machine learning model comprises a plurality of sublayers,
the plurality of sublayers each comprise a multi-head self-attention layer and a position wise feed forward network layer that uses a fully connected network to process the multi-head self-attention layer, and
processing by the generative pretrained transformer machine learning model comprises: executing the encoder and decoder sides of the generative pretrained transformer machine learning model on the first natural language text and the context, to insert a set of learnable prompt embeddings that are generated using a series of example prompts and corresponding completions, and to identify the first intent,
mapping, by a generative transformer machine learning model, the first intent to a first query associated with the confidence,
processing the first query to generate a clarification question in natural language text when the confidence fails to satisfy the threshold, and
populating the user interface with the clarification question when the confidence fails to satisfy the threshold wherein prompt and completion pairs are encoded in JAVASCRIPT® object notation language (JSON) to fine-tune the generative pretrained transformer machine learning model;
processing, by the generative transformer machine learning model, the first intent and the first query to generate a series of structured query language (SQL) type statements, wherein the processing is performed when the confidence satisfies the threshold;
parsing the series of SQL type statements into a query object comprising a JSON payload comprising a series of execution steps;
forwarding the query object to a reporting service;
mapping the execution steps to a reporting application programming interface (API) of the reporting service to generate a mapping; and
generating a report using the mapping.
2. The method of claim 1 , wherein parsing the series of SQL type statements comprises:
generating a JSON object according to an API schema, wherein the JSON object is based the series of SQL type statements; and
forwarding the JSON object to the reporting service.
3. The method of claim 1 , further comprising:
forwarding the report to the user interface.
4. The method of claim 3 , further comprising:
processing, by the generative pretrained transformer machine learning model, the report to generate a fourth natural language text; and
forwarding the fourth natural language text to the user interface.
5. A computer program product comprising non-transitory computer-readable program code that, when executed by a computer processor of a computing system, causes the computing system to perform operations comprising:
iteratively until a confidence satisfies a threshold:
receiving a first natural language text via a user interface,
processing, by a generative pretrained transformer machine learning model, the first natural language text and a context to identify a first intent, wherein:
the generative pretrained transformer machine learning model determines a latent space representation of both encoder and decoder sides of the generative pretrained transformer machine learning model,
each layer of the encoder and decoder sides of the generative pretrained transformer machine learning model comprises a plurality of sublayers,
the plurality of sublayers each comprise a multi-head self-attention layer and a position wise feed forward network layer that uses a fully connected network to process the multi-head self-attention layer, and
processing by the generative pretrained transformer machine learning model comprises: executing the encoder and decoder sides of the generative pretrained transformer machine learning model on the first natural language text and the context, to insert a set of learnable prompt embeddings that are generated using a series of example prompts and corresponding completions, and to identify the first intent,
mapping, by a generative transformer machine learning model, the first intent to a first query, wherein the mapping is associated with a first confidence,
processing the first query to generate a clarification question in natural language text when the confidence fails to satisfy the threshold, and
populating the user interface with the clarification question when the confidence fails to satisfy the threshold wherein prompt and completion pairs are encoded in JAVASCRIPT® object notation language (JSON) to fine-tune the generative pretrained transformer machine learning model;
processing, by the generative transformer machine learning model, the first intent and the first query to generate a series of structured query language (SQL) type statements, wherein the processing is performed when the confidence satisfies the threshold;
parsing the series of SQL type statements into a query object comprising a JSON payload comprising a series of execution steps;
forwarding the query object to a reporting service;
mapping the execution steps to a reporting application programming interface (API) of the reporting service to generate a mapping; and
generating a report using the mapping.
6. The computer program product of claim 5 , wherein parsing the series of SQL type statements comprises:
generating a JSON object according to an API schema, wherein the JSON object is based on the series of execution steps; and
forwarding the JSON object to the reporting service.
7. The computer program product of claim 5 , the operations further comprising:
forwarding the report to the user interface.
8. The computer program product of claim 7 , the operations further comprising:
processing, by the generative pretrained transformer machine learning model, the report to generate a fourth natural language text; and
forwarding the fourth natural language text to the user interface.
9. A system comprising:
a computer processor;
memory; and
instructions stored in the memory and executable by the computer processor to cause the computer processor to perform operations, the operations comprising:
iteratively until a confidence satisfies a threshold:
receiving a first natural language text via a user interface,
processing, by a generative pretrained transformer machine learning model, the first natural language text and a context to identify a first intent, wherein:
the generative pretrained transformer machine learning model determines a latent space representation of both encoder and decoder sides of the generative pretrained transformer machine learning model,
each layer of the encoder and decoder sides of the generative pretrained transformer machine learning model comprises a plurality of sublayers,
the plurality of sublayers each comprise a multi-head self-attention layer and a position wise feed forward network layer that uses a fully connected network to process the multi-head self-attention layer, and
processing by the generative pretrained transformer machine learning model comprises: executing the encoder and decoder sides of the generative pretrained transformer machine learning model on the first natural language text and the context, to insert a set of learnable prompt embeddings that are generated using a series of example prompts and corresponding completions, and to identify the first intent,
mapping, by a generative transformer machine learning model, the first intent to a first query associated with the confidence,
processing the first query to generate a clarification question in natural language text when the confidence fails to satisfy the threshold, and
populating the user interface with the clarification question when the confidence fails to satisfy the threshold wherein prompt and completion pairs are encoded in JAVASCRIPT® object notation language (JSON) to fine-tune the generative pretrained transformer machine learning model,
processing, by the generative transformer machine learning model, the first intent and the first query to generate a series of structured query language (SQL) type statements, wherein the processing is performed when the confidence satisfies the threshold,
parsing the series of SQL type statements into a query object comprising a JSON payload comprising a series of execution steps;
forwarding the query object to a reporting service,
mapping the execution steps to a reporting application programming interface (API) of the reporting service to generate a mapping, and
generating a report using the mapping.
10. The system of claim 9 , wherein parsing the series of SQL type statements comprises:
generating a JSON object according to an API schema, wherein the JSON object is based the series of execution steps; and
forwarding the JSON object to the reporting service.
11. The system of claim 9 , further comprising:
forwarding the report to the user interface.
12. The system of claim 11 , further comprising:
processing, by the generative pretrained transformer machine learning model, the report to generate a new natural language text; and
forwarding the new natural language text to the user interface.