Systems and methods for translating natural language queries into a constrained domain-specific language

A method of operating performing a natural language query (NLQ) to domain-specific language (DSL) translation, comprising:

• receiving a natural language query (NLQ) from a user;
• initializing a result data structure to store potential translations of the NLQ in a domain-specific language (DSL);
• and
• providing the NLQ and the result data structure as inputs to a language model, wherein the language model comprises a pre-trained neural sequence-to-sequence model, wherein the language model performs encoding of the NLQ and then sequentially determines a next potential token for each of the potential translations of the result data structure over a plurality of auto-regressive decoding steps, and wherein, after each auto-regressive decoding step of the plurality of auto-regressive decoding steps of the language model, the method further comprises:
• providing the result data structure as input to a DSL parser, wherein the result data structure includes a respective confidence score determined by the language model for each of the potential translations during each auto-regressive decoding step,
• discarding, via the DSL parser, a first one or more potential translations of the result data structure based on the respective confidence score of each of the first one or more potential translations,
• performing, via the DSL parser, an evaluation of the potential translations of the result data structure based on DSL rules of the DSL parser, and
• discarding, via the DSL parser, a second one or more of the potential translations of the result data structure based on the evaluation.

A non-transitory, computer-readable medium storing instructions of a natural language query (NLQ) to domain-specific language (DSL) translation system, the instructions executable by a processor of a computing system, and the instructions comprising instructions to:

• receive a natural language query (NLQ) from a user;
• initialize a result data structure to store potential translations of the NLQ in a domain-specific language (DSL);
• and
• provide the NLQ and the result data structure as inputs to a language model, wherein the language model comprises a pre-trained neural sequence-to-sequence model, wherein the language model is configured to perform encoding of the NLQ and then to sequentially determine a next potential token for each of the potential translations of the result data structure over a plurality of auto-regressive decoding steps, and wherein, after each auto-regressive decoding step of the plurality of auto-regressive decoding steps of the language model, the instructions further comprise instructions to:
• provide the result data structure as input to a DSL parser, wherein the result data structure includes a respective confidence score determined by the language model for each of the potential translations during each auto-regressive decoding step,
• discarding, via the DSL parser, a first one or more potential translations of the result data structure based on the respective confidence score of each of the first one or more potential translations,
• perform, via the DSL parser, an evaluation of the potential translations of the result data structure based on DSL rules of the DSL parser, and
• discard, via the DSL parser, a second one or more of the potential translations of the result data structure based on the evaluation.