Semantic parsing using embedding space representations of example natural language queries
AssigneeGoogle Inc.
InventorsWangqing Yuan, Bryan Christopher Horling
Life of this patent
Estimating the term…
Abstract
Techniques disclosed herein are directed towards generating structured data output based on processing a natural language user query using a semantic parser model. Many implementations include identifying one or more argument spans in the given natural language user query based on comparing an embedding space representation of a candidate argument with an embedding space representation of an example query, where the example query is provided by a developer. Various implementations include hotfixing an under-triggering model and/or an over-triggering model based on additional or alternative example queries provided by a developer.
Claims19
A method implemented by one or more processors, the method comprising:
- identifying a natural language user query, where the natural language query includes an intent and one or more arguments;
- processing the natural language user query using a semantic parser model to generate a structured data output based on the natural language user query, where processing the natural language user query using the semantic parser model to generate the structured data output comprises:
- processing the natural language user query using an embedding layer of the semantic parser model to generate a sequence of word embeddings based on the natural language user query;
- processing the sequence of word embeddings using an encoder layer of the semantic parser model to generate, for each word embedding in the sequence of word embeddings, a candidate start embedding and a candidate end embedding;
- processing the intent of the natural language user query using the encoder layer of the semantic parser model to generate an intent embedding;
- processing the candidate start embeddings and the candidate end embeddings to determine a start embedding and an end embedding for each argument;
- and
- generating the structured data output based on the natural language user query, the intent embedding, and the start embedding and the end embedding corresponding to each argument.
A method implemented by one or more processors, the method comprising:
- identifying a natural language user query, where the query includes an intent and one or more arguments corresponding to the intent;
- processing the natural language user query using a semantic parser model to generate structured data output;
- determining, based on processing the natural language user query using the semantic parser model, whether an embedding space representation of the natural language user query corresponds to a portion of an embedding space corresponding to the intent;
- in response to determining the embedding space representation of the natural language user query does not correspond to the portion of the embedding space corresponding to the intent:
- identifying a given portion of the embedding space corresponding to the embedding space representation of the natural language user query;
- and
- altering the portion of the embedding space corresponding to the intent based on the given portion of the embedding space corresponding to the embedding space representation of the natural language user query.
A method implemented by one or more processors, the method comprising:
- identifying a natural language user query, where the query includes an intent and one or more arguments corresponding to the intent;
- processing the natural language user query using a semantic parser model to generate structured data output;
- determining, based on processing the natural language user query using the semantic parser model, whether an embedding space representation of the natural language user query corresponds to a portion of the embedding space corresponding to an additional intent;
- in response to determining the embedding space representation of the natural language user query corresponds to the portion of the embedding space corresponding to the additional intent:
- identifying a given portion of the embedding space corresponding to the embedding space representation of the natural language user query;
- and
- altering the portion of embedding space corresponding to the additional intent based on the given portion of the embedding space corresponding to the embedding space representation of the natural language user query.
Dependencies are inferred from the claim text.