Model robustness on operators and triggering keywords in natural language to a meaning representation language system
Techniques are disclosed herein for improving model robustness on operators and triggering keywords in natural language to a meaning representation language system. The techniques include augmenting an original set of training data for a target robustness bucket by leveraging a combination of two training data generation techniques: (1) modification of existing training examples and (2) synthetic template-based example generation. The resulting set of augmented data examples from the two training data generation techniques are appended to the original set of training data to generate an augmented training data set and the augmented training data set is used to train a machine learning model to generate logical forms for utterances.
1 . A computer-implemented method comprising:
accessing training data comprising a plurality of training examples comprising a first training example, wherein the first training example comprises a first natural language utterance and a first logical form for the first natural language utterance, and wherein the first natural language utterance comprises one or more keywords associated with a target robustness bucket;
replacing, in the first natural language utterance, the one or more keywords with one or more replacement terms sampled from a list of replacement terms to generate a second natural language utterance;
generating a second training example comprising the second natural language utterance and the first logical form;
accessing a data manufacturing template that defines structure of a system operation and constraints imposed on operators of the system operation;
generating a second logical form by filling slots for one or more of the operators in the data manufacturing template based on a schema and values for a system, wherein the one or more of the operators are associated with the target robustness bucket;
translating the second logical form into a third natural language utterance based on a grammar data structure that includes a custom grammar and a set of rules for translating logical form statements into corresponding natural language expressions comprising one or more target keywords, wherein the custom grammar comprises target keywords including the replacement terms from the list of replacement terms, which are used to control the translating such that natural language utterances will contain one or more of the target keywords;
generating a third training example comprising the second logical form and the third natural language utterance;
augmenting the training data by adding the second training example and the third training example to the plurality of training examples to generate an augmented training data set; and
training a machine learning model to generate logical forms for utterances using the augmented training data set.
2 . The computer-implemented method of claim 1 , wherein accessing the training data comprises identifying, from the plurality of training examples, a set of target training examples that match patterns comprising the one or more keywords or the one or more of the operators associated with the target robustness bucket, wherein the set of target training examples comprises the first training example.
3 . The computer-implemented method of claim 1 , wherein the first training example comprises the first natural language utterance, the first logical form for the first natural language utterance, and metadata associated with the first natural language utterance, wherein generating the second training example comprises modifying the metadata associated with the first natural language utterance to generate updated metadata associated with the second natural language utterance, and wherein modifying the metadata comprises adjusting an offset value for schema linking to reflect values in the second natural language utterance.
4 . The computer-implemented method of claim 1 , wherein the data manufacturing template comprises a delexicalized version of the second logical form and filling the slots comprises: parsing the data manufacturing template, sampling the system and schema to identify a plurality of sampling components comprising the values, and lexicalizing the slots of the data manufacturing template with at least one sampling component of the plurality of sampling components.
5 . The computer-implemented method of claim 4 , wherein the data manufacturing template is parsed by applying a synchronous context-free grammar to the delexicalized version of the second logical form to generate an abstract syntax tree for the data manufacturing template, and lexicalizing the data manufacturing template comprises analyzing the data manufacturing template to identify the constraints imposed on the operators and lexicalizing the slots of the data manufacturing template with at least one sampling component of the plurality of sampling components based on the identified constraints.
6 . The computer-implemented method of claim 1 , wherein translating the second logical form into a third natural language utterance comprises:
converting the second logical form into a logical form structure comprising one or more logical form statements, wherein the converting comprises extracting one or more operations and associated one or more attributes from the second logical form and generating the one or more logical form statements from the one or more operations and associated one or more attributes;
translating each of the one or more logical form statements into a natural language expression based on the grammar data structure;
combining the one or more natural language expressions into the third natural language utterance; and
providing the third natural language utterance as an interpretation of the second logical form.
7 . The computer-implemented method of claim 1 , wherein training the machine learning model using the augmented training data set comprises:
sampling training values from the augmented training data set based on a sampling rate; and
training the machine learning model using the sampled training values.
8 . A system comprising:
one or more processors; and
one or more computer-readable media storing instructions which, when executed by the one or more processors, cause the system to perform operations comprising:
accessing training data comprising a plurality of training examples comprising a first training example, wherein the first training example comprises a first natural language utterance and a first logical form for the first natural language utterance, and wherein the first natural language utterance comprises one or more keywords associated with a target robustness bucket;
replacing, in the first natural language utterance, the one or more keywords with one or more replacement terms sampled from a list of replacement terms to generate a second natural language utterance;
generating a second training example comprising the second natural language utterance and the first logical form;
accessing a data manufacturing template that defines structure of a system operation and constraints imposed on operators of the system operation;
generating a second logical form by filling slots for one or more of the operators in the data manufacturing template based on a schema and values for a system, wherein the one or more of the operators are associated with the target robustness bucket;
translating the second logical form into a third natural language utterance based on a grammar data structure that includes a custom grammar and a set of rules for translating logical form statements into corresponding natural language expressions comprising one or more target keywords, wherein the custom grammar comprises target keywords including the replacement terms from the list of replacement terms, which are used to control the translating such that natural language utterances will contain one or more of the target keywords;
generating a third training example comprising the second logical form and the third natural language utterance;
augmenting the training data by adding the second training example and the third training example to the plurality of training examples to generate an augmented training data set; and
training a machine learning model to generate logical forms for utterances using the augmented training data set.
9 . The system of claim 8 , wherein accessing the training data comprises identifying, from the plurality of training examples, a set of target training examples that match patterns comprises the one or more keywords or the one or more of the operators associated with the target robustness bucket, wherein the set of target training examples comprise the first training example.
10 . The system of claim 8 , wherein the first training example comprises the first natural language utterance, the first logical form for the first natural language utterance, and metadata associated with the first natural language utterance, wherein generating the second training example comprises modifying the metadata associated with the first natural language utterance to generate updated metadata associated with the second natural language utterance, and wherein modifying the metadata comprises adjusting an offset value for schema linking to reflect values in the second natural language utterance.
11 . The system of claim 8 , wherein the data manufacturing template comprises a delexicalized version of the second logical form and filling the slots comprises: parsing the data manufacturing template, sampling the system and schema to identify a plurality of sampling components comprising the values, and lexicalizing the slots of the data manufacturing template with at least one sampling component of the plurality of sampling components.
12 . The system of claim 11 , wherein the data manufacturing template is parsed by applying a synchronous context-free grammar to the delexicalized version of the second logical form to generate an abstract syntax tree for the data manufacturing template, and lexicalizing the data manufacturing template comprises analyzing the data manufacturing template to identify the constraints imposed on the operators and lexicalizing the slots of the data manufacturing template with at least one sampling component of the plurality of sampling components based on the identified constraints.
13 . The system of claim 8 , wherein translating the second logical form into a third natural language utterance comprises:
converting the second logical form into a logical form structure comprising one or more logical form statements, wherein the converting comprises extracting one or more operations and associated one or more attributes from the second logical form and generating the one or more logical form statements from the one or more operations and associated one or more attributes;
translating each of the one or more logical form statements into a natural language expression based on the grammar data structure;
combining the one or more natural language expressions into the third natural language utterance; and
providing the third natural language utterance as an interpretation of the second logical form.
14 . The system of claim 8 , wherein training the machine learning model using the augmented training data set comprises:
sampling training values from the augmented training data set based on a sampling rate; and
training the machine learning model using the sampled training values.
15 . One or more non-transitory computer-readable media storing instructions which, when executed by one or more processors, cause a system to perform operations comprising:
accessing training data comprising a plurality of training examples comprising a first training example, wherein the first training example comprises a first natural language utterance and a first logical form for the first natural language utterance, and wherein the first natural language utterance comprises one or more keywords associated with a target robustness bucket;
replacing, in the first natural language utterance, the one or more keywords with one or more replacement terms sampled from a list of replacement terms to generate a second natural language utterance;
generating a second training example comprising the second natural language utterance and the first logical form;
accessing a data manufacturing template that defines structure of a system operation and constraints imposed on operators of the system operation;
generating a second logical form by filling slots for one or more of the operators in the data manufacturing template based on a schema and values for a system, wherein the one or more of the operators are associated with the target robustness bucket;
translating the second logical form into a third natural language utterance based on a grammar data structure that includes a custom grammar and a set of rules for translating logical form statements into corresponding natural language expressions comprising one or more target keywords, wherein the custom grammar comprises target keywords including the replacement terms from the list of replacement terms, which are used to control the translating such that natural language utterances will contain one or more of the target keywords;
generating a third training example comprising the second logical form and the third natural language utterance;
augmenting the training data by adding the second training example and the third training example to the plurality of training examples to generate an augmented training data set; and
training a machine learning model to generate logical forms for utterances using the augmented training data set.
16 . The one or more non-transitory computer-readable media of claim 15 , wherein accessing the training data comprises identifying, from the plurality of training examples, a set of target training examples that match patterns comprising the one or more keywords or the one or more of the operators associated with the target robustness bucket, wherein the set of target training examples comprises the first training example.
17 . The one or more non-transitory computer-readable media of claim 15 , wherein the first training example comprises the first natural language utterance, the first logical form for the first natural language utterance, and metadata associated with the first natural language utterance, wherein generating the second training example comprises modifying the metadata associated with the first natural language utterance to generate updated metadata associated with the second natural language utterance, and wherein modifying the metadata comprises adjusting an offset value for schema linking to reflect values in the second natural language utterance.
18 . The one or more non-transitory computer-readable media of claim 15 , wherein the data manufacturing template comprises a delexicalized version of the second logical form and filling the slots comprises: parsing the data manufacturing template, sampling the system and schema to identify a plurality of sampling components comprising the values, and lexicalizing the slots of the data manufacturing template with at least one sampling component of the plurality of sampling components.
19 . The one or more non-transitory computer-readable media of claim 18 , wherein the data manufacturing template is parsed by applying a synchronous context-free grammar to the delexicalized version of the second logical form to generate an abstract syntax tree for the data manufacturing template, and lexicalizing the data manufacturing template comprises analyzing the data manufacturing template to identify the constraints imposed on the operators and lexicalizing the slots of the data manufacturing template with at least one sampling component of the plurality of sampling components based on the identified constraints.
20 . The one or more non-transitory computer-readable media of claim 15 , wherein translating the second logical form into a third natural language utterance comprises:
converting the second logical form into a logical form structure comprising one or more logical form statements, wherein the converting comprises extracting one or more operations and associated one or more attributes from the second logical form and generating the one or more logical form statements from the one or more operations and associated one or more attributes;
translating each of the one or more logical form statements into a natural language expression based on the grammar data structure;
combining the one or more natural language expressions into the third natural language utterance; and
providing the third natural language utterance as an interpretation of the second logical form.