Intelligent training set augmentation for natural language processing tasks

A non-transitory computer-readable storage medium storing instructions for selectively augmenting utterances of a training set for a natural language task, the instructions when executed by a computer processor performing actions comprising:

• for a natural language utterance of the training set, using, by one or more processors, a task model for a natural language task to determine probabilities of possible task outputs;
• computing, by the one or more processors, importances of textual units of the utterance to the task model when determining the probabilities;
• identifying, by the one or more processors, a most important textual unit of the utterance based on the computed importances;
• accessing, by the one or more processors, task rules stored in system memory and corresponding to the natural language task;
• determining, by the one or more processors, using the task rules, whether to generate a variant of the utterance based on the identified textual unit;
• responsive to the determining, generating, by the one or more processors, a variant of the utterance based on the identified textual unit;
• generating, by the one or more computer processors, an augmented training set by adding the variant of the utterance to the training set;
• and
• re-training, by the one or more processors, the task model using the augmented training set and the variant of the utterance, the re-training performed using at least one of:
• neural networks, support vector machines (SVM), linear regression, or Naïve Bayes.

A computer system comprising:

• one or more computer processors;
• and
• non-transitory computer-readable storage medium storing instructions for selectively augmenting utterances of a training set for a natural language task, the instructions when executed by the one or more computer processor devices performing actions comprising:
• for a natural language utterance of the training set, using, by the one or more computer processors, a task model for a natural language task to determine probabilities of possible task outputs;
• computing, by the one or more computer processors, importances of textual units of the utterance to the task model when determining the probabilities;
• identifying, by the one or more computer processors, a most important textual unit of the utterance based on the computed importances;
• accessing, by the one or more computer processors, task rules stored in system memory and corresponding to the natural language task;
• determining, by the one or more computer processors, using the task rules, whether to generate a variant of the utterance based on the identified textual unit;
• responsive to the determining, generating, by the one or more computer processors, a variant of the utterance based on the identified textual unit;
• generating, by the one or more computer processors, an augmented training set by adding the variant of the utterance to the training set;
• and
• re-training, by the one or more processors, the task model using the augmented training set and the variant of the utterance, the re-training performed using at least one of:
• neural networks, support vector machines (SVM), linear regression, or Naïve Bayes.