Systems and methods for automated degradation-resistant tuning of machine-learning language processing models

A machine-learning (“ML”)-based system with automated tuning, said system comprising:

• an ML model that is configured to derive a comprehension of an input utterance, wherein:
• said ML model is trained using a training set;
• the training set is stored in a database and comprises a plurality of utterances mapped to a plurality of intents;
• and
• each of the plurality of utterances is assigned a set of features;
• and
• a tuning module, said tuning module for improving the comprehension of the ML model while resisting degradation in the comprehension, said tuning module configured to:
• flag an inoperative input utterance;
• retrieve a feature set associated with the inoperative input utterance;
• generate, based on the feature set, an utterance-feature-intent (“UFI”) mapping between the inoperative input utterance, the feature set, and an intent;
• update the ML model with the UFI mapping;
• compare a performance of the ML model before and after the update;
• and
• when the performance vis-à-vis the inoperative input utterance improves, and the performance vis-à-vis other input utterances is maintained or improves, deploy the update.

A machine-learning (“ML”)-based language-processing system with automated tuning, said system comprising:

• an ML model that is configured to derive a comprehension of an input utterance, wherein:
• said ML model is trained using a training set;
• the training set is stored in a database and comprises a plurality of utterances mapped to a plurality of intents;
• and
• each of the plurality of utterances is assigned a set of features;
• and
• a tuning module, said tuning module for improving the comprehension of the ML model while resisting degradation in the comprehension, said tuning module configured to:
• identify a set of inoperative input utterances and a set of operative input utterances;
• derive a first accuracy score for the set of inoperative input utterances and for the set of operative input utterances;
• retrieve a feature set associated with each inoperative input utterance from the set of inoperative input utterances;
• generate, based on the feature set, an updated utterance-feature-intent (“UFI”) mapping;
• retrain the ML model, in a controlled environment, using the updated UFI mapping;
• generate an updated accuracy score for the set of inoperative input utterances and for the set of operative input utterances, said updated accuracy score based on the retrained ML model;
• when the updated accuracy score for the set of inoperative input utterances is less than or equal to the first accuracy score for the set of inoperative input utterances, and the updated accuracy score for the set of operative input utterances is greater than or equal to the first accuracy score for the set of operative input utterances, amplify the impact of the updated UFI mapping by retraining the ML model with a plurality of instances of the updated UFI mapping;
• and
• when the updated accuracy score for the set of inoperative input utterances is greater than the first accuracy score for the set of inoperative input utterances, and the updated accuracy score for the set of operative input utterances is greater than or equal to the first accuracy score for the set of operative input utterances, deploy the updated UFI mapping.