Identification of candidate training utterances from human conversations with an intelligent interactive assistant

A method for creating a binary classification model, the method comprising:

• receiving a plurality of intents;
• for each intent, receiving a plurality of associated training utterances;
• combining all of the received training utterances into a training utterance collection;
• generating all n-grams included in all of the training utterances;
• assigning each n-gram a unique numeric identifier;
• for a first intent, included in the plurality of intents:
• assigning an entry for each of the training utterances included in the training utterance collection, wherein:
• each entry comprises the unique identifiers of each n-gram included in the training utterances;
• each unique identifier, included in each entry, is accompanied by a number of times the n-gram appears in the training utterance;
• and
• each entry is assigned either a “one” notation or a “zero” notation, the “one” notation indicating that the training utterance is associated with the first intent, the “zero” notation indicating that the training utterance is disassociated from the first intent;
• converting, by a support vector machine (“SVM”), each entry into a vector representation;
• separating, by the SVM, the vector representations into two groups, a first group and a second group, the first group being identified by the vector representations that are assigned the “one” notation, the second group being identified by the vector representations that are assigned to the “zero” notation;
• defining, by the SVM, a vector representation of a line of demarcation between the vector representations of the first group and the vector representations of the second group;
• creating a binary classification model for the first intent, said binary classification model comprising:
• the first group of vector representations;
• the second group of vector representations;
• and
• the vector representation of the line of demarcation.