Intent detection via masked models

A method of predicting an intent associated with a support ticket, comprising:

• processing, with a masked intent model, the support ticket, wherein:
• the masked intent model comprises a modified logit function and a classification head, and
• processing the support ticket with the masked intent model comprises:
• masking a subset of intents from a set of predictable intents by setting, based on the modified logit function, a probability of at least one intent of the subset of intents to a probability value such that the at least one intent is never predicted, wherein the at least one intent comprises an invalid intent for an organization;
• providing, as input to the classification head, an encoding of the support ticket and a second subset of intents of the set of predictable intents, the second subset of intents comprising one or more unmasked intents from the set of predictable intents;
• processing the encoding of the support ticket and the second subset of intents with at least one of an input mapper, configured to transform the encoding of the support ticket and the second subset of intents into the same dimensionalities, or a residual layer of the masked intent model;
• performing a dropout that drops at least one element of the residual layer to prevent the masked intent model from being overfitted with a portion of the set of predictable intents;
• and
• obtaining, as output from the classification head, at least one unmasked intent from the second subset of intents;
• and
• providing, as output from the masked intent model, one or more predicted intents associated with the support ticket,
• wherein the one or more predicted intents comprise the at least one unmasked intent from the second subset of intents.