Intent discovery with a prototype classifier

A system for updating intents of a communications system, the system comprising:

• at least one server computer comprising at least one processor and at least one memory, the at least one server computer configured to:
• deploy the communications system, comprising a neural network and a first plurality of prototype vectors for communicating with users, wherein:
• each prototype vector corresponds an intent of a first plurality of intents, and
• the communications system is configured to (i) process a message with the neural network to compute a message embedding that represents a meaning of the message and (ii) determine an intent of the message by comparing the message embedding with one or more of the first plurality of prototype vectors;
• collect usage data from the deployed communications system, wherein the usage data comprises a plurality of message embeddings computed by processing messages from users with the neural network;
• cluster the plurality of message embeddings into a plurality of clusters;
• determine to create a new intent using a first cluster of the plurality of clusters;
• compute a new prototype vector for the new intent by combining message embeddings of the first cluster;
• and
• deploy an updated communications system using the neural network and a second plurality of prototype vectors, wherein the second plurality of prototype vectors comprises at least a majority of the first plurality of prototype vectors and the new prototype vector.

One or more non-transitory, computer-readable media comprising computer-executable instructions that, when executed, cause at least one processor to perform actions comprising:

• deploying a communications system, comprising a neural network and a first plurality of prototype vectors for communicating with users, wherein:
• each prototype vector corresponds an intent of a first plurality of intents, and
• the communications system is configured to (i) process a message with the neural network to compute a message embedding that represents a meaning of the message and (ii) determine an intent of the message by comparing the message embedding with one or more of the first plurality of prototype vectors;
• collecting usage data from the deployed communications system, wherein the usage data comprises a plurality of message embeddings computed by processing messages from users with the neural network;
• clustering the plurality of message embeddings into a plurality of clusters;
• determining to create a new intent using a first cluster of the plurality of clusters;
• computing a new prototype vector for the new intent by combining message embeddings of the first cluster;
• and
• deploying an updated communications system using the neural network and a second plurality of prototype vectors, wherein the second plurality of prototype vectors comprises at least a majority of the first plurality of prototype vectors and the new prototype vector.