obtaining one or more historical conversation logs comprising a plurality of agent responses matched to a plurality of conversation contexts and a set of designated responses that are not matched to the plurality of conversation contexts;

replacing at least a portion of the plurality of agent responses with the designated responses in the set, thereby forming one or more modified conversation logs;

training (i) a first machine learning model, using the one or more modified conversation logs, to output respective ones of the designated responses in the set for corresponding text conversation contexts selected from a first portion of the plurality of conversation contexts and (ii) a second machine learning model, using the one or more historical conversation logs, to output respective ones of the agent responses for corresponding ones of the text conversation contexts;

determining that the trained first machine learning model does not output one of the designated responses in the set for at least one conversation context selected from a second portion of the plurality conversation contexts, wherein the second portion is different from the first portion;

identifying one or more new responses based at least in part on the output of the second machine learning model for the at least one conversation context selected from the second portion of the plurality of conversation contexts, wherein the identifying comprises selecting the one or more new responses from a plurality of clusters based on a similarity between the output of the second machine learning model for the at least one conversation context and the plurality of clusters; and

retraining the first machine learning model based at least in part on the one or more new responses;

wherein the method is carried out by at least one computing device.