Methods and apparatus related to determining a semantically diverse subset of candidate responses to provide for initial presentation to a user as suggestions for inclusion in a reply to an electroniccommunication. Selection of one of the provided responses by the user will reduce the number of user inputs that a user must make to formulate the reply, which may reduce the usage of various clientdevice computational resources and/or be of particular benefit to users that have low dexterity (or who otherwise have difficulties making user inputs to a client device). Some of the implementationsdetermine the semantically diverse subset of candidate responses based on generating, over a neural network response encoder model, embeddings that are each based on one of the plurality of the candidate responses. The embedding based on a given candidate response may be compared to embedding(s) of candidate response(s) already selected for the subset, and the given candidate response added to thesubset only if the comparing indicates a difference criterion is satisfied.