Method and system for automatically determining responses in a messaging platform
In variants, a method for automatically determining responses in a messaging platform can include: receiving a set of inputs; determining a set of response options based on the set of inputs; and providing the set of response options to a user. The method can optionally further include training and/or updating a set of models, and/or updating a corpus of historical responses.
1. A method comprising:
generating a query embedding of a query message from a conversation thread;
generating a set of context embeddings of a set of prior messages in the conversation thread;
with a trained model, determining a ranked list of response options for the query message based on the query embedding and the set of context embeddings;
determining a set of response options from the ranked list of response options, wherein each response option of the set of response options is distinct from each of the other members of the set, wherein determining the set of response options comprises applying a diversity sampling model, wherein applying the diversity sampling model comprises determining a set of semantic clusters from the ranked list of response options, and selecting one of the set of response options from each of the set of semantic clusters;
providing the set of response options, determined from the ranked list of response options, wherein each response option in the set of response options has a predetermined threshold dissimilarity from each of the other members of the set, and wherein the set of response options comprises a guidance response from a guidance response category; and
receiving a selection of one of the set of response options from a user, sending the selection as a message in the conversation thread, and training the trained model based upon the selection.
2. The method of claim 1 , wherein the set of response options comprises a listening response.
3. The method of claim 1 , wherein the set of response options comprises a prescriptive information response from a prescriptive information response category.
4. The method of claim 1 , wherein the set of response options comprises a follow-up question from a follow-up question response category.
5. The method of claim 1 , wherein a response of the set of response options is generated using a generative model.
6. The method of claim 1 , wherein training the trained model comprises training a response embedding model of the trained model used to determine the set of response options.
7. The method of claim 1 , wherein the trained model comprises a natural language embedding model.
8. The method of claim 1 , further comprising:
receiving a response selection from the set of response options from the user;
sending the selected response to a recipient; and
storing the selected response in a corpus of historical messages sent by a plurality of users, wherein the response options are determined from the corpus of historical messages.
9. The method of claim 1 , wherein response options of the set of response options are determined from a corpus of historical messages sent by a plurality of users.
10. The method of claim 1 , wherein the method is repeated for each message sent in the conversation thread, wherein a last message in the conversation thread is treated as the query message.
11. The method of claim 1 , further comprising inserting personalized information into a response option, with a personalization model structured to insert participant-specific information into the response option of the set of response options.
12. The method of claim 1 , further comprising inserting personalized information into a response option, with a personalization model structured to modify tone of the response option with a generative model.
13. The method of claim 1 , wherein an embedding model of the trained model is trained on cross-lingual word embeddings comprising mixed contents of a set of different languages.
14. The method of claim 13 , wherein the embedding model is structured to optimize a cross-lingual constraint between embeddings of different languages, and encourages close proximity of embeddings of similar words across languages in a shared embedding space.
15. The method of claim 1 , wherein an embedding model of the trained model is trained on to jointly optimize a combination of monolingual and cross-lingual losses.
16. A system for providing automatic message suggestions comprising:
a messaging platform that receives and transmits messages between users;
a query embedding model that determines a query embedding for a last message in a conversation thread;
a context embedding model that determines a set of context embeddings of a set of previously sent messages from the conversation thread;
a smart response model that determines a ranked list of response options based on the query embedding and the set of context embeddings, wherein the ranked list of response options comprises at least two of: a listening response, a guidance response, prescriptive information response, and a follow-up question response; and
a diversity sampling model that selects a set of response options from the ranked list, wherein each response option of the set of response options is distinct from each of the other members of the set, wherein the diversity sampling model determines a set of semantic clusters from the ranked list of response options and selects one of the set of response options from each of the set of semantic clusters.
17. The system of claim 16 , wherein the smart response model generates a response option of the ranked list of response options using a generative model.