Processing clusters with mathematical models for message suggestion
A message suggestion service may use clusters of pre-approved messages to improve the quality of messages suggested to users. During a conversation, messages of the conversation may be processed with a neural network to compute a conversation encoding vector. The neural network may also be used to compute pre-approved message encoding vectors of the pre-approved messages. Distances between the conversation encoding vector and the pre-approved message encoding vectors may be used to select one or more clusters. Distances between the conversation encoding vector and the pre-approved message encoding vectors may then be used to select one or more pre-approved messages from the selected clusters. The selected pre-approved messages may then be presented as suggested messages to a user.
1. A 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:
obtain one or more conversation messages from a conversation between a first user and a second user;
compute a conversation encoding vector by processing the one or more conversation messages with a neural network;
obtain a plurality of clusters of pre-approved message suggestions, wherein:
the plurality of clusters comprises a first cluster and a second cluster,
the first cluster comprises a first pre-approved message suggestion and a second pre-approved message suggestion, and
the second cluster comprises a third pre-approved message suggestion and a fourth pre-approved message suggestion;
obtain message encoding vectors for the pre-approved message suggestions, wherein a message encoding vector is computed by processing a corresponding pre-approved message suggestion with the neural network;
compute a first distance between the conversation encoding vector and a first message encoding vector corresponding to the first pre-approved message suggestion;
compute a second distance between the conversation encoding vector and a second message encoding vector corresponding to the second pre-approved message suggestion;
compute a third distance between the conversation encoding vector and a third message encoding vector corresponding to the third pre-approved message suggestion;
compute a fourth distance between the conversation encoding vector and a fourth message encoding vector corresponding to the fourth pre-approved message suggestion;
compute a first cluster selection score for the first cluster by processing a first feature with a first tree-based model, wherein the first feature is computed using at least one of the first distance or the second distance;
compute a second cluster selection score for the second cluster by processing a second feature with the first tree-based model, wherein the second feature is computed using at least one of the third distance or the fourth distance;
select the first cluster using the first cluster selection score and the second cluster selection score;
compute a first message selection score by processing the first distance with a second tree-based model, wherein the second tree-based model is different than the first tree-based model;
compute a second message selection score by processing the second distance with the second tree-based model;
select the first pre-approved message suggestion using the first message selection score and the second message selection score; and
presenting the first pre-approved message suggestion to the first user as a suggested message to send to the second user.
2. The system of claim 1 , wherein the at least one server computer is configured to:
select the second cluster;
select the third pre-approved message suggestion from the second cluster; and
presenting the third pre-approved message suggestion to the first user as the suggested message to send to the second user.
3. The system of claim 1 , wherein the message encoding vectors are computed in advance and obtained from storage.
4. The system of claim 1 , wherein the neural network comprises a conversation encoding model and a message encoding model.
5. The system of claim 1 , wherein the conversation encoding vector is computed by sequentially processing tokens of the one or more conversation messages with a recurrent neural network.
6. The system of claim 1 , wherein the first feature comprises one or more of:
a minimum distance between the conversation encoding vector and message encodings of the first cluster;
a maximum distance between the conversation encoding vector and the message encodings of the first cluster; or
an average distance between the conversation encoding vector and the message encodings of the first cluster.
7. A computer-implemented method for suggesting a message, comprising:
obtaining one or more conversation messages from a conversation between a first user and a second user;
computing a conversation encoding vector by processing the one or more conversation messages with a neural network;
obtaining a plurality of clusters of pre-approved message suggestions, wherein:
the plurality of clusters comprises a first cluster and a second cluster,
the first cluster comprises a first pre-approved message suggestion and a second pre-approved message suggestion, and
the second cluster comprises a third pre-approved message suggestion and a fourth pre-approved message suggestion;
obtaining message encoding vectors for the pre-approved message suggestions, wherein a message encoding vector is computed by processing a corresponding pre-approved message suggestion with the neural network;
computing a first distance between the conversation encoding vector and a first message encoding vector corresponding to the first pre-approved message suggestion;
computing a second distance between the conversation encoding vector and a second message encoding vector corresponding to the second pre-approved message suggestion;
computing a third distance between the conversation encoding vector and a third message encoding vector corresponding to the third pre-approved message suggestion;
computing a fourth distance between the conversation encoding vector and a fourth message encoding vector corresponding to the fourth pre-approved message suggestion;
computing a first cluster selection score for the first cluster by processing a first feature with a first probabilistic graphical model, wherein the first feature is computed using at least one of the first distance or the second distance;
computing a second cluster selection score for the second cluster by processing a second feature with the first probabilistic graphical model, wherein the second feature is computed using at least one of the third distance or the fourth distance;
selecting the first cluster using the first cluster selection score and the second cluster selection score;
computing a first message selection score by processing the first distance with a second probabilistic graphical model, wherein the second probabilistic graphical model is different than the first probabilistic graphical model;
computing a second message selection score by processing the second distance with the second probabilistic graphical model; and
selecting the first pre-approved message suggestion using the first message selection score and the second message selection score.
8. The computer-implemented method of claim 7 , wherein the first cluster selection score is computed by processing a third feature with the first probabilistic graphical model.
9. The computer-implemented method of claim 8 , wherein the third feature is independent of the conversation encoding vector.
10. The computer-implemented method of claim 8 , wherein the third feature comprises a frequency of use of the first cluster by any user when suggested; a frequency of use of the first cluster by the first user when suggested; an amount of time since a most recent message in the conversation; or a number of messages in the conversation.
11. The computer-implemented method of claim 7 , wherein the first message selection score is computed by processing a third feature with the second probabilistic graphical model.
12. The computer-implemented method of claim 11 , wherein the third feature is independent of the conversation encoding vector.
13. The computer-implemented method of claim 12 , wherein the third feature comprises a time of day; a sentiment of the one or more conversation messages; or an indication of current weather.
14. The computer-implemented method of claim 7 , wherein the first probabilistic graphical model comprises a tree-based model.
15. One or more non-transitory, computer-readable media comprising computer-executable instructions that, when executed, cause at least one processor to perform actions comprising:
obtaining one or more conversation messages from a conversation between a first user and a second user;
computing a conversation encoding vector by processing the one or more conversation messages with a neural network;
obtaining a plurality of clusters of pre-approved message suggestions, wherein:
the plurality of clusters comprises a first cluster and a second cluster,
the first cluster comprises a first pre-approved message suggestion and a second pre-approved message suggestion, and
the second cluster comprises a third pre-approved message suggestion and a fourth pre-approved message suggestion;
obtaining message encoding vectors for the pre-approved message suggestions, wherein a message encoding vector is computed by processing a corresponding pre-approved message suggestion with the neural network;
computing a first distance between the conversation encoding vector and a first message encoding vector corresponding to the first pre-approved message suggestion;
computing a second distance between the conversation encoding vector and a second message encoding vector corresponding to the second pre-approved message suggestion;
computing a third distance between the conversation encoding vector and a third message encoding vector corresponding to the third pre-approved message suggestion;
computing a fourth distance between the conversation encoding vector and a fourth message encoding vector corresponding to the fourth pre-approved message suggestion;
computing a first cluster selection score for the first cluster by processing a first feature with a first tree-based model, wherein the first feature is computed using at least one of the first distance or the second distance;
computing a second cluster selection score for the second cluster by processing a second feature with the first tree-based model, wherein the second feature is computed using at least one of the third distance or the fourth distance;
selecting the first cluster using the first cluster selection score and the second cluster selection score;
computing a first message selection score by processing the first distance with a second tree-based model, wherein the second tree-based model is different than the first tree-based model;
computing a second message selection score by processing the second distance with the second tree-based model; and
selecting the first pre-approved message suggestion using the first message selection score and the second message selection score.
16. The one or more non-transitory, computer-readable media of claim 15 , wherein the first tree-based model comprises an ensemble of decision trees or a random forest.
17. The one or more non-transitory, computer-readable media of claim 15 , wherein the first tree-based model is implemented using gradient boosting or bagging.
18. The one or more non-transitory, computer-readable media of claim 15 , wherein the first user is a customer support agent and the second user is a customer.
19. The one or more non-transitory, computer-readable media of claim 15 , wherein the first tree-based model and the second tree-based model are retrained more frequently than the neural network.