Customer care training using chatbots

1. A computer-implemented method comprising:

training, by one or more processors, a chatbot for a customer chat simulation based on customer service conversation data, task scenarios, and customer personas;

monitoring, by the one or more processors, interactions between customer service agents and the chatbot trained for customer chat simulation, wherein the interactions include requests provided by the chatbot and respective responses to the requests provided by the customer service agents in respective customer service agent styles;

determining, by the one or more processors, model customer service agent responses to the requests using a sequence-to-sequence model, wherein the sequence-to-sequence model receives, as input sequences, the requests in reverse order, and wherein the sequence-to-sequence model generates, as output sequences, the model customer service agent responses;

determining, by the one or more processors, assessments of the performance of the customer service agents, including respective training levels of the customer service agents, based on comparisons between the respective responses of the customer service agents and the model customer service agent responses;

generating, by the one or more processors, feedback for the customer service agents based on the assessments of the performance of the customer service agents; and

matching, by the one or more processors, the customer service agents to respective customers in a customer service routing system using the respective customer service agent styles and the performance of the respective customer service agents as ground truths,

wherein generating the feedback for the customer service agents includes generating a multiple-choice question for a first customer service agent having a first training level, generating a response template for a second customer service agent having a second training level, and generating a hint, extracted from a model customer service agent response, for a third customer service agent having a third training level.

2. The method of claim 1 , wherein determining, by the one or more processors, the model customer service agent responses further comprises:

determining, by the one or more processors, a plurality of model customer service agent responses for each interaction using the sequence-to-sequence model; and

selecting, by the one or more processors, a respective model customer service agent response from the plurality of model customer service agent responses based on a target style, a target tone, or a customer persona.

3. The method of claim 1 , wherein the sequence-to-sequence model: (i) uses an encoder neural network to map variable length input sequences to fixed length vectors, and (ii) uses a decoder neural network to map the fixed length vectors to variable length output sequences.

4. The method of claim 2 , wherein each plurality of model customer service agent responses is based on the respective customer persona, a respective task scenario, and at least one respective context data.

5. The method of claim 4 , wherein the respective customer persona, the respective task scenario, and the at least one respective context data for each interaction are applied as additional constraints for word embedding in the sequence-to-sequence model.

6. The method of claim 1 , wherein the comparisons between the respective responses and the model customer service agent responses are based, at least in part, on automatic metrics.

7. The method of claim 6 , wherein the automatic metrics include a BLEU score.

8. The method of claim 6 , wherein the automatic metrics include a Tf-idf score.

9. The method of claim 6 , wherein the automatic metrics include a word2vec score.

10. The method of claim 2 , wherein the selecting of the respective model customer service agent response from the plurality of model customer service agent responses is based on the target style, wherein the target style is selected from a group of customer chatbot styles, and wherein the group of customer chatbot styles is built as a bottom-up taxonomy by performing unsupervised machine learning to extract customer chatbot styles from the customer service conversation data.

11. The method of claim 2 , wherein the selecting of the respective model customer service agent response from the plurality of model customer service agent responses is further based on a beam search.

12. The method of claim 1 , wherein the determining of the assessments of the performance of the customer service agents includes comparing the performance of at least one customer service agent to a threshold level of performance.

13. The method of claim 12 , further comprising:

determining, by the one or more processors, that the performance of the at least one customer service agent is below the threshold level of performance; and

reducing, by the one or more processors, a difficulty level for training the at least one customer service agent.

14. The method of claim 12 , further comprising:

determining, by the one or more processors, that the performance of the at least one customer service agent is above the threshold level of performance; and

increasing, by the one or more processors, a level of uncertainty in responses provided by the chatbot to the at least one customer service agent.

15. The method of claim 1 , further comprising:

determining, by the one or more processors, a plurality of model customer service agent responses using a word-graph construction approach.

16. The method of claim 1 , further comprising:

determining, by the one or more processors, a plurality of model customer service agent responses using multi-sentence compression.

17. The method of claim 1 , wherein the generating of the feedback for the customer service agents further includes asking the first customer service agent to select the best response to the multiple-choice question, asking the second customer service agent to fill out the response template to respond to a customer request, and providing the hint to the third customer service agent.