Automated conversation review to surface virtual assistant misunderstandings
A scalable system provides automated conversation review that can identify potential miscommunications. The system may provide suggested actions to fix errors in intelligent virtual assistant (IVA) understanding, may prioritize areas of language model repair, and may automate the review of conversations. By the use of an automated system for conversation review, problematic interactions can be surfaced without exposing the entire set of conversation logs to human reviewers, thereby minimizing privacy invasion. A scalable system processes conversations and autonomously marks the interactions where the IVA is misunderstanding the user.
1. A method comprising:
receiving conversation data between a user and an intelligent virtual assistant, wherein the conversation data comprises a plurality of conversational turns;
identifying an intent for the conversation data, the intent being associated with a language model for natural language processing techniques;
generating a risk score for each of the plurality of conversational turns, wherein the risk score represents a risk of intent misclassification, wherein generating the risk score for each conversational turn comprises:
detecting a plurality of indications of misunderstanding in the conversational turn;
determining a plurality of indicator scores for the conversational turn by scoring each of the indications of misunderstanding of the plurality of indications of misunderstanding with an indicator score; and
determining the risk score for the conversational turn as the weighted sum of all the indicator scores in that conversational turn;
determining, using the intent and the risk score for each of the plurality of conversational turns, that the conversation data has a misunderstood intent;
determining an action to adjust the language model responsive to determining that the conversation data has the misunderstood intent; and
adjusting the language model using the action.
2. The method of claim 1 , wherein the determining that the conversation data has the misunderstood intent is performed without human review.
3. The method of claim 1 , wherein determining that the conversation data has the misunderstood intent comprises ranking the conversational turns by risk score for priority review.
4. The method of claim 3 , wherein determining that the conversation data has the misunderstood intent comprises detecting a plurality of features of intent error and aggregating the features into the risk score.
5. The method of claim 3 , further comprising reviewing the conversational turns based on ranking to determine whether each reviewed turn was misunderstood.
6. The method of claim 1 , wherein adjusting the language model comprises adjusting the intelligent virtual assistant.
7. The method of claim 1 , further comprising processing the conversation data with one or more natural language processing techniques to identify the intent for the conversation data.
8. The method of claim 7 , further comprising determining, based at least in part on a presence of one or more indicators of intent error for the conversation data, a measure of confidence that the intent is correctly identified for the user input, and associating the measure of confidence with the conversation data.
9. The method of claim 8 , further comprising:
receiving a selection of the intent to obtain feedback regarding the intent;
presenting a feedback interface that enables a voter to provide feedback regarding matching of the conversation data to the intent;
receiving feedback for the voter regarding an accuracy of matching the conversation data to the intent;
evaluating the intent based at least in part on the feedback;
determining that the feedback indicates that the matching of the conversation data to the intent is not accurate;
increasing a weighting to be applied to the one or more indicators of intent error based at least in part on the determining that the feedback indicates that the matching of the conversation data to the intent is not accurate; and
applying the weighting to the one or more indicators of intent error.
10. The method of claim 9 , wherein the feedback includes a vote indicating whether or not the conversation data matches the intent.
11. A system comprising:
one or more processors;
storage storing conversation data representing at least one conversation between a user and an intelligent virtual assistant, wherein the conversation data comprises a plurality of conversational turns, and one or more indicators of intent error;
memory communicatively coupled to the one or more processors and storing executable instructions that, when executed by the one or more processors, cause the one or more processors to perform acts comprising:
receiving the conversation data between the user and the intelligent virtual assistant;
identifying an intent for the conversation data, the intent being associated with a language model for natural language processing techniques;
generating a risk score for each of the plurality of conversational turns, wherein the risk score represents a risk of intent misclassification, wherein generating the risk score for each conversational turn comprises:
detecting a plurality of indications of misunderstanding in the conversational turn;
determining a plurality of indicator scores for the conversational turn by scoring each of the indications of misunderstanding of the plurality of indications of misunderstanding with an indicator score; and
determining the risk score for the conversational turn as the weighted sum of all the indicator scores in that conversational turn;
ranking the conversational turns by risk score for priority review;
determining, using the intent and the risk score for each of the plurality of conversational turns, that the conversation data has a misunderstood intent;
determining an action to adjust the language model responsive to determining that the conversation data has a misunderstood intent; and
adjusting the language model using the action.
12. The system of claim 11 , wherein the determining that the conversation data has the misunderstood intent is performed without human review.
13. The system of claim 11 , wherein determining that the conversation data has the misunderstood intent comprises detecting a plurality of features of intent error and aggregating the features into the risk score.
14. The system of claim 11 , wherein the acts further comprise reviewing the conversational turns based on ranking to determine whether each reviewed turn was misunderstood.
15. The system of claim 11 , wherein adjusting the language model comprises adjusting the intelligent virtual assistant.
16. A method comprising:
receiving conversation data between a user and an intelligent virtual assistant, wherein the conversation data comprises a plurality of conversational turns;
identifying an intent for the conversation data, the intent being associated with a language model for natural language processing techniques
identifying an intent for the conversation data by processing the conversation data with one or more of the natural language processing techniques;
generating a risk score for each of the plurality of conversational turns, wherein the risk score represents a risk of intent misclassification, wherein generating the risk score for each conversational turn comprises:
detecting a plurality of indications of misunderstanding in the conversational turn;
determining a plurality of indicator scores for the conversational turn by scoring each of the indications of misunderstanding of the plurality of indications of misunderstanding with an indicator score; and
determining the risk score for the conversational turn as the weighted sum of all the indicator scores in that conversational turn;
determining, using the intent and the risk score for each of the plurality of conversational turns, that the conversation data has a misunderstood intent;
determining, based at least in part on a presence of one or more indicators of intent error for the conversation data, a measure of confidence that the intent is correctly identified for the user input, and associating the measure of confidence with the conversation data; and
determining an action to adjust the language model responsive to the measure of confidence and responsive to determining that the conversation data has the misunderstood intent; and
adjusting the language model using the action.
17. The method of claim 16 , further comprising:
receiving a selection of the intent to obtain feedback regarding the intent;
presenting a feedback interface that enables a voter to provide feedback regarding matching of the conversation data to the intent;
receiving feedback for the voter regarding an accuracy of matching the conversation data to the intent;
evaluating the intent based at least in part on the feedback;
determining that the feedback indicates that the matching of the conversation data to the intent is not accurate;
increasing a weighting to be applied to the one or more indicators of intent error based at least in part on the determining that the feedback indicates that the matching of the conversation data to the intent is not accurate; and
applying the weighting to the one or more indicators of intent error.
18. The method of claim 17 , wherein the feedback includes a vote indicating whether or not the conversation data matches the intent.