Systems and methods for structured conversation summarization
Embodiments described herein provide a structured conversation summarization framework. A user interface may be provided which allows an agent to perform a conversation with a customer, for example regarding resolving a customer support issue. Utterances by both the agent and customer may be stored, and at the end of the conversation, the utterances may be used to generate a structured summary. The structured summary may include components such as a general summary, an issue summary, and a resolution summary. Using neural network models and heuristics, each component of the summary may be automatically generated.
1. A method of structured summarization of user-agent interactions on a service platform, the method comprising:
receiving, via a user interface, a plurality of user utterances;
generating, by a service agent and via the user interface, a plurality of responses in response to the plurality of user utterances, respectively;
retrieving a structured summary form based on a conversation topic of the plurality of user utterances and the plurality of responses;
generating, by a summarization model trained on a dataset of dialogue data and annotated reference summaries, a summary of the plurality of user utterances and the plurality of responses;
determining, for an issue summary field on the structured summary form, a first utterance selected from the plurality of user utterances as an issue summary based on ranking similarities between the summary and a subset of user utterances selected from the plurality of user utterances, respectively;
determining, for a resolution summary field on the structured summary form, a first response selected from the plurality of responses as a resolution summary based on a subset of responses selected from the plurality of responses; and
generating an engageable user interface element including data fields of the structured summary form incorporating the issue summary and the resolution summary.
2. The method of claim 1 , wherein retrieving the structured summary form comprises:
obtaining a user selected conversation topic through the user interface; and
querying a summary form database based on the user selected conversation topic.
3. The method of claim 1 , wherein generating, by the summarization model, the summary comprises:
concatenating the plurality of user utterances and the plurality of responses into an input sequence of tokens; and
generating, by the summarization model, the summary based on the input sequence of tokens.
4. The method of claim 1 , wherein determining the first utterance as the issue summary comprises:
filtering the plurality of user utterances based on relevance into the subset of user utterances;
encoding, via a sentence Transformer model, the subset of user utterances into a first plurality of utterance representations;
encoding, via the sentence Transformer model, the summary into a summary representation;
computing a set of cosine similarities between the summary representation and the first plurality of utterance representations, respectively; and
selecting the first utterance corresponding to a greatest cosine similarity with the summary representation.
5. The method of claim 4 , further comprising:
generating, via the summarization model, the issue summary based on the first utterance.
6. The method of claim 1 , wherein determining the first response as the resolution summary comprises:
selecting the first response from the subset of responses based on lengths of responses and heuristics of past responses.
7. The method of claim 1 , wherein the subset of responses comprise a second half of the plurality of responses in time.
8. The method of claim 1 , wherein the structured summary form further comprises a field indicating whether an issue discussed during the user-agent interactions has been resolved, and wherein the field is generated using a prediction model trained on a dataset of dialogue data and annotated resolution results.
9. A system for structured summarization of user-agent interactions on a service platform, the system comprising:
a memory that stores a summarization model trained on a dataset of dialogue data and annotated reference summaries and a plurality of processor-executable instructions;
a communication interface that receives a plurality of user utterances; and
one or more hardware processors that read and execute the plurality of processor-executable instructions from the memory to perform operations comprising:
generating, via the communication interface, a plurality of responses in response to the plurality of user utterances, respectively;
retrieving a structured summary form based on a conversation topic of the plurality of user utterances and the plurality of responses;
generating, by the summarization model, a summary of the plurality of user utterances and the plurality of responses;
determining, for an issue summary field on the structured summary form, a first utterance selected from the plurality of user utterances as an issue summary based on ranking similarities between the summary and a subset of user utterances selected from the plurality of user utterances, respectively;
determining, for a resolution summary field on the structured summary form, a first response selected from the plurality of responses as a resolution summary based on a subset of responses selected from the plurality of responses; and
generating an engageable user interface element including data fields of the structured summary form incorporating the issue summary and the resolution summary.
10. The system of claim 9 , wherein retrieving the structured summary form comprises:
obtaining a user selected conversation topic through the communication interface; and
querying a summary form database based on the user selected conversation topic.
11. The system of claim 9 , wherein generating, by the summarization model, the summary comprises:
concatenating the plurality of user utterances and the plurality of responses into an input sequence of tokens; and
generating, by the summarization model, the summary based on the input sequence of tokens.
12. The system of claim 9 , wherein determining the first utterance as the issue summary comprises:
filtering the plurality of user utterances based on relevance into the subset of user utterances;
encoding, via a sentence Transformer model, the subset of user utterances into a first plurality of utterance representations;
encoding, via the sentence Transformer model, the summary into a summary representation;
computing a set of cosine similarities between the summary representation and the first plurality of utterance representations, respectively; and
selecting the first utterance corresponding to a greatest cosine similarity with the summary representation.
13. The system of claim 12 , the operations further comprising:
generating, via the summarization model, the issue summary based on the first utterance.
14. The system of claim 9 , wherein determining the first response as the resolution summary comprises:
selecting the first response from the subset of responses based on lengths of responses and heuristics of past responses.
15. The system of claim 9 , wherein the subset of responses comprise a second half of the plurality of responses in time.
16. The system of claim 9 , wherein the structured summary form further comprises a field indicating whether an issue discussed during the user-agent interactions has been resolved, and wherein the field is generated using a prediction model trained on a dataset of dialogue data and annotated resolution results.
17. A non-transitory machine-readable medium comprising a plurality of machine-executable instructions which, when executed by one or more processors, are adapted to cause the one or more processors to perform operations comprising:
receiving, via a user interface, a plurality of user utterances;
generating, via the user interface, a plurality of responses in response to the plurality of user utterances, respectively;
retrieving a structured summary form based on a conversation topic of the plurality of user utterances and the plurality of responses;
generating, by a summarization model trained on a dataset of dialogue data and annotated reference summaries, a summary of the plurality of user utterances and the plurality of responses;
determining, for an issue summary field on the structured summary form, a first utterance selected from the plurality of user utterances as an issue summary based on ranking similarities between the summary and a subset of user utterances selected from the plurality of user utterances, respectively;
determining, for a resolution summary field on the structured summary form, a first response selected from the plurality of responses as a resolution summary based on a subset of responses selected from the plurality of responses; and
generating an engageable user interface element including data fields of the structured summary form incorporating the issue summary and the resolution summary.
18. The non-transitory machine-readable medium of claim 17 , wherein retrieving the structured summary form comprises:
obtaining a user selected conversation topic through the user interface; and
querying a summary form database based on the user selected conversation topic.