Machine learning divide and conquer techniques for long dialog summarization
Various embodiments of the present disclosure provide summarization techniques for summarizing complex documents, such as long unstructured call transcripts. The summarization techniques include generating a plurality of interaction topics for an interaction transcript and iteratively summarizing each interaction topic based on a preceding partial summary for the interaction transcript that corresponds to a preceding interaction topic that precedes the interaction topic in the interaction transcript. An abstractive summary is generated using a recursive abstractive model that is trained using training data generated based on holistic similarity scores between interaction topics of a call transcript and summary sentences of a corresponding target summary.
1 . A computer-implemented method comprising:
generating, by one or more processors, a plurality of interaction topics for an interaction transcript, wherein each interaction topic comprises one or more respective related sequential utterances of the interaction transcript;
receiving, by the one or more processors, a preceding partial summary for the interaction transcript, wherein the preceding partial summary corresponds to at least one preceding interaction topic that precedes an interaction topic in the interaction transcript; and
generating, by the one or more processors and using a machine learning recursive abstractive model, an abstractive summary for the interaction transcript based on the interaction topic and the preceding partial summary, wherein the machine learning recursive abstractive model comprises (i) a first encoder configured to generate a first encoding based on the interaction topic, and (ii) a second encoder configured to generate a second encoding based on the preceding partial summary.
2 . The computer-implemented method of claim 1 , wherein:
the interaction transcript comprises a plurality of interaction utterances,
one or more related sequential utterances of the interaction topic comprise a first subset of the plurality of interaction utterances, and
the preceding partial summary comprises one or more preceding summary sentences corresponding to a second subset of the plurality of interaction utterances that precede the first subset in the interaction transcript.
3 . The computer-implemented method of claim 2 , wherein generating the plurality of interaction topics for the interaction transcript comprises:
generating, by the one or more processors and using a machine learning topic segmentation model, an utterance similarity score between a pair of interaction utterances from the plurality of interaction utterances;
assigning, by the one or more processors, a first topic label to a first interaction utterance of the pair of interaction utterances based on the utterance similarity score; and
assigning, by the one or more processors, a second topic label to a second interaction utterance of the pair of interaction utterances based on the utterance similarity score.
4 . The computer-implemented method of claim 1 further comprising:
determining, by the one or more processors, a topic relevance for the interaction topic based on one or more related sequential utterances of the interaction topic, wherein the topic relevance is indicative of at least one of a relevant topic or an irrelevant topic; and
in response to determining that the interaction topic is the relevant topic, generating, by the one or more processors and using the machine learning recursive abstractive model, the abstractive summary for the interaction transcript based on the interaction topic.
5 . The computer-implemented method of claim 4 , wherein determining the topic relevance comprises:
generating, by the one or more processors and using a machine learning extractive summarizer model, an extractive summary for the interaction transcript;
generating, by the one or more processors, a topic similarity score between the one or more related sequential utterances and the extractive summary; and
determining, by the one or more processors, the topic relevance of the interaction topic based on the topic similarity score.
6 . The computer-implemented method of claim 5 , wherein the topic similarity score is indicative of a textual similarity between at least one of the one or more related sequential utterances and at least one extractive sentence from the extractive summary.
7 . The computer-implemented method of claim 1 , wherein the machine learning recursive abstractive model further comprises an attention layer configured to generate an attended encoding based on the first encoding and the second encoding.
8 . The computer-implemented method of claim 7 , wherein the machine learning recursive abstractive model further comprises a decoder configured to generate the abstractive summary based on the attended encoding.
9 . The computer-implemented method of claim 8 , wherein the machine learning recursive abstractive model comprises a neural network architecture.
10 . The computer-implemented method of claim 1 , wherein the first encoder comprises a first feed forward layer and a first self attention layer configured to generate the first encoding, and the second encoder comprises a second feed forward layer and a second self attention layer configured to generate the second encoding.
11 . The computer-implemented method of claim 10 , wherein the machine learning recursive abstractive model further comprises a joint decoder configured to generate the abstractive summary based on the first encoding and the second encoding.
12 . A system comprising:
one or more processors; and
one or more memories storing processor-executable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations comprising:
generating a plurality of interaction topics for an interaction transcript, wherein each interaction topic comprises one or more respective related sequential utterances of the interaction transcript;
receiving a preceding partial summary for the interaction transcript, wherein the preceding partial summary corresponds to at least one preceding interaction topic that precedes an interaction topic in the interaction transcript; and
generate, using a machine learning recursive abstractive model, an abstractive summary for the interaction transcript based on the interaction topic and the preceding partial summary, wherein the machine learning recursive abstractive model comprises (i) a first encoder configured to generate a first encoding based on the interaction topic, and (ii) a second encoder configured to generate a second encoding based on the preceding partial summary.
13 . The system of claim 12 , wherein:
the interaction transcript comprises a plurality of interaction utterances,
one or more related sequential utterances of the interaction topic comprise a first subset of the plurality of interaction utterances, and
the preceding partial summary comprises one or more preceding summary sentences corresponding to a second subset of the plurality of interaction utterances that precede the first subset.
14 . The system of claim 13 , wherein generating the plurality of interaction topics for the interaction transcript comprises:
generating, using a machine learning topic segmentation model, an utterance similarity score between a pair of interaction utterances from the plurality of interaction utterances;
assigning a first topic label to a first interaction utterance of the pair of interaction utterances based on the utterance similarity score; and
assigning a second topic label to a second interaction utterance of the pair of interaction utterances based on the utterance similarity score.
15 . The system of claim 12 , wherein the one or more processors are further configured to:
determine a topic relevance for the interaction topic based on one or more related sequential utterances of the interaction topic, wherein the topic relevance is indicative of at least one of a relevant topic or an irrelevant topic; and
in response to determining that the interaction topic is the relevant topic, generate, using the machine learning recursive abstractive model, the abstractive summary for the interaction transcript based on the interaction topic.
16 . The system of claim 15 , wherein determining the topic relevance comprises:
generating, using a machine learning extractive summarizer model, an extractive summary for the interaction transcript;
generating a topic similarity score between the one or more related sequential utterances and the extractive summary; and
determining the topic relevance of the interaction topic based on the topic similarity score.
17 . The system of claim 16 , wherein the topic similarity score is indicative of a textual similarity between at least one of the one or more related sequential utterances and at least one extractive sentence from the extractive summary.
18 . One or more non-transitory computer-readable media storing processor-executable instructions that, when executed by one or more processors, cause the one or more processors to perform operations comprising:
generating a plurality of interaction topics for an interaction transcript, wherein each interaction topic comprises one or more respective related sequential utterances of the interaction transcript;
receiving a preceding partial summary for the interaction transcript, wherein the preceding partial summary corresponds to at least one preceding interaction topic that precedes an interaction topic in the interaction transcript; and
generating, using a machine learning recursive abstractive model, an abstractive summary for the interaction transcript based on the interaction topic and the preceding partial summary, wherein the machine learning recursive abstractive model comprises (i) a first encoder configured to generate a first encoding based on the interaction topic, and (ii) a second encoder configured to generate a second encoding based on the preceding partial summary.
19 . The one or more non-transitory computer-readable storage media of claim 18 , wherein the machine learning recursive abstractive model further comprises an attention layer configured to generate an attended encoding based on the first encoding and the second encoding.
20 . The one or more non-transitory computer-readable media of claim 19 , wherein the machine learning recursive abstractive model further comprises a decoder configured to generate the abstractive summary based on the attended encoding.