Natural language processing for optimized extractive summarization
There is a need for more effective and efficient predictive natural language summarization. This need can be addressed by, for example, solutions for performing predictive natural language summarization using a constrained optimization model. In one example, a method includes identifying one or more per-party utterance subsets in a multi-party call transcript; generating a plurality of eligible extractive summaries that comply with one or more optimization constraints; for each eligible extractive summary of the plurality of eligible extractive summaries, determining an overall summary utility measure; generating the optimal extractive summary based at least in part on each overall summary utility measure for an eligible extractive summary of the plurality of eligible extractive summaries; and performing one or more summary-based actions based at least in part on the optimal extractive summary.
1. A computer-implemented method comprising:
determining, by one or more processors, one or more per-party utterance subsets from multi-party interaction transcript data, wherein: (1) the multi-party interaction transcript data comprises a plurality of interaction utterances associated with a plurality of interaction parties, and (2) each of the one or more per-party utterance subsets comprises one or more of the plurality of interaction utterances that are associated with one of the plurality of interaction parties;
generating, by the one or more processors and using an unsupervised machine learning model, an optimal extractive summary, wherein the optimal extractive summary comprises one of a plurality of eligible extractive summaries that is selected based at least in part on respective overall summary utility measures for each of the plurality of eligible extractive summaries, wherein:
(i) each of the plurality of eligible extractive summaries comprises a covered subset of the plurality of interaction utterances that complies with one or more optimization constraints, and the one or more optimization constraints comprise a similarity-based optimization constraint based at least in part on a covered subset for a particular one of the plurality of eligible extractive summaries comprising (a) a particular one of the plurality of interaction utterances that is in a particular one of the one or more per-party utterance subsets, and (b) additional ones of the plurality of interaction utterances, each of which is in any per-party utterance subset of the one or more per-party utterance subsets other than the particular one of the one or more per-party utterance subsets having a threshold-satisfying utterance similarity measure with respect to the particular one of the plurality of interaction utterances, and
(ii) each of the respective overall summary utility measures is based at least in part on an information quality measure and a linguistic quality measure for each of the plurality of interaction utterances in the covered subset for the particular one of the plurality of eligible extractive summaries; and
initiating, by the one or more processors, the performance of one or more summary-based actions based at least in part on the optimal extractive summary.
2. The computer-implemented method of claim 1 , wherein the one or more optimization constraints further comprise a content word coverage constraint requiring that, in an instance in which the covered subset for the particular one of the plurality of eligible extractive summaries comprises a particular interaction utterance having one or more per-utterance content words, then the covered subset for the particular one of plurality of eligible extractive summaries comprises all of the one or more per-utterance content words.
3. The computer-implemented method of claim 1 , wherein the one or more optimization constraints further comprise an utterance coverage constraint requiring that, in an instance in which the covered subset for the particular one of the purality of eligible extractive summaries comprises a particular covered content word, the covered subset for the particular one of the plurality of eligible extractive summaries further comprises at least one interaction utterance of the plurality of interaction utterances that is associated with the particular covered content word.
4. The computer-implemented method of claim 1 , wherein the one or more optimization constraints further comprise a party coverage constraint requiring that each covered subset for an eligible extractive summary comprises at least one interaction utterance from each of the one or more per-party utterance subsets.
5. The computer-implemented method of claim 1 , wherein the one or more optimization constraints further comprise a party size constraint requiring that each covered subset for an eligible extractive summary comprises a below-threshold count of interaction utterances from each of the one or more per-party utterance subsets.
6. The computer-implemented method of claim 1 further comprising: generating each information quality measure by:
generating a graph representation of one or more per-utterance content words that are associated with each of the plurality of interaction utterances in the covered subset for the particular one of the plurality of eligible extractive summaries;
for each per-utterance content word of one or more per-utterance content words that is associated with each of the plurality of interaction utterances in the covered subset for the particular one of the plurality of eligible extractive summaries:
generating a neighboring subset of the one or more per-utterance content words based at least in part on the graph representation, and
generating a per-word the information quality measure based at least in part on each per-utterance content word of the one or more per-utterance content words that is in the neighboring subset of the per-utterance content word; and
generating the information quality measure based at least in part on each per-word information quality measure for a per-utterance content word of the one or more per-utterance content words that is associated with each of the plurality of interaction utterances in the covered subset for the particular one of the plurality of eligible extractive summaries.
7. The computer-implemented method of claim 1 further comprising generating the linguistic quality measure by generating a probability aggregation measure based at least in part on a per-sequence probability for each n-gram word sequence of a set of n-gram word sequences that are associated with each of the plurality of interaction utterances in the covered subset for the particular one of the plurality of eligible extractive summaries.
8. The computer-implemented method of claim 7 , where the set of n-gram word sequences comprises a set of trigram word sequences.
9. An apparatus comprising one or more processors and memory including program code, the memory and the program code configured to, with the one or more processors, cause the apparatus to at least:
determine one or more per-party utterance subsets from multi-party interaction transcript data, wherein: (1) the multi-party interaction transcript data comprises a plurality of interaction utterances associated with a plurality of interaction parties, and (2) each of the one or more per-party utterance subsets comprises one or more of the plurality of interaction utterances that are associated with one of the plurality of interaction parties;
generate, using an unsupervised machine learning model, an optimal extractive summary, wherein the optimal extractive summary comprises one of a plurality of eligible extractive summaries that is selected based at least in part on respective overall summary utility measures for each of the plurality of eligible extractive summaries, wherein:
(i) each of the plurality of eligible extractive summaries comprises a covered subset of the plurality of interaction utterances that complies with one or more optimization constraints, and the one or more optimization constraints comprise a similarity-based optimization constraint based at least in part on a covered subset for a particular one of the plurality of eligible extractive summaries comprising (a) a particular one of the plurality of interaction utterances that is in a particular one of the one or more per-party utterance subsets, and (b) additional ones of the plurality of interaction utterances, each of which is in any per-party utterance subset of the one or more per-party utterance subsets other than the particular one of the one or more per-party utterance subsets having a threshold-satisfying utterance similarity measure with respect to the particular one of the plurality of interaction utterances, and
(ii) each of the respective overall summary utility measures is based at least in part on an information quality measure and a linguistic quality measure for each of the plurality of interaction utterances in the covered subset for the particular one of the plurality of eligible extractive summaries; and
initiate the performance of one or more summary-based actions based at least in part on the optimal extractive summary.
10. The apparatus of claim 9 , wherein the one or more optimization constraints further comprise a content word coverage constraint requiring that, in an instance in which the covered subset for the particular one of the plurality of eligible extractive summaries comprises a particular interaction utterance having one or more per-utterance content words, then the covered subset for the particular one of the plurality of eligible extractive summaries comprises all of the one or more per-utterance content words.
11. The apparatus of claim 9 , wherein the one or more optimization constraints further comprise an utterance coverage constraint requiring that, in an instance in which the covered subset for the particular one of the plurality of eligible extractive summaries comprises a particular covered content word, the covered subset for the particular one of the plurality of eligible extractive summaries further comprises at least one interaction utterance of the plurality of interaction utterances that is associated with the particular covered content word.
12. The apparatus of claim 9 , wherein the one or more optimization constraints further comprise a party coverage constraint requiring that each covered subset for an eligible extractive summary comprises at least one interaction utterance from each of the one or more per-party utterance subsets.
13. The apparatus of claim 9 , wherein the one or more optimization constraints further comprise a party size constraint requiring that each covered subset for an eligible extractive summary comprises a below-threshold count of interaction utterances from each of the one or more per-party utterance subsets.
14. The apparatus of claim 9 further caused to generate the information quality measure by:
generating a graph representation of one or more per-utterance content words that are associated with each of the plurality of interaction utterances in the covered subset for the particular one of the plurality of eligible extractive summaries;
for each per-utterance content word of one or more per-utterance content words that is associated with each of the plurality of interaction utterances in the covered subset for the particular one of the plurality of eligible extractive summaries:
generating a neighboring subset of the one or more per-utterance content words based at least in part on the graph representation, and
generating a per-word the information quality measure based at least in part on each per-utterance content word of the one or more per-utterance content words that is in the neighboring subset of the per-utterance content word; and
generating the information quality measure based at least in part on each per-word information quality measure for a per-utterance content word of the one or more per-utterance content words that is associated with each of the plurality of interaction utterances in the covered subset for the particular one of the plurality of eligible extractive summaries.
15. The apparatus of claim 9 further caused to generate the linguistic quality measure by generating a probability aggregation measure based at least in part on a per-sequence probability for each n-gram word sequence of a set of n-gram word sequences that are associated with each of the plurality of interaction utterances in the covered subset for the particular one of the plurality of eligible extractive summaries.
16. The apparatus of claim 15 , where the set of n-gram word sequences comprises a set of trigram word sequences.
17. A computer program product comprising at least one non-transitory computer-readable storage medium having computer-readable program code portions stored therein, the computer-readable program code portions configured to:
determine one or more per-party utterance subsets from multi-party interaction transcript data, wherein: (1) the multi-party interaction transcript data comprises a plurality of interaction utterances associated with a plurality of interaction parties, and (2) each of the one or more per-party utterance subsets comprises one or more of the plurality of interaction utterances that are associated with one of the plurality of interaction parties;
generate, using an unsupervised machine learning model, an optimal extractive summary, wherein the optimal extractive summary comprises one of a plurality of eligible extractive summaries that is selected based at least in part on respective overall summary utility measures for each of the plurality of eligible extractive summaries, wherein:
(i) each of the plurality of eligible extractive summaries comprises a covered subset of the plurality of interaction utterances that complies with one or more optimization constraints, and the one or more optimization constraints comprise a similarity-based optimization constraint based at least in part on a covered subset for a particular one of the p lurality of eligible extractive summaries comprising (a) a particular one of the plurality of interaction utterances that is in a particular one of the one or more per-party utterance subsets, and (b) additional ones of the plurality of interaction utterances, each of which is in any per-party utterance subset of the one or more per-party utterance subsets other than the particular one of the one or more per-party utterance subsets having a threshold-satisfying utterance similarity measure with respect to the particular one of the plurality of interaction utterances, and
(ii) each of the respective overall summary utility measures is based at least in part on an information quality measure and a linguistic quality measure for each of the plurality of interaction utterances in the covered subset for the particular one of the plurality of eligible extractive summaries; and
initiate the performance of one or more summary-based actions based at least in part on the optimal extractive summary.
18. The computer program product of claim 17 , wherein the one or more optimization constraints further comprise a content word coverage constraint requiring that, in an instance in which the covered subset for the particular one of the plurality of eligible extractive summaries comprises a particular interaction utterance having one or more per-utterance content words, then the covered subset for the particular one of the plurality of eligible extractive summaries comprises all of the one or more per-utterance content words.