Natural language processing techniques for document summarization using local and corpus-wide inferences

1. A computer-implemented method for generating an extractive summarization for a document data object, the computer-implemented comprising:

identifying, by one or more processors, a plurality of utterances associated with the document data object;

for each utterance, by the one or more processors:

generating, using a cross-utterance attention machine learning model, an attention-based utterance representation, wherein the cross-utterance attention machine learning model is configured to: (i) for each utterance pair, generate a cross-utterance self-attention weight, and (ii) generate the attention-based utterance representation for the utterance based at least in part on each cross-utterance self-attention weight that is associated with the utterance,

generating, based at least in part on the attention-based utterance representation and an utterance-based document representation that is generated based at least in part on each attention-based utterance representation, a document-utterance similarity score for the utterance, and

generating, based at least in part on a local utterance correlation graph data object and the document-utterance similarity score for the utterance, an utterance score for the utterance, wherein each utterance correlation edge of the local utterance correlation graph data object corresponds to a respective utterance pair and is associated with an utterance correlation edge weight that is generated based at least in part on the cross-utterance self-attention weight for the respective utterance pair;

generating, by the one or more processors, the extractive summarization based at least in part on each utterance score; and

performing, by the one or more processors, one or more prediction-based actions based at least in part on each utterance score.

2. The computer-implemented method of claim 1 , wherein generating the extractive summarization based at least in part on each utterance score comprises:

identifying a plurality of candidate keywords associated with the document data object, wherein each candidate keyword is associated with a keyword-related token subset of a group of text tokens in the document data object;

for each text token of the group of text tokens, generating, using a cross-token attention machine learning model, an attention-based token representation, wherein the cross-token attention machine learning model is further configured to generate: (i) for each token pair, a cross-token self-attention weight, and (ii) generate the attention-based token representation for the text token based at least in part on each cross-token self-attention weight that is associated with the text token;

for each candidate keyword:

generating, based at least in part on each attention-based token representation for the keyword-related token subset for the candidate keyword, a token-based keyword representation for the candidate keyword,

generating, based at least in part on the token-based keyword representation and a token-based document representation that is generated based at least in part on each attention-based token representation, a document-keyword similarity score for the candidate keyword, and

generating, based at least in part on a local keyword correlation graph data object and the document-keyword similarity score for the candidate keyword, a keyword score for the candidate keyword, wherein each keyword correlation edge of the local keyword correlation graph data object corresponds to a respective candidate keyword pair and is associated with a keyword correlation edge weight that is generated based at least in part on each cross-token self-attention weight for the respective candidate keyword pair; and

generating the extractive summarization based at least in part on each utterance score and each keyword score.

3. The computer-implemented method of claim 2 , wherein:

the document data object is a multi-party communication transcript data object that is associated with a plurality of party profiles, and

generating the extractive summarization based at least in part on each utterance score and each keyword score further comprises:

generating, using an integer linear programming joint keyword-utterance optimization model and based at least in part on each utterance score and each keyword score, a selected utterance subset of the plurality of utterances and a selected keyword subset of the plurality of candidate keywords, wherein: (i) the selected utterance subset comprises a party utterance summary for each party profile of the plurality of party profiles, (ii) the integer linear programming joint keyword-utterance optimization model is configured to generate the selected utterance subset and the selected keyword subset to maximize a joint keyword-utterance score for the selected utterance subset and the selected keyword subset given one or more integer linear programming optimization constraints, and (iii) the one or more integer linear programming optimization constraints comprise a party utterance summary length constraint requiring that each party utterance summary satisfies an upper-bound party utterance summary length threshold, and

generating the extractive summarization based at least in part on the selected utterance subset and the selected keyword subset.

4. The computer-implemented method of claim 3 , wherein the one or more integer linear programming optimization constraints comprise a keyword-based utterance coverage constraint requiring that, if the selected keyword subset comprises a particular candidate keyword, the selected utterance subset comprises at least one utterance that comprises the particular candidate keyword.

5. The computer-implemented method of claim 3 , wherein the one or more integer linear programming optimization constraints comprise an utterance-based keyword coverage constraint requiring that, if the selected utterance subset comprises a particular candidate keyword, the selected keyword subset comprises the particular candidate keyword.

6. The computer-implemented method of claim 3 , wherein the one or more integer linear programming optimization constraints comprise an utterance non-emptiness constraint requiring that, for each party profile of the plurality of party profiles, the selected utterance subset comprises at least one utterance related to the party profile.

7. The computer-implemented method of claim 3 , wherein the one or more integer linear programming optimization constraints comprise a pairwise utterance selection constraint requiring that, if a pairwise utterance similarity score for a cross-party utterance pair comprising a first utterance from a first party profile and a second utterance from a different party profile satisfies a lower-bound pairwise utterance similarity threshold, the selected utterance subset comprises both the first utterance and the second utterance.

8. The computer-implemented method of claim 7 , wherein the lower-bound pairwise utterance similarity threshold is generated based at least in part on a deviation measure between: (i) a maximal pairwise utterance similarity score for all cross-party utterance pairs, and (ii) a predefined maximal pairwise utterance similarity score deviation threshold.

9. The computer-implemented method of claim 3 , wherein the one or more integer linear programming optimization constraints comprise a keyword summary length constraint requiring that a selected keyword count of the selected keyword subset satisfies an upper-bound keyword selection count threshold.

10. An apparatus for generating an extractive summarization for a document data object, the apparatus comprising one or more processors and at least one memory including program code, the at least one memory and the program code configured to, with the one or more processors, cause the apparatus to:

identify a plurality of utterances associated with the document data object;

for each utterance:

generate, using a cross-utterance attention machine learning model, an attention-based utterance representation, wherein the cross-utterance attention machine learning model is configured to: (i) for each utterance pair, generate a cross-utterance self-attention weight, and (ii) generate the attention-based utterance representation for the utterance based at least in part on each cross-utterance self-attention weight that is associated with the utterance,

generate, based at least in part on the attention-based utterance representation and an utterance-based document representation that is generated based at least in part on each attention-based utterance representation, a document-utterance similarity score for the utterance, and

generate, based at least in part on a local utterance correlation graph data object and the document-utterance similarity score for the utterance, an utterance score for the utterance, wherein each utterance correlation edge of the local utterance correlation graph data object corresponds to a respective utterance pair and is associated with an utterance correlation edge weight that is generated based at least in part on the cross-utterance self-attention weight for the respective utterance pair;

generate the extractive summarization based at least in part on each utterance score; and

perform one or more prediction-based actions based at least in part on each utterance score.

11. The apparatus of claim 10 , wherein generating the extractive summarization based at least in part on each utterance score comprises:

identifying a plurality of candidate keywords associated with the document data object, wherein each candidate keyword is associated with a keyword-related token subset of a group of text tokens in the document data object;

for each text token of the group of text tokens, generating, using a cross-token attention machine learning model, an attention-based token representation, wherein the cross-token attention machine learning model is further configured to generate: (i) for each token pair, a cross-token self-attention weight, and

(ii) generate the attention-based token representation for the text token based at least in part on each cross-token self-attention weight that is associated with the text token;

for each candidate keyword:

generating, based at least in part on each attention-based token representation for the keyword-related token subset for the candidate keyword, a token-based keyword representation for the candidate keyword,

generating, based at least in part on the token-based keyword representation and a token-based document representation that is generated based at least in part on each attention-based token representation, a document-keyword similarity score for the candidate keyword, and

generating, based at least in part on a local keyword correlation graph data object and the document-keyword similarity score for the candidate keyword, a keyword score for the candidate keyword, wherein each keyword correlation edge of the local keyword correlation graph data object corresponds to a respective candidate keyword pair and is associated with a keyword correlation edge weight that is generated based at least in part on each cross-token self-attention weight for the respective candidate keyword pair; and

generating the extractive summarization based at least in part on each utterance score and each keyword score.

12. The apparatus of claim 11 , wherein:

the document data object is a multi-party communication transcript data object that is associated with a plurality of party profiles, and

generating the extractive summarization based at least in part on each utterance score and each keyword score further comprises:

generating, using an integer linear programming joint keyword-utterance optimization model and based at least in part on each utterance score and each keyword score, a selected utterance subset of the plurality of utterances and a selected keyword subset of the plurality of candidate keywords, wherein: (i) the selected utterance subset comprises a party utterance summary for each party profile of the plurality of party profiles, (ii) the integer linear programming joint keyword-utterance optimization model is configured to generate the selected utterance subset and the selected keyword subset to maximize a joint keyword-utterance score for the selected utterance subset and the selected keyword subset given one or more integer linear programming optimization constraints, and (iii) the one or more integer linear programming optimization constraints comprise a party utterance summary length constraint requiring that each party utterance summary satisfies an upper-bound party utterance summary length threshold, and

generating the extractive summarization based at least in part on the selected utterance subset and the selected keyword subset.

13. The apparatus of claim 12 , wherein the one or more integer linear programming optimization constraints comprise a keyword-based utterance coverage constraint requiring that, if the selected keyword subset comprises a particular candidate keyword, the selected utterance subset comprises at least one utterance that comprises the particular candidate keyword.

14. The apparatus of claim 12 , wherein the one or more integer linear programming optimization constraints comprise an utterance-based keyword coverage constraint requiring that, if the selected utterance subset comprises a particular candidate keyword, the selected keyword subset comprises the particular candidate keyword.

15. The apparatus of claim 12 , wherein the one or more integer linear programming optimization constraints comprise an utterance non-emptiness constraint requiring that, for each party profile of the plurality of party profiles, the selected utterance subset comprises at least one utterance related to the party profile.

16. The apparatus of claim 12 , wherein the one or more integer linear programming optimization constraints comprise a pairwise utterance selection constraint requiring that, if a pairwise utterance similarity score for a cross-party utterance pair comprising a first utterance from a first party profile and a second utterance from a different party profile satisfies a lower-bound pairwise utterance similarity threshold, the selected utterance subset comprises both the first utterance and the second utterance.

17. The apparatus of claim 16 , wherein the lower-bound pairwise utterance similarity threshold is generated based at least in part on a deviation measure between: (i) a maximal pairwise utterance similarity score for all cross-party utterance pairs, and (ii) a predefined maximal pairwise utterance similarity score deviation threshold.

18. A computer program product for generating an extractive summarization for a document data object, the 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:

identify a plurality of utterances associated with the document data object;

for each utterance:

generate, using a cross-utterance attention machine learning model, an attention-based utterance representation, wherein the cross-utterance attention machine learning model is configured to: (i) for each utterance pair, generate a cross-utterance self-attention weight, and (ii) generate the attention-based utterance representation for the utterance based at least in part on each cross-utterance self-attention weight that is associated with the utterance,

generate, based at least in part on the attention-based utterance representation and an utterance-based document representation that is generated based at least in part on each attention-based utterance representation, a document-utterance similarity score for the utterance, and

generate, based at least in part on a local utterance correlation graph data object and the document-utterance similarity score for the utterance, an utterance score for the utterance, wherein each utterance correlation edge of the local utterance correlation graph data object corresponds to a respective utterance pair and is associated with an utterance correlation edge weight that is generated based at least in part on the cross-utterance self-attention weight for the respective utterance pair;

generate the extractive summarization based at least in part on each utterance score; and

perform one or more prediction-based actions based at least in part on each utterance score.

19. The computer program product of claim 18 , wherein generating the extractive summarization based at least in part on each utterance score comprises:

identifying a plurality of candidate keywords associated with the document data object, wherein each candidate keyword is associated with a keyword-related token subset of a group of text tokens in the document data object;

for each text token of the group of text tokens, generating, using a cross-token attention machine learning model, an attention-based token representation, wherein the cross-token attention machine learning model is further configured to generate: (i) for each token pair, a cross-token self-attention weight, and (ii) generate the attention-based token representation for the text token based at least in part on each cross-token self-attention weight that is associated with the text token;

for each candidate keyword:

generating, based at least in part on each attention-based token representation for the keyword-related token subset for the candidate keyword, a token-based keyword representation for the candidate keyword,

generating, based at least in part on the token-based keyword representation and a token-based document representation that is generated based at least in part on each attention-based token representation, a document-keyword similarity score for the candidate keyword, and

generating, based at least in part on a local keyword correlation graph data object and the document-keyword similarity score for the candidate keyword, a keyword score for the candidate keyword, wherein each keyword correlation edge of the local keyword correlation graph data object corresponds to a respective candidate keyword pair and is associated with a keyword correlation edge weight that is generated based at least in part on each cross-token self-attention weight for the respective candidate keyword pair; and

generating the extractive summarization based at least in part on each utterance score and each keyword score.

20. The computer program product of claim 19 , wherein:

the document data object is a multi-party communication transcript data object that is associated with a plurality of party profiles, and

generating the extractive summarization based at least in part on each utterance score and each keyword score further comprises:

generating, using an integer linear programming joint keyword-utterance optimization model and based at least in part on each utterance score and each keyword score, a selected utterance subset of the plurality of utterances and a selected keyword subset of the plurality of candidate keywords, wherein: (i) the selected utterance subset comprises a party utterance summary for each party profile of the plurality of party profiles, (ii) the integer linear programming joint keyword-utterance optimization model is configured to generate the selected utterance subset and the selected keyword subset to maximize a joint keyword-utterance score for the selected utterance subset and the selected keyword subset given one or more integer linear programming optimization constraints, and (iii) the one or more integer linear programming optimization constraints comprise a party utterance summary length constraint requiring that each party utterance summary satisfies an upper-bound party utterance summary length threshold, and

generating the extractive summarization based at least in part on the selected utterance subset and the selected keyword subset.