Automated generation of fine-grained call reasons from customer service call transcripts

1. A computer-implemented method for automatically generating fine-grained call reasons from customer service call transcripts, the computer-implemented method comprising:

extracting, by an event extraction system using a natural language processing (NLP) technique, a set of events from a set of text strings of speaker turns, wherein said extracting comprises:

feeding, by the event extraction system, the set of text strings of speaker turns to a logistic regression classifier trained to generate a set of probability values comprising a respective probability value for each speaker turn in the set of text strings of speaker turns, wherein each probability value in the set of probability values indicates a likelihood that a respective speaker turn contains natural language expressing a respective call reason; and

extracting, by the event extraction system, each event in the set of events from the respective speaker turn in a respective text string of speaker turns having a highest-ranking respective probability value;

identifying, by a cluster generation system, a set of clusters of events based on the set of events;

labeling, by the cluster generation system, each cluster of events in the set of clusters of events to generate a set of labeled clusters of events; and

assigning, by a cluster assignment system, each event in the set of events to a respective labeled cluster of events in the set of labeled clusters of events.

2. The computer-implemented method of claim 1 , further comprising:

feeding, by the cluster generation system, the set of events to a role factor network trained to generate a set of event embeddings that capture (i) distributional information associated with the set of events and (ii) interactions between arguments within the set of events; and

identifying, by the cluster generation system, the set of clusters of events based on the set of event embeddings.

3. The computer-implemented method of claim 1 , further comprising labeling, by the cluster generation system, each cluster of events using a graph-based sentence compression algorithm that generalizes lexical variations in the events of each cluster of events.

4. The computer-implemented method of claim 1 , further comprising storing, by the cluster generation system, the set of labeled clusters of events in a cluster database.

5. The computer-implemented method of claim 1 , wherein each event in the set of events comprises a constituent phrase comprising a subject, a verb, and an object.

6. The computer-implemented method of claim 1 , further comprising:

preprocessing, by a speaker turn preprocessing system, the set of text strings of speaker turns to at least one of clean, autopunctuate, or resolve co-references in the set of text strings of speaker turns before extracting the set of events from the set of text strings of speaker turns.

7. A non-transitory computer readable medium including instructions for causing a processor to perform operations for automatically generating fine-grained call reasons from customer service call transcripts, the operations comprising:

extracting, using a natural language processing (NLP) technique, a set of events from a set of text strings of speaker turns, wherein said extracting comprises:

feeding the set of text strings of speaker turns to a logistic regression classifier trained to generate a set of probability values comprising a respective probability value for each speaker turn in the set of text strings of speaker turns, wherein each probability value in the set of probability values indicates a likelihood that a respective speaker turn contains natural language expressing a respective call reason; and

extracting each event in the set of events from the respective speaker turn in a respective text string of speaker turns having a highest-ranking respective probability value;

identifying a set of clusters of events based on the set of events;

labeling each cluster of events in the set of clusters of events to generate a set of labeled clusters of events; and

assigning each event in the set of events to a respective labeled cluster of events in the set of labeled clusters of events.

8. The non-transitory computer readable medium of claim 7 , wherein the operations further comprise:

feeding the set of events to a role factor network trained to generate a set of event embeddings that capture (i) distributional information associated with the set of events and (ii) interactions between arguments within the set of events; and

identifying the set of clusters of events based on the set of event embeddings.

9. The non-transitory computer readable medium of claim 7 , wherein the operations further comprise labeling each cluster of events using a graph-based sentence compression algorithm that generalizes lexical variations in the events of each cluster of events.

10. The non-transitory computer readable medium of claim 7 , wherein the operations further comprise storing the set of labeled clusters of events in a cluster database.

11. The non-transitory computer readable medium of claim 7 , wherein each event in the set of events comprises a constituent phrase comprising a subject, a verb, and an object.

12. The non-transitory computer readable medium of claim 7 , the operations further comprising:

preprocessing the set of text strings of speaker turns to at least one of clean, autopunctuate, or resolve co-references in the set of text strings of speaker turns before extracting the set of events from the set of text strings of speaker turns.

13. A computing system for automatically generating fine-grained call reasons from customer service call transcripts, the computing system comprising:

a storage unit configured to store instructions;

a control unit, coupled to the storage unit, configured to process the stored instructions to:

extract, using a natural language processing (NLP) technique, a set of events from a set of text strings of speaker turns, wherein, to extract the set of events, the control unit is further configured to:

feed the set of text strings of speaker turns to a logistic regression classifier trained to generate a set of probability values comprising a respective probability value for each speaker turn in the set of text strings of speaker turns, wherein each probability value in the set of probability values indicates a likelihood that a respective speaker turn contains natural language expressing a respective call reason; and

extract each event in the set of events from the respective speaker turn in a respective text string of speaker turns having a highest-ranking respective probability value;

identify a set of clusters of events based on the set of events;

label each cluster of events in the set of clusters of events to generate a set of labeled clusters of events; and

assign each event in the set of events to a respective labeled cluster of events in the set of labeled clusters of events.

14. The computing system of claim 13 , wherein the control unit is further configured to process the stored instructions to:

feed the set of events to a role factor network trained to generate a set of event embeddings that capture (i) distributional information associated with the set of events and (ii) interactions between arguments within the set of events; and

identify the set of clusters of events based on the set of event embeddings.

15. The computing system of claim 13 , wherein the control unit is further configured to process the stored instructions to label each cluster of events using a graph-based sentence compression algorithm that generalizes lexical variations in the events of each cluster of events.

16. The computing system of claim 13 , wherein the control unit is further configured to process the stored instructions to store the set of labeled clusters of events in a cluster database.

17. The computing system of claim 13 , wherein the control unit is further configured to:

preprocess the set of text strings of speaker turns to at least one of clean, autopunctuate, or resolve co-references in the set of text strings of speaker turns, and wherein, to extract the set of events from the set of text strings of speaker turns, the control unit is configured to extract the set of events from the preprocessed set of text strings of speaker turns.