Evaluation framework for intent authoring processes
Evaluating intent authoring processes, by a processor in a computing environment. Results are received of a simulated intent labeling effort of a dataset comprising utterances of interactive dialog sessions between agents and clients for a given product or service. Figures of merits for respective algorithms used to perform the simulated intent labeling effort are computed. Each of the respective algorithms are evaluated according to the computed figures of merits; and one of the respective algorithms is implemented for labeling intents of a remaining corpus of the synthesized dataset according to parameters evaluated in the computed figures of merits.
1. A method for evaluating intent authoring processes, by a processor, comprising:
receiving results of a simulated intent labeling effort of a dataset comprising utterances of interactive dialog sessions between agents and clients for a given product or service, wherein a plurality of snapshots of an intent state are generated during the simulated intent labeling effort;
computing figures of merits for respective algorithms used to perform the simulated intent labeling effort;
evaluating each of the respective algorithms according to the computed figures of merits; and
implementing one of the respective algorithms for labeling intents of a remaining corpus of the dataset according to parameters evaluated in the computed figures of merits.
2. The method of claim 1 , wherein the plurality of snapshots include:
an initial state snapshot taken during initiation of the simulated labeling effort;
a terminal state snapshot taken when a bounding constraint terminates the simulated labeling effort; and
a golden snapshot taken when all utterances in the corpus relevant to a target intent have been included in the simulated intent labeling effort.
3. The method of claim 1 , further including withholding at least a portion of the dataset for computing a recall rate of a classifier trained under one of the respective algorithms.
4. The method of claim 2 , wherein the implemented one of the respective algorithms comprises the algorithm which produces the classifier having an accuracy determined in the terminal state snapshot which substantially matches an accuracy determined in the golden snapshot.
5. The method of claim 2 , wherein the figures of merits are selected from a group consisting of a classifier accuracy, a linguistic diversity, and a data complexity measure of separable intent to off-intent utterances.
6. The method of claim 5 , wherein:
the linguistic diversity is determined by measuring a count of unique words in a respective one of the plurality of snapshots; and
the data complexity measure is determined by computing a minimum spanning tree (MST) in the respective one of the plurality of snapshots and counting a number of edges in the MST which connect a respective in-intent utterance to a respective off-intent utterance.
7. A system for evaluating intent authoring processes, comprising:
a processor executing instructions stored in a memory device; wherein the processor:
receives results of a simulated intent labeling effort of a dataset comprising utterances of interactive dialog sessions between agents and clients for a given product or service, wherein a plurality of snapshots of an intent state are generated during the simulated intent labeling effort;
computes figures of merits for respective algorithms used to perform the simulated intent labeling effort;
evaluates each of the respective algorithms according to the computed figures of merits; and
implements one of the respective algorithms for labeling intents of a remaining corpus of the synthesized dataset according to parameters evaluated in the computed figures of merits.
8. The system of claim 7 , wherein the plurality of snapshots include:
an initial state snapshot taken during initiation of the simulated labeling effort;
a terminal state snapshot taken when a bounding constraint terminates the simulated labeling effort; and
a golden snapshot taken when all utterances in the corpus relevant to a target intent have been included in the simulated intent labeling effort.
9. The system of claim 7 , wherein the processor withholds at least a portion of the dataset for computing a recall rate of a classifier trained under one of the respective algorithms.
10. The system of claim 9 , wherein the implemented one of the respective algorithms comprises the algorithm which produces the classifier having an accuracy determined in the terminal state snapshot which substantially matches an accuracy determined in the golden snapshot.
11. The system of claim 9 , wherein the figures of merits are selected from a group consisting of a classifier accuracy, a linguistic diversity, and a data complexity measure of separable intent to off-intent utterances.
12. The system of claim 11 , wherein:
the linguistic diversity is determined by measuring a count of unique words in a respective one of the plurality of snapshots; and
the data complexity measure is determined by computing a minimum spanning tree (MST) in the respective one of the plurality of snapshots and counting a number of edges in the MST which connect a respective in-intent utterance to a respective off-intent utterance.
13. A computer program product for evaluating intent authoring processes, the computer program product comprising a non-transitory computer-readable storage medium having computer-readable program code portions stored therein, the computer-readable program code portions comprising:
an executable portion that receives results of a simulated intent labeling effort of a dataset comprising utterances of interactive dialog sessions between agents and clients for a given product or service, wherein a plurality of snapshots of an intent state are generated during the simulated intent labeling effort;
an executable portion that computes figures of merits for respective algorithms used to perform the simulated intent labeling effort;
an executable portion that evaluates each of the respective algorithms according to the computed figures of merits; and
an executable portion that implements one of the respective algorithms for labeling intents of a remaining corpus of the synthesized dataset according to parameters evaluated in the computed figures of merits.
14. The computer program product of claim 13 , wherein the plurality of snapshots include:
an initial state snapshot taken during initiation of the simulated labeling effort;
a terminal state snapshot taken when a bounding constraint terminates the simulated labeling effort; and
a golden snapshot taken when all utterances in the corpus relevant to a target intent have been included in the simulated intent labeling effort.
15. The computer program product of claim 13 , further including an executable portion that withholds at least a portion of the dataset for computing a recall rate of a classifier trained under one of the respective algorithms.
16. The computer program product of claim 14 , wherein the implemented one of the respective algorithms comprises the algorithm which produces the classifier having an accuracy determined in the terminal state snapshot which substantially matches an accuracy determined in the golden snapshot.
17. The computer program product of claim 14 , wherein the figures of merits are selected from a group consisting of a classifier accuracy, a linguistic diversity, and a data complexity measure of separable intent to off-intent utterances.
18. The computer program product of claim 17 , wherein:
the linguistic diversity is determined by measuring a count of unique words in a respective one of the plurality of snapshots; and
the data complexity measure is determined by computing a minimum spanning tree (MST) in the respective one of the plurality of snapshots and counting a number of edges in the MST which connect a respective in-intent utterance to a respective off-intent utterance.