Ascribing ground truth performance to annotation blocks
Examples described herein provide a computer-implemented method that includes receiving a ground truth associated with a domain cartridge, the domain cartridge comprising a plurality of hierarchical layers. The method further includes analyzing annotation blocks in relation to data present in the ground truth to detect any errors in a set of natural language processing annotators. The analyzing includes computing a recall score, a precision score, and an F1 score for each annotation block in a lowest level layer of the plurality of hierarchical layers. The analyzing further includes determining whether an error is detected at the lowest level layer of the plurality of hierarchical layers based at least in part on the recall score, the precision score, and the F1 score. The analyzing further includes terminating the analyzing responsive to determining that the error is detected at the lowest level layer of the plurality of hierarchical layers.
1. A computer-implemented method for ascribing ground truth performance to annotation blocks, the method comprising:
receiving, by a processing device, a ground truth associated with a domain cartridge, the domain cartridge comprising a plurality of hierarchical layers, the plurality of hierarchical layers being generated by a set of natural language processing annotators; and
analyzing, by the processing device, annotation blocks in the plurality of hierarchical layers of the domain cartridge in relation to data present in the ground truth to detect any errors in the set of natural language processing annotators, wherein the analyzing comprises:
computing a recall score, a precision score, and an F1 score for each annotation block in a lowest level layer of the plurality of hierarchical layers;
determining whether an error is detected at the lowest level layer of the plurality of hierarchical layers based at least in part on the recall score, the precision score, and the F1 score; and
terminating the analyzing responsive to determining that the error is detected at the lowest level layer of the plurality of hierarchical layers,
wherein the plurality of hierarchical layers comprises:
a concept detection dictionary layer generated by a concept detection dictionary annotator using direct word matches;
an attribute set layer generated based on the concept detection dictionary layer;
an attribute values layer generated by a concept value annotator;
a negated true/false metadata layer generated by a negation annotator;
a hypothetical true/false metadata layer generated by a hypothetical annotator; and
a section metadata layer generated by a section annotator.
2. The computer-implemented method of claim 1 , wherein the analyzing further comprises:
repeating the analyzing at a next higher level layer of the plurality of hierarchical layers responsive to determining that the error is not detected at the lowest level layer of the plurality of hierarchical layers.
3. The computer-implemented method of claim 2 , wherein repeating the analyzing at a next higher level layer of the plurality of hierarchical layers is performed iteratively until it is determined that the error is detected.
4. The computer-implemented method of claim 1 , wherein the analyzing further comprises:
computing a total recall score, a total precision score, and a total F1 score for the domain cartridge across each of the plurality of hierarchical layers.
5. The computer-implemented method of claim 1 , wherein the detected error is a false positive.
6. The computer-implemented method of claim 1 , wherein the detected error is a false negative.
7. The computer-implemented method of claim 1 , wherein the plurality of hierarchical layers are each generated by a different natural language processing annotator of the set of natural language processing annotators.
8. A system comprising:
a memory comprising computer readable instructions; and
a processing device for executing the computer readable instructions, the computer readable instructions controlling the processing device to perform operations comprising:
receiving, by the processing device, a ground truth associated with a domain cartridge, the domain cartridge comprising a plurality of hierarchical layers, the plurality of hierarchical layers being generated by a set of natural language processing annotators; and
analyzing, by the processing device, annotation blocks in the plurality of hierarchical layers of the domain cartridge in relation to data present in the ground truth to detect any errors in the set of natural language processing annotators, wherein the analyzing comprises:
computing a recall score, a precision score, and an F1 score for each annotation block in a lowest level layer of the plurality of hierarchical layers;
determining whether an error is detected at the lowest level layer of the plurality of hierarchical layers based at least in part on the recall score, the precision score, and the F1 score; and
terminating the analyzing responsive to determining that the error is detected at the lowest level layer of the plurality of hierarchical layers,
wherein the plurality of hierarchical layers comprises:
a concept detection dictionary layer generated by a concept detection dictionary annotator using direct word matches;
an attribute set layer generated based on the concept detection dictionary layer;
an attribute values layer generated by a concept value annotator;
a negated true/false metadata layer generated by a negation annotator;
a hypothetical true/false metadata layer generated by a hypothetical annotator; and
a section metadata layer generated by a section annotator.
9. The system of claim 8 , wherein the analyzing further comprises:
repeating the analyzing at a next higher level layer of the plurality of hierarchical layers responsive to determining that the error is not detected at the lowest level layer of the plurality of hierarchical layers.
10. The system of claim 9 , wherein repeating the analyzing at a next higher level layer of the plurality of hierarchical layers is performed iteratively until it is determined that the error is detected.
11. The system of claim 8 , wherein the analyzing further comprises:
computing a total recall score, a total precision score, and a total F1 score for the domain cartridge across each of the plurality of hierarchical layers.
12. The system of claim 8 , wherein the detected error is a false positive.
13. The system of claim 8 , wherein the detected error is a false negative.
14. The system of claim 8 , wherein the plurality of hierarchical layers are each generated by a different natural language processing annotator of the set of natural language processing annotators.
15. A computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a processing device to cause the processing device to perform operations comprising:
receiving, by the processing device, a ground truth associated with a domain cartridge, the domain cartridge comprising a plurality of hierarchical layers, the plurality of hierarchical layers being generated by a set of natural language processing annotators; and
analyzing, by the processing device, annotation blocks in the plurality of hierarchical layers of the domain cartridge in relation to data present in the ground truth to detect any errors in the set of natural language processing annotators, wherein the analyzing comprises:
computing a recall score, a precision score, and an F1 score for each annotation block in a lowest level layer of the plurality of hierarchical layers;
determining whether an error is detected at the lowest level layer of the plurality of hierarchical layers based at least in part on the recall score, the precision score, and the F1 score; and
terminating the analyzing responsive to determining that the error is detected at the lowest level layer of the plurality of hierarchical layers,
wherein the plurality of hierarchical layers comprises:
a concept detection dictionary layer generated by a concept detection dictionary annotator using direct word matches;
an attribute set layer generated based on the concept detection dictionary layer;
an attribute values layer generated by a concept value annotator;
a negated true/false metadata layer generated by a negation annotator;
a hypothetical true/false metadata layer generated by a hypothetical annotator; and
a section metadata layer generated by a section annotator.
16. The computer program product of claim 15 , wherein the analyzing further comprises:
repeating the analyzing at a next higher level layer of the plurality of hierarchical layers responsive to determining that the error is not detected at the lowest level layer of the plurality of hierarchical layers.
17. The computer program product of claim 16 , wherein repeating the analyzing at a next higher level layer of the plurality of hierarchical layers is performed iteratively until it is determined that the error is detected.
18. The computer program product of claim 15 , wherein the analyzing further comprises:
computing a total recall score, a total precision score, and a total F1 score for the domain cartridge across each of the plurality of hierarchical layers.
19. The computer program product of claim 15 , wherein the detected error is a false positive.