Performance Tests on Other Test Sets
I. MEDLINE Title
- Informally tested to correct spelling errors in MEDLINE titles
- Test file includes 2585 MEDLINE titles with errors
- The preliminary results are similar to those on the consumer health questions
- Indicating that CSpell might potentially be used for other biomedical
spelling error corrections.
- Initial result for correction (10 instances):
Precision | Recall | F1
|
---|
1.00 | 0.60 | 0.75
|
II. CSpell Test Set on the important terms
- Data: Test Set
- Evaluated its performance on only those terms that are important for question understanding.
- The important terms were identified during manual annotation.
- The results on the important terms are very close to the overall results.
Tests | Precision | Recall | F1
|
---|
Non-word Detection | 0.8937 | 0.8259 | 0.8585
|
Non-word Correction | 0.7343 | 0.6786 | 0.7053
|
Real-word Detection | 0.9179 | 0.7422 | 0.8207
|
Real-word Correction | 0.7536 | 0.6094 | 0.6739
|