1. Improving Health Question Classification by Word Location Weights
Rey-Long Liu
Dept. of Medical Informatics
Tzu Chi University
Taiwan

2. Outline Background Problem definition The proposed approach: WLW
Empirical evaluation
Conclusion

3. Background

4. Categories of Health Questions

5. Classification of Health Questions
Why health questions?
Health questions provide both reliable and readable health information
Why classification of health questions?
Given a health question q, retrieve related questions (and their answers)

6. Problem Definition

7. Goal & Motivation Goal Motivation
Target: Chinese Health Questions (CHQs)
Contribution: Developing a technique WLW (Word Location Weight) that estimates the location weights of words in a CHQ based on their locations
Motivation
Location weights can be used by classifiers (e.g., SVM) to improve the classification
Classifying in-space CHQs (cause, diagnosis, process)
Filtering out-space CHQs (may be whatever)

8. Basic Idea
Those words that are more related to the category of a CHQ tend to appear at the beginning and end of the CHQ
Examples:
如何(how to)克服(deal with)緊張(nervous)的情緒(mood)?  process
嬰兒(infant)體溫(body temperature)太低(too low)怎麼辦(how to do)?  process

9. Related Work Recognition of question types (e.g., when, where)
Weakness: Types  Intended categories of CHQs
Classification by parsing
Weakness I: Parsing Chinese is still challenging
Weakness II: CHQs are NOT always well-formed
Classification by pattern matching
Weakness: Difficult to construct the string patterns

10. The Proposed Approach: WLW

11. Main Challenges
(1) Defining the two weights of a location p in a CHQ q

12. Main Challenges (cont.)
(2) Encoding the location weights of a word w into two features for the underlying classifier

13. Interesting Behaviors of WLW
A word w in a question q has two features
Fvaluefront and Fvaluerear
Applicable to different categories and languages (e.g., English)
When w is far from the front and the rear
Both features reduce to the term frequency (TF) of w
WLW reduces to traditional feature-encoding approach (using TF as the features)

14. Empirical Evaluation

15. Experimental Design
CHQs were downloaded from a health information provider
864 in-space CHQs
cause (category 1): 313
diagnosis (category 2): 92
process (category 3): 459
100 out-space CHQs
whatever (general description)
Five-fold cross validation

16. Underlying Classifiers
The Support Vector Machine (SVM) classifier

17. Results: Classification of In-Space CHQs
Evaluation criteria
Micro-averaged F1 (MicroF1)
Macro-averaged F1 (MacroF1)

18. SVM+WLW is significantly better than SVM

19. Results: Filtering of Out-Space CHQs
Evaluation criteria
Filtering ratio (FR) =
# out-space CHQs successfully rejected by all categories / # out-space CHQs
Average number of misclassifications (AM) =
# misclassifications for the out-space CHQs / # out-space CHQs

20. SVM+WLW achieves higher FR and lower AM

21. Conclusion

22. Healthcare consumers often read health information on the Internet
Health questions as the valuable resources for healthcare consumers
Providing both reliable and readable health information
Classification of health questions is basis for the retrieval of related questions
cause, diagnosis, process, whatever
WLW can help SVM to improve the classification of CHQs