Machine Learning and Word Sense Disambiguation in Biomedicine
Author Information
Author(s): Xu Hua, Markatou Marianthi, Dimova Rositsa, Liu Hongfang, Friedman Carol
Primary Institution: Columbia University
Hypothesis
The study investigates how sample size, sense distribution, and degree of difficulty affect the performance of machine learning classifiers in word sense disambiguation.
Conclusion
The study concludes that increasing sample size generally reduces error rates, but this effect is limited in difficult cases, and the sense distribution has minimal impact when senses are separable.
Supporting Evidence
- Increasing sample size generally reduces error rates, especially for well-separated senses.
- The sense distribution did not significantly affect performance when senses were separable.
- When a majority sense exceeds 90%, using the majority sense is as effective as machine learning methods.
Takeaway
This study shows that using more examples usually helps computers understand words better, but sometimes it's not enough if the words are too similar.
Methodology
The study used Support Vector Machine classifiers on datasets of ambiguous biomedical abbreviations to measure error rates under different conditions.
Potential Biases
The study acknowledges potential biases in the automated generation of training data and the selection of ambiguous terms.
Limitations
The experiments were limited to only four abbreviations and a small number of replications, which may affect the generalizability of the results.
Statistical Information
P-Value
p<0.05
Statistical Significance
p<0.05
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website