kornosk / log-odds-ratio Goto Github PK

I came across this repository after reading your NAACL '21 paper. It seems to be a simple method that captures the important words in the text dataset. But I have questions in two aspects: theory and usage.

Theory

• Does the log-odds-ratio method constrain the number of classes in the dataset? From what I could see, it would be difficult to apply this approach if I have a dataset with more than 3 classes (or equivalently, this method is only applicable to 2/3 class dataset).
• In your NAACL '21 paper, you mentioned the following. But I could not see what the score is compared to when you refer to "higher" and "lower". Is there a pre-specified threshold?

A higher score indicates more significance .... A lower score means ....

• In your NAACL '21 paper, you mentioned the following. What is this "sensitivity analysis"? Did you try to increase k to investigate its influence on the classification performance?

Based on a sensitivity analysis, we set k=10 to extract the top-10 significant words...

Usage

• The format of corpus i, corpus j, and background corpus is not provided in the README.md. Are they simply sentences separated by \n or some special format is required?