deeply analyse the features

The initial result of the coreference resolution has been got. But with the F-scores low leve, the problems appeared.

I believed that the precision was very high than the recall. The reasons were two. Firstly, the chosen twenty-four features were not good enough. There were some co-features had only been used inexplicit. They could be changed to be clearly. Secondly, the used research on my same problem displayed that the proportion of the positive and the negative was nearly 1:3. However my samples were in the proportion of 1:10.

The following task was to solve them.