lowinli / metric4coref Goto Github PK

Hi~

Appreciate so much for your excellent work! However, since it is quite different from the original paper, A Model-Theoretic Coreference Scoring Scheme, I wonder which implementation or paper did you refer to when realizing the MUC metric? After I checked several test cases, it seems that your implementation of the MUC metric is more critical than the paper I mentioned.

Hi, thank you for this great tool and the elaborated explanation of those metrics. :)

Just a kind reminder that in the calculation of f1, it is possible for both precision and recall to be zero and thus it might give an error as the denominator might be zero. The same for precision and recall, too.

Many thanks again!

Hi again, I rechecked the code and I found there was a tiny problem. After summing up all the precisions and recalls of all mentions, the value should be normalized by the number of mentions either in the predicted clusters (for precision) or in the gold clusters (for recall), but you the number you used is simply the number of the intersection between two clusters, which makes the denominator incorrect.

Regarding the MUC, I commented on another closed issue.

Hope it makes sense. :)

Hello

Thanks so much for making this available in python - it's really useful!

I think there might be a bit of a problem with the MUC measure - it gives different results depending on the order of the cluster elements.

For example:

predict_clusters = [["a", "b", "c"], ["d", "e", "f", "g"], ["h", "i", "j"], ["k"]]
predict_clusters_shuffled = [["c", "b", "a"], ["e", "d", "g", "f"], ["j", "i", "h"], ["k"]]
gold_clusters = [["a", "b", "d"], ["c", "e", "f", "g"], ["h", "i", "j", "k"]]

print(metric4coref.muc(predict_clusters, gold_clusters))
>>> (0.5833333333333334, 0.4666666666666667, 0.5185185185185186)

print(metric4coref.muc(predict_clusters_shuffled, gold_clusters))
>>> (0.16666666666666666, 0.13333333333333333, 0.14814814814814814)

This doesn't seem to be a problem with any of the other measures.

For now, I'm just sorting all my clusters before I evaluate them, but I just wanted to flag!

Thanks,
Emily