Problem with Binary classification

[navid5792]

If we make the fine_grain False, it performs the binary classification. But with binary, I can see there are three labels 0, 1, 2. So is it really a binary task?

[ttpro1995]

because in non-root node (which is part of the sentence), it is OK to be neutral (so does the dataset https://nlp.stanford.edu/sentiment/treebank.html)
We only filter out neutral sentence in Binary Task, but keep full sentence with positive, negative sentiment (Most of sentences have a neutral tree span).
So, the label 1 is necessary for the non-root node. At root node, it will ignore label 1 and consider whether 0 or 2 is more "score" to decide as sentiment of the sentence.

Look at this sample, you see the root is positive, but some others non-root node is neutral.

[navid5792]

ok, I get it now. You can have a softmax over 3 classes in the internal nodes of the tree but at the root, rather than keeping the softmax and overlooking the neutral class, it is possible to have a sigmoid layer and BCELoss as the loss function.