Questions on BobcatParser

[chin-jey]

Hi,
I have 2 questions on BobCatParser

• what is the use of Bert ? I see you have a parameter to specify a pre-trained model for huggingface, I'm not sure I understand what you do with it
• what is the purpose of suppressing exceptions and appending a None when calling sentences2diagrams ? Why don't you discard the diagram ?

Best Regards,
C.

[ianyfan]

what is the use of Bert ? I see you have a parameter to specify a pre-trained model for huggingface, I'm not sure I understand what you do with it

Bobcat parses a sentence by producing a bunch of candidate trees and then picking the tree with the best score. These scores are predicted by a BERT-based model specifically designed for this purpose.

what is the purpose of suppressing exceptions and appending a None when calling sentences2diagrams ? Why don't you discard the diagram ?

It makes it easier to match up each diagram in the list with the entry in the input (in case you have additional information attached to each entry). It's simple to discard the failed parses after the fact, by running filter(None, diagrams), but not so simple the other way round.

Please let me know if you have any more questions.

[chin-jey]

How is the score of the tree defined ?

Also, I'm trying to use TreeReader with Bobcat as a ccg_parser but I got an exception I don't understand

CCGRuleUseError: Illegal use of UNK: unknown CCG rule.

it's in general not clear to me in which cases grammar-aware readers fail.

[chin-jey]

I think I figured out the TreeReader problem. I'm still left with the doubt of the score.

[ianyfan]

The probability of the tree is determined by neural tagger as well as the probabilities of the subtrees:

prob(tree) = (prob from neural tagger) * product(probs of subtrees)

In the code, the score is the log probability, so it is implemented as:

score(tree) = (score from neural tagger) + sum(scores of subtrees)