search

DoodleJZ / HPSG-Neural-Parser

Source code for "Head-Driven Phrase Structure Grammar Parsing on Penn Treebank" published at ACL 2019
https://arxiv.org/abs/1907.02684
MIT License
107 stars 25 forks source link

Auto Tags Issue #10

Closed hankcs closed 4 years ago

hankcs commented 4 years ago

Hi, I'm a little bit confused if 02-21.10way.clean is auto tagged, how did you measure the constituency parsing score? I checked the 3th and 4th columns are the same. If they are auto tagged then you don't have gold truth for constituency parsing, otherwise the dependency parsing score is biased.

DoodleJZ commented 4 years ago

Which Tag do you mean? The POS tag or the label of constituency parsing?

DoodleJZ commented 4 years ago

I use the POS tags in 02-21.10way.clean which are not golden tag.

hankcs commented 4 years ago

I mean the POS tags. Yes, you mentioned the POS accuracy is 97.3%, so I assume the data contains no gold tag. Then could you explain how did you measure the acc for constituency parsing without knowing gold POS tags? In my understanding, the labels of constituency parsing are the superset of POS tags, as the labels to terminals are just the POS tags for them.

DoodleJZ commented 4 years ago

The acc of constituency parsing only measure the labels of constituency parsing without POS tags, otherwise only the label of internal node in the constituency tree.

hankcs commented 4 years ago

I see. Thank you for your clarification.