chanind / frame-semantic-transformer

Frame Semantic Parser based on T5 and FrameNet
https://chanind.github.io/frame-semantic-transformer
MIT License
51 stars 10 forks source link

Particularly low accuracy with passive sentences without complements #29

Open fatihbozdag opened 11 months ago

fatihbozdag commented 11 months ago

Greetings all,

Has anyone noticed particularly low accuracy with passive sentences? Also, the parser acts strangely in those cases.

For instance: Great results with

The participants' answers were looked at for general trends that might identify the differences in these emotions. 
{'Ground': "The participants'answers", 'Verb': 'look', 'Phenomenon': 'at for general trends that might identify the differences in these emotions'}

Also great with:

for example, specific words and phrases were looked for their accuracy.
{'Ground': "specific words and phrases", 'Verb': 'look', 'Phenomenon': 'for their accuracy'}
However: 
specific words and phrases were looked for
{'Sought_entity': 'for', 'Verb': 'look'}

P.S. I use my custom script to replace trigger_location with the actual verb, but this is handled after frame annotation.