LEAR
The implementation our EMNLP 2021 paper "Enhanced Language Representation with Label Knowledge for Span Extraction".
See below for an overview of the model architecture:
Requestments
- python 3.6
- pytorch 1.8
- transformers 3.9.2
- prefetch_generator
- tokenizers
Data
Note: The thunlp
has updated the repo HMEAE recently, which causing the mismatch of data. Make sure you use the earlier version for ED task.
Task | Url | Rreprocess scripts |
---|---|---|
NER | mrc-for-flat-nested-ner | - |
Event Detection | ACE 2005 | HMEAE |
Train
The details of hyper-parameter setting is listed in our paper - A.3.
- For NER task, run
run_ner.py
. - For Event Detection task, run
run_trigger_extraction.py
Evaluation
- Nested: set
exist_nested
=True. - Flat: set
span_decode_strategy
=v5.