dialogueRE
Re-implementation of "Dialogue-Based Relation Extraction" (ACL 2020) [paper]
Official code and dataset are avilable in [code] .
I want to make baseline code trainable in end-to-end for future works in this dataset.
I only implemented only bert baseline in DialogRE dataset v1 (English).
Result in devlopment set
Model | F1 | Precision | Recall |
---|---|---|---|
bert (my) | 58.0 | 59.2 | 56.7 |
bert (paper) | 60.6 | - | - |
Usage
Before run this code, construct dataset data/{train,dev,test}.json
with this format.
You can run this code:
python main.py
--transformer_type=bert \
--model_name=bert-base-cased \
--seed=42 \
--lr=3e-5 \
--wandb=True
Issue (check == solved)
- apply wandb and pytorch lighting - but it is my first attempt to use these libraries, please excuse my poor code :)
- roberta model training failure
- long sequence processing implementation
- long sequence processing harms performance