This is a Transformer-based model with Filtering and Aggregation mechanisms for the task of Relation Extraction. Previous methods adopt tree pruning to keep free from the noisy words, which may hurt the semantic integrity. FAT-RE treats a sentence as a fully-connected graph, and let the model decide which connections are important.
- Dataset
- TACRED
- SemEval2010 Task8
- Word Embedding
- Glove
- Word2vec
- BERT Embedding
- Prepare vocab
- scripts/prepare_vocab.sh
Please check if the resourses are prepared and the paths/arguments are specified. Example is shown in scripts/train.sh
python main.py --xxx
python eval.py --xxx
- https://github.com/qipeng/gcn-over-pruned-trees
- Yuhao Zhang, Peng Qi, Christopher D. Manning: Graph Convolution over Pruned Dependency Trees Improves Relation Extraction. EMNLP 2018: 2205-2215