KEFDA

Code and KGs for paper "Knowledge-Enhanced Domain Adaptation in Few-Shot Relation Classification" submitted in KDD'2021.

Data Preparation

FewRel and FewRel 2.0 Datasets are availble at here. Please download the data folder and place it in the home directory.

Use a pre-processed knowledge graph

The pre-processed Knowledge Graph can be downloaded directly from Google Drive.
Please place the ontology folder at ./fewshot_re_kit/kg_toolkit/.

Use your own knowledge graph

You can also choose to use your own knowledge graph by pre-processing it into a specific format.

Convert your knowledge graph into a low-dimensional representation, and we use OpenKE Toolkit for this process. Please place the learned embedding in the ./fewshot_re_kit/kg_toolkil/ontology/kg_embedding folder.
Please set the --kg_encoder argument in train.sh to the kg embedding file name, i.e. --kg_encoder my_transe.embedding.
Please place the following files in the ./fewshot_re_kit/kg_toolkil/ontology folder:
- entity2id.txt
```
# total entities
ENTITY_1  id_1
ENTITY_2  id_2
...
```
- ent2cnpt.txt
```
ENTITY  CNPT1   CNPT2   ...
```
- triple2id.txt
```
# total triples
head_entity_id rel_id tail_entity_id
...
```
- concept2id.txt
```
# total concepts
CNPT_1  id_1
...
```
- cnpt_id2def_fea.txt (The description features of concepts can be produced by BERT).
```
CNPT_id CNPT_DESC_FEATURES
```
- desc_feature/entity2id.txt
```
ENTITY id
```
- desc_feature/id2fea.txt
```
ENT_id ENT_DESC_FEATURES
```

How to Run

Training a Model

To run our model, use command

bash train.sh

This will start the training and evaluating process of KEDAF in a 5-way 1-shot setting.

You can train the model in different configurations by setting different args in train.sh.

train / val / test: Specify the training / validation / test set.
N: N in N-way K-shot.
K: N in N-way K-shot.
kg_encoder: Which knowledge graph encoder is used to encode the graph (for the provided pre-trained graph embedding, choose between transe/distmult/analogy/rotate; otherwise, set it to the name of the customized embedding file).
cnpt_att: Whether to aggregate relation meta using the attention mechanism.
lambda_para: Trade-off the importance of the knowledge enhanced prototype network and the relation-meta learning network, and the value should be picked between [0,1]. The higher the value, the more weight is given to the prototype network.

More arguments please see train.py for detail.

Inference

Run the following command to evaluate the existing checkpoint:

python train.py --only_test --load_ckpt {CHECKPOINT_PATH} {OTHER_ARGS}

njcx-ai / KEFDA