SDNet

An implementation for ACL 2022 paper Few-shot Named Entity Recognition with Self-describing Networks

Implementation

Quick links

Environment
Pretrained Data
Dataset
Pretrained SDNet
Fewshot Fine-tuning
Model Evaluation

Environment

conda create -n sdnet python=3.8.5
conda activate sdnet
bash env.sh

Pretrained Data

Part of pretrianed data is in the pretrain_data folder, the file includes 200k instances.

Dataset

Dataset is in the data folder:

data/DATASET/
├── test.json
├── kshot.json/full.json
└── mapping.json

test.json is data file, and each line is an Instance.

Instance format: Each instance is a Dict, containing tokens and entity fields, in which tokens is the list of tokens, and entity is the list of entity mentions.

{
    "tokens": [token1,token2,...],
    "entity": [
        [
            {"text":mention1, "type": type1, "offset":[startindex1,endindex1]},
            {"text":mention2, "type": type2, "offset":[startindex2,endindex2]},
            ...
        ]
},

kshot.json/full.json: the data file for k-shot fine-tuning, each line is a Dict, containing support and target_label fields, in which support is the list of instances in support set (full training set in full.json), and target_label is the list of target novel entity types.
mapping.json: a Dict mapping, the key is label name, the value is mapping words for each label (is commonly label name).

Pretrained SDNet

The pretrained SDNet (sdnet.th) should be putted in folder sdnetpretrain

You can download the pretrained SDNet in this link or line.

Fewshot Fine-tuning

run:

python main.py -dataset DATASET -K 5 -sdnet -cuda DEVICE

-dataset: DATASET is the dataset name in path data/DATASET
-sdnet: finetuning with our pre-trained SDNet, if not added, using t5-base
-K: control the shot number, default is 5
-full: if added, using the full training set to fine-tune
-cuda: is the GPU id

The predicted result is saved in tmp/dataset/...

Model Evaluation

just add -evalue:

python main.py -dataset DATASET -K 5 -sdnet -cuda DEVICE -evalue

License

The code is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International Public License for Noncommercial use only. Any commercial use should get formal permission first.

Shield:

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

chen700564 / sdnet