patrick-tssn/cdbert

[ACL2023] Shuo Wen Jie Zi is a new learning paradigm that enhances the semantics understanding ability of the Chinese PLMs with dictionary knowledge and structure of Chinese characters

CDBert

This is the official implementation of the paper "Shuo Wen Jie Zi: Rethinking Dictionaries and Glyphs for Chinese Language Pre-training".

Installation

pip install -r requirements.txt

Synthetic Chinese Character Data

This dataset can also be used for OCR

PolyMRC

A new machine reading comprehension task focusing on polysemy understanding

{"options": ["本着道义", "情义；恩情", "公正、合宜的道德、行为或道理", "坚持正义"], "sentence": ["汝之义绝高氏而归也，堂上阿奶仗汝扶持。"], "word": "义", "label": 0}

Download the dataset from huggingface dataset

Data Preparation

• Download Chinese Dictionary
• Download CLUE
• Download CCLUE

The dataset structure should look like the following:

| -- data
	| -- Pretrain
		| -- train.json
		| -- dev.json
	| -- CLUE
		| -- afqmc
			| -- train.json
			| -- dev.json
			| -- test.json
		| -- c3
		| -- chid
		| -- cmnli
		| -- cmrc
		| -- csl
		| -- iflytek
		| -- tnews
		| -- wsc
		| -- chid
	| -- CCLUE
		| -- fspc
		| -- mrc
		| -- ner
		| -- punc
		| -- seg
		| -- tc
	| -- PolyMRC
		| -- mrc
	| -- glyph_embedding.pt

  

CDBert

Pre-train

export MODEL_NAME=$1
export MODEL_PATH='prev_trained_models/'$1
export TRAIN=$2
export VAL=$3
export RADICAL=$4
export RID=$5
export LAN=$6
export BSZ=$8
export GLYPH=$9

CUDA_LAUNCH_BLOCKING=1 \
PYTHONPATH=$PYTHONPATH:. \
python -m torch.distributed.launch \
        --nproc_per_node=$7 \
        --master_port 40000 \
        pretrain.py \
        --distributed --multiGPU \
        --train datasets/$TRAIN \
        --valid datasets/$VAL \
        --batch_size $BSZ \
        --optim adamw \
        --warmup_ratio 0.05 \
        --clip_grad_norm 1.0 \
        --lr 5e-5 \
        --epoch 10 \
        --losses dict \
        --num_workers 1 \
        --backbone $MODEL_PATH \
        --individual_vis_layer_norm False \
        --output ckpts/$MODEL_NAME \
        --rid $RID \
        --max_text_length 256 \
        --radical_path $RADICAL \

CLUE (We only show the script for TNEW'S)

export MODEL_NAME=$1
export MODEL_PATH='prev_trained_models/'$1
export TASK_NAME=$2
export LR=$3
export EPOCH=$4
export BSZ=$5
export LEN=$6

PYTHONPATH=$PYTHONPATH:. \
python clue_tc.py \
        --task_name $TASK_NAME \
        --train train \
        --valid dev \
        --test test \
        --batch_size $BSZ \
        --valid_batch_size $BSZ \
        --optim adamw \
        --warmup_ratio 0.1 \
        --clip_grad_norm 1.0 \
        --lr $LR \
        --epoch $EPOCH \
        --num_workers 1 \
        --model_name $MODEL_NAME \
        --backbone $MODEL_PATH \
        --load ckpts/$MODEL_NAME/Epoch10 \
        --individual_vis_layer_norm False \
        --output outputs/CLUE/$TASK_NAME/$MODEL_NAME \
        --rid 368 \
        --embedding_lookup_table embedding/$MODEL_NAME/ \
        --fuse attn \
        --max_text_length  $LEN \
        --glyph radical \

CCLUE (We only show the script for MRC)

export MODEL_PATH='prev_trained_models/'$1
export TASK_NAME=$2
export LR=$3
export EPOCH=$4
export BSZ=$5
export LEN=$6

PYTHONPATH=$PYTHONPATH:. \
python cclue_mrc.py \
        --task_name $TASK_NAME \
        --train train \
        --valid dev \
        --test test \
        --batch_size $BSZ \
        --valid_batch_size $BSZ \
        --optim adamw \
        --warmup_ratio 0.05 \
        --clip_grad_norm 1.0 \
        --lr $LR \
        --epoch $EPOCH \
        --num_workers 1 \
        --model_name $MODEL_NAME \
        --backbone $MODEL_PATH \
        --load ckpts/$MODEL_NAME/Epoch10 \
        --individual_vis_layer_norm False \
        --output outputs/CCLUE/$TASK_NAME/$MODEL_NAME \
        --rid 233 \
        --embedding_lookup_table embedding/$MODEL_NAME/ \
        --fuse attn \
        --max_text_length  $LEN \
        --choices 4 \
        --glyph radical \

PolyMRC

export MODEL_NAME=$1
export MODEL_PATH='prev_trained_models/'$1
export TASK_NAME=$2
export LR=$3
export EPOCH=$4
export BSZ=$5
export LEN=$6

PYTHONPATH=$PYTHONPATH:. \
python dict_key.py \
        --task_name $TASK_NAME \
        --train train \
        --valid dev \
        --test test \
        --batch_size $BSZ \
        --valid_batch_size $BSZ \
        --optim adamw \
        --warmup_ratio 0.05 \
        --clip_grad_norm 1.0 \
        --lr $LR \
        --epoch $EPOCH \
        --num_workers 1 \
        --model_name $MODEL_NAME \
        --backbone $MODEL_PATH \
        --load ckpts/$MODEL_NAME/Epoch10 \
        --individual_vis_layer_norm False \
        --output outputs/CCLUE/$TASK_NAME/$MODEL_NAME \
        --rid 233 \
        --embedding_lookup_table embedding/$MODEL_NAME/ \
        --fuse attn \
        --max_text_length  $LEN \
        --choices 4 \
        --glyph radical \

Citation

@inproceedings{wang-etal-2023-rethinking,
    title = "Rethinking Dictionaries and Glyphs for {C}hinese Language Pre-training",
    author = "Wang, Yuxuan  and
      Wang, Jack  and
      Zhao, Dongyan  and
      Zheng, Zilong",
    booktitle = "Findings of the Association for Computational Linguistics: ACL 2023",
    month = jul,
    year = "2023",
    address = "Toronto, Canada",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.findings-acl.70",
    pages = "1089--1101",
    abstract = "We introduce CDBert, a new learning paradigm that enhances the semantics understanding ability of the Chinese PLMs with dictionary knowledge and structure of Chinese characters. We name the two core modules of CDBert as Shuowen and Jiezi, where Shuowen refers to the process of retrieving the most appropriate meaning from Chinese dictionaries and Jiezi refers to the process of enhancing characters{'} glyph representations with structure understanding. To facilitate dictionary understanding, we propose three pre-training tasks, i.e.„ Masked Entry Modeling, Contrastive Learning for Synonym and Antonym, and Example Learning. We evaluate our method on both modern Chinese understanding benchmark CLUE and ancient Chinese benchmark CCLUE. Moreover, we propose a new polysemy discrimination task PolyMRC based on the collected dictionary of ancient Chinese. Our paradigm demonstrates consistent improvements on previous Chinese PLMs across all tasks. Moreover, our approach yields significant boosting on few-shot setting of ancient Chinese understanding.",
}