Awesome

SAILER: Structure-aware Pre-trained Language Model for Legal Case Retrieval

The official repo for our SIGIR'23 Full paper: SAILER: Structure-aware Pre-trained Language Model for Legal Case Retrieval.

Introduction

SAILER is a structure-aware pre-trained language model. It is highlighted in the following three aspects:

• SAILER fully utilizes the structural information contained in legal case documents and pays more attention to key legal elements, similar to how legal experts browse legal case documents.

• SAILER employs an asymmetric encoder-decoder architecture to integrate several different pre-training objectives. In this way, rich semantic information across tasks is encoded into dense vectors.

• SAILER has powerful discriminative ability, even without any legal annotation data. It can distinguish legal cases with different charges accurately.

The following figure shows the model structure. SAILER consists of a deep encoder and two shallow decoders. The Reasoning and Decision section are aggressively masked, joined with the Fact embedding to reconstruct the key legal elements and the judgment results.

Installation

Requirements

python=3.8
transformers==4.27.1
tqdm==4.65.0 
datasets==2.10.1
torch==1.11.0
faiss==1.7.3 
pytorch==1.12.1
pyserini==0.20.0
jieba==0.42.1 

Released Models

We have uploaded some checkpoints to Huggingface Hub.

You can load them quickly with following codes:

from transformers import AutoModel
model = AutoModel.from_pretrained('CSHaitao/SAILER_zh')

Pretrain

Data format

Before pre-training, you need to process the data into the following form:

{   
    "fact": basic fact,
    "interpretation": reasoning section, 
    "articles": related criminal law articles,
    "judgment": decision section,
}

data_example.json provides some examples of training data.

The case law system does may not have relevant articles. We can correspondingly modify SAILER_Collator in data.py to implement pre-training.

The pre-trained corpus is not publicly available due to data permissions. The legal case documents can be downloaded from Einglish or Chinese.

Train

We can pre-train model with sh sailer.sh

    BATCH_SIZE_PER_GPU=36
    GRAD_ACCU=4

    CUDA_VISIBLE_DEVICES=0 python -m torch.distributed.launch \
    --nproc_per_node 1 \
    --master_port 29508 \
    run_pretraining.py \
    --model_name_or_path bert-base-chinese \
    --output_dir $OUTPUT_DIR/model/SAILER_bert \
    --do_train \
    --logging_steps 50 \
    --save_steps 500 \
    --fp16 \
    --logging_dir $OUTPUT_DIR/tfboard/$MODEL_NAME \
    --warmup_ratio 0.1 \
    --per_device_train_batch_size $BATCH_SIZE_PER_GPU \
    --gradient_accumulation_steps $GRAD_ACCU \
    --learning_rate 5e-5 \
    --overwrite_output_dir \
    --dataloader_drop_last \
    --dataloader_num_workers 4 \
    --max_seq_length 512 \
    --num_train_epochs 10 \
    --train_path ./data/data_example.json \
    --weight_decay 0.01 \
    --encoder_mask_ratio 0.15 \
    --decoder_mask_ratio 0.45 \
    --use_decoder_head \
    --enable_head_mlm \
    --ddp_find_unused_parameters False \
    --n_head_layers 1 

where encoder_mask_ratio and decoder_mask_ratio represent the mask rate of encoder and decoder respectively. n_head_layers is the number of decoder layers.

Finetune

Data format

Before finetuning, you need to process the data into the following form:

{'query': TEXT_TYPE, 'positives': List[TEXT_TYPE], 'negatives': List[TEXT_TYPE]}
...

Train

To finetune the dense retriever, call the dense.driver.train module:

python -m dense.driver.train \  
  --output_dir $OUTDIR \  
  --model_name_or_path bert-base-uncased \  
  --do_train \  
  --save_steps 20000 \  
  --train_dir $TRAIN_DIR \
  --fp16 \  
  --per_device_train_batch_size 8 \  
  --learning_rate 5e-6 \  
  --num_train_epochs 2 \  
  --dataloader_num_workers 2

More finetune code details can be found in Dense. Note that we simply modify the faiss_retriever.retriever, which may make a slight difference to the results.

Evaluation

trec_eval is employed to evaluate the model performance

You can run sh eval.sh to do a simple evaluation

trec_eval-9.0.7/trec_eval ../result/qrel.trec ../result/SAILER_LeCaRD -m all_trec

You will get:

P_5 0.9084
recall_5 0.1880
ndcg_cut_10 0.7979
ndcg_cut_20 0.8190
ndcg_cut_30 0.8514

Citations

If you find our work useful, please do not save your star and cite our work:

@misc{SAILER,
      title={SAILER: Structure-aware Pre-trained Language Model for Legal Case Retrieval}, 
      author={Haitao Li and Qingyao Ai and Jia Chen and Qian Dong and Yueyue Wu and Yiqun Liu and Chong Chen and Qi Tian},
      year={2023},
      eprint={2304.11370},
      archivePrefix={arXiv},
      primaryClass={cs.IR}
}