Skip to content

SIGIR'22 paper: Axiomatically Regularized Pre-training for Ad hoc Search

License

Notifications You must be signed in to change notification settings

xuanyuan14/ARES

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

24 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

img

THUIR License made-with-python code-size

Introduction

This codebase contains source-code of the Python-based implementation (ARES) of our SIGIR 2022 paper.

Requirements

  • python 3.7
  • torch==1.9.0
  • transformers==4.9.2
  • tqdm, nltk, numpy, boto3
  • trec_eval for evaluation on TREC DL 2019
  • anserini for generating "RANK" axiom scores

Why this repo?

In this repo, you can pre-train ARESsimple and TransformerICT models, and fine-tune all pre-trained models with the same architecture as BERT. The papers are listed as follows:

You can download the pre-trained ARES checkpoint ARESsimple from Google drive and extract it.

Pre-training Data

Download data

Download the MS MARCO corpus from the official website.
Download the ADORE+STAR Top100 Candidates files from this repo.

Pre-process data

To save memory, we store most files using the numpy memmap or jsonl format in the ./preprocess directory.

Document files:

  • doc_token_ids.memmap: each line is the token ids for a document
  • docid2idx.json: {docid: memmap_line_id}

Query files:

  • queries.doctrain.jsonl: MS MARCO training queries {"id" qid, "ids": token_ids} for each line
  • queries.docdev.jsonl: MS MARCO validating queries {"id" qid, "ids": token_ids} for each line
  • queries.dl2019.jsonl: TREC DL 2019 queries {"id" qid, "ids": token_ids} for each line

Human label files:

  • msmarco-doctrain-qrels.tsv: qid 0 docid 1 for training set
  • dev-qrels.txt: qid relevant_docid for validating set
  • 2019qrels-docs.txt: qid relevant_docid for TREC DL 2019 set

Top 100 candidate files:

  • train.rank.tsv, dev.rank.tsv, test.rank.tsv: qid docid rank for each line

Pseudo queries and axiomatic features:

  • doc2qs.jsonl: {"docid": docid, "queries": [qids]} for each line
  • sample_qs_token_ids.memmap: each line is the token ids for a pseudo query
  • sample_qid2id.json: {qid: memmap_line_id}
  • axiom.memmap: axiom can be one of the ['rank', 'prox-1', 'prox-2', 'rep-ql', 'rep-tfidf', 'reg', 'stm-1', 'stm-2', 'stm-3'], each line is an axiomatic score for a query

Quick Start

Example Usage

from model.modeling import ARESReranker

model = ARESReranker.from_pretrained(model_path).to(device)

query1 = "What is the best way to get to the airport"
query2 = "what do you like to eat?"

doc1 = "The best way to get to the airport is to take the bus"
doc2 = "I like to eat apples"

qd_pairs = [
        (query1, doc1), (query1, doc2),
        (query2, doc1), (query2, doc2)
]

score = model.score(qd_pairs)

You will get

scores: [ 41.60 -33.66 
          -38.00 30.03 ]

Note that to accelerate the training process, we adopt the parallel training technique. The scripts for pre-training and fine-tuning are as follow:

Pre-training

export BERT_DIR=/path/to/bert-base/
export XGB_DIR=/path/to/xgboost.model

cd pretrain

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5 NCCL_BLOCKING_WAIT=1 \
python  -m torch.distributed.launch --nproc_per_node=6 --nnodes=1 train.py \
        --model_type ARES \
        --PRE_TRAINED_MODEL_NAME BERT_DIR \
        --gpu_num 6 --world_size 6 \
        --MLM --axiom REP RANK REG PROX STM \
        --clf_model XGB_DIR

Here model type can be ARES or ICT.

Zero-shot evaluation (based on AS top100)

export MODEL_DIR=/path/to/ares-simple/
export CKPT_NAME=ares.ckpt

cd finetune

CUDA_VISIBLE_DEVICES=0 python train.py \
        --test \
        --PRE_TRAINED_MODEL_NAME MODEL_DIR \
        --model_type ARES \
        --model_name ARES_simple \
        --load_ckpt \
        --model_path CKPT_NAME

You can get:

#####################
<----- MS Dev ----->
MRR @10: 0.2991
MRR @100: 0.3130
QueriesRanked: 5193
#####################

on MS MARCO dev set and:

#############################
<--------- DL 2019 --------->
QueriesRanked: 43
nDCG @10: 0.5955
nDCG @100: 0.4863
#############################

on DL 2019 set.

Fine-tuning

export MODEL_DIR=/path/to/ares-simple/

cd finetune

CUDA_VISIBLE_DEVICES=0,1,2,3 NCCL_BLOCKING_WAIT=1 \
python -m torch.distributed.launch --nproc_per_node=4 --nnodes=1 train.py \
        --model_type ARES \
        --distributed_train \
        --PRE_TRAINED_MODEL_NAME MODEL_DIR \
        --gpu_num 4 --world_size 4 \
        --model_name ARES_simple

Visualization

export MODEL_DIR=/path/to/ares-simple/
export SAVE_DIR=/path/to/output/
export CKPT_NAME=ares.ckpt

cd visualization

CUDA_VISIBLE_DEVICES=0 python visual.py \
    --PRE_TRAINED_MODEL_NAME MODEL_DIR \
    --model_name ARES_simple \
    --visual_q_num 1 \
    --visual_d_num 5 \
    --save_path SAVE_DIR \
    --model_path CKPT_NAME

Results

Zero-shot performance:

Model Name MS MARCO MRR@10 MS MARCO MRR@100 DL NDCG@10 DL NDCG@100 COVID EQ
BM25 0.2962 0.3107 0.5776 0.4795 0.4857 0.6690
BERT 0.1820 0.2012 0.4059 0.4198 0.4314 0.6055
PROPwiki 0.2429 0.2596 0.5088 0.4525 0.4857 0.5991
PROPmarco 0.2763 0.2914 0.5317 0.4623 0.4829 0.6454
ARESstrict 0.2630 0.2785 0.4942 0.4504 0.4786 0.6923
AREShard 0.2627 0.2780 0.5189 0.4613 0.4943 0.6822
ARESsimple 0.2991 0.3130 0.5955 0.4863 0.4957 0.6916

Few-shot performance: img

Visualization (attribution values have been normalized within a document): img

Citation

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

@inproceedings{chen2022axiomatically,
  title={Axiomatically Regularized Pre-training for Ad hoc Search},
  author={Chen, Jia and Liu, Yiqun and Fang, Yan and Mao, Jiaxin and Fang, Hui and Yang, Shenghao and Xie, Xiaohui and Zhang, Min and Ma, Shaoping},
  booktitle={Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval},
  year={2022}
}

Notice

  • Please make sure that all the pre-trained model parameters have been loaded correctly, or the zero-shot and the fine-tuning performance will be greatly impacted.
  • We welcome anyone who would like to contribute to this repo. 🤗
  • If you have any other questions, please feel free to contact me via [email protected] or open an issue.
  • Code for data preprocessing will come soon. Please stay tuned~