LOUVRE

Weakly Supervised Pre-Training for Multi-Hop Retriever

Yeon Seonwoo, Sang-Woo Lee, Ji-Hoon Kim, Jung-Woo Ha, Alice Oh | ACL-Findings 2021 | paper

KAIST, NAVER Clova, NAVER AI Lab

Official implementation of weakly supervised pre-training for multi-hop retriever

Abstract

In multi-hop QA, answering complex questions entails iterative document retrieval for finding the missing entity of the question. The main steps of this process are sub-question detection, document retrieval for the sub-question, and generation of a new query for the final document retrieval. However, building a dataset that contains complex questions with sub-questions and their corresponding documents requires costly human annotation. To address the issue, we propose a new method for weakly supervised multi-hop retriever pre-training without human efforts. Our method includes 1) a pre-training task for generating vector representations of complex questions, 2) a scalable data generation method that produces the nested structure of question and sub-question as weak supervision for pre-training, and 3) a pre-training model structure based on dense encoders. We conduct experiments to compare the performance of our pre-trained retriever with several state-of-the-art models on end-to-end multi-hop QA as well as document retrieval. The experimental results show that our pre-trained retriever is effective and also robust on limited data and computational resources.

Overview of LOUVRE

Proposed pre-training data generation method

<p align="center"><img width="80%" src="img/data_generation.png" /></p>

The multi-hop retrieval performance of LOUVRE fine-tuned on the HotpotQA train set

<p align="center"><img width="60%" src="img/perf.png" /></p>

How to run

Setup

Download HotpotQA data

mkdir data
mkdir data/hotpotQA
cd data/hotpotQA
wget http://curtis.ml.cmu.edu/datasets/hotpot/hotpot_train_v1.1.json
wget http://curtis.ml.cmu.edu/datasets/hotpot/hotpot_dev_fullwiki_v1.json
cd ../../

Download HotpotQA datasets used in MDR.

cd data/hotpotQA
wget https://dl.fbaipublicfiles.com/mdpr/data/hotpot/hotpot_train_with_neg_v0.json
wget https://dl.fbaipublicfiles.com/mdpr/data/hotpot/hotpot_dev_with_neg_v0.json
cd ../../

mkdir data/wiki_corpus
cd data/wiki_corpus
wget https://dl.fbaipublicfiles.com/mdpr/data/hotpot_index/wiki_id2doc.json
cd ../../

Download the tfidf file and the Wikipedia database file used in PathRetriever.

1. Download the zip file from here
2. Unzip hotpot_models.zip
3. Locate hotpot_models/tfidf_retriever/wiki_open_full_new_db_intro_only-tfidf-ngram=2-hash=16777216-tokenizer=simple.npz to data/tfidf/
4. Locate hotpot_models/wiki_db/wiki_abst_only_hotpotqa_w_original_title.db to data/db/

Create a new folder to store the pre-trained model, LOUVRE

mkdir checkpoints

Download LOUVRE from here and locate the checkpoint folder to checkpoints/

Create a new folder to save prediction files of LOUVRE and LOUVRE-eff

mkdir outputs

Pre-Training

Generating the pre-training data

cd code/pretrain_data_generator/
CHUNK_IND=0 bash generate_pretrain_data.sh

Adjust CHUNK_IND and NCHUNK to proceed the generation process concurrently.

Pre-Training

cd code/louvre
bash scripts/louvre/louvre_pretrain.sh

Fine-tuning LOUVRE-eff, LOUVRE

Fine-tuning LOUVRE-eff

cd code/louvre
bash scripts/louvre/louvre_eff_finetune.sh

Fine-tuning LOUVRE

cd code/louvre
bash scripts/louvre/louvre_finetune.sh
bash scripts/louvre/louvre_momentum.sh

Test LOUVRE-eff, LOUVRE, LOUVRE_wiki

Testing LOUVRE-eff

cd code/louvre
bash scripts/louvre/louvre_eff_pred.sh

Testing LOUVRE

bash scripts/louvre/louvre_pred.sh

Testing LOUVRE_wiki

bash scripts/louvre/louvre_pred_wiki.sh

Evaluation

INPUT_FILE=../../outputs/louvre_eff_finetune_dev.jsonl bash eval_retriever.sh

Code Reference

https://github.com/facebookresearch/multihop_dense_retrieval

https://github.com/AkariAsai/learning_to_retrieve_reasoning_paths

https://github.com/easonnie/semanticRetrievalMRS

Copyright

Copyright 2021-present NAVER Corp. and KAIST(Korea Advanced Institute of Science and Technology)

Acknowledgement

This work was partly supported by NAVER Corp. and Institute for Information & communications Technology Planning & Evaluation(IITP) grant funded by the Korean government(MSIT) (No. 2017-0-01780, The technology development for event recognition/relational reasoning and learning knowledge based system for video understanding).