Skip to content

Latest commit

 

History

History
112 lines (102 loc) · 3.89 KB

README.md

File metadata and controls

112 lines (102 loc) · 3.89 KB
Raw

Improving Conversational Passage Re-ranking with View Ensemble

This repositary includes

  1. Step1-rerank: Retrieve and re-rank the passage with two views.
  2. Proposed pseudo labeling: Create the pseudo-labels for CANARD dataset based on the view-ensemble ranked list.
  3. Train conversational passage re-ranker
  4. Evaluation on TREC CAsT 2019 and 2020.

Step1-rerank

In this step, we need to generate three views of ranked list based on

  1. Manually reformulated query(CANARD)
  2. Manually reformulated query(CANARD) + Answer (QuAC)

Raw training data

  1. Download canard dataset (training set)
mkdir canard
Download link: https://sites.google.com/view/qanta/projects/canard
  1. Download canard dataset (training set and validation set)
mkdir quac
Downlaod link: https://s3.amazonaws.com/my89public/quac/train_v0.2.json
Downlaod link: https://s3.amazonaws.com/my89public/quac/val_v0.2.json
  1. Parsed and preprocess (and fix) the canard training set into data/canard/train.jsonl.
python3 tools/parse_canard.py
  1. Generated the first-stage ranked list using BM25 search
bash run_spr.sh

Convert the runs into monot5 input

Converiting the training triplet into

# Input format: Query: <q> Docuemnt: <d> Relevant: 
# Output format: true/false
bash prepare_ranking_sources.sh

monoT5 reranking

Followed the monoT5 paper, you can either using huggingface or GCP+TPU to get the results. Constructed the environments, See detail in T5-for-IR

mkdir monot5-probs
bash fetch_probs.sh

Proposed pseudo labeling (conversational hard positive/negative sampling)

  1. Using the proposed method to generate the pseudo conversational query passage pairs Run the bash files (note that the other ablation datasets are also included)
mkdir data/canard4ir
bash run_create_convir_dataset.sh
# or 
python3 tools/construct_convir_dataset.py \
  --topic data/canard/train.jsonl \
  --collections <corpus path>
  --run0 runs/cast20.canard.train.view0.monot5.top1000.trec \
  --run1 runs/cast20.canard.train.view1.monot5.top1000.trec \
  --output data/canard4ir/canard4ir.train.convrerank.txt \
  --topk_pool 200 \
  --topk_positive 20 \
  --n 20 \
  1. Training with the weakly-supervsied dataset using TPU or GPU See the T5_TPU_README.md for detail.

Evaluation on CAsT 2020 (eval)

  1. Download CAsT 2020 The evaluation files are processed and parsed from original CAsT'20 repositary, in data/cast2020 You can also download from the official CAsT repo, and follow our processing pipeline.
  2. First-stage retrieval using CQE The dense retrieval results are in this repo, including the top1000 passage ranklist which are in runs/cast2020
  3. Convert the runs into monot5 input
python3 tools/convert_runs_to_monot5.py \
  --run <run file> \
  --topic data/cast2020/cast2020.eval.jsonl \ 
  --collection <corpus path> \
  --output monot5/cast2020.eval.cqe.conv.rerank.txt 
  --conversational
  1. Predicted the relevance scores using fine-tuned t5. You can see our checkpoint at Google bucket.
  • monot5-large-canard4ir
  • monot5-base-canard4ir

Evaluation on CAsT 2019

  1. Download CAsT 2019 You may find out the download files in official CAsT repo Then, parse the evaluation topics into jsonl.
  • Qrels (already in this repo)
  • Collections (MARCO, TRECCAR, WAPO)
  1. First-stage retrieval using CQE We have inferenced several dense retrieval baseline in this repo, including the top1000 passage ranklist which are in cast2019 runs
  2. Convert the runs into monot5 input
python3 tools/convert_runs_to_monot5.py \
  --run <run file> \
  --topic data/cast2019/cast2019.eval.jsonl \ 
  --collection <corpus path> \
  --output monot5/cast2019.eval.cqe.conv.rerank.txt 
  --conversational
  1. Predicted the relevance scores using fine-tuned t5. You can see our checkpoint at bucket.
  • monot5-base-canard4ir
  • monot5-large-canard4ir