Skip to content

Latest commit

 

History

History
172 lines (148 loc) · 10.9 KB

README.md

File metadata and controls

172 lines (148 loc) · 10.9 KB

Open Table-and-Text Question Answering (OTT-QA)

This respository contains the OTT-QA dataset used in Open Question Answering over Tables and Text published in ICLR2021 and the baseline code for the dataset OTT-QA. This dataset contains open questions which require retrieving tables and text from the web to answer. This dataset is re-annotated from the previous HybridQA dataset. The dataset is collected by UCSB NLP group and issued under MIT license. You can browse the examples through our explorer.

overview

What's new compared to HybridQA:

  • The questions are de-contextualized to be standalone without relying on the given context to understand.
  • We add new dev/test set questions the newly crawled tables, which removes the potential bias in table retrieval.
  • The groundtruth table and passage are not given to the model, it needs to retrieve from 400K+ candidates of tables and 5M candidates of passages to find the evidence.
  • The tables in OTT-QA do not have groundtruth hyperlinks, which simulates a more general scenario outside Wikipedia.

Results

Table Retrieval: We use page title + page section title + table schema as the representation of a table for retrieval

Split HITS@1 HITS@5 HITS@10 HITS@20
Dev 41.0% 61.8% 68.5% 73.7%

QA Results: We use the retrieved table + retrieved text as the evidence to run HYBRIDER model (See https://arxiv.org/pdf/2004.07347.pdf for details), the results are shown as:

Model Dev-EM Dev-F1
BERT-based-uncased 8.7 10.9
BERT-large-uncased 10.9 13.1

Repo Structure

  • released_data: this folder contains the question/answer pairs for training, dev and test data.
  • data/all_plain_tables.json: this file contains the 400K+ table candidates for the dev/test set.
  • data/all_passages.json: this file contains the 5M+ open-domain passage candidates for the dev/test set.
  • data/traindev_tables_tok: this folder contains the train/dev tables.
  • data/traindev_request_tok: this folder cotains the linked passages for train/dev in-domain tables
  • table_crawling/: the folder contains the table extraction steps from Wikipedia.
  • retriever/: the folder contains the script to build sparse retriever index.

Requirements

We suggest using virtual environment to install these dependencies.

conda install pytorch torchvision cudatoolkit=10.2 -c pytorch
pip install transformers
pip install pexpect

Additional Information

If you want to know more about the crawling procedure, please refer to crawling for details.

If you want to know more about the retrieval procedure, please refer to retriever for details.

Or you can skip these two steps to directly download the needed files from AWS in Step1.

Step1: Preliminary Step

Step1-1: Download the necessary files

cd data/
wget https://opendomainhybridqa.s3-us-west-2.amazonaws.com/all_plain_tables.json
wget https://opendomainhybridqa.s3-us-west-2.amazonaws.com/all_passages.json
cd ../

This command will download the crawled tables and linked passages from Wikiepdia in a cleaned format.

Step1-2: Build inedx for retriever

cd retriever/
python build_tfidf.py --build_option text_title --out_dir text_title_bm25 --option bm25
python build_tfidf.py --build_option title_sectitle_schema --out_dir title_sectitle_schema

This script will generate index files under retriever/ folder, which are used in the following experiments

Step1-3: Reproducing the retrieval results

python evaluate_retriever.py --split dev --model retriever/title_sectitle_schema/index-tfidf-ngram\=2-hash\=16777216-tokenizer\=simple.npz  --format question_table

This script will produce the table retrieval results in terms of HITS@1,5,10,20,50.

Step2: Training

Step2-0: If you want to download the model from Google Drive, you can skip the following training procedure.

unzip models.zip

Step2-1: Preprocess the training data

python retrieve_and_preprocess.py --split train

This command will generate training data for different submodules in the following steps.

Step2-2: Train the three modules in the reader.

python train_stage12.py --do_lower_case --do_train --train_file preprocessed_data/stage1_training_data.json --learning_rate 2e-6 --option stage1 --num_train_epochs 3.0 --model_name_or_path bert-large-uncased
python train_stage12.py --do_lower_case --do_train --train_file preprocessed_data/stage2_training_data.json --learning_rate 5e-6 --option stage2 --num_train_epochs 3.0 --model_name_or_path bert-large-uncased
python train_stage3.py --do_train  --do_lower_case   --train_file preprocessed_data/stage3_training_data.json  --per_gpu_train_batch_size 12   --learning_rate 3e-5   --num_train_epochs 4.0   --max_seq_length 384   --doc_stride 128  --threads 8 --model_name_or_path bert-large-uncased

The three commands separately train the step1, step2 and step3 neural modules, all of them are based on BERT-uncased-base model from HugginFace implementation.

Step3: Evaluation

Step3-1: Reconstruct Hyperlinked Table using built text title index

python evaluate_retriever.py --format table_construction --model retriever/text_title_bm25/index-bm25-ngram\=2-hash\=16777216-tokenizer\=simple.npz
python retrieve_and_preprocess.py --split dev_retrieve --model retriever/title_sectitle_schema/index-tfidf-ngram\=2-hash\=16777216-tokenizer\=simple.npz
python retrieve_and_preprocess.py --split test_retrieve --model retriever/title_sectitle_schema/index-tfidf-ngram\=2-hash\=16777216-tokenizer\=simple.npz

This step can potentially take a long time since it matches each cell in the 400K tables against the whole passage title pool.

Step3-2: Evaluate with the trained model

python train_stage12.py --stage1_model stage1/[YOUR-MODEL-FOLDER] --stage2_model stage2/[YOUR-MODEL-FOLDER] --do_lower_case --predict_file preprocessed_data/dev_inputs.json --do_eval --option stage12 --model_name_or_path bert-large-uncased --table_path data/all_constructed_tables.json --request_path data/all_passages.json
python train_stage3.py --model_name_or_path stage3/[YOUR-MODEL-FOLDER] --do_stage3   --do_lower_case  --predict_file predictions.intermediate.json --per_gpu_train_batch_size 12  --max_seq_length 384   --doc_stride 128 --threads 8 --request_path data/all_passages.json

Once you have generated the predictions.json file, you can use the following command to see the results.

python evaluate_script.py predictions.json released_data/dev_reference.json

To replicate my results, please see the generated predictions.dev.json by my model.

python evaluate_script.py predictions.dev.json released_data/dev_reference.json

CodaLab Evaluation

To obtain the score on the test set (released_data/test.blind.json), you need to participate the CodaLab challenge in OTT-QA Competition. Please submit your results to obtain your testing score. The submitted file should first be named "test_answers.json" and then zipped. The required format of the submission file is described as follows:

[
  {
    "question_id": xxxxx,
    "pred": XXX
  },
  {
    "question_id": xxxxx,
    "pred": XXX
  }
]

The reported scores are EM and F1.

Link Prediction in Table

We also provide the script to predict the links from the given table based on the context using GPT-2 model. To train the model, please use the following command.

python link_prediction.py --dataset data/traindev_tables.json --do_train --batch_size 512

To generate links, please run

python link_prediction.py --do_all --load_from link_generator/model-ep9.pt --dataset data/all_plain_tables.json --batch_size 256

This command will generate all the link mapping in the link_generator/ folder.

Visualization

If you want to browse the tables, please go to this website and type in your table_id like 'Serbia_at_the_European_Athletics_Championships_2', then you will see all the information related to the given table.

Recent Papers

Model Organization Reference Dev-EM Dev-F1 Test-EM Test-F1
COS CMU + Microsoft Research + UIUC Ma et al. (2023) 56.9 63.2 54.9 61.5
CORE CMU + Microsoft Research Ma et al. (2022) 49.0 55.7 47.3 54.1
OTTeR MSRA + Beihang Huang et al. (2022) 37.1 42.8 37.3 43.1
RINK JBNU + NAVER Park et al. (2023) 36.7 42.4 35.5 41.5
CARP MSRA + Sun Yet-sen University Zhong et al. (2021) 33.2 38.6 32.5 38.5
Fusion+Cross-Reader Google Chen et al. (2021) 28.1 32.5 27.2 31.5
Dual Reader-Parser Amazon Alexander et al. (2021) 15.8 - - -
BM25-HYBRIDER UCSB Chen et al. (2021) 10.3 13.0 9.7 12.8

Reference

If you find this project useful, please cite it using the following format

  @article{chen2021ottqa,
  title={Open Question Answering over Tables and Text},
  author={Wenhu Chen, Ming-wei Chang, Eva Schlinger, William Wang, William Cohen},
  journal={Proceedings of ICLR 2021},
  year={2021}
}