Awesome

OmniTab: Omnivorous Pretraining for Table-based QA

This repository contains the code, pre-trained models, and data for our paper OmniTab: Pretraining with Natural and Synthetic Data for Few-shot Table-based Question Answering by Zhengbao Jiang, Yi Mao, Pengcheng He, Graham Neubig, Weizhu Chen.

Overview

We propose an omnivorous pretraining approach that consumes natural data to endow models with the ability to understand and align natural language with tables, and synthetic questions to train models to perform reasoning.

Install environment with Conda

Create a conda env with the name omnitab using ./setup.sh.

Quick start using Transformers 🤗

You can directly load the OmniTab model (neulab/omnitab-large-finetuned-wtq) from HuggingFace's model hub.

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import pandas as pd

tokenizer = AutoTokenizer.from_pretrained('neulab/omnitab-large-finetuned-wtq')
model = AutoModelForSeq2SeqLM.from_pretrained('neulab/omnitab-large-finetuned-wtq')

data = {
    'year': [1896, 1900, 1904, 2004, 2008, 2012],
    'city': ['athens', 'paris', 'st. louis', 'athens', 'beijing', 'london']
}
table = pd.DataFrame.from_dict(data)

query = 'In which year did beijing host the Olympic Games?'
encoding = tokenizer(table=table, query=query, return_tensors='pt')

outputs = model.generate(**encoding)

print(tokenizer.batch_decode(outputs, skip_special_tokens=True))
# [' 2008']

Model list

• Pretrained models (w/o finetuning)
  • neulab/omnitab-large: Pretrained on natural and synthetic data generated with a SQL2NL model trained in the full setting.
  • neulab/omnitab-large-16shot: Pretrained on natural and synthetic data generated with a SQL2NL model trained in the 16-shot setting.
  • neulab/omnitab-large-128shot: Pretrained on natural and synthetic data generated with a SQL2NL model trained in the 128-shot setting.
  • neulab/omnitab-large-1024shot: Pretrained on natural and synthetic data generated with a SQL2NL model trained in the 1024-shot setting.
• Finetuned models
  • neulab/omnitab-large-finetuned-wtq: neulab/omnitab-large finetuned on WTQ in the full setting.
  • neulab/omnitab-large-16shot-finetuned-wtq-16shot: neulab/omnitab-large-16shot finetuned on WTQ in the 16-shot setting.
  • neulab/omnitab-large-128shot-finetuned-wtq-128shot: neulab/omnitab-large-128shot finetuned on WTQ in the 128-shot setting.
  • neulab/omnitab-large-1024shot-finetuned-wtq-1024shot: neulab/omnitab-large-1024shot finetuned on WTQ in the 1024-shot setting.

Performance reference

The table below contains the peformance of OmniTab models of various settings on validation/test split of WTQ before (omnitab-large-{f}shot) and after finetuning (...-finetuned-wtq-{f}shot).

Download pretraining data and WikiTableQuestions dataset

Download the pretraining data and the WikiTableQuestions dataset from Google Drive. You can download it programmatically with gdrive using gdrive download -r 14IAqJb9ObVDE5oOJouhkqgd_mn11PkYY. It includes:

|-- pretrain_data
    |-- natural.jsonl # natural pretraining data (generated from a subset of the TAPAS pretraining data (https://github.com/google-research/tapas/blob/master/PRETRAIN_DATA.md))
    |-- synthetic.jsonl # synthetic pretraining data (generated from sql.jsonl using SQL2NL model training in the full setting)
    |-- sql.jsonl # SQL pretraining data (a subset of the TAPEX pretraining data (https://github.com/microsoft/Table-Pretraining#pre-training-corpus))
|-- wtq # the WikiTableQuestions dataset
    |-- fewshot_ids # ids of training examples used in few-shot finetuning
    |-- predictions_validation # predictions of various OmniTab models on the WTQ validation split
    |-- predictions_test # predictions of various OmniTab models on the WTQ test split
    |-- tagged # annotation files used in computing metrics
    |-- validation_ids.txt # ids of validation examples used in computing metrics

Pretraining data format

Each example contains 4 fields context, table, mentions, and answers, where

• context is the natural language sentence relevant to the table.
• table is a 2-dimensional table with a header and one or multiple rows.
• mentions is a list of char-indexed spans that indicate mentions in the context aligned with the table.
• answers is a list of answers.

context, table, and mentions are required for natural pretraining examples, while context, table, and answers are required for synthetic pretraining examples. Take a look at the pretraining data processing code to get a better idea. Below is a natural pretraining example:

{
  "context": " actresses (catalan: actrius) is a 1997 catalan language spanish drama film ...", 
  "mentions": [[40, 47], [101, 113], [164, 181]], 
  "table": {
    "header": ["Directed by", "Produced by", "Written by", ...], 
    "rows": [
      ["Ventura Pons", "Ventura Pons", "Josep Maria Benet i Jornet", ...]
    ]
  },
  "answers": []
}

Experiments

The scripts below by defaults use 8 GPUs and assume they are 32GB V100. An example SLURM header is included at the top of each script if you want to submit the script using sbatch. Modify these headers and the commands to activate the environemnt based on the configuration of your cluster.

Finetuning

Full setting

Finetune pretrained OmniTab models on WikiTableQuestions dataset in the full setting:

./train_wikitablequestions.sh neulab/omnitab-large output/omnitab-large-finetuned-wtq

which finetunes neulab/omnitab-large and saves the model to output/omnitab-large-finetuned-wtq. All hyperparameters are set to reproduce the experiments in the paper, and you should be able to get performance close to 0.633 as listed in the "performance reference" table

Few-shot settings

Finetune pretrained OmniTab models on WikiTableQuestions dataset in few-shot settings:

./train_wikitablequestions.sh neulab/omnitab-large-${f}shot output/omnitab-large-${f}shot-finetuned-wtq-${f}shot ${f}

which takes the pretrained model neulab/omnitab-large-${f}shot and finetunes it in f-shot setting where f=[16|32|64|128|256|512|1024].

Inference

Run inference using OmniTab models on WikiTableQuestions validation or test split and save the prediction results to outupt/predictions.txt:

./test_wikitablequestions.sh validation|test neulab/omnitab-large-finetuned-wtq output

Evaluation

We use answer accuracy (i.e., denotation accuracy) as the metric and the numbers produced by run.py is less lenient to small variances (such as different formats of dates). To get a more accurate metric (usually higher), we use the official WTQ evaluation script and modifies it slightly to handle predictions from generative models which are a single string potentially containing multiple answers concatenated by a separator.

python evaluator.py --split test omnitab_download/wtq/predictions_test/omnitab-large-finetuned-wtq.txt
python evaluator.py --split validation omnitab_download/wtq/predictions_validation/omnitab-large-finetuned-wtq.txt

Pretraining

The best OmniTab model is initialized with microsoft/tapex-large, and continuously pretrained on both natural data, synthetic data, and SQL data in omnitab_download/pretrain_data:

./pretrain.sh omnitab_download/pretrain_data microsoft/tapex-large output/omnitab-large

Hyperparameters are the same as the ones used in the paper.

Reference

@inproceedings{jiang-etal-2022-omnitab,
  title = "{O}mni{T}ab: Pretraining with Natural and Synthetic Data for Few-shot Table-based Question Answering",
  author = "Jiang, Zhengbao and Mao, Yi and He, Pengcheng and Neubig, Graham and Chen, Weizhu",
  booktitle = "Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
  month = jul,
  year = "2022",
}

Acknowledgement

The codebase is inspired by and based on TAPEX and its Huggingface version, which is well-organized and easy to use. Take a look their repositories if you want to explore more details!