Question Answering with T5 model using 🤗Transformers and Pytorch Lightning

Overview

Question Answering (QA) is the task of automatically answering questions given a paragraph, document, or collection. This open-source project aims to provide simplified training & inference routines, and QA fine-tuned models to facilitate and speed up research and experimentation within the QA task. Initial experiments use the T5 model and the Stanford Question Answering Dataset (SQuAD) v1.1. Here the goal is to generate an answer given the pair (question, text).

Example

...As this was the 50th Super Bowl, the league emphasized the "golden anniversary" with various gold-themed initiatives, as well as temporarily suspending the tradition of naming each Super Bowl game with Roman numerals (under which the game would have been known as "Super Bowl L"), so that the logo could prominently feature the Arabic numerals 50...

Question: If Roman numerals were used, what would Super Bowl 50 have been called?

Generated Answer: Super Bowl L

Main Features

Training, inference and evaluation scripts for QA
Fine-tuned QA T5 models for both English and Portuguese

Prerequisites

Python 3 (tested with version 3.9.13 on Windows 10)

Installation and Configuration

Clone this project:

git clone https://github.com/bernardoleite/question-answering-t5-pytorch-lightning

Install the Python packages from requirements.txt. If you are using a virtual environment for Python package management, you can install all python packages needed by using the following bash command:
```
cd question-answering-t5-pytorch-lightning/
pip install -r requirements.txt
```

Usage

You can use this code for data preparation, training, inference/predicting (full corpus or individual sample), and evaluation.

Data preparation

Current experiments use the SQuAD v1.1 dataset for English (original) and Portuguese (machine-translated) versions. So the next steps are specifically intended to preparing this dataset, but the same approach is applicable to other data types.

Example for preparing the English (original) SQuAD v1.1 dataset:

Download train-v1.1.json and dev-v1.1.json data files from here.
Create squad_en_original/raw/ and squad_en_original/processed/ foders inside data/ and copy previous files to data/squad_en_original/raw/.
Go to src/data. By running src/data/pre_process_squad_en_original.py the following dataframes (pickle format) will be created inside data/squad_en_original/processed/: df_train_en.pkl and df_validation_en.pkl.

Example for preparing the Portuguese (machine-translated) SQuAD v1.1 dataset:

Download train-v1.1-pt.json and dev-v1.1-pt.json data from here.
Create squad_br_v2/raw and squad_br_v2/processed folders inside data/ and copy previous files to data/squad_br_v2/raw/.
Go to src/data. By running src/data/pre_process_squad_br.py and then src/data/pre_process_squad_br_processed.py the following dataframes (pickle format) will be created inside data/squad_br_v2/processed/: df_train_br.pkl and df_validation_br.pkl.

Important note: Regardless of the data type, make sure the dataframe columns follow this scheme: [document_title, context, qa_id, question, answer].

Training

Go to src/model_qa. The file train.py is responsible for the training routine. Type the following command to read the description of the parameters:

python train.py -h

You can also run the example training script (linux and mac) train_qa_en_t5_base_512_96_32_10.sh:

bash train_qa_en_t5_base_512_96_32_10.sh

The previous script will start the training routine with predefined parameters:

#!/usr/bin/env bash

for ((i=42; i <= 42; i++))
do
    CUDA_VISIBLE_DEVICES=1 python train.py \
    --dir_model_name "qa_en_t5_base_512_96_32_10_seed_${i}" \
    --model_name "t5-base" \
    --tokenizer_name "t5-base" \
    --train_df_path "../../data/squad_en_original/processed/df_train_en.pkl" \
    --validation_df_path "../../data/squad_en_original/processed/df_validation_en.pkl" \
    --test_df_path "../../data/squad_en_original/processed/df_test_en.pkl" \
    --max_len_input 512 \
    --max_len_output 96 \
    --batch_size 32 \
    --max_epochs 10 \
    --patience 3 \
    --optimizer "AdamW" \
    --learning_rate 0.0001 \
    --epsilon 0.000001 \
    --num_gpus 1 \
    --seed_value ${i}
done

In the end, all model information is available at checkpoints/checkpoint-name. The information includes models checkpoints for each epoch (*.ckpt files), tensorboard logs (tb_logs/) and csv logs (csv_logs/).
Previous steps also apply to train_qa_br_v2_ptt5_base_512_96_32_10.sh for the training routine in Portuguese.

Inference (full corpus)

Go to src/model_qa. The file inference_corpus.py is responsible for the inference routine (full corpus) given a certain model checkpoint. Type the following command to read the description of the parameters:

python inference_corpus.py -h

You can also run the example inference corpus script (linux and mac) inference_corpus_qa_en_t5_base_512_96_10.sh:

bash inference_corpus_qa_en_t5_base_512_96_10.sh

The previous script will start the inference routine with predefined parameters for the model checkpoint model-epoch=00-val_loss=0.32.ckpt:

#!/usr/bin/env bash

for ((i=42; i <= 42; i++))
do
    CUDA_VISIBLE_DEVICES=1 python inference_corpus.py \
    --checkpoint_model_path "../../checkpoints/qa_en_t5_base_512_96_32_10_seed_42/model-epoch=00-val_loss=0.32.ckpt" \
    --predictions_save_path "../../predictions/qa_en_t5_base_512_96_32_10_seed_42/model-epoch=00-val_loss=0.32/" \
    --test_df_path "../../data/squad_en_original/processed/df_validation_en.pkl" \
    --model_name "t5-base" \
    --tokenizer_name "t5-base" \
    --batch_size 32 \
    --max_len_input 512 \
    --max_len_output 96 \
    --num_beams 5 \
    --num_return_sequences 1 \
    --repetition_penalty 1 \
    --length_penalty 1 \
    --seed_value ${i}
done

In the end, predictions will be available at predictions/checkpoint-name. The folder contains model predictions (predictions.json), and parameters (params.json).
Previous steps also apply to inference_corpus_qa_br_v2_ptt5_base_512_96_10.sh for the inference routine in Portuguese.

Inference (individual sample)

Go to src/model_qa. The file inference_example.py is responsible for the inference routine (individual sample) given a certain model checkpoint, CONTEXT and QUESTION. Type the following command to read the description of the parameters:

python inference_example.py -h

Example/Demo:

Change QUESTION and CONTEXT variables in inference_example.py:

QUESTION = 'If Roman numerals were used, what would Super Bowl 50 have been called?'
CONTEXT = """Super Bowl 50 was an American football game to determine the champion of the National Football League (NFL) for the 2015 season. The American Football Conference (AFC) champion Denver Broncos defeated the National Football Conference (NFC) champion Carolina Panthers 24–10 to earn their third Super Bowl title. The game was played on February 7, 2016, at Levi's Stadium in the San Francisco Bay Area at Santa Clara, California. As this was the 50th Super Bowl, the league emphasized the "golden anniversary" with various gold-themed initiatives, as well as temporarily suspending the tradition of naming each Super Bowl game with Roman numerals (under which the game would have been known as "Super Bowl L"), so that the logo could prominently feature the Arabic numerals 50."""

Run inference_example.py (e.g., using model-epoch=00-val_loss=0.32.ckpt as model checkpoint and 2 for num_return_sequences).

See output:

QUESTION:  If Roman numerals were used, what would Super Bowl 50 have been called?

Answer prediction for model_returned_sequence 0: Super Bowl L
Answer prediction for model_returned_sequence 1: Super Bowl L"

Evaluation

Go to src/. The file eval_qa_squad1.1.py is responsible for the evaluation procedure given a certain test_file_path and predictions_file. Type the following command to read the description of the parameters:
```
python eval_qa_squad1.1.py -h
```
Run eval_qa_squad1.1.py (e.g., using "../data/squad_en_original/raw/dev-v1.1.json as test_file_path and "../predictions/qa_en_t5_base_512_96_32_10_seed_42/model-epoch=00-val_loss=0.32/predictions.json" predictions_file)
```
python eval_qa_squad1.1.py -tfp "../data/squad_en_original/raw/dev-v1.1.json" -pf "../predictions/qa_en_t5_base_512_96_32_10_seed_42/model-epoch=00-val_loss=0.32/predictions.json"
```

Important note: Set variable LANG to "Portuguese" or "English" for correct answer normalization.

Checkpoints

We present the model checkpoints and their corresponding results. We show the results taking into account the full and partial validation dataset. We do this since there is an issue with the Portuguese machine-translated of SQuAD v1.1 dataset. The problem is that 39% of the translated answers are not found in the corresponding translated contexts. This happens because both answers and contexts are translated separately. Therefore, the translation of the individual answer does not always match the answer included in the translated context. Indeed, if we evaluate only the validation set instances where all the answers can be found in the contexts (6486 from 10570 instances), the results are 74.39% for Exact Match and 83.06% for F1 score. On the contrary, if we evaluate only the validation set instance where all answers are not found in the contexts (4084 from 10570 instances), the results are 26.81% for Exact Match and 55.16% for F1 score.

Checkpoint Name	Validation File	Model Type	Language	Exact Match	F1 Score
epoch=00_loss=0.32	Full val. corpus (10570 instances)	T5-Base	English	79.32	88.47
epoch=02_loss=1.12	Full val. corpus (10570)	PTT5-Base	Portuguese	56.01	72.28
epoch=02_loss=1.12	val. corpus w/out translation errors (6486)	PTT5-Base	Portuguese	74.39	83.06
epoch=02_loss=1.12	val. corpus with translation errors (4084)	PTT5-Base	Portuguese	26.81	55.16

Links for model checkpoint download:

epoch=00_loss=0.32
epoch=02_loss=1.12

Issues and Usage Q&A

To ask questions, report issues or request features, please use the GitHub Issue Tracker.

Contributing

Contributions are what make the open source community such an amazing place to learn, inspire, and create. Any contributions you make are greatly appreciated.

If you have a suggestion that would make this better, please fork the repo and create a pull request. You can also simply open an issue with the tag "enhancement". Don't forget to give the project a star! Thanks in advance!

Fork the Project
Create your Feature Branch (git checkout -b feature/AmazingFeature)
Commit your Changes (git commit -m 'Add some AmazingFeature')
Push to the Branch (git push origin feature/AmazingFeature)
Open a Pull Request

License

Project

This project is released under the MIT license. For details, please see the file LICENSE in the root directory.

Commercial Purposes

A commercial license may also be available for use in industrial projects, collaborations or distributors of proprietary software that do not wish to use an open-source license. Please contact the author if you are interested.

Acknowledgements

This project is inspired by/based on the implementations of Venelin Valkov, Ramsri Golla, Suraj Patil and Kristiyan Vachev.

Contacts

Bernardo Leite, bernardo.leite@fe.up.pt

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

README.md

README.md

Question Answering with T5 model using 🤗Transformers and Pytorch Lightning

Overview

Example

Main Features

Prerequisites

Installation and Configuration

Usage

Data preparation

Training

Inference (full corpus)

Inference (individual sample)

Evaluation

Checkpoints

Issues and Usage Q&A

Contributing

License

Project

Commercial Purposes

Acknowledgements

Contacts

Files

README.md

Latest commit

History

README.md

File metadata and controls

Question Answering with T5 model using 🤗Transformers and Pytorch Lightning

Overview

Example

Main Features

Prerequisites

Installation and Configuration

Usage

Data preparation

Training

Inference (full corpus)

Inference (individual sample)

Evaluation

Checkpoints

Issues and Usage Q&A

Contributing

License

Project

Commercial Purposes

Acknowledgements

Contacts