This repository contains the code package for the NAACL 2021 paper:

Technical Question Answering across Tasks and Domains. Wenhao Yu (ND), Lingfei Wu (IBM), Yu Deng (IBM), Qingkai Zeng (ND), Ruchi Mahindru (IBM), Sinem Guven (IBM), Meng Jiang (ND).

After cloning this repo, install dependencies using

pip install -r requirements.txt

If you want to run with fp16, you need to install Apex

The dataset can be found here: http://ibm.biz/Tech_QA

The dataset contains three files (i) training_Q_A.json; (ii) dev_Q_A.json; (iii) training_dev_technotes.json. After downloading the dataset, you should put them in the TransTD folder and set their paths to the training script.

In order to train a model on TechQA, use the script below.

Note: Since TechQA is smaller dataset, it is better to start with a model that is already trained on a bigger QA dataset. Here, we start with BERT-Large trained on Squad.

python run_techqa.py \
    --model_type bert \
    --model_name_or_path bert-large-uncased-whole-word-masking-finetuned-squad \
    --do_lower_case \
    --num_train_epochs 20 \
    --learning_rate 5.5e-6 \
    --do_train \
    --train_file <PATH TO training_Q_A.json> \
    --do_eval \
    --predict_file <PATH TO dev_Q_A.json> \
    --input_corpus_file <PATH TO training_dev_technotes.json> \
    --per_gpu_train_batch_size 4 \
    --predict_batch_size 16 \
    --overwrite_output_dir \
    --output_dir <PATH TO OUTPUT FOLDER> \ 
    --add_doc_title_to_passage \

You can add the --fp16 flag if you have apex installed.

To evaluate a model, you can run:

python run_techqa.py \
    --model_type bert \
    --model_name_or_path <PATH TO TRAINED MODEL FOLDER> \
    --do_lower_case \
    --do_eval \
    --predict_file <PATH TO dev_Q_A.json> \
    --input_corpus_file <PATH TO training_dev_technotes.json> \
    --predict_batch_size 16 \
    --overwrite_output_dir \
    --output_dir <PATH TO OUTPUT FOLDER> \ 
    --add_doc_title_to_passage 

To reproduce our results, you can directly download our checkpoints and evaluate the model.

We provide two checking points trained with different seeds.

To get the evaluation file (output a json file including all metrics), you can run:

python techqa_evaluation.py \
    --data_file <PATH TO dev_Q_A.json> \
    --output_dir <PATH TO OUTPUT FOLDER> \ 

For help or issues, please submit a GitHub issue by clicking here.

For direct communication, please contact Wenhao Yu ([email protected]).

Many thanks to the Github repository of TechQA Baseline provided by IBM.

Part of our codes are modified based on their codes.

If you find this repository useful in your research, please consider to cite our paper:

@inproceedings{yu2021technical,
  title={Technical Question Answering across Tasks and Domains},
  author={Yu, Wenhao and Wu, Lingfei and Deng, Yu and Zeng, Qingkai and Mahindru, Ruchi and Guven, Sinem and Jiang, Meng},
  booktitle={NAACL},
  year={2021}
}