dm2-nd / transtqa

Author: Wenhao Yu ([email protected]). EMNLP'20. Transfer Learning for Technical Question Answering.

TransTQA data and code

This repository contains the code package for the EMNLP'20 paper:

A Technical Question Answering System with Transfer Learning. Wenhao Yu (ND), Lingfei Wu (IBM), Yu Deng (IBM), Ruchi Mahindru (IBM), Qingkai Zeng (ND), Sinem Guven (IBM), Meng Jiang (ND).

• Our demonstration website is avaiable at http://159.89.9.22:5800/

• We also provide a video for the paper at https://vimeo.com/431118548

System overall and Framework

Figure: High level architecture of our proposed TransTQA system. First, the pre-trained ALBERT model is fine tuned with unstructured source technical corpus with masked language model (MLM) task, i.e., θ → θ+. Second, a siamese ALBERT takes fine tuned ALBERT and fine tunes with source technical QA, i.e., θ+ → θ++. Third, the siamese ALBERT further fine tunes with target QA, i.e., θ++ → θ+++. Our deployed system takes θ+++. Given a query, our system first calculates similarity scores between the query and each candidate answer, then ranks all scores from highest to lowest. Finally, the system returns top-3 ranked answers.

Environment settings

A detailed dependencies list can be found in requirements.txt and can be installed by:

pip install -r requirements.txt

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

Run the code

For fine-tuning BERT masked language model on tech corpus (before you run the model, you have to first download the technical corpus from here and put it into mlm/corpus folder ):

./script/run_mlm.sh

For pre-training the model (we use askubuntu for technical domain QA pre-training as default):

./script/run_pretrain.sh

For model transfer learning (we have two target datasets: stackunix and techqa (ACL 2020)):

./script/run_transfer.sh

Note that you should specify the path of pre-trained model and dataset.

Citation

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

@inproceedings{yu2020technical,
  title={A Technical Question Answering System with Transfer Learning},
  author={Yu, Wenhao and Wu, Lingfei and Deng, Yu and Mahindru, Ruchi and Zeng, Qingkai and Guven, Sinem and Jiang, Meng},
  booktitle={Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year={2020}
}

transtqa's Issues

Checkpoints for the models

Is there a way to obtain checkpoints for any of the models (M1/M2/M3/M4, ideally M4) listed in the paper ?

How to get the technical corpus

When I open your link to download the corpus, it jumped to the google drive page. The name is TTQA-TechCorpus, but the dir
is empty. So how can I get the corpus? Or if it's not convenient to put the corpus into the cloud, could please send it to my email or share a private link? Thanks a lot.