GFC

Pytorch implementation for EMNLP 2022 paper

A Sequential Flow Control Framework for Multi-hop Knowledge Base Question Answering

Minghui Xie, Chuzhan Hao, and Peng Zhang

Overall architecture of our proposed GFC model.

The schematic diagram of the GRU-inspired Flow Control Framework.
If you find this code useful in your research, please cite ```bib @InProceedings{xie2022gfc, author = {Minghui Xie and Chuzhan Hao and Peng Zhang}, title = {A Sequential Flow Control Framework for Multi-hop Knowledge Base Question Answering}, year = {2022}, booktitle = {Empirical Methods in Natural Language Processing (EMNLP)}, } ``` ## dependencies - pytorch>=1.2.0 - [transformers](https://github.com/huggingface/transformers) - tqdm - nltk - shutil ## Prepare Datasets For all raw data files and their corresponding preprocessed data files, we have uploaded them to [google drive](https://drive.google.com/drive/folders/1ur-tSF_A1AkQWLkARDSMMQYBGODjKzwY?usp=sharing) ### [WebQSP](https://drive.google.com/drive/folders/1RlqGBMo45lTmWz9MUPTq-0KcjSd3ujxc?usp=sharing) We use script files from [EmbedKGQA](https://github.com/malllabiisc/EmbedKGQA) to process completed dataset of WebQSP. ### WebQSP-half We use the dataset preprocessed by [EmbedKGQA](https://github.com/malllabiisc/EmbedKGQA). ### MetaQA You should preprocess MetaQA using MetaQA/preprocess.py ### [CompWebQ](https://drive.google.com/file/d/1ua7h88kJ6dECih6uumLeOIV9a3QNdP-g/view?usp=sharing) We use dataset processed by [NSM](https://github.com/RichardHGL/WSDM2021_NSM). ## Demo To ensure the reproducibility, we write demo.py for all 4 main tasks. You can use pretrained model checkpoints to reproduce the results. You should put the checkpoints files into corresponding path. The checkpoints files are uploaded to [google drive](https://drive.google.com/drive/folders/1s0SwIALbgpJfaT800TEKwp8P2YOMc09A?usp=sharing). ### WebQSP Enter the directory GFC/WebQSP, then input the following command. ```shell python demo_wsp.py --input_dir data/WebQSP --save_dir checkpoints/WebQSP --ckpt checkpoints/WebQSP/model_wqsp.pt ``` ### WebQSP_half Enter the directory GFC/WebQSP_half, then input the following command ```shell python demo_half.py --input_dir data/WebQSP_half --save_dir checkpoints/WebQSP_half --ckpt checkpoints/WebQSP_half/model_wqsp_half.pt ``` ### CompWebQ Enter the directory GFC/CWQ, then input the following command ```shell python demo_cwq.py --input_dir data/CWQ --save_dir checkpoints/CWQ --ckpt checkpoints/CWQ/model_cwq.pt ``` ### MetaQA Enter the directory GFC/CWQ, then input the following command ```shell python demo_metaqa.py --input_dir data/Metaqa --save_dir checkpoints/Metaqa_ps --ckpt checkpoints/Metaqa/model_metaqa.pt ``` ## Experiments We train and test simultaneously. You should enter the corresponding directory of different datasets. ### WebQSP Enter the directory GFC/WebQSP, then input the following command. ```shell python train_hop_final.py --input_dir data/WebQSP --save_dir checkpoints/WebQSP ``` ### WebQSP_half Enter the directory GFC/WebQSP_half, then input the following command ```shell python train_half_hop_final.py --input_dir data/WebQSP_half --save_dir checkpoints/WebQSP_half ``` ### CompWebQ Enter the directory GFC/CWQ, then input the following command. ```shell python train_final.py --input_dir data/CWQ --save_dir checkpoints/CWQ --rev ``` ### MetaQA Enter the directory GFC/MetaQA, then input the following command. ```shell python train_final.py --glove_pt data/glove/glove.840B.300d.pickle --input_dir data/MetaQA --save_dir checkpoints/MetaQA ``` ## Acknowledgement This repo is built upon the following work: ``` TransferNet: An Effective and Transparent Framework for Multi-hop Question Answering over Relation Graph. Jiaxin Shi, Shulin Cao, Lei Hou1∗, Juanzi Li1 and Hanwang Zhang. EMNLP 2021. https://github.com/shijx12/TransferNet ``` Many thanks to the authors and developers! [](https://star-history.com/#Xie-Minghui/GFC&Date)