Are you sure you want to delete this task? Once this task is deleted, it cannot be recovered.
You can not select more than 25 topics
Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.
tbright17
9a225df8ed
|
4 years ago | |
|---|---|---|
| LICENSE.txt | 4 years ago | |
| README.md | 4 years ago | |
| WikiAnalyzerUtils.py | 4 years ago | |
| allennlp_lr_scheduler.py | 4 years ago | |
| coattn.py | 4 years ago | |
| config.py | 4 years ago | |
| dev_pred.json | 4 years ago | |
| dev_predictions.json | 4 years ago | |
| gnn.py | 4 years ago | |
| label_smooth_criterion.py | 4 years ago | |
| model_graph_comb.py | 4 years ago | |
| my_embedding.py | 4 years ago | |
| preprocessing_graph_comb.py | 4 years ago | |
| requirements.txt | 4 years ago | |
| run.py | 4 years ago | |
| sort_helper.py | 4 years ago | |
| submission68_1.py | 4 years ago | |
| trainer_graph_comb.py | 4 years ago | |
| vocab.pkl | 4 years ago | |
README.md
HDEGraph
Code for ACL 2019 paper "Multi-hop Reading Comprehension across Multiple Documents by Reasoning over Heterogeneous Graphs"
Based on PyTorch
Overview
Evaluation code for HDEGraph on WikiHop leaderboad with pretrained models.
Installation
-
git clone
-
Install PyTorch. The code has been tested with PyTorch >= 1.0
-
Install the requirements
-
Download pretrained models. Put zip file into the same folder with
run.py, and unzip it.
Running
Run
python run.py input_file output_file
input_file can be WikiHop dev file or other data sets organized in the same format with WikiHop.
output_file is the file where predictions locate at
Citation
@inproceedings{tu2019multi,
title={Multi-hop Reading Comprehension across Multiple Documents by Reasoning over Heterogeneous Graphs},
author={Tu, Ming and Wang, Guangtao and Huang, Jing and Tang, Yun and He, Xiaodong and Zhou, Bowen},
booktitle={Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics},
pages={2704--2713},
year={2019}
}
该项目针对跨文档的多跳阅读理解任务开源了一种新类型的图模型：异构文档-实体网络（Heterogeneous Document-Entity graph）。该算法包含不同粒度级别的信息，包括特定文档上下文、文档和实体， HDEGraph算法使用基于共同注意和自注意的上下文编码器初始化整个HDE 图, 然后通过基于图神经网络的传播算法在整个HDE 图上进行推断过程，从而得到各个结点的最终表示，最终用HDE 图的结点表示对候选答案进行打分，筛选出最终答案。在公开数据集上，HDEGraph算法表现出了优秀的性能。
Pickle Python