TriB-QA
Here is our group's project for CCF & Baidu 2019 Reading Comprehension Competition, dataset is Baidu Dureader.
The code will be completed once the competition finishes. The whole project is based on pytorch-version BERT: Passage Rerank, Answer predcition and YES/NO answer classification. So you may need to download pretrained language model,config file and vocab list in advance, or use our pretrained model to get final prediction.
Later, if possible, we will build a simple pipeline to ease the complicated procedures, also a web API may be built for it.
(起初以为参加就是白给,没想到或许可以嫖上一笔零花钱!o( ̄▽ ̄)ブ)
谢谢Naturali高抬一手,如愿以偿! 🙇
百度阅读理解竞赛官网查看具体要求.
当前训练数据集可以从我这边用U盘拷贝。
项目注意事项、进展情况可以由这里跟踪。
数据格式可以从这里 查看并补全。
竞赛的关键时间如下图:
Event | 时间 |
---|---|
报名 训练数据发放 | 02/25 |
报名截止 开发数据、测试集1发放 | 03/31 |
测试集2发放 | 05/13 |
结果提交截止 | 05/20 |
公布结果,接受报告论文 | 05/31 |
名字 | Rouge | Bleu | 时间 | Rouge提升 | Bleu 提升 |
---|---|---|---|---|---|
BIT_03/31(simple) | 37.15 | 23.41 | 2019/3/31 | ||
冲鸭04/01(single) | 40.84 | 24.82 | 2019/4/1 | 3.69 | 1.41 |
冲鸭^2(single) | 44.54 | 27.6 | 2019/4/2 | 3.7 | 2.78 |
冲鸭^3(single) | 45.91 | 35 | 2019/4/8 | 1.37 | 7.4 |
冲鸭^4(single) | 47.85 | 45.61 | 2019/4/9 | 1.94 | 10.61 |
冲鸭^5(single) | 48.03 | 46.09 | 2019/4/10 | 0.18 | 0.48 |
冲鸭^6(果断就会白给) | 48.13 | 46.7 | 2019/4/12 | 0.1 | 0.61 |
冲鸭^7(single) | 50.3 | 52.77 | 2019/4/15 | 2.17 | 6.07 |
冲鸭^8(single) | 48.27 | 49.65 | 2019/5/5 | -2.03 | -3.12 |
冲鸭^8(single) | 46.35 | 48.07 | 2019/5/6 | -1.92 | -1.58 |
冲鸭^8(single) | 50.46 | 52.37 | 2019/5/7 | 4.11 | 4.3 |
冲鸭^9(single) | 52.5 | 54.3 | 2019/5/8 | 2.04 | 1.93 |
冲鸭^10(single) | 53.13 | 54.63 | 2019/5/12 | 0.63 | 0.33 |
冲鸭^11(single) | 54.12 | 55.82 | 2019/5/13 | 0.99 | 1.19 |
果断就会白给(single) | 54.54 | 55.87 | 2019/5/15 | 0.42 | 0.05 |
果断就会白给(single) | 54.47 | 55.67 | 2019/5/16 | -0.07 | -0.2 |
果断就会白给(single) | 54.97 | 56.05 | 2019/5/18 | 0.5 | 0.38 |
果然还是白给了吗(single) | 55.3 | 56.09 | 2019/5/19 | 0.33 | 0.04 |
.... | 18.15 | 32.68 |
我们的模型初步定为三个bert,简称Tri-Bert。
Since Our model was initialy desgined as a pipileline include three BERT, we also call it TriB(ert) in our group :>.
Sounds rough but superisely effective.
小组目前成员5名:任慕成、魏然、柏宇、王洋、刘宏玉。
人人都是炼丹师