zhixiuye / mlman

ACL 2019 paper:Multi-Level Matching and Aggregation Network for Few-Shot Relation Classification

Multi-Level Matching and Aggregation Network

Dependencies

The code is written in Python 3.6 and pytorch 1.0.0.

Evaluation Results

Model	5 Way 1 Shot	5 Way 5 Shot	10 Way 1 Shot	10 Way 5 Shot
MLMAN	82.98 ± 0.20	92.66 ± 0.09	75.59 ± 0.27	87.29 ± 0.15

Usage

1. download train.json and val.json from here

2. download glove.6B.50d.json from here

3. make data folder in the following structure

MLMAN
|-- data
    |-- glove.6B.50d.json
    |-- train.json
    |-- val.json
|-- models
    |-- data_loader.py
    |-- embedding.py
    |-- framework.py
    |-- MLMAN.py
    |-- utils.py
|-- README.md
|-- train_demo.py

3. train model

CUDA_VISIBLE_DEVICES=0 python train_demo.py --N_for_train 20 --N_for_test 5 --K 1 --Q 5 --batch 1

Cite

If you use the code, please cite the following paper: "Multi-Level Matching and Aggregation Network for Few-Shot Relation Classification" Zhi-Xiu Ye, Zhen-Hua Ling. ACL (2019)

@inproceedings{ye-ling-2019-multi,
    title = "Multi-Level Matching and Aggregation Network for Few-Shot Relation Classification",
    author = "Ye, Zhi-Xiu  and
      Ling, Zhen-Hua",
    booktitle = "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
    month = jul,
    year = "2019",
    address = "Florence, Italy",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/P19-1277",
    pages = "2872--2881",
}

mlman's Issues

label数值和具体类别不是固定对应的？

MLMAN/models/data_loader.py

Line 220 in 43cbbdb

query_label += [i] * Q

target_classes = random.sample(self.rel2scope.keys(), N)
类别是随机选的每次类别顺序不一样，但是query label数值却是递增的。所以不同的batch，相同类别对应label数值是不同的，也就是不固定的。请问模型为什么能收敛呢？

预处理的mask没有用上？

https://github.com/ZhixiuYe/MLMAN/blob/master/models/MLMAN.py#L63
input_mask = (input['mask'] != 0).float()

这里只取了非零即文本长度的mask，且bool结果都为1，所以原先data_loader里的2,3值没有用上？

https://github.com/ZhixiuYe/MLMAN/blob/master/models/data_loader.py#L172
self.data_mask[i][j] = 2

效果

您好，根据您的ReadMe，代码运行，最后线下的效果5-way-1-shot大概78.3左右，请问正常吗，是提交后，test结果普遍比线下高吗？

在1 shot设置下的实验结果

您好，我重新运行了您的代码，并提交测试，在1 shot上的结果和您提供的结果有些不同，
5 way 1 shot的结果为 81.54，
10 way 1 shot的结果为 72.31，
请问是在训练1 shot时有需要修改的地方吗？

glove.6B.50d.json数据问题

你好，请问这个数据不存在了，是移动到哪里了吗？

请问能解释一下全连接层和lstm这样初始化的原因吗？

全连接层初始化代码
bias = np.sqrt(6.0 / (input_linear.weight.size(0) + input_linear.weight.size(1))) nn.init.uniform_(input_linear.weight, -bias, bias) if input_linear.bias is not None: input_linear.bias.data.zero_()
lstm层初始化代码
for ind in range(0, input_lstm.num_layers): weight = eval('input_lstm.weight_ih_l' + str(ind)) bias = np.sqrt(6.0 / (weight.size(0) / 4 + weight.size(1))) nn.init.uniform_(weight, -bias, bias) weight = eval('input_lstm.weight_hh_l' + str(ind)) bias = np.sqrt(6.0 / (weight.size(0) / 4 + weight.size(1))) nn.init.uniform_(weight, -bias, bias)