stephaniewyt / rdgcn Goto Github PK

The idea of ​​the article is great！
I understand the paper but I have a problem when I refer to the source code.Main.py 140 lines
what is logits = f_1 + tf.transpose(f_2) mean?

Dear author, thanks for your great work!

I would like to know how the initial vector X_q^{e_init} is obtained from the entity name in equation (8). I just found in the code that it is obtained by reading the file "fr_vectorList.json" directly, but how to generate this file?

您好，请问该论文代码有实现pytorch版本吗？

请问一下为什么运行后只有嵌入指标hits@,没有实体对齐的结果文件。

对glove.840B.300d进行了什么操作构造了实体表示。

出现了以下问题，请问您在运行代码时有这个问题吗？
Process finished with exit code -1073740791 (0xC0000409)

A very nice work!

I replaced the initialization method with the random initialization and found that RDGCN failed to achieve promising performance, So I want to know what's the effect of the translated name embeddings?

Thanks!

Hi,
Thanks for sharing your code.

I tried to conduct ablation studies that only uses the structure information. I made some modifications on the following function.

After modification, the entity embedding are not initialized from pretrained word embeddings. Instead, they are randomly initialized. However, ramdom initialization leads to significant decrement of the performance.

def get_input_layer(e, dimension, lang):
    print('adding the primal input layer...')
    with open(file='data/' + lang + '_en/' + lang + '_vectorList.json', mode='r', encoding='utf-8') as f:
        embedding_list = json.load(f)
        print(len(embedding_list), 'rows,', len(embedding_list[0]), 'columns.')
    input_embeddings = tf.convert_to_tensor(embedding_list)
    # ent_embeddings = tf.Variable(input_embeddings)
    ent_embeddings = tf.Variable(tf.random.uniform(shape=tf.shape(input_embeddings)))
    return tf.nn.l2_normalize(ent_embeddings, 1)

On JA_EN dataset, H@1 only reaches 0.53% after 150/600 epochs. Did I miss something? I would be grateful if you could reply to this issue.

Thanks and regrads,

We use your way to use google translate the Japanese into english and inferences the embedding with the GloVe. However, we used those embedding as the input for RDGCN and didn't get the same result as you proposed in the paper.

Caused by op 'gradients/concat', defined at:
File "main.py", line 34, in
Config.epochs, train, e, Config.k, test)
File ".\include\Model.py", line 275, in training
train_step = tf.train.AdamOptimizer(learning_rate).minimize(loss)

Datasets can not be downloaded, is there any other download link?