Hi,

I run the code using this command

!CUDA_VISIBLE_DEVICES=1 python3 train_bert.py --model_name senticgcn_bert --dataset rest15 --save True --learning_rate 2e-5 --seed 39 --batch_size 16 --hidden_dim 768

But I got a very low accuracy compared to the reported results in ur paper

the obtained accuracy is 61.09% and 25.28% for f1-score and these values are the same for all the running steps and epochs !!

Do u have any idea plz?

Thank u

Hi,

First, thank u for sharing this great work with us.

I have a question about this model sdgcn.py is this an implementation of this paper "Modeling Sentiment Dependencies with Graph Convolutional Networks for Aspect-level Sentiment Classification" If no could u please provide a description of this model?

Many thanks in advance.

Hi,

Thank u for sharing this code with us. However, I got this error when trying to run this code

FileNotFoundError: [Errno 2] No such file or directory: 'state_dict/senticgcn_bert_rest14.pkl'

Could u please tell us how to generate the pkl files?

Thank u in advance

您好！阅读了您的论文之后，对于情感增强依存图的意义还不是很理解，添加了情感知识后的邻接矩阵中的数值有什么含义，在图卷积中是不是把这个数值看作两个单词间的相关性。以下面的例子为例： 这句话中的方面词分别是food和service,根据依存树得到的邻接矩阵中food和delicious,以及service和terrible的边的权重为1，在senticnet词典中，food、delicious、serviceh和terrible对应的情感值分别为0.054、0.106、0.84和-0.90，加入情感知识后，food和delicious,以及service和terrible的边的权重变为2.16，和1.94，请问这个数值的含义是什么，希望能尽快得到您的回复。

First of all, thank you for releasing the program on your paper. What I'm curious about is that SpaCy divides sentences into word units, but BERT divides them into WordPiece units, so I think there will be a problem that the tokens are not accurately mapped to each other. I wonder which part of the program you uploaded deals with these problems.

您好，我用您的代码跑中文数据集出现了以下错误，请问您知道解决方案吗？ Traceback (most recent call last): File "E:/Sentic-GCN-main/Sentic-GCN-main/train_bert.py", line 281, in main() File "E:/Sentic-GCN-main/Sentic-GCN-main/train_bert.py", line 276, in main ins = Instructor(opt) File "E:/Sentic-GCN-main/Sentic-GCN-main/train_bert.py", line 53, in init self.trainset = ABSADataset(opt.dataset_file['train'], tokenizer) File "E:\Sentic-GCN-main\Sentic-GCN-main\data_utils_bert.py", line 183, in init ((0,tokenizer.max_seq_len-idx2graph[graph_id].shape[0]),(0,tokenizer.max_seq_len-idx2graph[graph_id].shape[0])), 'constant') File "<array_function internals>", line 6, in pad File "D:\anaconda\envs\torch16\lib\site-packages\numpy\lib\arraypad.py", line 746, in pad pad_width = _as_pairs(pad_width, array.ndim, as_index=True) File "D:\anaconda\envs\torch16\lib\site-packages\numpy\lib\arraypad.py", line 517, in _as_pairs raise ValueError("index can't contain negative values") ValueError: index can't contain negative values

进程已结束，退出代码 1

Hello, after I clone your code, follow the readme file and run run_ senticgcn.sh , but the scores on the four data sets are 1-2% lower than those reported in your paper. Isn't your code the latest version. I run the code on the RTX 6000 GPU

The eq. 2 is Si,j = SenticNet(wi) + SenticNet(wj) in your paper, however, the code only computed SenticNet(wi) for S_i,j. Meanwhile w_i or w_j is a aspect word, the eq.4 need add T_ij=1 only once, however, the code here computed twice T_ij . Please check it.