meituan-dianping / asap Goto Github PK

Thanks for your extremely great work.

I have a question about the number of labels C mentioned in 4.2 section.

In your dataset, it seems that there are 4 labels for one aspect-category (1 - positive, 0 - neural, -1 - negative, -2 - not mentioned), but in the ACSA part, it looks like C is equal to 3 (positive, neural, negative)

C is the number of labels (i.e, 3 in our task).

so, I wonder why the label not mentioned was scrapped.

I have noticed that the gate function p_i is used to ensure only the mentioned aspect categories can participate in the calculation of the loss function.
but it seems like the gate function p_i is not a trainable parameter, so when using the trained model to do prediction, how the model knows which aspect-category are mentioned in a review?

for example, for an input review R written by a user,
the output of the ACSA part might be:

\hat{y}_1 = [0.7, 0.2, 0.1] 
\hat{y}_2 = [0.5, 0.2, 0.3]
...
\hat{y}_18 = [0.05, 0.05, 0.9]

(Please correct me if I have misunderstood something)

in this case, we can know that this user has positive sentiment with the first and second aspect-category, and has negative sentiment with the 18th aspect-category, but how do we know which aspect-category are mentioned in this review?

I have noticed that the dataset in this repo only contained a few samples. How can I get a full one for research purpose?

如题

Your work is extremely valuable!

If possible, could you make your experimental model publicly available for replication?

Thanks in advance!

本文的BERT模型应该是采用单阶段的方法来解决方面级的情感分类吧，一共有18个方面，每个方面有3类。
但是，不应该再多出一个类来表示评论中是否包含该方面的信息吗？针对每个方面进行情感分类的时候，用的是交叉熵损失函数吧。这样的话，不就默认评论包含所有的方面信息吗？