wangpeiyi9979 / esd Goto Github PK

Hi! I have a detailed question that the code prints
Batch f1 [ SNIPS ]: [Beam Soft Nms]: ( p: {:.4f}; r: {:.4f}; f1: {:.4f} ) beam_size :{} k:{}, u:{}, delta:{}
Batch f1 [ SNIPS ]: [Origin ]: ( p: {:.4f}; r: {:.4f}; f1: {:.4f} )
All f1 [FewNERD]: [Beam Soft Nms]: ( p: {:.4f}; r: {:.4f}; f1: {:.4f} ) beam_size :{} k:{}, u:{}, delta:{}
All f1 [FewNERD]: [Origin ]: ( p: {:.4f}; r: {:.4f}; f1: {:.4f} ).

What does this mean? I guess that:
If I use SNIPS dataset, I can get the results through the corresponding Batch F1. If I use FewNERD dataset, I can get the results through the corresponding All F1. Is my guess correct? Thanks.

你好作者！如果用其他数据集，需要更改类别数量的地方在哪里？

Hi, peiyi. I have a detailed question about CSA.
It seems that we can directly use this to compute that does not need loop:

support_span_enhance4one_query = self.cross_attentioner(support_span_rep.unsqueeze(0),
one_query_spans_squeeze.unsqueeze(0),
one_query_spans_squeeze.unsqueeze(0)).squeeze(0)
query_span_enhance_rep = self.cross_attentioner(one_query_spans_squeeze.unsqueeze(0),
support_span_rep.unsqueeze(0),
support_span_rep.unsqueeze(0)).squeeze(0)

And it seems that the implementation is different from paper.
I don't know if there is something wrong with my understanding.

在处理含有不同数目span的句子时，您是较少的span数目句子添加pad，然后再统计那些span是pad。
比如当前句子有12个span，batch句子中最大span数是14，那么就会把12个扩充成14个，最后2个是0。
然而这个函数在标记0的位置时，用“>”判断，utils.py第16行，是否应该换成">="呢，因为
seq_range_expand = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10,11,12,13]
seq_length_expand = [12,12,12,12,12,12,12,12,12,12,12,12,12,12]
如果用“>”，会少标记一个0

①if self.is_support == False:
tag = tag - self.opt.O_class_num + 1 # for query, O1, O2, O3 -> O, entity_tag -> entity_tag - 2
请问这里，为什么要减2啊？之前根据support set，得到了tag2label, label2tag。这两个字典，就是下标和槽名的一一对应，那query set里获取tag的时候，减2不会把这个对应关系搞混吗？
②上面那两句代码的下一行，gold_entitys.add((b, e - 1, tag))。请问这里为什么还要减1？在函数 get_entity里，
if end_of_chunk(prev_tag, tag, prev_type, type_):
chunks.append((prev_type, begin_offset, i-1))
此处加到chunks里的，不是已经i-1了吗？
期待您的解答，谢谢

您好！我发现span这个变量里，一个句子中总是把它正确的槽标签起始位置放在前面，后面才是枚举了一些别的span。那在进行预测的时候，在计算增强后的span representation，以及每类的原型向量，不就相当于已经提前知道了query set的正确的槽边界，并且利用它来计算了representation，后面才进行span的分类？就是指query set虽然不知道槽的类别，但是它提前得到了边界信息。不知道我有没有理解错，期待您的回复

Hi, peiyi.
I have two detailed questions about paper and code:

1. Is SoftNMS equivalent to a greedy (i.e., we pick the span with the highest probability each time, and the tokens in this span will not be selected again) strategy under FewNERD and SNIPS dataset? There is no nested entity on these two datasets.
2. The input of bert needs mask:
   support_out = self.word_encoder(support['word'], output_hidden_states=True, return_dict=True) # [num_sent, number_of_tokens, 768]
   query_out = self.word_encoder(query['word'], output_hidden_states=True, return_dict=True) # [num_sent, number_of_tokens, 768]

query_out = self.word_encoder(query['word'], query['word'] != 0,output_hidden_states=True, return_dict=True)

代码似乎没有分开计算span f1和type f1,在尝试分别计算时遇到了点问题,想请问一下,怎么获取模型预测的span呢?似乎logits和pred的长度只与query的span长度相关,但是没有找到pred span相关的内容