leepleased / stackpropagation-slu Goto Github PK

你好，请问该模型如何与BERT同时使用呢？怎样添加BERT到模型中？相关联合模型是否有代码可共享呢？谢谢！

您好，我在使用process.validate对test数据进行测试的时候，发现save中的测试结果中第二列的标签与原始测试数据中的标签不一致，请问这是什么原因呢？

ATIS 数据集上： 训练参数 python train.py -wed 256 -ehd 256 -aod 128

Accepted performance: (0.95540278512251, 0.9664053751399776, 0.8622620380739082) at test dataset;

When checking the test result on test data of atis, I found 97 size difference of predicted slot and sentence. Details as following, in each tuple I list (len(w_list), len(r_slot_list), len(p_slot_list)). So the r_slot_list size is always the same as p_slot_list, but the w_list length can be shorter or longer. May I know what causes this?

A typical example is the first test result, which has longer prediction than the sentence
0 i O O atis_flight
1 would B-depart_date.date_relative B-depart_date.date_relative atis_flight
2 like B-depart_date.day_name B-depart_date.day_name atis_flight
3 to O O atis_flight
4 find O O atis_flight
5 a B-fromloc.city_name B-fromloc.city_name atis_flight
6 flight I-fromloc.city_name I-fromloc.city_name atis_flight
7 from O O atis_flight
8 charlotte B-toloc.city_name B-toloc.city_name atis_flight
9 to O O atis_flight
10 las O O atis_flight
11 vegas O O atis_flight
12 that B-toloc.city_name B-toloc.city_name atis_flight
13 makes B-arrive_time.time_relative B-arrive_time.time_relative atis_flight
14 a B-arrive_time.time B-arrive_time.time atis_flight
15 stop I-arrive_time.time I-arrive_time.time atis_flight
16 in O O atis_flight
17 st. O O atis_flight
18 louis O O atis_flight
atis_flight atis_flight
There is slot error in the prediction: Yes
slot diff is [(19, 'B-return_date.day_name', 'B-arrive_date.day_name')]

Full list of differences:
[(19, 20, 20), (11, 16, 16), (17, 16, 16), (13, 14, 14), (12, 14, 14), (8, 17, 17), (9, 13, 13), (8, 13, 13), (13, 15, 15), (16, 17, 17), (9, 11, 11), (16, 17, 17), (14, 16, 16), (12, 29, 29), (10, 18, 18), (11, 14, 14), (16, 14, 14), (9, 12, 12), (7, 9, 9), (9, 8, 8), (15, 7, 7), (13, 9, 9), (10, 7, 7), (14, 7, 7), (8, 15, 15), (9, 13, 13), (7, 8, 8), (14, 8, 8), (6, 11, 11), (8, 16, 16), (9, 12, 12), (13, 12, 12), (8, 11, 11), (8, 11, 11), (12, 8, 8), (6, 14, 14), (11, 14, 14), (9, 13, 13), (12, 5, 5), (8, 17, 17), (9, 11, 11), (12, 11, 11), (4, 10, 10), (11, 8, 8), (10, 17, 17), (10, 17, 17), (10, 15, 15), (7, 8, 8), (11, 6, 6), (10, 21, 21), (9, 10, 10), (16, 21, 21), (18, 20, 20), (21, 18, 18), (7, 8, 8), (2, 5, 5), (10, 4, 4), (13, 3, 3), (8, 3, 3), (3, 13, 13), (8, 12, 12), (7, 9, 9), (8, 2, 2), (8, 2, 2), (7, 15, 15), (6, 12, 12), (14, 12, 12), (10, 8, 8), (12, 6, 6), (5, 15, 15), (12, 17, 17), (7, 13, 13), (7, 8, 8), (13, 7, 7), (11, 7, 7), (12, 16, 16), (8, 11, 11), (9, 11, 11), (11, 6, 6), (8, 10, 10), (8, 10, 10), (8, 9, 9), (10, 6, 6), (8, 6, 6), (9, 19, 19), (11, 13, 13), (11, 12, 12), (14, 11, 11), (11, 10, 10), (11, 9, 9), (7, 11, 11), (12, 8, 8), (10, 21, 21), (20, 29, 29), (9, 10, 10), (9, 15, 15), (15, 11, 11)]

您好，看到您论文中有使用bert作为词向量的测试结果，但是在module中没有找到相关代码，可否补充一下？

您好，我看了一下源码，对模型选择那部分有些疑惑。论文中提到，模型是根据验证集结果来选择的，但是代码实现中，只要三个度量中任意一个有提升，就更新测试集结果和对应验证集上的度量，这样做的合理性如何解释？之前有类似的做法吗？

Hi, The paper team
I have gone through your code in detail, seems that the input tensor of the forward function of SelfAttention (the main member is a QKVAttention object) class is using padding digits. But in this case, I suppose we also need to put a mask_tensor, so that the padding digits will not affect the sentence meaning, right? This may help to improve the performance of the model.