CS224n (Spring 2017) assignment 3-----1. A window into NER

作业中提到实体识别的几个评价指标:Percision, Recall, F1. 第一题实现一个baseline模型,

CS224n (Spring 2017) assignment 3-----1. A window into NER

(a)比较简单,参考课后答案。


CS224n (Spring 2017) assignment 3-----1. A window into NER

(i) e(t): 1 * (2w+1)D;  W: D(2w+1) * H; U: H * C

(ii) 就是根据各个矩阵的维度来计算复杂度


CS224n (Spring 2017) assignment 3-----1. A window into NER

CS224n (Spring 2017) assignment 3-----1. A window into NER

(i)

CS224n (Spring 2017) assignment 3-----1. A window into NER

注:红框中的将嵌套的列表变成一个列表,以满足输出的格式。

(ii)

CS224n (Spring 2017) assignment 3-----1. A window into NER

CS224n (Spring 2017) assignment 3-----1. A window into NER

CS224n (Spring 2017) assignment 3-----1. A window into NER

CS224n (Spring 2017) assignment 3-----1. A window into NER

注:这里直接用了dropout,用1-dropout在测试集上的F1等指标不变化,没太搞清楚原因,明明注释中说要用1-dropout

CS224n (Spring 2017) assignment 3-----1. A window into NER

(iii)

CS224n (Spring 2017) assignment 3-----1. A window into NER

注: *符号的使用。

(iv)

CS224n (Spring 2017) assignment 3-----1. A window into NER


CS224n (Spring 2017) assignment 3-----1. A window into NER

这个题参考了答案,大概意思就是window base的模型不能使用句子中其他部分的信息。

CS224n (Spring 2017) assignment 3-----1. A window into NER