![](data:image/png;base64,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)
392 Chapter 11. 基于卷积神经网络的模型 肖桐 朱靖波
11.1.4 面向序列的卷积操作
对比于图像处理任务中二维图像数据,自然语言处理任务中主要处理一维序列,
如单词序列。由于单词序列长度往往是不固定的,很难使用全连接网络处理它,因
为变长序列无法用固定大小的全连接网络进行直接建模,而且过长的序列也会导致
全连接网络参数量的急剧增加。
e
1
e
2
e
3
e
4
e
5
e
6
e
7
e
8
e
9
(a) 循环神经网络的串行结构
0
e
1
e
2
e
3
e
4
e
5
e
6
e
7
e
8
e
9
0
0
2 2 2 2 2 2 2 2 2
0
0
3 3 3 3 3 3 3 3 3
0
0
4 4 4 4 4 4 4 4 4
0
(b) 卷积神经网络的层级结构
图 11.9 串行及层级结构对比(e
i
表示词嵌入,0 表示 0 向量,方框里的 2、3、4 表示层次编号)
针对不定长序列,一种可行的方法是使用之前介绍过的循环神经网络进行信息
提取,其本质也是基于权重共享的想法,在不同的时间步复用相同的循环神经网络
单元进行处理。但是,循环神经网络最大的弊端在于每一时刻的计算都依赖于上一
时刻的结果,因此只能对序列进行串行处理,无法充分利用硬件设备进行并行计算,
导致效率相对较低。此外,在处理较长的序列时,这种串行的方式很难捕捉长距离
的依赖关系。相比之下,卷积神经网络采用共享参数的方式处理固定大小窗口内的
信息,且不同位置的卷积操作之间没有相互依赖,因此可以对序列进行高效地并行
处理。同时,针对序列中距离较长的依赖关系,可以通过堆叠多层卷积层来扩大感
受野 (Receptive Field) ,这里感受野指能够影响神经元输出的原始输入数据区域的大
小。图11.9对比了这两种结构,可以看出,为了捕捉 e
2
和 e
8
之间的联系,串行结构
需要顺序地进行 6 次操作,操作次数与序列长度相关。而该卷积神经网络中,卷积
操作每次对三个词进行计算,仅需要 4 层卷积计算就能得到 e
2
和 e
8
之间的联系,其
操作数和卷积核的大小相关,相比于串行的方式具有更短的路径和更少的非线性计
算,更容易进行训练。因此,也有许多研究人员在许多自然语言处理任务上尝试使
用卷积神经网络进行序列建模
[497, 498, 500, 507, 508]
。
区别于传统图像上的卷积操作,在面向序列的卷积操作中,卷积核只在序列这
一维度进行移动,用来捕捉连续的多个词之间的特征。需要注意的是,由于单词通
常由一个实数向量表示(词嵌入),因此可以将词嵌入的维度看作是卷积操作中的
通道数。图11.10就是一个基于序列卷积的文本分类模型,模型的输入是维度大小为