Parsing Natural Scenes and Natural Language with Recursive Neural Networks
作者信息:
Richard Socher richard@socher.org
Cliff Chiung-Yu Lin chiungyu@stanford.edu
Andrew Y. Ng ang@cs.stanford.edu
Christopher D. Manning manning@stanford.edu
Computer Science Department, Stanford University, Stanford, CA 94305, USA
年份:2011
代码和数据公开:
https://www.socher.org/index.php/Main/ParsingNaturalScenesAndNaturalLanguageWithRecursiveNeuralNetworks
理解什么是semantic representation(word index向量-》~,图片的segment-》~)
Syntactic parsing of natural language sentences:
its importance in mediating between linguistic expression and meaning.
Our RNN architecture jointly learns how to parse and how to represent phrases in a continuous vector space of features.
优点: This allows us to embed both single lexical units and unseen, variable-sized phrases in a syntactically coherent order. The learned feature representations capture syntactic and compositional-semantic information. We show that they can help inform accurate parsing decisions and capture interesting similarities between phrases and sentences.
4 recursive neural networks for structure predication
我们的discriminative parsing architecture的目标就是要学到一个函数f:X->Y,Y是所有可能的binary parse trees,输入X包含两部分1)activation vectors的集合,代表图片块或者句子的单词,2)对称矩阵A,当segmentI和segmentJ相邻时A(i,j)=1,所以主对角线上下的两个对角线diagonal的值一定为1
句子的ground truth tree只有一个,但图片的ground可能有多个
所有可能的parser tree中,只有当在和不同类merge之前所有属于同一类的相邻的部分都merge到一起时,才算正确
4.1 max-margin estimation
▲:合并错了就给惩罚
f:只有当算法认为tree y正确时才得分高 = 最大化s,s即得分,下面会详细
根据2007年的max-margin estimation framework,确保正确的树才是得分最高的树,我们设置正确的树的得分至少比错误的树大🔺这么多的loss
正则化的风险函数J,最大化正确树的得分,最小化错误树的最大得分
4.2 greedy structure predicting rnns
指数级别个数的parse trees,没有rnn有效的动态算法,因此采用贪婪算法来找出预测可能的tree--相邻的得分最高时就合并,更新矩阵A,上面说的s就是sd的和,sd就是d节点在不断合并过程中的得分
4.3 category classifiers in the tree/master/word_complexity_lexiconsyntactic categories:矩阵和上面向量p的乘积在经过激活函数softmax,最小化softmax后的交叉熵
4.4 improvement for language parsing
因为一个句子只有两个相邻的,所以计算量没那么大,用beam search,CKY算法
5 learning
反向传播。。。
参数theda=(Wsem,W,Wscore,Wlabel),分别是输入时映射到句法-语义空间的矩阵,连接相邻向量的矩阵,求合并得分的矩阵,合并处分类的矩阵,Wsem子语言的parse中,就是look-up table L。
6 实验
需要调的参数:隐藏层的大小n(句法-语义表示),不正确parse中的过程的惩罚项(求🔺),正则化参数郎姆拉