DL | Embedded Reber Grammar原理简述及一种python实现

1. Reber Grammar

如上图所示，它基本是个‘有环的有向图’。从B开始经过各种可能的有向边（所代表的字母）到达E组成的字符串，就是满足Reber Grammar的字符串。例如 BTSXSE，BPTVVE等。

2. Embedded Reber Grammar

把满足Reber Grammar的字符串Embed到一定的格式中得到的字符串满足Embedded Reber Grammar，有两种格式。

上图中的上方的路径的格式：

上图中的下方的路径的格式：

例如BTBTSXSETE或BPBTSXSEPE都是满足Embedded Reber Grammar的。

如果要判断字符串满足Embedded Reber Grammar，需要确定第二个字母和倒数第二个字母相同。对于一个学习模型，需要有某种记忆（第2个字母和倒数第2个字母相同）才能正确判断一个字符串是否满足Embedded Reber Grammar。如下图格式则不满足

3. Embedded Reber Grammar代码实现

下面的代码是计算n个满足Embedded Reber Grammar的字符串，基本思路就是用dict来刻画Reber Grammar图中边的前后关系，遇到多个后续的边可以选择的时候，用均匀概率来随机选取。

import random
import pandas as pd
import numpy as np

def ReberGrammar(n):
    '''
    :param n:
        how many reber string to generate
    :return:
        a list of unembedded reber strings
    '''
    graph={"B":["T1","P1"],
             "T1":["S1","X1"],
             "P1":["T1","V1"],
             "S1":["S1","X1"],
             "T2":["T2","V1"],
             "X1":["X2","S2"],
             "X2":["T2","V1"],
             "V1":["P2","V2"],
             "P2":["X2","S2"],
             "S2":["E"],
             "V2":["E"],
             "E":["end"]}
    rebers=[]
    for i in range(n):
        str_i=""
        edge_ = "B"
        while edge_ != "end":
            str_i+=edge_[0]
            sub_edge_=graph[edge_]
            edge_=random.sample(sub_edge_,1)[0]
        rebers.append(str_i)
    return(rebers)

def EmbeddedReberGrammar(rebers):
    '''
    :param rebers:
        unembedded reber strings
    :return:
        embedded reber strings
    '''
    newRebers=[]
    for _,str_ in enumerate(rebers):
        type=random.randint(0,1)
        if type==0:
            newRebers.append("BT"+str_+"TE")
        else:
            newRebers.append("BP"+str_+"PE")
    return(newRebers)

4. illegal Embedded Reber Grammar代码实现

用一个机器学习模型来判断一个字符串是否满足Embedded Reber Grammar时，需要一些不满足该法则的字符串作为负样本。
这个实现方法有很多，我用的方法是：
（1）0.5的概率篡改第2个字母，P改成T或者T改成P；
（2）0.5的概率篡改第3个～倒数第3个字母中的随机的一个，改成不合法的一个。

def illegalReberGrammarString(rebers):
    '''
    :param rebers:
        a list of legal reber strings
    :return:
        a list of illegal reber strings
    '''
    graph={"B":["T1","P1"],
             "T1":["S1","X1"],
             "P1":["T1","V1"],
             "S1":["S1","X1"],
             "T2":["T2","V1"],
             "X1":["X2","S2"],
             "X2":["T2","V1"],
             "V1":["P2","V2"],
             "P2":["X2","S2"],
             "S2":["E"],
             "V2":["E"],
             "E":["end"]}
    all_edges=["B","T","P","S","X","V","E"]
    illegalRebers=[]
    for i in range(len(rebers)):
        str_i=rebers[i]
        type=random.randint(0,1)
        if type==0:
            if str_i[1]=="P":
                str_i=str_i[0]+"T"+str_i[2:]
            else:
                str_i=str_i[0]+"P"+str_i[2:]
            illegalRebers.append(str_i)
        if type==1:
            L=len(str_i)
            edge_index=random.sample(range(2,L-2),1)[0]
            edge_=str_i[edge_index]
            if edge_ in ["B","E"]:
                sub_edge_=["P","T"]
            else:
                sub_edge_=graph[edge_+"1"]+graph[edge_+"2"]
            wrong_edge_=set(all_edges)-set([_[0] for _ in sub_edge_])
            replace=random.sample(wrong_edge_,1)[0]
            str_i=str_i[0:edge_index]+replace+str_i[edge_index+1:]
            illegalRebers.append(str_i)
    return(illegalRebers)

以下代码为生成一个csv文件：
第1列为标签y：1表示满足，0表示不满足;
第2列为是否满足Embed Reber Grammer的字符串x；

num=5000
####生成 num 个 Reber Grammar String
rebers=ReberGrammar(num)
####将Reber Grammar String嵌入格式中
embeddedRebers=EmbeddedReberGrammar(rebers)
####根据篡改规则，生成num个illegal Embed Reber Grammer
illegalRebers=illegalReberGrammarString(embeddedRebers)
####生成dataframe，写入csv
labels=[1]*num+[0]*num
strings=embeddedRebers+illegalRebers
####
index=np.random.permutation(2*num)

df=pd.DataFrame()
df['y']=labels
df['x']=strings
df=df.iloc[index,:]
df.to_csv('chapter14/reberStings.csv',index=0)

5. 参考资料

一个简洁的说明网页