Lesson-01 AI Introduction: Syntax Tree and Probability Model
如果给定一个语法,我们怎么生成语法呢?
import random
sentence = """
句子 = 主 谓 宾
主 = 你 | 我 | 他
"""
two_number = """
num* => num num* | num
num => 0 | 1 | 2 | 3 | 4
"""
def two_num(): return num() + num()
def num():
return random.choice("0 | 1 | 2 | 3 | 4 ".split('|'))
def numbers():
if random.random() < 0.5:
return num()
else:
return num() + numbers()
for i in range(10):
print(numbers())
1. 语法可以通过定义最简单的函数来实现
2. 我们可以通过递归,来生成更复杂,“无限”长的文字
simple_grammar = """
sentence => noun_phrase verb_phrase
noun_phrase => Article Adj* noun
Adj* => Adj | Adj Adj*
verb_phrase => verb noun_phrase
Article => 一个 | 这个
noun => 女人 | 篮球 | 桌子 | 小猫
verb => 看着 | 坐在 | 听着 | 看见
Adj => 蓝色的 | 好看的 | 小小的
"""
another_grammar = """
#
"""
import random
def adj(): return random.choice('蓝色的 | 好看的 | 小小的'.split('|')).split()[0]
def adj_star():
# 为什么如果不用if-else 的random,我们需要用lambda
return random.choice([lambda : '', lambda : adj() + adj_star()])()
def adj_star():
return random.choice([lambda : '', lambda : adj() + adj_star()])()
for i in range(10):
print(adj_star())
But the question is ?
如果我们更换了语法,会发现所有写过的程序,都要重新写。😦
number_ops = """
expression => expression num_op | num_op
num_op => num op num
op => + | - | * | /
num => 0 | 1 | 2 | 3 | 4
"""
def generate_grammar(grammar_str: str, target, split='=>'):
grammar = {}
for line in grammar_str.split('\n'):
if not line: continue
# two => num + num
expression, formula = line.split(split)
formulas = formula.split('|')
formulas = [f.split() for f in formulas]
grammar[expression.strip()] = formulas
return grammar
choice_a_expr = random.choice
def generate_by_grammar(grammar: dict, target: str):
if target not in grammar: return target
# the above line is to test if target is a key
expr = choice_a_expr(grammar[target])
return ''.join(generate_by_grammar(grammar, t) for t in expr)
def generate_by_str(grammar_str, split, target):
grammar = generate_grammar(grammar_str, target, split)
return generate_by_grammar(grammar, target)
generate_by_str(number_ops, split='=>', target='expression')
two => num + num | num - num
num => 0 | 1 | 2 | 3 | 4
#在西部世界里,一个”人类“的语言可以定义为:
human = """
human = 自己 寻找 活动
自己 = 我 | 俺 | 我们
寻找 = 找找 | 想找点
活动 = 乐子 | 玩的
"""
假如既然 = """
句子 = if someone state , then do something
if = 既然 | 如果 | 假设
someone = one 和 someone | one
one = 小红 | 小蓝 | 小绿 | 白白
state = 饿了 | 醒了 | 醉了 | 癫狂了
then = 那么 | 就 | 可以
do = 去
something = 吃饭 | 玩耍 | 去浪 | 睡觉
"""
#一个“接待员”的语言可以定义为
host = """
host = 寒暄 报数 询问 业务相关 结尾
报数 = 我是 数字 号 ,
数字 = 单个数字 | 数字 单个数字
单个数字 = 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9
寒暄 = 称谓 打招呼 | 打招呼
称谓 = 人称 ,
人称 = 先生 | 女士 | 小朋友
打招呼 = 你好 | 您好
询问 = 请问你要 | 您需要
业务相关 = 具体业务
具体业务 = 喝酒 | 打牌 | 打猎 | 赌博
结尾 = 吗?
"""
for i in range(10):
print(generate_by_str(假如既然, split='=', target='句子'))
for i in range(10):
print(generate_by_str(host, split='=', target='host'))
希望能够生成最合理的一句话?
Eliza
Data Driven
我们的目标是,希望能做一个程序,然后,当输入的数据变化的时候,我们的程序不用重写。Generalization.
AI? 如何能自动化解决问题,我们找到一个方法之后,输入变了,我们的这个方法,不用变。
simpel_programming = '''
programming => if_stmt | assign | while_loop
while_loop => while ( cond ) { change_line stmt change_line }
if_stmt => if ( cond ) { change_line stmt change_line } | if ( cond ) { change_line stmt change_line } else { change_line stmt change_line }
change_line => /N
cond => var op var
op => | == | < | >= | <=
stmt => assign | if_stmt
assign => var = var
var => var _ num | words
words => words _ word | word
word => name | info | student | lib | database
nums => nums num | num
num => 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 0
'''
print(generate_by_str(simpel_programming, target='programming', split='=>'))
def pretty_print(line):
# utility tool function
lines = line.split('/N')
code_lines = []
for i, sen in enumerate(lines):
if i < len(lines) / 2:
#print()
code_lines.append(i * " " + sen)
else:
code_lines.append((len(lines) - i) * " " + sen)
return code_lines
generated_programming = []
for i in range(20):
generated_programming += pretty_print(generate_by_str(simpel_programming, target='programming', split='=>'))
for line in generated_programming:
print(line)
Language Model
1. 条件概率
2. 独立概率
Review: 1. 条件概率?
假设你 365天,迟到30次
Pr(迟到) = 30/365
假如,你这一年中,有60次拉肚子,其中20次迟到了
Pr(迟到 | 拉肚子) = 20 / 60
= Pr(迟到&拉肚子) / Pr(拉肚子)
= (20 / 365) / (60 / 365)
= 20 / 60
Pr(你迟到 | 伊利诺伊发生车祸) = Pr(你迟到)
= Pr(你迟到&伊利诺伊车祸) / Pr(伊利诺伊车祸)
Pr(你迟到&伊利诺伊车祸) = Pr(你迟到) * Pr(伊利诺伊车祸)
Pr(你迟到 | 肚子痛 & 伊利诺伊发生车祸) = Pr(你迟到 | 肚子痛)
= Pr(你迟到&肚子痛) / Pr(肚子疼)
~ Count(你迟到且肚子痛) / Count(肚子痛)
Pr(其实就和随机森林原理一样)
-> Pr(其实&就和&随机森林&原理&一样)
-> Pr(其实|就和&随机森林&原理&一样)Pr(就和&随机森林&原理&一样)
—> Pr(其实|就和)Pr(就和&随机森林&原理&一样)
-> Pr(其实|就和)Pr(就和|随机森林&原理&一样)Pr(随机森林&原理&一样)
-> Pr(其实|就和)Pr(就和|随机森林)Pr(随机森林&原理&一样)
-> Pr(其实|就和)Pr(就和|随机森林)Pr(随机森林|原理)Pr(原理&一样)
-> Pr(其实|就和)Pr(就和|随机森林)Pr(随机森林|原理)Pr(原理|一样)Pr(一样)
语言学家们做了一个简化
P r ( s e n t e n c e ) = P r ( w 1 w 2 w 3 w 4 ) = ∏ i n # W i W i + 1 # W i + 1 ∗ P r ( w n ) Pr(sentence) = Pr(w_1w_2w_3w_4) = \prod_i^{n} \frac{\# W_iW_{i+1}}{\# W_{i+1}} * Pr(w_n) Pr(sentence)=Pr(w1w2w3w4)=i∏n#Wi+1#WiWi+1∗Pr(wn)
l a n g u a g e _ m o d e l ( S t r i n g ) = P r o b a b i l i t y ( S t r i n g ) ∈ ( 0 , 1 ) language\_model(String) = Probability(String) \in (0, 1) language_model(String)=Probability(String)∈(0,1)
P r o ( w 1 w 2 w 3 w 4 ) = P r ( w 1 ∣ w 2 w 3 w 4 ) ∗ P ( w 2 ∣ w 3 w 4 ) ∗ P r ( w 3 ∣ w 4 ) ∗ P r ( w 4 ) Pro(w_1 w_2 w_3 w_4) = Pr(w_1 | w_2 w_3 w_ 4) * P(w2 | w_3 w_4) * Pr(w_3 | w_4) * Pr(w_4) Pro(w1w2w3w4)=Pr(w1∣w2w3w4)∗P(w2∣w3w4)∗Pr(w3∣w4)∗Pr(w4)
P r o ( w 1 w 2 w 3 w 4 ) ∼ P r ( w 1 ∣ w 2 ) ∗ P ( w 2 ∣ w 3 ) ∗ P r ( w 3 ∣ w 4 ) ∗ P r ( w 4 ) Pro(w_1 w_2 w_3 w_4) \sim Pr(w_1 | w_2 ) * P(w2 | w_3 ) * Pr(w_3 | w_4) * Pr(w_4) Pro(w1w2w3w4)∼Pr(w1∣w2)∗P(w2∣w3)∗Pr(w3∣w4)∗Pr(w4)
import random
random.choice(range(100))
filename = '/Users/gaominquan/Downloads/sqlResult_1558435.csv'
import pandas as pd
content = pd.read_csv(filename, encoding='gb18030')
content.head()
articles = content['content'].tolist()
len(articles)
invalid
articles[0]
import re # 正则表达式
def token(string):
# we will learn the regular expression next course.
return re.findall('\w+', string)
token(articles[0])
import jieba
list(jieba.cut('这个是用来做汉语分词的'))
from collections import Counter
with_jieba_cut = Counter(jieba.cut(articles[110]))
with_jieba_cut.most_common()[:10]
''.join(token(articles[110]))
articles_clean = [''.join(token(str(a)))for a in articles]
len(articles_clean)
len(articles_clean)
假如,你做了很久的数据预处理
AI的问题里边,65%都是在做数据预处理
我们要养成一个习惯,就是把重要的信息,及时保存起来
存到硬盘里
with open('article_9k.txt', 'w') as f:
for a in articles_clean:
f.write(a + '\n')
!ls
def cut(string): return jieba.cut(string)
import jieba
def cut(string): return jieba.cut(string)
ALL_TOKEN = cut(open('article_9k.txt').read())
TOKEN = []
for i, t in enumerate(ALL_TOKEN):
if i > 50000: break
# 大家把它变成20万
if i % 1000 == 0: print(i)
TOKEN.append(t)
len(TOKEN)
from functools import reduce
from operator import add, mul
reduce(add, [1, 2, 3, 4, 5, 8])
[1, 2, 3] + [3, 43, 5]
from collections import Counter
words_count = Counter(TOKEN)
words_count.most_common(100)
frequiences = [f for w, f in words_count.most_common(100)]
x = [i for i in range(100)]
%matplotlib inline
import matplotlib.pyplot as plt
plt.plot(x, frequiences)
NLP比较重要的规律:在很大的一个text corpus,文字集合中,出现频率第二多的单词,是出现频率第一多单词的频率的1/2, 出现频率第n多的单词,是出现频率最高的单词的1/n.
import numpy as np
P r ( s e n t e n c e ) = P r ( w 1 w 2 w 3 w 4 ) = ∏ i n # W i W i + 1 # W i + 1 ∗ P r ( w n ) Pr(sentence) = Pr(w_1w_2w_3w_4) = \prod_i^{n} \frac{\# W_iW_{i+1}}{\# W_{i+1}} * Pr(w_n) Pr(sentence)=Pr(w1w2w3w4)=i∏n#Wi+1#WiWi+1∗Pr(wn)
words_count['我们']
def prob_1(word):
return words_count[word] / len(TOKEN)
prob_1('我们')
TOKEN = [str(t) for t in TOKEN]
TOKEN_2_GRAM = [''.join(TOKEN[i:i+2]) for i in range(len(TOKEN[:-2]))]
TOKEN_2_GRAM[10:]
words_count_2 = Counter(TOKEN_2_GRAM)
def prob_1(word): return words_count[word] / len(TOKEN)
def prob_2(word1, word2):
if word1 + word2 in words_count_2:
return words_count_2[word1+word2] / words_count[word2]
else: # out of vocabulary problem
return 1 / len(words_count)
prob_2('我们', '在')
prob_2('在', '吃饭')
prob_2('用', '手机')
P r ( s e n t e n c e ) = P r ( w 1 w 2 w 3 w 4 ) = ∏ i n # W i W i + 1 # W i + 1 ∗ P r ( w n ) Pr(sentence) = Pr(w_1w_2w_3w_4) = \prod_i^{n} \frac{\# W_iW_{i+1}}{\# W_{i+1}} * Pr(w_n) Pr(sentence)=Pr(w1w2w3w4)=i∏n#Wi+1#WiWi+1∗Pr(wn)
def get_probablity(sentence):
words = list(cut(sentence))
sentence_pro = 1
for i, word in enumerate(words[:-1]):
next_ = words[i+1]
probability = prob_2(word, next_)
sentence_pro *= probability
sentence_pro *= prob_1(words[-1])
return sentence_pro
get_probablity('小明今天抽奖抽到一台苹果手机')
get_probablity('小明今天抽奖抽到一架波音飞机')
get_probablity('洋葱奶昔来一杯')
get_probablity('养乐多绿来一杯')
host
need_compared = [
"今天晚上请你吃大餐,我们一起吃日料 明天晚上请你吃大餐,我们一起吃苹果",
"真事一只好看的小猫 真是一只好看的小猫",
"今晚我去吃火锅 今晚火锅去吃我",
"洋葱奶昔来一杯 养乐多绿来一杯"
]
for s in need_compared:
s1, s2 = s.split()
p1, p2 = get_probablity(s1), get_probablity(s2)
better = s1 if p1 > p2 else s2
print('{} is more possible'.format(better))
print('-'*4 + ' {} with probility {}'.format(s1, p1))
print('-'*4 + ' {} with probility {}'.format(s2, p2))