最近在看莫烦的NLP的课程,其中关于TF-IDF算法实际编程的时候还是遇到一些小问题,主要是计算方法问题,整理后放上来,加深记忆。
TF-IDF的计算方法有很多种,这里主要用的是SKlearn中的计算方式,根示例代码不太一样,费了点劲儿才搞明白。
目录
2. IDF:Inverse Document Frequency,逆向文本频率
一、 TF-IDF算法简介
TF-IDF(term frequency–inverse document frequency,词频-逆向文件频率)算法是一种用于信息检索与文本数据挖掘的常用加权技术。它用统计学方法评估一个词对某篇文章的重要程度,常用来提取文章的关键词,算法简单高效,因此常用于信息检索的粗排阶段。
TF-IDF算法的核心思想是通过统计的方法,评估一个词对一个文件集或者语料库的重要程度。一个词的重要程度跟它在文章中出现的次数成正比,跟它在语料库出现的次数成反比。这种计算方式能有效避免常用词对关键词的影响,提高了关键词与文章之间的相关性。
1. TF:Term Frequency,词频
指的是某个词在某篇文章中出现的次数, 计算公式为:TF = 某词在某文档中出现的次数
也就是说,就一篇文章局部来看,一个单词出现的次数越多就越重要,但这并不是绝对的。比如,a、the、of等单词出现的次数一定不会少,很显然它们并没有什么重要信息。所以,我们接下来要引入IDF。
注意:也有 “TF = 某词在某文档中出现的次数 / 该文档的总词量” 这种计算,但SKLEARN中是采用直接计次。
2. IDF:Inverse Document Frequency,逆向文本频率
指的是某个词在一个文件集或者语料库中区分力指标。计算公式为:
其中,Nd是训练集文档总数量,df(d,t)是包含某个单词的文档数量,+1的原因是避免分母为0。
也就是说,对一个文件集或者语料库而言,包含某个单词的文档越少,IDF的值越大,这个词的区分力越强,就越重要。
特别需要注意的是,IDF是针对文件集或者语料库而言的。计算机领域的IDF用在医学领域往往是不合适的。
3. TF-IDF = TF * IDF
综合考虑以某篇文章为中心的局部信息TF,和以某个语料库全局信息为基础的IDF,得到以下公式:
TF-IDF = TF * IDF
特别注意:
在sklearn中,上述计算的TF-IDF会经过一个欧几里得范数归一化:
二、代码示例
以下代码改编自莫烦的NLP课程中的源码。
输入15篇文章,形成一个44个单词的词汇表(去掉两个高频词,a 和 i),计算这15篇文章的tf-idf矩阵。再输入查询语句,计算该语句的tf-idf向量,然后求该语句的tf-idf向量和每一篇文章tf-idf向量的cosin距离,找出距离最近的三篇文章即是搜索结果。
核心思想--向量化。将文章向量化,将待查询语句也向量化,就可以通过计算余弦距离来比较相近程度。注意这里使用两个向量的夹角的余弦值来衡量两个文本间的相似度,而不是常用的欧氏距离,余弦相似度更加注重两个向量在方向上的差异,而不是实际距离差异。
1. 自编函数
import numpy as np
from collections import Counter
import itertools
from sklearn import preprocessing
from plot import show_tfidf
# from sklearn.metrics.pairwise import cosine_similarity
#15 docs
docs = [
"it is a good day, I like to stay here",
"I am happy to be here",
"I am bob",
"it is sunny today",
"I have a party today",
"it is a dog and that is a cat",
"there are dog and cat on the tree",
"I study hard this morning",
"today is a good day",
"tomorrow will be a good day",
"I like coffee, I like book and I like apple",
"I do not like it",
"I am kitty, I like bob",
"I do not care who like bob, but I like kitty",
"It is coffee time, bring your cup",
]
#vocablist包括44 words 去掉两个超高频单词
docs_words = [d.lower().replace(",", "").split(" ") for d in docs]
wordlist = list(itertools.chain(*docs_words)) #遍历对象,去除内嵌,为什么需要加*没细看
vocablist = list(set(wordlist))
vocablist.sort(key=wordlist.index) #转set去重,保持原序 => 全部单词表
vocablist.remove('a') #为了根sklearn保持一致,去掉两个超高频单词
vocablist.remove('i')
#print(vocablist)
v2i = {v: i for i, v in enumerate(vocablist)} #给单词编索引,eg: 'tree': 0 #enumerate函数 index, value
i2v = {i: v for v, i in v2i.items()} #逆索引,eg: 0: 'tree' #items函数,value, index
#print(v2i)
#print(i2v)
# tf = 每个单词出现频率
def get_tf():
# term frequency: how frequent a word appears in a doc
_tf = np.zeros((len(vocablist), len(docs)), dtype=np.float64) # [n_vocab, n_doc] =》 [44 * 15]矩阵
for i, d in enumerate(docs_words): #循环每篇文章
counter = Counter(d)
for v in counter.keys(): #统计每篇文章单词计数
if v in v2i:
_tf[v2i[v], i] = counter[v] #每个单词出现频率
return _tf
# idf = 1 + np.log((len(docs) + 1) / (该单词在几篇文章中出现 + 1))
def get_idf(method="sklearn"):
# inverse document frequency: low idf for a word appears in more docs, mean less important
df = np.zeros((len(i2v), 1))
for i in range(len(i2v)): #循环词汇表每一个单词
d_count = 0
for d in docs_words:
d_count += 1 if i2v[i] in d else 0 #该单词在几篇文章中出现
df[i, 0] = d_count
idf_fn = lambda x: 1 + np.log((len(docs) + 1) / (x+1))
if idf_fn is None:
raise ValueError
return idf_fn(df)
def cosine_similarity(_tf_idf, q):
unit_ds = _tf_idf / np.sqrt(np.sum(np.square(_tf_idf)))
unit_q = q / np.sqrt(np.sum(np.square(q)))
similarity = unit_ds.T.dot(unit_q).ravel()
return similarity
#根据输入,比较每篇文章的相似度,不考虑输入句子中新加入的单词
def docs_score(q):
q_words = q.replace(",", "").split(" ")
counter = Counter(q_words)
q_tf = np.zeros((len(idf), 1), dtype=float)
for v in counter.keys():
if v in v2i:
q_tf[v2i[v], 0] = counter[v] # 每个单词出现频率
q_vec = q_tf * idf
q_tf_idf = preprocessing.normalize(q_vec, norm='l2', axis=0) # 欧几里得范数归一化
#q_scores = cosine_similarity(tf_idf.transpose(), q_tf_idf.transpose()) #如果用库中函数,就用归一化后的
q_scores = cosine_similarity(origin_tf_idf, q_vec) #如果自已写,就用未归一化的
return q_scores
#获得tf_idf最高的n个词
def get_keywords(n=2):
for c in range(15):
col = tf_idf[:, c]
idx = np.argsort(col)[-n:][::-1] #从小到大排列,提取其对应的index, 从后向前反向取
print("doc{}, top{} keywords {}".format(c, n, [i2v[i] for i in idx])) #TOP2,TOP1
#----------TEST
tf = get_tf() # [n_vocab, n_doc] 44*15
idf = get_idf() # [n_vocab, 1] 44*1
origin_tf_idf = tf * idf # [n_vocab, n_doc] 44*15
tf_idf = preprocessing.normalize(origin_tf_idf, norm='l2', axis=0) #欧几里得范数归一化
print("\ntf samples:\n", tf[:2])
print("\nidf sample:\n", idf[:2])
print("\ntf_idf sample:\n", tf_idf[:2])
#--- 提取关键词
get_keywords()
#--- 搜索最相似的句子
q = "I get a coffee cup"
scores = docs_score(q)
d_ids = scores.ravel().argsort()[-3:][::-1]
print("\ntop 3 docs for '{}':\n{}".format(q, [docs[i] for i in d_ids]))
show_tfidf(tf_idf.T, [i2v[i] for i in range(tf_idf.shape[0])], "tfidf_matrix")
2. 使用sklearn库
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.pairwise import cosine_similarity
from plot import show_tfidf
docs = [
"it is a good day, I like to stay here",
"I am happy to be here",
"I am bob",
"it is sunny today",
"I have a party today",
"it is a dog and that is a cat",
"there are dog and cat on the tree",
"I study hard this morning",
"today is a good day",
"tomorrow will be a good day",
"I like coffee, I like book and I like apple",
"I do not like it",
"I am kitty, I like bob",
"I do not care who like bob, but I like kitty",
"It is coffee time, bring your cup",
]
vectorizer = TfidfVectorizer()
tf_idf = vectorizer.fit_transform(docs)
# print("idf: ", [(n, idf) for idf, n in zip(vectorizer.idf_, vectorizer.get_feature_names())])
# print("v2i: ", vectorizer.vocabulary_)
# print(tf_idf)
q = "I get a coffee cup"
qtf_idf = vectorizer.transform([q])
res = cosine_similarity(tf_idf, qtf_idf)
res = res.ravel().argsort()[-3:]
print("\ntop 3 docs for '{}':\n{}".format(q, [docs[i] for i in res[::-1]]))
i2v = {i: v for v, i in vectorizer.vocabulary_.items()}
dense_tfidf = tf_idf.todense() #tf_idf为稀疏矩阵
show_tfidf(dense_tfidf, [i2v[i] for i in range(dense_tfidf.shape[1])], "tfidf_sklearn_matrix")