如何用gensim过滤掉语料库中tf-idf低的单词？

Question

我正在使用gensim 执行一些 NLP 任务。我从dictionary.doc2bow 创建了一个语料库，其中dictionary 是corpora.Dictionary 的对象。现在我想在运行 LDA 模型之前过滤掉具有低 tf-idf 值的项。我查看了语料库类的documentation，但找不到访问这些术语的方法。有任何想法吗？谢谢。

Answer 1

假设您有一个由 gensim 的 TfidfModel() 生成的文档 tfidf_doc 和相应的词袋文档 bow_doc，并且您想要过滤 tfidf 值低于该文档中 cut_percent% 单词的单词，你可以调用tfidf_filter(tfidf_doc, cut_percent)，然后它会返回tfidf_doc的剪切版本：

def tfidf_filter(tfidf_doc, cut_percent):

    sorted_by_tfidf = sorted(tfidf_doc, key=lambda tup: tup[1])
    cut_value = sorted_by_tfidf[int(len(sorted_by_tfidf)*cut_percent)][1]

    #print('before cut:',len(tfidf_doc))

    #print('cut value:', cut_value)
    for i in range(len(tfidf_doc)-1, -1, -1):
        if tfidf_doc[i][1] < cut_value:
            tfidf_doc.pop(i)

    #print('after cut:',len(tfidf_doc))

    return tfidf_doc

然后你想通过生成的tfidf_doc过滤文档bow_doc，只需调用filter_bow_by_tfidf(bow_doc, tfidf_doc)，它将返回bow_doc的剪切版本：

def filter_bow_by_tfidf(bow_doc, tfidf_doc):
    bow_idx = len(bow_doc)-1
    tfidf_idx = len(tfidf_doc)-1

    #print('before :', len(bow_doc))

    while True:
        if bow_idx < 0: break

        if tfidf_idx < 0:
            #print('pop2 :', bow_doc.pop(bow_idx))
            bow_doc.pop(bow_idx)
            bow_idx -= 1
        if bow_doc[bow_idx][0] > tfidf_doc[tfidf_idx][0]:
            #print('pop1 :', bow_doc.pop(bow_idx))
            bow_doc.pop(bow_idx)
            bow_idx -= 1
        if bow_doc[bow_idx][0] == tfidf_doc[tfidf_idx][0]:
            #print('keep :', bow_doc[bow_idx])
            bow_idx -= 1
            tfidf_idx -= 1

    #print('after :', len(bow_doc))

    return bow_doc

Answer 2

这与之前的答案基本相同，但额外处理了 tf-idf 表示中由于 0 分而缺失的单词（所有文档中都存在术语）。以前的答案没有过滤这些术语，它们仍然出现在最终的语料库中。

#Same as before

dictionary = corpora.Dictionary(doc_list)
corpus = [dictionary.doc2bow(doc) for doc in doc_list]
tfidf = models.TfidfModel(corpus, id2word = dictionary)


#Filter low value words and also words missing in tfidf models.

low_value = 0.025

for i in range(0, len(corpus)):
    bow = corpus[i]
    low_value_words = [] #reinitialize to be safe. You can skip this.
    tfidf_ids = [id for id, value in tfidf[bow]]
    bow_ids = [id for id, value in bow]
    low_value_words = [id for id, value in tfidf[bow] if value < low_value]
    words_missing_in_tfidf = [id for id in bow_ids if id not in tfidf_ids] # The words with tf-idf socre 0 will be missing

    new_bow = [b for b in bow if b[0] not in low_value_words and b[0] not in words_missing_in_tfidf]  

#reassign        
corpus[i] = new_bow

Answer 3

这是旧的，但如果您想查看每个文档级别的内容，请执行以下操作：

#same as before
dictionary = corpora.Dictionary(doc_list)
corpus = [dictionary.doc2bow(doc) for doc in doc_list]
tfidf = models.TfidfModel(corpus, id2word = dictionary)

#filter low value words
low_value = 0.025

for i in range(0, len(corpus)):
    bow = corpus[i]
    low_value_words = [] #reinitialize to be safe. You can skip this.
    low_value_words = [id for id, value in tfidf[bow] if value < low_value]
    new_bow = [b for b in bow if b[0] not in low_value_words]

    #reassign        
    corpus[i] = new_bow

Answer 4

假设你的语料库如下：

corpus = [dictionary.doc2bow(doc) for doc in documents]

运行 TFIDF 后，您可以检索低值单词列表：

tfidf = TfidfModel(corpus, id2word=dictionary)

low_value = 0.2
low_value_words = []
for bow in corpus:
    low_value_words += [id for id, value in tfidf[bow] if value < low_value]

然后在运行 LDA 之前将它们从字典中过滤掉：

dictionary.filter_tokens(bad_ids=low_value_words)

现在重新计算语料库，过滤掉低价值的词：

new_corpus = [dictionary.doc2bow(doc) for doc in documents]