为什么我的.pickle保存的分类器给出的结果与在一个程序中训练和分类不同

Question

我以前在stack上见过这个问题，但是解决方案对我不起作用，请参见Save and Load testing classify Naive Bayes Classifier in NLTK in another method。当我加载pickled分类器而不是在同一个程序中进行训练和分类时，为什么我的准确度相差甚远。第一个代码块调用pickled分类器，第二个代码块一起进行所有的训练和分类。第二种方法的准确度为99%，而第一种方法的准确度为81%。。。在

Academic_classifier= None
Academic_classifier=pickle.load(open('Academic_classifier.pickle','rb'))
tweets=[]

readdata=csv.reader(open('C:\Users\Troy\Documents\Data\Gold_test.csv','r'))
for row in readdata:
    tweets.append(row)
Header = tweets[0]
tweets.pop(0)
Academic_test_tweets=tweets[:]
Tweets=[]
for (words, sentiment) in tweets:
    bigram=[]
    bigram_list=[]
    words_filtered = [e.lower() for e in WordPunctTokenizer().tokenize(words) if len(e) >= 3]
    words_filtered=[re.sub(r'(.)\1+', r'\1\1', e) for e in words_filtered if len(e)>=3]
    bigram_words=bigrams(words_filtered)
    for x in bigram_words:
        bigram.append(x)
    for (bi) in bigram:
        bigram_word=bi[0]+bi[1]
        bigram_list.append(bigram_word)
    list_to_append=words_filtered+bigram_list
    Tweets.append((list_to_append, sentiment))
Academic_test_tweets_words=Tweets[:]


word_features = get_word_features(get_words_in_tweets(Academic_test_tweets_words))
Academic_test_set = nltk.classify.apply_features(extract_features,Academic_test_tweets_words)

print(nltk.classify.accuracy(Academic_classifier, Academic_test_set), 'tweet corpus used in academic paper Sentiment Analysis on the Social Networks Using Stream Algorithms Authors: Nathan Aston, Timothy Munson, Jacob Liddle, Garrett Hartshaw, Dane Livingston, Wei Hu   *compare to their accuracy of 87.5%')

与我训练和测试准确性的代码相反。我对每件事都使用相同的定义，所以我知道问题不在于定义。唯一的区别是腌制分类器。。。发生什么事了？在

^{pr2}$