Python NLTK 情感分析不正确

1、问题背景

一位 Reddit 用户使用 Python 的 NLTK 库来训练一个朴素贝叶斯分类器以研究其他句子的情感，但是无论输入什么句子，分类器总是预测为正面。

2、解决方案

经过仔细检查，发现原始代码中的问题在于 wordList 为空。因此，需要将 wordList 赋值为从推文中提取的单词特征。修改后的代码如下：

wordList = getwordfeatures(getwords(tweets))
wordList = [i for i in wordList if not i in stopwords.words('english')]
wordList = [i for i in wordList if not i in customstopwords]

以下是完整的修复代码：

import nltk
import math
import re
import sys
import os
import codecs
reload(sys)
sys.setdefaultencoding('utf-8')

from nltk.corpus import stopwords

__location__ = os.path.realpath(
    os.path.join(os.getcwd(), os.path.dirname(__file__)))

postweet = __location__ + "/postweet.txt"
negtweet = __location__ + "/negtweet.txt"

customstopwords = ['band', 'they', 'them']

# Load positive tweets into a list
p = open(postweet, 'r')
postxt = p.readlines()

# Load negative tweets into a list
n = open(negtweet, 'r')
negtxt = n.readlines()

neglist = []
poslist = []

# Create a list of 'negatives' with the exact length of our negative tweet list.
for i in range(0, len(negtxt)):
    neglist.append('negative')

# Likewise for positive.
for i in range(0, len(postxt)):
    poslist.append('positive')

# Creates a list of tuples, with sentiment tagged.
postagged = zip(postxt, poslist)
negtagged = zip(negtxt, neglist)

# Combines all of the tagged tweets to one large list.
taggedtweets = postagged + negtagged

tweets = []

# Create a list of words in the tweet, within a tuple.
for (word, sentiment) in taggedtweets:
    word_filter = [i.lower() for i in word.split()]
    tweets.append((word_filter, sentiment))

# Pull out all of the words in a list of tagged tweets, formatted in tuples.
def getwords(tweets):
    allwords = []
    for (words, sentiment) in tweets:
        allwords.extend(words)
    return allwords

# Order a list of tweets by their frequency.
def getwordfeatures(listoftweets):
    # Print out wordfreq if you want to have a look at the individual counts of words.
    wordfreq = nltk.FreqDist(listoftweets)
    words = wordfreq.keys()
    return words

# Calls above functions - gives us list of the words in the tweets, ordered by freq.
print(getwordfeatures(getwords(tweets)))

wordList = getwordfeatures(getwords(tweets))
wordList = [i for i in wordList if not i in stopwords.words('english')]
wordList = [i for i in wordList if not i in customstopwords]

def feature_extractor(doc):
    docwords = set(doc)
    features = {}
    for i in wordList:
        features['contains(%s)' % i] = (i in docwords)
    return features

# Creates a training set - classifier learns distribution of true/falses in the input.
training_set = nltk.classify.apply_features(feature_extractor, tweets)
classifier = nltk.NaiveBayesClassifier.train(training_set)

print(classifier.show_most_informative_features(n=30))

while True:
    input = raw_input('ads')
    if input == 'exit':
        break
    elif input == 'informfeatures':
        print(classifier.show_most_informative_features(n=30))
        continue
    else:
        input = input.lower()
        input = input.split()
        print('\nWe think that the sentiment was ' + classifier.classify(feature_extractor(input)) + ' in that sentence.\n')

p.close()
n.close()

用户可以根据需要调整 customstopwords 列表以过滤掉不相关的词语。