在上一节的情感分类当中，有些评论是负面的，但预测的结果是正面的，比如，"this movie was shit"这部电影是狗屎，很明显就是对这部电影极不友好的评价，属于负类评价，给出的却是positive。

所以这节我们通过专门的“分词”和“扩大词向量维度”这两个途径来改进，提高预测的准确率。

spaCy分词

我们用spaCy分词工具来进行分词看是否能提高准确性。

推荐带上镜像站点来下载并安装。

pip install spacy -i http://pypi.douban.com/simple/  --trusted-host pypi.douban.com

import spacy
>>> spacy.__version__
'3.0.9'

安装英文包

python -m spacy download en

这种方法我没有安装成功，于是我选择直接下载安装，感觉太慢选择迅雷下载：https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-3.0.0/en_core_web_sm-3.0.0-py3-none-any.whl

或者：

pip install https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-3.0.0/en_core_web_sm-3.0.0-py3-none-any.whl

这里选择的是en_core_web_sm语言包，所以也可以直接选择豆瓣镜像下载《推荐这种方法》

pip install en_core_web_sm-3.0.0-py3-none-any.whl  -i http://pypi.douban.com/simple/  --trusted-host pypi.douban.com

安装好之后，就可以通过spacy来加载这个英文包

spacy_en = spacy.load("en_core_web_sm")
>>> spacy_en._path
WindowsPath('D:/Anaconda3/envs/pygpu/lib/site-packages/en_core_web_sm/en_core_web_sm-3.0.0')

然后进行分词，将上一节或者说自带的get_tokenized_imdb函数修改下，使用修改的这个函数：

def get_tokenized_imdb(data):def tokenizer(text):return [tok.text for tok in spacy_en.tokenizer(text)]return [tokenizer(review) for review, _ in data]

我们训练看下效果如何：

print(d2l.predict_sentiment(net, vocab, ["this", "movie", "was", "shit"]))
print(d2l.predict_sentiment(net, vocab, ["this", "movie", "is", "not", "good"]))
print(d2l.predict_sentiment(net, vocab, ["this", "movie", "is", "so", "bad"]))
'''
training on [gpu(0)]
epoch 1, loss 0.5781, train acc 0.692, test acc 0.781, time 66.0 sec
epoch 2, loss 0.4024, train acc 0.822, test acc 0.839, time 65.4 sec
epoch 3, loss 0.3465, train acc 0.852, test acc 0.844, time 65.6 sec
epoch 4, loss 0.3227, train acc 0.861, test acc 0.856, time 65.9 sec
epoch 5, loss 0.2814, train acc 0.880, test acc 0.859, time 66.2 sec
negative
positive
negative
'''

可以看到准确率有提高，而且第一条影评在上一节预测是positive，这里预测为negative，正确识别了这条影评的负类评价。第二条影评的预测错误了，说明没有识别出not good属于负类评价，接下来我们再叠加一个方法来提高准确率。

300维度的词向量

我们将预处理文件的词向量从100维度提高到300维度看下准确度有没有上升，也就是选择glove.6B.300d.txt来替换glove.6B.100d.txt

glove_embedding = text.embedding.create("glove", pretrained_file_name="glove.6B.300d.txt", vocabulary=vocab
)

选择更高维度的词向量文档之后，我们做下训练测试看下：

print(d2l.predict_sentiment(net, vocab, ["this", "movie", "was", "shit"]))
print(d2l.predict_sentiment(net, vocab, ["this", "movie", "is", "not", "good"]))
print(d2l.predict_sentiment(net, vocab, ["this", "movie", "is", "so", "bad"]))
print(d2l.predict_sentiment(net, vocab, ["this", "movie", "is", "so", "good"]))
'''
training on [gpu(0)]
epoch 1, loss 0.5186, train acc 0.734, test acc 0.842, time 74.7 sec
epoch 2, loss 0.3411, train acc 0.854, test acc 0.862, time 74.8 sec
epoch 3, loss 0.2851, train acc 0.884, test acc 0.863, time 75.6 sec
epoch 4, loss 0.2459, train acc 0.903, test acc 0.843, time 75.3 sec
epoch 5, loss 0.2099, train acc 0.917, test acc 0.853, time 75.8 sec
negative
negative
negative
positive
'''

准确度再次有了提升，四条影评都被正确识别了情绪。

全部代码

import collections
import d2lzh as d2l
from mxnet import gluon, init, nd
from mxnet.contrib import text
from mxnet.gluon import data as gdata, loss as gloss, nn, rnn
import spacy#spacy_en = spacy.load("en")
spacy_en = spacy.load("en_core_web_sm")def get_tokenized_imdb(data):def tokenizer(text):return [tok.text for tok in spacy_en.tokenizer(text)]return [tokenizer(review) for review, _ in data]def get_vocab_imdb(data):"""Get the vocab for the IMDB data set for sentiment analysis."""tokenized_data = get_tokenized_imdb(data)counter = collections.Counter([tk for st in tokenized_data for tk in st])return text.vocab.Vocabulary(counter, min_freq=5, reserved_tokens=[""])# d2l.download_imdb(data_dir='data')
train_data, test_data = d2l.read_imdb("train"), d2l.read_imdb("test")
tokenized_data = get_tokenized_imdb(train_data)
vocab = get_vocab_imdb(train_data)
features, labels = d2l.preprocess_imdb(train_data, vocab)
batch_size = 64
# train_set = gdata.ArrayDataset(*d2l.preprocess_imdb(train_data, vocab))
train_set = gdata.ArrayDataset(*[features, labels])
test_set = gdata.ArrayDataset(*d2l.preprocess_imdb(test_data, vocab))
train_iter = gdata.DataLoader(train_set, batch_size, shuffle=True)
test_ieter = gdata.DataLoader(test_set, batch_size)"""
for X,y in train_iter:print(X.shape,y.shape)break
"""class BiRNN(nn.Block):def __init__(self, vocab, embed_size, num_hiddens, num_layers, **kwargs):super(BiRNN, self).__init__(**kwargs)# 词嵌入层self.embedding = nn.Embedding(input_dim=len(vocab), output_dim=embed_size)# bidirectional设为True就是双向循环神经网络self.encoder = rnn.LSTM(hidden_size=num_hiddens,num_layers=num_layers,bidirectional=True,input_size=embed_size,)self.decoder = nn.Dense(2)def forward(self, inputs):# LSTM需要序列长度(词数)作为第一维，所以inputs[形状为：(批量大小,词数)]需做转置# 输出就是(词数,批量大小,词向量维度)(500, 64, 100)->全连接层之后的形状(5,1,100)embeddings = self.embedding(inputs.T)# 双向循环所以乘以2(词数,批量大小,词向量维度*2)(500, 64, 200)->全连接层之后的形状(5,1,200)outputs = self.encoder(embeddings)# 将初始时间步和最终时间步的隐藏状态作为全连接层输入# (64, 400)->全连接层之后的形状(1,400)encoding = nd.concat(outputs[0], outputs[-1])outs = self.decoder(encoding)return outs# 创建一个含2个隐藏层的双向循环神经网络
embed_size, num_hiddens, num_layers, ctx = 300, 100, 2, d2l.try_all_gpus()
net = BiRNN(vocab=vocab, embed_size=embed_size, num_hiddens=num_hiddens, num_layers=num_layers
)
net.initialize(init.Xavier(), ctx=ctx)glove_embedding = text.embedding.create("glove", pretrained_file_name="glove.6B.300d.txt", vocabulary=vocab
)
net.embedding.weight.set_data(glove_embedding.idx_to_vec)
net.embedding.collect_params().setattr("grad_req", "null")lr, num_epochs = 0.01, 5
trainer = gluon.Trainer(net.collect_params(), "adam", {"learning_rate": lr})
loss = gloss.SoftmaxCrossEntropyLoss()
d2l.train(train_iter, test_ieter, net, loss, trainer, ctx, num_epochs)print(d2l.predict_sentiment(net, vocab, ["this", "movie", "was", "shit"]))
print(d2l.predict_sentiment(net, vocab, ["this", "movie", "is", "not", "good"]))
print(d2l.predict_sentiment(net, vocab, ["this", "movie", "is", "so", "bad"]))
print(d2l.predict_sentiment(net, vocab, ["this", "movie", "is", "so", "good"]))

其中需要注意的是embed_size的大小需设定为300，跟新选择的文件的词向量维度保持一致。

小结：从目前实验结果来看对词语的分词做的更好，对于理解词义是很有帮助的，另外将词映射成的向量维度越高，准确度也在提升。