1 目标

从中文文本中识别出句子里的情绪。和上一章节单车预测回归问题相比，这个问题是分类问题，不是回归问题

2 神经网络分类器

2.1 如何用神经网络分类

第二章节用torch.nn.Sequantial做的回归预测器，输出神经元只有一个。分类器和其区别如下

1 分类器输出单元有几个，就有几个分类

2 分类器各输出单元输出值加和为1（值表示某个预测分类的概率，概率求和为1）

3 回归预测最后一层可以用sigmoid，分类不行，因为sigmoid无法保证分类器输出单元值求和为1，可以用softmax，表达式如下

2.2 分类问题损失函数

分类问题损失函数一般用交叉熵：

N表示样本数量，Ci表示第i个样本的类别标签，y表示每个样本的损失值

3 词袋模型分类器

直白理解：就是统计句子里各词语出现数量，忽略语法，语义等因素

思想：1 将句子转化为词袋，然后可以onehot处理，方便用神经网络预测；2 大概思想：在词袋里寻找和输出相关关键词语，然后统计词语数量，哪个输出类别的敏感词语统计数量高，结果就是哪个输出类别

数据预处理：可以归一化处理，网络计算会更快

3.1 简单文本分类器

背景与数据：获取网购商城顾客评论信息，从评论信息获取句子情绪

神经网络组成：用前馈神经网络，有3层，分别是输入，中间，输出层，输出层数量有两个，分别代表情绪是正面还是负面

3.2 程序实现

3.2.1 数据处理

主要步骤：处理标点，句子分词，建立词表

标点处理：正则，涉及的标点全部删除

句子分词：使用jieba模块对句子分词

建立词表：python字典建立词表

def deal_punc(sentence):sentence = re.sub('[\s+\.\!\/_,$%^*(+\"\'“”《》?]+|[+——！，。？、~@#￥%……&*（）：)]+', '', sentence)return sentencedef prepare_data(good_file, bad_file, is_filter=True):print('prepare data begin')all_words = []pos_sen, neg_sen = [], []for path, sen in zip((good_file, bad_file), (pos_sen, neg_sen)):with open(path, 'r', encoding='utf-8') as f:for index, line in enumerate(f):if is_filter:line = deal_punc(line)words = jieba.lcut(line)if len(words) > 0:all_words += wordssen.append(words)print(f'{path} include {index} rows, all words:{len(all_words)}')print(f'pos_sen len:{len(pos_sen)}, neg_sen len:{len(neg_sen)}')diction = {}cnt = Counter(all_words)for word, freq in cnt.items():diction[word] = [len(diction), freq]print(f'diction len:{len(diction)}')return (pos_sen, neg_sen, diction)def word2index(word, diction):if word in diction:value = diction[word][0]else:value = -1return (value)def index2word(index, diction):for w, v in diction.items():if v[0] == index:return (w)return (None)

3.2.2 文本数据向量化

    pos_sen, neg_sen, diction = prepare_data_sample(good_file, bad_file)st = sorted([(v[1], w) for w, v in diction.items()])datasets, labels, sentences = [], [], []'''for sens, tag in zip((pos_sen, neg_sen), (0, 1)):for sen in sens:new_sen = []for l in sen:if l in diction:new_sen.append(word2index_sample(1, diction))datasets.append(sentence2vec_sample(new_sen, diction))labels.append(tag)sentences.append(sen)'''for sentence in pos_sen:new_sentence = []for l in sentence:if l in diction:new_sentence.append(word2index(l, diction))datasets.append(sentence2vec_sample(new_sentence, diction))labels.append(0) #正标签为0sentences.append(sentence)# 处理负向评论for sentence in neg_sen:new_sentence = []for l in sentence:if l in diction:new_sentence.append(word2index(l, diction))datasets.append(sentence2vec_sample(new_sentence, diction))labels.append(1) #负标签为1sentences.append(sentence)indices = np.random.permutation(len(datasets))datasets = [datasets[i] for i in indices]labels = [labels[i] for i in indices]sentences = [sentences[i] for i in indices]

3.3.3 划分数据集

训练集，校验集，测试集，比例一般为10：1：1

    # split train, test, verify datasetstest_size = int(len(datasets) // 10)train_data = datasets[2 * test_size :]train_label = labels[2 * test_size : ]valid_data = datasets[: test_size]valid_label = labels[: test_size]test_data = datasets[test_size : 2 * test_size]test_label = labels[test_size : 2 * test_size]

3.3.4 建立神经网络

结构：输入层（约7000个样本），一个中间层（10个隐单元），输出层（2个单元）

注意，此处用的是relu，不是sigmoid。原因：虽然x>0时没处理数据，但x<0时为0让relu有了非线性的能力，且运算比sigmoid简单，速度更快，方便反向误差传导

def plot(records):print('plot begin')a = [i[0] for i in records]b = [i[1] for i in records]c = [i[2] for i in records]pyplot.plot(a, label='train loss')pyplot.plot(b, label='verify loss')pyplot.plot(c, label='valid accuracy')pyplot.xlabel('step')pyplot.ylabel('losses & accuracy')pyplot.legend()pyplot.show()def main():model = torch.nn.Sequential(torch.nn.Linear(len(diction), 10),torch.nn.ReLU(),torch.nn.Linear(10, 2),torch.nn.LogSoftmax(dim=1),)cost = torch.nn.NLLLoss()optimizer = torch.optim.SGD(model.parameters(), lr=0.01)records = []losses = []for epoch in range(3):for i, data in enumerate(zip(train_data, train_label)):x, y = datax = torch.tensor(x, requires_grad=True, dtype=torch.float).view(1, -1)y = torch.tensor(np.array([y]), dtype=torch.long)optimizer.zero_grad()predict = model(x)loss = cost(predict, y)losses.append(loss.data.numpy())loss.backward()optimizer.step()val_losses = []rights = []for j, val in enumerate(zip(valid_data, valid_label)):x, y = valx = torch.tensor(x, requires_grad=True, dtype=torch.float).view(1, -1)y = torch.tensor(np.array([y]), dtype = torch.long)predict = model(x)right = rightness_sample(predict, y)rights.append(right)loss = cost(predict, y)val_losses.append(loss.data.numpy())right_ratio = 1.0 * np.sum([i[0] for i in rights]) / np.sum([i[1] for i in rights])print(f'No.{epoch}, train loss:{np.mean(losses):.4f}, verify loss:{np.mean(val_losses):.4f}, verify accuracy:{right_ratio}')records.append([np.mean(losses), np.mean(val_losses), right_ratio])# plotplot(records)

3.3 运行结果