BiLSTM+CRF基于pytorch实现命名实体识别，对pytorch官网给出的例子实现优化，附学习思路和源代码

学习思路

BiLSTM + CRF

本文建立在读者已有一定深度学习知识的基础上，可以看看pytorch实现验证码图片识别，基本的反向传播思想，以及机器学习的相关概念（损失函数、随机梯度下降、训练集测试集等等）。

LSTM看这一篇就够了。

接下来给出我学习CRF的思路

• 看这篇的Section1,2,3，了解基本的朴素贝叶斯和最大熵模型，我个人觉得这篇CRF的部分讲得不清楚
• 看这篇全部，了解CRF的特征向量、CRF建模的对象、对数极大似然函数以及CRF的损失函数、以及viterbi解码（decode）和迭代计算所有路径分数的思想
• 再对照这篇博客看pytorch官网BiLSTM+CRF教程中对CRF的路径分数等的定义，初步实现模型
• 最后，基于这篇博客实现简化版模型，但我觉得这篇博客没搞清楚特征向量和分数的关系，所以只要看迭代计算那里就行

Requirements

• python==3.7.6，
• torch==1.4.0+cu92
• 安装环境可以参考这篇博客

运行结果

随便编了3个句子，用前两句训练100轮后预测第三句，结果如下：

可以看到效果还是不错的。

源代码

import torch
import torch.autograd as autograd
import torch.nn as nn
import torch.optim as optim
from utils import prepare_sequence
def prepare_sequence(seq, to_ix):
    idxs = [to_ix[w] for w in seq]
    return torch.tensor(idxs, dtype=torch.long)

class BiLSTM_CRF(nn.Module):
    def __init__(self, vocab_size, tag_to_ix, embedding_dim, hidden_dim):
        super().__init__()

        self.embedding_dim = embedding_dim
        self.hidden_dim = hidden_dim
        self.vocab_size = vocab_size
        self.tag_to_ix = tag_to_ix
        self.tagset_size = len(tag_to_ix)

        self.word_embeds = nn.Embedding(vocab_size, embedding_dim)
        self.lstm = nn.LSTM(embedding_dim, hidden_dim // 2,
                            num_layers=1, bidirectional=True)

        self.hidden2tag = nn.Linear(hidden_dim, self.tagset_size)

        self.transitions = nn.Parameter(
            torch.randn(self.tagset_size, self.tagset_size))
        # never transfer to the start tag and never transfer from stop tag
        self.transitions.data[tag_to_ix[START_TAG], :] = -10000
        self.transitions.data[:, tag_to_ix[STOP_TAG]] = -10000

    def _viterbi_decode(self, feats):
        # backpointers:(n,m) where n is length of feats and m is tagset_size
        # 存储feat中到当前所有tag最优的tag（相当于一步延迟）
        # backpointers[i]意味着到feat_i的各个tags的最大可能的tag_pre
        backpointers = []

        # 初始化一个表示开始前的feat的向量，除了start_tag为0，其余都是-10000，这样由于argmax则feat_0的各个tags都以start_tag为前序
        init_vvars = torch.full((1, self.tagset_size), -10000.)
        init_vvars[0][self.tag_to_ix[START_TAG]] = 0.

        # 将初始状态寄存在forward_var中，用作后续更新
        pre_feat_var = init_vvars[0]

        # 对于句子中每一个字的标签特征表示
        for feat in feats:
            pre_compute_var = torch.stack([pre_feat_var]*self.tagset_size)
            crt_feat_var = pre_compute_var + self.transitions
            pre_feat_var, best_pre = torch.max(crt_feat_var, dim=1)
            backpointers.append(best_pre.tolist())

            pre_feat_var += feat

        terminal_var = pre_feat_var + \
            self.transitions[self.tag_to_ix[STOP_TAG]]
        total_var, best_tag_id = torch.max(terminal_var, dim=0)

        best_path = [best_tag_id.item()]
        for backpointer in reversed(backpointers):
            best_tag_id = backpointer[best_tag_id]
            best_path.append(best_tag_id)
        start = best_path.pop()
        assert start == self.tag_to_ix[START_TAG]
        best_path.reverse()
        return total_var, best_path

    def _forward_alg(self, feats):
        # 初始化，是否需要初始化为-10000还需研究
        init_vars = torch.full((1, self.tagset_size), -10000)
        init_vars[0][self.tag_to_ix[START_TAG]] = 0
        #init_vars = torch.zeros((1,self.tagset_size))

        # pre_feat:(1,tagset_size),用来保存到前序feat的各个tag的score
        pre_feat = init_vars[0]

        # 对于每一个feat
        for feat in feats:
            # 将pre_feat铺为tagset_size*tagset_size，方便计算
            pre_compute_var = torch.stack([pre_feat]*self.tagset_size)
            # 当前feat的Emission Vector即crt_emit_var[i]为当前feat被标为第i个tag的非归一化概率
            crt_emit_var = torch.unsqueeze(feat, dim=0).view(-1, 1)
            # 当前feat的总特征向亮=Emission + Transition，crt_feat_var[i][j]为当前feat的前序feat被标为第j个tag的情况下当前feat被标为第i个tag的特征值
            crt_feat_var = crt_emit_var+self.transitions
            # 迭代计算，因为log(exp(log(exp(\sum{x}))+y)) = log(exp(\sum{x})+exp(y))，得到当前feat的总score
            feat_var = pre_compute_var + crt_feat_var
            # 更新pre_feat，计算logsumexp即到当前feat的各个tag的各种情况的score
            pre_feat = feat_var.logsumexp(dim=1)

        # 同理，计算到终点的总score，即所有情况（len(feats)^{tagset_size}种）的总score
        terminal_var = pre_feat + self.transitions[self.tag_to_ix[STOP_TAG]]
        total_score = terminal_var.unsqueeze(dim=0).logsumexp(dim=1)[0]
        return total_score

    def _get_lstm_features(self, sentence):

        embeds = self.word_embeds(sentence).view(len(sentence), 1, -1)
        lstm_out, _ = self.lstm(embeds)
        lstm_out = lstm_out.view(len(sentence), self.hidden_dim)
        lstm_feats = self.hidden2tag(lstm_out)
        return lstm_feats

    def _score_sentence(self, feats, tags):
        # Gives the score of a provided tag sequence
        score = torch.zeros(1)
        tags = torch.cat(
            [torch.tensor([self.tag_to_ix[START_TAG]], dtype=torch.long), tags])
        for i, feat in enumerate(feats):
            score = score + \
                self.transitions[tags[i + 1], tags[i]] + feat[tags[i + 1]]
        score = score + self.transitions[self.tag_to_ix[STOP_TAG], tags[-1]]
        return score

    # 损失函数，来源于log的最大似然函数
    def neg_log_likelihood(self, sentence, tags):
        feats = self._get_lstm_features(sentence)
        total_score = self._forward_alg(feats)
        tagged_score = self._score_sentence(feats, tags)
        return total_score - tagged_score

    def forward(self, sentence):  # dont confuse this with _forward_alg above.
        # Get the emission scores from the BiLSTM
        lstm_feats = self._get_lstm_features(sentence)

        # Find the best path, given the features.
        score, tag_seq = self._viterbi_decode(lstm_feats)
        return score, tag_seq


if __name__ == "__main__":
    START_TAG = ""
    STOP_TAG = ""
    EMBEDDING_DIM = 300
    HIDDEN_DIM = 256

    training_data = [(
        "参 加 北 京 智 源 大 会".split(),
        "O O B I I I I I I ".split()
    ), (
        "他 出 席 计 算 机 协 会 会 长".split(),
        "O O O B I I I I O O".split()
    ),
        (
        "作 为 主 讲 人 参 加 教 研 会 应 对 疫 情".split(),
        "O O O O O O O B I I O O O O".split()
    )]

    word_to_ix = {}
    for sentence, tags in training_data:
        for word in sentence:
            if word not in word_to_ix:
                word_to_ix[word] = len(word_to_ix)
    tag_to_ix = {"B": 0, "I": 1, "O": 2, START_TAG: 3, STOP_TAG: 4}

    model = BiLSTM_CRF(len(word_to_ix), tag_to_ix, EMBEDDING_DIM, HIDDEN_DIM)
    optimizer = optim.SGD(model.parameters(), lr=0.01, weight_decay=1e-4)

    # 测试为训练模型
    with torch.no_grad():
        precheck_sent = prepare_sequence(training_data[-1][0], word_to_ix)
        precheck_tags = torch.tensor(
            [tag_to_ix[t] for t in training_data[0][1]], dtype=torch.long)
        print(model(precheck_sent))

    for epoch in range(100):
        for sentence, tags in training_data[0:-1]:
            model.zero_grad()

            sentence_in = prepare_sequence(sentence, word_to_ix)
            targets = torch.tensor([tag_to_ix[t]
                                    for t in tags], dtype=torch.long)
            loss = model.neg_log_likelihood(sentence_in, targets)

            loss.backward()
            optimizer.step()

    with torch.no_grad():
        for i in range(len(training_data)):
            precheck_sent = prepare_sequence(training_data[i][0], word_to_ix)
            print("predict:{}\n target:{}".format(model(precheck_sent),
                                                    prepare_sequence(training_data[i][1], tag_to_ix)))