信息抽取（二）实体识别之经典方法介绍

regex = "d{17}[d|x|X]"text = "身份证号码信息是410403198805135531"id_number = re.findall(regex, text)print(id_number)
#抽取结果：['410403198805135531']

regex = "(?:(?:[0,1]?d?d|2[0-4]d|25[0-5]).){3}(?:[0,1]?d?d|2[0-4]d|25[0-5])"text = "ip地址信息是127.0.0.1"ip = re.findall(regex, text)print(ip)
#抽取结果：['127.0.0.1'']

regex = "1[34578]d{9}"text = "电话号码信息是18768875303"phone_number = re.findall(regex, text)print(phone_n umber)
#抽取结果：['18768875303']

 def _viterbi_decode(self, feats):        backpointers = []
        # 在对数空间初始化维特比变量        init_vvars = torch.full((1, self.tagset_size), -10000.)        init_vvars[0][self.tag_to_ix[START_TAG]] = 0
        # 第i步的 forward_var 存放第i-1步的维特比变量        forward_var = init_vvars
        for feat in feats:            bptrs_t = []  # 存放这一步的后指针            viterbivars_t = []  # 存放这一步的维特比变量
            for next_tag in range(self.tagset_size):
                # next_tag_var[i] 存放先前一步标签i的维特比变量, 加上了从标签i到next_tag的转移                # （这里暂时没有将发射score添加进来，因为最大值并不依赖它们。我们下面会加它。）                next_tag_var = forward_var + self.transitions[next_tag]                best_tag_id = argmax(next_tag_var)                bptrs_t.append(best_tag_id)                viterbivars_t.append(next_tag_var[0][best_tag_id].view(1))
            # 现在将所有发射score相加，更新forward_var            forward_var = (torch.cat(viterbivars_t) + feat).view(1, -1)            backpointers.append(bptrs_t)
        # 转移到STOP_TAG        terminal_var = forward_var + self.transitions[self.tag_to_ix[STOP_TAG]]        best_tag_id = argmax(terminal_var)        path_score = terminal_var[0][best_tag_id]
        # 使用后指针解码最优路径        best_path = [best_tag_id]        for bptrs_t in reversed(backpointers):            best_tag_id = bptrs_t[best_tag_id]            best_path.append(best_tag_id)
        # 弹出开始标签，无需返回它        start = best_path.pop()
        assert start == self.tag_to_ix[START_TAG]        best_path.reverse()        return path_score, best_path

class BiLSTM_CRF(nn.Module):
    def __init__(self, vocab_size, tag_to_ix, embedding_dim, hidden_dim):        """        :param vocab_size: 词库大小        :param tag_to_ix: tag id映射        :param embedding_dim: embedding维度        :param hidden_dim: 隐含层维度        """        super(BiLSTM_CRF, self).__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)  # 获得tag size
        # 定义embedding层，        # 设置        #   词库大小（输入序列的最大id不能超过它）        #   embedding维度（一个词被表示为向量的维度）        self.word_embeds = nn.Embedding(vocab_size, embedding_dim)
        # 定义lstm层，        # 设置        #   输入dim为embedding_dim        #   隐含层单元为 1/2 hidden_dim（双向拼接之后为hidden_dim，送入下一层）        #   一层lstm        #   设置双向        self.lstm = nn.LSTM(embedding_dim, hidden_dim // 2, num_layers=1, bidirectional=True)
        # 将bilstm的输出通过全连接层映射到标签空间        self.hidden2tag = nn.Linear(hidden_dim, self.tagset_size)
        # CRF层的核心参数，转移矩阵        # transitions[i,j]表示从j转移到i的score        self.transitions = nn.Parameter(            torch.randn(self.tagset_size, self.tagset_size)        )
        # 使不能出现 转移到START_TAG        self.transitions.data[tag_to_ix[START_TAG], :] = -10000        # 使不能出现 从STOP_TAG转移        self.transitions.data[:, tag_to_ix[STOP_TAG]] = -10000

class BertBiLSTMCRFModel(nn.Module):
    def __init__(self, bert_base_model_dir, label_size, drop_out_rate=0.5):        super(BertCRFModel, self).__init__()        self.label_size = label_size
        if 'albert' in bert_base_model_dir.lower():            # 注意albert base使用bert tokenizer，参考https://huggingface.co/voidful/albert_chinese_base            self.bert_tokenizer = BertTokenizer.from_pretrained(bert_base_model_dir)            self.bert_model = AlbertModel.from_pretrained(bert_base_model_dir)        elif 'electra' in bert_base_model_dir.lower():            self.bert_tokenizer = ElectraTokenizer.from_pretrained(bert_base_model_dir)            self.bert_model = ElectraModel.from_pretrained(bert_base_model_dir)        else:            self.bert_tokenizer = BertTokenizer.from_pretrained(bert_base_model_dir)            self.bert_model = BertModel.from_pretrained(bert_base_model_dir)
        self.lstm=nn.LSTM(self.bert_model.config.hidden_size,384,num_layers=1, bidirectional=True,batch_first=True) ##双向lstm        self.dropout = nn.Dropout(drop_out_rate)        #self.linear = nn.Linear(self.bert_model.config.hidden_size, label_size)        self.linear = nn.Linear(768, label_size)##bilstm层        self.crf = CRF(label_size)  # # 定义CRF层