TPLinker 联合抽取 实体链接方式+源码分析

关系抽取–TPLinker： https://blog.csdn.net/weixin_42223207/article/details/116425447

Tagging
TPLinker模型需要对关系三元组(subject, relation, object)进行手动Tagging，过程分为三部分：
（1）entity head to entity tail (EH-TO-ET)
（2）subject head to object head (SH-to-OH)
（3）subject tail to object tail (ST-to-OT)
标记示例见下图，EH-TO-ET用紫色表示，SH-to-OH用红色表示，ST-to-OT用蓝色表示。

论文笔记：路由器 TPLinker 也来做关系抽取：https://zhuanlan.zhihu.com/p/304104571

关系抽取之TPLinker解读加源码分析：https://zhuanlan.zhihu.com/p/342300800

关于解码过程
通过实体抽取得到字典D后，遍历关系，通过关系得到有关系的两个实体的尾部E， 再通过关系得到有关系的两个实体的头部，结合字典D得到这两个实体尾部set(s),set(o)(此为真实的标签），再通过set(s),set(o)在不在E里面，来判断是否成功抽取了一条三元组。
其中相关的公式：
文章的大概内容进行了解读，现在到了代码解读环节：
输入数据部分：
输入的seq的长度为seq_len ，获取句子的最大长度

train and valid max token num

max_tok_num = 0
all_data = train_data + valid_data 
    
for sample in all_data:
    tokens = tokenize(sample["text"])
    max_tok_num = max(max_tok_num, len(tokens))
max_tok_num # 获取句子的最大长度

接下来对文本超过长度的进行划窗处理：

if max_tok_num > hyper_parameters["max_seq_len"]: # 截断长度
    train_data = preprocessor.split_into_short_samples(train_data, 
                                                          hyper_parameters["max_seq_len"], 
                                                          sliding_len = hyper_parameters["sliding_len"], 
                                                          encoder = config["encoder"] #超过长度则滑动窗口得到新的样本
                                                         )                                           
    valid_data = preprocessor.split_into_short_samples(valid_data, 
                                                          hyper_parameters["max_seq_len"], 
                                                          sliding_len = hyper_parameters["sliding_len"], 
                                                          encoder = config["encoder"]
                                                         )

接下来看划窗的具体操作

def split_into_short_samples(self, sample_list, max_seq_len, sliding_len = 50, encoder = "BERT", data_type = "train"):
    new_sample_list = []
    for sample in tqdm(sample_list, desc = "Splitting into subtexts"):
        text_id = sample["id"]
        text = sample["text"]
        tokens = self._tokenize(text)
        tok2char_span = self._get_tok2char_span_map(text) #返回句子中单词的偏移量

        # sliding at token level
        split_sample_list = []
        for start_ind in range(0, len(tokens), sliding_len): #sliding_len 滑动窗口的大小，
            if encoder == "BERT": # if use bert, do not split a word into two samples
                while "##" in tokens[start_ind]:
                    start_ind -= 1
            end_ind = start_ind + max_seq_len # 结束的长度

            char_span_list = tok2char_span[start_ind:end_ind] #截断
            char_level_span = [char_span_list[0][0], char_span_list[-1][1]] #第一个词到最后一个词的长度
            sub_text = text[char_level_span[0]:char_level_span[1]]#原始文本截断

            new_sample = {
                "id": text_id,
                "text": sub_text,
                "tok_offset": start_ind, #token的偏移量
                "char_offset": char_level_span[0], #每个字符的偏移量
                }
            if data_type == "test": # test set
                if len(sub_text) > 0:
                    split_sample_list.append(new_sample)
            else: # train or valid dataset, only save spo and entities in the subtext
                # spo
                sub_rel_list = []
                for rel in sample["relation_list"]:
                    subj_tok_span = rel["subj_tok_span"]
                    obj_tok_span = rel["obj_tok_span"]
                    # if subject and object are both in this subtext, add this spo to new sample
                    if subj_tok_span[0] >= start_ind and subj_tok_span[1] <= end_ind \
                        and obj_tok_span[0] >= start_ind and obj_tok_span[1] <= end_ind: 
                        new_rel = copy.deepcopy(rel)
                        new_rel["subj_tok_span"] = [subj_tok_span[0] - start_ind, subj_tok_span[1] - start_ind] # start_ind: 单词级别的偏移量
                        new_rel["obj_tok_span"] = [obj_tok_span[0] - start_ind, obj_tok_span[1] - start_ind]
                        new_rel["subj_char_span"][0] -= char_level_span[0] # 字符级别的偏移量
                        new_rel["subj_char_span"][1] -= char_level_span[0]
                        new_rel["obj_char_span"][0] -= char_level_span[0]
                        new_rel["obj_char_span"][1] -= char_level_span[0]
                        sub_rel_list.append(new_rel)
                
                # entity
                sub_ent_list = []
                for ent in sample["entity_list"]:
                    tok_span = ent["tok_span"]
                    # if entity in this subtext, add the entity to new sample
                    if tok_span[0] >= start_ind and tok_span[1] <= end_ind: 
                        new_ent = copy.deepcopy(ent)
                        new_ent["tok_span"] = [tok_span[0] - start_ind, tok_span[1] - start_ind]
                        
                        new_ent["char_span"][0] -= char_level_span[0]
                        new_ent["char_span"][1] -= char_level_span[0]

                        sub_ent_list.append(new_ent)
                
                # event
                if "event_list" in sample:
                    sub_event_list = []
                    for event in sample["event_list"]:
                        trigger_tok_span = event["trigger_tok_span"]
                        if trigger_tok_span[1] > end_ind or trigger_tok_span[0] < start_ind:
                            continue
                        new_event = copy.deepcopy(event)
                        new_arg_list = []
                        for arg in new_event["argument_list"]:
                            if arg["tok_span"][0] >= start_ind and arg["tok_span"][1] <= end_ind:
                                new_arg_list.append(arg)
                        new_event["argument_list"] = new_arg_list
                        sub_event_list.append(new_event)
                    new_sample["event_list"] = sub_event_list # maybe empty
                    
                new_sample["entity_list"] = sub_ent_list # maybe empty
                new_sample["relation_list"] = sub_rel_list # maybe empty
                split_sample_list.append(new_sample)
            
            # all segments covered, no need to continue
            if end_ind > len(tokens):
                break
                
        new_sample_list.extend(split_sample_list)
    return new_sample_list

输入数据，DataMaker4Bert中定义：

class DataMaker4Bert():
    def __init__(self, tokenizer, handshaking_tagger):
        self.tokenizer = tokenizer
        self.handshaking_tagger = handshaking_tagger
    
    def get_indexed_data(self, data, max_seq_len, data_type = "train"): #index转换为data
        indexed_samples = []
        for ind, sample in tqdm(enumerate(data), desc = "Generate indexed train or valid data"):
            text = sample["text"]
            # codes for bert input
            codes = self.tokenizer.encode_plus(text, 
                                    return_offsets_mapping = True, 
                                    add_special_tokens = False,
                                    max_length = max_seq_len, 
                                    truncation = True,
                                    pad_to_max_length = True)


            # tagging
            spots_tuple = None
            if data_type != "test":
                spots_tuple = self.handshaking_tagger.get_spots(sample) #获取实体,头，尾标签

            # get codes
            input_ids = torch.tensor(codes["input_ids"]).long()
            attention_mask = torch.tensor(codes["attention_mask"]).long()
            token_type_ids = torch.tensor(codes["token_type_ids"]).long()
            tok2char_span = codes["offset_mapping"]

            sample_tp = (sample,
                     input_ids,
                     attention_mask,
                     token_type_ids,
                     tok2char_span,
                     spots_tuple,
                    )
            indexed_samples.append(sample_tp)       
        return indexed_samples

输入的是tokenizer和handshakingtagger，tokenizer为bert等一系列模型的标准输入，而get_spots函数获取了实体，头，尾的标签，具体看下代码

    def get_spots(self, sample):
        '''
        entity spot and tail_rel spot: (span_pos1, span_pos2, tag_id)
        head_rel spot: (rel_id, span_pos1, span_pos2, tag_id)
        '''
        ent_matrix_spots, head_rel_matrix_spots, tail_rel_matrix_spots = [], [], [] 

        for rel in sample["relation_list"]:
            subj_tok_span = rel["subj_tok_span"]
            obj_tok_span = rel["obj_tok_span"]
            ent_matrix_spots.append((subj_tok_span[0], subj_tok_span[1] - 1, self.tag2id_ent["ENT-H2T"])) #sub token的[起始位置，尾部位置，实体标签（1)]
            ent_matrix_spots.append((obj_tok_span[0], obj_tok_span[1] - 1, self.tag2id_ent["ENT-H2T"]))# obj token的[起始位置，尾部位置，实体标签（1)]

            if  subj_tok_span[0] <= obj_tok_span[0]:
                head_rel_matrix_spots.append((self.rel2id[rel["predicate"]], subj_tok_span[0], obj_tok_span[0], self.tag2id_head_rel["REL-SH2OH"]))#【关系类别，实体_1 头部，实体_2头部，关系标签（1）】
            else:
                head_rel_matrix_spots.append((self.rel2id[rel["predicate"]], obj_tok_span[0], subj_tok_span[0], self.tag2id_head_rel["REL-OH2SH"]))#【关系类别，实体_1 头部，实体_2头部，关系标签（2）】
                
            if subj_tok_span[1] <= obj_tok_span[1]:
                tail_rel_matrix_spots.append((self.rel2id[rel["predicate"]], subj_tok_span[1] - 1, obj_tok_span[1] - 1, self.tag2id_tail_rel["REL-ST2OT"]))#【关系类别，实体_1 尾部，实体_2尾部，关系标签（1）】
            else:
                tail_rel_matrix_spots.append((self.rel2id[rel["predicate"]], obj_tok_span[1] - 1, subj_tok_span[1] - 1, self.tag2id_tail_rel["REL-OT2ST"]))#【关系类别，实体_1 尾部，实体_2尾部，关系标签（2）】
                
        return ent_matrix_spots, head_rel_matrix_spots, tail_rel_matrix_spots
获取输入的数据

indexed_train_data = data_maker.get_indexed_data(train_data, max_seq_len) #获取输入
# index_train_data = data_maker.get_indexed_data(train_test_data,max_seq_len)
indexed_valid_data = data_maker.get_indexed_data(valid_data, max_seq_len)
tokenizer = BertTokenizerFast.from_pretrained(config["bert_path"], add_special_tokens = False, do_lower_case = False)
data_maker = DataMaker4Bert(tokenizer, handshaking_tagger) #(sample,input_ids，attention_mask,token_type_ids,tok2char_span,spots_tuple,)

接下来则是定义HandshakingTaggingScheme

max_seq_len = min(max_tok_num, hyper_parameters["max_seq_len"]) #max_len 长度
rel2id = json.load(open(rel2id_path, "r", encoding = "utf-8")) 
handshaking_tagger = HandshakingTaggingScheme(rel2id = rel2id, max_seq_len = max_seq_len) #初始化

查看具体的定义

class HandshakingTaggingScheme(object):
    """docstring for HandshakingTaggingScheme"""
    def __init__(self, rel2id, max_seq_len):
        super(HandshakingTaggingScheme, self).__init__()
        self.rel2id = rel2id
        self.id2rel = {ind:rel for rel, ind in rel2id.items()}

        self.tag2id_ent = { #实体头尾
            "O": 0,
            "ENT-H2T": 1, # entity head to entity tail
        }
        self.id2tag_ent = {id_:tag for tag, id_ in self.tag2id_ent.items()}

        self.tag2id_head_rel = { #sub,obj头对头标识1，obj头对sub头标识2
            "O": 0,
            "REL-SH2OH": 1, # subject head to object head
            "REL-OH2SH": 2, # object head to subject head
        }
        self.id2tag_head_rel = {id_:tag for tag, id_ in self.tag2id_head_rel.items()}

        self.tag2id_tail_rel = {
            "O": 0,    
            "REL-ST2OT": 1, # subject tail to object tail
            "REL-OT2ST": 2, # object tail to subject tail
        }
        self.id2tag_tail_rel = {id_:tag for tag, id_ in self.tag2id_tail_rel.items()}

        # mapping shaking sequence and matrix
        self.matrix_size = max_seq_len
        # e.g. [(0, 0), (0, 1), (0, 2), (1, 1), (1, 2), (2, 2)] #转换成矩阵上三角矩阵平铺
        self.shaking_ind2matrix_ind = [(ind, end_ind) for ind in range(self.matrix_size) for end_ind in list(range(self.matrix_size))[ind:]]

        self.matrix_ind2shaking_ind = [[0 for i in range(self.matrix_size)] for j in range(self.matrix_size)]
        for shaking_ind, matrix_ind in enumerate(self.shaking_ind2matrix_ind): #上三角矩阵，上三角每个元素储存着上三角铺平序列的相对应的位置序号
            self.matrix_ind2shaking_ind[matrix_ind[0]][matrix_ind[1]] = shaking_ind

这里比较关键的是shaking_ind2matrix_ind，与matrix_ind2shaking_ind，其中shaking_ind2matrix_ind如下所示，是一个上三角铺平序列

[(0, 0), (0, 1), (0, 2), (0, 3), (0, 4), (0, 5), (0, 6), (0, 7), (0, 8), (0, 9), (0, 10), (0, 11), (0, 12), (0, 13), …]
而matrix_ind2shaking_ind为优化前的二维矩阵，其中上三角每个元素储存着上三角铺平序列的相对应的位置序号

[[0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], …]
最后组成上三角矩阵（打印的结果没有填满）

[[0, 1, 2, 3, 4, 5, 6, 7, 8, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], [0, 0, 0, 0, 0, 0, 0, 0, 0, …], …]
载入数据

train_dataloader = DataLoader(MyDataset(indexed_train_data), 
                                  batch_size = hyper_parameters["batch_size"], 
                                  shuffle = False, 
                                  num_workers = 5,
                                  drop_last = False,
                                  collate_fn = data_maker.generate_batch,

                             )

查看DataLoader的返回值：

def generate_batch(self, batch_data, data_type = "train"):
        sample_list = []
        input_ids_list = []
        attention_mask_list = []
        token_type_ids_list = [] 
        tok2char_span_list = []
        
        ent_spots_list = []
        head_rel_spots_list = []
        tail_rel_spots_list = []

        for tp in batch_data:
            sample_list.append(tp[0])
            input_ids_list.append(tp[1])
            attention_mask_list.append(tp[2])        
            token_type_ids_list.append(tp[3])        
            tok2char_span_list.append(tp[4])
            
            if data_type != "test":
                ent_matrix_spots, head_rel_matrix_spots, tail_rel_matrix_spots = tp[5]
                ent_spots_list.append(ent_matrix_spots)
                head_rel_spots_list.append(head_rel_matrix_spots)
                tail_rel_spots_list.append(tail_rel_matrix_spots)

        # @specific: indexed by bert tokenizer
        batch_input_ids = torch.stack(input_ids_list, dim = 0)
        batch_attention_mask = torch.stack(attention_mask_list, dim = 0)
        batch_token_type_ids = torch.stack(token_type_ids_list, dim = 0)
        
        batch_ent_shaking_tag, batch_head_rel_shaking_tag, batch_tail_rel_shaking_tag = None, None, None
        if data_type != "test":
            batch_ent_shaking_tag = self.handshaking_tagger.sharing_spots2shaking_tag4batch(ent_spots_list)
            batch_head_rel_shaking_tag = self.handshaking_tagger.spots2shaking_tag4batch(head_rel_spots_list)
            batch_tail_rel_shaking_tag = self.handshaking_tagger.spots2shaking_tag4batch(tail_rel_spots_list)

        return sample_list, \
              batch_input_ids, batch_attention_mask, batch_token_type_ids, tok2char_span_list, \
                batch_ent_shaking_tag, batch_head_rel_shaking_tag, batch_tail_rel_shaking_tag

其中比较重要的是batch_ent_shaking_tag, batch_head_rel_shaking_tag, batch_tail_rel_shaking_tag这三个函数,在实体batch_ent_shaking_tag中用到了handshaking_tagger的sharing_spots2shaking_tag4batch的函数

   def sharing_spots2shaking_tag4batch(self, batch_spots):
        '''
        convert spots to batch shaking seq tag
        因长序列的stack是费时操作，所以写这个函数用作生成批量shaking tag
        如果每个样本生成一条shaking tag再stack，一个32的batch耗时1s，太昂贵
        spots: [(start_ind, end_ind, tag_id), ], for entiy
        return: 
            batch_shake_seq_tag: (batch_size, shaking_seq_len)
        '''
        shaking_seq_len = self.matrix_size * (self.matrix_size + 1) // 2
        batch_shaking_seq_tag = torch.zeros(len(batch_spots), shaking_seq_len).long()
        for batch_id, spots in enumerate(batch_spots):
            for sp in spots:
                shaking_ind = self.matrix_ind2shaking_ind[sp[0]][sp[1]] #在矩阵中找到实体的的start_int，跟end_int的位置值
                tag_id = sp[2]
                batch_shaking_seq_tag[batch_id][shaking_ind] = tag_id #铺平上三角矩阵，标注序列中的实体标识1的位置
        return batch_shaking_seq_tag

在关系batch_head_rel_shaking_tag与batch_tail_rel_shaking_tag中用到了spots2shaking_tag4batch

def spots2shaking_tag4batch(self, batch_spots):
‘’’
convert spots to batch shaking seq tag
spots: [(rel_id, start_ind, end_ind, tag_id), ], for head relation and tail_relation
return:
batch_shake_seq_tag: (batch_size, rel_size, shaking_seq_len)
‘’’
shaking_seq_len = self.matrix_size * (self.matrix_size + 1) // 2
batch_shaking_seq_tag = torch.zeros(len(batch_spots), len(self.rel2id), shaking_seq_len).long()
for batch_id, spots in enumerate(batch_spots):
for sp in spots:
shaking_ind = self.matrix_ind2shaking_ind[sp[1]][sp[2]]
tag_id = sp[3]
rel_id = sp[0]
batch_shaking_seq_tag[batch_id][rel_id][shaking_ind] = tag_id
return batch_shaking_seq_tag

跟实体类似，只是多了关系总数，然后整个函数的返回为

return sample_list,
batch_input_ids, tok2char_span_list,
batch_ent_shaking_tag, batch_head_rel_shaking_tag, batch_tail_rel_shaking_tag

初始化模型

rel_extractor = TPLinkerBert(encoder,
len(rel2id),
hyper_parameters[“shaking_type”],
hyper_parameters[“inner_enc_type”],
hyper_parameters[“dist_emb_size”],
hyper_parameters[“ent_add_dist”],
hyper_parameters[“rel_add_dist”],
)
模型的具体定义

class TPLinkerBert(nn.Module):
    def __init__(self, encoder, 
                 rel_size, 
                 shaking_type,
                 inner_enc_type,
                 dist_emb_size,
                 ent_add_dist,
                 rel_add_dist
                ):
        super().__init__()
        self.encoder = encoder
        hidden_size = encoder.config.hidden_size
        
        self.ent_fc = nn.Linear(hidden_size, 2) #实体预测，0，1
        self.head_rel_fc_list = [nn.Linear(hidden_size, 3) for _ in range(rel_size)] #rel_size多少种关系
        self.tail_rel_fc_list = [nn.Linear(hidden_size, 3) for _ in range(rel_size)]# 对每个关系进行个linear层的3分类【0，1，2】
        
        for ind, fc in enumerate(self.head_rel_fc_list):
            self.register_parameter("weight_4_head_rel{}".format(ind), fc.weight) #过3层全连接层
            self.register_parameter("bias_4_head_rel{}".format(ind), fc.bias) #偏差
        for ind, fc in enumerate(self.tail_rel_fc_list):
            self.register_parameter("weight_4_tail_rel{}".format(ind), fc.weight)
            self.register_parameter("bias_4_tail_rel{}".format(ind), fc.bias)
            
        # handshaking kernel
        self.handshaking_kernel = HandshakingKernel(hidden_size, shaking_type, inner_enc_type)
        
        # distance embedding
        self.dist_emb_size = dist_emb_size
        self.dist_embbedings = None # it will be set in the first forwarding
        
        self.ent_add_dist = ent_add_dist
        self.rel_add_dist = rel_add_dist

self.head_rel_fc_list与self.tail_rel_fc_list相同，是关系的一个全连接（标签有三个0,1,2），每一种关系有一个独立的MLP层，self.head_rel_fc_list是列表的形式。

上面为关系和实体关系预测，实体和各个关系都经过了mlp层，我们假设有5种关系，则会有11层MLP即为：1个实体预测层+（1个头部层+1个尾部层）*5

def forward(self, input_ids, attention_mask, token_type_ids):
        # input_ids, attention_mask, token_type_ids: (batch_size, seq_len)
        context_outputs = self.encoder(input_ids, attention_mask, token_type_ids) # 0 last_hidden 1 pooled
        # last_hidden_state: (batch_size, seq_len, hidden_size)
        last_hidden_state = context_outputs[0]
        
        # shaking_hiddens: (batch_size, 1 + ... + seq_len, hidden_size)
        shaking_hiddens = self.handshaking_kernel(last_hidden_state) #铺平上三角矩阵
        shaking_hiddens4ent = shaking_hiddens
        shaking_hiddens4rel = shaking_hiddens
        
        # add distance embeddings if it is set
        if self.dist_emb_size != -1:
            # set self.dist_embbedings
            hidden_size = shaking_hiddens.size()[-1]
            if self.dist_embbedings is None:
                dist_emb = torch.zeros([self.dist_emb_size, hidden_size]).to(shaking_hiddens.device)
                for d in range(self.dist_emb_size):
                    for i in range(hidden_size):
                        if i % 2 == 0:
                            dist_emb[d][i] = math.sin(d / 10000**(i / hidden_size))
                        else:
                            dist_emb[d][i] = math.cos(d / 10000**((i - 1) / hidden_size))
                seq_len = input_ids.size()[1]
                dist_embbeding_segs = []
                for after_num in range(seq_len, 0, -1): #铺平
                    dist_embbeding_segs.append(dist_emb[:after_num, :])
                self.dist_embbedings = torch.cat(dist_embbeding_segs, dim = 0)
            
            if self.ent_add_dist:
                shaking_hiddens4ent = shaking_hiddens + self.dist_embbedings[None,:,:].repeat(shaking_hiddens.size()[0], 1, 1)
            if self.rel_add_dist:
                shaking_hiddens4rel = shaking_hiddens + self.dist_embbedings[None,:,:].repeat(shaking_hiddens.size()[0], 1, 1)
                
#         if self.dist_emb_size != -1 and self.ent_add_dist:
#             shaking_hiddens4ent = shaking_hiddens + self.dist_embbedings[None,:,:].repeat(shaking_hiddens.size()[0], 1, 1)
#         else:
#             shaking_hiddens4ent = shaking_hiddens
#         if self.dist_emb_size != -1 and self.rel_add_dist:
#             shaking_hiddens4rel = shaking_hiddens + self.dist_embbedings[None,:,:].repeat(shaking_hiddens.size()[0], 1, 1)
#         else:
#             shaking_hiddens4rel = shaking_hiddens
            
        ent_shaking_outputs = self.ent_fc(shaking_hiddens4ent) #实体预测，（0，1）
            
        head_rel_shaking_outputs_list = []
        nn.ModuleList()
        for fc in self.head_rel_fc_list:
            head_rel_shaking_outputs_list.append(fc(shaking_hiddens4rel)) #对每一种关系头进行分类
            
        tail_rel_shaking_outputs_list = []
        for fc in self.tail_rel_fc_list: #对每一种关系尾进行分类
            tail_rel_shaking_outputs_list.append(fc(shaking_hiddens4rel))
        
        head_rel_shaking_outputs = torch.stack(head_rel_shaking_outputs_list, dim = 1) #n种关系拼接在一起
        tail_rel_shaking_outputs = torch.stack(tail_rel_shaking_outputs_list, dim = 1) #n种关系拼接在一起
        
        return ent_shaking_outputs, head_rel_shaking_outputs, tail_rel_shaking_outputs

ent_shaking_outputs为实体预测，head_rel_shaking_outputs_list对关系头进行分类，tail_rel_shaking_outputs_list对关系尾进行分类，其中关键函数shaking_hiddens4ent中的HandshakingKernel函数定义如下

class HandshakingKernel(nn.Module):
    def __init__(self, hidden_size, shaking_type, inner_enc_type):
        super().__init__()
        self.shaking_type = shaking_type
        if shaking_type == "cat":
            self.combine_fc = nn.Linear(hidden_size * 2, hidden_size) #fc层
        elif shaking_type == "cat_plus":
            self.combine_fc = nn.Linear(hidden_size * 3, hidden_size)
        elif shaking_type == "cln":
            self.tp_cln = LayerNorm(hidden_size, hidden_size, conditional = True)
        elif shaking_type == "cln_plus":
            self.tp_cln = LayerNorm(hidden_size, hidden_size, conditional = True)
            self.inner_context_cln = LayerNorm(hidden_size, hidden_size, conditional = True)
            
        self.inner_enc_type = inner_enc_type #一层单向lstm
        if inner_enc_type == "mix_pooling":
            self.lamtha = Parameter(torch.rand(hidden_size))
        elif inner_enc_type == "lstm":
            self.inner_context_lstm = nn.LSTM(hidden_size, 
                           hidden_size, 
                           num_layers = 1, 
                           bidirectional = False, 
                           batch_first = True)
     
    def enc_inner_hiddens(self, seq_hiddens, inner_enc_type = "lstm"):
        # seq_hiddens: (batch_size, seq_len, hidden_size)
        def pool(seqence, pooling_type):
            if pooling_type == "mean_pooling":
                pooling = torch.mean(seqence, dim = -2)
            elif pooling_type == "max_pooling":
                pooling, _ = torch.max(seqence, dim = -2)
            elif pooling_type == "mix_pooling":
                pooling = self.lamtha * torch.mean(seqence, dim = -2) + (1 - self.lamtha) * torch.max(seqence, dim = -2)[0]
            return pooling
        if "pooling" in inner_enc_type:
            inner_context = torch.stack([pool(seq_hiddens[:, :i+1, :], inner_enc_type) for i in range(seq_hiddens.size()[1])], dim = 1)
        elif inner_enc_type == "lstm":
            inner_context, _ = self.inner_context_lstm(seq_hiddens)
            
        return inner_context
    def forward(self, seq_hiddens):
        '''
        seq_hiddens: (batch_size, seq_len, hidden_size)
        return:
        shaking_hiddenss: (batch_size, (1 + seq_len) * seq_len / 2, hidden_size) (32, 5+4+3+2+1, 5)
        '''#一句话中每个字与剩下的字构成上三角矩阵如：长度为5则的到的为[[batch,5,hidden_size],[batch,4,hidden_size]...]
        seq_len = seq_hiddens.size()[-2] #句子的长度
        shaking_hiddens_list = []
        for ind in range(seq_len):
            hidden_each_step = seq_hiddens[:, ind, :] #取每个batch的每个字的维度
            visible_hiddens = seq_hiddens[:, ind:, :] # 从当前取到最后
            repeat_hiddens = hidden_each_step[:, None, :].repeat(1, seq_len - ind, 1) #复制dim=1的维度跟visible维度保持一致
            if self.shaking_type == "cat":#选择的是cat模式，可以在配置文件中设置
                shaking_hiddens = torch.cat([repeat_hiddens, visible_hiddens], dim = -1) #将当前每个字的维度与其后的每个字的维度拼接在一起
                shaking_hiddens = torch.tanh(self.combine_fc(shaking_hiddens))#过一个线性层
            elif self.shaking_type == "cat_plus":
                inner_context = self.enc_inner_hiddens(visible_hiddens, self.inner_enc_type)
                shaking_hiddens = torch.cat([repeat_hiddens, visible_hiddens, inner_context], dim = -1)
                shaking_hiddens = torch.tanh(self.combine_fc(shaking_hiddens))
            elif self.shaking_type == "cln":
                shaking_hiddens = self.tp_cln(visible_hiddens, repeat_hiddens)
            elif self.shaking_type == "cln_plus":
                inner_context = self.enc_inner_hiddens(visible_hiddens, self.inner_enc_type)
                shaking_hiddens = self.tp_cln(visible_hiddens, repeat_hiddens)
                shaking_hiddens = self.inner_context_cln(shaking_hiddens, inner_context)
            shaking_hiddens_list.append(shaking_hiddens) #添加到列表中
        long_shaking_hiddens = torch.cat(shaking_hiddens_list, dim = 1)#铺平上三角矩阵
        return long_shaking_hiddens

输入的seq_hiddens维度是[batch,seq_len, hiddensize]，是一句话经过bert编码过后的值，而HandshakingKernel函数的作用是将矩阵变为上三角矩阵，即本身矩阵为[seq_len * seq_len]，在经过函数过后为每一行都减1，最后通过long_shakinghiddens把函数把结果铺平，得到[seq_len+(seq_len -1) + (seq_len -2)…+1],对应了图片部分。

整个函数先是循环每句话中的词，当ind是0时，hidden_each_step代表了循环的每个词的编码[batch,1,hidden_size]，visiblehiddens是循环到的这个单词以及之后的单词的编码，维度就是[batch,seq_len,hidden_size],repeat_hiddens对hidden_each_step的第二个维度进行了复制，维度为[batch,seq_len,hidden_size]，将当前单词和其后的各个单词的编码进行拼接维度是[batch,seq_len,hidden_size*2]组成上三角矩阵的一行,在经过MLP层后shakinghiddens的维度是[batch,seq_len,hidden_size]，之后每一行依次类推。

关于loss部分

total_loss, total_ent_sample_acc, total_head_rel_sample_acc, total_tail_rel_sample_acc = 0., 0., 0., 0.
        for batch_ind, batch_train_data in enumerate(dataloader):
            t_batch = time.time()
            z = (2 * len(rel2id) + 1) # 2倍的关系
            steps_per_ep = len(dataloader) #有多少数据
            total_steps = hyper_parameters["loss_weight_recover_steps"] + 1 # + 1 avoid division by zero error #加速loss在一定的步数回归
            current_step = steps_per_ep * ep + batch_ind # ？
            w_ent = max(1 / z + 1 - current_step / total_steps, 1 / z)
            w_rel = min((len(rel2id) / z) * current_step / total_steps, (len(rel2id) / z))
            loss_weights = {"ent": w_ent, "rel": w_rel} #给予不同任务的权重
            
            loss, ent_sample_acc, head_rel_sample_acc, tail_rel_sample_acc = train_step(batch_train_data, optimizer, loss_weights)
            scheduler.step()
            
            total_loss += loss
            total_ent_sample_acc += ent_sample_acc
            total_head_rel_sample_acc += head_rel_sample_acc
            total_tail_rel_sample_acc += tail_rel_sample_acc
            
            avg_loss = total_loss / (batch_ind + 1)
            avg_ent_sample_acc = total_ent_sample_acc / (batch_ind + 1)
            avg_head_rel_sample_acc = total_head_rel_sample_acc / (batch_ind + 1)
            avg_tail_rel_sample_acc = total_tail_rel_sample_acc / (batch_ind + 1)

随着step加大，w_ent的权重递减，w_rel权重递增。先关注实体，保证实体抽准确，后面关注关系的抽取，由于目前工作原因，更多细节待闲时在进行解读。

百度信息抽取Lic2020关系抽取：https://zhuanlan.zhihu.com/p/138858558