【实体关系抽取】之——CasRel关系抽取（Duie数据集）

lxjEL

已于 2023-09-06 21:23:39 修改

阅读量2.5k

点赞数 4

文章标签：服务器 python bert

于 2023-04-06 21:44:59 首次发布

本文链接：https://blog.csdn.net/m0_62339601/article/details/129998027

版权

文章描述了一个数据集的预处理过程，将Duie数据集中的样本转换为与NYT数据集类似的triple形式，同时根据关系模式将triple分为SEO（单实体多关系）、EPO（多关系同一实体对）和正常类型。处理后的数据集可用于后续的分析或模型训练。

摘要由CSDN通过智能技术生成

数据集

Duie数据集

例：

{"text": "《步步惊心》改编自著名作家桐华的同名清穿小说《甄嬛传》改编自流潋紫所著的同名小说电视剧《何以笙箫默》改编自顾漫同名小说《花千骨》改编自fresh果果同名小说《裸婚时代》是月影兰析创作的一部情感小说《琅琊榜》是根据海宴同名网络小说改编电视剧《宫锁心玉》，又名《宫》《雪豹》，该剧改编自网络小说《特战先驱》《我是特种兵》由红遍网络的小说《最后一颗子弹留给我》改编电视剧《来不及说我爱你》改编自匪我思存同名小说《来不及说我爱你》", "spo_list": [{"predicate": "作者", "object_type": {"@value": "人物"}, "subject_type": "图书作品", "object": {"@value": "顾漫"}, "subject": "何以笙箫默"}, {"predicate": "改编自", "object_type": {"@value": "作品"}, "subject_type": "影视作品", "object": {"@value": "最后一颗子弹留给我"}, "subject": "我是特种兵"}, {"predicate": "作者", "object_type": {"@value": "人物"}, "subject_type": "图书作品", "object": {"@value": "桐华"}, "subject": "步步惊心"}, {"predicate": "作者", "object_type": {"@value": "人物"}, "subject_type": "图书作品", "object": {"@value": "流潋紫"}, "subject": "甄嬛传"}, {"predicate": "作者", "object_type": {"@value": "人物"}, "subject_type": "图书作品", "object": {"@value": "fresh果果"}, "subject": "花千骨"}, {"predicate": "作者", "object_type": {"@value": "人物"}, "subject_type": "图书作品", "object": {"@value": "月影兰析"}, "subject": "裸婚时代"}, {"predicate": "作者", "object_type": {"@value": "人物"}, "subject_type": "图书作品", "object": {"@value": "海宴"}, "subject": "琅琊榜"}, {"predicate": "改编自", "object_type": {"@value": "作品"}, "subject_type": "影视作品", "object": {"@value": "特战先驱"}, "subject": "雪豹"}, {"predicate": "改编自", "object_type": {"@value": "作品"}, "subject_type": "影视作品", "object": {"@value": "来不及说我爱你"}, "subject": "来不及说我爱你"}, {"predicate": "作者", "object_type": {"@value": "人物"}, "subject_type": "图书作品", "object": {"@value": "匪我思存"}, "subject": "来不及说我爱你"}]}

训练样本格式：

text和spo_list（predict,object_type,subject_type,object,subject）两部分组成，每行一个样本。

一、数据预处理

目标：将数据处理为样例代码中NYT处理后的triple的样子（训练用的数据为triple类型）

处理后的NYT：
原文：

{"sentText": "The only greater tragedy than the death of Ousmane Zongo would be to convict an innocent man , who at the time was acting reasonably pursuant to his training as a New York City police officer , '' Mr. London said . ''", "relationMentions": [{"em1Text": "Zongo", "em2Text": "City", "label": "/people/person/place_lived"}]}

triple：

		{
        "text": "The only greater tragedy than the death of Ousmane Zongo would be to convict an innocent man , who at the time was acting reasonably pursuant to his training as a New York City police officer , '' Mr. London said . ''",
        "triple_list": [
            [
                "Zongo",
                "/people/person/place_lived",
                "City"
            ]
        ]
    }

包含两部分：text和triple_list(subject,relation,objective)
（1）将数据集提取成triple的形式

import json
from tqdm import tqdm
import codecs


train_data = []
dev_data = []
test_data = []

with open('train_dev.json') as f:
    for l in tqdm(f):
        a = json.loads(l)
        if not a['spo_list']:
            continue
        line = {
            'text': a['text'].lstrip('\"').strip('\r\n').rstrip('\"'),
            'triple_list': [(i['subject'], i['predicate'], i['object']['@value']) for i in a['spo_list'] if i['predicate'] != 'None']
            }
        if not line['triple_list']:
            continue
        train_data.append(line)

with open('dev_dev.json') as f:
    for l in tqdm(f):
        a = json.loads(l)
        if not a['spo_list']:
            continue
        line = {
            'text': a['text'].lstrip('\"').strip('\r\n').rstrip('\"'),
            'triple_list': [(i['subject'], i['predicate'], i['object']['@value']) for i in a['spo_list'] if i['predicate'] != 'None']
            }
        if not line['triple_list']:
            continue
        dev_data.append(line)

with open('test_dev.json') as f:
    for l in tqdm(f):
        a = json.loads(l)
        if not a['spo_list']:
            continue
        line = {
            'text': a['text'].lstrip('\"').strip('\r\n').rstrip('\"'),
            'triple_list': [(i['subject'], i['predicate'], i['object']['@value']) for i in a['spo_list'] if i['predicate'] != 'None']
            }
        if not line['triple_list']:
            continue
        test_data.append(line)

with codecs.open('train_triples.json', 'w', encoding='utf-8') as f:
    json.dump(train_data, f, indent=4, ensure_ascii=False)
    
with codecs.open('dev_triples.json', 'w', encoding='utf-8') as f:
    json.dump(dev_data, f, indent=4, ensure_ascii=False)


with codecs.open('test_triples.json', 'w', encoding='utf-8') as f:
    json.dump(test_data, f, indent=4, ensure_ascii=False)

（2）将triple划分为SEO,EPO,Normal三种类型
SEO为一个实体与多个实体之间都存在关系的情况
例如：

{
        "text": "《步步惊心》改编自著名作家桐华的同名清穿小说《甄嬛传》改编自流潋紫所著的同名小说电视剧《何以笙箫默》改编自顾漫同名小说《花千骨》改编自fresh果果同名小说《裸婚时代》是月影兰析创作的一部情感小说《琅琊榜》是根据海宴同名网络小说改编电视剧《宫锁心玉》，又名《宫》《雪豹》，该剧改编自网络小说《特战先驱》《我是特种兵》由红遍网络的小说《最后一颗子弹留给我》改编电视剧《来不及说我爱你》改编自匪我思存同名小说《来不及说我爱你》",
        "triple_list": [
            [
                "何以笙箫默",
                "作者",
                "顾漫"
            ],
            [
                "我是特种兵",
                "改编自",
                "最后一颗子弹留给我"
            ],
            [
                "步步惊心",
                "作者",
                "桐华"
            ],
            [
                "甄嬛传",
                "作者",
                "流潋紫"
            ],
            [
                "花千骨",
                "作者",
                "fresh果果"
            ],
            [
                "裸婚时代",
                "作者",
                "月影兰析"
            ],
            [
                "琅琊榜",
                "作者",
                "海宴"
            ],
            [
                "雪豹",
                "改编自",
                "特战先驱"
            ],
            [
                "来不及说我爱你",
                "改编自",
                "来不及说我爱你"
            ],
            [
                "来不及说我爱你",
                "作者",
                "匪我思存"
            ]
        ]
    }

EPO为两个实体之间存在有不同的关系的情况
例如：

{
        "text": "2019年2月25日和26日，温氏股份实控人之一、前任董事长温鹏程之妻伍翠珍分别减持公司股票608万股和256万股，成交均价分别为30.78元/股和30.02元/股，共计套现约2.64亿元",
        "triple_list": [
            [
                "温氏股份",
                "董事长",
                "温鹏程"
            ],
            [
                "伍翠珍",
                "丈夫",
                "温鹏程"
            ],
            [
                "温鹏程",
                "妻子",
                "伍翠珍"
            ]
        ]
    }

代码如下：

import json
from tqdm import tqdm
import codecs
import numpy as np

test_normal = []
test_seo = []
test_epo = []

def is_seo_triple(triples): # 头实体相同，尾实体不同
    s_to_o = {}
    for triple in triples:

        if triple[0] in s_to_o.keys():
            cur = set(s_to_o[triple[0]])
            if triple[2] not in cur:
                return 1    

        elif triple[2] in s_to_o.keys():
            cur = set(s_to_o[triple[2]])
            if triple[0] not in cur:
                return 1

        else: 
            s_to_o[triple[0]] = triple[2]
            s_to_o[triple[2]] = triple[0]
        # print(s_to_o.items())
    return 0


def is_epo_triple(triples): # 头尾实体均相同，关系不同
    subject_object = {}
    for triple in triples:
        if triple[0] in subject_object.keys():
            cur = set(subject_object[triple[0]])
            if triple[2] in cur:
                return 1
            else:
                cur.add(triple[2])
                subject_object[triple[0]] = cur
        elif triple[2] in subject_object.keys():
            cur = set(subject_object[triple[2]])
            if triple[0] in cur:
                return 1
            else:
                cur.add(triple[0])
                subject_object[triple[2]] = cur
        else:
            subject_object[triple[0]] = triple[2]
            subject_object[triple[2]] = triple[0]
        # print(subject_object.items())
        
    return 0

def is_normal_triple(tripels):
    if is_epo_triple(tripels) or is_seo_triple(tripels):
        return 0
    return 1

if __name__ == '__main__':

    with open('duie_test.json') as f:
        for l in tqdm(f):
            a = json.loads(l)
            if not a['spo_list']:
                continue
            line = {
                'text': a['text'].lstrip('\"').strip('\r\n').rstrip('\"'),
                'triple_list': [(i['subject'], i['predicate'], i['object']['@value']) for i in a['spo_list'] if i['predicate'] != 'None']
                }
            # 划分关系类型
            if is_epo_triple(line['triple_list']):
                # epo += 1
                test_epo.append(line)
            elif is_seo_triple(line['triple_list']):
                # seo += 1
                test_seo.append(line)
            else:
                # normal += 1
                test_normal.append(line)
    #测试集按照type进行划分
    with codecs.open('test_normal_triples.json', 'w', encoding='utf-8') as f:
        json.dump(test_normal, f, indent=4, ensure_ascii=False)

    with codecs.open('test_seo_triples.json', 'w', encoding='utf-8') as f:
        json.dump(test_seo, f, indent=4, ensure_ascii=False)

    with codecs.open('test_epo_triples.json', 'w', encoding='utf-8') as f:
        json.dump(test_epo, f, indent=4, ensure_ascii=False)

这样数据集就处理好啦
(ps.附上split_by_num

#! -*- coding:utf-8 -*-

import json
from tqdm import tqdm
import codecs

test_1 = []
test_2 = []
test_3 = []
test_4 = []
test_other = []

with open('duie_test.json') as f:
    for l in tqdm(f):
        a = json.loads(l)
        if not a['spo_list']:
            continue
        line = {
                'text': a['text'].lstrip('\"').strip('\r\n').rstrip('\"'),
                'triple_list': [(i['subject'], i['predicate'], i['object']['@value']) for i in a['spo_list'] if i['predicate'] != 'None']
               }
        if not line['triple_list']:
            continue
        spo_num = len(line['triple_list'])
        if spo_num == 1:
            test_1.append(line)
        elif spo_num == 2:
            test_2.append(line)
        elif spo_num == 3:
            test_3.append(line)
        elif spo_num == 4:
            test_4.append(line)
        else:
            test_other.append(line)


with codecs.open('test_triples_1.json', 'w', encoding='utf-8') as f:
    json.dump(test_1, f, indent=4, ensure_ascii=False)

with codecs.open('test_triples_2.json', 'w', encoding='utf-8') as f:
    json.dump(test_2, f, indent=4, ensure_ascii=False)

with codecs.open('test_triples_3.json', 'w', encoding='utf-8') as f:
    json.dump(test_3, f, indent=4, ensure_ascii=False)

with codecs.open('test_triples_4.json', 'w', encoding='utf-8') as f:
    json.dump(test_4, f, indent=4, ensure_ascii=False)

with codecs.open('test_triples_5.json', 'w', encoding='utf-8') as f:
    json.dump(test_other, f, indent=4, ensure_ascii=False)