如何使用 Neo4J 和 Transformer 构建知识图谱（Project neo）

本文最初发布于 Towards Data Science。

图片由作者提供：Neo4j中的知识图谱

简介

在这篇文章中，我将展示如何使用经过优化的、基于转换器的命名实体识别（NER）以及 spaCy 的关系提取模型，基于职位描述创建一个知识图谱。这里介绍的方法可以应用于其他任何领域，如生物医学、金融、医疗保健等。

以下是我们要采取的步骤：

在 Google Colab 中加载优化后的转换器 NER 和 spaCy 关系提取模型；
创建一个 Neo4j Sandbox，并添加实体和关系；
查询图，找出与目标简历匹配度最高的职位，找出三个最受欢迎的技能和共现率最高的技能。

要了解关于如何使用 UBIAI 生成训练数据以及优化 NER 和关系提取模型的更多信息，请查看以下文章。

UBIAI：简单易用的NLP应用程序文本标注
如何使用BERT转换器与spaCy3训练一个联合实体和关系提取分类器
如何使用spaCy3优化BERT转换器

职位描述数据集可以从Kaggle获取。

在本文结束的时候，我们就可以创建出如下所示的知识图谱。

图片由作者提供：职位描述的知识图谱

命名实体和关系提取

首先，我们加载 NER 和关系模型的依赖关系，以及之前优化过的 NER 模型本身，以提取技能、学历、专业和工作年限：

!pip install -U pip setuptools wheel !python -m spaCy project clone tutorials/rel_component !pip install -U spaCy-nightly --pre !!pip install -U spaCy transformers import spaCy #安装完依赖项后重启运行时 nlp = spaCy.load("[PATH_TO_THE_MODEL]/model-best")

加载我们想从中提取实体和关系的职位数据集：

import pandas as pd def get_all_documents(): df = pd.read_csv("/content/drive/MyDrive/job_DB1_1_29.csv",sep='"',hea der=None) documents = [] for index,row in df.iterrows(): documents.append(str(row[0])) return documents documents = get_all_documents() documents = documents[:]

从职位数据集中提取实体：

import hashlib def extract_ents(documents,nlp): docs = list() for doc in nlp.pipe(documents, disable=["tagger", "parser"]): dictionary=dict.fromkeys(["text", "annotations"]) dictionary["text"]= str(doc) dictionary['text_sha256'] = hashlib.sha256(dictionary["text"].encode('utf-8')).hexdigest() annotations=[] for e in doc.ents: ent_id = hashlib.sha256(str(e.text).encode('utf-8')).hexdigest() ent = {"start":e.start_char,"end":e.end_char, "label":e.label_,"label_upper":e.label_.upper(),"text":e.text,"id":ent_id} if e.label_ == "EXPERIENCE": ent["years"] = int(e.text[0]) annotations.append(ent) dictionary["annotations"] = annotations docs.append(dictionary) #print(annotations) return docs parsed_ents = extract_ents(documents,nlp)

在将实体提供给关系提取模型之前，我们可以看下提取出的部分实体：

[('stock market analysis', 'SKILLS'), ('private investor', 'SKILLS'), ('C ', 'SKILLS'), ('Investment Software', 'SKILLS'), ('MS Windows', 'SKILLS'), ('web development', 'SKILLS'), ('Computer Science', 'DIPLOMA_MAJOR'), ('AI', 'SKILLS'), ('software development', 'SKILLS'), ('coding', 'SKILLS'), ('C', 'SKILLS'), ('C ', 'SKILLS'), ('Visual Studio', 'SKILLS'), ('2 years', 'EXPERIENCE'), ('C/C development', 'SKILLS'), ('data compression', 'SKILLS'), ('financial markets', 'SKILLS'), ('financial calculation', 'SKILLS'), ('GUI design', 'SKILLS'), ('Windows development', 'SKILLS'), ('MFC', 'SKILLS'), ('Win', 'SKILLS'), ('HTTP', 'SKILLS'), ('TCP/IP', 'SKILLS'), ('sockets', 'SKILLS'), ('network programming', 'SKILLS'), ('System administration', 'SKILLS')]

我们现在准备好预测关系了；首先加载关系提取模型，务必将目录改为 rel_component/scripts 以便可以访问关系模型的所有必要脚本。

cd rel_component/

import random import typer from pathlib import Path import spaCy from spaCy.tokens import DocBin, Doc from spaCy.training.example import Example #使factory生效 from rel_pipe import make_relation_extractor, score_relations #使config生效 from rel_model import create_relation_model, create_classification_layer, create_instances, create_tensors #安装完依赖项后重启运行时 nlp2 = spaCy.load("/content/drive/MyDrive/training_rel_roberta/model-best") def extract_relations(documents,nlp,nlp2): predicted_rels = list() for doc in nlp.pipe(documents, disable=["tagger", "parser"]): source_hash = hashlib.sha256(doc.text.encode('utf-8')).hexdigest() for name, proc in nlp2.pipeline: doc = proc(doc) for value, rel_dict in doc._.rel.items(): for e in doc.ents: for b in doc.ents: if e.start == value[0] and b.start == value[1]: max_key = max(rel_dict, key=rel_dict. get) #print(max_key) e_id = hashlib.sha256(str(e).encode('utf-8')).hexdigest() b_id = hashlib.sha256(str(b).encode('utf-8')).hexdigest() if rel_dict[max_key] >=0.9 : #print(f" entities: {e.text, b.text} --> predicted relation: {rel_dict}") predicted_rels.append({'head': e_id, 'tail': b_id, 'type':max_key, 'source': source_hash}) return predicted_rels predicted_rels = extract_relations(documents,nlp,nlp2)

Predicted relations: entities: ('5 years', 'software engineering') --> predicted relation: {'DEGREE_IN': 9.5471655e-08, 'EXPERIENCE_IN': 0.9967771} entities: ('5 years', 'technical management') --> predicted relation: {'DEGREE_IN': 1.1285037e-07, 'EXPERIENCE_IN': 0.9961034} entities: ('5 years', 'designing') --> predicted relation:{'DEGREE_IN': 1.3603304e-08, 'EXPERIENCE_IN': 0.9989103} entities: ('4 years', 'performance management') --> predicted relation: {'DEGREE_IN': 6.748373e-08, 'EXPERIENCE_IN': 0.92884386}Neo4j

现在，我们可以加载职位数据集，并将数据提取到 Neo4j 数据库中了。

首先，我们创建一个空的neo4j Sandbox，并添加连接信息，如下所示：

documents = get_all_documents() documents = documents[:] parsed_ents = extract_ents(documents,nlp) predicted_rels = extract_relations(documents,nlp,nlp2) #neo4j的基础查询功能 from neo4j import GraphDatabase import pandas as pd host = 'bolt://[your_host_address]' user = 'neo4j' password = '[your_password]' driver = GraphDatabase.driver(host,auth=(user, password)) def neo4j_query(query, params=None): with driver.session() as session: result = session.run(query, params) return pd.DataFrame([r.values() for r in result], columns=result.keys())

接下来，我们将文档、实体和关系添加到知识图谱中。注意，我们需要从实体 EXPERIENCE 的 name 中提取出整数年限，并将其作为一个属性存储起来。

#清空当前的Neo4j Sandbox db (删除所有东西) neo4j_query(""" MATCH (n) DETACH DELETE n; """) #创建第一个主节点 neo4j_query(""" MERGE (l:LaborMarket {name:"Labor Market"}) RETURN l """) #向KG中添加实体：技能、经验、学历、专业 neo4j_query(""" MATCH (l:LaborMarket) UNWIND $data as row MERGE (o:Offer{id:row.text_sha256}) SET o.text = row.text MERGE (l)-[:HAS_OFFER]->(o) WITH o, row.annotations as entities UNWIND entities as entity MERGE (e:Entity {id:entity.id}) ON CREATE SET e.name = entity.text, e.label = entity.label_upper MERGE (o)-[m:MENTIONS]->(e) ON CREATE SET m.count = 1 ON MATCH SET m.count = m.count 1 WITH e as e CALL apoc.create.addLabels( id(e), [ e.label ] ) YIELD node REMOVE node.label RETURN node """, {'data': parsed_ents}) #为实体EXPERIENCE添加属性'name' res = neo4j_query(""" MATCH (e:EXPERIENCE) RETURN e.id as id, e.name as name """) #从EXPERIENCE name中提取工作年限，并保存在属性years中 import re def get_years(name): return re.findall(r"\d ",name)[0] res["years"] = res.name.map(lambda name: get_years(name)) data = res.to_dict('records') #为实体EXPERIENCE添加属性'years' neo4j_query(""" UNWIND $data as row MATCH (e:EXPERIENCE {id:row.id}) SET e.years = row.years RETURN e.name as name, e.years as years """,{"data":data}) #将关系添加到KG neo4j_query(""" UNWIND $data as row MATCH (source:Entity {id: row.head}) MATCH (target:Entity {id: row.tail}) MATCH (offer:Offer {id: row.source}) MERGE (source)-[:REL]->(r:Relation {type: row.type})-[:REL]->(target) MERGE (offer)-[:MENTIONS]->(r) """, {'data': predicted_rels})

现在开始进入有趣的部分了。我们可以启动知识图谱并运行查询了。让我们运行一个查询，找出与目标简历最匹配的职位：

#在表中显示最佳匹配项 other_id = "8de6e42ddfbc2a8bd7008d93516c57e50fa815e64e387eb2fc7a27000ae904b6" query = """ MATCH (o1:Offer {id:$id})-[m1:MENTIONS]->(s:Entity)<- [m2:MENTIONS]-(o2:Offer) RETURN DISTINCT o1.id as Source,o2.id as Proposed_Offer, count(*) as freq, collect(s.name) as common_terms ORDER BY freq DESC LIMIT $limit """ res = neo4j_query(query,{"id":other_id,"limit":3}) res #在neo4j浏览器中，使用该查询显示最佳匹配项的图 """MATCH (o1:Offer {id:"8de6e42ddfbc2a8bd7008d93516c57e50fa815e64e387eb2fc7a27000ae904b6"})-[m1:MENTIONS]->(s:Entity)<- [m2:MENTIONS]-(o2:Offer) WITH o1,s,o2, count(*) as freq MATCH (o1)--(s) RETURN collect(o2)[0], o1,s, max(freq)"""

以表格形式显示的结果中的公共实体：

以可视化形式显示的图：

图片由作者提供：基于最佳匹配职位

虽然这个数据集只有 29 个职位描述，但这里介绍的方法可以应用于有成千上万个职位的大规模数据集。只需几行代码，我们立马就可以提取出与目标简历匹配度最高的工作。

下面，让我们找出最需要的技能：

query = """ MATCH (s:SKILLS)<-[:MENTIONS]-(o:Offer) RETURN s.name as skill, count(o) as freq ORDER BY freq DESC LIMIT 10 """ res = neo4j_query(query) res

以及需要最高工作年限的技能：

query = """ MATCH (s:SKILLS)--(r:Relation)--(e:EXPERIENCE) where r.type = "EXPERIENCE_IN" return s.name as skill,e.years as years ORDER BY years DESC LIMIT 10 """ res = neo4j_query(query) res

Web 开发和技术支持需要的工作年限最高，然后是安全设置。

最后，让我们查下共现率最高的技能对：

neo4j_query(""" MATCH (s1:SKILLS)<-[:MENTIONS]-(:Offer)-[:MENTIONS]->(s2:SKILLS) WHERE id(s1) < id(s2) RETURN s1.name as skill1, s2.name as skill2, count(*) as cooccurrence ORDER BY cooccurrence DESC LIMIT 5 """)