【知识图谱】实践篇——基于医疗知识图谱的问答系统实践（Part5-完结）：信息检索与结果组装

前序文章：

【知识图谱】实践篇——基于医疗知识图谱的问答系统实践（Part1）：项目介绍与环境准备
【知识图谱】实践篇——基于医疗知识图谱的问答系统实践（Part2）：图谱数据准备与导入
【知识图谱】实践篇——基于医疗知识图谱的问答系统实践（Part3）：基于规则的问题分类
【知识图谱】实践篇——基于医疗知识图谱的问答系统实践（Part4）：结合问题分类的问题解析与检索语句生成

背景

在前面几个模块中我们已经完成了问题分类、问题解析以及问题所需要的信息检索语句的生成。下面就把这些模块串起来，然后将结果进行组装。

结果组装

结果组装就是根据不同类型的问题输出对应的结果，具体实现如下：

KGQAMedicine\answer_search\raw_answer_search.py

from utils.config import SysConfig
from py2neo import Graphclass RawAnswerSearcher(object):def __init__(self):self.graph = Graph(SysConfig.NEO4J_HOST + ":" + str(SysConfig.NEO4J_PORT), auth=(SysConfig.NEO4J_USER,SysConfig.NEO4J_PASSWORD))self.num_limit = 20def search(self, sql_list: list):final_answers = []for sql in sql_list:question_kind = sql['question_kind']answers = []for query in sql['sql']:query_result = self.graph.run(query).data()answers += query_resultfinal_answer = self._answer_standard(question_kind, answers)if final_answer:final_answers.append(final_answer)return final_answersdef _answer_standard(self, question_kind, answers):final_answer = []if not answers:return ''if question_kind == 'disease_symptom':desc = [i['n.name'] for i in answers]subject = answers[0]['m.name']final_answer = '{0}的症状包括：{1}'.format(subject, '；'.join(list(set(desc))[:self.num_limit]))elif question_kind == 'symptom_disease':desc = [i['m.name'] for i in answers]subject = answers[0]['n.name']final_answer = '症状{0}可能染上的疾病有：{1}'.format(subject, '；'.join(list(set(desc))[:self.num_limit]))elif question_kind == 'disease_cause':desc = [i['m.cause'] for i in answers]subject = answers[0]['m.name']final_answer = '{0}可能的成因有：{1}'.format(subject, '；'.join(list(set(desc))[:self.num_limit]))elif question_kind == 'disease_prevent':desc = [i['m.prevent'] for i in answers]subject = answers[0]['m.name']final_answer = '{0}的预防措施包括：{1}'.format(subject, '；'.join(list(set(desc))[:self.num_limit]))elif question_kind == 'disease_lasttime':desc = [i['m.cure_lasttime'] for i in answers]subject = answers[0]['m.name']final_answer = '{0}治疗可能持续的周期为：{1}'.format(subject, '；'.join(list(set(desc))[:self.num_limit]))elif question_kind == 'disease_cureway':desc = [';'.join(i['m.cure_way']) for i in answers]subject = answers[0]['m.name']final_answer = '{0}可以尝试如下治疗：{1}'.format(subject, '；'.join(list(set(desc))[:self.num_limit]))elif question_kind == 'disease_cureprob':desc = [i['m.cured_prob'] for i in answers]subject = answers[0]['m.name']final_answer = '{0}治愈的概率为（仅供参考）：{1}'.format(subject, '；'.join(list(set(desc))[:self.num_limit]))elif question_kind == 'disease_easyget':desc = [i['m.easy_get'] for i in answers]subject = answers[0]['m.name']final_answer = '{0}的易感人群包括：{1}'.format(subject, '；'.join(list(set(desc))[:self.num_limit]))elif question_kind == 'disease_desc':desc = [i['m.desc'] for i in answers]subject = answers[0]['m.name']final_answer = '{0},熟悉一下：{1}'.format(subject, '；'.join(list(set(desc))[:self.num_limit]))elif question_kind == 'disease_acompany':desc1 = [i['n.name'] for i in answers]desc2 = [i['m.name'] for i in answers]subject = answers[0]['m.name']desc = [i for i in desc1 + desc2 if i != subject]final_answer = '{0}的症状包括：{1}'.format(subject, '；'.join(list(set(desc))[:self.num_limit]))elif question_kind == 'disease_not_food':desc = [i['n.name'] for i in answers]subject = answers[0]['m.name']final_answer = '{0}忌食的食物包括有：{1}'.format(subject, '；'.join(list(set(desc))[:self.num_limit]))elif question_kind == 'disease_do_food':do_desc = [i['n.name'] for i in answers if i['r.name'] == '宜吃']recommand_desc = [i['n.name'] for i in answers if i['r.name'] == '推荐食谱']subject = answers[0]['m.name']final_answer = '{0}宜食的食物包括有：{1}\n推荐食谱包括有：{2}'.format(subject, ';'.join(list(set(do_desc))[:self.num_limit]),';'.join(list(set(recommand_desc))[:self.num_limit]))elif question_kind == 'food_not_disease':desc = [i['m.name'] for i in answers]subject = answers[0]['n.name']final_answer = '患有{0}的人最好不要吃{1}'.format('；'.join(list(set(desc))[:self.num_limit]), subject)elif question_kind == 'food_do_disease':desc = [i['m.name'] for i in answers]subject = answers[0]['n.name']final_answer = '患有{0}的人建议多试试{1}'.format('；'.join(list(set(desc))[:self.num_limit]), subject)elif question_kind == 'disease_drug':desc = [i['n.name'] for i in answers]subject = answers[0]['m.name']final_answer = '{0}通常的使用的药品包括：{1}'.format(subject, '；'.join(list(set(desc))[:self.num_limit]))elif question_kind == 'drug_disease':desc = [i['m.name'] for i in answers]subject = answers[0]['n.name']final_answer = '{0}主治的疾病有{1},可以试试'.format(subject, '；'.join(list(set(desc))[:self.num_limit]))elif question_kind == 'disease_check':desc = [i['n.name'] for i in answers]subject = answers[0]['m.name']final_answer = '{0}通常可以通过以下方式检查出来：{1}'.format(subject, '；'.join(list(set(desc))[:self.num_limit]))elif question_kind == 'check_disease':desc = [i['m.name'] for i in answers]subject = answers[0]['n.name']final_answer = '通常可以通过{0}检查出来的疾病有{1}'.format(subject, '；'.join(list(set(desc))[:self.num_limit]))return final_answer

各模块组装与问答类构建

该模块就是将pipline中的各个模块组装起来。具体如下：

KGQAMedicine\chatbot.py

from question_classify.rule_question_classify import RuleQuestionClassifier
from question_parser.rule_question_parser import RuleQuestionParser
from answer_search.raw_answer_search import RawAnswerSearcherclass ChatBot(object):def __init__(self):self.classifier = RuleQuestionClassifier()self.parser = RuleQuestionParser()self.answer_generate = RawAnswerSearcher()self.common_answer = "您好，我是科皮子菊的医药私人助手，希望可以为您解答。如果答案不满意，可以通过：https://github.com/Htring 联系我哦。祝您身体健康，远离我哦！"def answer(self, question):question_classify = self.classifier.classify(question)if not question_classify:return self.common_answerres_sql = self.parser.parser(question_classify)final_answers = self.answer_generate.search(res_sql)if not final_answers:return self.common_answerelse:return "\n".join(final_answers)if __name__ == '__main__':chat_bot = ChatBot()while True:question = input("用户：")answer = chat_bot.answer(question)print("科皮子菊：", answer)

效果展示：

总结

总得来说，这个项目把使用知识图谱进行QA的一些流程介绍的比较清楚，但是在完成问答的过程中技术相对老旧，不过效果依然还不错。源码已经放到我的github上：https://github.com/Htring/KGQAMedicine，有兴趣的可以下载运行看看哦，上面有运行介绍哦。

为了能够进一步提升效果的话可以引入很多新技术。例如在问题分类环节可以引入基于深度学习的问题分类方法，在进行问题解析的时候，可以引入基于深度学习的NER实体识别方式以及进一步处进行实体对齐等，这里不作进一步展开。

除此之外，知识图谱在构建是需要结合业务需求，也就是在接到业务的时候以及对现有数据进行分析然后构建基于业务的schema，再通过自然语言处理相关技术进行知识图谱的构建。在原项目中，使用爬虫的方式进行数据爬取，其也可以使用NLP相关的基础，优化提取的数据等等。

雄关漫道真如铁，而今迈步重头越。刚入门，一个新的开始。往后和增加更多基于深度学习算法的内容到知识图谱的建设，应用等。