知识图谱问答_biji
基于模版查询
基于语义解析
Semantic Parsing 将自然语言转换为SPARQL/SQL。这是一种以语言学为基础的方法。它将自然语言转换为一系列形式化表示这些概念与关系的逻辑结构。通过对这些逻辑结构进行自底向上的分析。最终构建出能够完整表达问题语义的统一的形式化模型。利用相应的查询指令在知识库中检索相关信息。从而完成对问题的答案求解过程。
基于检索排序
信息抽取通过在知识库中检索问题中的实体来实现,在检索到该实体后可以从知识库中获取以该实体为中心的知识子图。这些子图中的节点及其周边节点都被视为候选答案选项。根据观察到的问题并遵循特定规则或模板对问题进行分析从而完成信息提取工作。随后基于这些分析结果构建特征向量并训练分类模型最终通过对候选答案进行筛选确定最终答案。
流程包括: entity extraction, attribute extraction, and knowledge base linking. 具体步骤如下: 1) 实体抽取; 2) 属性提取; 3) 知识库构建关联; 4) 构建知识子图; 5) 建立预测模型; 6) 输出最终答案.

基于深度学习
Vector Modeling 基于问题生成候选回答,并将这些信息转化为分布式表示形式;通过利用训练数据对这种表示方式进行优化,在测试阶段计算生成的答案与真实结果之间的向量相似度;建立评估机制以选择得分最高的选项作为最终输出结果;结合点在于从知识库中提取关键实体并构建相关的子图结构;随后运用深度学习算法对实体与其关联的关系进行建模,并根据计算出的评分排序来确定最优解
- Entity Linking in 100 Languages. (EMNLP 2020).
- An End-to-End Model for Question Answering over Knowledge Base with Cross-Attention Combining Global Knowledge. (ACL 2017)
- Key-Value Memory Networks for Directly Reading Documents. (ACL 2016)
- Enhancing Key-Value Memory Neural Networks for Knowledge Based Question Answering. (NAACL2019)
- Neural Symbolic Machines: Learning Semantic Parsers on Freebase with Weak Supervision. (ACL 2017)
- Improving Question Answering over Incomplete KBs with Knowledge-Aware Reader. (ACL 2019)
- UHop: An Unrestricted-Hop Relation Extraction Framework for Knowledge-Based Question Answering, employing a search framework based on transformation to overcome the limitations of hop-based search, achieving notable success on long relation chains.
- Go for a Walk and Arrive at the Answer: A Reinforcement Learning Approach to Reason Over Paths in Knowledge Bases.
- In addressing complex question answering tasks over knowledge graphs, the study integrates knowledge graph embeddings with problem embeddings, utilizing the Roberta model to encode inputs into representations capturing semantic meaning through self-supervised pre-training strategies.
In evaluating performance, we construct a scoring function based on triplets composed of head entities, questions, and candidate entities within two hops of the head entity's subgraph within the knowledge graph.
This approach mirrors training mechanisms that utilize triplet-based scoring functions during model optimization,
thus establishing a robust framework for multi-hop question answering over knowledge graphs.

借鉴NLG领域中一篇核心论文的思想框架,在本研究中我们主要关注的是通过图神经网络对知识图谱进行建模,并运用语言预处理模型对文本进行初步清洗。随后将这两者有机融合以实现综合分析功能。
该方法专注于生成用于回答多轮复杂问题的Query Graph(QG),其核心在于基于Query机制结合extend、concat和aggregate操作来扩展种子实体,并构建了一个Query模板。通过beam search策略保留前k个最优模式,并将提取的7维特征向量传递给神经网络模型以便对其进行排序处理。

