Semantic Knowledge 💬

Abstract

The demand for clinical decision support systems in medicine and self-diagnostic symptom checkers has substantially increased in recent years. They depend on large biomedical databases, such as knowledge graphs (KG), to derive inferences. KGs express medical concepts as connections between multiple nodes in a graph. Most biomedical KGs are incomplete and miss relationships between nodes.
This work employed a deep medical reasoning system to infer missing relationships from a biomedical KG by synthesizing existing information. A question answering (QA) framework is conducted to explicitly infer relationships that are not directly linked in the KG. An embedding-based model was trained to answer questions based on semantic similarities. We were able to achieve high accuracy scores using the statistic metrics Hits@k and MRR. Further investigations revealed limitations when facing more complex chains of reasoning and the lack of transparency in cases of misclassifications. Following the latest research, we were able to overcome these limitations by using a path-based, multi-hop reinforcement learning (RL) approach. We trained the RL agent and presented its ability to use the symbolic compositionality of relations in a qualitative analysis. Given a question, the agent 'walks' on the graph by choosing relations at each step. Thus, providing an explainable reasoning path. Such computational methods can help in a variety of downstream tasks, such as generate hypotheses of potential protein associations or predict explainable interactions in drug-symptom networks.

Biomedical Knowledge Graphs

A Knowledge Graph represents knowledge in the form of interconnections between nodes and edges in a graph. It provides a structured representation of facts drawn from the relationships between entities. Where entities can be real-world objects or abstract concepts and Relationships represent the relation between them. The concept of representing knowledge in a graph initially gained traction in the web community by shifting from a ‘web of documents’ towards a semantic web. This new form of web content is structured and tagged to provide a meaningful representation for machines to explore the data. To regulate the way publishing and interlinking data on the web in relational manner works, the World Wide Web Consortium (W3C) released a collection of international standards, such as the Resource Description Framework (RDF). It defines the representation of a fact in the form of a triple HEAD, RELATION, TAIL, (h,r,t), or subject, predicate, object, (s,p,o). Appreciations are (h,r,t) and (s,p,o), respectively. A combination of multiple facts stored in a list is also known as a knowledge base (KB). The term knowledge base is synonymous with a knowledge graph with a minor difference. A knowledge graph can be viewed as a graph when considering its graph structure. When it involves formal semantics, it can be taken as a knowledge base for interpretation and inference over facts.

Graph embeddings

Graph embeddings mainly focuses on the process of learning low dimensional distributed embedding for entities and relations. Current literature uses real-valued point-wise space including vector, matrix and tensor space, while other kinds of space such as complex vector space, Gaussian space, and manifold are utilized as well. Models used can be categorized in translational distance an semantic matching. For further technical details please check out my corresponding blog post on Medium.

Path-based Reasoning in Knowledge Graphs

Learning embeddings of entities and relations using tensor factorization or neural methods to optimize a predefined scoring function between facts in the KG, achieves outstanding results on standard benchmark datasets. However, these methods cannot capture chains of reasoning expressed by paths in the KG and cannot capture the full underlying semantics. Path-based models address these problems by scoring facts based either on random walks over the KBs (path ranking) or by encoding entire paths in a vector space (path encoding).

Framework