A framework for structured semantic representation capable of active sensing and interpretable inference: A cancer prognostic analysis case study

Abstract

Precise semantic representation is important for allowing machines to truly comprehend the meaning of natural language text, especially biomedical literature. Although the semantic relations among words in a single sentence may be accurately represented with existing approaches, relations between two sentences cannot yet be accurately modeled, which leads to a lack of contextual information and difficulty in performing interpretable semantic inference. Additionally, it is challenging to merge semantic representations curated by different experts. These critical challenges are insufficiently addressed by existing methods. In this paper, we present a framework for structured semantic representation (FSSR) to address these issues. FSSR uses a double-layer structure Construct that combines Paradigm and Instance to represent the semantics of a word or a sentence. It uses six types of rules to represent the semantic relations between sentence Constructs and uses a Computational Model to represent an action. FSSR is a graph-based representation of semantics, in which a node represents a Construct or a Paradigm. Two nodes are connected by an edge (a rule). In addition, FSSR enables interpretable inference and active acquisition of new information, as illustrated in a case study. This case study models the semantics of a cancer prognostic analysis article and reproduces its text results and charts. We provide a website that visualizes the inference process (http://cragraph.synergylab.cn).