Representing unstructured text semantics for reasoning purpose

摘要

To interpret a natural language text using a machine, we need to convert its semantics into structured information. In the field of Natural Language Processing, multiple tasks have been designed and developed to interpret the semantics of an unstructured text, and change words into meanings. However, there are some challenges in directly using the output of these tasks in subsequent applications such as logical inference. There has been a growing interest in building and enhancing state-of-the-art semantic representation systems in recent years. However, most of these systems involve supervised models that benefit from manually annotated data, which is not accessible for a wide range of languages. This paper presents a new framework for modeling text in order to extract its information, and through an inference system, obtain new information that is not explicitly stated in the text, but could be logically inferred. This framework is based on Open Information Extraction and Semantic Web techniques for machine reading. We translate the text into a machine-readable representation by using Semantic Types Identification and Question-based Semantic Role Labeling, which could be used in low-resource languages. We integrate the extracted information into the background knowledge by using existing Semantic Web standards. The proposed framework could increase generalization of labelling and reduce ambiguities, therefore, it is an appropriate solution for preparing text for reasoning systems.