Robot Learning of Domain Specific Knowledge from Natural Language Sources

Abstract

The belief that problem solving systems would require only processing power was proven false. Actually almost the opposite is true: for even the smallest problems vast amounts of knowledge are necessary. So the key to systems that would aid humans or even replace them in some areas is knowledge. Humans use texts written in natural language as one of the primary knowledge sources. Natural language is by definition ambiguous and therefore less appropriate for machine learning. For machine processing and use the knowledge must be in a formal; machine readable format. Research in recent years has focused on knowledge acquisition and formalization from natural language sources (documents, web pages). The process requires several research areas in order to function and is highly complex. The necessary steps usually are: natural language processing (transformation to plain text, syntactic and semantic analysis), knowledge extraction, knowledge formalization and knowledge representation. The same is valid for learning of domain specific knowledge although the very first activity is the domain definition. These are the areas that this chapter focuses on; the approaches, methodologies and techniques for learning from natural language sources. Since this topic covers multiple research areas and every area is extensive, we have chosen to segment this chapter into five content segments (excluding introduction, conclusion and references). In the second segment we will define the term domain and provide the reader with an overview of domain engineering (domain analysis, domain design and domain implementation). The third segment will present natural language processing. In this segment we provide the user with several levels of natural language analysis and show the process of knowledge acquirement from natural language (NL). Sub segment 3.1 is about theoretical background on syntactic analysis and representational structures. Sub segment 3.2 provides a short summary of semantic analysis as well as current sources for semantic analysis (WordNet, FrameNet). The fourth segment elaborates on knowledge extraction. We define important terms such as data, information and knowledge and discuss on approaches for knowledge acquisition and representation. Segment five is a practical real world (although on a very small scale) scenario on learning from natural language. In this scenario we limit ourselves on a small segment of health/nutrition domain as we acquire, process and f…