Natural Language Generation (NLG) is the process of constructing natural language outputs from non-linguistic inputs. NLG process can be usefully decomposed into three stages: macroplanning, microplanning and linguistic realization. In linguistic realization stage, a natural language generator usually receives a fully specified discourse plan and generates individual sentences as constrained by its lexical and grammatical knowledge resources. This thesis focuses on this sentence realization process. We need linguistic theory to describe Korean grammatical knowledge and perform well-organized linguistic processing. We adopted Systemic-Functional grammar (SFG) that has emerged as being well suited to the needs of NLG. SFG is primarily concerned with the functions of languages and describes how the functions may be mapped into or expressed by surface forms. A sentence is constructed as a consequence of the choices of a linguistic knowledge. The selection process and relevant realization rules are represented in a graph called a system network in SFG. This thesis shows functional analysis of Korean grammar based on SFG (Halliday, 1994)’s metafunctions and Korean system network. The generation grammar is organized into sub-grammars. Each of them handles a relevant syntactic category.
The sentence generator in this thesis goes through three stages. In syntactic processing stage, fully specified syntactic descriptions are produced by traversing system network recursively. Next, a linealizer produces ordered list of constituents that have morphological description and spacing information. Finally, this generator fulfills morphological processing for inflecting endings and concatenating morphemes. Experimental results show that our sentence generator produces grammatically correct and natural sentences. Our approach based on SFG brings sentence generator to encapsulation of syntactic knowledge and good extensibility.