Fast, Scalable and Reliable Generation of Controlled Natural Language

Software Engineering, Testing, and Quality Assurance for Natural Language Processing, pages 40–48, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Fast, Scalable and Reliable Generation of Controlled Natural Language David Hardcastle Faculty of Maths, Computing and Technology The Open University Milton Keynes, UK d.w.hardcastle@open.ac.uk Richard Power Faculty of Maths, Computing and Technology The Open University Milton Keynes, UK r.power@open.ac.uk Abstract In this paper we describe a natural language generation system which takes as its input a set of assertions encoded as a semantic graph and outputs a data structure connecting the se- mantic graph to a text which expresses those assertions, encoded as a TAG syntactic tree. The scope of the system is restricted to con- trolled natural language, and this allows the generator to work within a tightly restricted domain of locality. We can exploit this fea- ture of the system to ensure fast and efficient generation, and also to make the generator re- liable by providing a rapid algorithm which can exhaustively test at compile time the com- pleteness of the linguistic resources with re- spect to the range of potential meanings. The system can be exported for deployment with a minimal build of the semantic and linguistic resources that is verified to ensure that no run- time errors will result from missing resources. The framework is targeted at using natural lan- guage generation technology to build semantic web applications where machine-readable in- formation can be automatically expressed in natural language on demand. 1 Introduction This paper describes a fast, reliable and scalable framework for developing applications supporting tactical generation – by which we mean applications which take as their input some semantic structure that has already been organised at a high level, and choose the syntactic structures and words required to express it. The framework takes as input a semantic graph representing a set of assertions in Description Logic (DL) (Baader et al., 2003) and transforms it into a tree which encodes the grammar rules, syn- tactic subcategorisations, orderings and lexical an- chors required to construct a textual representation of the input data. The resulting text is conceptu- ally aligned, by which we mean that each compo- nent of the text structure (such as words, clauses or sentences, for example) is linked back to the medi- ating structure from which the text was generated, and from there back to vertices and edges in the semantic graph received as input. The target con- text for the framework is the construction of se- mantic web (Berners-Lee et al., 2001) resources us- ing Natural Language Generation (NLG) technology which extends the notion of semantic alignment de- veloped in the WYSIWYM system (Power and Scott, 1998; Power et al., 2003). In this context the text is ephemeral and is generated on demand, while the document content is fully machine-readable, sup- porting tasks such as automated consistency check- ing, inferencing and semantic search/query. Since the text is fully linked to the underlying semantic representation it supports a rich user interface …

Content truncated to 500 words