Approximating a Zulu GF concrete syntax with a neural network for natural language understanding

Abstract

Multilingual Grammatical Framework (GF) domain grammars have been used in a variety of different applications, including question answering, where concrete syntaxes for parsing questions and generating answers are typically required for each supported language. In low-resourced settings, grammar engineering skills, appropriate knowledge of the use of supported languages in a domain, and appropriate domain data are scarce. This presents a challenge for developing domain specific concrete syntaxes for a GF application grammar, on the one hand, while on the other hand, machine learning techniques for performing question answering are hampered by a lack of sufficient data. This paper presents a method for overcoming the two challenges of scarce or costly grammar engineering skills and lack of data for machine learning. A Zulu resource grammar is leveraged to create sufficient data to train a neural network that approximates a Zulu concrete syntax for parsing questions in a proof-of-concept question-answering system.