ATMS-Based architecture for stylistics-aware text generation

Abstract

This thesis is concerned with the effect of surface stylistic constraints (SSC) on syntactic and lexical choice within a unified generation architecture. Despite the fact that these issues have been investigated by researchers in the field, little work has been done with regard to system architectures that allow surface form constraints to influence earlier linguistic or even semantic decisions made throughout the NLG process. By SSC we mean those stylistic requirements that are known beforehand but cannot be tested until after the utterance or — in some lucky cases — until a proper linearised part of it has been generated. These include collocational constraints, text size limits, and poetic aspects such as rhyme and metre to name a few. This thesis introduces a new NLG architecture that can be sensitive to surface stylistic requirements. It brings together a well-founded linguistic theory that has been used in many successful NLG systems (Systemic Functional Linguistics, SFL) and an exist¬ ing AI search mechanism (the Assumption-based Truth Maintenance System, ATMS) which caches important search information and avoids work duplication. To this end, the thesis explores the logical relation between the grammar formalism and the search technique. It designs, based on that logical connection, an algorithm for the automatic translation of systemic grammar networks to ATMS dependency networks. The generator then uses the translated networks to generate natural language texts with a high paraphrasing power as a direct result of its ability to pursue multiple paths simultaneously. The thesis approaches the crucial notion of choice differently to previ¬ ous systems using SFL. It relaxes the choice process in that choosers are not obliged to deterministically choose a single alternative allowing SSC to influence the final lexical and syntactic decisions. The thesis also develops a situation-action framework for the specification of stylistic requirements independently of the micro-semantic input. The user or application can state what surface requirements they wish to impose and the ATMS-based generator then attempts to satisfy these constraints. Finally, a prototype ATMS-based generation system embodying the ideas presented in this thesis is implemented and evaluated. We examine the system's stylistic sensitivity by testing it on three different sets of stylistic requirements, namely: collocational, size, and poetic constraints.