Amalgam: A Machine-Learned Generation Module

Amalgam is a novel system for sentence realization during natural language generation. Amalgam takes as input a logical form graph, which it transforms through a series of modules involving machine-learned and knowledge-engineered sub-modules into a syntactic representation from which an output sentence is read. Amalgam constrains the search for a fluent sentence realization by following a linguistically informed approach that includes such component steps as raising, labeling of phrasal projections, extraposition of relative clauses, and ordering of elements within a constituent. In this technical report we describe the architecture of Amalgam based on a complete implementation that generates German sentences. We describe several linguistic phenomena, such as relative clause extraposition, that must be handled in order to successfully generate German.