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.

Word document
PDF file


InstitutionMicrosoft Research
> Publications > Amalgam: A machine-learned generation module