French Amalgam: a quick adaptation of a sentence realization system to French

The adaptation of a machine-learned sentence realization system to French Martine Smets, Michael Gamon, Simon Corston-Oliver and Eric Ringger Microsoft Research One Microsoft Way Redmond, WA 98052, USA { martines, mgamon, simonco, ringger } @microsoft.com Abstract We describe the adaptation to French of a machine-learned sentence realization system called Amalgam that was originally developed to be as language independent as possible and was first implemented for German. We discuss the development of the French implementation with particular attention to the degree to which the original system could be re- used, and we present the results of a human evaluation of the quality of sentence realization using the new French system. Introduction Recently, statistical and machine-learned approaches have been applied to the sentence realization phase of natural language generation. The Nitrogen system, for example, uses a word bigram language model to score and rank a large set of alternative sentence realizations (Langkilde and Knight, 1998a, 1998b). Other recent approaches use syntactic representations. FERGUS (Bangalore and Rambow, 2000), Halogen (Langkilde 2000, Langkilde-Geary 2002) and Amalgam (Corston-Oliver et al., 2002) use syntactic trees as an intermediate representation to determine the optimal string output. The Amalgam system discussed here is a sentence realization system which maps a semantic representation to a surface syntactic tree via intermediate syntactic representations. The mappings are performed with linguistic operations, the context for which is primarily machine-learned. The resulting syntactic tree contains all the necessary information on its leaf nodes from which a surface string can be read. The promise of machine-learned approaches to sentence realization is that they can easily be adapted to new domains and ideally to new languages merely by retraining The architecture of Amalgam was intended to be language- independent, although the system has previously only been applied to German sentence realization. Adapting this system to French allows us to assess which aspects of the system are truly language-independent and what must be added in order to account for French. The purpose of this paper is to focus on the adaptation of Amalgam to French. Discussions about the general architecture of the system can be found in Corston-Oliver et al. (2002) and Gamon et al. (2002b). 1 Overview of German Amalgam Amalgam takes as its input a logical form graph, i.e., a sentence-level dependency graph with fixed lexical choices for content words. This graph represents the predicate-argument structure of a sentence and includes semantic information concerning relations between nodes of the graph (Heidorn, 2002). Examples of French logical forms are given in section 3. Amalgam first degraphs the logical form into a tree and then augments it by the insertion of function words, 323 assignment of case and verb position features, syntactic labels, etc., to produce an unordered syntax tree. Amalgam then establishes intra- constituent order. After syntactic aggregation, insertion of punctuation, morphological inflection, and capitalization, an output string is read off the leaf nodes. The contexts for most of these linguistic operations are machine-learned (Gamon et al., 2002a). Figure 1 lists the eight stages …

Content truncated to 500 words