French Amalgam: A machine-learned sentence realization system

TALN 2003, Batz-sur-Mer, 11-14 juin 2003 French Amalgam: A machine-learned sentence realization system Martine Smets, Michael Gamon, Simon Corston-Oliver and Eric Ringger Microsoft Research One Microsoft Way Redmond, WA 98052, U.S.A. {martines, mgamon, simonco, ringger}@microsoft.com Résumé – Abstract Cette communication présente la version pour le français d’Amalgam, un système de réalisation automatique de phrases. Deux des modèles du système sont décrits en détail, et nous expliquons comment la performance des modèles peut être améliorée en combinant connaissances et intuition linguistiques et méthodes statistiques. This paper presents the French implementation of Amalgam, a machine-learned sentence realization system. It presents in some detail two of the machine-learned models employed in Amalgam and shows how linguistic intuition and knowledge can be combined with statistical techniques to improve the performance of the models. Keywords – Mots Clés Réalisation de phrase, génération automatique, arbres de décision, français. Sentence realization, generation, machine-learning, decision trees, French. 1 Introduction Amalgam is a multilingual sentence realization system. Developed originally for German (Corston-Oliver et al. 2002, Gamon et al. 2002), it has been adapted to French (Smets et al. 2003). Amalgam maps a representation of propositional content (henceforth "logical form") to a surface syntax tree via intermediate syntactic representations. The mappings are performed with linguistic operations, the context for which is primarily machine-learned. The leaf nodes of the resulting syntax tree contain all necessary information from which to generate the surface string. Other systems use machine learning techniques for sentence realization in 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 Smets, Gamon, Corston-Oliver and Ringger approaches use syntactic representations: FERGUS (Bangalore and Rambow 2000) and Halogen (Langkilde 2000, Langkilde-Geary 2002) use syntax trees as an intermediate representation to determine the optimal string output. The adaptation of German Amalgam to French has been discussed elsewhere (Smets et al. 2003). In this paper, we discuss French Amalgam in some detail by presenting two of the machine-learned models and a tool which allows the researcher to manually inspect the relevant training data for each model. In this way, the researcher can understand why the model makes the decisions it makes and can improve the model by adding relevant linguistic features. The researcher can thus leverage the richness of the linguistic information to optimize the models. 2 French Amalgam Amalgam takes as its input a logical form graph, i.e., a sentence-level dependency graph with fixed lexical choices for content words. Function words are represented as features or attributes (c.f. Heidorn 2000). The logical form represents the predicate-argument structure of the sentence and also includes some semantic information (e.g., the meaning of time and location prepositions). During sentence realization, the logical form is first degraphed into a tree and then augmented by the insertion of function words (determiners, auxiliaries, some prepositions, etc.) and syntactic labels. Linguistic operations such as the introduction of coordination, raising, ordering, aggregation, punctuation, inflection, etc. are performed to produce …

Content truncated to 500 words