Machine-learned contexts for linguistic operations in German sentence realization

Machine-learned contexts for linguistic operations in German sentence realization Michael GAMON, Eric RINGGER, Simon CORSTON-OLIVER, Robert MOORE Microsoft Research Microsoft Corporation Redmond, WA 98052 {mgamon, ringger, simonco, bobmoore}@microsoft.com Abstract We show that it is possible to learn the contexts for linguistic operations which map a semantic representation to a surface syntactic tree in sentence realization with high accuracy. We cast the problem of learning the contexts for the linguistic operations as classification tasks, and apply straightforward machine learning techniques, such as decision tree learning. The training data consist of linguistic features extracted from syntactic and semantic representations produced by a linguistic analysis system. The target features are extracted from links to surface syntax trees. Our evidence consists of four examples from the German sentence realization system code-named Amalgam: case assignment, assignment of verb position features, extraposition, and syntactic aggregation 1 Introduction The last stage of natural language generation, sentence realization, creates the surface string from an abstract (typically semantic) representation. This mapping from abstract representation to surface string can be direct, or it can employ intermediate syntactic representations which significantly constrain the output. Furthermore, the mapping can be performed purely by rules, by application of statistical models, or by a combination of both techniques. Among the systems that use statistical or machine learned techniques in sentence realization, there are various degrees of intermediate syntactic structure. Nitrogen (Langkilde and Knight, 1998a, 1998b) produces a large set of alternative surface realizations of an input structure (which can vary in abstractness). This set of candidate surface strings, represented as a word lattice, is then rescored by a word- bigram language model, to produce the best- ranked output sentence. FERGUS (Bangalore and Rambow, 2000), on the other hand, employs a model of syntactic structure during sentence realization. In simple terms, it adds a tree-based stochastic model to the approach taken by the Nitrogen system. This tree-based model chooses a best-ranked XTAG representation for a given dependency structure. Possible linearizations of the XTAG representation are generated and then evaluated by a language model to pick the best possible linearization, as in Nitrogen. In contrast, the sentence realization system code-named Amalgam (A Machine Learned Generation Module) (Corston-Oliver et al., 2002; Gamon et al., 2002b) employs a series of linguistic operations which map a semantic representation to a surface syntactic tree via intermediate syntactic representations. The contexts for most of these operations in Amalgam are machine learned. The resulting syntactic tree contains all the necessary information on its leaf nodes from which a surface string can be read. The goal of this paper is to show that it is possible to learn accurately the contexts for linguistically complex operations in sentence realization. We propose that learning the contexts for the application of these linguistic operations can be viewed as per-operation classification problems. This approach combines advantages of a linguistically informed approach to sentence realization with the advantages of a machine Computational Linguistics (ACL), Philadelphia, July 2002, pp. 25-32. Proceedings of the 40th Annual Meeting of the Association …

Content truncated to 500 words