Exploring Model Building for Natural Language Understanding Johan Bos Institute for Communicating and Collaborative Systems Division of Informatics, University of Edinburgh 2 Buccleuch Place, Edinburgh EH8 9LW, Scotland, United Kingdom jbos@ inf. ed. ac. uk Abstract In this paper the use of model building for natural language understanding in practi- cal systems is investigated. After outlining several interpretation tasks where model construction would be useful, we perform experiments using two state-of-the-art model builders on the interpretation of imperatives in discourse. The results are acceptable for small discourses and small domains, but don’t scale up for larger domains or longer discouses: the complexity of model generation rises sharply in relation to the difficulty of a task and the size of the dialogue. Several suggestions are provided for future research to improve the performance of model building for larger domains and richer texts. 1 Introduction This paper introduces a discipline in the field of automated reasoning that has received only little attention within computational semantics: model building (sometimes also referred to as model generation, model construction, or model searching ) for first-order logic. Automatic model builders offer a positive handle on the satisfiability problem and are therefore often used in tandem with theorem provers (who offer a negative handle on satisfiability). But a model builder, as its name suggests, has another attractive property: it is able to construct concrete models for first-order theories. As I will show in this paper, the representations of such models are not only flat and extremely easy to process—they also embody the information required for many natural language understanding tasks. Although it is certainly true that automated model building hasn’t reached a state of matureness that automated theorem proving has achieved in the last decades, it is also fair to say that performances in model building have sig- nificantly improved the last years and have reached a level close to be useful in linguistic applications. These recent developments haven’t gone entirely