Case Retrieval Nets for Heuristic Lexicalization in Natural Language Generation Raquel Herv´as and Pablo Gerv´as Departamento de Sistemas Inform´aticos y Programaci´on Universidad Complutense de Madrid, Spain raquelhb@fdi.ucm.es,pgervas@sip.ucm.es Abstract. In this paper we discuss the use of Case Retrieval Nets, a par- ticular memory model for implementing case-base reasoning solutions, for implementing a heuristic lexicalisation module within a natural language generation application. We describe a text generator for fairy tales im- plemented using a generic architecture, and we present examples of how the Case Retrieval Net solves the Lexicalization task. 1 Introduction Natural Language Generation (NLG) is divided into various specific tasks [1], each one of them operating at a different level of linguistic representation (dis- course, semantics, lexical,...). NLG can be applied in domains where communi- cation goals and features of generated texts are diverse, from transcription into natural language of numerical contents [2] to literary texts generation [3]. Each kind of NLG application may need a different division into modules [4]. Given a specific organization (or architecture ) of the system, it may occur that diverse classes of application require different solutions when facing each of the specific tasks involved in the generation process. For a particular task, in processes where a quick answer is required (for instance, in interactive commu- nication between user and machine in real time) it can be useful to use simple solutions based on heuristics, that provide quick answers even if the achieved quality is poor. On the other hand, in situations where long texts of high quality are needed with no constraints on response time it would be better to draw on knowledge-based techniques that exhaustively consider more possibilities. The present paper proposes a case-based approach to decide which words should be used to pick out or describe particular domain concepts or entities in the generated text. The idea is that people do not create new words each time they need to express an idea not used before, but rather they appeal to the lexicon they have acquired throughout time looking for the best way to express the new idea, always taking into account existing relations between the elements of the lexicon they already know. The paper starts with a revision of the Case-Based Reasoning and Lexicaliza- tion fields. Then we expose the fairy tale text generator where the work presented in this paper is implemented, and we consider the performance of the CBR mod- ule. Finally, the obtained results and future research lines are discussed.