A “natural” Lexicalization Model for Language Generation

We propose a general lexicalization model which accounts for how lexical units are selected and introduced in linguistic utterances during language generation. This model aims at “naturalness” by being based on actual lexical knowledge used in speech; consequently, it should be compatible with standard patterns of behavior shown by humans when they speak (flexibility in computing both content and form of linguistic utterances, prototypical types of mistakes and backtracking, etc.). The main advantage of our model, once implemented in automatic language generation, is that it takes into account fundamental differences that exist between lexical units, with regard to why and how they are used in texts. This is achieved by means of a stratificational approach to lexicalization, where each type of lexical unit is introduced at a proper level of representation, according to the role it plays in the enunciation. Section 1 offers a general characterization of the approach and makes explicit its main assumptions. Sections 2 to 4 successively examine the three levels of transition implied by the stratificational structuring of the model. Section 5 concludes with an examination of its relevance to the design of text generation systems.