This paper presents an encoding of Generation-TAG (G-TAG) within Abstract Categorial Grammars (ACG). We show how the key notions of G-TAG have a natural interpretation in ACG, allowing us to use its reversibility property for text generation. It also offers solutions to several limitations of G-TAG.
MotivationsG-TAG (Danlos, 1998;Danlos, 2000) is a formalism based on the Tree Adjoining Grammar (TAG) formalism (Joshi et al., 1975;Joshi and Schabes, 1997) dedicated to text generation. It focuses on providing several notions to support useful data structures, such as g-derivation trees or lexical databases, to effectively relate a surface form (a derived tree or a string) to a conceptual representation. An actual implementation in ADA was first provided for French (Meunier, 1997), and it has recently been implemented in the .NET framework as the EasyText NLG system and is operational at Kantar Media, a French subsidiary company of TNS-Sofres .The G-TAG proposal can be seen as a result of the observation of the mismatch between the derivation tree notion of TAG and the expected semantic dependencies (Schabes and Shieber, 1994) from a generation perspective. Several approaches that extend the derivation tree notion of TAG have been proposed to overcome this difficulty. Other approaches showed that the derivation trees still could be used without additional modifications. Such approaches rely on unification (Kallmeyer and Romero, 2004;Kallmeyer and Romero, 2007) or a functional approach to TAG (Pogodalla, 2004; * This work has been supported by the French agency Agence Nationale de la Recherche (ANR-12-CORD-0004).Pogodalla, 2009) 1 based on Abstract Categorial Grammars (ACG) (de Groote, 2001). The latter is intrinsically reversible: the grammars and the algorithms are the same for parsing and for generation.We propose then to study G-TAG under the ACG perspective. We show that the key notion of g-derivation tree naturally express itself in this framework. The surface form construction from a conceptual representation can then use the general algorithms of ACG, the very same ones that can be used in parsing to analyze mildly context sensitive languages (TAG generated language, LCFRS) (de Groote and Pogodalla, 2004), following (Kanazawa, 2007)'s proposal here applied to give an ACG account of G-TAG. We do not consider here the G-TAG treatment of preferences between the different realizations of the same input. Similarly, we do not consider the generation of pronouns used in G-TAG and we will work on integrating a theory of generation of referring expressions.2 Sketching G-TAG G-TAG deals with the How to say it? task of generation. The input is a conceptual representation. A G-TAG grammar includes elementary trees, as any TAG grammar. But it also makes g-derivation trees primary objects, relating them to the elementary trees and considering them as pivot to the conceptual representation level.Conceptual Representation G-TAG conceptual representation makes use of notions as second order relation, first order relat...