GenDR: A Generic Deep Realizer with Complex Lexicalization

GenDR: A Generic Deep Realizer with Complex Lexicalization Franc¸ois Lareau, Florie Lambrey, Ieva Dubinskaite, Daniel Galarreta-Piquette, Maryam Nejat OLST, D´epartement de linguistique et de traduction, Universit´e de Montr´eal C.P. 6128, succ. Centre-Ville, Montr´eal QC, H3C 3J7, Canada {francois.lareau, florie.lambrey, ieva.dubinskaite, daniel.galarreta-piquette, maryam.sadat.nejat}@umontreal.ca Abstract We present a generic deep realizer called GenDR, which takes as input an abstract semantic representation of predicate-argument relations, and produces corresponding syntactic dependency structures in English, French, Lithuanian and Persian, with the possibility to fairly easily add more languages. It is generic in that it is designed to operate across a wide range of languages and applications, given the appropriate lexical resources. The focus is on the lexicalization of multiword expressions, with built-in rules to handle thousands of different cross-linguistic patterns of collocations (intensifiers, support verbs, causatives, etc.), and on rich paraphrasing, with the ability to produce many syntactically and lexically varied outputs from the same input. The system runs on a graph transducer, MATE (Bohnet et al., 2000; Bohnet and Wanner, 2010), and its grammar design is directly borrowed from MARQUIS (Lareau and Wanner, 2007; Wanner et al., 2009; Wanner et al., 2010), which we have trimmed down to its core and built upon. The grammar and demo dictionaries are distributed under a CC-BY-SA licence (http://bit.ly/2x8xGVO). This paper explains the design of the grammar, how multiword expressions (especially collocations) are dealt with, and how the syntactic structure is derived from the relative communicative salience of the meanings involved. Keywords: multilingual natural language generation, deep realization, lexicalization, multiword expressions, collocations 1. Introduction Natural language generation (NLG) is one of the rare tasks in natural language processing (NLP) that is not yet com- pletely dominated by statistical or neuronal methods. We believe the main reason for this is that for real-life applica- tions, the output of an NLG system must often be flawless, so that extremely high precision is required, and rule-based methods still outperform other approaches in that respect. A number of domain-independent text realizers have been developed to provide an easier way of producing text au- tomatically. While some have impressive coverage, many expect as input a syntactic structure, thus offering very little flexibility in terms of lexical choice and structure. Others take a more abstract input, offering more syntactic flexibility, but their lexicalization model is rather rigid. We propose a multilingual generic deep realizer, GenDR, that is a platform for the modeling of the semantics-syntax interface in lan- guages. In this paper, we will show how GenDR deals with two crucial tasks: arborization (building a syntactic struc- ture that reflects the semantic structure) and lexicalization (picking the right words to express the desired meanings). 2. Previous work Most realizers that we know of expect an input where both lexical choice and syntactic structure have already been com- puted, leaving the user with two particularly complex tasks. This is the case of FUF/SURGE (Elhadad, 1993; Elhadad and Robin, 1996), RealPro (Lavoie and Rambow, 1997; Co- GenTex, 1998), SimpleNLG (Gatt and Reiter, 2009), its bilingual version, …

Content truncated to 500 words