Section: Application Domains

Empirical linguistics

Participants : Benoit Crabbé, Benoît Sagot, Alexandra Simonenko, Sarah Beniamine.

Alpage is a team that dedicates efforts in producing ressources and algorithms for processing large amounts of textual materials. These ressources can be applied not only for purely NLP purposes but also for linguistic purposes. Indeed, the specific needs of NLP applications led to the development of electronic linguistic resources (in particular lexica, annotated corpora, and treebanks) that are sufficiently large for carrying statistical analysis on linguistic issues. In the last 10 years, pioneering work has started to use these new data sources to the study of English grammar, leading to important new results in such areas as the study of syntactic preferences [51] , [112] , the existence of graded grammaticality judgments [72] .

The reasons for getting interested for statistical modelling of language can be traced back by looking at the recent history of grammatical works in linguistics. In the 1980s and 1990s, theoretical grammarians have been mostly concerned with improving the conceptual underpinnings of their respective subfields, in particular through the construction and refinement of formal models. In syntax, the relative consensus on a generative-transformational approach [57] gave way on the one hand to more abstract characterizations of the language faculty [57] , and on the other hand to the construction of detailed, formally explicit, and often implemented, alternative formulation of the generative approach [50] , [83] . For French several grammars have been implemented in this trend, such as the tree adjoining grammars of [54] , [61] among others. This general movement led to much improved descriptions and understanding of the conceptual underpinnings of both linguistic competence and language use. It was in large part catalyzed by a convergence of interests of logical, linguistic and computational approaches to grammatical phenomena.

However, starting in the 1990s, a growing portion of the community started being frustrated by the paucity and unreliability of the empirical evidence underlying their research. In syntax, data was generally collected impressionistically, either as ad-hoc small samples of language use, or as ill-understood and little-controlled grammaticality judgements [98] . This shift towards quantitative methods is also a shift towards new scientific questions and new scientific fields. Using richly annotated data and statistical modelling, we address questions that could not be addressed by previous methodology in linguistics.

In this line, at Alpage we have started investigating the question of choice in French syntax with a statistical modelling methodology. In the perspective of better understanding which factors influence the relative ordering of post verbal complements across languages and through language evolution.

On the other hand we are also collaborating with the Laboratoire de Sciences Cognitives de Paris (LSCP/ENS) where we explore the design of algorithms towards the statistical modelling of language acquisition (phonological acquisition). This has been supported in the past years by one PhD project, whose defense has now taken place.

In parallel, quantitative methods are applied to computational morphology, in particlar in relation with Sarah Beniamine's PhD supervised by Olivier Bonami (LLF, CNRS, U. Paris Diderot and U. Paris Sorbonne) [31] , [20] , [32] . Collaborative work in this area is also conducted in collaboration with descriptive linguists from CRLAO (CNRS and Inalco; Guillaume Jacques) and HTL (CNRS, U. Paris Diderot and U. Sorbonne Nouvelle; Aimée Lahaussois) and formal linguists from DDL (CNRS and Université Lyon 2; Géraldine Walther).