Section: New Results

Long-range diachronic variation

Participants : Benoît Sagot, Laurent Romary, Éric Villemonte de La Clergerie, Clémentine Fourrier, Gaël Guibon, Mathilde Regnault, Kim Gerdes.

ALMAnaCH members have resumed their work on longer-range diachronic variation, in two distinct directions:

• Firstly, we have been working on resources and tools for Old French, using contemporary French as a starting point for which resources and tools are available. This work is carried out within the ANR project “Profiterole”, whose goal is to automatically annotate a large corpus of medieval French (9th-15th centuries) in dependency syntax and to provide a methodology for dealing with heterogeneous data as found in such a corpus. Indeed, Old French does not only involve diachronic variation when contrasted with contemporary French. It also involve large internal variation, notably because of diachronic (within Old French), dialectal, geographic, stylistic and genre-based variation. We have carried out experiments on morphosyntactic tagging by trying to determine which parameters and which training sets are the best ones to use when annotating a new text. We explored two approaches for parsing. On the one hand, an ongoing thesis aims at adapting the FRMG metagrammar to medieval French, notably by changing the constraints on certain syntactic phenomena and relaxing the order of words [31], [30]. This work relies on the new morphological and syntactic lexicon for Old French, OFrLex, developed at ALMAnaCH [34]. On the other hand, we conducted parsing experiments with neural models (DyALog's SRNN models).

• Secondly, we have started experiments to investigate whether and under which conditions neural networks can be used for learning sound correspondences between two related languages, i.e. for predicting cognates of source language words in a related target language. In order to obtain suitably large homogeneously phonetised data, we extracted bilingual lexicons and cognate sets from available resources, including our EtymDB etymological database, of which a new, extended version was created in 2019. This data was then used to train and evaluate several neural architectures (seq2seq, Siamese). Preliminary results are promising, but further investigation is required.

These two research directions will find a common ground now that we have begun to investigate, in the context of the Profiterole ANR project, how we can model the diachronic evolution of the lexicon from Old French to contemporary French. Moreover, our work on Basnage's 1701 Dictionnaire Universel, in the context of the BASNUM ANR project might draw some inspiration from the Profiterole project. But since 1700's French is much closer from contemporary French than Old French, another source of inspiration for BASNAGE might come from our work on sociolinguistic variation in contemporary French and more generally on our work on User-Generated Content (UCG).