Abstract

Sign languages express meaning through sequentially and simultaneously produced manual and non-manual articulations, making them structurally different from spoken languages. However, many linguistic theories remain biased towards spoken language structures. An exception is construction grammar, a linguistic theory which analyses language through form-meaning mappings that can involve manual and non-manual articulations and can combine specific and schematic structures. While construction grammar is increasingly used to describe sign language constructions theoretically, computational implementations remain sparse, limiting its systematic application to corpora. The theoretical applicability of construction grammar to signed languages, however, does not guarantee that the theory can be computationally operationalised, which introduces challenges in representation and processing. This thesis addresses these challenges by introducing SignFCG, an extension of Fluid Construction Grammar (FCG). SignFCG introduces phonetic representations of manual articulations that account for their sequential and simultaneous organisation. It also provides avatar visualisations for signed forms and tools to transform linguistic annotations into the proposed format. To investigate the types of constructions that can be operationalised for a restricted domain of French-Belgian Sign Language (LSFB), a computational grammar comprising 904 constructions was manually developed using GeoQuery-LSFB, an extension of the multilingual GeoQuery corpus created as part of this thesis. The grammar was evaluated in comprehension and production, achieving high coverage. A user study based on semi-structured interviews with sign language researchers identified perceptions of SignFCG’s usability and potential adoption. Participants identified promising applications, including exploring and formalising constructions and studying constructional variation across contexts. The user study also identified a need for greater flexibility in SignFCG’s form representation, which may require varying levels of detail to answer different research questions. The contributions of this thesis are situated at the intersection of sign language linguistics research, construction grammar, and computational linguistics. For construction grammar, it demonstrates how constructional principles can be operationalised for the visual-gestural modality and how the combination of specific and schematic structures within constructions can facilitate this. For computational linguistics, it demonstrates the feasibility of bidirectional processing between manually signed forms and structured meanings within a restricted domain, while providing resources for further research. Extending FCG to signed languages also provides a basis for exploring other computational research areas in which FCG has been applied.

Jury

  • Prof. Katrien BEULS - University of Namur, Belgium
  • Prof. Bruno DUMAS - University of Namur, Belgium
  • Prof. Anthony CLEVE - University of Namur, Belgium
  • Prof. Laurence MEURANT - University of Namur, Belgium
  • Prof. Paul VAN EECKE - Artificial Intelligence Laboratory, Vrije Universiteit Brussel, Belgium
  • Dr. Remi VAN TRIJP - Sony Computer Science Laboratories, France
  • Dr. Michael FILHOL - LISN, CNRS, France