Section: New Results
Common Basic Resources
Participants : Maxime Amblard, Clément Beysson, Philippe de Groote, Bruno Guillaume, Maxime Guillaume, Guy Perrier, Sylvain Pogodalla, Nicolas Lefebvre.
Application of Graph Rewriting to Natural Language Processing
Guillaume Bonfante, Bruno Guillaume and Guy Perrier collected their work on the application of graph rewriting to Natural Language Processing (NLP) in a book written in French  and translated to English  by the editor. This book shows how graph rewriting can be used as a computational model adapted to NLP. Currently, there is no standard model for graph rewriting and, as such, the authors have conceived one that is specifically adapted to NLP, proposing their own implementation: the GREW system. In addition to the application to Syntax-Semantic Interface mentioned above, the book presents applications in syntactic parsing and in syntactic corpus conversion.
In , Guillaume Bonfante and Bruno Guillaume describe some mathematical properties of the Graph Rewriting framework used in GREW. The previous experiments on NLP tasks have shown that Graph Rewriting applications to Natural Language Processing do not require the full computational power of the general Graph Rewriting setting. The most important observation is that all graph vertices in the final structures are in some sense "predictable" from the input data and so, it is possible to consider the framework of Non-size increasing Graph Rewriting. The paper concerns the theoretical aspect of termination with respect to this calculus. It is shown that uniform termination is undecidable and that non-uniform termination is decidable. We define termination techniques based on weight, we prove the termination of weighted rewriting systems and we give complexity bounds on derivation lengths for these rewriting systems.
Building Linguistics Resources with Crowdsourcing
In the Joint Workshop on Linguistic Annotation, Multiword Expressions and Constructions, Karën Fort (Sorbonne Université), Bruno Guillaume, Matthieu Constant (ATILF, Nancy), Nicolas Lefebvre and Yann-Alan Pilatte (Sorbonne Université) presented the results obtained in crowdsourcing French speakers’ intuition concerning multi-word expressions (MWEs) . They developed a slightly gamified crowdsourcing platform, part of which is designed to test users’ ability to identify MWEs with no prior training. The participants perform relatively well at the task, with a recall reaching 65% for MWEs that do not behave as function words.
Kim Gerdes (Sorbonne nouvelle, Paris 3), Bruno Guillaume, Sylvain Kahane (Université Paris Nanterre) and Guy Perrier proposed a surface-syntactic annotation scheme called Surface Universal Dependencies (SUD) that is near-isomorphic to the Universal Dependencies (UD) annotation scheme. The SUD scheme follows distributional criteria for defining the dependency tree structure and the naming of the syntactic functions . Rule-based graph transformation grammars allow for a bi-directional transformation of UD into SUD. The back-and-forth transformation can serve as an error-mining tool to assure the intra-language and inter-language coherence of the UD treebanks. The UD corpora are available on gitlab.inria.fr.
Bruno Guillaume and Guy Perrier used the GREW system for the development of the French part of the Universal Dependencies project (UD) . They focused in particular on correcting the annotation of two French corpora, UD_French-GSD and UD_French-Sequoia. For the correction, they first used the tool Grew-match (based on the pattern matching part of GREW) to detect error patterns, but also the GREW rewriting rule system to transform the annotation from one format to another one . Version 2.3 of the UD corpora was released on 15 November 2018.
Maxime Amblard, Clement Beysson, Philippe de Groote, Bruno Guillaume and Sylvain Pogodalla continue their work on the FR-Fracas project. There are two major levels of processing that are significant in the use of a computational semantics framework: semantic composition, for the construction of meanings, and inference, either to exploit those meanings, or to assist the determination of contextually sensitive aspects of meanings. FraCas is an inference test suite for evaluating the inferential competence of different NLP systems and semantic theories. Providing an implementation of the inference level was beyond the scope of FraCaS, but the test suite nevertheless provides an overview of a useful and theory- and system-independent semantic tool .
There currently exists a multilingual version of the resource for Farsi, German, Greek, and Mandarin. Sémagramme completed the translation into French of the test suite. All translations were subject to a bidding phase by two project members. Then the cases that were identified as difficult were discussed by all project members. An adjudication step finally ensured the quality of the translation. In order to evaluate the inference mechanism triggered by the translated sentences, a web interface is being developed.
Large Coverage Abstract Categorial Grammars
Maxime Amblard, Maxime Guillaume, and Sylvain Pogodalla have worked on the automatic translation of large coverage Tree-Adjoining grammars into Abstract Categorial Grammars. On the theoretical side, this work hinges on the encoding proposed by Philippe de Groote and Sylvain Pogodalla , . On the implementation side, the starting point are TAG grammars generated from meta-grammars by XMG , . This generates Abstract Categorial grammars containing about 23 000 entries, and was used as a test bed for the ACGtk toolkit, some parts of which have been rewritten to scale up.