FCGlight: Making the bridge between Fluid Construction Grammars and main-stream unification grammars by using feature constraint logics

Fluid Construction Grammars (FCGs) are a flavour of Construction Grammars, which themselves are unification-based grammars. Its syntax is (only) up to certain extent similar to other unification-based grammars. However, it lacks a declarative semantics, while its procedural semantics is truly particular, compared to other unification-based grammar formalisms. Here we propose the re-definition of a subset of the FCG formalism (henceforth called FCGlight) within the framework of (order-sorted) feature constraint logics (OSF-logic) that would assign FCG a rigorous (both declarative a and procedural) semantics, which is suitable for both parsing, generation and grammar learning. In this way we will be able to compare FCG to other unification-based grammars. We will also get the advantage of associating it another classical paradigm for learning (“evolving”) new grammars. This is learning in hierarchies (lattices) of concepts which exploits a partial order relation (generalisation / specialisation) between grammars, instead of the method currently used by FCG, which is (inspired by) reinforcement learning. This will enable a rather natural link with the linguistic background knowledge when devising the grammar repair strategies and a clear way to compare different grammars that could be learned by an agent at each step during the grammar evolution process. In return, inspired by the FCG approach to grammar learning, we are able to define new ways for learning grammars in other unification-based formalisms. In particular, ILP-like learning of HPSGs can be substantially improved by using ideas borrowed from FCG: − instead of using a given “golden” test suite (against which the learning is performed), this test suite is dynamically produced during the language game (played by two agents); − the grammar learning process will be “grounded”, something which was not considered before; − we will learn also new rules (instead of modifying the given ones, as currently done in ilpLIGHT) − learning new rules by generalising upon already learned rules, will also be a significant step forward.