My current research project consists in using graph alignment techniques to:
- improve parse ranking,
- suggest new transfer rules for rule-based machine translation systems,
- pack semantic structures.
Before that, for my PhD, I worked on nonstatistical supertagging techniques and graph rewriting for the syntax-semantics interface.
Both directions aimed at providing general answers to questions raised while working within the Interaction Grammars formalism.
Modular graph rewriting for computational linguistics
Part of my PhD reports on the use of modular graph rewriting for computational linguistics in general and the syntax-semantics interface in particular. Graph rewriting is a high-level model of computation which proves to be both powerful and convenient to describe complex systems. In particular, the non-confluence of graph rewriting enables to account for the pervasive ambiguity of natural languages.
Several practical questions however arise when working on a large scale, such as how to control the global behavior of the system, how to ensure the linguistic consistency of the output, or how to keep a reasonable computation cost. A good answer to all these questions is to use modules. This article explains how to use modular graph rewriting to compute underspecified semantic representations, namely Dependency MRSes, from syntactic dependency graphs.
Nonstatistical parsing-based supertagging
I also worked on supertagging techniques that follow Pierre Boullier’s approach in that they are nonstatistical, strict and parsing-based. The general idea is to automatically extract, from a grammar, a simpler grammar that is a superset approximation of the former. This simple grammar is used to filter, before parsing, rich lexicalized syntactic descriptions.
Our approach consists in using the dependency constraints that rich syntactic descriptions contain. Most lexicalized grammatical formalisms rely on such constraints: for example, substitution nodes in Tree Adjoining Grammars (TAG) require the presence of another tree with a compatible root node. Dependency constraints can be further specified to account for the linear ordering of the words ; our method uses this information as well. This article describes the method and two implementations.