Prédiction structurée pour l'analyse syntaxique en constituants par transitions : modèles denses et modèles creux

The article introduces a transition-based constituent parsing method relying on a deep-learning weighting scheme. This weighting method is compared to a structured perceptron, which is a more traditional method. First of all, we introduce a syntactic parser weighted by a local greedy neural network based on symbol embeddings. Then, we extend this model to a global beam-search based model. Our experiments highlight the surprisingly good properties of the greedy local neural parser.