Tailoring dependency models to NLP tasks
Benjamin Kolz · LA Referencia (Red Federada de Repositorios Institucionales de Publicaciones Científicas) · 2016
Currently available dependency structures differ significantly in the linguistic criteria they are based on, but are not always adequate for their later use in natural language processing tasks. This dissertation analyses the needs of some of these tasks, in particular temporal and discourse parsing, and suggests task-based dependency structures. A surface-syntax dependency structure is taken as base version, which is then tailored to the needs of the corresponding task by means of head selection, customised syntactic function tagset and collapsed dependencies. The work is grounded on the Spanish corpus AnCora, establishing a surface-syntax base version from its constituent structure level. Two dependency models are created, Temporal and Discourse Dependencies, which take the base version as input and adapt it automatically to the task-based versions. The resulting versions are evaluated by network analysis methods, which confirm the adequacy of these new dependency structures with respect to the specific tasks.