Joint Learning of Syntactic and Semantic Dependencies

In this master’s thesis we designed, implemented and evaluated a novel joint syntactic and semantic parsing model. Syntactic and semantic parsing have been and are still being addressed as sequence or pipeline of tasks. As far as we know, the only open domain exception to this pipeline approach was published by Musillo and Merlo (2006). The pipeline processing implies the undesirable and hard to recover effect of error propagation across components. Furthermore, syntax and semantics are assumed to interact between themselves at some degree and these interactions cannot be modeled by a pipeline system. Thus the main objective of this work is to design a joint model and compare the pipeline and joint approaches. Also and in spite of a fair comparison, the computing costs are evaluated. This master’s thesis is concerned to the joint syntactic and semantic parsing under the machine learning paradigm. Therefore a dataset must be available and evaluation measures provided. The CoNLL-2008 shared task (Surdeanu et al., 2008), devoted to joint parsing of syntactic and semantic dependencies, brings these elements and presents an excellent framework to train and evaluate our model. Shared task organizes merged together several data sources to provide training and testing datasets. Furthermore, the shared task evaluation framework comprises widely used measures and in addition our system can be compared to other shared task contributing teams. Our proposed model for joint parsing relies on an on-line structured perceptron for learning (Collins, 2002) and on the Eisner algorithm (Eisner, 1996) for inference. The Eisner algorithm is a bottom-up parser previously successfully extended in the context of syntactic parsing (McDonald and Pereira, 2006; Carreras, 2007). Our proposal is one step further and, as far as we know, for the first time the Eisner algorithm is applied to jointly parse syntactic and semantic dependencies. At each algorithm step a syntactic dependency and some semantic dependencies