Joint parsing of syntactic and semantic dependencies

Zusammenfassung

Syntactic Dependency Parsing and Semantic Role Labeling (SRL) are two main problems in Natural Language Understanding. Both tasks are closely related and can be regarded as parsing on top of a given sequence. In the data-driven approach context, these tasks are typically addressed sequentially by a pipeline of classifiers. A syntactic parser is run in the first stage, and then given the predicates, the semantic roles are identified and classified (Gildea and Jurafsky, 2002). An appealing and largely unexplored idea is to jointly process syntactic dependencies and semantic roles. A joint process could capture some interactions that pipeline systems are unable to model. We expect joint models to improve on syntax based on semantic cues and also the reverse. Despite this potential advantage and the interest in joint processing stimulated by the CoNLL-2008 and 2009 Shared Tasks (Surdeanu et al., 2008; Hajic et al., 2009), very few joint models have been proposed to date, few have achieved attention and fewer have obtained competitive results. This thesis presents three contributions on this topic. The first contribution is to frame semantic role labeling as a linear assignment task. Under this framework we avoid assigning repeated roles to the arguments of a predicate. Our proposal follows previous work on enforcing constraints on the SRL analysis (Punyakanok et al., 2004; Surdeanu et al., 2007). But in our case, we enforce only a relevant subset of these constraints. We solve this problem with the efficient O(n^3) Hungarian algorithm. Our next contributions will rely on this assignment framework. The second contribution of this thesis is a joint model that combines syntactic parsing and SRL (Lluís et al., 2013). We solve it by using dual-decomposition techniques. A strong point of our model is that it generates a joint solution relying on largely unmodified syntactic and SRL parsers. We train each component independently and the dual-decomposition method finds the optimal joint solution at decoding time. Our model has some optimality and efficiency guarantees. We show experiments comparing the pipeline and joint approaches on different test sets extracted from the CoNLL-2009 Shared Task. We observe some improvements both in syntax and semantics when our syntactic component is a first-order parser. Our results for the English language are competitive with respect to other state-of-the-art joint proposals such as Henderson et al., (2013). The third contribution of this thesis is a model that finds semantic roles together with syntactic paths linking predicates and arguments (Lluís et al., 2014). We frame SRL as a shortest-path problem. Our method instead of conditioning over complete syntactic paths is based on the assumption that paths can be factorized. We rely on this factorization to efficiently solve our problem. The approach represents a novel way of exploiting syntactic variability in SRL. In experiments we observe improvements in the robustness of classifiers.