Joint Labeling of Syntactic Function and Semantic Role Using Probabilistic Finite State Automata

Syntactic parsing and semantic labeling are common natural language processing tasks. Traditionally, they were mostly arranged in a pipeline architecture. During the last decade, however, different joint parsing approaches have been introduced where the sequential dependency between the two levels is reduced. In this paper, we present a model for a simplified joint parsing, namely, labeling, based on probabilistic finite state automata through the extended label set paradigm. The parsing (labeling) we present in this research considers syntactic dependency annotation and semantic role labeling without constructing a complete dependency hierarchy. In our experiment, we show that the proposed model outperforms the standard finite transducer approach (Hidden Markov Model). In spite of the considerably increased search space for the joint syntactic and semantic labeling, the proposed solution keeps a high accuracy of the syntactic labeling on par with the quality of syntax-only models. In addition to that it provides a reasonable semantic annotation quality without a separate processing step.