A High-Performance Syntactic and Semantic Dependency Parser

This demonstration presents a highperformance syntactic and semantic dependency parser. The system consists of a pipeline of modules that carry out the tokenization, lemmatization, part-of-speech tagging, dependency parsing, and semantic role labeling of a sentence. The system’s two main components draw on improved versions of a state-of-the-art dependency parser (Bohnet, 2009) and semantic role labeler (Björkelund et al., 2009) developed independently by the authors. The system takes a sentence as input and produces a syntactic and semantic annotation using the CoNLL 2009 format. The processing time needed for a sentence typically ranges from 10 to 1000 milliseconds. The predicate–argument structures in the final output are visualized in the form of segments, which are more intuitive for a user. 1 Motivation and Overview Semantic analyzers consist of processing pipelines to tokenize, lemmatize, tag, and parse sentences, where all the steps are crucial to their overall performance. In practice, however, while code of dependency parsers and semantic role labelers is available, few systems can be run as standalone applications and even fewer with a processing time per sentence that would allow a Authors are listed in alphabetical order. user interaction, i.e. a system response ranging from 100 to 1000 milliseconds. This demonstration is a practical semantic parser that takes an English sentence as input and produces syntactic and semantic dependency graphs using the CoNLL 2009 format. It builds on lemmatization and POS tagging preprocessing steps, as well as on two systems, one dealing with syntax and the other with semantic dependencies that reported respectively state-of-the-art results in the CoNLL 2009 shared task (Bohnet, 2009; Björkelund et al., 2009). The complete system architecture is shown in Fig. 1. The dependency parser is based on Carreras’s algorithm (Carreras, 2007) and second order spanning trees. The parser is trained with the margin infused relaxed algorithm (MIRA) (McDonald et al., 2005) and combined with a hash kernel (Shi et al., 2009). In combination with the system’s lemmatizer and POS tagger, this parser achieves an average labeled attachment score (LAS) of 89.88 when trained and tested on the English corpus of the CoNLL 2009 shared task (Surdeanu et al., 2008). The semantic role labeler (SRL) consists of a pipeline of independent, local classifiers that identify the predicates, their senses, the arguments of the predicates, and the argument labels. The SRL module achieves an average labeled semantic F1 of 80.90 when trained and tested on the English corpus of CoNLL 2009 and combined with the system’s preprocessing steps and parser.