Assigning Deep Lexical Types Using Structured Classifier Features for Grammatical Dependencies

Deep linguistic grammars are able to provide rich and highly complex grammatical representations of sentences, capturing, for instance, long-distance dependencies and returning a semantic representation. These grammars lack robustness in the sense that they do not gracefully handle words missing from their lexicon. Several approaches have been explored to handle this problem, many of which consist in pre-annotating the input to the grammar with shallow processing machine-learning tools. Most of these tools, however, use features based on a fixed window of context, such as n-grams. We investigate whether the use of features that encode discrete structures, namely grammatical dependencies, can improve the performance of a machine learning classifier that assigns deep lexical types. In this paper we report on the design and evaluation of this classifier.