Using shallow parsing to improve robustness of hand-crafted grammars

Abstract

The goal of this project is to use supervised learning to train a model which will be able to create logical forms, relying purely on shallow methods. For the supervised portion, we used annotated data from the Redwoods Treebank as the source of the gold-standard semantic representations. We then used the raw text of each sentence as input into our shallow processing component and used the output to create our underspecified semantic representations. We used the fully specified semantic representations to train a maximum entropy model which then predicted which elements should be added to the underspecified representation. This involved the creation of maximum entropy models for each of the conditions in question, and then collating them to create (more) fully specified representations from our underspecified ones. The new representations were evaluated against the gold-standard Redwoods representations, showing that the general principles here do provide reasonable results, though there is much room for improvement and future work.