Automatic distinction of arguments and modifiers

The automatic distinction of arguments and modifiers is a necessary step for the automatic acquisition of subcategorisation frames and argument structure. In this work, we report on supervised learning experiments to learn this distinction for the difficult case of prepositional phrases attached to the verb. We develop statistical indicators of linguistic diagnostics for argumenthood, and we approximate them with counts extracted from an annotated corpus. We reach an accuracy of 86.5%, over a baseline of 74%, showing that this novel method is promising in solving this difficult problem.