Towards Robust Linguistically Deep Processing
Speaker: Yi Zhang
Institution: Saarland University
Abstract:
The improvement of computer hardware and processing algorithm is now shifting the focus of deep processing to achieving robustness/board-coverage while maintaining accuracy. For a pure symbolic system (as most of the traditional deep grammars are), enhancing the coverage while maintaining the accuracy is extremely difficult. However, we do see a more promising approach of relaxing the accuracy of the grammar to achieve better robustness/coverage, and relying on the disambiguation model to filter out the inappropriate analyses.
Despite the controversy of whether the grammar should achieve accuracy by itself or via extra mechanism, we see the efficiency problem reemerges as a more practical difficulty. When the grammar gets more coverage and less accuracy, more analyses will be generated (both in parsing and generation). And if not handled appropriately, the situation can easily run out of control due to the combinatory explosion.
In this talk, I will start with the presentation of my approach towards lexical acquisition for deep grammars, followed by a further investigation on acquiring the multiwords expressions, and conclude by addressing some on-going work of selective unpacking.