Tutorial: Practical Parsing for Downstream Applications

at COLING

August 20, 2018

Santa Fe, New Mexico

Parsing is a fundamental aspect for many NLP applications. However, it is often nor clear how to best incorporate parsers in downstream applications. Using parsers off the shelf is simple but often leads to bad performance. But even for researchers familiar with parsing and language issues, there are many decisions, often overlooked, concerning the algorithms themselves as well as the interaction between parser and language/treebank. This tutorial is intended to give researchers not familiar with parsing a better understanding of the practical implications of the individual decisions made when using parsers for their downstream application. We will cover dependency as well as constituent parsers. In the tutorial, we will cover topics that walk a potential parsing user through the different steps from choosing between dependencies and constituents, choosing a parser, preprocessing, parser parameters, postprocessing, to evaluation and domain adaptation. We want to make sure that dialogs like the following do not happen as often anymore:

• “I am working on question answering for medical texts, and one of the problems we encounter is that the answer is often present, but in a different form from what we expected.”
• “Sounds like a syntax problem. Have you tried using a parser?”
• “Yup, we took the XX parser with the pretrained model for English, but the results were really bad, so we gave up on that idea.”
• “Have you tried retraining the parser for the medical domain?”
• “No, why? It comes with a pretrained model ...”

Organizers:

Daniel Dakota, Indiana University

Sandra Kübler, Indiana University

 
For more information about the tutorial, please send an email to ddakota@indiana.edu