The goal is to categorize questions into different semantic classes based on the possible semantic types of the answers. We develop a hierarchical classifier guided by a layered semantic hierarchy of answer types that makes use of a sequential model for multi-class classification Question classification would benefit question answering process further if it has the capacity to distinguish between a large and complex set of finer classes.
We define a two-layered taxonomy, which represents a natural semantic classification for typical answers. The hierarchy contains 6 coarse classes: (ABBREVIATION, DESCRIPTION, ENTITY, HUMAN, LOCATION and NUMERIC VALUE) and 50 fine classes.
A recommended way to use this system us through the pipeline. Please take a look into the pipeline readme for more details.
The dataset used for training the models is provided in this publication.
Also, you have to uncomment the lines in the lbjava definitions that are specify training and testing.
After that, you can simply run and mvn compile -pl question-typer to train the model.
If you use this tool, please cite the following paper:
@inproceedings{LiRo02,
author = {X. Li and D. Roth},
title = {Learning Question Classifiers},
booktitle = {COLING},
pages = {556--562},
year = {2002},
acceptance = {198-435 (45\%)},
url = "http://cogcomp.org/papers/qc-coling02.pdf",
funding = {NSF98,MURI,ITR-MIT},
projects = {CCR,QA,TE},
comment = {Classifying Answer Type for Question Answering. Sequential Classification. Multiclass Classification.},
}