Verb Prediction Experiment
This is code for an experiment meant to demonstrate:
- clustering verbs using topic models for the purpose of learning scripts or frames
- evaluating the learned topic models by perplexity counting only verbs
It was supposed to show that adding argument and coreference information increases the ability of a model to predict the verbs in a test correctly, and that perplexity is a useful measure of this ability. Unfortunately it wasn't successful. The model with verb arguments only beat the baseline sometimes, and the coreference model completely failed. This is largely because the models are very unsophisticated, but I had predicted that even an unsophisticated model should be able to beat the baseline verb-only LDA model.
Since I've lost interest in continuing this line of research, I am making the code available as-is, on the off-chance that someone may find it useful.
preprocess/
This directory contains a program for preprocessing the Annotated English Gigaword corpus, written in Scala. It uses the Java API for reading in documents from the corpus, and processes the documents using POS tags and dependencies to extract verbs and create data for training the topic models.
src/
This directory contains the Julia code for training and evaluating
four LDA-based topic models. lda.jl and verbonly.jl implement LDA
using Gibbs sampling; verbargs.jl and verbchain.jl implement
variants of the LDA model including verb arguments and verb-argument
coreference relations, respectively.
tools/
This directory contains some scripts written in Julia for setting up experiments and formatting results in a table. These are good examples of how useful Julia is for system scripting; it's much nicer than bash.
Permissions
All source code and documentation in this repository, including this file, was written by Joseph Irwin. No rights reserved.
To the extent possible under law, the author(s) have dedicated all copyright and related and neighboring rights to this software to the public domain worldwide. This software is distributed without any warranty.
You should have received a copy of the CC0 Public Domain Dedication along with this software. If not, see http://creativecommons.org/publicdomain/zero/1.0/.
JAR files
agiga-1.2.jar and vtd-xml-2.10.jar are included from the Annotated
English Gigaword Java API. Both files are licensed under the GNU GPL,
and the source code can be found here: