riedelcastro / ucleed

UCL (BioNLP) (E)vent (E)xtractor based on dual (d)ecomposition

UCL/UMass BioNLP Event Extractor

Installation

Get maven, untar/unzip and then run

$ mvn compile

If it looks like dependencies cannot be found, first try this.

ucleed stores preprocessed data in a mongo database. Hence you need to get mongo, and run the mongo server

$ mongod

You should also have an installation of the BioNLP reranking parser by David McClosky on your machine.

You also need to configure a few directory locations. Copy the example in src/main/resources/props/example.prop and modify as needed.

Setting up the syntactic parser

ucleed uses the reranking parser by David McClosky, in combination with his Improved self-trained biomedical parsing model. In the configuration file, set rerankparser to the main directory of the parser, and biomodel to the directory of the biomedical parsing model. Note that for some odd reason, the recent versions of the bllip reranker expects bzip files and not the gzip files provided in the biomodel. You can fix this by calling

$ gunzip *.gz; bzip *

in the biomodel/reranker directory.

Preprocessing

Before we train, we need to go through two preprocessing steps that prepare the data.

Data preprocessing

First call

$ mvn exec:exec -Dexec.executable="java" -Dexec.args="-Xmx1g -Dprop=props/example.prop  -cp %classpath cc.refectorie.proj.bionlp2011.ClearRaw"

to clear the database (this is actually only necessary if you want to rerun experiments but it shouldn't hurt). Then do

$ mvn exec:exec -Dexec.executable="java" -Dexec.args="-Xmx1g -Dprop=props/example.prop  -cp %classpath cc.refectorie.proj.bionlp2011.LowLevelAnnotation dev train test"

This will add tokenize, sentence-split etc. the data specified in the prop file.

Feature preprocessing

Next we run

$ mvn exec:exec -Dexec.executable="java" -Dexec.args="-Xmx1g -Dprop=props/example.prop  -cp %classpath cc.refectorie.proj.bionlp2011.ClearAnnotated"

to initialize the feature preprocessing database. Then do:

$ mvn exec:exec -Dexec.executable="java" -Dexec.args="-Xmx1g -Dprop=props/example.prop  -cp %classpath cc.refectorie.proj.bionlp2011.App dev train test"

This will prepare some candidate structures that are used during inference/learning.

Learning

Now copy data with features to the learning KB:

$ mvn exec:exec -Dexec.executable="java" -Dexec.args="-Xmx1g -Dprop=props/example.prop  -cp %classpath cc.refectorie.proj.bionlp2011.ClearLearningKB"

Finally, you're ready to train the model

$ mvn exec:exec -Dexec.executable="java" -Dexec.args="-Xmx8g -Dprop=props/example.prop -cp %classpath cc.refectorie.proj.bionlp2011.BioNLPLearner"

This will store weights for different epochs into $UMASSDIR/weights/[epoch]

Learning also runs evaluation on test and development sets. The results will appear in the outDir specified in the prop file.

Testing

You can use the stored weights in a standalone tool that applies the complete preprocessing chain and the event extractor model to input files. For this first set weightsSrc=weights/[epoch of choice] in the prop file. Generally epoch 4 or 5 seems to give good results, but can check what works best on the dev set.

Then run the standalone tool as follows:

$ mvn exec:exec -Dexec.executable="java" -Dexec.args="-Xmx80g -Dprop=props/example.prop -cp %classpath cc.refectorie.proj.bionlp2011.UMassBioEventExtractor [txt file] [a1file] [destfile]"

Further Reading and Citations

The most relevant citation for this work is our EMNLP paper. Further details can be found in our BioNLP shared task papers on system combination and dual decomposition.

ucleed's Issues

Installation trouble

Firstly, Maven was unable to find certain packages such as factorie, mongodb, etc. I installed them as required by Maven. However, while compiling the sources, mvn comes up with the following error :-

[ERROR] ~/ucleed/src/main/scala/cc/refectorie/proj/bionlp2011/AntiTransitivityModule.scala:5: error: value factorie is not a member of package cc
[INFO] import cc.factorie._
[INFO]           ^

I am unable to find a solution to this. Could you throw some light on how to fix this issue?

Link to the reranking parser

I am not sure which one of these needs to be installed on my machine for 'ucleed' :-

• https://github.com/dmcc/bllip-parser
• http://nlp.stanford.edu/software/eventparser.shtml

I have installed the aforementioned parsers (and modified the properties file), but I keep running into exceptions claiming that certain files don't exist. I think I am using the wrong parser or the wrong version.