LlamaFur: Learning Latent Category Matrix to Find Unexpected Relations

In this repository, you will find dataset and code to be able to evaluate algorithms which aim to discover unexpected links, and to repeat LlamaFur experiments.

The human-evaluated dataset we used to evaluate the unexpectedness of links in Wikipedia is human-evaluation.tsv. In this file, the first column is the title of the Wikipedia page that contains the link, the second column is the title of the target page, while the third column is TE, E, U or TU depending on whether the link is Totally Expected, Expected, Unexpected or Totally Unexpected.

The code we used is available in the directory llamafur/java.

How to replicate experiments

Set up the environment

In order to compile LlamaFur code, you'll need Java 8, Ant and Ivy. To install those inside a clean Vagrant box with ubuntu/trusty64, you should use these lines:

sudo apt-get --yes update
sudo apt-get install -y software-properties-common python-software-properties
echo oracle-java8-installer shared/accepted-oracle-license-v1-1 select true | sudo /usr/bin/debconf-set-selections
sudo add-apt-repository ppa:webupd8team/java -y
sudo apt-get update
sudo apt-get --yes install oracle-java8-installer
sudo apt-get --yes install oracle-java8-set-default
sudo apt-get --yes install ant ivy
sudo ln -s -T /usr/share/java/ivy.jar /usr/share/ant/lib/ivy.jar

Compile LlamaFur code

If the environment is set up properly, you should install git and download this repo with

sudo apt-get install git
git clone https://github.com/corradomonti/llamafur.git

and then go to the directory llamafur/java. There, run:

ant ivy-setupjars to download dependencies
ant to compile
. setcp.sh to include the produced Jar unexpectedness-1.0.jar inside the Java classpath.

Please check that a command like java efen.scorers.llamafur.LatentMatrixEstimator --help correctly shows help.

Download the wikipedia dump

You should be able download the Wikipedia articles dump we used, named enwiki-20140203-pages-articles.xml.bz2 from AcademicTorrents. If this torrent is not working, please email us at {corrado.monti, paolo.boldi} @unimi.it. Please unzip this dump, and put it in a directory named llamafur/dump.

Run experiments

Then, cd to the llamafur/ directory where you can proceed to run the script with sh run-experiments.sh. Parsing the Wikipedia snapshot will take many hours (at least 10 usually), while the LlamaFur experiments won't take as long. See the comment in the script file (llamafur/run-experiments.sh) for more info.