This repository provides code to replicate the experiments from:

• Ander Barrena, Aitor Soroa and Eneko Agirre. Learning text representations for 500K classification tasks on Named Entity Disambiguation. In the SIGNLL Conference on Computational Natural Language Learning CONLL 2018. https://aclweb.org/anthology/K18-1017

The initial prototype of the word expert model was implemented during november of the 2016, close to the release of the single "Moth into Flame" by Metallica. That's why the Word Expert models are named as m@ths (moths).

Download the data from:

http://ixa2.si.ehu.es/500kNED-download/source.tar.gz Unzip the data in the main folder. Contains dictionary, vector and cluster files.

http://ixa2.si.ehu.es/500kNED-download/entities.tar.gz Unzip the data in ssv/train/ folder. Contains all the preprocessed contexts for all the entities in Wikipedia (06Nov2014 dump)

Bash Scripts

---

00-do.all.sh: runs sequentially all scripts bellow to reproduce the experiments in "Learning text representations for 500K classification tasks on Named Entity Disambiguation" paper.

---

00-do.gather.sh: the script creates the training input files for the "Word Expert Models" tested on the paper (Aidatesta, Aidatestb, Tac2010test, Tac2011test and Tac2012test datasets), around 6K mentions. If you want to create the 500K full dataset, uncomment the last lines in the script. This may take so much time, we recommend to divide the dictionary file and run many processes in parallel. Training input files are stored in ssv/train/mentions/

The trainining data is stored as:

• one file per mention in the dictionary
• in the first line, the entity priors (not used)
• the rest of the lines gather all the mention contexts
• the contexts are represented as the 100 non stop words sorrounding the mention
• the words are replaced by their corresponding index in the vectors file /source/vectors/wikipedia.d300.w2v.bz2

00-do.m@th.cBoW.sh: the script trains and tests the Continious Bag of Words Word Expert model. Only for the mentions occurring in test datasets. Models are stored in m@ths/ and debug files in ssv/train/ and ssv/test/. Output files are stored in m@th/.

00-do.m@th.sBow.sh: the script trains and tests the Sparse Bag of Words Word Expert model. Only for the mentions occurring in test datasets. Models are stored in m@ths/ and debug files in ssv/train/ and ssv/test/. Output files are stored in m@th/.

00-do.m@th.lstm.sh: the script trains and tests the LSTM Word Expert model. Only for the mentions occurring in test datasets. Models are stored in m@ths/ and debug files in ssv/train/ and ssv/test/. Output files are stored in m@th/.

00-do.m@th.lstm.SingleModel.sh: the script trains the LSTM Single model. Models are stored in m@ths/ and debug files in ssv/train/. Output files are stored in m@th/.

00-do.m@th.transferLstm.sh: the script trains and tests the transferLSTM Word Expert model. Only for the mentions occurring in test datasets. Models are stored in m@ths/ and debug files in ssv/train/ and ssv/test/. Output files are stored in m@th/.

01-do.mix.sh: the script mixes original and augmented model results of Word Experts, then evals the system output against the gold standard. Eval results are stored in results/ folder.

01-do.mix.KB.sh: the script mixes original and augmented model results of Word Experts, and removes those entities not belonging to a Knowledge Base (AidaMeans or TacKB). Then evals the system output against the gold standard. Eval results are stored in results/ folder.

Perl Scripts

m@th.[deep].pl: Perl script for preproccesing Wikipedia, it creates the training dataset used for training and testing the "Word Expert Models" (a.k.a. moths).

m@thNster.[deep].pl: Perl script for preproccesing of the data, training and testing the "Single Model" or (a.k.a. [moth]nster).

m@th.mxr.pl: Perl script to mix the results of the "Word Expert Models" (original and augmented versions)

eval.pl: Perl script to eval the system results.

Pytorch Scripts

The following pytorch scripts are used in m@th.[deep].pl and m@thNster.[deep].pl to train and test the word expert models.

Continious Bag of Words Word Expert pytorch code:

• @cBoW.py

LSTM Word Expert pytorch code:

• @lstm.py

Sparse Bag of Words Word Expert pytorch code:

• @sBoW.py

Transfer LSTM Word Expert pytorch code:

• @transferLstm.py

Single model LSTM pytorch code:

• %lstm.py

Folders

Some of the folders will be created or filled during previous script executions:

debug: Contains debugging information from scripts

ed: Contains the preprocessed input and gold standard files from test data (Aidatesta, Aidatestb, Tac2010test, Tac2011test and Tac2012test datasets).

gs: Contains the gold standard files.

m@th: Contains the output files of the system for each test dataset

m@ths: Contains the "Word Expert Models" (or moths) and the "Single Model" (or [moth]nster) see paper for details

results: Contains the evaluation results

source: Contains preprocessed Wikipedia files (download from link above)

• clusters: word to cluster files used in sparse experiments
• dict: name to entity dictionaries
• vectors: word vectors used for the continious and LSTM based experiments (word2vec)

ssv: Contains the train and test input data (in ssv format) for the classifiers and some debugging information

• train: debuggin files for training and input files in train/mentions/ folder, and train/entities/ contains entity context files, one file per entity (download from link above)
• test: debuggin and input files for test