A Comprehensive Comparison of Word Embeddings in Event & Entity Coreference Resolution.

Introduction

This repo contains experimental code derived from :

"Revisiting Joint Modeling of Cross-document Entity and Event Coreference Resolution"
Shany Barhom, Vered Shwartz, Alon Eirew, Michael Bugert, Nils Reimers and Ido Dagan. ACL 2019.

Please go to the original repo for more information on the original code.

Embeddings used

Character : https://github.com/minimaxir/char-embeddings
Word2vec : https://github.com/mmihaltz/word2vec-GoogleNews-vectors
FastText : https://pypi.org/project/FastText/
GloVe : https://nlp.stanford.edu/projects/glove/
GPT-2 : https://huggingface.co/transformers/model_doc/gpt2.html
BERT : https://pypi.org/project/bert-embedding/
Elmo : https://allennlp.org/elmo

Branches

Original model

Original : Code optimized compared to the the original repo leading to faster training time
Uses GloVe, Elmo, and a fine tuned character embedding.

Ablative models

NoStatic : Removed GloVe embedding from orignal model
NoContext : Removed Elmo embedding from orignal model
NoChar : Removed character embedding from orignal model
noctx-static : Removed Elmo and GloVe embedding from orignal model
noctx-static-char : Removed all embedding from orignal model

Comparative models

GPT-2 : Replace Elmo with GPT-2
BERT : Replace Elmo with BERT
FastText : Replace GloVe with FastText
Word2Vec : Replace GloVe with Word2Vec

Comparative ablative models

Onlybert : Removed GloVe and character embedding from orignal model + Replace Elmo with BERT
OnlyGPT : Removed GloVe and character embedding from orignal model + Replace Elmo with GPT-2
OnlyELMO : Removed GloVe and character embedding from orignal model
onlyfasttext : Removed Elmo and character embedding from orignal model + Replace GloVe with FastText
onlyword2vec : Removed Elmo and character embedding from orignal model + Replace GloVe with Word2Vec
onlyglove : Removed Elmo and character embedding from orignal model

Contact info

Contact Judicaël POUMAY at [email protected] for questions about this repository.