This repository contains the preliminary code for reproducing the results reported in the paper "[GraphNER: Named Entity Recognition as Graph Classification](https://openreview.net/forum?id=vfpW-kRvLgu)" (currently under review).

The code is organized as notebooks, to be used as follows:

* `final_generate_gazetteers.ipynb`: to generate the gazeteers from Wikidata (by specifying a list of QIDs correspoding to the classes)

* `graph_embeddings_generation.ipynb`: to generate node embeddings using [GEM library](https://github.com/palash1992/GEM) algorithms (e.g. node2ve, SDNE..)

* `nodes_classifier.ipynb`: to train a model for the node embeddings

* `v2.0/autoencoder_embeddings.ipynb`: to generate auto-encoder embeddings from the binary graph representations.

* `v2.0/autoencoder_nn_classification.ipynb`: to train a model for the auto-encoder embeddings

* `v2.0/GCN-node-simple-features.ipynb`: to train a GCN on the CoNLL-2003 task

The code will be streamlined into stand-alone configurable scripts and fully documented soon.

* Python 3.8

* PyTorch 1.7

* [GEM](https://github.com/palash1992/GEM)

* [PyTorch Geometric](https://pytorch-geometric.readthedocs.io/en/latest/)

* [SPARQLWrapper](https://github.com/RDFLib/sparqlwrapper)

* tqdm

* Numpy

* Pandas

:warning: This code runs on a CUDA11.0-enabled GPU, please install the compatible version of the modules for your hardware.

Method | Accuracy | Micro-F1 | Macro-F1

-----------------|----------|----------|---------

Binary | 91.0 | 90.7 | 77.9

Binary+ | 94.4 | 94.2 | 81.9

Binary++ | 94.3 | 93.8 | 82.3

Auto-encoder-100 | 87.2 | 86.7 | 57.6

Auto-encoder-500 | 90.4 | 89.9 | 68.3

Auto-encoder-2000| 91.8 | 91.5 | 71.7

Node2Vec-300 | 93.8 | 94.1 | 82.0

Node2Vec-500 | 93.8 | 94.1 | 82.5

Node2Vec-1000 | 93.8 | 94.1 | 82.1

GCN | 96.1 | 96.1 | 86.3

**GCN+** | **96.5** | **96.5** | **88.8**