Graph Convolutional Networks - Application

This is a TensorFlow implementation of Graph Convolutional Networks for SemEval 2019 Task 6: OffensEval: Identifying and Categorizing Offensive Language in Social Media and in specific Sub-task A - Offensive language identification.

Requirements

• tensorflow (>0.12)
• networkx

Run

• Use (Clean_Tweets.ipynb)[data/Clean_Tweets.ipynb] to clean your .csv or .tsv files and generate corresponding .txt files. Place the generated files appropritely in corpus or R8 sub-folder under the data folder.

• Use (pre-processing.ipynb)[pre-processing.ipynb] to pre-processing the data and generate required files.

• Use (build_graph.ipynb)[build_graph.ipynb] to generate the graph.

• Finally run:

python3 train.py R8

Data

I have used a part of the OLID dataset (2499 tweets) (accepted in NAACL in 2019) for this experiment. I have converted the tsv file to Graph using (build_graph.ipynb)[build_graph.ipynb]

Papers blogs referred

I have used information from the following papers:

• gcn: Graph convolutional network (Thomas N. Kipf, Max Welling, Semi-Supervised Classification with Graph Convolutional Networks, 2016) (ICLR 2017)
• gcn_cheby: Chebyshev polynomial version of graph convolutional network as described in (Michaël Defferrard, Xavier Bresson, Pierre Vandergheynst, Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering, NIPS 2016)
• Thomas Kipf, Graph Convolutional Networks (2016)

Graph classification

Our framework also supports batch-wise classification of multiple graph instances (of potentially different size) with an adjacency matrix each. It is best to concatenate respective feature matrices and build a (sparse) block-diagonal matrix where each block corresponds to the adjacency matrix of one graph instance. For pooling (in case of graph-level outputs as opposed to node-level outputs) it is best to specify a simple pooling matrix that collects features from their respective graph instances, as illustrated below:

Output

Here is an instance of output:

I had forcibly labelled all the test set instances as NOT offensive and the model was allowed to predict. Accordingly the confusion matrix was generated.

Note

This is my first GCN Experiment. 🎉 Any constructive criticism is highly appreciated.