jokra1 / ltp_final

LTP_Final

Introduction

For the course "Language technology project" at the RUG (academic year 20/21) we compared recurrent neural networks and transformers in their ability to correctly classify russian troll tweets into 5 distinct categories, which were originally described in the paper by Linvill & Warren. The data-set was taken from Kaggle. Our results are described in a final report.

Requirements

The python libraries outlined in the requirements_conda.txt document are the minimum requirements to replicate our findings. We relied on R version 4.0.5. Only base R is required to generate the data-files. However, recreating the word clouds requires the "wordcloud" package (2.6), the "tm" package (0.7-8), and the "RColorBrewer" package (1.1-2) also listed in the tutorial which includes the original code used to generate the clouds here as well.

Setup

To replicate our results first download the original data-set from Kaggle and place the individual IRA* .csv files in the data-folder. Then:

• Use the R-script in the helpers folder to generate the necessary train, dev, and test sets.
• Subsequently, use the train_* files to train and fine-tune the individual models discussed in the report. Start with the train_embeddings file.
• Finally, use the test_* notebooks to evaluate the models on the test-set.
• The embedding_analysis notebook in Results/embedding_results can also be used to inspect the learned embeddings.