Human Language Technologies [2019-2020] project implementation: Deep Learning Topic Based Sentiment Analysis

This model can perform Aspect Category Detection and Aspect Category Polarity.

How to tun the code

To install the requirements:

pip install -r requirements.txt

From the train files, you can say to the model which task it needs to learn (modifying TASK with acd or acd), which embedding to use (modifying EMB with w2v or alberto), if you want to perform validation (setting VALIDATION to true, this will save the grid search results inside logs/EMB/TASK/hparam_tuning that can be visualize with Tensorboard) and which is the hidden dimensions that you want to use (these are effective only if VALIDATION is false). To test a model with the Absita official evaluation script, you need to run

python test.py

Settings the right values of the hidden dimensions, that can be seen from the checkpoint directory (*checkpoints/EMB/TASK/model_rnn_cells_final_cells). To visualize the training history obtained, you just need to do

tensorboard --logdir logs/EMB/TASK/final_training

this will create the temporary files data/emb_tas_results.csv and call the evaluation Absita script with the right parameters.