Identifying Aspect Categories and Their Sentiments in Scientific Peer Reviews

This repository contains dataset and code of Identifying Aspect Categories and Their Sentiments in Scientific Peer Reviews

Download the project source folder

Can download the source code using git clone or the zip file.

Dataset:

The dataset contains of all the annotated review sentences, SciBERT embeddings, BERT_BASE embeddings, POS one-hot-encoded data, reviews(X) and annotations and it's polarity(Y). We have also stored the embeddings to expedite the training process.

NOTE: Since the entire dataset sums upto 1.92 GB, we have uploaded it HERE

Notebooks:

1) Our Proposed Model:

This notebook consists our proposed multi-task model for aspect category classification and sentiment detection.

2) POS-SciBERT Model:

This notebook consists another competitve variant of our proposed model, wherein we feed POS one-hot-encoded and SciBERT embeddings parallelly.

3) Ablation Variants Models:

This notebook consists 3 variants viz, WithoutBiLSTM, WithoutAttention, WithoutBoth. In the notebook we have 3 different cells in section 5) Define Ablation Models for initialising each variant. Uncomment the one that you want to reproduce and let the other two be commented.

4) Attention Analysis:

This notebook loads our proposed model and produces a sentence-wise heatmap distribution for aspect categories for 2 selected Reviews which is discussed in depth in our paper.

IMPORTANT POINTS BEFORE RUNNING:

a) Change the URL PATH accordingly before loading the dataset(pickle files)
b) The 1st tab in the notebook consists all the additional dependencies required and will be downloaded on running the cell
c) For SAVE_PATH set the URL path where you want to save the trained model

Once all the setup is complete then execute run all.

Libraries & Dependencies used:

TensorFlow

Keras

Hugging Face

Matplotlib, numpy, pandas, pickle