🕵️ Introduction

We all refer to reviews and ratings for movies & shows, either on IMDB or through our friends before we binge them. In this assignment, you’re tasked with identifying the rating of a review using sentiment analysis.

The training dataset consists of {sentence id, review, rating}. Given a review, your neural network should be able to assign a rating score between 1 to 5 - low to high, respectively.

Do note:

• Assignment1 (First submission)
  • Perform sentiment analysis using a neural network with just one input and output layer (with softmax as activation function).
  • You’re not allowed to have hidden layers.
  • You need fill the pre-written functions as provided in the starter skeleton code.
• Assignment1 (Second submission)
  • Introduce hidden layers
  • Use sigmoid as well as RELU for the hidden neurons and study the difference    
  • Develop a good GUI for input mechanism
  • Use word2vec, GloVe, FastText for word embedding and compare results
  • Solve the data imbalance problem (class V has 65% of the data)
  • Achieve at least 85% accuracy (currently for many groups it is in the range of 40s)
  • You can also use RNN in this assignment, but FFNN results (compulsory) needs to be shown

💾 Dataset

The dataset contains training data, test data, and a skeleton code. You can find it under Resources. Use the auto.py to execute your code.

📁 Files

• train.csv - (50000 samples) This file contains training data with sentence ids, reviews & corresponding ratings (1 to 5).
• test.csv - (10000 samples) This file contains test data with sentence ids & reviews.
• code.py- This file contains sample skeleton code with pre-written function & their descriptions.
• auto.py - This file will be used to execute the code. No edits to this file is permitted.

🚀 Submission

Submissions will be in the form of csv file containing your predictions for test data in the form of {sentence id, reviews, rating} like the following example:

🖊 Evaluation Criteria

The evaluation will be done according to the following criterion:

• F1 score (Harmonic Mean of precision & recall)

📱 Contact

• DL for NLP TAs