Aspect-sentiment-classification

By Unaiza Faiz(ufaiz2@uic.edu), Vijaya Nandhini Sivaswamy (vsivas2@uic.edu)

Our goal was to train and build a model to predict positive, negative and neutral polarity labels of an unseen test data based on a given aspect term of an opinionated sentence. Sentiment Analysis refers to the computational process of identifying the emotions or opinions based on a given text.

This project aims at building a supervised learning classification model identifying the polarity of an aspect term provided for a given statement. We used context window to extract words within the window range of the aspect term i.e 4 words to the left and right of the aspect term. Vectorization is performed using TF-IDF scheme and different models are evaluated based on 10-fold cross validation.

Data pre-processing techniques:

• Replacing punctuations and special characters
• Stop word removal
• Lemmatization
• Tokenization

Feature Engineering:

• Using windowing technique
• TF-IDF

Models Attempted:

We implemented 8 models using scikit-learn to classify our training set. The models attempted were:

• LinearSVC.
• Naive Bayes Classifier.
• Multinomial Naive Bayes Classifier
• MLP Classifier
• SGD Classifier
• Adaboost
• K-neighbours classifier.
• Logistic Regression

Results

LinearSVC was considered as the best model with respect to overall accuracy, precision, recall and F-score for the positive and negative classes, and hence used as the classifier on the held-out set.

Steps to run the code:

1. Download and unzip the folder
2. Edit the variable file_no in predicttest.py as 1 or 2 (depending on test set) and save
3. Run the command - python predicttest.py
4. All the employed classifiers are commented in classmode.py