Sentiment Analysis

NLP task : postive or negative Review Sentiment Analyzer using IMDB dataset in Python .

Overview

1. Dataset

   The 100K reviews are divided into three datasets: 25K labeled training instances, 25K labeled test instances and 50K unlabeled training instances. Each review has one label representing the sentiment of it: Positive or Negative. These labels are balanced in both the training and the test set.

2. Text Pre-processing

   Text preprocessing was done in order to remove unecessary words or turn verbs and adjectives into a unified word to make the model less confused . Text PreProcessing included

   • spell checker
   • stop words removal (is,are,the,this ....etc)
   • Lemmatization

3. Vector Spaces and Data Matrix

   This step aimed to convert the text of the review after pre-processing into a vector form for further steps.Different alternatives for text representation inclued:

   • n-grams.
   • Count, TF-IDF.
   • DocToVec

4. Supervised Learning Step

   We tried multiple of classification models. We tuned their parameters. We tuned ​these models parameters on 10% subset of the training set. Model choices are:

   • Decision Trees
   • Random Forests
   • SVM Non-Linear
   • Adaboost
   • Logistic Regression
   • Bagging

5. Results

   With All mentioned Preprocessing:

   • Logistic regression :87.5%,
   • Adaboost 80%
   • Decision trees was 52% before turning and reached 70% after

   With preprocessing without lemmatization:

   • Random forests was 54% before tuning and reached 85.5% after
   • Adaboost 79.852%

   Without preprocessing at all expept html removal:

   • Naive Bayes 85.4%
   • Logistic Regression 88.6% with c set to 1000
   • Adaboost 80.2%
   • Random Forests was 83.6% before tuning and 86.6% after

   With stop words removal and html only:

   • Logistic Regression 88.3%

   Without any preprocessiong:

   • Logistic Refression 88.5%