Skip to content
This repository has been archived by the owner on Dec 20, 2023. It is now read-only.

Latest commit

 

History

History
113 lines (72 loc) · 7.35 KB

README.md

File metadata and controls

113 lines (72 loc) · 7.35 KB
Raw

Sentiment Analysis with Naive Bayes Classifier

Coding up a Naive Bayes Sentiment Classifier from scratch using python

How to get working:

  • install conda and then use:

      conda env create -f condaEnv.yml

  • or use the requirements.txt to get the necessary components

ISyE 6420

(Dated: November 30, 2023)

Abstract

This project focuses on sentiment analysis using a Naive Bayes classifier applied to the IMDB Dataset of 50K Movie Reviews. The goal is to predict the sentiment (positive or negative) based on the text content of the reviews. The Naive Bayes classifier is implemented using natural language processing techniques, including text preprocessing and Laplace smoothing.

Introduction

Sentiment analysis is a crucial aspect of natural language processing, with applications ranging from product reviews to social media comments. The IMDB Dataset provides an extensive collection of movie reviews labeled with sentiments, making it an ideal dataset for training a sentiment analysis model.

Methodology

The project employs a Naive Bayes classifier, a probabilistic machine learning model known for its simplicity and effectiveness in text classification tasks. The classifier is trained on a preprocessed subset of the IMDB Dataset, with a focus on feature extraction using Count Vectorization and Laplace smoothing to handle unseen words.

Data

The IMDB Dataset of 50K Movie Reviews is utilized for this project. The dataset contains reviews labeled as positive or negative sentiments.

Analysis/Implementation

Text Preprocessing:

  • HTML tags, URLs, and non-alphanumeric characters are removed from the reviews.
  • Text is converted to lowercase.
  • Lemmatization is applied to find the root form of words.
  • Stop words are removed (common words that don't add to the meaning i.e: for, an, nor, but, or, yet, so…)

Naive Bayes Classifier:

  • The foundation of the Naive Bayes classifier lies in Bayes' Theorem, a fundamental concept in probability theory.

  • Bayes' Theorem is expressed as follows:

       P(A|B) = \frac{P(B|A) \cdot P(A)}{P(B)}

  • Let A be the sentiment label (positive or negative) and B be the words present in a movie review. The goal is to calculate the probability of a particular sentiment given the observed words in a review; ( P(A|B) ).

  • The "Naive" assumption the Naive Bayes algorithm makes is that the features (words in this case) are conditionally independent given the class label. This allows for a more tractable calculation, greatly simplifying the process.

  • The classifier computes the log likelihood of each label given the words in a review and combines them with the log prior probabilities.

  • Laplace smoothing is applied to handle words not present in the training set, preventing zero probabilities.

  • The fit method calculates the log prior probability for each label based on the training data.

  • The fit_word_counts method extracts word counts for each label using Count Vectorization.

Prediction:

  • The predict method utilizes the trained classifier to predict the sentiment of input texts.

  • Predictions are made based on the label with the highest log probability.

Evaluation:

  • The model is evaluated on a test set, and accuracy is calculated using scikit- learn's accuracy_score.

Discussion

  • The Naive Bayes classifier is grounded in Bayes' Theorem, leveraging conditional probabilities to make predictions about sentiment based on observed words in a movie review. The model's assumptions, particularly the independence assumption, simplify the computation and allow for efficient training and prediction.

  • In the context of this project, Bayes' Theorem is employed to calculate the probability of a particular sentiment given the words present in a review. The log likelihoods of each sentiment class are combined with the log prior probabilities to make a final prediction. Laplace smoothing is applied to address the issue of zero probabilities for words not present in the training set.

  • While Naive Bayes classifiers are known for their simplicity, they often perform surprisingly well in text classification tasks. The approach adopted in this project showcases the application of Bayes' Theorem in a practical and effective manner, providing a solid foundation for sentiment analysis on textual data.

Results

Training Set Accuracy:

  • The Naive Bayes classifier achieved an accuracy of 85.24% on the split IMDB data. This indicates that the model performs well on the seen data, effectively capturing the sentiment patterns present in the IMDB Movie Reviews dataset.

Robustness of the Model:

  • The model exhibits robust performance, particularly in cases where the language used is straightforward and unambiguous. It performs best when the review is conveyed in a straightforward manner. However, the model might encounter challenges when faced with nuanced language or instances where the sentiment is expressed in a counterintuitive manner, such as the phrase "the movie is so bad it is good." In such cases, where the sentiment is inherently contradictory, the model may struggle to accurately predict the sentiment.

User Input Section:

  • The user input section provides an interactive way to test the model in real-time. The output demonstrates the model's capability to predict sentiments for various input movie reviews.

Positive Predictions:

  • Example: "Because the film is about Kishi Bashi and his inspiration for musical composition, the first hurdle an audience must clear is totally subjective: Do you like Kishi Bashi's music? I did, and this song film was thoughtful and inspiring."

Negative Predictions:

  • Example: "the movie is great if you want to be bored out of your mind and wishing you were elsewhere."

Challenges:

  • The model struggles with reviews containing contradictory sentiments or complex language. Example: "terribly fantastic movie" may confuse the model due to the conflicting terms.

Overall Observations:

  • The model tends to perform well on reviews with clear sentiments, making it suitable for applications where straightforward opinions are prevalent.

Conclusion

  • The model works quite well and achieves a high accuracy on the training set. However, caution is advised when dealing with reviews containing contradictory language or nuanced expressions. This could likely be improved by providing more training examples of similar language. The interactive user input section provides insights into the model's real-world application, showcasing its strengths and potential limitations.

  • Further exploration and refinement could enhance the model's ability to handle diverse linguistic nuances and improve overall accuracy.

References

  • IMDB Dataset of 50K Movie Reviews. Link
  • Scikit-learn Documentation. Link

Appendices

Code Output:

Accuracy of prediction on the test set: 0.8524

Enter a movie review (type 'exit' to stop): Because the film is about Kishi Bashi and his inspiration for musical composition, the first hurdle an audience must clear is totally subjective: Do you like Kishi Bashi's music? I did and this song film was thoughtful and inspiring
Predicted sentiment: positive

... (other user input and predictions)

Enter a movie review (type 'exit' to stop): this moive is so bad it's good
Predicted sentiment: negative