Natural Language Processing Project - Sentiment Analysis on Amazon Reviews

Introduction

• Instructor: Rob
• Content Overview: Sentiment analysis using VADER and RoBERTa models on Amazon reviews.
• Additional resources: Everything shared in a Kaggle notebook (link in description).

Libraries and Tools Used

• Python Libraries:
  • pandas: For data manipulation.
  • numpy: For numerical operations.
  • matplotlib, seaborn: For plotting and visualization.
  • nltk: Natural language toolkit for text processing.
  • transformers (Hugging Face): For using pre-trained deep learning models.
  • tqdm: For progress bar visualization.

Data Overview

• Dataset: Amazon fine food reviews.
• Content: Reviews and ratings (1-5 stars) in CSV format.
• Data Size: Approximately half a million reviews, downsampled to 500 for this project.

Data Import and Basic Analysis

• Data Reading:
  • Example: pd.read_csv('reviews.csv')
  • Preview data using df.head().
• Exploratory Data Analysis (EDA):
  • Score counts visualization using a bar plot.
  • Findings: Most reviews are 5 stars, followed by 1 star.

Basic NLTK Operations

• Tokenization:
  • Example: nltk.word_tokenize(text), splits text into tokens.
• Part of Speech Tagging:
  • Example: nltk.pos_tag(tokens), assigns POS tags to tokens.
• Chunking:
  • Example: nltk.ne_chunk(tagged), groups tokens into chunks.

Sentiment Analysis with VADER

• VADER: Valence Aware Dictionary and Sentiment Reasoner.
  • Features: Analyzes text based on a lexicon of positive, negative, and neutral words.
  • Process: Removes stop words, computes sentiment scores.
  • Implementation:
    • Create analyzer: sia = SentimentIntensityAnalyzer().
    • Example: sia.polarity_scores(text).
• Analysis:
  • Running sentiment analysis on each review.
  • Store results in a pandas DataFrame.
  • Findings: Comparison of sentiment scores with star ratings. Positive correlation between high star ratings and positive sentiment.

Sentiment Analysis with RoBERTa

• RoBERTa Model: A transformer-based model for sentiment analysis.
  • From Hugging Face’s library.
  • Pretrained on large datasets (e.g., Twitter data).
• Steps:
  • Import tokenizer and model: AutoTokenizer, AutoModelForSequenceClassification.
  • Tokenize text and run the model to get sentiment scores.
  • Normalize output using softmax.
  • Example function for sentiment scoring: polarity_scores_roberta(text).
• Analysis:
  • Comparison with VADER results.
  • Findings: RoBERTa shows more distinct and confident sentiment separation.

Comparative Analysis

• Pair Plot: Visualize the comparison between VADER and RoBERTa sentiment scores.
• Key Observations:
  • Positive correlation with star ratings for both models.
  • RoBERTa shows clearer distinction in sentiment compared to VADER, especially for 1-star and 5-star reviews.

Error Analysis

• Misclassifications: Examples where sentiment analysis results diverge from expected sentiment based on star ratings.
• Insights: Context and nuance in language can cause discrepancies, highlighting the limitations of simpler models like VADER.

Hugging Face Pipelines

• Usage: Extremely simple setup for sentiment analysis.
  • Example: pipeline('sentiment-analysis').
  • Automatically downloads and configures models.
• Performance: Quick and easy for out-of-the-box sentiment analysis with reasonable accuracy.

Conclusion

• Summary: Walkthrough of traditional and modern sentiment analysis approaches, comparison of different models, and practical implementation using Python libraries.
• Call to Action: Subscribe for more content and follow live coding sessions on Twitch.