Threads Reviews Text Mining

Overview

This project analyzes reviews from the Threads platform to extract insights using various text mining and NLP techniques. The dataset comprises user reviews along with their ratings, offering an opportunity to perform sentiment analysis, clustering, and advanced classification tasks.

The project is structured into three questions that explore different approaches to understanding and modeling textual data:

1. Traditional Machine Learning Models with N-grams
2. Clustering Techniques for Unsupervised Learning
3. Fine-Tuning DistilBERT for Sentiment Classification

Each question emphasizes unique aspects of natural language processing (NLP) and leverages different algorithms to solve distinct problems, including sentiment classification, clustering, and interpretability.

The repository is available at Threads Reviews Text Mining in Spanish.

Dataset

The dataset contains user reviews from Threads, focusing on the review text and associated ratings.

• Text Field: Represents the review content.
• Rating Field: A numeric score indicating user satisfaction.

To standardize tasks across questions:

• Reviews are categorized into three sentiment classes:
  • Positive: Ratings above 4.
  • Neutral: Ratings between 3 and 4.
  • Negative: Ratings of 2 or below.

This dataset is ideal for text mining as it combines structured (ratings) and unstructured (text) data. Preprocessing steps include cleaning text, creating n-grams, tokenizing for transformers, and balancing classes for fair evaluation.

Question 1: Traditional Machine Learning with N-grams

In this phase, the focus is on extracting insights from reviews using traditional NLP techniques:

1. N-grams: Reviews are represented as unigrams, bigrams, and trigrams using CountVectorizer.
2. Model Training: Four classifiers were applied to these n-grams:
   • Decision Tree Classifier
   • Multinomial Naive Bayes
   • Logistic Regression
   • Support Vector Machines (SVM)
3. Metrics: Evaluation metrics included accuracy, F1-score, precision, and recall to compare model performance across n-gram variations.

Outcome:

• Logistic Regression and SVM showed the best results.
• N-grams provided valuable insights into the textual structure of reviews.

Question 2: Clustering Techniques

Here, the goal was to explore unsupervised learning for grouping reviews:

1. Dimensionality Reduction:
   • PCA (Principal Component Analysis) was applied to reduce the dataset’s dimensionality for visualization.
2. Clustering Algorithms:
   • K-Means: Optimal clusters were identified using the elbow/knee method.
   • Agglomerative Clustering: Hierarchical clustering was visualized through dendrograms.
   • DBSCAN: Density-based clustering grouped similar reviews.
3. Validation: Silhouette scores were used to assess clustering quality.

Outcome:

• Clusters revealed patterns in review sentiments, highlighting differences in user preferences and feedback themes.

Question 3: Fine-Tuning DistilBERT

This phase tackled advanced sentiment classification using a transformer-based approach:

1. Dataset Preparation:
   • Reviews were tokenized and converted into a Hugging Face dataset format.
   • Sentiments were labeled as Positive, Neutral, or Negative.
2. Fine-Tuning:
   • DistilBERT, a lightweight transformer model, was fine-tuned for this task.
   • Training leveraged weighted F1-scores to address class imbalance.
3. Interpretability:
   • The Transformers Interpret library visualized token contributions to model predictions, ensuring transparency and trustworthiness.

Outcome:

• DistilBERT achieved high accuracy and interpretability, outperforming traditional models.
• Visualizations provided insights into how the model processed user reviews.