Beyond Keywords: BERTopic and Mistral-8B Integration for Smartphone User Perception Analysis

Abstract

Online forums provide rich but unstructured feedback from smartphone users. Traditional analysis methods struggle to extract meaningful themes from such content. This study introduces a hybrid analytical pipeline that combines BERTopic for clustering and Mistral-8B for semantic refinement. The system analyzes discussions about the iPhone 15 and Galaxy S23 series from Reddit and Whirlpool forums, delivering clear topic labels and sentiment summaries to enhance user perception analysis.

Methodology

The pipeline consists of five main steps:

1. Data collection and preprocessing
2. Sentence embedding using all-MiniLM-L6-v2
3. Dimensionality reduction with UMAP
4. Clustering using HDBSCAN
5. Topic labeling and summarization using Mistral-8B and sentiment analysis via VADER

This method allows for interpretable and scalable insight extraction from large-scale user discussions.