Applying NLP for topic modelling & customer sentiment analysis

The aim of this project was to process and analyse the customer reviews of a Multinational Gym provider using NLP and LLM-based pipelines to pinpoint key areas for improvements, propose actionable solutions for addressing the prevalent problems and improve customer satisfaction.

🔹 Data

Over 20,000 customer reviews collected annually from Google Reviews and Trustpilot.

🔹 Methodology

• Performed the necessary text preprocessing and EDA, identified and visualised the most frequently used words in the reviews.

• Applied BERTopic for topic modelling, keeping track of gym locations, to locate common topics and words in the negative reviews.

• Aggregated the datasets on locations and extracted the top 30 venues with the most negative reviews.

• Used the built-in visualisation functions in BERTopic to cluster and visually represent the topics and words in these reviews, thereby identifying specific themes.

• Performed emotion analysis to determine the emotions associated with the reviews.

• Filtered the angry reviews and applied BERTopic to discover prevalent topics and words in these negative reviews.

• Summarised the outputs into actionable insights.

• Leveraged the multi-purpose capability of Microsoft's Phi-4-mini-instruct model https://huggingface.co/microsoft/Phi-4-mini-instruct to compare and interpret the identified dominant topics.

• Utilised the created pipeline with additional prompting to generate additional improvement suggestions, based on the main topics identified from the negative reviews.

• Conducted a comparative analysis between BERTopic and Gensim’s LDA model to validate topic coherence and ensure methodological robustness.

🔹 Tools

• Python, NumPy, Pandas
• Matplotlib, Seaborn, Scikit-learn
• NLTK, WordCloud
• BERTopic, Bert-based-uncased-emotion, Phi-4
• PyTorch & HuggingFace tools: transformers, datasets, evaluate
• Gensim & pyLDavis

🔹 Results and Findings

• Key negative themes: poor equipment maintenance, hygiene, unhelpful staff, and contract cancellation issues making up over 65% of negative feedback.

• Emotion analysis revealed anger and frustration dominating negative reviews, particularly in large urban gyms.

• Actionable improvements included:

  • Enhance staff and customer support training
  • Devise flexible membership policies
  • Improve equipment upgrade schedules
  • Redesign schedule of offered classes for improved engagement and attendance

---