Board Game Review NLP Pipeline: Hands-On Fine-Grained Sentiment Analysis with Gemini

This project is a complete data engineering and NLP project that demonstrates how to scrape board game reviews from BoardGameGeek, perform aspect-based sentiment analysis (ABSA) using the Google Gemini API, and convert unstructured text into structured business insights. The original project is maintained by HSTutida, sourced from GitHub (link: https://github.com/HSTutida/boardgame-nlp-pipeline), and was published on June 17, 2026.

This is a graduation thesis project for an MBA in Data Science and Analytics, aiming to build an end-to-end data pipeline for processing unstructured text. Board game reviews were chosen because BoardGameGeek is the world's largest board game community, with a vast accumulation of user review data. Traditional sentiment analysis only provides binary "positive/negative" judgments; this project uses an aspect-based sentiment analysis (ABSA) framework to identify sentiment tendencies in reviews toward specific game dimensions (e.g., rules, components, replayability), providing more granular market feedback for game designers and publishers.

An automated crawler was built using Python and Selenium WebDriver, with core features including:

• Stealthy Automation: Implements headless browsing, WebDriver camouflage, and User-Agent modification to bypass basic anti-crawling mechanisms
• Dynamic Pagination Scraping: Automatically traverses BGG pages to collect target game IDs and URLs
• Precise Forum Mining: Filters and extracts titles and full review texts from popular/pinned posts for each game
• Fault-Tolerant Data Engineering: Micro-batch processing strategy (each game written to CSV independently), network issue tolerance, and polite delay mechanisms

Deep content analysis was performed using the Google Gemini API:

• Advanced Prompt Engineering: Defines an expert computational linguist role, constraining the use of a predefined system of 6 game aspects and 3 sentiment polarity categories
• Deterministic JSON Generation: Sets a low temperature (0.1) to ensure output consistency, enforcing application/json output mode
• Robust Data Processing: Uses Pandas to load and clean CSV data, including exception handling and missing file fallback mechanisms
• API Rate Management: Implements automatic pauses and error capture to ensure reliable calls on large datasets

ABSA Analysis Dimensions:

1. Rules: Clarity, complexity, learning curve
2. Components: Accessory quality, art design, material craftsmanship
3. Replayability: Replay value, variability
4. Gameplay: Core mechanism fun level, smoothness
5. Balance: Fairness, strategic depth
6. Cost-effectiveness: Price-to-content ratio Each dimension corresponds to 3 sentiment polarities: positive, neutral, and negative.

Tech Stack:

• Data Scraping Layer: Python 3.x, Selenium WebDriver, webdriver_manager, csv module
• Analysis Processing Layer: Google Gemini API (gemini-2.5-pro), Pandas, JSON, ABSA

Insights for Data Engineers:

1. Dynamic JavaScript Rendering Handling: Requires explicit waiting and DOM manipulation skills
2. Anti-Crawling Countermeasures: Simulate real user behavior, balance data acquisition and website friendliness
3. Structured Data Storage: Conversion from unstructured text to CSV/JSON is the first critical step in a machine learning pipeline

Insights for NLP Practitioners:

1. Power of Zero-Shot Classification: Achieve fine-grained classification without large amounts of labeled data
2. Importance of Structured Output: Guide generative models to produce controllable results through strict constraints
3. Integration of Domain Knowledge: Predefined aspect classification systems reflect the importance of domain expert knowledge for NLP tasks

This pipeline architecture can be migrated to other domains:

• E-commerce Review Analysis: Extract user feedback on different product features
• Hotel Review Processing: Analyze satisfaction across dimensions like location, service, and facilities
• App Store Review Analysis: Identify user pain points in aspects like functionality, UI, and performance
• Social Media Monitoring: Track brand sentiment trends across different topics

boardgame-nlp-pipeline is a complete data science and NLP engineering example, from web scraping to LLM analysis, demonstrating the full-link development capabilities of modern AI applications. It is not only an academic project but also provides directly referenceable technical solutions for text analysis tasks in industry. By combining traditional data engineering methods with cutting-edge LLM technology, the project successfully converts massive unstructured reviews into actionable business insights.