Optimizing Text Classification in Construction Data Using FastText and BERT Models

Explore a detailed conference talk from Data Science Conference Europe 2023 that tackles the complex challenge of text classification in construction industry tenders. Learn how the speakers addressed the massive task of classifying over a million tender-related paragraphs across 22,000 potential products and 100 categories and suppliers. Discover their innovative approach using a fasttext model with custom embeddings for predicting product and supplier categories, including their strategy of merging less common categories to reduce errors. Understand how they implemented probability thresholds based on uniform distribution to enhance model performance, achieving impressive accuracy and F1 Micro scores. Gain insights into their experimental use of BERT and other Large Language Models for product prediction, which achieved 70% accuracy for top 30 product matches. Examine their future improvement strategies, including deeper domain data analysis and potential integration of rule-based approaches, providing practical solutions for business applications in construction data classification.