Optimizing Text Classification in Construction Data
summary
Over a million tender paragraphs with ~22,000 potential products, ~100 categories and suppliers, and ~300 positions per tender require a month of manual tagging; a Random Forest offer filter plus customized fastText embeddings then runs region-specific fastText classifiers (six regions, 19 models) and label-merging to merge rare suppliers, lifting supplier/category accuracy to ~80% and F1-micro high; a fine-tuned BERT bi-encoder top-30 70% product retrieval ranks candidates for human review, with tuned probability thresholds to limit error cascades.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:27Presenters
- 3:35Construction tender challenges
- 8:19Model architecture
- 12:18Key challenges
- 14:32Achievements
- 17:28FastText performance
- 20:20BERT encoder
- 23:56Model advantages
- 26:29Acknowledgments
- 27:17Q&A
related talks
38mE-Commerce Product Categorization at Scale
45mTransformer based clustering: Identifying product clusters f...
30mA study in classification
44mExtracting Structured Data from Legal Documents
31mWhen humans teach machines
22mImplementing a Text Classifier based on Lucene & LIBSVM
22mHarnessing image and text data for product matching
38mAutomatically categorizing news articles on nrk.no