BI Toolkit Powered by Natural Language Processing
summary
Prototype wraps PDF-to-text and machine translation into English, uses YAKE keyword extraction to compress documents, then two Random Forest classifiers (one for stakeholders, one for domains) with probability vectors and SMOTE balancing, plus word2vec embeddings to infer ~100 latent impacts via semantic expansion and cosine/Jaccard similarity. Output is a georeferenced heatmap BI toolkit (React frontend, Python/Django backend). Evaluation reports surprisingly high macro F1 scores of 93–97% on the 3-sector and 10-domain labels.
AI-written summary · about
chapters (5)
- 0:00Start
- 13:18Keyword extraction
- 13:36Multi-label Document Classification
- 16:08Automatic inference of MESOC impacts
- 17:37Impact inference
related talks
01h 00mData, visualization, and designing AI
26mAdding Narrative to BI Dashboards with Natural Language...
29mConquering PDFs- document understanding beyond plain text
24mBertTopic, From Free-Text feedbacks to Calls for Action
35mUsing Scattertext and the Python NLP Ecosystem for Text Visualization
19mHow to look at your data
22mOne-click analysis: Detecting and visualizing insights automatically
38mAutomatically categorizing news articles on nrk.no