Data-Driven NLP: From Statistics to GenAI
summary
Patent intelligence product Vantage for IT processes hundreds of thousands of patents using NLP and is presented as a working system. Model-size progression is concrete: GPT1 125M, GPT2 1.5B, GPT3 175B and training corpora reported up to 13 trillion tokens. Early perplexities (n‑gram >120, RNN ~73.4, LSTM ~44.9) contrast with benchmarks like BLEU 41.8, VQA 84.3% and a study claiming 80+% multimodal assessments outperform physician diagnosis. Discussed tools include pruning, distillation and quantization.
AI-written summary · about
chapters (12)
- 0:00Introduction to NLP Evolution
- 0:47Speaker Introduction and Background
- 1:36NLP's Rise and Investment Surge
- 3:12Early NLP Techniques and Challenges
- 5:49Shift to Machine Learning and Deep Learning
- 8:00Transformers and Attention Mechanisms
- 9:48BERT and Bidirectional Training
- 12:25GPT Series and Scaling Trends
- 14:16Multimodal Models and Applications
- 16:44Efficiency Techniques in NLP
- 19:11Future Metrics and Ethical Considerations
- 20:34Conclusion and Future of NLP
