HU Berlin: Industrial-Strength Natural Language Processing with spaCy and Prodigy

Guest lecture for "Deep Learning and Natural Language Processing" at Humboldt University Berlin.

Transcript

1. Industrial-Strength Natural Language Processing with spaCy & Prodigy Ines Montani Founder & CEO Dr. Matthew Honnibal Founder & CTO
2. 480m+ downloads Open-source library for industrial-strength natural language processing spacy.io spaCy Modern scriptable annotation tool for machine learning developers prodigy.ai Prodigy 12.000+ users
3. “ Requirements: We’re building a crime database based on news reports. We want to extract the following: • victim name • perpetrator name • crime location • o ff ence date • arrest date “
6. spacy.fyi/displacy
7. Levels of linguistic annotations application- oriented linguistic descriptions Alex Smith was stabbed in East London
8. Levels of linguistic annotations tokens application- oriented linguistic descriptions Alex Smith was stabbed in East London
9. Levels of linguistic annotations tokens application- oriented linguistic descriptions spans Alex Smith was stabbed in East London
10. Levels of linguistic annotations tokens application- oriented linguistic descriptions spans parse trees Alex Smith was stabbed in East London
11. Levels of linguistic annotations tokens application- oriented linguistic descriptions spans parse trees semantic roles PATIENT ADJUNCT Alex Smith was stabbed in East London
12. Levels of linguistic annotations tokens application- oriented linguistic descriptions spans parse trees semantic roles PATIENT ADJUNCT entities PERSON LOCATION Alex Smith was stabbed in East London
13. Levels of linguistic annotations tokens application- oriented linguistic descriptions spans parse trees semantic roles PATIENT ADJUNCT entities PERSON LOCATION text categories CRIME Alex Smith was stabbed in East London
14. Levels of linguistic annotations tokens application- oriented linguistic descriptions spans parse trees semantic roles PATIENT ADJUNCT entities PERSON LOCATION text categories CRIME relations VICTIM Alex Smith was stabbed in East London
15. spacy.io/usage
16. spacy.io/usage
17. spacy.io/usage
18. spacy.io/usage Doc Doc.ents Doc.noun_chunks Doc.spans Doc.cats …
19. spacy.io/usage Doc Doc.ents Doc.noun_chunks Doc.spans Doc.cats …
20. spacy.io/usage Doc Doc.ents Doc.noun_chunks Doc.spans Doc.cats … Token Token.pos_ Token.dep_ …
21. COMPANY COMPANY MONEY INVESTOR “Hooli raises $5m to revolutionize search, led by ACME Ventures” 5923214 1681056 Database
22. COMPANY COMPANY MONEY INVESTOR “Hooli raises $5m to revolutionize search, led by ACME Ventures” 5923214 1681056 Database named entity recognition
23. COMPANY COMPANY MONEY INVESTOR “Hooli raises $5m to revolutionize search, led by ACME Ventures” 5923214 1681056 Database named entity recognition entity disambiguation
24. COMPANY COMPANY MONEY INVESTOR “Hooli raises $5m to revolutionize search, led by ACME Ventures” 5923214 1681056 Database named entity recognition entity disambiguation custom database lookup
25. COMPANY COMPANY MONEY INVESTOR “Hooli raises $5m to revolutionize search, led by ACME Ventures” 5923214 1681056 Database named entity recognition entity disambiguation custom database lookup currency normalization
26. COMPANY COMPANY MONEY INVESTOR “Hooli raises $5m to revolutionize search, led by ACME Ventures” 5923214 1681056 Database named entity recognition entity disambiguation custom database lookup currency normalization entity relation extraction
27. Case Study: S&P Global explosion.ai/blog/sp-global-commodities ⚠ data-private ⚠ real-time “heards”: trading activities
28. Human-in-the-loop distillation explosion.ai/blog/human-in-the-loop-distillation LLM
29. Human-in-the-loop distillation explosion.ai/blog/human-in-the-loop-distillation continuous evaluation baseline LLM
30. Human-in-the-loop distillation explosion.ai/blog/human-in-the-loop-distillation continuous evaluation baseline LLM prompting
31. Human-in-the-loop distillation explosion.ai/blog/human-in-the-loop-distillation continuous evaluation baseline LLM prompting
32. Human-in-the-loop distillation explosion.ai/blog/human-in-the-loop-distillation continuous evaluation baseline LLM prompting transfer learning COMPONENT
33. Human-in-the-loop distillation explosion.ai/blog/human-in-the-loop-distillation continuous evaluation baseline LLM prompting transfer learning COMPONENT distilled model
34. Human-in-the-loop distillation explosion.ai/blog/human-in-the-loop-distillation continuous evaluation baseline LLM prompting transfer learning COMPONENT distilled model deploy 🚀
35. Data refactoring explosion.ai/blog/sp-global-commodities 99% F-score 6mb model size 16k+ words/second
36. Data refactoring explosion.ai/blog/sp-global-commodities 99% F-score 6mb model size 16k+ words/second
37. Data refactoring explosion.ai/blog/sp-global-commodities 30min per attribute GPT-5 API 99% F-score 6mb model size 16k+ words/second
38. Data refactoring explosion.ai/blog/sp-global-commodities 10× faster reduce cognitive load 30min per attribute GPT-5 API 99% F-score 6mb model size 16k+ words/second
39. Factor out business logic MODEL result = business_logic(classification(text))
40. Factor out business logic MODEL words, grammar, syntax information in the text result = business_logic(classification(text))
41. Factor out business logic MODEL external knowledge facts that can change over time words, grammar, syntax information in the text result = business_logic(classification(text))
42. At their core, many NLP systems consist of flat classifications. You can shove them into a single prompt, or you can decompose them into smaller pieces. Many classification tasks are straightforward to solve nowadays – but they become vastly more complicated if one model needs to do them all at once. explosion.ai/blog/human-in-the-loop-distillation “ “
43. prodigy.ai/docs Intro to Prodigy Annotate
44. prodigy.ai/docs Intro to Prodigy recipe Annotate
45. prodigy.ai/docs Intro to Prodigy dataset recipe Annotate
46. prodigy.ai/docs Intro to Prodigy dataset recipe model / tokenizer Annotate
47. prodigy.ai/docs Intro to Prodigy dataset recipe model / tokenizer input data Annotate
48. prodigy.ai/docs Intro to Prodigy dataset recipe model / tokenizer input data labels Annotate
49. prodigy.ai/docs Intro to Prodigy dataset recipe model / tokenizer input data labels Python function Customize Annotate
50. prodigy.ai/docs Intro to Prodigy dataset recipe model / tokenizer input data labels Python function Customize Annotate Automate model API rules
51. prodigy.ai/docs Intro to Prodigy dataset recipe model / tokenizer input data labels Python function Customize web server & annotation UI + database Annotate Automate model API rules
52. Train prodigy.ai/docs Intro to Prodigy dataset recipe model / tokenizer input data labels Python function Customize web server & annotation UI + database Annotate Automate model API rules
53. Train prodigy.ai/docs Intro to Prodigy dataset recipe model / tokenizer input data labels Python function Customize recipe web server & annotation UI + database Annotate Automate model API rules
54. Train prodigy.ai/docs Intro to Prodigy dataset recipe model / tokenizer input data labels Python function Customize output recipe web server & annotation UI + database Annotate Automate model API rules
55. Train prodigy.ai/docs Intro to Prodigy dataset recipe model / tokenizer input data labels Python function Customize dataset output recipe web server & annotation UI + database Annotate Automate model API rules
56. Train prodigy.ai/docs Intro to Prodigy dataset recipe model / tokenizer input data labels Python function Customize dataset evaluation % output recipe web server & annotation UI + database Annotate Automate model API rules