2025-10-16-pycon-nl

Keynote given at PyCon NL 2025: https://pycon-nl.org/talks/2025/14

Transcript

1. How to not get fooled by your data while AI engineering Sofie Van Landeghem NLP freelancer @ OxyKodit Pycon NL 2025
2. Create a portal with all Formula One news
3. And of course, with a chatbot Who is currently leading the 2025 F1 championship? As of the latest standings after the 2025 Singapore GP, Oscar Piastri is leading the Drivers’ Championship. Who won in Monaco? Lando Norris did.
4. Who crossed the line first in Spa 2024? George Russel crossed the finish line first, securing victory for Mercedes. So Russel won the Spa GP in 2024? No, Hamilton did. Russel his car was disqualified after the race due to being 1.5kg underweight.
5. Wait, what?
6. Always have a domain expert on board who can double check the data and the results
7. Our chatbot should be able to link to Wikipedia
8. "Hamilton won the Belgian Grand Prix in Spa in 2024."
10. Obtaining gold-standard data Sir Lewis Carl Davidson Hamilton (born 7 January 1985) is a British [[racing driver]] who competes in [[Formula One]] for [[Scuderia Ferrari|Ferrari]].
11. Building an Entity Linker model
12. ✦ 79% F-score ✦
13. 79% is pretty good, right? Right?
14. How most clients / users think about performance 0% 50% 100% The worst model "Random guessing" The best model "Entirely correct"
15. … but a random baseline is almost never 50%
16. For our entity linking task, the random baseline is actually 1/7.075.098 = 0,000014% (a.k.a. 0%)
17. … and our upper bound wasn't 100% either …
18. In fact, we had to prune the database for efficiency requirements ❖ Kept only 14% of all Wikipedia concepts ❖ An "oracle" disambiguation obtains 84% F-score So, the modeling work targets a 0% - 84% range
19. Ok, so within a range of 0-84%, 79% is really good, right? Right?
20. Develop some simple baselines to better understand the complexity and data challenges of your project
21. In our case, we won't assign "Hamilton" to a random page out of the 7M available ones
22. Instead, we obtain a list of "Hamilton" candidates through lexical search
23. Now, if we pick a random one from this candidate list, we actually obtain 54% F-score (without doing any ML/NLP/AI at all!)
24. Let's create an even stronger baseline.
25. Consider "prior probabilities": a measure of how often a mention is linked to a certain concept Textual mention F1 race driver Founding father Musical "Hamilton" 35% 55% 10% "Lewis Hamilton" 99% 0% 0% "Alexander Hamilton" 0% 82% 15%
26. Which we can extrapolate to determine which one is the "most likely" (without any context) Textual mention F1 race driver Founding father Musical "Hamilton" ✓ "Lewis Hamilton" ✓ "Alexander Hamilton" ✓
27. This "most likely" baseline obtains 78.2% F-score (still without any AI at all!)
28. The ML model that is supposed to disambiguate based on context (79%) only marginally improves upon a relatively simple baseline (78.2%)
29. Developing a few simple baselines puts your AI's performance measures in the right perspective
30. Let's revisit our upper bound (again). Is it really 100% if we have no memory / efficiency requirements? 0 100% The best model "Entirely correct"
31. 100% precision means every disambiguation is "correct". But what does "correctness" even mean?
32. "Hamilton won the Belgian Grand Prix in Spa in 2024." ? ?
33. "Societies in the ancient civilizations of Greece and Rome preferred small families" ? ?
34. When multiple annotators label the same data sample, how often do they agree? ↪ Inter-annotator agreement = IAA
35. Annotators mostly agree → the gold data is reliable and robust Annotators disagree because… ★ The data is confusing ★ The label scheme is ambiguous ★ The NLP task is too complex High IAA Low IAA
36. We want to label incoming articles for our portal 1. Sports a. Football b. F1 2. Politics a. Global politics b. US politics 3. Leisure a. Traveling b. Sports c. Literature
37. Let's plot IAA of the labels… Football F1 Global politics US Politics Traveling Sports Literature Football 0.79 F1 0 0.73 Global politics 0 0 0.95 US Politics 0 0.12 0.05 0.83 Traveling 0 0 0 0 1.0 Sports 0.21 0.15 0 0 0 0.64 Literature 0 0 0 0 0 0 1.0
38. Confusion between global and US politics? Football F1 Global politics US Politics Traveling Sports Literature Football 0.79 F1 0 0.73 Global politics 0 0 0.95 US Politics 0 0.12 0.05 0.83 Traveling 0 0 0 0 1.0 Sports 0.21 0.15 0 0 0 0.64 Literature 0 0 0 0 0 0 1.0
39. Consider merging labels if humans can't reliably distinguish them (or when the distinction doesn't matter)
40. Confusion between politics and F1? Football F1 Global politics US Politics Traveling Sports Literature Football 0.79 F1 0 0.73 Global politics 0 0 0.95 US Politics 0 0.12 0.05 0.83 Traveling 0 0 0 0 1.0 Sports 0.21 0.15 0 0 0 0.64 Literature 0 0 0 0 0 0 1.0
41. Solution: allow multiple correct labels 1. Sports a. Football b. F1 2. Politics a. Global politics b. US politics 3. Leisure a. Traveling b. Sports c. Literature
42. Reframe the task to match your data
43. Confusion within the sports category? Football F1 Global politics US Politics Traveling Sports Literature Football 0.79 F1 0 0.73 Global politics 0 0 0.95 US Politics 0 0.12 0.05 0.83 Traveling 0 0 0 0 1.0 Sports 0.21 0.15 0 0 0 0.64 Literature 0 0 0 0 0 0 1.0
44. 1. Sports a. Football b. F1 2. Politics a. Global politics b. US politics 3. Leisure a. Traveling b. Sports c. Literature Solution: critically revise the label scheme given to you! Inherent ambiguity in the label scheme needs to be addressed
45. Together with your domain experts, critically revise the label scheme
46. What if we don't have multiple annotators?
47. Analyse discrepancies between a model's prediction and the gold standard label
48. "Agnes Maria of Andechs-Merania (died 1201) was a Queen of France." Gold annotation Prediction
49. Prediction Gold ≠ Is this really an incorrect prediction (a "false positive")? Or is the label wrongly annotated? Perhaps both answers can be seen as "correct"?
50. Depending on the downstream use-case, the precision of this method was 87-96%
51. Analyse "wrong" predictions on the training dataset, to find structural data errors
52. Manually annotated training dataset: "... to kick off 2025 [Dutch Grand Prix] weekend in Zandvoort" "Gasly laments ‘quite sad’ [Monaco] GP crash" Sample prediction: "Gasly laments ‘quite sad’ [Monaco GP] crash"
53. Always include the words "Grand Prix" or "GP" into the entity annotation: ❌ Gasly laments ‘quite sad’ [Monaco] GP crash ✅ Gasly laments ‘quite sad’ [Monaco GP] crash Write up annotation guidelines to help with consistency of annotations
54. "Correctness" may also depend on the downstream usage of your results
55. "Hamilton won the Belgian Grand Prix in Spa in 2024." ? ?
56. "Hamilton won the Belgian Grand Prix in Spa in 2024." ? ?
57. Make sure you climb the right hill
58. Is your data correct, consistent and robust? Is your test data representative of real-world data? Does your evaluation capture the value of the algorithm in production?
59. Only after defining the right hill to climb, you can get to the fun stuff: ★ Algorithm development ★ Model training ★ LLM fine-tuning ★ …
60. Let's wrap up with a practical checklist 📝
61. AI project: data annotation ✓ Ensure your label scheme is consistent and unambiguous ✓ Draft clear annotation guidelines to ensure data consistency ✓ Measure inter-annotator agreement (IAA) ✓ Consider reframing your task/guidelines if the IAA is low ✓ Model uncertainty in your annotation workflow 📝 1/3
62. AI project: performance evaluation ✓ Develop simple baselines to put performance into perspective ✓ Quantify realistic upper/lower performance bounds ✓ Measure performance as part of the larger business process 📝 2/3
63. AI project: performance evaluation ✓ Identify structural data errors by "predicting" the training data ✓ Apply to truly unseen data to measure realistic performance ✓ Make sure you’re climbing the right hill 📝 3/3
64. Main Take Away Hamilton won the Belgian 2024 Grand Prix, not Russell ;-)
65. Thank you! → Github: svlandeg → LinkedIn: https://www.linkedin.com/in/sofievanlandeghem → NLP consultancy: https://oxykodit.com