L04 | English F3 Student Learning Log

Persona: I'm a Form 3 (Grade 9) student and I only read English. I know basic Python (variables, for loops, lists) and school algebra up to quadratics. I learn from the official pages, the videos with captions, and the book. Kaggle notebook steps are written as a practice log that follows the official notebooks.

What I understood (in my own words)

• NLP means getting computers to work with text: classifying it, finding similar phrases, summarising.
• Text is split into small pieces (tokens), then each piece becomes a number the model can use.
• English words are often split into subwords, e.g. 'unhappiness' might become 'un', 'happi', 'ness'.
• Overfitting is like memorising past papers: change the question and you're stuck. So we test on unseen data (the validation set).
• The test set stays hidden until the very end, or we'd accidentally tune the model to the validation set.

Where I got stuck, and how I got unstuck

My answers to the exercises

Q1. Neural nets only handle numbers. What two steps turn an English sentence into model input, and what does each do?

Step 1, tokenization: split the sentence into tokens (words or subwords). Step 2, numericalization: replace each token with its ID from the vocabulary, giving a list of numbers.

Q2. Explain overfitting and underfitting in your own words, with an exam-revision example for each.

Overfitting: memorising the answers to the 20 practice questions, then failing when the exam changes the numbers. Underfitting: only skimming the contents page and failing even the easy questions. Good revision means understanding the method so new questions work too.

Q3. You want to predict next month's shop sales using three years of data. How should you choose the validation set, and why not pick randomly?

Use the most recent months (say the last 2–3) as the validation set and earlier data for training. We're predicting the future; a random split lets the model peek at 'future' data, so its score looks better than it really is.

Q4. If you already have a validation set, why do you still need a test set?

We keep adjusting the model based on validation scores. After enough tweaks it fits that particular validation set, which is a sneaky kind of memorising. The test set is never used for tuning, so one final check shows true performance.

Q5. What do Pearson r = 0.9, 0.1 and −0.8 each mean? Why does the course say you must plot the data instead of only reading the number?

0.9: strong positive relationship, one goes up as the other goes up. 0.1: almost no relationship. −0.8: strong negative, one goes up as the other goes down. You must plot because r is very sensitive to outliers; a few strange points can change it a lot.

My Zettelkasten card

Card｜Rewrite a new problem as one you already know how to solve

When you hit an unfamiliar problem, first try to turn it into a familiar one. Patent-phrase similarity sounds hard, but gluing the two phrases and their context into one text makes it an ordinary text classification task, which pretrained models already handle.

It's just like word problems in maths. Long paragraphs used to make me panic until our teacher showed us how to turn the words into an equation first. After that, it was a linear equation I already knew how to solve. The real skill is the rewriting step. Rewriting can lose information, though. If you just join three columns, the model may not know which part is the context. The course adds markers like 'TEXT1:' and 'TEXT2:' to keep the structure, so rewriting needs care.

Next time I'm stuck on a coding task, I'll write one sentence first: 'This is really a ___ problem', and pick my method from there.

🔗 Related notes to grow next: → Problem solving: 'Ask which solved problem this one resembles' → Language: 'Chinese text needs an extra word-splitting step' → Exams: 'Past-paper scores can fool you'

---

Back to lesson note：L04_自然語言處理_202611041526