11.0 Learning Checklist: Natural Language Processing

Use this page as a printable checklist. If you need the full explanation, return to the Chapter 11 entry page.

Two-Hour First Pass

Time box	Do this	Stop when you can say
20 min	Read the text-to-task pipeline	”NLP starts with raw text and ends with evaluable outputs.”
25 min	Run the label evaluation script	”I can compare predicted labels with expected labels.”
25 min	Skim 11.1 text preprocessing	”Cleaning can help or harm depending on meaning.”
25 min	Skim classification, extraction, and generation roadmaps	”The task is defined by the output.”
25 min	Read the task output map	”I can choose metrics from the output type.”

Required Evidence

Evidence	Minimum version
text_cleaning.py	cleaning, tokenization, before/after examples
label_guide.md	label definitions, boundary cases, positive and negative examples
classification_report.md	metrics, confusion matrix or error table, model comparison
extraction_examples.jsonl	source text, extracted fields, validation result
failure_cases.md	confusing labels, missing fields, unsupported facts, bad summaries
README.md	task goal, run command, input/output, metrics, limitations

Quality Gates

Gate	Pass condition
Label/schema boundary	Labels or fields include positive, negative, and edge examples.
Baseline	Rule, TF-IDF, simple model, or LLM baseline runs on the same fixed eval cases.
Factuality	Generated summaries or answers are checked against source evidence, not only fluency.
Error review	Confusion, missing fields, unsupported facts, and bad summaries have a cause and next test.

Exit Questions

• Can you explain how raw text becomes tokens and model input?
• Can you define label boundaries before training or prompting?
• Can you decide whether a task needs classification, extraction, retrieval, or generation?
• Can you evaluate factual consistency for summaries or answers?
• Can you explain when a traditional NLP method is enough and when an LLM is helpful?

If the answer is yes, you can use NLP ideas more confidently in Prompt, RAG, Agent memory, and multimodal work.

Check reasoning and explanation

1. A strong answer explains the path from raw text to tokens, representation, model input, prediction, metric, and failure case.
2. Label boundaries are ready only when you have positive examples, negative examples, edge cases, and a written rule for disagreements.
3. Choose classification for fixed labels, extraction for fields, retrieval for evidence lookup, generation for new text, and hybrids when outputs require multiple steps.
4. Factual consistency means each generated summary or answer can be traced to source evidence; fluency alone is not enough.
5. Traditional NLP is enough when the task is small, transparent, and stable; LLMs help when language variation, generation, or reasoning over context dominates.

Evidence to Keep

Keep this page’s proof of learning as a small evidence card:

Task Output

label, entity fields, summary, answer, retrieval result, or semantic graph

Artifacts

raw text, processed text, predictions, metrics, and failure cases

Metric

accuracy/F1, precision/recall, retrieval hit rate, faithfulness, or schema validity

Failure Check

unclear labels, over-cleaning, boundary errors, hallucination, or unsupported answer

Expected Output

reproducible text pipeline folder with metrics and examples