Summary and Revision Notes

Core idea

Raw text is not automatically a dataset. It becomes one through boundaries, cleanup, views, and splits.

Terms

TermMeaning
corpusthe collection of text for the task
documentone unit inside the corpus
sentencea smaller linguistic or punctuation-based unit
normalizationmaking chosen text forms consistent
leakageaccidental access to evidence that should be held out

Split roles

train      -> fit the model
validation -> choose settings
test       -> final check

Before moving on

You should be able to describe a tiny corpus, name its document units, explain one normalization choice, and identify a simple leakage risk.