NCP-GENL Data Preparation Practice Question

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Hash every token in the corpus so that the original text cannot be reconstructed, and train on the hashed sequences.

  Why it's wrong here

  Hashing every token destroys the semantic and syntactic structure the model needs to learn language from; the resulting sequences carry no transferable linguistic signal. It also does not reliably anonymize free-form PII because hashing is deterministic and can be attacked with dictionary lookups for common names or emails. The approach fails both the privacy guarantee and the context preservation requirement.

• ✗

  Replace all digits and capitalized words with a generic mask token across the entire corpus before training.

  Why it's wrong here

  Masking all digits and capitalized words destroys legitimate content such as product names, dates, and technical terms, producing text that no longer resembles real support language. The model would learn from a distorted distribution, and the masking is not targeted enough to guarantee that every PII span is caught, since names can appear in lowercase. Privacy improves only incidentally while data quality suffers broadly.

• ✓

  Detect PII spans with a combination of regular expressions and a named-entity recognition model, then replace each detected span with a consistent placeholder token that preserves the surrounding sentence structure.

  Why this is correct

  Regexes reliably catch structured identifiers such as emails and account numbers, while a named-entity recognition model covers names and locations that patterns miss. Replacing spans with consistent placeholders removes the sensitive values but keeps sentence structure and surrounding context intact, which is exactly what continued pretraining needs to learn language patterns without memorizing private data.

• ✗

  Drop every transcript that contains any detected PII so that no sensitive information reaches the training pipeline.

  Why it's wrong here

  Discarding all transcripts with PII removes a large fraction of the corpus, shrinking the training set and biasing it toward interactions that happen to lack names or contact details. Customer support text almost always contains such information, so the remaining data would be unrepresentative. This approach satisfies the privacy requirement but fails the goal of preserving linguistic context for the model to learn from.