NCA-GENM - Data Analysis and Visualization (10% of the exam) - Section 5.2
Manipulate and analyse text-based data with LLMs.
Use LLMs to perform text analysis tasks such as summarisation, classification, entity extraction, and sentiment analysis on unstructured datasets. Recognise how prompt design and output parsing affect the reliability and structure of results returned from the model.
Practice question for this objective
lock_openFree sampleData Analysis and Visualizationmedium
A team analysing free-text survey responses wants to extract named entities such as organisations, dates, and monetary amounts from each response using an LLM. Conceptually, what does the named entity recognition step add to the raw text that makes the output useful for downstream tabular analysis?
AIt rewrites each response into a grammatically simpler paraphrase that is easier for humans to read
BIt translates the response into a single canonical language so all entries share one vocabulary
CIt locates spans of text and tags each with a type label, turning unstructured prose into typed, extractable fieldscheck_circle Correct
DIt assigns a sentiment polarity score to the response so it can be aggregated numerically
Understand that named entity recognition labels typed spans, turning unstructured text into structured fields for analysis. Named entity recognition adds structure by detecting spans and tagging each with a category, producing typed fields that can populate table columns. It neither paraphrases, translates, nor scores sentiment, so only span detection with type labelling describes what it contributes to downstream analysis.
Why A is wrong: Tempting because LLMs can paraphrase, but named entity recognition does not simplify or rewrite the prose; it identifies and labels spans, leaving the surrounding text untouched.
Why B is wrong: Tempting because normalising language aids analysis, but translation is a separate task; named entity recognition operates within the source language and tags spans rather than converting the whole text.
Why C is correct: Correct: named entity recognition detects spans and assigns category labels such as organisation, date, or money, converting unstructured text into typed key-value fields that map cleanly onto columns for tabular analysis.
Why D is wrong: Tempting because both are common text-analytics steps producing structured fields, but sentiment scoring is a distinct task; named entity recognition labels entity spans and does not output a polarity value.
Answers that look right on this material and are not. Each one is a distractor from a different question in the NCA-GENM bank for this domain.
The service label is a faithful reading because the review never names the word service, so the analyst should accept neutral and leave the prompt unchanged.
errorWhy it is wrong: Tempting because the literal word service is absent, but the forty-minute wait is a clear negative service signal; accepting neutral treats an implicit but evident aspect as absent, which is the faithfulness failure the analyst must correct.
The OCR engine lacks a language model for the document language, so it cannot map glyph shapes to characters on any of the files.
errorWhy it is wrong: A missing language pack is a real OCR failure mode, but it would degrade recognition across all scanned files uniformly rather than affecting only the subset that already contains a clean digital text layer.
Raise the sampling temperature so the model commits more decisively to one of the allowed labels and stops adding commentary.
errorWhy it is wrong: Tempting because temperature controls decisiveness, but raising it increases output variability and would make stray commentary more likely, not less; lower temperature would be the decisiveness lever, and neither constrains the response shape.