Which is the best description of an NL2VIS system?
Q2
Microsoft's LIDA organizes chart generation into stages. Which stage describes the dataset (column types, example values) before charts are proposed?
Q3
A user types "show me the data". What should a good NL2VIS system do?
Q4
Why is generating a Vega-Lite specification often safer than generating arbitrary Python plotting code?
Q5
An LLM returns "aggregate": "avg" in a Vega-Lite encoding. What is wrong?
Q6
The dataset has a column district, but the generated spec uses "field": "District". What happens?
Q7
A generated spec computes mean(station_id). Which guardrail catches it?
Q8
Why should prompts to an AI chart generator include the schema (column names, types, a few example values) rather than the full dataset?
Q9
What is prompt injection through data?
Q10
A predicted spec draws horizontal bars (x = mean(pm25), y = district) while the gold spec draws vertical bars (x = district, y = mean(pm25)). Which metric counts it as correct?
Q11
Gold: bar chart of mean PM2.5 by district. Prediction: line chart of mean PM2.5 by district. Execution accuracy says "same data". Why is that not enough?
Q12
Four NL2VIS predictions: 1 exact match, 3 with the same computed data. What are exact-match and execution accuracy?
Q13
"Readings per district" — gold: count(); prediction: sum(pm25). What does the evaluation show?
Q14
What is the ChartQA benchmark about?
Q15
Which chart is generally easier for both people and multimodal AI models to read values from accurately?
Q16
A generated caption says "Ouhai fell by 20 %"; the data shows 38 → 36 µg/m³. What is the verdict?
Q17
"Air quality improved because of the new bike lanes" appears in an AI-generated summary. Why can an automatic fact-checker not verify it?
Q18
Which recommendation system uses design rules plus constraint optimization over Vega-Lite specifications?
Q19
In an AI-assisted dashboard, what should be recorded as provenance?
Q20
A BI copilot is only allowed to use measures defined in the company's semantic model. What does this achieve?
Q21
Which workflow best reflects "AI may draft, humans decide" for a CityPulse report chart?
Q22
Why can over-reliance on AI chart suggestions be a problem even when the suggestions are usually correct?
Q23
NL4DV (Narechania et al., IEEE VIS 2020) is notable because, in addition to Vega-Lite charts, it returns:
Q24
"Busiest stations" is ambiguous for a bike dataset. Which interpretation issue is it?
Q25
Why should captions for AI-generated charts be produced from computed numbers inserted into templates (or tightly constrained generation) on a public portal?
Q26
A chart-recommendation system trained on millions of online charts (e.g., VizML) mostly learns:
Q27
After a model update, the city's chatbot starts using "aggregate": "average" instead of "mean". What should the evaluation pipeline do?
Q28Short answer
Design a guardrail pipeline for CityPulse's "ask a question, get a chart" feature for citizens. List each stage, what it checks or does, and what happens when a check fails. Include privacy and prompt-injection protections.
Q29Short answer
You evaluate an NL2VIS model on 200 CityPulse questions with gold specifications. Describe the metrics you would report, how you would compute execution accuracy, and why you would add a human evaluation. Give an example where each metric misleads.
Q30Short answer
Fact-check this AI-generated caption against the data (monthly mean PM2.5): Lucheng 52, 47, 40 (Jan–Mar); Ouhai 38, 37, 36. Caption: "Lucheng fell by 23 % between January and March. Ouhai fell by 20 %. Lucheng had the highest average PM2.5. The city average was 42. Air quality improved because of the new bike lanes." Rewrite the caption so that every sentence is supported.