IC-848GPT-3.5 achieves 98.3% precision and 96.0% recall for automatic question-tuple matching but makes errors when questions differ in wording yet are semantically unique

Jaemin Cho, Yushi Hu, Jason Michael Baldridge, Roopal Garg, Peter Anderson, Ranjay Krishna, Mohit Bansal, Jordi Pont-Tuset, Su Wang

SourceDavidsonian Scene Graph: Improving Reliability in Fine-grained Evaluation for Text-to-Image Generation

The paper uses GPT-3.5 with in-context examples to automatically evaluate whether generated DSG questions match their source semantic tuples (precision/recall) and whether questions are unique (no duplicates). On the full TIFA160 set, GPT-3.5 achieves 98.3% precision and 96.0% recall for matching, and 90.5% uniqueness for DSG questions. However, the paper identifies a specific error pattern: GPT-3.5 is 'distracted by text overlap' and flags semantically distinct questions as duplicates when they share surface wording, or misses true duplicates that are phrased differently.

Evidence
correlational
Key metric
precision 98.3%, recall 96.0% (automatic, 160 prompts); uniqueness 90.5% for DSG (automatic, 160 prompts); manual: precision 92.2%, recall 100.0% (30 prompts)
Caveat
The automatic evaluation uses GPT-3.5 as both the system under test and the evaluator for some metrics, and the paper notes disagreement between automatic and human evaluation in specific cases (Figure 9).
Model
GPT-3.5 / ChatGPT-3.5
Concepts
Failure mode
Related findings
IC-847
Extraction
automatic-extraction