Which Wikipedia articles did we use?
We downloaded 142 introductions. Spark retained 141 articles and split them into 228 passages for MiniLM embeddings.
| Article | What it explains | Use in the project |
| Longline fishing | A main line carries many baited hooks. | Maps drifting and set longlines to hooks. |
| Gillnetting | Fish become caught in a wall of netting. | Maps set gillnets to nets. |
| Seine fishing | A surrounding net encloses fish. | Maps seine and purse-seine categories to nets. |
| Trawling | Boats tow a fishing net through the water. | Maps trawlers to nets. The table cannot separate bottom from midwater trawling. |
| North Sea | Describes the sea’s location and geography. | Background only; it does not assign records to a region. |
Only checked method-to-gear mappings affect the SQL. Background articles are retained for context.
Does a definition library help generated SQL?
Schema-only SQL answered 3/5 correctly; adding the definition library answered 5/5.
| Question | Schema only | With definitions |
|---|
| Northern hook/net activity | Match | Match |
| Southern net-fishing hotspots | Mismatch | Match |
| Vessel activity by gear | Match | Match |
| Missing registry information | Match | Match |
| Monthly changes by fishing method | Mismatch | Match |
GPT-5.2; same questions and schemas, only the definition library changes. First responses, no repairs. Scores describe these five questions.
See the questions, schema, reference answers, and both GPT queries →
Question 5 accepts month-start dates and month numbers as equivalent representations. Strict formatting scored B 4/5; answer correctness scored B 5/5.