Voice49/dber

DB-ER — Dataset for Database Entity Recognition

Dataset Summary

DB-ER is a token-level dataset for Database Entity Recognition (DB-ER) in natural-language queries (NLQs) paired with SQL. The task is to tag each token as one of Table, Column, Value, or O (non-entity).
Each example includes: the NLQ, database identifier, a canonical dataset id, the paired SQL query, a tokenized question, a compact entity→token reverse index, an explicit entities table (typed schema/value items), and CoNLL-style DB‑ER tags.

---

Fields

• question_id (int) — Example id
• db_id (str) — Database identifier
• dber_id (str) — Canonical id linking back to the source file/split (BIRD, SPIDER)
• question (str) — NLQ text
• SQL (str) — Paired SQL query
• tokens (List[str]) — Tokenized NLQ
• entities (List[Object]) — Typed DB items referenced in the SQL; each item has:
  • id (int) — Local entity id (unique within the example)
  • type ("table"|"column"|"value")
  • value (str) — Surface form from the DB schema or literal value
• entity_to_token (List[Object]) — Reverse index:
  • entity_id (int) — Refers to an entities[*].id
  • token_idxs (List[int]) — Token indices composing that entity in tokens
• dber_tags (List[str]) — CoNLL-style IOB2 tags over tokens

---

Splits

Entity token prevalence is consistent across splits: ~29% entity vs. ~71% O.

Split	# Examples
human_train	500
human_test	500
synthetic_train	15,026
human_gold_100	100

synthetic_train is produced via our auto-annotation pipeline, which aligns SQL-referenced entities to NLQ spans using string-similarity candidates (Jaccard 3-gram / Levenshtein) and a non-overlapping ILP selection objective. See Annotation below.

human_gold_100 A 100-sample gold standard covering 50 distinct questions. Each question appears twice: once with the original gold annotation and once with an independent blind re-annotation. Although the question text is repeated, the two annotations are presented and counted as separate samples. Unlike the other splits, annotators labeled these questions with access to the actual database instance (BIRD-train and Spider SQLite databases), so labels are grounded in the concrete tables, columns, and stored values of each database.

• 100 annotation samples for 50 questions over 45 distinct databases; 1,500 annotated token instances, 463 entity-token labels (142 Table / 206 Column / 115 Value), and 376 IOB2 mentions.
• Reliability between the two annotations of each question: Cohen's κ = 0.90 (4-class; 95% CI [0.86, 0.93]) and 95.1% raw agreement.
• Rows 1–50 contain the original annotations; rows 51–100 contain the independent re-annotations.

---

Example

{
  "question_id": 13692,
  "db_id": "retail_complains",
  "dber_id": "bird:train.json:282",
  "question": "Among the clients born between 1980 and 2000, list the name of male clients who complained through referral.",
  "SQL": "SELECT T1.first, T1.middle, T1.last FROM client AS T1 INNER JOIN events AS T2 ON T1.client_id = T2.Client_ID WHERE T1.year BETWEEN 1980 AND 2000 AND T1.sex = 'Male' AND T2.`Submitted via` = 'Referral'",
  "tokens": ["Among","the","clients","born","between","1980","and","2000",",","list","the","name","of","male","clients","who","complained","through","referral","."],
  "entities": [
    {"id": 0, "type": "column", "value": "first"},
    {"id": 1, "type": "column", "value": "middle"},
    {"id": 2, "type": "column", "value": "last"},
    {"id": 3, "type": "table", "value": "client"},
    {"id": 4, "type": "table", "value": "events"},
    {"id": 5, "type": "column", "value": "client_id"},
    {"id": 6, "type": "column", "value": "year"},
    {"id": 7, "type": "value", "value": "1980"},
    {"id": 8, "type": "value", "value": "2000"},
    {"id": 9, "type": "column", "value": "sex"},
    {"id": 10, "type": "value", "value": "Male"},
    {"id": 11, "type": "column", "value": "Submitted via"},
    {"id": 12, "type": "value", "value": "Referral"}
  ]
  "entity_to_token": [
    ...,
    {"entity_id":3,"token_idxs":[2]},
    {"entity_id":5,"token_idxs":[14]},
    {"entity_id":7,"token_idxs":[5]},
    {"entity_id":8,"token_idxs":[7]},
    {"entity_id":10,"token_idxs":[13]},
    {"entity_id":12,"token_idxs":[18]},
    ...
  ],
  "dber_tags": ["O","O","B-TABLE","O","O","B-VALUE","O","B-VALUE","O","O","O","O","O","B-VALUE","B-COLUMN","O","O","O","B-VALUE","O"]
}

---

Annotation

• Human: collaborative web UI with schema and SQL visible during labeling.
• Gold (human_gold_100): 100 samples from two annotations of the same 50 questions, annotated against the live database instance (read-only). The original annotation used the paired SQL for adjudication; the independent re-annotation was blind to the paired SQL and original labels (κ = 0.90).
• Synthetic: for each NLQ–SQL pair, generate candidate spans with Jaccard/Levenshtein, then solve a non-overlapping ILP to select spans maximizing similarity. Hyperparameters are validated on human data.

---

Data provenance

• Sources: text-to-SQL benchmarks BIRD (https://bird-bench.github.io/) and Spider (https://yale-lily.github.io/spider).
• Transform: NLQ–SQL pairs → DB-ER annotations via the synthetic pipeline; human annotations provide gold labels and validation.

---

Release notes

• v1.3 (2026-07-24): Expanded the gold split to human_gold_100 by adding the independent re-annotation as 50 additional samples
• v1.2 (2026-07-18): Added human_gold_50 — 50-question database-scoped gold standard (evaluation only); explicit split configs for the Data Viewer
• v1.1 (2025-08-26): HF Data Viewer compatibility update
• v1.0: Initial public release