data/docs/statistics.md

# InfoSeek statistics

최신 local InfoSeek 파일과 새로 추가된 `infoseek_val_qtype.jsonl` 기준으로 다시 계산했다.

Raw 계산 결과:

```text
dataset_explanation/infoseek/infoseek_stats_full.json
```

사용한 KnowCoL entity set:

```text
KnowCoL/knowcol/datasets/wikidata_subgraph_v1/entity.txt
```

주의:

```text
train/val은 with-KB 파일이 있으므로 gold QID가 명시되어 있다.
test/human은 local에 explicit with-KB 파일이 없다.
또한 최신 OVEN test/human 파일에도 entity_id가 없다.
따라서 test/human의 entity coverage는 현재 local 파일만으로는 공식적으로 계산할 수 없다.
```

---

## 1. Local files

```text
infoseek_train.jsonl
infoseek_train_withkb.jsonl
infoseek_val.jsonl
infoseek_val_withkb.jsonl
infoseek_val_qtype.jsonl
infoseek_test.jsonl
infoseek_human.jsonl
infoseek_human_qtype.jsonl
```

---

## 2. Train

Gold entity source:

```text
infoseek_train_withkb.jsonl
```

```text
rows / Q-A pairs: 934,048
unique images: 770,602
unique entities: 5,549
missing entity rows: 0
```

KnowCoL coverage:

```text
in KnowCoL entity.txt: 5,549
not in KnowCoL entity.txt: 0
```

Answer type:

```text
list[str] / string-or-time style: 743,065
numerical dict: 190,983
```

---

## 3. Validation

Gold entity source:

```text
infoseek_val_withkb.jsonl
```

Question type source:

```text
infoseek_val_qtype.jsonl
```

Overall:

```text
rows / Q-A pairs: 73,620
unique images: 37,999
unique entities: 1,794
missing entity rows: 0
```

KnowCoL coverage:

```text
in KnowCoL entity.txt: 1,794
not in KnowCoL entity.txt: 0
```

Answer type:

```text
list[str] / string-or-time style: 57,724
numerical dict: 15,896
```

Question type:

```text
String: 54,319
Numerical: 15,896
Time: 3,405
```

### val_unseen_entity

```text
rows: 54,964
unique images: 26,278
unique entities: 1,192

KnowCoL coverage:
  in KnowCoL entity.txt: 1,192
  not in KnowCoL entity.txt: 0

answer type:
  list[str] / string-or-time style: 43,455
  numerical dict: 11,509

question type:
  String: 40,864
  Numerical: 11,509
  Time: 2,591
```

### val_unseen_question

```text
rows: 18,656
unique images: 11,727
unique entities: 602

KnowCoL coverage:
  in KnowCoL entity.txt: 602
  not in KnowCoL entity.txt: 0

answer type:
  list[str] / string-or-time style: 14,269
  numerical dict: 4,387

question type:
  String: 13,455
  Numerical: 4,387
  Time: 814
```

---

## 4. Test

Current local status:

```text
explicit test_withkb file: not present
latest OVEN test files with entity_id: not present
```

Therefore:

```text
rows / Q-A pairs: 347,980
unique images: 193,793
gold entity coverage: not reliably available from current local files
```

Only a small subset can be weakly resolved by accidental image-id overlap with labeled OVEN train/val files:

```text
resolved rows by image_id overlap: 5,679
missing entity rows: 342,301
resolved unique entities: 41
```

This is not an official test entity mapping and should not be used as a main coverage claim.

Answer type:

```text
list[str] / string-or-time style: 268,361
numerical dict: 79,619
```

Split:

```text
test_unseen_entity:
  rows: 249,079
  unique images: 130,517
  weakly resolved rows: 4,300
  missing entity rows: 244,779
  answer type:
    list[str] / string-or-time style: 194,666
    numerical dict: 54,413

test_unseen_question:
  rows: 98,901
  unique images: 63,531
  weakly resolved rows: 1,379
  missing entity rows: 97,522
  answer type:
    list[str] / string-or-time style: 73,695
    numerical dict: 25,206
```

---

## 5. Human

Current local status:

```text
explicit human_withkb file: not present
latest OVEN human file with entity_id: not present
```

Therefore:

```text
rows / Q-A pairs: 8,931
unique images: 8,884
gold entity coverage: not reliably available from current local files
```

Only a very small subset can be weakly resolved by accidental image-id overlap with labeled OVEN train/val files:

```text
resolved rows by image_id overlap: 116
missing entity rows: 8,815
resolved unique entities: 14
```

This is not an official human entity mapping and should not be used as a main coverage claim.

Answer type:

```text
list[str] / string-or-time style: 6,551
structured numeric dict: 2,380
```

Question type from `infoseek_human_qtype.jsonl`:

```text
String: 4,049
Time: 2,502
Numerical: 2,380
```

Split:

```text
human_unseen_entity:
  rows: 5,683
  unique images: 5,682
  weakly resolved rows: 68
  missing entity rows: 5,615
  question type:
    String: 2,600
    Time: 1,598
    Numerical: 1,485

human_unseen_question:
  rows: 3,248
  unique images: 3,203
  weakly resolved rows: 48
  missing entity rows: 3,200
  question type:
    String: 1,449
    Time: 904
    Numerical: 895
```

---

## 6. Reliable entity coverage claim

Reliable coverage should currently be stated only for train/val:

```text
InfoSeek train unique entities: 5,549
InfoSeek val unique entities: 1,794
train/val union: available in raw stats

all explicit train/val entities are in KnowCoL entity.txt.
```

Do not claim full test/human entity coverage unless we restore:

```text
test_withkb / human_withkb
or
older OVEN evaluation files that contain gold entity_id
or
an official image_id -> entity_id mapping.
```

---

## 7. Interpretation for our framework

Usable immediately:

```text
train:
  gold QID available
  can build entity/fact training data

val:
  gold QID available
  qtype available
  can evaluate fact recovery by String / Numerical / Time
```

Not reliable yet:

```text
test/human entity/fact coverage
```

For test/human QA evaluation, we can still use the question-answer files as QA benchmarks, but entity-level or fact-level oracle analysis requires restoring gold QID mappings.