infoseek-numeric-fact-store / data /docs /fact_coverage.md
Kyle-han's picture
Upload folder using huggingface_hub
908f811 verified
|
Raw
History Blame Contribute Delete
13.9 kB

InfoSeek fact coverage

이 문서는 InfoSeek 질문에서 필요한 positive fact를 어디서 복원할 수 있는지 정리한다.

현재 기준:

train/val:
  gold QID exists via *_withkb.jsonl
  fact coverage analysis is reliable

test/human:
  no explicit *_withkb file locally
  latest OVEN test/human files also have no entity_id
  fact coverage analysis is not reliable yet

Raw structural stats:

dataset_explanation/infoseek/fact_coverage_structural.json
dataset_explanation/infoseek/infoseek_stats_full.json

Generated fact-store artifacts:

codebook-contrastive/build_fact_store/output_facts/infoseek_facts.jsonl
codebook-contrastive/build_fact_store/output_facts/qid_to_fact_ids.json
codebook-contrastive/build_fact_store/output_facts/pid_to_fact_ids.json
codebook-contrastive/build_fact_store/output_facts/qid_label_alias_cache.json
codebook-contrastive/build_fact_store/output_facts/infoseek_train_question_fact_map.jsonl
codebook-contrastive/build_fact_store/output_facts/infoseek_val_question_fact_map.jsonl
codebook-contrastive/build_fact_store/output_facts/infoseek_question_fact_map_summary.json
codebook-contrastive/build_fact_store/output_facts/fact_coverage_by_qtype.json
codebook-contrastive/build_fact_store/output_facts/infoseek_numeric_facts.jsonl
codebook-contrastive/build_fact_store/output_facts/infoseek_train_numeric_question_fact_map.jsonl
codebook-contrastive/build_fact_store/output_facts/infoseek_val_numeric_question_fact_map.jsonl
codebook-contrastive/build_fact_store/output_facts/infoseek_numeric_fact_summary.json
codebook-contrastive/build_fact_store/output_facts/wikidata_claim_cache/

Current label cache status:

qid_label_alias_cache entries: 16,786
entries with label: 16,605

fact store unique object QIDs: 18,584
object QIDs with label: 16,743
object QIDs without label: 1,841
fact rows missing object label: 28,785 / 1,141,235

1. What counts as fact coverage?

There are two different levels.

Entity-level KG coverage

Does the gold entity QID exist in our entity universe?

For train/val, all explicit InfoSeek entities are in KnowCoL:

train unique entities: 5,549 / 5,549 in KnowCoL
val unique entities: 1,794 / 1,794 in KnowCoL

Query-specific positive fact coverage

Does the KG contain the exact fact needed to answer this question?

Example:

question: What country does this river belong to?
gold entity: Q214924 / Rhine Falls
answer: Switzerland

positive fact:
  (Q214924, P17, Q39)
  Rhine Falls --country--> Switzerland

This is the useful coverage for FactHead training.


2. KnowCoL local KG limitation

KnowCoL local KG stores triples like:

(subject_qid, property_pid, object_qid)

So it is naturally useful for entity-valued facts:

(Q214924, P17, Q39)
Rhine Falls --country--> Switzerland

But it does not directly store literal quantity/time values such as:

height = 23 metre
date of birth = 1901-01-01
average flow = 600 m^3/s

For those, we need full Wikidata claims:

wikibase-item claims
quantity claims
time claims

Therefore:

STRING:
  often recoverable from KnowCoL local KG if object label/alias matches answer_eval

NUMERICAL:
  usually requires full Wikidata quantity claims

TIME:
  usually requires full Wikidata time claims

3. Train structural coverage

Gold entity source:

infoseek_train_withkb.jsonl
rows: 934,048
resolved rows with entity: 934,048
unique entities: 5,549

entities in KnowCoL entity.txt: 5,549
entities not in KnowCoL entity.txt: 0

entities with local outgoing KnowCoL facts: 5,544
entities without local outgoing KnowCoL facts: 5
rows whose entity has no local head entry: 948

Answer type:

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

Interpretation:

Most train samples have a local entity-object fact pool.
But 190,983 numerical rows need full Wikidata quantity matching for real positive fact recovery.

Current local positive matching result:

rows: 934,048
rows with matched local positive fact: 240,946
positive rate: 25.80%

positive_status:
  matched_local_tail_surface: 240,946
  no_match_local_tail_surface: 692,154
  no_local_head_entry: 948

entity-level:
  unique entities: 5,549
  entities with at least one local positive: 1,921
  entity-major status:
    matched_local_tail_surface: 1,394
    no_match_local_tail_surface: 4,150
    no_local_head_entry: 5

average negatives per row:
  same_entity_different_fact: 15.62
  same_property_different_subject_object: 15.90

This is still conservative. A local positive is counted only when an object QID label/alias exactly matches answer_eval.

Type-level breakdown:

train total:
  rows: 934,048
  matched: 240,946
  rate: 25.80%

Numerical:
  rows: 190,983
  matched: 0
  rate: 0.00%
  reason: local KnowCoL KG does not encode quantity literals

StringOrTime:
  rows: 743,065
  matched: 240,946
  rate: 32.43%
  note: train qtype file is not used here, so String and Time are not separated

4. Validation structural coverage

Gold entity source:

infoseek_val_withkb.jsonl

Question type source:

infoseek_val_qtype.jsonl

Overall:

rows: 73,620
resolved rows with entity: 73,620
unique entities: 1,794

entities in KnowCoL entity.txt: 1,794
entities not in KnowCoL entity.txt: 0

entities with local outgoing KnowCoL facts: 1,794
entities without local outgoing KnowCoL facts: 0
rows whose entity has no local head entry: 0

Question type:

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

val_unseen_entity

rows: 54,964
unique entities: 1,192
entities with local outgoing KnowCoL facts: 1,192

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

val_unseen_question

rows: 18,656
unique entities: 602
entities with local outgoing KnowCoL facts: 602

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

Interpretation:

Val is fully usable for fact coverage evaluation because gold QID exists.
Local KnowCoL gives a fact pool for every val entity.
However, query-specific positive matching still requires answer-to-claim matching.

Current local positive matching result:

rows: 73,620
rows with matched local positive fact: 14,445
positive rate: 19.62%

positive_status:
  matched_local_tail_surface: 14,445
  no_match_local_tail_surface: 59,175

entity-level:
  unique entities: 1,794
  entities with at least one local positive: 813
  entity-major status:
    matched_local_tail_surface: 371
    no_match_local_tail_surface: 1,423

average negatives per row:
  same_entity_different_fact: 15.78
  same_property_different_subject_object: 15.95

The previous no_object_label_cache issue is mostly resolved for InfoSeek train/val tails. Remaining unmatched rows are now mostly true no_match_local_tail_surface: the local entity-object KG fact pool exists, but no object label/alias exactly matches the answer.

Type-level breakdown:

val total:
  rows: 73,620
  matched: 14,445
  rate: 19.62%

String:
  rows: 54,319
  matched: 14,445
  rate: 26.59%

Numerical:
  rows: 15,896
  matched: 0
  rate: 0.00%

Time:
  rows: 3,405
  matched: 0
  rate: 0.00%

Split-level type breakdown:

val_unseen_entity:
  String:    10,864 / 40,864 = 26.59%
  Numerical:     0 / 11,509 = 0.00%
  Time:          0 /  2,591 = 0.00%

val_unseen_question:
  String:     3,581 / 13,455 = 26.61%
  Numerical:     0 /  4,387 = 0.00%
  Time:          0 /    814 = 0.00%

5. Full Wikidata numeric-like claim coverage

For the next stage, Numerical and Time are merged into one numeric-like group.

numeric-like = quantity claims + time claims

Reason:

The retriever does not need to treat time and quantity as fundamentally different classes.
Both are structured value facts attached to a subject entity and property.

Generated artifacts:

infoseek_numeric_facts.jsonl
infoseek_train_numeric_question_fact_map.jsonl
infoseek_val_numeric_question_fact_map.jsonl
infoseek_numeric_fact_summary.json
wikidata_claim_cache/{QID}.json

Numeric-like fact schema:

quantity fact:
  subject_qid
  property_id
  amount
  unit / unit_qid
  lower_bound / upper_bound if available

time fact:
  subject_qid
  property_id
  time
  year
  precision

Positive matching:

Numerical:
  match Wikidata quantity amount to InfoSeek answer_eval range

Time:
  match Wikidata time year to answer_eval year variants

Negative facts:

same_entity_different_numeric_time_fact:
  another quantity/time claim from the same gold entity

same_property_different_subject_value:
  same property, different subject/value

Combined train+val Wikidata claim store:

unique InfoSeek train/val entities: 6,741
numeric-like facts: 7,939
  quantity: 6,296
  time: 1,643

Train numeric-like coverage

Train currently uses only explicit Numerical rows because no train qtype file is available locally.

Numerical rows: 190,983
matched Wikidata numeric-like positive: 158,870
coverage: 83.19%

status:
  matched_wikidata_numeric_like_claim: 158,870
  no_wikidata_numeric_time_claims: 27,796
  no_match_wikidata_numeric_like_claim: 4,317

Validation numeric-like coverage

Validation uses infoseek_val_qtype.jsonl, so Numerical and Time are both included.

Numeric-like rows: 19,301
matched Wikidata numeric-like positive: 18,912
coverage: 97.98%

status:
  matched_wikidata_numeric_like_claim: 18,912
  no_match_wikidata_numeric_like_claim: 369
  no_wikidata_numeric_time_claims: 20

Breakdown:

Numerical:
  rows: 15,896
  matched: 15,526
  coverage: 97.67%

Time:
  rows: 3,405
  matched: 3,386
  coverage: 99.44%

Interpretation:

Local KnowCoL KG cannot cover Numerical/Time because it mostly stores QID-object triples.
Full Wikidata claims recover most numeric-like positives.

Remaining misses are mostly:

1. the entity has no quantity/time claim of the required kind
2. the answer value is present in Wikipedia text but not structured Wikidata
3. the matching rule is too strict for ranges, units, or temporal precision

6. Test and human status

Current local files do not support reliable fact coverage for test/human.

Reason:

InfoSeek test/human:
  no local test_withkb / human_withkb file

latest OVEN test/human:
  no entity_id field

Weak image-id overlap with labeled OVEN train/val resolves only a tiny subset:

test:
  rows: 347,980
  weakly resolved rows: 5,679
  missing entity rows: 342,301

human:
  rows: 8,931
  weakly resolved rows: 116
  missing entity rows: 8,815

Therefore:

Do not report test/human fact coverage from current local files.
Use test/human for QA-only evaluation unless official QID mappings are restored.

7. Positive fact recovery logic

For each InfoSeek train/val sample:

input:
  gold entity QID S
  question q
  answer / answer_eval A

Candidate facts:

local KnowCoL facts:
  all triples with subject S

full Wikidata claims:
  all wikibase-item / quantity / time claims for S

Positive fact matching:

STRING:
  match answer_eval against object label / alias / normalized surface

NUMERICAL:
  match answer_eval numeric range against quantity claim amount

TIME:
  match answer_eval string/date/year against time claim value

Property-question filtering:

Use property label/description and question text to rank or filter candidates.

Example:

question:
  What is the height of this river in metre?

gold entity:
  Q214924 / Rhine Falls

answer_eval:
  {"wikidata": 23.0, "range": [20.7, 25.3]}

positive fact source:
  full Wikidata quantity claim, not KnowCoL local entity-object KG

8. FactHead training data

For reliable MVP construction, use train/val with gold QID.

Fact item input:

subject:
  QID
  title / label
  short text metadata if available

property:
  PID
  property description from KnowCoL relation file when available
  Wikidata property label/description when using full claims

object/value:
  if wikibase-item:
    object QID
    object label / aliases

  if quantity:
    amount
    unit QID if available
    normalized numeric value

  if time:
    timestamp
    precision

Positive:

the matched claim/triple answering the sample

Negatives:

same entity, different fact
same property, different subject/object or different value
same datatype, different value
same-code-prefix hard fact

9. What to report in experiments

Report fact recovery separately from final QA.

Recommended metrics:

Fact Recall@K:
  whether a matched positive fact appears in top-K retrieved facts

Positive construction coverage:
  how many samples have at least one high-confidence positive fact

Coverage by question type:
  String
  Numerical
  Time

Coverage by source:
  KnowCoL local KG
  full Wikidata claims
  not matched / text evidence needed

Important:

KnowCoL local KG coverage is not the same as positive fact coverage.
The real metric is whether the query-specific answer fact can be matched.

10. Immediate next implementation target

Do this first:

InfoSeek train/val only
gold QID from *_withkb.jsonl
qtype from infoseek_val_qtype.jsonl for val

Then build:

1. object QID -> label/alias cache
2. property PID -> label/description cache
3. full Wikidata claim cache for unique train/val entities
4. positive_fact_candidates.jsonl
5. fact coverage report by qtype and source

Do not spend time on test/human fact coverage until QID mappings are restored.