markozubac/tripletsgoldantecedent

Pronoun Resolution in LLM Triplet Extraction: Human Evaluation Set

Dataset Summary

This dataset holds human-annotated gold standards for checking whether pronoun resolution during LLM-based knowledge-graph triplet extraction picks the right antecedent.

When an LLM pulls (Subject, Relation, Object) triplets from chunked text, it often produces pronoun entities such as he, she, it or they. These entities break the knowledge graph. Counting how many pronouns are left only shows whether a method replaced them, not whether the replacement is correct. This dataset measures the second question directly.

Each row is one baseline triplet that has a pronoun in its subject or object. Next to it are the matching triplets from one or more pronoun-resolution methods. A human annotator wrote down the true antecedent (Gold_Antecedent) and marked each method's substitution as correct (Y) or incorrect (N).

Config	Source corpus	Rows	Methods evaluated
hotpotqa	HotpotQA (dev, distractor setting)	400	Method 1, Method 2, Method 3, Method 4 (FastCoref)
novelqa	NovelQA (5 full-length novels)	300	Method 3

Methods Compared

Every method starts from the same baseline. The text is split into chunks and an LLM (Cohere command-a-03-2025) extracts triplets from each chunk on its own. If a triplet's subject or object contains a pronoun, the chunk is processed again with context:

Method	Strategy
Baseline	Direct triplet extraction from the chunk with no pronoun handling.
Method 1	An LLM first rewrites the chunk, replacing pronouns with their antecedents using the K = 2 previous chunks from the same document. Triplets are then extracted from the rewritten text.
Method 2	Re-extraction with raw context: the K = 2 previous chunks plus the current chunk go to the LLM in one prompt that tells it to resolve pronouns in the output triplets.
Method 3	Re-extraction with structured context: the triplets already extracted from the K = 2 previous chunks (not the raw text) are given as context for resolving pronouns in the current chunk.
Method 4	FastCoref (FCoref), a neural coreference model, rewrites the chunk using the previous chunks as context. Triplets are then extracted with the LLM. This is a non-LLM coreference baseline.

On NovelQA only the Baseline and Method 3 were run.

Dataset Structure

hotpotqa config

Column	Type	Description
case_id	int	Case identifier (1–400).
chunk_id	int	ID of the source text chunk (paragraph) that the triplet was extracted from.
question_id	int	HotpotQA question / source document the chunk belongs to (1-indexed position in hotpot_dev_distractor_v1.json).
Baseline_Subject, Baseline_Relation, Baseline_Object	str	Baseline triplet that contains the pronoun.
M{1..4}_Subject, M{1..4}_Relation, M{1..4}_Object	str	Matching triplet produced by Method 1–4.
M{1..4}_Resolution_Correct(Y/N)	str	Human judgment: Y if the method's substitution names the same real-world entity as the gold antecedent, otherwise N.
Gold_Antecedent	str	The true antecedent of the pronoun, set by the annotator.

novelqa config

Column	Type	Description
case_id	int	Case identifier (not contiguous; see Dataset Creation).
chunk_id	int	ID of the source chunk within the book.
book_id	str	NovelQA book ID (see table below).
Baseline_Subject, Baseline_Relation, Baseline_Object	str	Baseline triplet that contains the pronoun.
M3_Subject, M3_Relation, M3_Object	str	Matching triplet produced by Method 3.
M3_Resolution_Correct(Y/N)	str	Human judgment of Method 3's substitution.
Gold_Antecedent	str	The true antecedent of the pronoun.

book_id	Title	Author	Rows
B03	Gone with the Wind	Margaret Mitchell	66
B28	The History of Tom Jones, a Foundling	Henry Fielding	75
B39	Can You Forgive Her?	Anthony Trollope	74
B42	Les Misérables	Victor Hugo	82
B54	Middlemarch	George Eliot	3

Example (hotpotqa)

Baseline:  (She, voiced by, April Stewart)
Method 1:  (Wendy Testaburger, voiced by, April Stewart)   -> Y
Method 2:  (Wendy Testaburger, voiced by, April Stewart)   -> Y
Method 3:  (Wendy Testaburger, voiced by, April Stewart)   -> Y
Method 4:  (Wendy Testaburger, voiced by, April Stewart)   -> Y
Gold antecedent: Wendy Testaburger

Resolution Accuracy (from the annotations)

HotpotQA (N = 400)

Method	Correct	Accuracy
Method 1	382 / 400	95.50%
Method 2	392 / 400	98.00%
Method 3	395 / 400	98.75%
Method 4 (FastCoref)	308 / 400	77.00%

NovelQA (N = 300)

Method	Correct	Accuracy
Method 3	268 / 300	89.33%

Every HotpotQA case was judged for all four methods. That gives a fully paired design, so methods can be compared directly with paired tests such as McNemar's.

Dataset Creation

Source data

• HotpotQA
• NovelQA:

Sampling

• HotpotQA: for every baseline triplet with a pronoun in its subject or object, we manually paired it with the corresponding triplet produced by each of Methods 1–4 for the same source chunk. From the pool of cases that could be paired in all four methods, we drew a stratified sample of 400 cases, spread across different values of question_id so that the sample covers 398 distinct source documents (2 chunks are appearing two times).
• NovelQA: the same manual pairing procedure was applied between the Baseline and Method 3, and 300 cases were sampled across the five novels. Cases where Method 3 left the pronoun unresolved were removed and replaced. As a result, every row has Method 3 making some named substitution, so the NovelQA accuracy measures antecedent-selection correctness specifically.

Annotation

For each case, the annotator read the target chunk and up to three preceding chunks from the same document, and from that context recorded the true antecedent (Gold_Antecedent). Each method's output was marked:

• Y: the substituted entity refers to the same real-world entity as the gold antecedent.
• N: a wrong entity was substituted, the pronoun was left unresolved, or the substitution was too generic (e.g. album instead of the album's title).

Reflexive and possessive constructions whose referent is already clear from the sentence's own subject (e.g. "Mr Allworthy never used his power") were marked correct when the substitution kept that reference, even if the possessive was just dropped.

Licensing

The annotations are released under CC BY-SA 4.0, matching the HotpotQA license. Short text excerpts from the NovelQA novels appear inside triplets. Use of the underlying book texts is subject to the NovelQA benchmark's terms and to the copyright status of each work in your jurisdiction.

Citation

Please cite the source benchmarks if you use this dataset:

@inproceedings{yang2018hotpotqa,
  title     = {{HotpotQA}: A Dataset for Diverse, Explainable Multi-hop Question Answering},
  author    = {Yang, Zhilin and Qi, Peng and Zhang, Saizheng and Bengio, Yoshua and Cohen, William W. and Salakhutdinov, Ruslan and Manning, Christopher D.},
  booktitle = {Proceedings of EMNLP},
  year      = {2018}
}

@inproceedings{wang2025novelqa,
  title     = {{NovelQA}: Benchmarking Question Answering on Documents Exceeding 200K Tokens},
  author    = {Wang, Cunxiang and Ning, Ruoxi and Pan, Boqi and Wu, Tonghui and Guo, Qipeng and Deng, Cheng and Bao, Guangsheng and Hu, Xiangkun and Zhang, Zheng and Wang, Qian and Zhang, Yue},
  booktitle = {ICLR},
  year      = {2025}
}