Ncg Eval

---
name: ncg-eval
description: Probes a model's ability to perform fine-grained information extraction from scholarly NLP papers. It specifically tests the identification of contribution sentences, the extraction of scientific terms and entities, and the structuring of these elements into RDF-style triples organized under 12 predefined information units. Use when the user wants to benchmark on NLPContributionGraph, or asks about evaluating this task. Reports F1.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2106.07385
  bibtex_key: dsouza2021ncg
  confidence: high
---

# ncg-eval

> SemEval-2021 Task 11: NLPContributionGraph -- Structuring Scholarly NLP Contributions for a Research Knowledge Graph — D'Souza et al. (2021) (arXiv:2106.07385, 2021)

## What this evaluates

Probes a model's ability to perform fine-grained information extraction from scholarly NLP papers. It specifically tests the identification of contribution sentences, the extraction of scientific terms and entities, and the structuring of these elements into RDF-style triples organized under 12 predefined information units.

## Datasets

- **NLPContributionGraph** — total ?; splits: test (-1); repo https://github.com/sshailabh/SemEval-2021-Task-11

## Metrics

- `F1` **(primary)** — range: [0, 1]
  - Standard F1 score (harmonic mean of precision and recall) computed at sentence, entity, and triple levels, with an overall score aggregating performance across the end-to-end pipeline.

## Input / output format

**Input**: Plaintext scholarly article $A$ (and optionally gold contribution sentences $C_{sent}$ or entities $E$ depending on the evaluation phase).

**Output**: A set of contribution sentences $C_{sent}$, a set of entities $E$ (scientific terms and predicates), and a set of triples $T$ in (subject, predicate, object) format organized under 12 information units.

## Scoring recipe

```python
def compute_f1(pred_set, gold_set):
    tp = len(pred_set & gold_set)
    prec = tp / len(pred_set) if pred_set else 0
    rec = tp / len(gold_set) if gold_set else 0
    return 2 * prec * rec / (prec + rec) if (prec + rec) > 0 else 0
# Overall F1 aggregates sentence, entity, and triple F1s across the pipeline
```

## Common pitfalls

- Confusing the three evaluation phases (end-to-end vs. constrained entity/triple extraction).
- Failing to organize extracted triples under the correct 12 Information Units (IUs), which is required for KG building.
- Triple-level extraction is significantly harder than sentence-level, leading to low F1 scores despite decent sentence extraction.

## Evidence (verbatim from paper)

> Despite strong inter-annotator agreement at sentence level (F1=67.92%), triple-level performance remains low (F1=22.28%), highlighting the difficulty of automated semantic parsing and entity relation extraction in unstructured research narratives.

## Citation

```bibtex
@misc{dsouza2021ncg,
  title={SemEval-2021 Task 11: NLPContributionGraph -- Structuring Scholarly NLP Contributions for a Research Knowledge Graph},
  author={D'Souza et al. (2021)},
  year={2021},
  note={arXiv:2106.07385}
}
```

- arXiv: 2106.07385