Dip Benchmark Eval

---
name: dip-benchmark-eval
description: Evaluates machine translation systems on four discourse phenomena: anaphora resolution, lexical consistency, coherence/readability, and discourse connectives. It probes whether context-aware models can maintain discourse-level quality and consistency across different language pairs beyond standard n-gram overlap. Use when the user wants to benchmark on DiP Benchmark, or asks about evaluating this task. Reports BLEU.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2004.14607
  bibtex_key: jwalapuram2020dipbenchmark
  confidence: medium
---

# dip-benchmark-eval

> Can Your Context-Aware MT System Pass the DiP Benchmark Tests? : Evaluation Benchmarks for Discourse Phenomena in Machine Translation — Jwalapuram et al. (2020) (arXiv:2004.14607, 2020)

## What this evaluates

Evaluates machine translation systems on four discourse phenomena: anaphora resolution, lexical consistency, coherence/readability, and discourse connectives. It probes whether context-aware models can maintain discourse-level quality and consistency across different language pairs beyond standard n-gram overlap.

## Datasets

- **DiP Benchmark** — total ?; splits: test (-1)

## Metrics

- `BLEU` **(primary)** — range: percent
  - Standard n-gram overlap metric between reference and hypothesis translations. Computed as geometric mean of modified n-gram precisions (typically 1-4 grams) with a brevity penalty applied if the hypothesis is shorter than the reference. The paper uses it as a baseline that often fails to capture discourse-level deficiencies.

## Input / output format

**Input**: Source sentence paired with preceding context (automatically extracted from WMT system outputs), target language.

**Output**: Translated target sentence.

## Scoring recipe

```python
def compute_bleu(predictions, references):
    # Standard sentence-level or corpus-level BLEU calculation
    # Counts matching n-grams (typically up to 4-grams) between hypothesis and reference
    # Applies brevity penalty if hypothesis length < reference length
    # Returns score in [0, 100] percent
    return standard_bleu_score(predictions, references)
```

## Common pitfalls

- Test sets are automatically generated from existing WMT system errors, which may bias evaluation towards specific error types rather than general discourse quality.
- BLEU scores may not capture discourse-level deficiencies, leading to misleading performance assessments for context-aware models.
- French test sets are smaller and based on older data (2013-2015) due to WMT discontinuation, making cross-lingual comparisons uneven.

## Evidence (verbatim from paper)

> We include French since Fr-En is a popular translation pair that results in some of the highest BLEU scores. Other source languages that are part of WMT can be extracted as needed; the testsets can also be expanded if older data were to be considered.

## Citation

```bibtex
@misc{jwalapuram2020dipbenchmark,
  title={Can Your Context-Aware MT System Pass the DiP Benchmark Tests? : Evaluation Benchmarks for Discourse Phenomena in Machine Translation},
  author={Jwalapuram et al. (2020)},
  year={2020},
  note={arXiv:2004.14607}
}
```

- arXiv: 2004.14607