PatSnap/patent-translation-bench

PatSnap Patent Translation Bench

English | 中文

A benchmark for evaluating patent machine translation systems, covering both Chinese-to-English (CN→EN) and English-to-Chinese (EN→CN) directions. It assesses translation quality across six dimensions: translation accuracy, terminology accuracy, terminology consistency, patent writing conventions, hallucination, and omission.

Dataset Overview

Attribute	Value
Total samples	881
Translation directions	CN→EN / EN→CN
Rows with both Chinese and English text	854
Rows without an English reference (content_en: Zero)	27
IPC coverage	All 8 sections: A / B / C / D / E / F / G / H
Text granularity	Word, character, sentence, paragraph, document
License	CC BY-NC 4.0

Use Cases

This bench evaluates patent translation systems on the following capabilities:

1. Translation accuracy: Semantic alignment with the reference translation at word, character, sentence, paragraph, and document granularity
2. Terminology accuracy: Whether patent-specific terms are translated correctly
3. Terminology consistency: Whether the same term is translated consistently throughout the text
4. Patent writing conventions: Whether the translation conforms to patent document writing norms
5. Hallucination detection: Whether the translation introduces content not present in the source (e.g., source-language characters mixed in, abnormal length inflation)
6. Omission detection: Whether the translation is abnormally shorter than the source

Data Fields

Field	Type	Description
pn	string	Patent publication number (PatSnap normalized PN)
ipc	string	IPC top-level class (A–H; Zero if unclassified)
content_cn	string	Chinese source text
content_en	string	English source text (reference answer for CN→EN; Zero if unavailable)
label_1	string	Text granularity: word / character / sentence / paragraph / document
label_2	string	Evaluation dimension (see table below)
label_3	nullable string	Text context, such as abstract or claims; Zero or null if not applicable
special_cn	nullable string	Chinese special terms (used for professional metrics); Zero or null if none
special_en	nullable string	English special terms (used for professional metrics); Zero or null if none
domain	nullable string	Domain label; Zero or null if none

label_2 Values

label_2	Category	Description	Count
terminology_accuracy	Professional	Whether patent terminology is accurately translated	205
terminology_consistency	Professional	Whether the same term is translated consistently	124
normal_sentence	General accuracy	Translation quality of regular sentences	129
normal_character	General accuracy	Translation quality of regular characters/words	125
paragraph_accuracy	General accuracy	Paragraph-level translation quality	110
special_character	Professional	Whether special symbols/characters are handled correctly	80
special_sentence	Professional	Whether special sentence structures are translated correctly	57
patent_writing_norm	Professional	Whether the translation conforms to patent writing norms	19
title_writing_norm	Professional	Whether the translation follows patent title writing norms	8
document_accuracy	General accuracy	Full-document translation quality	24

Data Distribution

By Text Granularity (label_1)

Granularity	Count	Ratio
Word (词)	329	37.3%
Character (字)	205	23.3%
Sentence (句)	194	22.0%
Paragraph (段)	110	12.5%
Document (篇)	43	4.9%

By IPC Technical Domain

IPC	Domain	Count
H	Electricity	89
G	Physics	90
C	Chemistry; Metallurgy	70
B	Performing Operations; Transporting	66
A	Human Necessities	65
D	Textiles; Paper	52
F	Mechanical Engineering; Lighting; Heating; Weapons; Blasting	52
E	Fixed Constructions	49
—	No IPC label	348

Evaluation Metrics

Accuracy Metrics (General translation quality — applies to label_2: normal_sentence / normal_character / paragraph_accuracy / special_sentence / document_accuracy)

BLEU (Bilingual Evaluation Understudy)

Measures translation accuracy and fluency by computing n-gram precision between the hypothesis and reference, with a brevity penalty. Scores range from 0 to 1; higher is better.

SacreBLEU

A standardized BLEU implementation using consistent tokenization and smoothing methods, ensuring reproducible and comparable results across different research groups.

METEOR (Metric for Evaluation of Translation with Explicit ORdering)

Considers stemming, synonyms, and word order; more sensitive to semantics than BLEU. Aligns tokens via exact match, stem match, and synonym match, then computes the harmonic mean of precision and recall with a fragmentation penalty.

ROUGE (Recall-Oriented Understudy for Gisting Evaluation)

Evaluates semantic similarity via n-gram overlap. ROUGE-N measures n-gram overlap; ROUGE-L measures the longest common subsequence (LCS).

Metric	Description	Direction
BLEU-1 / 2 / 4	n-gram precision	CN→EN
SacreBLEU	Standardized corpus-level BLEU	CN→EN / EN→CN (document-level)
METEOR	n-gram metric with stemming and synonym matching	CN→EN / EN→CN
ROUGE-1 / 2 / L	Recall-oriented n-gram overlap	EN→CN

Composite score calculation:

• Regular sentences / characters / paragraphs / special sentences: score = (BLEU-1 + METEOR) / 2 (CN→EN); score = (ROUGE-1 + METEOR) / 2 (EN→CN)
• Document accuracy: score = SacreBLEU

Hallucination and omission metrics are computed only for entries whose label_2 is: normal_sentence, normal_character, paragraph_accuracy, special_sentence, document_accuracy.

Professional Metrics

Terminology Accuracy

Ensures professional terms are translated correctly, avoiding ambiguity.

label_2	Calculation	Description
terminology_accuracy	Whether special_en (CN→EN) / special_cn (EN→CN) appears in the translation (0/1)	ACC: whether the terminology is correctly translated

Terminology Consistency

The same term should be translated consistently throughout the document.

label_2	Calculation	Description
terminology_consistency	Whether the term appears ≥ 2 times in the translation (0/1)	ACC: consistency of term translation across different parts (only counted when term is correctly translated)

Special Characters

Whether the translation correctly preserves special symbols/characters from the source.

label_2	Calculation	Description
special_character	Whether special_en (CN→EN) / special_cn (EN→CN) appears in the translation (0/1)	ACC: whether special symbols/characters are correctly preserved

Patent Writing Conventions

Whether the translation meets USPTO patent writing requirements.

label_2	Calculation	Description
patent_writing_norm	Ratio of matched convention phrases (0–1)	ACC: whether patent section names (abstract, claims, technical field, background, summary, detailed description, etc.) are correctly translated

Hallucination and Omission Metrics

All values are reported as percentages (%) — lower is better.

Omission

Whether the translation is complete. Computed only for entries whose label_2 is normal_sentence, normal_character, paragraph_accuracy, special_sentence, or document_accuracy.

Metric	Calculation
Omission rate (%)	count(translation length / reference length < 0.5) / total × 100

Hallucination

Whether the model output contains hallucinated content. Computed only for entries whose label_2 is normal_sentence, normal_character, paragraph_accuracy, special_sentence, or document_accuracy.

Metric	Calculation
Length hallucination rate (%)	count(translation length / reference length > 5) / total × 100
Source-language leakage rate — CN→EN (%)	count(translation contains Chinese characters) / total × 100
Source-language leakage rate — EN→CN (%)	count(translation contains English letters AND source has no English) / total × 100

Length ratio for CN→EN is computed in words; for EN→CN in characters.

Dataset Construction

1. Data Sources

Bilingual (CN/EN) patent text pairs were collected from the PatSnap patent database across all eight IPC sections (A–H). Text sources include patent abstracts, claims, and all description sections (background, summary, brief description of drawings, detailed description).

2. Stratified Annotation

Samples were annotated by text granularity (word / character / sentence / paragraph / document) and evaluation dimension (general accuracy / terminology accuracy / terminology consistency / patent writing conventions / special characters / special sentences) to ensure sufficient coverage of each dimension.

3. Professional Annotation

For samples containing patent-specific terms, special characters, or writing conventions, the special_cn / special_en fields were manually annotated for use in exact-match professional metrics.

4. Using the bilingual pairs

For evaluation that requires both languages, use the 854 rows whose content_en is not Zero. Keep the 27 remaining rows if your task uses their Chinese text or annotations.

Evaluation Example

import json, sys

sys.path.insert(0, "../common/metrics")
from translation_metrics import evaluate

# Add translation result field to each record in test_dataset.jsonl
# CN→EN: add content_cn_translate field
# EN→CN: add content_en_translate field

summary = evaluate("your_results.jsonl", direction="cn2en")
print(json.dumps(summary, ensure_ascii=False, indent=2))

Example output:

{
  "direction": "cn2en",
  "total": 854,
  "accuracy_by_label": {
    "normal_sentence": 72.34,
    "terminology_accuracy": 85.10,
    "...": "..."
  },
  "hallucination_pct_by_label": {
    "normal_sentence": 1.20,
    "...": "..."
  },
  "miss_translation_pct_by_label": {
    "normal_sentence": 0.80,
    "...": "..."
  }
}

CLI usage:

python ../common/metrics/translation_metrics.py \
    --input your_results.jsonl \
    --direction cn2en \
    --output result_cn2en.json

Score Grade Reference (composite score, as percentage)

Grade	Score	Description
A	≥ 80	Excellent — ready for professional patent use
B	≥ 65	Good — suitable as a translation aid
C	≥ 50	Acceptable — key terms require manual review
D	< 50	Below standard — model improvement recommended

Citation

If you use this dataset, please cite:

@dataset{patsnap_patent_translation_bench_2026,
  title={PatSnap Patent Translation Bench},
  author={PatSnap},
  year={2026},
  note={A benchmark for evaluating patent machine translation systems, covering CN↔EN bidirectional translation}
}

License

This dataset is released under CC BY-NC 4.0 for research and non-commercial evaluation purposes.

Try the Production System

Experience PatSnap AI Translation — the patent translation system evaluated by this bench, offering CN↔EN translation for patent documents from major patent offices worldwide.

Try it now: PatSnap Eureka