EVSeManip

EVSeManip is an English–Vietnamese benchmark for detecting simulated semantic manipulation in translations of U.S. presidential speeches. The paper describes 7,604 aligned triplets from 980 speeches, giving 22,812 labeled instances. Each triplet contains an English source excerpt, a Vietnamese faithful translation, and one low-severity and one high-severity manipulated Vietnamese variant. The final triplets passed unanimous semantic review by three LLM judges (DeepSeek V4 Pro, Mistral Small 2603, and Qwen3 30B Instruct).

Data files

Configuration	Split	Rows	Contents
benchmark (default)	test	22,812	One English source and Vietnamese candidate per row, with a three-class gold_label
triplets	test	7,604	The aligned faithful, low-severity, and high-severity translations with edit metadata

The test split represents the full benchmark evaluated in the paper. No official training or validation split is provided. The two configurations are alternative views of the same examples; do not combine them as independent observations. If creating new splits, group by speech_id to keep chunks of one speech together.

The labels are Faithful, Low-severity, and High-severity, with 7,604 instances in each class. Low severity means an edit changes meaning without changing the speaker's main position or message. High severity means an edit changes important information or a central statement. The labels and edit techniques reflect the controlled generation and LLM review process; they are not human annotations.

Fields

The benchmark configuration has:

• instance_id: unique variant identifier (triplet_id plus variant suffix)
• triplet_id, speech_id: grouping identifiers
• source_url: canonical Miller Center page for the original speech (the source records used an older Acquia host)
• english_source: source excerpt
• vietnamese_candidate: translation being classified
• gold_label: one of the three labels above
• technique, source_span_en, original_span_vi, replacement_span_vi: edit information; empty strings for faithful instances

The triplets configuration contains the same source identifiers and english_source, plus vietnamese_faithful, vietnamese_low_severity, vietnamese_high_severity, and separate low_* and high_* technique and span fields.

Construction and intended use

English excerpts were taken from the Miller Center Presidential Speeches archive. TranslateGemma-4B generated the faithful Vietnamese translations. Gemma3:12B proposed one local span edit per manipulated variant, using six low-severity or six high-severity techniques. Structural checks, Qwen3:8B validation, and the three-judge unanimous filter produced this release. See the accompanying paper for the full methodology and baseline results.

This dataset is intended for research and evaluation of source-grounded translation integrity and severity classification. The manipulated Vietnamese text is intentionally inaccurate and should not be quoted as an authentic translation of a president's speech. The source is limited to U.S. presidential speeches, the language pair is English–Vietnamese, and some edit techniques have very few examples. Model-judge acceptance does not replace human validation.

Source attribution and licensing

Miller Center of Public Affairs, University of Virginia. "Presidential Speeches: Downloadable Data." data.millercenter.org. The Miller Center states that the speeches in its downloadable collection are in the public domain and requests this citation. Each record also links to its source speech. The repository's Apache-2.0 metadata applies to the contributed dataset packaging and annotations; it does not claim ownership of the source speeches.

Loading

from datasets import load_dataset

benchmark = load_dataset("nblong415/EVSeManip", "benchmark", split="test")
triplets = load_dataset("nblong415/EVSeManip", "triplets", split="test")

Citation

Please cite the paper An English–Vietnamese Benchmark for Simulated Semantic Manipulation in Political Speech Translation and the Miller Center source attribution above. A paper DOI or archival link can be added here when available.