💻 computer science

ANCHOR: An English-to-Chinese Transcreation Method for Transferring Check-Worthiness Estimation Data

The paper proposes ANCHOR, an adaptive transcreation method that classifies cultural dependence to either domesticate or foreignize English check-worthiness data, effectively transferring annotated resources to Chinese while preserving factual scope, assertion strength, and label accuracy.

Original authors: Yinglong Yu, Fang Yuan, Yu Huang, Yuchang Duan, Zhengyi Lyu, Yu Wang

Published 2026-10-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yinglong Yu, Fang Yuan, Yu Huang, Yuchang Duan, Zhengyi Lyu, Yu Wang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the digital age, the internet is a constant stream of claims, from political speeches to viral social media posts. Before we can verify if a statement is true or false, we must first decide if it is even worth checking. This initial step, known as "check-worthiness estimation," acts as a filter, sorting out the noise to focus human and machine effort on the statements that matter most. For years, researchers have built sophisticated tools to perform this sorting in English, creating large datasets where experts have labeled thousands of sentences as either "check-worthy" or "not worth checking." However, a significant gap exists: these tools and datasets are largely silent in other languages, particularly Chinese. Simply translating these English examples into Chinese seems like an obvious solution, but language is not a simple code to be swapped. A phrase that sounds like a harmless joke in one culture might be a serious accusation in another, and a vague complaint in English could accidentally become a specific, verifiable fact when translated. If the translation changes the nature of the claim, the original label becomes useless, and the data is broken.

To solve this problem, a team of researchers from the Communication University of China developed a new method called ANCHOR. Their goal was not just to translate text, but to "transcreate" it—a process that adapts the cultural flavor of a sentence while strictly preserving its core meaning and its original status as a claim. The researchers recognized that not all sentences require the same treatment. Some English texts are deeply rooted in specific cultural contexts, using idioms, political metaphors, or regional references that a direct translation would confuse a Chinese reader. Others are straightforward, dealing with universal facts or emotions that do not depend on a specific background. ANCHOR acts as a traffic director, first analyzing each sentence to determine its cultural dependence. If a sentence is highly culture-dependent, the system takes a "domestication" path, rewriting the expression to fit Chinese cultural norms while carefully ensuring that no new facts are invented and no existing facts are lost. If the sentence is culturally neutral, it takes a "foreignization" path, using a library of professional bilingual examples to guide a translation that stays close to the original structure.

The team tested this approach on three major English datasets containing thousands of sentences about politics, social media, and public health. They compared ANCHOR against eight other translation models, including major commercial services and open-source systems. The results were striking. While standard translation models often failed to preserve the original label—sometimes turning a non-factual opinion into a factual claim, or erasing the specific target of a claim—ANCHOR succeeded in keeping the original label applicable in 99.38% of cases. This performance tied with the top-performing commercial model, DeepL, but ANCHOR achieved this while specifically handling the tricky cultural nuances that usually trip up automated systems. In a direct comparison against a powerful AI model translating without any special instructions, ANCHOR improved the success rate by nearly two percentage points. More importantly, it made far fewer mistakes with proper names, numbers, and specific entities, which are the building blocks of any fact-checking claim.

The researchers also looked closely at what happens when the system makes a mistake. They found that the most common errors involved adding extra conclusions or changing the direction of an argument, such as turning a statement about something being "under control" into a claim that someone is "in control." These subtle shifts can completely alter whether a statement needs to be verified. By breaking down the process into two distinct paths based on cultural need, the team ensured that the system knew when to adapt and when to stay literal. For the culturally complex sentences, human evaluators rated the translations highly for meaning accuracy and fluency, confirming that the system could sound natural without sacrificing the truth. The study demonstrates that reusing English data for Chinese fact-checking is possible, but it requires a method that respects the boundaries of the original claim. ANCHOR provides a practical blueprint for this, showing that with the right constraints, we can build robust resources for languages that have historically been left behind in the fight against misinformation.

Technical Summary: ANCHOR – An English-to-Chinese Transcreation Method for Transferring Check-Worthiness Estimation Data

Problem Statement
Translating English annotated data into Chinese offers a resource-efficient path for building check-worthiness estimation datasets, a critical upstream task in automated fact-checking. However, standard machine translation often fails to preserve the "check-worthiness" label of the source text. Translation can alter textual properties by omitting entities, shifting claim scopes, or adding information, rendering the original label inapplicable. Furthermore, culturally loaded expressions (e.g., political metaphors, regional references, idioms) pose a specific challenge: literal translation may obscure the original function for Chinese readers, while excessive rewriting risks altering factual scope, assertion strength, or pragmatic function. Existing methods often struggle to balance semantic preservation, cultural adaptation, and label consistency simultaneously.

Methodology: ANCHOR
The authors propose ANCHOR (Adaptive Naturalization for Check-worthiness-Oriented Rewriting), a transcreation pipeline designed to transfer English check-worthiness data into Chinese while ensuring the source label remains applicable. The method operates on a dual-path architecture determined by cultural dependence classification:

Cultural Dependence Classification:
- An LLM (GPT-5.2) classifies input texts as either High or Low cultural dependence.
- High Dependence: Texts containing culture-specific metaphors, regional references, or idioms that require adaptation for target-language intelligibility.
- Low Dependence: General emotional expressions or factual claims understandable across cultures.
- This step routes texts to either a Domestication or Foreignization path.
Foreignization Path (Low Cultural Dependence):
- Designed to preserve context, register, and semantic boundaries with minimal alteration.
- Retrieval: Utilizes a self-built parallel corpus of 4,987 bilingual entries from Chinese Ministry of Foreign Affairs press conferences (2024–2025).
- Hybrid Retrieval: Employs a two-stage retrieval system combining dense vector similarity (using text-embedding-3-large) with sparse and multi-vector scores to select top- $k$ bilingual demonstrations.
- Generation & Selection: An LLM (Qwen-3 Max) generates $n=5$ candidate translations based on retrieved demonstrations. A self-supervised reranker (frozen sentence encoder + MLP scoring head), trained on perturbed target-side texts, selects the highest-scoring candidate.
Domestication Path (High Cultural Dependence):
- Designed to adapt expressions while strictly bounding the scope of rewriting.
- Constraints: The generation prompt (using GPT-5.2) enforces four rules:
  1. Retain existing verifiable facts; do not introduce new factual claims.
  2. Preserve claim targets (entities, numbers, times, causal relations) and assertion scope.
  3. Maintain emotional tone and pragmatic function without expanding opinions into verifiable details.
  4. Adapt cultural expressions to functionally equivalent Chinese expressions while retaining the factual elements above.

Key Contributions

ANCHOR Framework: A novel transcreation method that dynamically routes texts based on cultural dependence, combining demonstration-assisted foreignization with constraint-based domestication to preserve check-worthiness labels.
Evaluation Scheme: A dual-evaluation approach separating general translation quality (measured by COMET-QE, COMETKiwi, and LLM Score) from check-worthiness preservation (human judgment on whether the translation supports the original claim). This design explicitly distinguishes between claim loss and the rewriting of non-factual expressions into facts.
Empirical Validation: Comprehensive baseline and ablation experiments demonstrating that ANCHOR maintains high label preservation rates while achieving competitive general translation quality.

Results

Label Preservation: In baseline experiments against seven translation models (including Baidu, Google, DeepL, and NMT models), ANCHOR achieved a label preservation rate of 99.38%, tying with DeepL for the highest among all models. Notably, ANCHOR preserved 100.00% of negative samples (non-check-worthy claims), outperforming DeepL (99.44%) and significantly outperforming models like OPUS-MT (98.33%) and NLLB-600M (97.41%).
Ablation vs. Direct Translation: Compared to GPT-5.2 direct translation, ANCHOR improved the preservation rate by 1.88 percentage points (97.19% vs. 95.31%). This gain was driven by fewer errors in handling proper names, quantities, and entities (9 failed units for ANCHOR vs. 15 for direct translation).
General Translation Quality: ANCHOR ranked second in LLM Score (90.63) among eight models, slightly below DeepL (91.78). However, the authors note that general quality metrics (COMET-QE, COMETKiwi) do not perfectly correlate with label preservation; for instance, Google Translate had higher COMET-QE scores but lower label preservation.
Path Performance: The foreignization path performed competitively with DeepL on NLP4IF data. The domestication path showed lower scores on general quality metrics compared to direct translation, attributed to the necessary rewriting of cultural expressions which reduces sentence-level correspondence.
Human Evaluation: Cultural dependence classification achieved 0.8225 accuracy (with 100% recall for high-dependence texts). Domesticated translations received a mean human rating of 4.386/5 for meaning accuracy, cultural relevance, and fluency.

Significance and Claims
The paper claims that ANCHOR offers a practical way to reuse English annotated resources for building Chinese check-worthiness datasets. By explicitly addressing the tension between cultural adaptation and factual preservation, the method mitigates the risk of "label drift" where translation alters the verifiability or scope of a claim. The authors emphasize that for check-worthiness data transfer, label preservation is the primary objective, and general translation quality serves as a secondary indicator of linguistic usability. The study demonstrates that a generation pipeline constrained by factual and pragmatic boundaries can successfully transfer annotated data across languages with high fidelity to the original labeling intent, a capability that standard translation models and direct LLM translation lack. The work highlights that maintaining the correspondence between texts and source labels requires specific strategies (like cultural routing and constraint-based rewriting) rather than relying solely on high-quality translation metrics.

More like this