Lost in Translation: Investigating Systematic Discrepancies between Parallel English and Chinese Names of American Chinese Restaurants

Hang Jiang; Nanxi Liu; Tiantian Wang; Cassandra Overney; Rukun Zhang; Artemisia Luk; Diyi Yang; Deb Roy

cs.CLOct 26, 2026

Lost in Translation: Investigating Systematic Discrepancies between Parallel English and Chinese Names of American Chinese Restaurants

Authors: Hang Jiang, Nanxi Liu, Tiantian Wang, Cassandra Overney, Rukun Zhang, Artemisia Luk, Diyi Yang, Deb Roy,

Organizations: Massachusetts Institute of Technology · Northeastern University · Wellesley College · Stanford University

Abstract

Given that the Chinese are one of the oldest ethnic groups to immigrate to the US, their cuisine have become central to the US dining scene and an interesting avenue to study how Chinese immigrants have culturally adapted into American life. In that regard, previous research has focused on analyzing the English names that Chinese restaurants adopt, e.g., finding the use of familiar and typical words. To investigate how Chinese immigrants have culturally integrated in the US depending on social, geographic and demographic contexts, we investigate in this work the under-explored relationship between the Chinese and English names of American Chinese restaurants. Specifically, we conduct a study on 3,172 Chinese restaurants (10% of the total on Google Maps) sampled across the US. We present 8-dimension name frames to capture the underlying meanings behind names and create the NameFrames dataset containing 884 annotated parallel Chinese/English names. We also built an NLP pipeline to analyze the semantic discrepancies between names with both frame-based and embedding-based distance metrics. Our frame analysis shows that English names more often emphasize Location and Specialty, while Chinese names more often emphasize Positivity and Ambiance. Furthermore, we observe that for Chinese restaurants located in urban areas, those in counties with a higher concentration of Chinese population, or those that are relatively new, the discrepancies between the English and Chinese names are noticeably larger. Finally, interviews with 10 Chinese restaurant owners reveal additional interesting insights regarding naming conventions.

Figures & tables

Figure 1: An example of mismatched English and Chinese ACR names. We hired Prolific workers to transcribe the names by browsing photos on Google Maps. The English name encodes the Personal Name frame, while the Chinese name encodes the Positivity frame.

Frame	Simplified Definition	Selected Examples
Personal Name	first or last names with or without possessives	Lin’s Kitchen , P.F. Chang’s
Specialty	special dishes, cooking methods, or ingredients	Peking Duck House , Noodle Galaxy
Positivity	superlatives or other positive modifiers	Happy Lamb , Best Chinese Buffet
Culture	references to cultural associations	Panda Express , The Phoenix Club
Location	place names or regional origins	China Garden , Oriental Kitchen
Ambiance	dining atmosphere or intimate words of family or relationship	Jade Garden , Friendship BBQ

Table 1: This table shows 8 frames with their simplified definitions and selected examples. The bold text spans indicate where the corresponding frame is encoded in the names. “Ding Tai Fung” means “wok, peace, and abundance”. “Hai Di Lao” means “catch from the deep ocean”. See full details of the frames in Appendix A.3 .

Model	Personal Name	Speciality	Positivity	Culture	Location	Ambiance	Creativity	Romanization	Average
Chinese Names (Weighted F1)
Expert Rules	$88.09$	$97.39$	$90.81$	$94.79$	$88.18$	$93.44$	$95.20$	$-$	$92.56$
Qwen3.7-Max	$95.14$	$\mathbf{98.86}$	$92.38$	$94.15$	$92.80$	$\mathbf{96.08}$	$96.89$	$-$	$95.18$
Kimi K2.6	$94.23$	$97.42$	$88.64$	$94.93$	$94.37$	$93.46$	$\mathbf{97.05}$	$-$	$94.30$
GPT-5.5	$\mathbf{96.01}$	$98.02$	$\mathbf{92.68}$	$95.38$	$93.99$	$94.71$	$96.88$	$-$	$\mathbf{95.38}$
Claude Sonnet 4.6	$94.83$	$97.75$	$90.13$	$\mathbf{96.66}$	$\mathbf{95.20}$	$94.95$	$96.88$	$-$	$95.20$

Table 2: Performance of different approaches on Name Frame Detection in both English and Chinese ACR names. “Expert Rules” represents the rule-based system based on lexicons and heuristics. Qwen3.7-Max, Kimi K2.6, GPT-5.5, and Claude Sonnet 4.6 represent the best performance of each model across two prompt settings: (1) zero-shot, (2) few-shot, as described in Section 4.1 .

Figure 2: Percentage of unique English and Chinese names tagged with each frame.

Panel of Figure 3 — Figure 3: These three figures show the semantic distance scores between parallel English and Chinese names across different location, Chinese percentage, and opening date groups. Jaccard distance is the frame-based $FD$ score and Cosine distance is translation-based embedding cosine distance with all-mpnet-base-v2 .

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

state	sample_frequency	national_frequency
California	357	4265
New York	260	2979
Texas	165	1990
Pennsylvania	153	1830
New Jersey	130	1546
Florida	123	1596

Appendix

Table 6

	Chinese		English
	Cohen’s $\bar{\kappa}$	Fleiss $\bar{\kappa}$	Cohen’s $\bar{\kappa}$	Fleiss $\bar{\kappa}$
Personal Name	0.8235	0.8240	0.812	0.8119
Specialty	0.9034	0.9032	0.8294	0.8297
Positivity	0.7989	0.7989	0.8011	0.8002
Culture	0.8387	0.8381	0.8545	0.8542
Location	0.8856	0.8858	0.935	0.935

Appendix

Table 4: This table shows the inter-annotator agreement (IAA) scores in Cohen’s Kappa and Fleiss Kappa on the English and Chinese names of 1000 ACRs. Romanization is not applicable to Chinese names, so we used “-”.

Model	Personal Name	Speciality	Positivity	Culture	Location	Ambiance	Creativity	Romanization
Chinese Names (Weighted F1)
Population Avg.	$94.93$	$95.45$	$88.13$	$81.07$	$91.50$	$86.65$	$96.30$	$-$
Expert Rules	$88.09$	$97.39$	$90.81$	$94.79$	$88.18$	$93.44$	$95.20$	$-$
Qwen3.7-Max (zero-shot)	$95.14$	$97.55$	$86.49$	$64.45$	$88.25$	$77.50$	$96.89$	$-$
Qwen3.7-Max (few-shot)	$94.99$	$\mathbf{98.86}$	$92.38$	$94.15$	$92.80$	$\mathbf{96.08}$	$96.46$	$-$
Kimi K2.6 (zero-shot)	$94.06$	$94.29$	$81.48$	$62.38$	$87.82$	$73.86$	$96.66$	$-$

Appendix

Table 5: Performance of different approaches on Name Frame Detection in both English and Chinese ACR names. “Expert Rules” represents the rule-based system based on lexicons and heuristics. We report performance for Qwen3.7-Max, Kimi K2.6, GPT-5.5, and Claude Sonnet 4.6 under two prompt settings: (1) “zero-shot”, (2) “few-shot”, respectively, as described in Section 4.1 .

	Google Translate	GPT-5.5	Claude Sonnet 4.6
Cosine	$0.6550$	$0.7598$	$\mathbf{0.7706}$
BERT F1	$0.9198$	$0.9404$	$\mathbf{0.9432}$
BERT Precision	$0.9161$	$0.9417$	$\mathbf{0.9434}$
BERT Recall	$0.9241$	$0.9392$	$\mathbf{0.9431}$
ROUGE-1	$0.4722$	$\mathbf{0.5975}$	$0.5920$
ROUGE-2	$0.3003$	$0.3692$	$\mathbf{0.3781}$

Appendix

Table 6: Translation quality comparison across Google Translate, GPT-5.5, and Claude Sonnet 4.6.

Respondent	Gender	English Proficiency	Customer Base	Area	Opening Year	Name Match
R1	Female	Proficient	Weekdays: 20% Chinese; Weekends: 50%	Urban	2017	Close match
R2	Female	Advanced	50% Chinese	Urban	2011	Mismatch
R3	Female	Intermediate	45% Chinese	Urban	2000	Mismatch
R4	Female	Intermediate	Weekdays: 50% Chinese; Weekends: 75%	Urban	2003	Close match
R5	Male	Basic	80% Chinese	Urban	1993	Mismatch
R6	Male	Proficient	40% Chinese	Urban	2010	Mismatch

Appendix

Table 7: Information about 10 ACR respondents that we interviewed. Each respondent is an owner, co-owner, or founder of an ACR in Massachusetts. The gender, English proficiency, percentage of Chinese customers, whether the restaurant is located in an urban or rural area, and the closeness between the Mandarin and English names are reported.

Figure 5: Qualtrics survey English instructions

Figure 6: Qualtrics survey Chinese instructions

Figure 7: Qualtrics survey response instructions

Figure 8: Qualtrics survey sample question

Panel of Figure 9 — Figure 9: Demographic characteristics of Prolific workers

	Jaccard Distance		Cosine Distance
Predictor	$\beta$ (SE)	$p$	$\beta$ (SE)	$p$
Intercept	0.5706 (0.010)	0.000	0.5165 (0.007)	0.000
Chinese population % (scaled)	0.0128 (0.010)	0.199	-0.0085 (0.007)	0.195
Time since opening (scaled)	-0.0128 (0.010)	0.194	-0.0179 (0.007)	0.006
Location (rural)	-0.1682 (0.069)	0.014	0.0181 (0.045)	0.688
Location (Suburb)	-0.0388 (0.039)	0.323	-0.0360 (0.026)	0.165

Appendix

Table 8: Multi-factor regression results

DV	CV $R^{2}$	Train / Test $R^{2}$
Jaccard distance	$-0.0782$	$-0.0164$ / $-0.0131$
Cosine distance	$-0.0040$	$-0.0294$ / $-0.0761$

Appendix

Table 9: Random Forest regression results.

Semantic Alignment

See all 17 papers

CulturalMenuBench: Probing the Knowledge-Application Gap in Multimodal Culinary Reasoning
Sep 3, 2026Bo Zeng, Linfeng Gao, Peiqin Lin +9Multimodal Reasoning Multilingual VLM Evaluation
Alibaba Group · Xiamen University · University of Hamburg
CIBuzzBench: A Benchmark for Cross-Lingual Understanding of Chinese Internet Buzzwords
Sep 18, 2026Yifan Wang, Junyu Lu, Qifan Wang +6Cross-Cultural Language Model Evaluation Cultural Knowledge in Language Models
People’s Public Security University of China · Dalian University of Technology · Meta AI +3
PragAlign: Evidence-Sensitive Reply Assistance Across Chinese and Japanese Appropriateness Judgments
Aug 27, 2026Xin Zhong, Satori HachisukaPragmatic Reasoning in Language Models Cultural Alignment
The University of Tokyo

Sep 3, 2026J/K move · Enter open · S save

CulturalMenuBench: Probing the Knowledge-Application Gap in Multimodal Culinary Reasoning

Bo Zeng, Linfeng Gao, Peiqin Lin, Yu Zhao+8

Alibaba Group · Xiamen University · University of Hamburg

Multimodal Reasoning Multilingual VLM Evaluation VLM Reasoning Cultural Knowledge in Language Models Multimodal Model Evaluation

Open paper AskPDF arXiv

Multimodal language models achieve near-ceiling scores on food recognition benchmarks, yet it remains unclear whether this success reflects genuine cultural understanding or mere visual matching. To probe this distinction, we introduce CulturalMenuBench, a benchmark of 4,870 items in 10 languages across 18 regions; its 10 tasks pair final-dish and step-by-step cooking images with ingredients, procedural text, and regional labels, spanning basic recognition to process-grounded cultural attribution. Evaluating 12 models exposes a substantial knowledge-application gap: models exceeding 94% on standard multiple-choice tasks drop to at most 56% when attributing dishes to Chinese regional cuisines, despite an identical four-way format. Diagnostic analyses explain why: error patterns are consistent with random guessing, accuracy tracks visual distinctiveness rather than cultural structure, and models classify cuisines more accurately from dish names alone than from images (+7-18 points). The knowledge is thus present but cannot be activated through visual input. An ablation confirms these tasks genuinely require procedural evidence: removing sequential cooking images selectively degrades process-grounded tasks while others remain stable. Overall, CulturalMenuBench shows that near-perfect recognition can conceal an inability to apply cultural knowledge, motivating training that explicitly connects perception, procedure, and cultural context. Code and data are publicly available.

Abstract

Semantic Alignment

Explore similar work

CulturalMenuBench: Probing the Knowledge-Application Gap in Multimodal Culinary Reasoning

CIBuzzBench: A Benchmark for Cross-Lingual Understanding of Chinese Internet Buzzwords

PragAlign: Evidence-Sensitive Reply Assistance Across Chinese and Japanese Appropriateness Judgments