CulturePark: Boosting Cross-cultural Understanding in Large Language Models

W1: ome of the comparisons may seem unfair, for example:

• Comparing SeaLLM (which is LLaMA-2-based) with GPT-3.5-based models in Table 2.
  • We agree that it is necessary to fine-tune gpt models on the training data of SeaLLM. However, a sad story is that their training data is not publicibly accessible. Moreover, we realize that it is never easy to reach a "fair" comparison since if we fine-tune the same models on their data, it is unfair for our approach since their pre-training data is significantly larger than us. We would like to claim that given limited budget and GPU hardware, CulturePark remains a cost-effective solution to fastly build a cultural-specific LLM for low-resource culture. This is the main contribution of the paper.
• In Section 5.1, it’s unclear if the data generated from GPT-4 is about the same size as the data generated from GPT-3.5, and if the same verification and diversification strategies are applied to the GPT-4-generated data.
  • Yah! Both the data size and process strategies are the same.
• While it is interesting to know that the proposed framework works for GPT-3.5 (and to LLaMA-2-70B to a certain extent), it would be great if the authors studied a range of open-source models.
  • Interesting advice! However, we can't implement our method on all open-source models due to time and computing resource limit. LLaMA is the one of the most popular and best open-source models. We think our experiments on LLaMA can verify the effectiveness of our method. And we will apply our method on more open-source models in the next version of our paper.

Q1: Currently, we lack rigorous studies on how accurately LLMs can simulate different cultures. Some analysis would be great, such as the percentage of data filtered after verification and human validation of the generated reasoning/perspectives, etc.

Yah! To explore how accurately LLMs can simulate different cultures, we conducted two new experiments. For the first experiment, we analysis the percentage of data filtered after verification. For the second experiment, we hire four human pariticipants to help us validate the generated dialogues.

For the first experiment, we generate 500 dialogues between the "Arabic delegate" and "Main Contact" agents. Each dialogue can be extracted 5-10 cultural specific opinions. In this experiment, we extracted 3725 cultural specific opinions in total. After factual verification, we get 3216 samples. Consequently, 86.34% of data can pass the verification.

For the second experiment, we generate 50 dialogues between the "Chinese delegate" and "Main Contact" agents, and 50 dialogues between the "Korean delegate" and "Main Contact" agents. And we hire two Chinese persons to check if they agree with the opinions of "Chinese delegate" and two Korean persons to check if they agree with the opinions of "Korean delegate". They were asked to score each dialogue with 1-10 ( 1-strongly disagree, 10-strongly agree). The table below shows the average score of generated dialogues in Chinese and Korean, indicating our method perform well in simulating different cultures.

Q2: What would the results look like if we don’t use multi-agent role-playing but just use GPT-3.5 to directly generate diverse reasons? Do you have insights on the performance of generation and diversification (i.e., no verification)? It would be interesting to know how important it is to just get a diverse perspective versus having the correct perspective.

We have experiments on using GPT-3.5 to directly generate diverse reasons. The results are shown in Fig.5(a). And we also conducted a new experiment to explore how important it is to get a diverse perspective versus have the correct perspective. The setting of this part is the same with that of Ablation study in our paper. The table below shows the results in 4 cultures. Verify and Diversify show comparable performance, while Verify performs better in our experiments.

Q3: In the 41K data you collected, what should be the input and output to fine-tune an LLM? How are the non-factual and redundant sentences removed? It’s confusing because both operations are performed on the extracted opinions. Additionally, if you remove some of these sentences, some of the utterances may be removed, does the dialogue still make sense?

• Examples on input-output pairs for fine-tuning. The table below shows an example.

  Input	How do you think about "one of my main goals in life has been to make my parents proud?"
  Output	Strongly agree. I believe that pleasing parents and elders is a sign of respect and love.

• Details on factual verification
  • We have seed data and extracted opinions which expresses the opinions of people from different cultures and is generated by our algorithm, respectively. Then we juage the relationship of seed data and extracted opinions from Entail, Contradict and Irrelevant. This step is implemented by prompting GPT-4 with "What's the relationship between the two opinion? Direct answer with Contradict, Entail or Irrelevant.\nOpinion 1: {seed_data}\nOpinion 2: {extracted opinions}". After that, we get the relationship between the seed data and extracted opinions. If relationship equals to Entail, we will save the extracted opinions. If relationship equals to Contradict, we will rewrite the opinions and check the relationship again. If relationship equals to Irrelevant, we will discard the extracted opinions. More details can be found in Sec.A.3.
• Details on redundancy removal
  • We first get the embeddings of extracted opinions via Openai API and cluster them leveraging k-means. For each cluster, we randomly select one as represenative opinion and discard others.
• More explanation
  • The dialogue is the intermediate products of our method. We use the extracted opinions to fine-tune models instead of the dialogues. And the dialogues can be reserved for other usages, such as extracting other cultural information.

Q1: How is content moderation relevant to culture understanding? The only reason that I thought it could be relevant is that some offensive words may be exclusive to certain languages. However, you only use English to do the role playing.

Content moderation is a popular type of tasks highly relevant to culture understanding:

• Offensive language detection is a task of content moderation. For different cultures, they have different values, traditions and customs. For example, some cultures may embrace Christianity. It would be offensive If you ask people from those cultures to eat bloody food. Content moderation is relevant to offensive word, but not limit to. English can also catch this kind of offensive content.
• Other works on culture understanding also stress the application of cultural models for content moderation. For example, [1] proposes an approach to train cultural attuned models and explore their application in content moderation. This approach contains pre-training, fine-tuning and content violation prediction.

[1] Chan, Alex J., et al. "Harmonizing global voices: Culturally-aware models for enhanced content moderation." arXiv preprint arXiv:2312.02401 (2023).

Q2: The idea that we need to have culture-specific LLM seems counter-intuitive. [...]

Indeed, everyone wants to have one unified model that handles all cultures perfectly well. However, we are not alone in finding that one model cannot do well in all cultures. Recent work by Stanford [1] and other institutions [2] have also found that it is of significant importance to train culture-specific LLMs. The reason is that there is cultural conflicts existing. To be specific, people from different cultures have different values, norms and customs. And those contribute to their diverse opinions to the same thing. According to the World Value Survey, Arabic culture believes that men are better political leaders than women, while people in the United States disagree. The cultural conflicts can not be solved in one model, so culture-specific models are required.

To further answer your question about training one unified LLM to handle all the culture, we also train a model on all our generated data. The table below shows the performance. "ours" means our culture-specific model, and "ours(all)" means the model which is trained on all generated data. The results show that "ours(all)" outperform gpt-3.5-turbo and gpt-4 on most cultures, while perform worse than "ours".

[1] Ryan, Michael J., William Held, and Diyi Yang. "Unintended impacts of llm alignment on global representation." ACL (2024).

[2] Li, Cheng, et al. "Culturellm: Incorporating cultural differences into large language models." arXiv preprint arXiv:2402.10946 (2024).

W1: The culture defined in the paper is too coarse-grained. [...].

We strongly agree that language is not equal to, but only a part of culture. But using language to study culture is possible due to the following aspects:

• Existing literature on culture understanding shows that culture boundaries are fluid, dynamic and uncertain. Delanoy emphasizes that cultures are not homogeneous or static entities but are fluid and dynamic. He critiques essentialist views that rigidly define cultural boundaries and instead promotes a more nuanced understanding that considers the intersections of various cultural factors, such as ethnicity, language, religion, and socio-economic conditions [1]. Appadurai also discusses the fluidity of cultural boundaries and the creation of new cultural forms [2]. Cultural boundaries can be geographical regions, language, religion and so on. Based on above statements, using language as cultural boundaries is reasonable.
• Existing NLP works on culture also leverage labguage as culture boundaries. [3] focuses on Arabic and English culture. [4] focuses on 8 different cultures: English, Chinese, French, Russian, German, Arabic, Japanese and Korean. [5] also use language to split different cultures. The authors work on English, German, Russian, Bengali, Chinese, and Indonesian culture. [6] is a hand-crafted benchmark for evaluate diverse cultures. They also use languages as culture boundaries.
• Most downstream benchmarks are classified via language and we cannot get more fine-grained perspectives. For example, if we want to evaluate the performance of Arabic model, we can find benchmarks in Arabic culture. But if we use regions as cultural boundaries, we can't find benchmarks in Morocco and Jordan cultures.
• Note that the main contribution of the paper is to present a general algorithm that can augment LLM culture data but not specific to any cultures. In the future, if more fine-grained culture data are available, our algorithm can also work well.

[1] Delanoy, Werner. "What is culture." The Cambridge handbook of intercultural communication (2020): 17-34.

[2] Appadurai, Arjun. Modernity at large: Cultural dimensions of globalization. Vol. 1. U of Minnesota Press, 1996.

[3] Naous, Tarek, et al. "Having beer after prayer? measuring cultural bias in large language models." ACL (2024).

[4] Wang, Wenxuan, et al. "Not all countries celebrate thanksgiving: On the cultural dominance in large language models." arXiv preprint arXiv:2310.12481 (2023).

[5] Liu, Chen Cecilia, et al. "Are multilingual llms culturally-diverse reasoners? an investigation into multicultural proverbs and sayings." arXiv preprint arXiv:2309.08591 (2023).

[6] Myung, Junho, et al. "BLEnD: A Benchmark for LLMs on Everyday Knowledge in Diverse Cultures and Languages." arXiv preprint arXiv:2406.09948 (2024).

W2: Incomparable baseline comparison. The LLMs you compared in Table 1 all have different size and training recipes. Perhaps it makes more sense to fix a LLM and compare training it using yours or baseline data.

We agree that it is necessary to fine-tune gpt models on the training data of TaiwanLLM and SeaLLM. However, a sad story is that their training data is not publicibly accessible. Moreover, we realize that it is never easy to reach a "fair" comparison since if we fine-tune the same models on their data, it is unfair for our approach since their pre-training data is significantly larger than ours. We would like to claim that given limited budget and GPU hardware, CulturePark remains a cost-effective solution to fastly build a cultural-specific LLM for low-resource culture. This is the main contribution of the paper. As for "fixing an LLM and compare training it using yours", our main experiments are done in this manner: fix gpt-3.5-turbo and fine-tune it on our platform, which has proved its effectiveness. We further provide experiments on Llama-2, which also shows improvement by our platform.

W3: Problematic evaluation settings. Some of the evaluations do not quite make sense. For example, TaiwanLLM was mainly trained on Traditional Chinese data, but was evaluated on a Simplified Chinese dataset. Additionally, SeaLLM focuses on South East Asian languages, but was tested on Korean.

This is a good advice! In short, these comparison seems not meaningful, but they are just an act of negotiation due to shortage of benchmarks:

• For comparison on Chinese culture, we cannot find any models claiming great performance on Chinese culture. But TaiwanLLM is closer. We are aware of the difference (traditional vs. simple Chinese) and to avoid that, we have converted the test samples into Traditional Chinese via GPT-4 and evaluate on TaiwanLLM and our models. Sorry for leaving the details behind.
• For Korean culture, we also cannot find any "KoreanLLM" to compare. Therefore, since Korean belongs to the far east which is close to southeast asia, we use SeaLLM as a baseline for Korean culture. We realize that this comparison setting is not appropriate. We will remove this result in the next version of our paper.
• Finally, we would like to claim that these two comparisons are just a small part of our experimental results among our extensive experiments and removing them does not necessarily hurt our major contribution. We thank the reviewer for pointing this out and we will try to make modifications in the future version of the paper. Please do not blame the whole paper only for these two results:)

We write remaining rebuttal in Comment due to character limit. We apologize for the inconvenience!

Firstly, I'd like to bring to the AC's attention that the authors' rebuttal appears to exceed the 6,000 characters limit according to the guidelines, which may be unfair to other authors. I will address the authors' responses point by point:

Regarding the coarse-grained culture definition, the authors have provided substantial support from the literature to justify their approach. While I acknowledge that previous works have also used language as a proxy for culture, this methodology remains contentious. Simplifying culture by using language as a boundary may facilitate implementation, but it does not fully capture the complexity and diversity of cultures that share the same language.

On the matter of unfair comparisons, it is indeed unfortunate that those LLMs did not release their training data. However, to achieve a fairer evaluation, it would make sense to compare models that have the same base architecture, such as LLAMA-2, trained on your data versus SeaLLM's data, assuming both are LLAMA-2-based.

Addressing the problematic evaluation settings, your explanation regarding Chinese evaluations has clarified some concerns. Nonetheless, using SeaLLM as a baseline for evaluating Korean remains problematic due to the significant linguistic differences between Korean and Southeast Asian languages. Importantly, I have identified two Korean LLMs: maywell/Synatra-7B-v0.3-dpo and yanolja/EEVE-Korean-10.8B-v1.0. Testing against these models should provide a more accurate and fair comparison for the Korean evaluation.

The NeurIPS guidelines only mention that authors may use the comment box instead of the rebuttal box, but they do not specify that the comment box is excluded from the 6,000-character limit. This ambiguity could still lead to unfairness towards other authors. However, I understand the ambiguity and will simply bring this issue to the AC’s attention.

Regarding your point:

We never claimed that using language as a proxy can fully capture the complexity and diversity of cultures, and this is not the main contribution of our work, we hope that this point is not used against our work.

While you did not explicitly claim this, the use of language to denote cultures in your paper implies reliance on this assumption. Furthermore, the work you cited [1] states, “cultures are not homogeneous or static entities but are fluid and dynamic. It critiques essentialist views that rigidly define cultural boundaries and instead promotes a more nuanced understanding that considers intersections of various cultural factors such as ethnicity, language, religion, and socio-economic conditions.” This is in line with my concerns, as understanding culture should take into account multiple factors beyond just language, including but not limited to ethnicity, religion, and socio-economic conditions. This undermines the assumption that studying culture through language alone is holistic.

[1] Delanoy, Werner. "What is culture?" The Cambridge handbook of intercultural communication (2020): 17-34.

Regarding the additional experiments, I appreciate your efforts. However, to ensure fairness, such comparisons should be extended to all baselines.

In conclusion, since my concerns about the assumption/settings of the paper and fair comparisons remain, I will keep my review score unchanged.

W1: There are missing details on: (1) how the authors use the extracted opinions on target culture to filter out examples (L176) (2) how are the LLMs finetuned (e.g., are they finetuned only on the "culture delegate" agent utterances?) (L187)

Thanks for the reminder! We will append those details into the next version of paper.

• (1) We have seed data and extracted opinions which expresses the opinions of people from different cultures and is generated by our algorithm, respectively. Then we juage the relationship of seed data and extracted opinions from Entail, Contradict and Irrelevant. This step is implemented by prompting GPT-4 with "What's the relationship between the two opinion? Direct answer with Contradict, Entail or Irrelevant.\nOpinion 1: {seed_data}\nOpinion 2: {extracted opinions}". After that, we get the relationship between the seed data and extracted opinions. If relationship equals to Entail, we will save the extracted opinions. If relationship equals to Contradict, we will rewrite the opinions and check the relationship again. If relationship equals to Irrelevant, we will discard the extracted opinions. More details can be found in Sec.A.3.
• (2) For LLMs finetuned, we use the utterances of both "Culture delegate" agent and "Main contact" agent. Because agents have cross-cultural understanding ability. They try to comprehend the value, norm and custom of other cultures and interpret in their own words. So we leverage utterances of both "Culture delegate" and "Main contact" agents to extract cultural specific opinions.

W2: L168 mentions 41k dialogues generated (presumably before filtering), but then L185 mentions 41k after filtering. Authors should include how many data instances are filtered out after generation.

41k dialogue is truly the data before filtering. Note that each dialogue is multi-turn, which has several utterances (samples) from "Culture delegate" and "Main contact" agents. Probably, each diglogue can extract 5-10 culture-specific opinions. Those opinions will be refined. After that, we can get 41k high-quality culture-specific opinions.

W3: The dataset analysis section has some issues of claims that are too strong and causal that do not have experiments that back them up:

• L193: While I appreciate the investigation into cross-cultural understanding, the analyses done in the paper seem to be more qualitative [...] I suggest discussing the baseline results together with the fact that the conversations have rich cultural understanding.

  • Strongly agree! We will refine this part and discuss the baseline results in the part.

• L205: Again, no experiment was done to test this hypothesis; the setup based on CCT/SCT seems to have led to good conversations, [...] CCT/SCT (e.g., within-culture conversations, as the authors mentioned).

  • To verify the effectiveness of CCT/SCT seeting, we generate training data in four different settings: "In-cultural+Same gender", "In-cultural+Different gender", "Cross-cultural+Same gender", "Cross-cultural+Different gender". For each, we generate 500 samples to fine-tune models. Other settings of this experiment are the same with the experiments in Sec 5.2. As shown in the table below, the setting which has less conflict perform worse while the setting which shows conflicts in culture and gender perform better.

    Setting	Performance
    In-cultural+Same gender	0.411
    In-cultural+Different gender	0.435
    Cross-cultural+Same gender	0.472
    Cross-cultural+Different gender	0.480

• L212: Same comment, I would tone down the causal claims and just focus on reporting observations from the dataset.

  • Agreed. We will modify the statement in the final version of the paper and just focus on the observations from our experimental datasets.

  • Furthermore, just to claim the "novel questions" here, we compute the perplexity and diversity gain of the generated dataset. Fig.5(b) presents some detailed results on diversity gain, showing that the generated data has significantly larger diversity. The results of our method on perplexity are shown in the table below. The perplexity of CulturePark is higher than that of CultureLLM, indicating high-quality information is brought by our method.

    Number of data	150	550	1050
    CultureLLM	15.06(0.43)	15.39(0.05)	14.78(0.81)
    CulturePark	46.02(0.14)	52.41(0.52)	61.22(0.37)

W4: More details are needed on Hofstede's cultural dimensions theory. [...].

In our evaluation on cultural alignment via Hofstede’s Cultural Dimensions Theory, we evaluate models in the same setting on our models and GPT-4/3.5. For example, we will prompt GPT-4/3.5 and our models with "you are a Korean chatbot that knows Korean culture very well" when we want to verify their cultural understanding in Korean. In short, we use the latter setup in your question. We will append the details in the next version of our paper.

W5: L360: The societal impact statement needs to be drastically toned down. [...]

Nice advice! We rewrite this paragraph as follows: CulturePark aims to promote fairness and inclusivity by ensuring accurate cultural representation. It helps improve global communication, fosters cross-cultural understanding, and supports multilingual societies. By minimizing biases in language models, it builds trust and aligns with responsible principles. Additionally, it supports inclusive education and promotes cultural awareness. Addressing cultural biases in language models leads to more reliable and beneficial AI systems.

We write remaining rebuttal in Comment due to character limit. We apologize for the inconvenience!

Minor issue1: L107: saying generated data is "more natural" is factually wrong, human-generated data is by definition more natural.

There is a little misunderstanding here about the word "natural" since this is used to state that the data generated by CulturePark is more natural than CultureLLM, but not compared to humans.

Minor issue2: L159: it's unclear what the authors mean by "does not need human involved", as I thought there was no human involved.

Your understanding is true. The claim "there is no human involved" just means the entire platform is automatic without any human efforts.

There is a big missing limitation, which is that simulated conversations do not reflect how people actually would talk (e.g., w.r.t. information asymmetry and style of utterances; Zhou et al. 2024; http://arxiv.org/abs/2403.05020), and could contain biases or stereotypes w.r.t. the culture represented (see Cheng et al 2023, marked personas paper).

Thank you for your great advice! We also noticed this phenomenon in our experiments. We found that cultural bias and stereotypes emerge in LLMs even be prompted to simulate specific culture. To mitigate the cultural bias in role-playing and make the simulation real, we devise self-calibration prompts to calibrate their outputs and different conversation styles to guide the conversation. More details can be found in Sec 3.1. Furthermore, experimental results can verify the effectiveness of CulturePark and the authenticity of the cross-cultural communication.

We would like to thank the reviewer for the positive support of our work. If you have any new questions, please do not hesitate to let us know.