CultureLLM: Incorporating Cultural Differences into Large Language Models

W1: Rely on human-annotated dataset (i.e., WVS). The author used WVS dataset as seed data to augment their fine-tuning dataset. This limits the applicability of the proposed method.

Using human-annotated data to help research is a popular choice for most of the LLM papers, as shown in [1-7], where they all use human-annotated dataset as seed data. On this point, one noteworthy advantage of our approach is that we only use 50 samples, which is significantly less that existing literature, claiming our less reliance on annotations.

[1] Wang, Ruida, Wangchunshu Zhou, and Mrinmaya Sachan. "Let's Synthesize Step by Step: Iterative Dataset Synthesis with Large Language Models by Extrapolating Errors from Small Models." arXiv preprint arXiv:2310.13671 (2023).

[2] Chen, Zixiang, et al. "Self-play fine-tuning converts weak language models to strong language models." arXiv preprint arXiv:2401.01335 (2024).

[3] Yu, Longhui, et al. "Metamath: Bootstrap your own mathematical questions for large language models." arXiv preprint arXiv:2309.12284 (2023).

[4] Li, Xian, et al. "Self-alignment with instruction backtranslation." arXiv preprint arXiv:2308.06259 (2023).

[5] Liu, Bingbin, et al. "Tinygsm: achieving> 80% on gsm8k with small language models." arXiv preprint arXiv:2312.09241 (2023).

[6] Singh, Avi, et al. "Beyond human data: Scaling self-training for problem-solving with language models." arXiv preprint arXiv:2312.06585 (2023).

[7] Huang, Yiming, et al. "Key-point-driven data synthesis with its enhancement on mathematical reasoning." arXiv preprint arXiv:2403.02333 (2024).

W2: Most evaluation downstream tasks are anti-social detection tasks (such as offensive language, hate speech, toxicity and abusive language detections). The results may be biased to such tasks. I also wonder why these tasks are culture-related tasks. Are there any particular examples?

• First of all, the downstream tasks are just evaluation, but not fine-tuning datasets. That being said, we only perform evaluation on those downstream datasets. Therefore, your comment "the results may be biased to such tasks" does not hold.
• Second, it is actually challenging to find enough public datasets to evaluate the cultural models. Those downstream tasks are what we have found. Most of them may relate to anti-social, but our models are not. In fact, those extreme cases are what cultural misunderstanding could possibly occur, which can help "stress test" our models and verify its real application.
• Third, the table below shows the input-output pairs of "offensive language detect" task. The offensive contents are different for different cultures. The model needs to judge if the input content is offensive and label it with "OFF" or "NOT". And the label is related to the cultural context.

W3: Some experiment details need clarification. 1. What are your fine-tuning hyperparamters? Like learning rate and training steps? 2. When you evaluate CultureLLM the downstream tasks, are the input samples in English or the particular language?

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

• Fine-tuning hyperparamters. For gpt-3.5-turbo, the number of training epochs is 3. And we use default settings for other hyperparamters. For Llama-70b-Chat, the details on setup are shown in Sec.D.1.
• Language of input sample: The evaluation benchmarks are all in particular languages.

Q1: Did you check if any fine-tuning samples are overlapped with your downstream samples? Or some samples are very similar to downstream tasks?

We checked manually and it turns out the fine-tuning samples are not overlapped with the downstream samples. It is easy to understanding since the seed data from World Value Survey(WVS) are related to cultural value, which focus on more abstract information. The evaluation downstream tasks care about the meaning of a specific sentence.

Q2: In your ablation studies, WVS+a only used a semantic template. Do you mean that LLMs are fine-tuned on semantic templates?

No. "WVS+a" means the output of Step 2 instead of output of Step 3 in Fig 2.

If you are satisfied with the answers, please consider improving the rating to support our work! If you have more questions, please do not hesitate to let us know:)

Q1: I am uncertain about the relevancy between the evaluated tasks and the fine-tuning data. How are values in the World Value Survey/extracted opinions relevant to offensive language/ hate speech/ stance? Can you elaborate?

We analyze the performance for each task and report the WinRate in the table below.

The relevance of each task with WVS can be described in the following:

• offensive language detect:

1. Cultural Context and Sensitivity to Offensive Language: The World Values Survey aims to capture cultural values and beliefs across different societies. One aspect of cultural values is the tolerance or acceptance of offensive language. In some cultures, certain words or expressions may be considered highly offensive, while in others they may be more tolerated or even commonly used.

• hate speech detect:

1. Societal Norms and Attitudes: The WVS provides data on societal norms, attitudes towards minorities, and levels of societal trust. This data can help understand the underlying societal conditions that might foster hate speech or, conversely, promote tolerance and inclusivity.
2. Cultural Context: Understanding the cultural context is crucial for effectively detecting and interpreting hate speech. The WVS offers a rich dataset for understanding cultural differences in values and norms, which can inform more nuanced hate speech detection algorithms.

• stance detect:

1. Understanding Contextual Influences on Stance: The WVS can provide the cultural and societal background needed to understand why certain stances are more prevalent in specific regions or among certain demographic groups. This context can be invaluable for interpreting the results of stance detection analyses, especially when comparing stances across different cultures and societies.

• toxicity detect:

1. Reflection of Societal Norms in Online Behavior: The WVS provides insights into the prevailing norms and values within societies, which can indirectly inform the context within which toxic behavior manifests online. Understanding societal attitudes towards diversity, authority, individual freedom, and tolerance can help in interpreting the root causes of toxic behavior and devising appropriate responses.

• threat detect:

1. Understanding Motivations and Behaviors: Insights from the WVS can help understand the cultural and societal contexts that may influence the behavior of individuals or groups posing threats. This knowledge can inform more targeted and effective threat detection and mitigation strategies that consider the root causes of conflict or aggression.
2. Cultural Sensitivity in Security Measures: Incorporating findings from the WVS can lead to more culturally sensitive security practices that respect local values and norms. This is crucial in global operations where misunderstanding cultural nuances can lead to ineffective or counterproductive security measures.

• bias detect:

1. Understanding Societal Norms and Attitudes: Insights from the WVS can help in understanding the cultural and societal norms that underlie biases. By analyzing patterns in global values and beliefs, we can identify prevalent stereotypes, prejudices, and discriminatory attitudes that may need to be addressed in bias detection efforts
2. Injection of More cultural nuances: The WVS data can provide valuable context that are sensitive to cultural differences in values and norms. This is better equipped to detect and mitigate biases in data sets that reflect cultural nuances, ensuring that AI-driven decisions are fair and equitable across different societal contexts.

• abusive detect:

1. Cultural Contexts of Abuse: The WVS can help identify cultural norms that influence perceptions of what constitutes abusive behavior. This is crucial for developing detection systems that are sensitive to cultural differences, ensuring that they can effectively identify abuse without mistakenly flagging culturally specific but non-abusive interactions.
2. Injection of More cultural nuances: Insights from the WVS can inform the development of more nuanced algorithms for detecting abusive behavior by providing context on societal values and norms.

W1: Assumption of Language as Culture. The paper equates languages with cultures, which is an oversimplification. Cultures are multi-faceted and cannot be fully encapsulated by language alone. There are significant cultural differences within the same language-speaking regions that may not be adequately captured by this approach. Even worse, on line 205, the authors said they picked “representative countries''. But how do you define “representative”? More importantly, this selection method will likely cause the fine-tuned LLMs to be biased towards these “representative” countries for certain languages.

Why using one language to denote one culture: 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.
• It’s interesting to incorporate more fine-grained cultural differences within the same language-speaking regions into LLMs. Our method is an initial attempt, which can generalize to more fine-grained cultures. Just changing the source of seed data can achieve this. However, we can not implement this method for all fine-grained cultures in one paper, because of time and resource limit.

Explanation on "representative countries":

In terms of “representative countries”, our selection criterion is that we choose the country which has the most population. We agree that this can cause the fine-tuned LLMs to be biased towards the “representative countries”. However, we think the criterion may be the best way to align with majority of people from certain cultures.

Finally, 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 or contries. 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: Limited Scope of Seed Data. The methodology relies heavily on a small seed dataset of only 50 samples from the World Value Survey (WVS). While the approach is cost-effective, it may not capture the full breadth and nuances of cultural diversity.

• First, although the 50 data seem narrow, they however carry resourceful information since they are carefully designed and validaded by human experts. The seed data covers 7 topics: social values, security, science and technology, religious values, ethical values and norms, political interest and political participation, and migration. The data is high quality and have been widely used by a lot of literature. This statement can also be verified by the excellent performance of CultureLLM on downstream tasks.
• Second, we have conduct pilot experiments on selecting the seed data from WVS. In our study, we explore the effect of sample’s quantity and different composition of samples on the performance of model. The results indicate that the 50 samples perform better than other settings. The reasons may be that fine-tuned LLMs perform better when the distribution of fine-tuned data similar to that of pre-training data. The distribution of those 50 seed data is similar to the pre-training data. So we select those samples.
• Finally, the main contribution of the paper is not the seed data, but our data augmentation algorithm that remains general to any seed data and topics. That being said, in the future, with new topics and seed data available, users can easily develop their own versions of LLMs using our algorithm.

W3: Unfair Baseline Comparison.[...]Until then, “cost-effective” cannot be claimed.

We response in general response for this part.

While I acknowledge that prior works have used language as a proxy for culture, the validity of this approach remains debatable. Using language as a cultural boundary can simplify the implementation, but it doesn't fully address the complexity and diversity of cultures that share a common language. Regarding the choice of “representative countries,” the authors mentioned, "we think the criterion may be the best way to align with the majority of people from certain cultures." However, isn’t that exactly the flip side of having a biased model?

In terms of the seed data, your explanations have raised more questions. First, in your pilot study, what other settings did you compare these 50 seed data against? Second, it is confusing when you refer to the seed data as being "similar to that of pre-training data." Are you referring to the pre-training data of GPT-3.5-turbo? If so, how do you know the distribution of proprietary data? Your response implies that you already know the pre-training data of GPT-3.5-turbo, which raises concerns about the methodology.

As for your justification regarding fair comparison, there are also some issues. First, referencing other leaderboards does not justify your experimental setup. Second, acknowledging that a fair comparison cannot be achieved due to data unavailability suggests that it is premature to label your approach as "cost-effective." We need to clarify which costs and what effectiveness metrics you are comparing against. Lastly, the mention of “CulturePark” in your response makes it seem like the responses are being reused for multiple submission, which may come across as unprofessional.

I am unsure why the use of language as a proxy to study culture cannot be judged here, as it is a fundamental assumption of your work.

Regarding the selection of the 50 seed data, there seems to be a contradiction. Initially, you mentioned that "the distribution of those 50 seed data is similar to the pre-training data," implying that Shi et al.’s method was used for selection. However, your experiments with settings other than your 50 seed data suggest that these examples were chosen based on average performance. Could you clarify which criteria were ultimately used for the selection?

Moreover, Shi et al.’s method requires computing token probability, which is not available in GPT-3.5-turbo's output. How did you obtain the probability for each token?

Thank you for providing the OpenAI API call.

However, Shi's method requires obtaining the (log) probability of the input prompt (i.e. the input tokens), while the API you shared only provides probabilities for the output tokens.

Additionally, it appears this API call supports only the top 20 highest probability tokens. Could you please clarify how you obtained the minimum 30% and 40% log probabilities in your experiments?

We thank reviewer XR5a for the improved rating. There seems to be two remaining concerns:

the fundamental assumption of using language to denote culture

As we explained previously, for this open question, we are not the first work to use this assumption. We respect the reviewer's opinion on this point.

the issue of fair comparison

For this comment, we have answered reviewer XR5a's demand on fine-tuning GPT models on the pre-training data of SeaLLM and Taiwan LLM by stating that their data is not publicly available.

In fact, the major experiments in the paper are conducted under fair comparison: we compare GPT-3.5-turbo with our fine-tuned GPT-3.5-turbo version, ensuring that they are using the same backbone models. If one only cares about absolute performance, we still compare with GPT-4, the most advanced model to date. We hope that the reviewer can acknowledge this.

In the future, with more multilingual data publicly released, we will continue more comparisons using the same backbone models.

Authors appreciate the multiple rounds of discussion with reviewer XR5a, which makes the paper more sound. We will include all discussion results and analysis into the final version of the paper.

Thank you for your response. The authors have addressed several of my concerns. However, the fundamental assumption of using language to denote culture, along with the issue of unfair comparison, remains challenging to resolve.

Nonetheless, considering the authors’ efforts in their rebuttal, I have decided to increase my rating.

W1: It is important to heavily note that language is not equal to culture and this tends to cause confusion in this paper. Culture is way more complex than language and it might have been easier to call the language splits as culture for writing but this will introduce misunderstandings that will muddy your message.

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: Augmentation with semantic similarity checks is something that has been worked on before for NLP augmentaiton. How do you deal with the challenge that even with high cosine similarity because of BERT embeddings, sentences that have very dissimilar meanings will be counted as similar or pass through your filter.

There are two tricks that work in our study

• Threshold Tuning and Filtering: Adjust the similarity threshold to balance between false positives and false negatives. We choose a more conservative threshold, which can reduce the likelihood of accepting semantically dissimilar sentences.
• Use of Cultural Contextual Information: BERT embeddings are context-sensitive, but sometimes context nuances are not fully captured. To mitigate this, we consider incorporating more cultural context.

Q1: A questions that comes to mind, that you partly address, is the crosslingual nature of your finetuning, instead of translation. Would not having both even be better? And what would the effect also of using more local language based LLMs be? e.g. LeoLLM for german langauge.

Nice question! We fine-tune LeoLLM with our generated data in both English and German version. Results are shown below:

The results indicate that the fine-tuned model perform better when the language of fine-tuned data is same with that of pre-training data. This could be due to the fact that English is much more than the local languages, which can naturally assist fine-tuning.

Q2: Augmentation with semantic similarity checks is something that has been worked on before for NLP augmentaiton. How do you deal with the challenge that even with high cosine similarity because of BERT embeddings, sentences that have very dissimilar meanings will be counted as similar or pass through your filter.

The response is shown in the response of W2.

Note: You refer to the crowdsource study as having details including IRB/Ethics information in Appendix E, but this is not so. You just describe the task questions. Was IRB/Ethics approval sought given that the nature of the questions may for some people find offensive?

Thanks for the reminder! We have gotten IRB/Ethics approval for the human study.

Q1: I'm wondering how the performance (or errors) are associated with the coverage of the 50 questions used for fine-tuning.

We analyze the performance for each task and report the WinRate in the table below.

The relevance of each task with WVS can be described in the following:

• offensive language detect:

1. Cultural Context and Sensitivity to Offensive Language: The World Values Survey aims to capture cultural values and beliefs across different societies. One aspect of cultural values is the tolerance or acceptance of offensive language. In some cultures, certain words or expressions may be considered highly offensive, while in others they may be more tolerated or even commonly used.

• hate speech detect:

1. Societal Norms and Attitudes: The WVS provides data on societal norms, attitudes towards minorities, and levels of societal trust. This data can help understand the underlying societal conditions that might foster hate speech or, conversely, promote tolerance and inclusivity.
2. Cultural Context: Understanding the cultural context is crucial for effectively detecting and interpreting hate speech. The WVS offers a rich dataset for understanding cultural differences in values and norms, which can inform more nuanced hate speech detection algorithms that are sensitive to context and do not inadvertently suppress legitimate expressions of cultural or political dissent.

• stance detect:

1. Understanding Contextual Influences on Stance: The WVS can provide the cultural and societal background needed to understand why certain stances are more prevalent in specific regions or among certain demographic groups. This context can be invaluable for interpreting the results of stance detection analyses, especially when comparing stances across different cultures and societies.

• toxicity detect:

1. Reflection of Societal Norms in Online Behavior: The WVS provides insights into the prevailing norms and values within societies, which can indirectly inform the context within which toxic behavior manifests online. Understanding societal attitudes towards diversity, authority, individual freedom, and tolerance can help in interpreting the root causes of toxic behavior and devising appropriate responses.

• threat detect:

1. Understanding Motivations and Behaviors: Insights from the WVS can help understand the cultural and societal contexts that may influence the behavior of individuals or groups posing threats. This knowledge can inform more targeted and effective threat detection and mitigation strategies that consider the root causes of conflict or aggression.
2. Cultural Sensitivity in Security Measures: Incorporating findings from the WVS can lead to more culturally sensitive security practices that respect local values and norms. This is crucial in global operations where misunderstanding cultural nuances can lead to ineffective or counterproductive security measures.

• bias detect:

1. Understanding Societal Norms and Attitudes: Insights from the WVS can help in understanding the cultural and societal norms that underlie biases. By analyzing patterns in global values and beliefs, we can identify prevalent stereotypes, prejudices, and discriminatory attitudes that may need to be addressed in bias detection efforts

• abusive detect:

1. Cultural Contexts of Abuse: The WVS can help identify cultural norms that influence perceptions of what constitutes abusive behavior. This is crucial for developing detection systems that are sensitive to cultural differences, ensuring that they can effectively identify abuse without mistakenly flagging culturally specific but non-abusive interactions.
2. Injection of More cultural nuances: Insights from the WVS can inform the development of more nuanced algorithms for detecting abusive behavior by providing context on societal values and norms.
3. Evaluating Tolerance Levels: The WVS data can provide insights into societal tolerance levels towards different forms of behavior, including what might be considered abusive. This can help in assessing the urgency and type of interventions needed to address abusive behaviors in various cultural contexts.

• spam detect

1. Cultural Variations in Communication: The WVS can shed light on cultural differences in communication styles and preferences, which can inform more nuanced spam detection algorithms that are better able to distinguish between legitimate mass communications and spam in different cultural contexts.
2. Attitudes Towards Technology and Privacy: Insights from the WVS regarding societal attitudes towards technology use, privacy, and data protection can help in tailoring spam detection efforts to respect cultural norms and expectations. For instance, societies with a high value on privacy might be more receptive to stringent spam filters.
3. Attitudes Towards Technology and Privacy: Insights from the WVS regarding societal attitudes towards technology use, privacy, and data protection can help in tailoring spam detection efforts to respect cultural norms and expectations. For instance, societies with a high value on privacy might be more receptive to stringent spam filters.