"Globally, people express pride, celebrate, and respect cultural diversity, while acknowledging and working towards reducing cultural bias" -a data point in CultureBank
Large Language Models (LLMs) have become instrumental in various applications to interact with diverse user populations, such as in recommender systems (Li et al., 2023;Fan et al., 2023) and customer service (Pandya and Holia, 2023). However, these models often mirror Western-centric perspectives (Santurkar et al., 2023;Durmus et al., 2023b), as they are predominantly trained on data that reflect these values and behaviors. Such cultural bias can lead to unintended consequences (Ryan et al., 2024), e.g., reinforcing stereotypes, alienating non-Western users, hindering global deployment and so on. Therefore, it becomes increasingly important to develop language technologies that are aware of diverse cultures. We acknowledge that culture is a broad concept, and prior work has attempted to operationalize culture via different proxies (Nangia et al., 2020;Nguyen et al., 2023;Dwivedi et al., 2023;Adilazuarda et al., 2024). In the following paper, we use culture to refer to the knowledge of traditional behavior, norms, values, shared meanings, and patterned ways of behaving which are characteristic of a given society, a group of societies, a certain race or a certain period of time (Kroeber and Kluckhohn, 1952;Boyd and Richerson, 1988;Bozeman, 1992;Brumann, 1999;Mead et al., 2000;McFate, 2005;Birukou et al., 2013).
To enhance LLMs' culture awareness, existing studies have developed cultural knowledge databases to represent culture-related knowledge and norms, but they have several limitations. (1) They often rely on formal knowledge sources like Wikipedia and online articles (Nguyen et al., 2023;Fung et al., 2024), which miss the rich, evolving and long-tailed cultural nuances experienced by local communities. (2) Secondly, these methods tend to present cultural knowledge in an assertive manner (Nguyen et al., 2023;Fung et al., 2024;Yin et al., 2022), failing to capture the fact that cultural practices and values can vary among individuals within the same cultural group. (3) Besides, their evaluation methods often rely on classification tasks and question answering (Naous et al., 2023;Afina Putri et al., 2024;Shafayat et al., 2024), which is very different from how LLMs are deployed in the real world and hence cannot reflect their cultural awareness in practice.
To tackle these challenges, we utilize online communities, an important complement to existing work, where people share their cultural ex-Figure 1: Overview. Our goal is culturally-aware language technologies. To do so, we develop a pipeline and construct CultureBank with structured cultural descriptors. Each descriptor has a grounded scenario, persona, and question for evaluation. We fine-tune a model on CultureBank and improve its performance on two cultural tasks.
periences, and develop a bottom-up approach to process noisy self-narratives on a massive scale. Using this pipeline, we develop CultureBank, a cultural knowledge base with 12K cultural descriptors sourced from TikTok (Figure 1 shows one example) 3 . Besides, to address the limitation on assertiveness, we gather diverse views on similar cultural practices, and calculate an agreement level to enable inclusive cultural understanding. Moreover, to facilitate contextualized evaluation on LLMs' cultural awareness, we provide a related situation grounded in real-world settings for each cultural descriptor (e.g., travel consultation in Figure 1). Then we evaluate state-of-the-art LLMs' cultural awareness on CultureBank, and the results show room for improvement. Additionally, we demonstrate that training LLMs on CultureBank enhances their performance on downstream culture-related tasks. We also show that our pipeline can be easily generalized to Reddit, another online community, illustrating its potential for future expansions.
In short, we make the following contributions. • A general framework to collect cultural knowledge from online communities ( §4). • CultureBank, an open-source cultural knowledge base with 12K cultural descriptors from TikTok and 11K from Reddit ( §5 and §8) 4 . 3 We note that our goal is not to develop a comprehensive cultural knowledge base, but rather, to propose a scalable way to extract cultural knowledge from online communities. 4 We note that CultureBank is not a comprehensive and exclusive cultural knowledge base, but rather it is a necessary complement to existing cultural knowledge base by utilizing social media for cultural knowledge acquisition. With Cul-tureBank and prior cultural knowledge base that rely on more formal source, one can learn cultural knowledge in a more holistic way. And for cultural learning, we need cultural data from different sources.
• Grounded evaluation on existing LLMs' cultural awareness ( §6) and a more culturally-aware language model fine-tuned on CultureBank ( §7).
Cultural knowledge bases. There have been many cultural knowledge base efforts in different domains (Fung et al., 2023;Lee et al., 2023;Kim et al., 2024;Jin et al., 2023;Fung et al., 2024).
With traditional ethnographic methods, social scientists recorded cultural knowledge through existing historical accounts, ethnographic data, and cultural documents. For instance, behavioral scientists compiled a collection of cultural materials, and released an online database named eHRAF (the Human Relations Area Files) (Gehring and Sinnott, 2018). In computer science studies, researchers employ computational methods to automatically construct datasets (Penta et al., 2011) from large sources or curate data from crowd source workers (Lee et al., 2023). (Nguyen et al., 2023) built a pipeline to extract assertive cultural commonsense knowledge from C4 (Raffel et al., 2020), a large collection of Internet data, and (Fung et al., 2024) used Wikipedia and navigated to related online documents to extract cultural knowledge. Data from these sources are much cleaner compared to online communities, and often focus more on normative cultural indicators. Since culture is highly heterogeneous, we also need descriptive cultural expressions from sources like online communities. StereoKG (Deshpande et al., 2022) used Reddit and Twitter to extract cultural stereotypes for 5 religious groups and 5 nationalities, but due to the lack of proper filtering, the results are noisy. As an important complement to existing data sources, our work proposes a pipeline to process highly noisy online communities data on a large scale, and show that it can be easily generalized across different platforms, to provide valuable descriptive cultural knowledge.
Cultural-awareness in language models. Previous works have studied cultural dimensions in language models (Gutiérrez et al., 2016;Jha et al., 2023;Ramezani and Xu, 2023;Jiang et al., 2020;Adewole et al., 2021;Yao et al., 2023;Li et al., 2024;Adilazuarda et al., 2024;Cao et al., 2023;Liu et al., 2021;Hämmerl et al., 2022;Huang and Yang, 2023;Wang et al., 2023;Köksal et al., 2023;Hershcovich et al., 2022;Mukherjee et al., 2023). On the evaluation side, prior studies have measured subjective global opinions from LLMs (Durmus et al., 2023a;Santurkar et al., 2023), and probed cultural value differences in these models (Arora et al., 2022;Yin et al., 2022;Roberts et al., 2023). On the model side, CultureLLM (Li et al., 2024) proposed a cost-effective method to integrate cultural differences into language models with augmented data. This work proposes a grounded way to evaluate cultural awareness to match real-world use cases, and fine-tune a more culturally aware language model with descriptive cultural behaviors constructed from online communities.
Prior efforts on cultural knowledge base (Nguyen et al., 2023) often represent cultural knowledge in free-text sentences. But free-text contents on online communities are often noisy, and such an unstructured representation hinders further computational operation such as search and filter. Therefore, besides providing free-text description, we also develop a taxonomy (shown in Table 1) for more structured cultural knowledge representation, based on the taxonomy of social factors (Hovy and Yang, 2021), the taxonomy of social norms (Ziems et al., 2023;Goffman et al., 2002), and other literature (Goffman et al., 2002;Nguyen et al., 2023;Fung et al., 2024;Yin et al., 2022).
Table 7 lists the mapping between our proposed taxonomy and previous taxonomies, and shows that our taxonomy covers a broad range of cultural fields in existing literature. It has the following fields: (1) cultural group, (2) context, (3) goal, (4) actor, (5) recipient, (6) relation, (7) actor's behavior, (8) recipient's behavior, (9) other description, (10) topic, and (11) agreement. For all these fields, we provide in-context examples and let the model extract any related information without constraint. This could improve diversity in the data: for instance, examples for cultural group include typical cultural groups by countries such as "American", as well as more fine-grained ones by regions or ethnicity groups such as "Californian" and "Asian American", and more broad social groups such as "people from cold weather" which can be overlooked before (Barth, 2010;Stenou, 2002).
Centering on the proposed taxonomy, we propose a bottom-up pipeline to construct cultural descriptors from online communities (TikTok in the main (3) post-process them for agreement levels and remove bad contents. For each step, we sample a random subset of the output, and validate it with human evaluation.
experiment, and we also apply it on Reddit to see the generalizability of the pipeline).
In the first part of the pipeline, we extract cultural descriptors from self-narrative data like comments and posts on online communities, and organize them into our taxonomy. More detail in §B.
1. Culture relevance classifier. Given the large amounts of noisy data, the first step is to get the culturally-relevant portion. To do so, three student researchers annotate a subset with 280 training examples. The instruction to the annotators is "In this step, we will remove completely culturallyirrelevant comments such as 'cool post', 'thanks for sharing'. Please be open-minded, if you think it's slightly culturally relevant, such as "we also do the same", we will keep it.". With these examples, we train a distill-bert-based (Sanh et al., 2019) cultural relevance classifier. Then the binary classifier is applied on the entire dataset to get the subset related to culture. The classifier achieves an accuracy of 79% on a held out test set with 100 examples. Cultural descriptor extractor. After obtaining the cultural comments, we employ Llama-2-70B (Touvron et al., 2023), one of the best open-source LLMs at the task time, to extract values for each field in our taxonomy, by conditioning on the definition of fields and in-context examples. Listing 2 shows the prompt used. Human evaluation shows that this extractor achieves an accuracy of 82% across fields in the taxonomy on a test set with 240 examples, annotated by three researchers (80 distinct examples per person) with Kappa=0.74 on a held-out set with 80 examples.
After the extraction step, we have many cultural descriptors, but the same cultural behavior can be expressed in many different ways, for instance, "Japanese do not tip service staff", or "In Japan, people do not give tips". So naturally in the second part, we need to first cluster the extracted descriptors, and summarize each cluster afterwards.
Clustering. For the clustering step, we concatenate the extracted fields, encode the concatenated contents with SentenceBert (Reimers and Gurevych, 2019), and perform Hierarchical Agglomerative Clustering (HAC) clustering. We use the cluster size as the support value and remove clusters with less than 5 comments to ensure that each cultural descriptor is supported by enough evidences, because if only one or two comments mention certain practice, it could be very specific personal experiences. The clustering parameters are chosen based on the performance on a validation set. One student researcher annotated the validation set, and it achieve an accuracy of 93% with the human annotated clusters. The clustered results achieve an average Silhouette score of 0.14 within the clusters.
Cluster summarizer. After clustering, each cluster contains multiple cultural descriptors, so the next step is to summarize and generate a representative descriptor for each cluster. We use Mixtral-8X7B (Jiang et al., 2024), a state-of-the-art open-source language model at the task time, to summarize each cluster. Since the clusters contain noisy opinions, the vanilla model often fails to output a comprehensive summarization with in-context examples. To achieve a better performance, we ask GPT-4 to generate 1K high-quality summarizations, and distilled those samples to train our own Mixtral summarizer. Listing 3 shows the prompt used for the summarizer. Human evaluation by two student researchers shows the cluster summarizer achieves a fidelity score of 89.7% and coherence score of 96.6%. Definitions of these metrics are in §B.2.
Table 6 shows examples of the summary.
The final step is to post-process the clustered data. Agreement calculator. People may have different opinions regarding the same cultural behaviors, so instead of assertive statements like "Japanese people don't tip", we provide agreement levels for each cultural descriptor in our CultureBank.
Each cluster contains → 5 comments with extracted values, and each comment is associated with an agreement score of 0 or 1 (extracted by the LLMs) during knowledge extraction, so we compute the average of these agreement scores among commenters as the agreement level. Besides, the cluster size can also reflect the agreement level. Content moderation. Finally, online platforms can contain controversial contents. So the last step is content moderation. To do so, we first use the perspective APIfoot_2 , a machine-learning-based content moderation tool, and filter out contents with scores above 0.2 for every category (toxicity, profanity, insult, identity attack, threat, severe toxicity). For more nuanced controversial contents, three researchers annotate 800 examples, and then we train a distill-bert-based classifier (test acc=0.77 on 117 examples), and employ a list of keywords to further identify them. Next, we manually label these identified contents, and remove inappropriate ones (e.g., toxic, biased descriptors). See § B.3 for more detail on the content moderation process, like keywords, annotation rubrics, etc. Finally, we use the Presidio Analyzerfoot_3 to detect and remove Personal Identifiable Information (PII).
TikTok is a popular social media platform with users from different cultural backgrounds, so we apply our pipeline on data from TikTok to construct our CultureBank dataset. We obtain TikTok data via their official research APIfoot_4 and collect a total of 34K posts and 720K English comments from 2019/05 to 2023/08 with the hashtags "#culturaldifference" and "#cultureshock". Table 2 shows
CultureBank basic statistics after construction: for Tiktok, there are 12K cultural descriptors, 730 cultural groups, and 36 topics. Table 10 shows the topic distribution.
Table 8 shows the running time and the data volume after each stepfoot_5 . Statistics Metrics # cultural descriptors 11,754 Well-formatted 98.5% # cultural groups 730 Traceable 93.3% # cultural topics 36 Meaningful 84.5%
Table 2: CultureBank-TikTok statistics and quality.
To assess the dataset quality quantitatively, we select a random subset with 200 samples, and four researchers experienced in cultural topics annotated them (50 distinct examples per person, Kappa=0.8 on a held out set of 50 examples) for their (1) format (if the descriptor is well-formatted), ( 2) traceability (if it is possible to trace the cultural knowledge on the Internet) and ( 3) meaningfulness (if the descriptor provides meaningful cultural insights rather than generic ones) 9 . Since these descriptors are self-reported and the annotators may not come from the same cultural group, we use the term traceability instead of factualness: we perform Internet search, and as long as there is sufficient related information online, we consider it traceable and meaningful to be included. Table 2 shows Culture-Bank has well-formatted, traceable and meaningful cultural descriptors with moderate noise levels.
Table 6 shows qualitative examples in Culture-Bank. It presents interesting features, such as: ( 1) cross-culture behaviors: e.g., Americans in France experience culture shock in terms of electricity bills and driving habits; (2) linguistics variations: e.g., Americans use "chickpeas" or "garbanzo beans" interchangeably;
(3) diverse ethnic groups: e.g., Italian Americans identify themselves as Italian American with varying connection to Italy heritage; (4) recent cultural information: e.g., Chinese people heavily rely on mobile payment; and (5) cultural nuances hard to obtain from formal sources like Wikipedia: e.g., in South Africa, some people express frustration over having to calculate prices and taxes separately while others do not think so.
For the following evaluation ( §6) and fine-tuning ( §7) steps, we split CultureBank-TikTok by cultural descriptors into 9402 train, 1183 validation, and 1169 test samples.
With CultureBank-TikTok, we evaluate LLMs' cultural awareness. Prior work asks LLMs to answer cultural true/false questions (Fung et al., 2024). But LLMs are used in contextualized settings like a dialogue agent. So we propose a grounded evaluation, that grounds cultural knowledge in a real-world scenario, to test LLMs' ability to integrate cultural knowledge into their responses. We also perform classification-based direct evaluation in § D.2. Grounded data generation. For each descriptor in CultureBank, we first use a Mixtral-8x7B model fine-tuned on GPT-4-generated examples to generate a relevant consulting scenario, a client persona, and a grounded evaluation question. Then we employ a self-refinement method to improve the model generation based on two quality-control metrics at inference time. Figure 3 shows a generated example (For the "No tipping in Japan" descriptor, the grounded question is "what gesture says 'thanks you' in Japan?"). Human annotation shows 86% questions are correctly grounded on the original descriptor on 50 random samples. See §D.3 for more details. Grounded evaluation. As shown in Figure 3, we present the generated grounded question to the LLM for an answer. Given the answer, we perform (1) automatic evaluation that uses GPT-4 to judge if the answer entails the original cultural descriptor (entailment score); and ( 2) human evaluation where two experts compare answers from two LLMs and select the more culturally-aware one (win rate).
Llama-
2-7B-chat-vanilla 71.2 66.0 61.2 62.5 Llama-2-70B-chat 74.9 66.2 64.2 65.1 Mistral-7B-Instruct 72.9 67.2 63.4 64.5 Mixtral-8x7B-Instruct-vanilla 73.9 67.4 66.3 66.9 GPT-3.5 71.4 66.4 61.8 62.6 GPT-4 75.8 67.9 65.0 66.1 Fine-tuned models Mixtral-8x7B-CANDLE-60k 69.7 68.8 64.2 65.2 Llama-2-7B-SFT (Ours) 75.7 67.1 63.8** 64.7** Mixtral-8x7B-SFT (Ours) 73.3 70.3 66.6 67.5 Mixtral-8x7B-DPO (Ours) 72.4 70.5 68.1** 68.7** Table 3: Automatic evaluation on LLMs' cultural awareness, evaluated by knowledge entailment scores on our grounded evaluation benchmark by support. High support: cluster size > 50 (70 examples). Mid: cluster size between 20 and 50 (175 examples). Low: cluster size ↑ 20 (924 examples). We perform t-test between the vanilla model and the fine-tuned counterpart (** indicates p < 0.01). We bold the best results in vanilla models and fine-tuned models.
We evaluate open-source (Llama-2, Mixtral), close-source (GPT families (Achiam et al., 2023)), and our own fine-tuned models (See §7). See Table 11 for the model version details. All the results are on the test set.
Automatic entailment results. Table 3 shows the average entailment score of each model split by the cluster size (level of support). Mixtral-8X7B and GPT-4 are the best but still has a relatively low overall score of 66.9 and 66.1, suggesting room for improvements. Larger models have slightly better performance than their smaller versions. For more long-tailed cultural descriptors with fewer supports, the performance drops as expected.
7 Fine-tuning a More Culturally Aware Language Model
Our ultimate goal is to develop more culturallyaware language technologies. So we train on our CultureBank dataset to see if such a resource can improve LLMs' cultural awareness. Baselines. As a baseline, we trained a vanilla Mixtral model on an existing cultural knowledge base, CANDLE (Nguyen et al., 2023) with 60K cultural knowledge descriptions on five topics (Mixtral-CANDLE-60k). We note our models are only finetuned on 2K cultural descriptors, and due to the difference in data size, topic coverage, etc, such a comparison may not be fair. The comparison is only to show show the utility of CultureBank on specific datasets, and we appreciate all kinds of data efforts. The vanilla models without fine-tuning can also be viewed as baselines, and we perform t-test to show if the fine-tuned version is better. Training process. The training has two steps. First, we train a model on the 9402 cultural descriptors in the training set via supervised fine-tuning (SFT). In the second step, we select a 2K subset where the model performs poorly, and train on the grounded questions and answers augmented by the golden cultural descriptors in our CultureBank via SFT ("model-SFT (Ours)" in the tables) or DPO (Rafailov et al., 2024) ("model-DPO (Ours)") . See § E for more details. Automatic entailment results. Table 3 shows the results on the test set, which mostly contains outof-domain cultural descriptors, but the fine-tuned models can still achieve a better performance than their vanilla versions, suggesting that CultureBank can improve models' cultural awareness. Human-evaluated win rate. Figure 4 shows the human evaluation win-rate (two researchers, kappa=0.87). Compared to the base Mixtral model, Mixtral-DPO is more culturally aware 44% of the time, and equally good 38% of the time; compared to GPT-4, Mixtral-SFT wins 46% of the time and ties 38% of the time. This trend also aligns with the automatic evaluation in Table 3, indicating that the automatic entailment evaluation makes sense. Qualitatively (Table 15), we find that our fine-tuned 28% 46% 44% 58% 68% 38% 38% 20% 4% 16% 18% 22% 0% 20% 40% 60% 80% 100% Win, Tie, and Lose Rates by Model Win Rate Tie Rate Loss Rate Mixtral-DPO (Ours) vs GPT4 Mixtral-SFT (Ours) vs GPT4 Mixtral-DPO (Ours) vs Mixtral-Vanilla Mixtral-SFT (Ours) vs Mixtral-Vanilla Figure 4: Human evaluation on win rates between different LLMs (50 examples per pair) evaluated by humans on cultural-awareness in grounded scenarios. The two annotators achieved a Kappa score of 0.87. models generate shorter, and more culturally specific answers tailored to the user's inquiry (e.g., "in France, you might seek out artisanal cheese shops"), whereas RLHF-ed models like GPT-4 give generic templated answers (e.g., "1. Research local specialties, 2. Portion control...").
§D.3.3 has more details on the human evaluation, rubrics, models' win rates, and qualitative analysis.
Ideally, a more culturally-aware language model can achieve a better performance on different cultural tasks. So, we evaluate the fine-tuned models on two downstream tasks, to see if our CultureBank can help other cultural tasks in a zero-shot fashion. Downstream tasks. We choose two tasks: (1) GlobalOpinionQA (Durmus et al., 2023a), which has questions from world value survey (Inglehart et al., 2000) to measure how similar LLMgenerated responses are towards different countries;
(2) CultureNLI (Huang and Yang, 2023), which contains premise-hypothesis pairs labeled by two cultural groups, American and Indian. See Appendix E.2 for the detailed prompts and settings. Results. Table 4 shows the results. On both datasets, the models fine-tuned on CultureBank achieves a better performance than the vanilla counterparts (e.g., in GlobalOpinionQA, 79.5 VS. 81.8 for Mixtral-SFT; in Culture NLI, 59.9 VS. 61.5 in the US and 60.8 VS. 61.3 in India for Mixtral-SFT). These results suggest that CultureBank can be used to improve models' cultural awareness in downstream tasks even in a zero-shot setting.
There are different online communities, so it is important to test if our pipeline can be transferred to other platforms. So we apply our pipeline on Reddit, another online community. Table 9 shows the running time and data volume on Reddits. We made the following customization on the pipeline:
• Culture Relevance Classifier: similar to TikTok, we first search for tags like "#culturaldifference" on Reddit, but tags are not used frequently on Reddit. And because Reddit comments are much longer than TikTok comments, the TikTok-based Cultural Relevance Classifier does not work well for Reddit. Due to these reasons, we decide to directly search for cultural keywords on both submissions and comments to identify cultural contents. We first source discussions with tags related to cultural differences. Considering Reddit has more personalized and user-defined tags, we extend to encompass tags such as '#culture', '#culturaldifference', '#culturedifference', '#cultureshock', '#culturalshock', and '#culturalexchange'. Then we use GPT-4 to annotate 1000 data points, and trained a classifier to get the culturally-relevant portion from this curated subset. Finally, we obtain 7M comments after keyword filtering, send them to the trained classifier and get 2.6M cultural comments after the classification step. Considering the computation cost, we take a random subset of 528K cultural com-ments for the following processing steps. The top five subreddits with the highest frequency of relevant comments were "AskAnAmerican," "france," "explainlikeimfive," "AskEurope," and "thenetherlands". • Descriptor Extractor: to achieve a better performance, instead of using few-shot Llama-2 extractor, we fine-tune a Mixtral-based extractor on 1K GPT-4-generated extraction examples to extract structured cultural descriptors from Reddit comments.
Table 5 shows the basic statistics: CultureBank-Reddit contains 11K cultural descriptors and 2K cultural groups. Human annotation shows that it also contains high-quality data, suggesting that our pipeline can be easily generalized to a different platform.
Statistics Metrics # cultural descriptors 11,236 Well-formatted 95.5% # cultural groups 1,850 Traceable 94.0% # cultural topics 36 Meaningful 85.0%
Table 5: CultureBank-Reddit statistics and quality.
9 Recommendations for Culturally aware Language Technologies
Informed by results on CultureBank construction and analysis, cultural awareness evaluation, and fine-tuning, we outline insights towards future culturally-aware language technologies.
We show that fine-tuning on CultureBank can improve the cultural-awareness on various downstream tasks, so it remain critical to keep developing cultural knowledge databases. Data source. Prior work often relies on formal data sources and collapse different sources together: e.g., (Fung et al., 2024) started from WikiPedia and continued to scrape any related websites to construct cultural knowledge bases. But different data sources cover various aspects of culture: official documents like textbooks provide factual cultural knowledge, while online communities like social media offer insights on everyday cultural practices. So we should invite diverse data sources to capture the full spectrum of culture in the future. Besides, different data sources host different populations: Table 10 shows that topic-wise, CultureBank-Reddit contains more contents on community and identify, while CultureBank-TikTok is more about daily life like social norms and etiquette. So future datasets should keep data source as an important attribute to allow further analysis.
Data contents. Culture is multifaceted, so it is also important to factor in various dimensions.
Here is an example list of attributes to consider.
• Cross-culture behavior. In a globalized world, it is crucial to understand cross-culture behaviors to facilitate effective communication (Watkins, 2012). CultureBank contains some cross-culture behaviors but we need more efforts on it. • Perspectives. It is also important to track through whose lens we are looking at certain culture behaviors, as different perspectives may lead to different understanding of the same cultural practice (Iyengar et al., 1999;Brewer, 1999). • Time. Culture changes over time. CultureBank contains the time range associated with the cultural descriptors. Future efforts should also consider the time factor to enable temporal analysis. • Multilingual. Culture and language are deeply intertwined. But many existing cultural knowledge bases still rely on English. To capture the cultural nuances, in the future, we should develop multilingual multicultural knowledge banks. • Multimodality. Cultural knowledge goes beyond text information. So it is essential to include different modalities to capture the full spectrum of culture, from non-verbal communication cues, to rituals and arts, and so on in the future.
Data analysis. On data analysis, future research should consider temporal change rather than focusing on static data, as culture is evolving over time. For instance, we perform preliminary temporal analysis in § F.1 and find there are more discussions around studying abroad, LGBTQ+ rights, and technology over the years. Besides, existing research still categorizes culture by country, but we need to attend to more fine-grained cultural groups (ethnicity, generation, regions, ethnolinguistics, immigrants, socioeconomics, etc), to fully understand cultural diversity. Moreover, the study of cultural adaptation becomes important, as it reveals how culture changes in response to global influences. These focus areas -temporal dynamics, cultural group diversity, and adaptation processesoffers a comprehensive understanding of the fluid nature of culture in a globalized world.
We highlight two findings in evaluation. First, in our evaluation, humans also find it difficult to decide which model response is more culturally aware, partly because they are not from the presented cultural group. As we spend more effort on cultural data resources, it is also increasingly important to involve global annotators to enable more accurate evaluation. Secondly, as shown in our findings, direct and grounded evaluations give different results. So during evaluation, it is important to be more grounded on the end applications.
We realize that when fine-tuning models for cultural awareness, training only on the cultural knowledge or the grounded QA tasks could be insufficient. Take training a culturally-aware conversational assistant as an example. First, it requires appropriate cultural data grounded in multi-turn conversational settings. In addition, it requires a well-designed training paradigm to attend to the cultural nuances potentially implicit in the dialogue context. It also needs a solid evaluation method to rate the culture awareness of the generated responses, to help the model improve and evolve. Such a model needs to have a holistic view of the user cultural background, a personalized recognition of individual differences, and an inclusive mind for new cultural concepts and practices.
To conclude, our study introduces a generalizable pipeline for creating cultural knowledge bases from online communities. Using the pipeline, we develop CultureBank, a cultural knowledge database with 12K cultural descriptors sourced from TikTok, and 11K from Reddit. CultureBank features agreement levels for nuanced cultural interpretation and contextualized scenarios for grounded evaluation.
With CultureBank, we assess the cultural awareness of various LLMs, showing room for improvement. Further, fine-tuning an LLM with Culture-Bank leads to better performance on downstream cultural tasks, which showcases the potential of CultureBank. Finally, drawing from our findings, we close the paper by presenting insights towards future culturally-aware language technologies.
Despite our efforts, we recognize several limitations of our work. First, in our pipeline, we utilize open-source LLMs in various steps. Constrained by opensource LLMs' ability to process non-English languages, we process English-only data. But many cultural nuances cannot be fully expressed or captured by English. This limitation inherently restricts our ability to grasp and represent the full spectrum of cultural contexts and meanings, potentially leading to oversimplifications of certain cultural aspects. Besides, although we attempt to minimize bias with various efforts, these open-source LLMs could still extract biased information, and generate biased summarization, which could lead to biased final outcome in the data. Also, new models are released very frequently and they (e.g., Mistral) are usually better than the older models (e.g., Llama-2), we used the best available models at the time of experiments for better performances (e.g., when we were at the summarization step, Mixtral came out and was better than LLaMA2, so we decided to use the best model instead of an older one), but these open-source LLMs might be trained on different datasets, which could also have an impact on the extracted outputs.
Second, our dataset is subject to sample bias. Because we scrape the data with certain keywords and hashtags like "cultural difference" and "culture shock". Oftentimes people only post on online platforms when they have strong reactions such as surprise or shock towards certain cultural phenomena. This bias means that our findings might overemphasize aspects of cultural difference that are more likely to stand out to individuals, while underrepresenting more mundane or universally shared aspects of culture. Such a bias can skew the perception of cultural diversity and difference, potentially reinforcing stereotypes or overlooking the subtleties of cultural exchange and adaptation.
Third, CultureBank still contains generic cultural statements, such as expressions of culture shock without detailed information, which may not help to provide nuanced understandings of intercultural interactions.
In this work, we construct a cultural knowledge base from online communities. Given the large size of the dataset, we acknowledge that stereo-types, controversial, and negative content may still exist in our dataset, despite our rigorous efforts to filter the data and minimize the impact of such content. We want to emphasize that the cultural descriptors in CultureBank are not intended to reflect, nor should they be interpreted as reflecting, the personal views or opinions of the authors or the online platforms. We call for a better approach for content moderation in the future and hope that researchers will use our data with a discerning perspective, and always consider the broader implications of its application and the potential for reinforcing harmful biases.
We also recognize the responsibility that comes with handling cultural data, especially from diverse and broad communities like those on TikTok and Reddit. In our method, we have strived not only for technological innovation but also for a conscious approach that respects the dignity, privacy, and cultural sensitivities of individuals and groups represented in the data. This includes anonymizing data where possible, ensuring compliance with platform terms of service, and engaging with ethical guidelines that govern research in social sciences and humanities.
In conclusion, while we acknowledge the limitations and challenges inherent in our work, we believe in its potential to contribute positively to the field of culturally-aware language technology. We encourage the community to join us in these efforts, to promote cultural diversity, inclusivity, and sensitivity. We discuss limitations of this work in §10.
Table 6 shows qualitative examples in our Culture-Bank-TikTok, and Table 7 presents the taxonomy of current popular datasets.
We present more details on the construction pipeline in this section. Culture relevance classifier. The classifier achieves an accuracy of 79% on a held out test set with 100 examples. For the cultural class, precision is 0.69, recall is 0.78, and f1-score is 0.73; for the non-cultural class, precision is 0.86, recall is 0.79, and f1-score is 0.83.
Cultural descriptor extractor. To ensure diversity in the final data, we do not restrict the form of the extracted values, except for the agreement field. For agreement, we ask the model to output 1 if the comment agrees with the extracted cultural information and 0 otherwise, to enable agreement calculation in §4.3.
Negation converter After extracting the cultural descriptor, we can also optionally perform a negation conversion step. This is because cultural descriptions can be diverse and dynamic, it is important to calculate the agreement levels to certain cultural behavior inside a cultural group. This requires both positive and negative examples for one cultural behavior. But due to the limitation of clustering algorithms, opposite sentences such as "Japaneses do not give tips" and "Japaneses give tips" may be placed into the different cluster in latter steps. So after the extraction and before clustering, we use different heuristics to convert negative sentences to their positive forms, and also flip the agreement field, i.e., "Japanese do not give tips, agreement=1" will become "Japaneses give tips, agreement=0", to help future clustering and agreement calculation steps.
Clustering. To manage the computation, we perform two rounds of clusterings. We first conduct Hierarchical Agglomerative Clustering (HAC) on the cultural groups to identify similar ones. Then we enumerate the identified cultural groups. Within each cultural group, we concatenate the extracted
Table 7: The coverage between our taxonomy an taxonomies proposed in previous literature. We list the equivalent field names in other work. Our taxonomy can cover fields proposed in existing datasets, and also proposes more details fields such as "agreement", "relation", "other description" and so on.
fields, encode the concatenated sentence with Sen-tenceBert (Reimers and Gurevych, 2019), and perform another round of HAC clustering. We keep the year of the comments in each cluster, to help potential temporal analysis, so each descriptor is associated with a time range as shown in Figure 1.
To ensure enough supporting evidence for each cluster, we remove clusters with less than 5 data points, and discretized the cluster size into 10-unit intervals like [10, 20) as the support value for each cultural descriptor. Summarization. To summarize each cluster into a single, high-level cultural indicator, we leverage a Mixtral 8x7B model finetuned on 1k summarization examples generated by GPT-4. To evaluate the performance of our summarizer, we performed human annotation on the following two metrics:
• Fidelity: extent to which the summary accurately represents the main ideas, facts, and figures from the original text without introducing inaccuracies or distortions;
• Coherence: How well the summary flows from one sentence to another, maintaining logical progression and clear connections between points.
Normalization on cultural group and topic.
Since we leverage LLMs to perform cluster summarization, and models can extract and summarize to generate any information, similar values may appear in different forms, for instance, "Aboriginal Australians" and "Indigenous Australians" are synonyms, and "fashion and attire" and "clothing" are two similar topics. Having such synonyms makes it hard to query and manage our dataset, so we normalize values in the two fields of "cultural group" and "topic", to collapse synonyms. Similar to the previous clustering step, we use HAC to group synonyms in cultural groups together and take the majority vote as the representative cultural group. Grouping together similar topics turns out to be a more involved task, as many topics overlap with and differ from each other in subtle ways, so embedding-based clustering methods (e.g., HAC) often yield noisy results that fail to align with human judgements. Instead, we use an LLM to detect 100 recurring themes from our original topics, and then manually merge, prune, and rephrase these themes into 36 high-level cultural topics. Finally, we ask an LLM to classify each summarized cluster into one of these topics. Table 10 shows the details of our final list of 36 cultural topics.
Agreement calculator. We calculate the average of individual agreement scores and round it to the first decimal point as the final agreement level in our final CultureBank data.
Content moderation. The table below show examples of the keywords we used for the filtering.
To ensure a high quality in content moderation, we manually review all potentially controversial content flagged by the keyword list, the classifier, and the Perspective API. This manual evaluation is conducted by authors who are trained and experience people in controversial content detection, with the following rubric: Keywords: shooting, alcohol, drug, prison, legal, negative, sentiment, illegal, cult, slang, wtf, nudity, shame, safety, cleavage, sex, communism, socialism, racist, danger, gun, swinging, polyamorous, n-word, colorism, ...
We will annotate and remove the following contents:
1. very personal experience (e.g., toxic behavior people experience in family), 2. very subjective opinions (e.g., negative sentiments/opinions about something),
3. controversial and non-culture related, 4. related to discrimination, bullying, or the use of derogatory terms
Listing 2 shows the prompt used for the descriptor extraction step, and Listing 3 shows the prompt used for the cluster summarize.
Table 8 and 9 shows the running time and data volumn after each step for TikTok and Reddit.
Table 10 shows the topic distribution in Culture-Bank.
Table 11 shows the mapping between the model name mentioned in the paper and exact model versions and how they are trained.
Prior work (Fung et al., 2024) evaluates LLMs' culture awareness via directly asking true/false ques-
# Step (TikTok) Time # gpus Output/Data size Input: total comments --720K 1 Cultural relevance classifier 2 hours 1 400K 2 Cultural descriptor extractor 4 days 32 400K 3 Clustering 2 hours 1 13K 4 Cluster summarizer 13 hours 4 13K 5 Agreement calculator < 1 min 0 13K 6 Content moderation 3 hours + human annotation 0 12K Table 9: Reddit running time of each processing step and the amount of data afterwards.
tions. Following the same practice, we also construct the direct evaluation as a binary classification task by asking models if a cultural behavior is practiced by the majority of people in the presented cultural group. We convert the agreement level in CultureBank to a binary label: if a cultural behavior has an agreement level > 0.5, then we label it as positive. Otherwise, we label it as negative.
Listing 4 shows the prompts for direct evaluation.
Table 12 shows the macro F1 as we have highly unbalanced agreement levels. We also further categorize the cultural descriptors into three bins: high support (cluster size > 50), mid support (cluster size between 20 and 50), low support (more longtailed behavior with cluster size ↑ 20).
Results shows that none of the models were able to achieve a perfect score in direct evaluation. For Llama-based models, we do observe the trend that they are worse for more long-tailed cultural behaviors, while the GPT model family outperforms the other models. Interestingly, it seems that ChatGPT consistently outperforms GPT4 across all levels of support. Qualitative analysis found that it is partially because there is often times no clean-cut thresholds on whether a cultural behav-ior is a common practice, and in ambiguous cases GPT4 tends to behave more conservatively, and this demonstrates one major limitation of such a simplistic approach of direct evaluation. Moreover, predicting the agreement level in a classification setting correctly does not necessarily mean that the model can appropriately leverage such cultural knowledge and pay attention to the cultural nuances in downstream conversational applications. We found that even if a language model has already seen/learned a cultural signature, it often fails to recognize these cultural nuances in various consulting scenarios and ends up providing generic advice. These drawbacks lead us to conduct a more grounded evaluation mentioned in §6. 5 shows a more detailed procedure to generate scenarios, personas, and questions for the grounded evaluation. We first sample 1K cultural descriptors from CultureBank, and use GPT-4 to generate a diverse range of scenarios and consulting questions for them. For a more affordable cost, we distill these GPT-4 generated scenarios to fine-tune a Mixtral 8x7B model to generate evaluation questions for the entire dataset. Given a piece of cultural knowledge, we ask the model to generate: 1. a consulting scenario, 2. a client persona, and 3. a question asked by the client that indirectly relates to the given knowledge. Listing 5 shows the prompt used to generate questions in the grounded evaluation.
Self-refinement. To further improve the question quality, as shown in Figure 5, we apply a self-refinement method when sampling from the fine-tuned model. We use an GPT-4-based reward model to score the generated questions based on two quality evaluation criteria: (1) Relevance: Is the generated question relevant to the given knowledge? (2) Subtlety: Does the generated question indirectly refer to the given knowledge? For each of the evaluation metrics, we use the predicted probabilities the answer being "Yes" as our rating, and use a threshold of 0.95 for Relevance and 0.8 for Subtlety. Listing 6 shows the prompt used for the question evaluation. Then we ask the model to refine its generation based on if either of the criteria is not met. Lastly we use this improved model to generated data for the entire set. Table 13 and 14 shows two generated examples.
To evaluate LLMs' culture awareness, we ask the LLM the generated question grounded in a practical scenario. Listing 7 shows the prompt used for the grounded evaluation. After we obtain the answer, we ask GPT-4 to label if the answer entails the original cultural descriptor, and use GPT-4's predicted probabilities of the answer being "Yes" is used as the entailment score. Listing 8 shows the prompt used for the knowledge entailment task.
<s > [ INST ] You are a helpful AI assistant trained to help users on their personal issues . Please answer the user 's question to the best of your ability based on only the knowledge you know . Do NOT make up any unfounded
In Japan, particularly in restaurants, it is customary for customers to attempt to tip service staff as a gesture of gratitude, despite it being considered rude or unnecessary in Japanese culture. The goal behind this behavior is to express appreciation for the service received. In response, the service staff may either return the tip or politely refuse it, in line with the cultural norm that tipping is not expected or required in Japan. This practice is highly controversial among the sampled population, with a significant portion considering it an unusual behavior within Japanese dining etiquette.
A curious and generous American traveler, planning a first-time trip to Japan, eager to show appreciation for the country's rich culture and impeccable service, yet unsure about the proper etiquette in regards to tipping
Question I'm planning my first trip to Japan and I've heard so many great things about the hospitality and service there. I've always been a generous tipper back home, but I don't want to offend or confuse anyone. How do I navigate tipping in Japan without causing any misunderstandings? Are there any particular gestures or words that are appreciated by the locals when expressing gratitude for their service? ------------------------------------User Prompt : Response : { model_response } Knowledge : { cultural_knowledge_description } Does the given response entail the provided knowledge ? Entailment ( Yes / No ):
Listing 8: Prompt for calculating knowledge entailment in grounded evaluation
To calculate the model win rates, we employed a single-blind human annotation process involving two annotators and 200 randomly selected results.
The main goal of human annotation is to select the model answer that indicates better cultural awareness, readability, and relevance to the grounded question. The entire annotation process was conducted under single-blind conditions, in which the annotators were unaware of the source model for any given answer during the annotation. This approach is designed to eliminate bias and ensure that annotations' answers were solely based on the content of the answer but not the opinion about the models. Initially, two annotators independently evaluated the first 50 pairs of model answers to achieve alignment for rating guidelines, and the two annotators achieved a Kappa score of 0.87, which indicated a high agreement between the annotators. Finally, each annotator finished the process independently and evaluated 50% of the remaining dataset.
Qualitative insights from human evaluation. Table 15 shows a qualitative dialogue example generated by our fine-tuned model and GPT-4. Here are some qualitative insights from the annotation process.
• Our finetuned models tend to give shorter and more straight-to-the-point answers, whereas the standard RLHF-ed models almost always
In Icelandic culture, especially in cold weather and outdoor settings, it is customary for parents to leave their babies outside for naps, allowing them to enjoy the fresh air and sleep peacefully. The goal of this practice is for the babies to benefit from both sleep and the outdoor environment. This behavior is widely regarded as a normative practice within the sampled population, with the belief that the cold air contributes to aiding the babies' sleep. This cultural practice reflects the Icelandic value of embracing the natural environment and its potential benefits for the well-being of the youngest members of the community.
A warm-climate native newly employed in Reykjavik, concerned about the well-being of colleagues' children during outdoor team-building events in the cold weather
Question Hey there, I recently started working in Reykjavik after moving from a much warmer climate. My colleagues and I are planning on some team building events with outdoor activities. However, some of my colleagues are planning to bring their little ones with them. It's quite chilly outside, and some of the kids are really young. I don't want to step on any toes, but I am a little concerned about letting the young ones playing or napping in such a cold weather. What should I do?
Table 14: Another example of the generated scenario, persona, and question grounded on a cultural descriptor.
provide a long list in their responses, which are often more comprehensive but less specific to the user's inquiry.
• Our finetuned models generally provide more culturally-specific advice tailored towards the user's question and cultural context, while the standard RLHF-ed models, especially the GPT family, sometimes fail to acknowledge the cultural aspects and output generic guidelines instead.
• Even though GPT-4 is quite knowledgeable with a variety of cultures if being directly asked about them, it sometimes fails to integrate the knowledge into the downstream consulting scenarios in our grounded evaluation. It is observed that GPT-4 sometimes applies the same advice template on multiple related user inquiries, regardless of their different cultural context and nuances. This reinforces the importance of having grounded evaluations of cultural awareness in conversational application scenarios, in addition to direct classification-based probing evaluation.
E Fine-tuning Culturally-aware Language models
Our model fine-tuning process contains two phases.
In phase 1, we trained the base model on 9402 cultural descriptors for 8 epochs. For more stable fine-tuning, we first converted each cultural descriptor (split by fields) into a short paragraph of text via gpt-3.5-turbo before using it to train our model. The prompt template for this conversion is available in Listing 9.
In phase 2, we then sampled 2000 grounded questions from the training set where the base model performed poorly according to the reward model (i.e., with knowledge entailment score < 0.6). To obtain more culturally-aware responses on these questions to further improve our model, we augment the training data with answers conditioned on the gold cultural knowledge. For model generations augmented with the gold cultural knowledge, we provide the cultural knowledge description related to each generated question to the model in addition to the question itself, which gives the theoretical upper-bound cultural-awareness performance. Listing 10 shows the prompt used for the augmented upper bound model. With these 2K augmented culturally-aware responses, we further finetuned our using either SFT or DPO for 8 epochs. E.2 Downstream cultural tasks E.2.1 Global Opinions QA We follow the same evaluation settings as the original paper (Durmus et al., 2023b). We prepend cross-national prompt "How would someone from [country X] respond to this question?" for each multiple-choice question paired with a given country query. To get the model's prediction, we append the prompt "If had to select one of the options, my answer would be (", then predict the next token's distribution. Listing 11 shows the prompt used.
For each multiple-choice question q, denote the distribution of the respondents from country c as P c (q|c), and model's predicted distribution for country c as P m (q|c), we use Jensen-Shannon Distance (JSD) to evaluate the distance between p c and p m . Specifically, we calculate
The definition of JSD is:
(2) in which D represents Kullback-Leibler divergence, and M = 1 2 (P + Q) is the midpoint distribution between P and Q.
A higher cultureSim score represents a higher cultural similarity between the model's response and the country queried.
In addition, to measure how much the model's representation is biased towards a certain subset of countries, we calculate the Standard Deviation Errors of the cultureSim score across different countries. The lower the score, the more balanced the model's representation.
For the evaluation on CulturalNLI, we prompt the models with cultural indicators (e.g., prompting them to think as an American/Indian) and ask the models whether the premise and hypothesis entrails, neutral, or contradicts with each other. Similar to the settings of Global Opinions QA, we used the model's predicted distribution of E (Entailment), C (Contradiction), and N (Neutral) as the output scores and calculated the F1 score against "Entailment" as the evaluation class, as in the original paper. Listing 12 shows the prompt used. Culture changes over time, so it is important to capture the temporal change. Since Reddit contains many historical data (2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015)(2016)(2017)(2018)(2019)(2020)(2021)(2022), we perform a preliminary temporal analysis on CultureBank-Reddit by searching for related keywords in Cul-tureBank-Reddit. Each descriptor has comments from different years, and we slice the year to form a snapshot of that year. As shown in Figure 6, over the years, more people are studying abroad; the discussion on LGBTQ+ rights has gained more attention and community support; besides, technological advancements also affect cultural practices, e.g., people mention a notable shift within Dutch society towards cashless payment and a growing acceptance among Swiss society. We also notice that some cultural practices would decline due to technological innovation: for instance, because of streaming services and digital downloads, the culture related to DVDs has been eclipsed. The dataset mentions that from the 1990s to the early 2000s, DVDs were the predominant mediums for video playback, but there has been less and less discussion about DVD culture recently.
Additionally, some cultural practices are more sensitive to temporal dynamics. In Figure 7, we discuss some entertainment topics as examples. Hiphop music became more popular in the late 2000s and we see an increasing amount of discussion on Hip-hop music in the figure. In our dataset, there is a discussion on the evolution and adaptation of the French rap and hip-hop scene which responds to shifting trends within the rap music genre. Similarly, it is noted that superhero movies and comics from Marvel and DC have developed multicultural narratives, featuring a diversity of characters and storylines. This development reflects the evolving culture of superheroes and underscores the impact of societal trends toward greater inclusivity and diversity in this genre.
The work was done when WS was at Stanford. We release the CultureBank dataset, code, and models at https://github.com/SALT-NLP/CultureBank.
Our project page: https://culturebank.github.io/.
https://perspectiveapi.com/
https://microsoft.github.io/presidio/analyzer/
https://developers.tiktok.com/products/research-api/
In the main text, we focus on Tiktok, but show in §8 the pipeline can be generalized to Reddit, another platform.
For example, "US people experience culture shock when traveling" is generic insight because it does not give any detail on what makes people feel culture shock; "US people notice an measurement unit difference when traveling" is meaningful, because it mentions the specific detail (measurement unit).