Investigating Cultural Alignment of Large Language Models (2402.13231v2)

Abstract: The intricate relationship between language and culture has long been a subject of exploration within the realm of linguistic anthropology. LLMs, promoted as repositories of collective human knowledge, raise a pivotal question: do these models genuinely encapsulate the diverse knowledge adopted by different cultures? Our study reveals that these models demonstrate greater cultural alignment along two dimensions -- firstly, when prompted with the dominant language of a specific culture, and secondly, when pretrained with a refined mixture of languages employed by that culture. We quantify cultural alignment by simulating sociological surveys, comparing model responses to those of actual survey participants as references. Specifically, we replicate a survey conducted in various regions of Egypt and the United States through prompting LLMs with different pretraining data mixtures in both Arabic and English with the personas of the real respondents and the survey questions. Further analysis reveals that misalignment becomes more pronounced for underrepresented personas and for culturally sensitive topics, such as those probing social values. Finally, we introduce Anthropological Prompting, a novel method leveraging anthropological reasoning to enhance cultural alignment. Our study emphasizes the necessity for a more balanced multilingual pretraining dataset to better represent the diversity of human experience and the plurality of different cultures with many implications on the topic of cross-lingual transfer.

Cultural Alignment in LLMs: A Detailed Examination

The paper "Investigating Cultural Alignment of LLMs" offers an in-depth exploration into the degree to which LLMs reflect the cultural knowledge and values specific to different societies. The paper bridges the domains of linguistic anthropology and artificial intelligence to elucidate whether LLMs, which are often heralded as comprehensive repositories of human knowledge, can genuinely align with the diverse cultural paradigms embedded in various languages.

Overview of Methodology

The authors propose a novel framework to assess cultural alignment, operationalizing it by simulating sociological surveys and using cultural personas as a reference. These personas encapsulate different demographic attributes such as age, social class, and education. The core method involves probing multiple LLMs with prompts that mirror a variety of sociodemographic personas, derived from survey responses collected in Egypt and the United States. The authors employ a variety of LLMs, including GPT-3.5, mT0-XXL, LLaMA-2-Chat, and AceGPT-Chat—each characterized by distinct pretraining language distributions.

Key Research Areas

Four primary research questions guide the investigation:

1. Prompting Language Impact: The hypothesis posits that prompting an LLM with a culture's native language enhances cultural alignment compared to using a secondary language.
2. Pretraining Data Composition: It is hypothesized that models pretrained with a predominant focus on the data of a particular culture will align more closely with that culture's survey results.
3. Profile Representation and Variability: The authors explore whether LLMs show higher misalignment for underrepresented backgrounds and culturally sensitive topics, using personas to simulate diverse demographic variables.
4. Cross-Lingual Transfer through Finetuning: By examining the effects of finetuning models in a secondary language, the paper seeks to understand cross-lingual knowledge transfer capabilities.

Results

The results reveal significant findings about cultural alignment:

• Cultural Bias: It was found that models exhibit a notable Western bias. Even models positioned as multilingual or specifically finetuned on Arabic cultural data demonstrate more substantial alignment with US survey responses than with those from Egypt.
• Influence of Language in Prompting: For some models, using the dominant language of a culture in prompts significantly improved the alignment, especially for GPT-3.5 and AceGPT-Chat. However, this was less effective for models like LLaMA-2-Chat, primarily pretrained on English.
• Disparity in Demographic Representation: Findings indicated that LLMs captured a narrower spectrum of responses from digitally underrepresented groups. The alignment was significantly lower for personas representing lower social classes and educational levels.
• Anthropological Prompting: A novel method termed Anthropological Prompting was proposed, harnessing anthropological reasoning to improve cultural alignment. Encouraging the model to consider nuanced social contexts resulted in better alignment with underrepresented groups.

Implications and Future Directions

The implications of this paper are profound, emphasizing the urgent need for balanced multilingual datasets in the pretraining of LLMs and the importance of culturally diverse pretraining data. The paper also opens new pathways for improving cross-lingual knowledge transfer, a critical capability for developing more culturally adept LLMs.

In future research endeavors, expanding the data sources and languages, incorporating non-script-based languages, and further refining anthropological prompting mechanisms are promising. This work sets a foundational trajectory for designing AI systems that can ethically and effectively operate within diverse cultural landscapes, highlighting the intersection of computational models with anthropological insights.