LLMs as Cultural Archives: Cultural Commonsense Knowledge Graph Extraction

Junior Cedric Tonga, Chen Cecilia Liu, Iryna Gurevych, Fajri Koto · 2026

Read LLMs as Cultural Archives: Cultural Commonsense Knowledge Graph Extraction by Junior Cedric Tonga, Chen Cecilia Liu, Iryna Gurevych, Fajri Koto (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-17971.txt.

Introduction

Junior Cedric Tonga, Chen Cecilia Liu, Iryna Gurevych, and Fajri Koto turn language models into sources of cultural knowledge. Their 2026 preprint starts from a simple fact. Models soak up cultural know-how from vast and varied web text. That know-how is rich, but it sits inside the model in a loose, hidden form. Hidden knowledge is hard to check, share, or build on.

The team draws the knowledge out into a graph of cultural commonsense. Nodes and links state what people in a culture take for granted. An iterative process pulls the facts out step by step. A structured graph can serve search, teaching, and cross-cultural tools. It also lets scholars inspect what the model has absorbed.

Catalog record

CREATORJunior Cedric Tonga, Chen Cecilia Liu, Iryna Gurevych, Fajri Koto
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE57 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-17971

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

LLMs as Cultural Archives: Cultural Commonsense Knowledge Graph Extraction Junior Cedric Tonga, Chen Cecilia Liu, Iryna Gurevych, Fajri Koto Abstract Large language models (LLMs) encode rich cultural knowledge learned from diverse web-scale data, offering an unprecedented opportunity to model cultural commonsense at scale. Yet this knowledge remains mostly implicit and unstructured, limiting its interpretability and use. We present an iterative, prompt-based framework for constructing a Cultural Commonsense Knowledge Graph (CCKG) that treats LLMs as cultural archives, systematically eliciting culture-specific entities, relations, and practices and composing them into multi-step inferential chains across languages. We evaluate CCKG on five countries with human judgments of cultural relevance, correctness, and path coherence. We find that the cultural knowledge graphs are better realized in English, even when the target culture is non-English (e.g., Chinese, Indonesian, Arabic), indicating uneven cultural encoding in current LLMs. Augmenting smaller LLMs with CCKG improves performance on cultural reasoning and story generation, with the largest gains from English chains. Our results show both the promise and limits of LLMs as cultural technologies and that chain-structured cultural knowledge is a practical substrate for culturally grounded NLP. 1 1 1 Code available at https://github.com/JuniorTonga/Cultural_Commonsense_Knowledge_Graph Figure 1: Application of our framework for constructing a partial Cultural Commonsense Knowledge Graph (CCKG) capturing culturally grounded reasoning about breakfast in Indonesia. Given an input prompt specifying the subtopic, language, country, and task-specific constraints, GPT-4o generates English if–then commonsense assertions (\textit{action}_{i},\textit{relation},\textit{action}_{j}) to form an initial knowledge base (KB). Assertions with relations ( xNext , oNext ) are iteratively expanded by re-prompting GPT-4o to generate intermediate action expansions that decompose \textit{action}_{i} into finer-grained steps leading to \textit{action}_{j} and forward actions occurring after \textit{action}_{j} . In this example, only the first assertion in the expansion list is expanded for a single iteration. The resulting assertions are added to the KB, post-processed and composed into the final CCKG subgraph. 1 Introduction Culture and commonsense reasoning are deeply intertwined, as culture shapes how people interpret everyday situations, social conventions, and causal regularities Koto et al. 2024 ; Sadallah et al. 2025 ; Sap et al. 2020 . Culture encompasses shared and learned values, norms, and practices that guide interpretation and action within a community Hershcovich et al. 2022 ; Adilazuarda et al. 2024 ; Liu et al. 2025 . Because commonsense is grounded in cultural experience rather than universal logic, what seems self-evident in one community can be unfamiliar, or even misleading, in another Naous et al. 2024 ; Almheiri et al. 2025 . While early computational approaches largely treated commonsense as culture-neutral Bisk et al. 2020 ; Sap et al. 2019b , recent works have begun to recognize and model its cultural dimensions, highlighting the linguistic and cognitive variation that arises across communities Koto et al. 2024 ; Sadallah et al. 2025 . We argue that language models can reason more appropriately across contexts when equipped with structured representations of culture, especially given its complexity and uneven representation in online data. Early efforts such as ATOMIC Sap et al. 2019a demonstrated the value of modeling inferential knowledge between everyday events Hwang et al. 2021 ; Bosselut et al. 2019 , yet these resources remain largely Western-centric and lack cross-cultural generalizability. Extending this idea to cultural contexts requires uncovering and organizing the implicit cultural knowledge within LLMs into interpretable, multi-step structures that reflect local norms and practices. In this work, we study how far large language models serve as cultural archives—examining the extent and accuracy of the cultural information they encode. Since large language models are pretrained on culturally diverse corpora Jiang et al. 2021 ; Sun et al. 2024 ; Brown et al. 2020 , much of this knowledge already exists implicitly and can be systematically extracted in structured form. Specifically, we address three research questions: (1) To what extent do LLMs encode cultural knowledge that aligns with real-world cultural relevance? (2) Which language best represents cultural knowledge when extracted from LLMs? and (3) Can the extracted cultural knowledge be used to enhance the cultural reasoning ability of smaller or weaker models? While our initial hypothesis assumes that the language spoken by the culture provides the most authentic representation, our findings reveal an interesting contrast: English often serves as a more coherent medium for representing cultural knowledge graphs. Prior work has constructed cultural knowledge bases from sources such as Wikipedia, CommonCrawl, and social media Nguyen et al. 2024 ; Fung et al. 2024 ; Shi et al. 2024 . While valuable, these approaches face two key limitations. First, they represent culture as isolated, static facts Yin et al. 2022 ; Nguyen et al. 2024 ; Fung et al. 2024 , overlooking the procedural and contextual nature of cultural practices. Many traditions consist of ordered sequences of actions—such as the proposal-to-marriage process in Indonesian weddings—and reducing them to atomic statements omits crucial context, limiting applications like reasoning and story generation. Second, most cultural knowledge bases are built in English, despite the fact that many cultural nuances are best expressed in native languages. Our contributions can be summarized as follows: • We propose an iterative, prompt-based framework for constructing Cultural Commonsense Knowledge Graphs (CCKG) , which extract multilingual, culture-specific if–then inferential knowledge chains from large language models. Following the ATOMIC -style formulation Sap et al. 2019a , we compose these relations into multi-step cultural knowledge chains (Figure 1 ). • Through extensive human evaluations of cultural relevance, correctness, and path coherence, we find that while native languages capture richer cultural detail, English extractions are more coherent and consistently preferred. • We show that augmenting smaller or weaker LLMs with CCKG improves their cultural reasoning and story generation performance, highlighting the value of inferential cultural knowledge for developing culturally grounded NLP systems. 2 Related work Knowledge Bases for Commonsense Reasoning. Early commonsense knowledge bases (KBs) such as WebChild Tandon et al. 2014 linked nouns with descriptive adjectives to encode physical and perceptual properties, providing one of the first large-scale automatically constructed commonsense resources. ConceptNet Speer et al. 2017 aggregated human-authored assertions but primarily captured lexical relations between words rather than inferential or situational knowledge. Quasimodo Romero et al. 2019 further expanded coverage by mining commonsense assertions from QA forums and web text through large-scale automated extraction. ATOMIC Sap et al. 2019a marked a shift toward event-level reasoning, introducing large-scale if–then relations that capture causes, effects, and social intentions. While these knowledge bases advanced commonsense modeling, they remain largely English-centric and culturally neutral. Recent work has begun integrating culture into knowledge representation: CANDLE Nguyen et al. 2023 performs large-scale web extraction to probe factual knowledge across regional contexts, Mango Nguyen et al. 2024 extracts cross-cultural facts from LLMs, CultureAtlas Fung et al. 2024 constructs a cultural knowledge base from Wikipedia and Wikidata, and CultureBank Shi et al. 2024 collects cultural descriptors from social media such as TikTok and Reddit. However, these efforts primarily focus on factual or descriptive knowledge—typically short, unstructured snippets without inferential chains or sequential reasoning. In contrast, we treat large language models as cultural archives, leveraging their internalized knowledge to construct structured, inferential representations. We introduce the Cultural Commonsense Knowledge Graph (CCKG)–a graph-based resource that models if–then reasoning chains involving human actions, intentions, and consequences. Evaluating Cultural Commonsense Knowledge. Early research on

Read the complete text