SEADialogues: Culturally Grounded Dialogue Dataset
- SEADialogues is a multilingual, culturally grounded dialogue dataset featuring 32,000 persona-rich conversations in eight Southeast Asian languages.
- It embeds local cultural references, honorifics, and natural topic transitions to create authentic multi-turn dialogues.
- Designed with rigorous human annotation and evaluation, it advances research in culturally aware, human-centric dialogue systems.
SEADialogues is a multilingual, culturally grounded, persona-rich multi-turn dialogue dataset for Southeast Asian languages. It was introduced to address a persistent limitation of existing dialogue resources: many are English-centric, rely on translation from English, or omit the local cultural texture of everyday conversation. The dataset contains 32,000 dialogues in eight languages from six Southeast Asian countries, and is intended to support research on culturally aware and human-centric LLMs, including conversational dialogue agents (Kautsar et al., 9 Aug 2025).
1. Conceptual scope and research motivation
SEADialogues is positioned at the intersection of multilingual dialogue modeling, persona conditioning, and culturally grounded generation. Its motivating claim is that dialogue systems should not only be fluent and personalized, but also culturally aware in ways that existing dialogue datasets generally are not. The paper argues that translation-based multilingual dialogue resources often produce unnatural conversations and fail to capture everyday local referents such as food, celebrations, folklore, cinema, TV, music, travel habits, communication style, honorifics, and region-specific named entities.
In the paper’s comparison table, SEADialogues is contrasted with DailyDialog, MultiWOZ, PERSONA-CHAT, XPersona, GlobalWOZ, Multi3WOZ, and XDailyDialog. Those datasets may be multi-turn, multilingual, or persona-driven, but SEADialogues is described as the only one in that comparison that explicitly embeds cultural relevance into each conversation while also avoiding pure translation from an existing dataset. The dataset is also framed as especially important for the Global South, where culturally grounded dialogue generation has been understudied (Kautsar et al., 9 Aug 2025).
A central conceptual point is that culture is not treated as a superficial label. The dataset does not merely replace generic place names with local ones. Instead, it grounds dialogue in ordinary and festive life, local media, mythological references, regional customs, and politeness norms. This gives SEADialogues a more specific objective than generic open-domain conversation modeling: it targets dialogue generation that reflects linguistic and cultural realities of Southeast Asian communities.
2. Dataset composition and dialogue structure
SEADialogues contains 32,000 dialogues evenly distributed across eight languages: Indonesian, Javanese, Minangkabau, Thai, Malay, Vietnamese, Tamil, and Tagalog. The paper repeatedly states that these dialogues span six Southeast Asian countries. Each language contributes 4,000 dialogues, and the generation process uses four models, with each model contributing 1,000 dialogues per language.
The dataset statistics reported in the appendix are as follows:
| Property | Value |
|---|---|
| No. of languages | 8 |
| No. of dialogues | 32,000 |
| Average dialogues per language | 4,000 |
| Average utterances per dialogue | 13.86 |
| Average words per utterance | 21.69 |
| No. of topics | 100 |
| Topics per dialogue | 2 |
| No. of scenarios | 300 |
| No. of personas | 210 |
Each example is a conversation between two people. The prompt requires a multi-turn conversation in the target language, formatted as name: utterance. The dialogue must involve two speaker personas, discuss Topic 1, then transition to Topic 2, remain in a polite setting, use honorifics, include a marked topic change with the token [TRANSITION], and stay within a specified maximum number of turns. This design makes the resource multi-turn not only because each conversation contains many utterances, but also because each dialogue must execute an explicit topic shift rather than remaining fixed on a single subject (Kautsar et al., 9 Aug 2025).
The paper states that this multi-scenario structure is meant to mirror real conversations, which “often shift fluidly between different topics rather than remaining fixed on a single subject.” Topic transition is therefore a dataset-level design variable rather than an incidental conversational property.
3. Cultural grounding, topics, and personas
Cultural grounding is the defining principle of SEADialogues. The paper states that the authors manually curate local knowledge and inject it into scenario and persona templates through lexicalization. This includes food, ceremonies and holidays, travel destinations, cities and countries, TV shows and media genres, mythological or folk characters, traditional instruments, movie types, regional customs, and communication norms such as politeness and honorifics.
Every dialogue contains two topics, implemented through two scenarios selected from a pool of 300 scenarios generated from 100 curated topics. GPT-4.1 mini is used to generate the 300 scenario templates, roughly three scenarios per topic. To make topic transitions natural, the authors cluster scenarios by semantic similarity using Top2Vec, then pair scenarios from the same cluster so that the [TRANSITION] shift is smoother. The prompt requires that both specified topics be addressed, and this is later checked by the Correctness annotation criterion.
The paper gives concrete lexicalization examples. For Indonesian, a template such as “Person A describes a family trip to the [TRAVEL_DESTINATION]” can be lexicalized with “iconic rice paddies of Ubud-ind” and realized as “Person A describes a family trip to the iconic rice paddies of Ubud.” Other topic examples include “Favorite TV Shows from Childhood,” “Favorite Musicians or Bands,” “Movie or Series Characters That Inspire You,” “The Most Interesting Local Folk Tales or Myths,” and “First Experience Watching a Movie in the Cinema.”
Persona design is equally explicit. Each dialogue includes two personas, one for each speaker, constructed from gender, name, persona description, and a personality trait. The paper reports 210 diverse personas; 300 were originally planned, but 90 were removed due to low quality. Personality traits include values such as Active, Appreciative, Considerate, Creative, Friendly, Honest, Imaginative, Open, Patient, Witty, Ambitious, Amusing, Boyish, Businesslike, and Determined.
The persona templates are also culturally lexicalized. Examples include a person fascinated by traditional or mythological characters, a classically trained musician interested in folk instruments, a person who appreciates movie characters inspired by folklore and traditional values, and a skeptic who enjoys listening to stories of a myth character. The paper’s ceremony-food mapping table further shows grounded pairings such as Loy Krathong-tha with pad thai, green curry, and mango sticky rice; Pasko-tag with adobo and lechon; Deepavali-tam with fish head curry and roti prata; and Tet-vie with pho, banh mi, and goi cuon. These examples indicate that cultural grounding in SEADialogues is operationalized through structured local knowledge rather than generic topic labels (Kautsar et al., 9 Aug 2025).
4. Construction pipeline and human annotation
The construction pipeline has four stages: template generation, lexicalization of cultural elements, synthetic dialogue generation using LLMs, and final human annotation. The authors first curate scenario templates, persona templates, Southeast Asian names, topic lists, and cultural entity pools. GPT-4.1 mini is used to generate 300 scenario templates from 100 curated topics, after which human annotators revise low-quality or inappropriate templates. Persona templates are constructed in parallel, with culture-specific references masked for later lexicalization.
Lexicalization fills placeholders using manually curated cultural entities. The paper emphasizes preserving contextual dependencies via JSON dictionaries that constrain valid slot combinations. The example given is country-city compatibility: Jakarta, Bandung, and Denpasar only pair with Indonesia, while Bangkok, Chiang Mai, and Songkhla only pair with Thailand. Ceremony-food combinations are similarly curated so that festivals are paired with plausible local foods. This step is intended to prevent nonsensical combinations and strengthen authenticity.
Dialogue generation then proceeds by prompting both open-source and proprietary LLMs with the target language, Topic 1 and Topic 2, the personas for speakers A and B, the politeness and honorific requirement, the turn limit, and the mandatory [TRANSITION] token. The prompt explicitly instructs the model to speak only in the target language and to generate nothing except the conversation itself.
Human annotation is performed with three annotators per language. The paper states that they were hired through the authors’ contacts, were indigenous people from the country, and were fluent in the native language. Annotation is carried out on a proprietary platform built by the authors. The rubric includes conversational quality criteria—Fluency, Engagingness, Coherence, and Naturalness on a 1–3 scale, and Cultural Relevance on a 0–3 scale—plus two instruction-following criteria: Profile Detection and Correctness, both binary. The guideline for Cultural Relevance is explicit: a score of 0 indicates no cultural aspect, 1 indicates culturally irrelevant or stereotyped content, 2 indicates somewhat relevant but inaccurate or shallow content, and 3 indicates culturally correct and authentic content (Kautsar et al., 9 Aug 2025).
5. Models, automatic evaluation, and empirical findings
The generation experiments compare four models: Llama-3.1-8B Instruct, Aya-8B Expanse, Gemini 1.5 Flash, and GPT-4o mini. The reported decoding settings are temperature and top-, and the hardware is NVIDIA A100 40GB. The experimental focus is not downstream task training, but data-generation quality assessment: which LLMs produce better culturally grounded, persona-consistent dialogues for the target languages.
Automatic evaluators are G-Eval with GPT-4.1 mini, M-Prometheus-7B, and R3-Qwen-14B-14k. They score the same dimensions used in human annotation: Coherence, Cultural Relevance, Engagingness, Fluency, Naturalness, Correctness, and Profile Detection. Agreement with human judgment is measured using Pearson correlation, Spearman correlation, and Kendall Tau; for the binary instruction-following metrics, the paper reports Precision, Recall, F1, and Accuracy.
The main empirical pattern is that closed-weight models outperform open-weight models overall. Gemini 1.5 Flash and GPT-4o mini perform best in both human and automatic evaluation. Among open models, Aya-8B Expanse is competitive on several Southeast Asian languages, especially Indonesian, Javanese, and Minangkabau, though the paper notes fluency issues. Llama-3.1-Instruct is generally the weakest and most variable. Representative human results include Indonesian coherence scores of 2.88 for Gemini, 2.84 for GPT-4o, 2.79 for Aya, and 2.70 for Llama; Javanese coherence scores of 2.92 for GPT-4o, 2.90 for Gemini, 2.84 for Aya, and 1.17 for Llama; and Thai cultural relevance scores of 1.59 for Gemini, 1.42 for GPT-4o, 1.37 for Llama, and 0.71 for Aya.
The paper also reports cross-language variation. Indonesian, Javanese, Minangkabau, Vietnamese, and Malay appear easier for some models—especially Aya—to handle well, whereas Thai, Tamil, and Tagalog are more challenging for some open models. Javanese is particularly difficult for Llama-3.1-Instruct, and Thai functions as a stress test for cultural relevance and profile following. On automatic evaluation, Gemini and GPT-4o are nearly saturated on coherence, fluency, and correctness in many languages, while Aya remains strong on Indonesian, Javanese, Minangkabau, Vietnamese, and Malay but weak on Thai cultural relevance.
A major secondary result concerns judge alignment. G-Eval with GPT-4.1 mini correlates best with human judgment overall, but automatic judges remain weaker on Fluency and Cultural Relevance. For Minangkabau, G-Eval Pearson correlation reaches 0.5876 on Coherence and 0.6124 on Naturalness, but only 0.1528 on Fluency. For Javanese, G-Eval Pearson correlation is 0.7899 on Coherence and 0.8338 on Engagingness. Instruction-following alignment is also uneven: on Minangkabau, G-Eval F1 is 0.986 for Correctness and 0.971 for Profile Detection, but on Indonesian the reported Profile Detection accuracy is 0.42 for G-Eval, 0.41 for M-Prometheus, and 0.52 for R3. The paper concludes that LLM judges are useful, but still lag human evaluation, especially on culturally sensitive dimensions (Kautsar et al., 9 Aug 2025).
6. Significance, use cases, limitations, and ethics
SEADialogues is significant because it fills a gap in dialogue research: the absence of datasets that are simultaneously multilingual, culturally grounded, persona-conditioned, and multi-turn for underrepresented Southeast Asian languages. Its strongest contribution is not only scale, but the operationalization of culture through localized entities, paired topics, persona design, politeness norms, honorific usage, and native-language human evaluation.
The paper identifies multiple use cases. These include training or fine-tuning culturally aware chatbots, studying persona-consistent dialogue generation, evaluating multilingual LLMs on Southeast Asian conversational competence, building human-centric assistants for local communities, researching topic transition and honorific usage, benchmarking LLM-as-judge methods on multilingual dialogue, and improving open-weight models for low-resource Southeast Asian languages.
The dataset’s limitations are also explicit. The dialogues are synthetic rather than collected from natural human-human interaction. Even with manual curation, generated dialogue may still miss subtle cultural nuance. The paper does not benchmark downstream tasks such as topic transition detection or persona detection. Some languages and settings may remain underrepresented in cultural depth, and automatic judges do not fully match human evaluation, especially on Fluency and Cultural Relevance.
The ethical notes are narrowly scoped but concrete. Names were manually curated from public, non-sensitive sources, and any resemblance to real individuals is stated to be coincidental. Topics were manually filtered to avoid violence and other sensitive content. Annotators were compensated fairly according to local wage standards. Only aggregated annotation scores are released, and no personal identifying information is included. Taken together, these design choices suggest a dataset intended as a foundation for culturally aware and human-centered dialogue systems in Southeast Asia, while also retaining the standard caveats associated with synthetic data (Kautsar et al., 9 Aug 2025).
