How to Collect Conversational Data Across Cultures

Short answer. Across the major dialogue datasets — DailyDialog, MultiWOZ, PERSONA-CHAT, XPersona, GlobalWOZ, XDailyDialog — cultural relevance is largely absent, and SEADialogues describes itself as the first to represent cultural aspects explicitly within each conversation. Translation cannot fix this: it preserves the language and imports the scenario, so a translated dialogue is a foreign situation conducted in the target language. Hershcovich and colleagues argue that collecting inside large local communities produces culturally richer data and avoids imposing English-derived scenarios on every market.

Across Cultures?

There is a comparison table in the SEADialogues paper that makes the state of this field embarrassingly clear.

It lists the major dialogue datasets: DailyDialog, MultiWOZ, PERSONA-CHAT, XPersona, GlobalWOZ, Multi2WOZ, Multi3WOZ, XDailyDialog. Some are large. XPersona covers six languages with 104,600 dialogues. GlobalWOZ covers 21 languages.

And in the final column, cultural relevance, every one of them is marked with a cross.

SEADialogues describes itself as the first dialogue dataset to explicitly represent cultural aspects within each conversation. Several of the multilingual entries carry a second cross too, in the column marking whether the dataset avoided translation. GlobalWOZ, XPersona and XDailyDialog are all translated.

So the honest starting position for anyone scoping this work: multilingual dialogue data mostly exists, culturally grounded dialogue data mostly does not, and the two have been treated as the same thing.

Why translation does not produce cross-cultural dialogue

The reason is that the scenarios themselves carry culture, and translation preserves the words while leaving the situation intact.

Take a task-oriented dialogue about booking a restaurant table. Translate it into Javanese and you have a Javaneselanguage conversation about a Western restaurant booking flow, with Western assumptions about reservation norms, party sizes, payment and how one addresses staff. The language is right. The situation is imported.

The research community has named this directly. Hershcovich and colleagues argue that collecting multilingual data within large local communities results in culturally richer data and avoids imposing English-driven use cases.

The COD project's approach involves cultural adaptations and replacements of foreign concepts with those common in the annotators' culture and environment, and its authors note that the next step should be careful selection of dialogue scenarios based on their relevance and plausibility in the culture in question.

Scenario selection, not translation quality, is where cross-cultural dialogue data succeeds or fails.

What a culturally grounded pipeline looks like

Two recent datasets show the construction method in enough detail to copy.

SEADialogues covers eight Southeast Asian languages, Indonesian, Javanese, Malay, Minangkabau, Tagalog, Tamil, Thai and Vietnamese, across six countries, producing 32,000 dialogues.

Its pipeline begins with supporting resources rather than with conversations: scenario templates, persona templates, and culturally relevant Southeast Asian names. For each dialogue, two domain-relevant scenarios and corresponding personas are selected to ensure consistency and coherence across both intra-scenario and inter-scenario persona relationships, followed by manual lexicalisation.

The dataset ships 300 scenarios and 210 personas. Those two numbers are the actual cultural content. Everything downstream is generation and annotation against them.

CultureTalk-ID goes deeper geographically within a single country. Built through a multi-stage human pipeline involving native speakers, it covers general Indonesian culture plus the cultures of ten provinces, Aceh, West Sumatra, West Java, Central Java, East Java, Bali, Nusa Tenggara Timur, South Kalimantan, South Sulawesi and West Papua, spanning Indonesian and ten local languages across thirteen cultural topics.

The design point worth extracting: they treated "Indonesian culture" as insufficient granularity and built provincially. The same logic that applies to dialect scoping applies to cultural scoping, and for the same reason.

Social norms, and why they need explicit labels

The most technically interesting strand of this work concerns norms, because norms are what make a conversation feel right or wrong to a participant, and they are almost never labelled.

NormDial produced 4,231 dyadic dialogues totalling 29,550 conversational turns across Chinese and American cultures, with social norm adherences and violations labelled on a dialogue-turn basis.

Their verification process is the part worth copying. Native speakers in each culture manually evaluated whether each generated norm was factually correct according to their own lived experiences, in line with the defined norm category, specific to the culture, and detailed in its description, removing those that failed any criterion. That produced 133 Chinese and 134 American norms as the grounded foundation.

Note the phrase "according to their own lived experiences." Not "correct according to a reference work." That is the right standard for cultural content and it is only available from people who live it.

RENOVI extends this to repair, containing 9,258 multi-turn dialogue instances and described as the first dataset exploring the remediation of norm violations based on Chinese cultural norms. For annotation they invited 20 university lecturers and students familiar with Chinese culture into a structured training procedure.

NormGenesis shows the annotation depth that supports this kind of modelling. Dialogues run 5 to 15 turns, and each utterance is annotated with norm adherence, speaker reaction including intent and emotional state, and a justification for the assigned label, with reaction labels grounded in dialogue act theory.

That third element, the justification, is unusual and valuable. A label without a rationale cannot be audited, and in cultural annotation the rationale is frequently the only way to tell a genuine cultural judgement from a personal one.

The finding that motivates all of it: existing models often fail to reason correctly about norm adherence and violation in conversational contexts.

The localisation method that works

For task-oriented dialogue specifically, there is a two-stage approach documented in the multilingual dialogue literature that is worth knowing because it is more robust than single-pass translation.

Stage one: native speakers translate and localise the slot values. Restaurant names, dish names, currencies, addresses, times, honorifics.

Stage two: a different group of human subjects translates or localises the entire phrase, using the slot output from stage one.

Separating slot localisation from phrase localisation prevents the common failure where a translator preserves the English slot values because they appear to be proper nouns, leaving a Thai-language dialogue about ordering a Caesar salad from a place called The Golden Lion.

The same source contrasts two philosophies: substituting English slot values with target-language counterparts under a controlled automatic procedure, versus a more human-driven approach that provides closer contact with the local community speaking the target language. The second is slower and produces the culturally richer result.

Naturalistic dyadic collection

For spontaneous conversational speech rather than constructed dialogue, the 2026 Hume-DaiKon corpus shows a collection design worth knowing.

It contains 945 sessions totalling 743.4 hours across German, English, Spanish, Dutch and Polish, collected through a dual-channel conversational platform that connects pairs of participants from similar geographic regions.

Three design details are worth copying.

Participants complete an audio quality screening before participation, ensuring a minimum standard of microphone clarity rather than discovering the problem in post-processing.

They respond to a short prompt in their native language, with the example given being "How was your weekend?" A prompt rather than a script, which produces spontaneous speech.

The response is automatically checked using a language model to verify it is both relevant to the prompt and linguistically fluent, which is automated triage before human review, the same architecture that works throughout data operations.

And the splits are stratified by language, with the test set kept blind. Per-language stratification again, which is the recurring discipline in every part of multilingual data work.

Where synthetic generation fits, honestly

Several of the datasets above are wholly or partly synthetic, generated by language models under expert prompting with human verification. That deserves a clear-eyed assessment rather than either dismissal or enthusiasm.

The argument for it is stated plainly by the NormDial authors: gathering realistic data at scale in this domain is challenging and potentially cost-prohibitive, particularly for identifying norm adherences and violations across multiple cultural contexts. They report that their synthetic bilingual conversations were comparable to or exceeded the quality of existing naturally occurring datasets under interactive human evaluation and automatic metrics.

The critical qualifier is where the humans sit. NormDial's pipeline has human verification at every stage, and the norms themselves were validated by native speakers against lived experience before any dialogue was generated. The synthesis operates within a human-authored cultural frame.

That is the distinction that matters when evaluating a supplier or a dataset. Synthetic dialogue grounded in native-verified cultural norms is a legitimate method. Synthetic dialogue generated by prompting a model to "write a conversation between two Indonesian friends" is the model's stereotype of Indonesian conversation, and it will read as such to any Indonesian.

Where our own work fits

Declaring the interest: Lifewood collects conversational and dialogue data across 50-plus languages and dialects through delivery centres in more than 30 countries, including several of the Southeast Asian markets these datasets cover.

Two observations.

The first is that the scenario library is the deliverable, not the dialogues. SEADialogues built 300 scenarios and 210 personas and generated 32,000 dialogues from them. If the scenarios are culturally accurate, the dialogues can scale. If the scenarios were imported and translated, no amount of dialogue volume fixes it. When scoping this work, the question to ask a supplier is how the scenarios were sourced, not how many dialogues they will deliver.

The second is that cultural granularity needs deciding as explicitly as dialect granularity. CultureTalk-ID built provincially within Indonesia because "Indonesian culture" was insufficient resolution. That decision has a cost and a rationale, and it should appear in a scope document rather than being resolved by default. A dataset labelled "Indonesian" that was collected entirely in Jakarta is a Jakarta dataset, and nothing in the delivery statistics will say so.

A scoping checklist

Specify the scenarios, not just the languages. Where do they come from, and who validated that they are plausible in that culture?

Decide cultural granularity explicitly, at national, provincial or community level, with a stated rationale.

Recruit for lived experience, since the verification standard that works is whether a norm is correct according to the annotator's own life, not according to a reference.

Localise slot values separately from phrases, in two stages with different people.

Annotate norms explicitly where the application involves social appropriateness, including adherence, reaction and a justification for the label.

Use prompts rather than scripts for spontaneous collection, with audio screening before the session and automated relevance checking after it.

Stratify everything by language, including evaluation splits.

Be precise about synthetic content. Human-verified norms grounding model-generated dialogue is a method. Unverified generation is a stereotype.

Key takeaways

Across the major dialogue datasets including DailyDialog, MultiWOZ, PERSONA-CHAT, XPersona, GlobalWOZ and XDailyDialog, cultural relevance is absent, and several of the multilingual ones are translated rather than natively created.
SEADialogues describes itself as the first dialogue dataset to explicitly represent cultural aspects within each conversation.
Translation preserves the language and imports the scenario, so a translated dialogue is a foreign situation conducted in the target language.
Hershcovich and colleagues argue that collecting within large local communities produces culturally richer data and avoids imposing English-driven use cases.
SEADialogues covers eight Southeast Asian languages across six countries with 32,000 dialogues, built from 300 scenarios and 210 personas plus culturally relevant names.
CultureTalk-ID covers general Indonesian culture plus ten provinces, Indonesian plus ten local languages, across thirteen cultural topics, treating national-level culture as insufficient granularity.
NormDial produced 4,231 dyadic dialogues and 29,550 turns across Chinese and American cultures with norm adherence and violation labelled per turn.
NormDial's norm validation standard was whether native speakers judged each norm factually correct according to their own lived experiences, culture-specific, in category and sufficiently detailed, yielding 133 Chinese and 134 American norms.
RENOVI contains 9,258 multi-turn instances and is described as the first dataset addressing remediation of norm violations, annotated by 20 university lecturers and students familiar with Chinese culture.
NormGenesis annotates each utterance in 5 to 15 turn dialogues with norm adherence, speaker reaction including intent and emotional state, and a justification for the label.
Existing models often fail to reason correctly about norm adherence and violation in conversational settings.
Task-oriented localisation works best in two stages: native speakers localise slot values first, then a different group localises the full phrase using that output.
Hume-DaiKon collected 945 naturalistic dyadic sessions totalling 743.4 hours across five languages, with presession audio screening, native-language prompts rather than scripts, and automated relevance and fluency checking.
Synthetic dialogue is legitimate when grounded in native-verified cultural norms with human verification at every stage, and produces stereotypes when generated without that frame.

Sources and further reading

Frequently asked questions

You will get the right language and the wrong situation. The scenarios carry cultural assumptions about roles, norms and procedures, and translation leaves those intact.

The scenarios and personas. SEADialogues generated 32,000 dialogues from 300 scenarios and 210 personas built with culturally relevant names and situations. The scenario library is where the cultural content lives.

More granular than national, usually. CultureTalk-ID built across ten Indonesian provinces because "Indonesian culture" was insufficient resolution, which mirrors the same decision required in dialect scoping.

Against native speakers' lived experience. NormDial required each norm to be factually correct per the annotator's own life, in category, culture-specific and detailed, discarding those that failed.

When it is grounded in human-verified cultural norms with verification at every stage, yes. NormDial reported synthetic conversations comparable to or exceeding naturally occurring datasets under human evaluation. Unverified generation produces the model's stereotype of a culture.

With prompts rather than scripts. Hume-DaiKon used nativelanguage prompts such as "How was your weekend?", with audio quality screening before participation and automated checking of relevance and fluency afterwards.

Have an AI or visibility project in mind?

From AI evaluation and human-in-the-loop review to GEO and AEO strategy, our team can help you deploy with confidence and get found in the AI search era.

Talk to our team