How to Collect Conversational Data Across Cultures

A comparison table in the SEADialogues paper makes the state of the field clear. It lists the major dialogue datasets — DailyDialog, MultiWOZ, PERSONA-CHAT, XPersona, GlobalWOZ, Multi2WOZ, Multi3WOZ, XDailyDialog — and in the column for cultural relevance, every one is marked with a cross. Some are large: XPersona covers six languages with 104,600 dialogues, and GlobalWOZ covers 21 languages. 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 rather than natively created.

A dialogue dataset is a structured collection of conversational exchanges used to train or evaluate conversational AI, typically organised around scenarios, personas and turns. 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 interchangeable.

Take a task-oriented dialogue about booking a restaurant table. Translate it into Javanese and the result is a Javanese-language 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. Hershcovich and colleagues argue that collecting multilingual data within large local communities results in culturally richer data and avoids imposing English-driven use cases on every market. The COD project's approach involves cultural adaptations and replacements of foreign concepts with those common in the annotators' own culture and environment, and its authors note that scenario selection based on cultural relevance and plausibility should come before dialogue generation, not after. Scenario selection, not translation quality, is where cross-cultural dialogue data succeeds or fails — the same lesson explored in why culturally relevant data matters for AI.

Cultural grounding, in this context, means that the scenarios, personas and norms underlying a dialogue were sourced from and verified by people who live the culture being represented. 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 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 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, and 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: the team treated "Indonesian culture" as insufficient granularity and built provincially, the same logic that applies to scoping language coverage at locale level.

Social norm annotation is the practice of tagging each conversational turn for whether it adheres to or violates a culturally specific behavioural expectation. NormDial produced 4,231 dyadic dialogues totalling 29,550 conversational turns across Chinese and American cultures, with social norm adherences and violations labelled turn by turn. 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 any norm that failed a criterion, and arriving at 133 Chinese and 134 American norms as the grounded foundation. RENOVI extends this to repair, containing 9,258 multi-turn dialogue instances and described as the first dataset exploring remediation of norm violations based on Chinese cultural norms, annotated by 20 university lecturers and students familiar with Chinese culture. NormGenesis shows the annotation depth this work can reach: 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. That justification is unusual and valuable — a label without a rationale cannot be audited, and in cultural annotation the rationale is often the only way to tell a genuine cultural judgement from a personal one. The finding that motivates all of this work: existing models often fail to reason correctly about norm adherence and violation in conversational contexts.

In stage one, native speakers translate and localise the slot values — restaurant names, dish names, currencies, addresses, times, honorifics. In stage two, a different group of human subjects translates or localises the entire phrase, using the slot output from stage one. Separating the two steps prevents the common failure where a translator preserves English slot values because they read as proper nouns, leaving a Thai-language dialogue about ordering a Caesar salad from a place called The Golden Lion. The alternative — substituting English slot values with target-language counterparts under a controlled automatic procedure — is faster but produces a thinner result than the slower, more human-driven approach that keeps close contact with the local community speaking the target language.

The 2026 Hume-DaiKon corpus 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. 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 — the example given is "How was your weekend?" — which produces spontaneous speech rather than a scripted read. The response is then automatically checked using a language model to verify it is both relevant to the prompt and linguistically fluent, an automated-triage-before-human-review pattern that recurs throughout how speech data is collected for low-resource languages. Splits are stratified by language, with the test set kept blind — per-language stratification is the recurring discipline across multilingual data work.

Synthetic dialogue data is conversational text or speech generated by a language model rather than recorded from real participants. The NormDial authors argue plainly for it: 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 qualifier that matters 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 — synthesis operating within a human-authored cultural frame. Synthetic dialogue generated by prompting a model to "write a conversation between two Indonesian friends," with no such frame underneath it, is the model's stereotype of Indonesian conversation, and it reads as such to any Indonesian.

If the scenarios are culturally accurate, dialogue volume scales cleanly from them, as SEADialogues demonstrates by generating 32,000 dialogues from 300 scenarios and 210 personas. If the scenarios were imported and translated, no amount of dialogue volume fixes it — which is why the question to ask a supplier is how the scenarios were sourced, not how many dialogues they will deliver. Cultural granularity also needs deciding as explicitly as dialect granularity: CultureTalk-ID built provincially within Indonesia because national-level culture was insufficient resolution, a decision that belongs in a scope document rather than being resolved by default, much like the community-partnership questions covered in partnering with universities and local communities for language data. A dataset labelled "Indonesian" that was collected entirely in Jakarta is a Jakarta dataset, and nothing in the delivery statistics will say so — the same reasoning behind what a multilingual data collection service should include and Lifewood's multilingual data collection services more broadly, as well as recruitment practices such as recruiting native contributors for language data.

• Specify the scenarios, not just the languages — where 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 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, and be precise about synthetic content — human-verified norms grounding model-generated dialogue is a method; unverified generation is a stereotype, a distinction that applies equally to collecting speech and text that mixes languages and to collecting accented or non-native speech.