Transferring When to Switch Dialects, Not Which Variety Is Prestigious

Many speakers move between a standard variety and a local dialect depending on who they are talking to and why. Examples include Modern Standard Arabic and Egyptian Arabic, Literary and Spoken Tamil, Standard and Swiss German, and Mainstream U.S. English and African American English. Language models are increasingly asked to write in these varieties, and nearly all the evidence about when users want a dialect comes from a single language. The obvious move is to teach a multilingual model the English pattern ("standard for the boss, vernacular for friends") and rely on zero-shot cross-lingual transfer to carry it everywhere. Zero-shot cross-lingual transfer means training on one language and running in another without target-language labels. That move is likely to be wrong in a specific, predictable way. The prestige ordering of varieties is not universal. Swiss listeners chose a Zurich-dialect speaker over a Standard German speaker even for hiring a secretary.

We propose to split dialect choice into two parts. The first is a model of the situation: how hierarchical, how public, how evaluative, how intimate it is. We expect this part to transfer across languages, because it is about social structure rather than words. The second is a small community profile that maps those situation features to a preferred variety. We expect this part not to transfer and to need a handful of local judgments. The core claim is testable. A situation encoder trained only on English preferences, combined with a profile fit from about ten community judgments, should match or beat both direct English transfer and few-shot prompting on the cases where the community disagrees with English.

Core Claim and Motivation

Two findings in the literature pull against each other. Muennighoff et al. show that instruction-following learned from English prompts transfers to other languages (Crosslingual Generalization through Multitask Finetuning), with a 50% zero-shot gain on non-English XNLI for English-only instruction tuning. Earlier work shows the same kind of transfer for reading comprehension (Zero-shot Reading Comprehension by Cross-lingual Transfer Learning with Multi-lingual Language Representation Model) and parsing (75 Languages, 1 Model). In all of these cases, though, the task logic is the same in every language: an answer span is an answer span everywhere.

Dialect appropriateness is different. The English user study My LLM might Mimic AAE - But When Should It? found a clear gradient, with the standard preferred in formal and workplace settings and the vernacular in casual and personal ones. The Swiss matched-guise study The effect of speakers' regional varieties on listeners' decision-making played one actress speaking each variety, so that only the variety differed between recordings. Standard German was chosen by only 15.3% of listeners for the job scenario and 9.7% for the travel scenario. It rose to 36.1% only for choosing a surgeon. So the direction of the mapping flips between communities. The axis still matters, though: the more consequential and evaluative the scenario, the more the standard gains. Social-media studies support the idea that situation axes such as audience breadth are shared across communities. Scots speakers moved toward Standard English when addressing the international independence debate (Aye or naw, whit dae ye hink?), and standard Buenos Aires Spanish variants clustered in tourist and commercial zones (Social Context and User Profiles of Linguistic Variation on a Micro Scale).

We formalize this split. Let $x$ be a described situation, $f(x)$ a vector of situation features produced by a multilingual encoder trained only on English preference data, and $c$ a speech community. The probability that community $c$ prefers the dialect over the standard is

$$p(\text{dialect} \mid x, c) = \sigma\big(w_c^\top f(x) + b_c\big),$$

where $\sigma$ is the logistic function and $(w_c, b_c)$ is the community profile. The profile has a few dozen parameters and is fit from $k$ judgments by community members. Plain English transfer corresponds to using English's $(w, b)$ for every community. Our hypothesis is that $f$ transfers and $(w_c, b_c)$ does not. We also predict the Swiss case is mostly a shift in $b_c$ (the dialect is the default) together with a reweighting of the "high-stakes" direction in $w_c$.

Relation to Existing Work

The closest dialect-and-transfer work studies comprehension or fidelity, not appropriateness. DIALECT-COPA (DIALECT-COPA) and the JSI/WüNLP system paper (JSI and WüNLP at the DIALECT-COPA Shared Task) measure reasoning over South Slavic dialect text. The system paper shows that ten dialect examples raise GPT-4 from 76.0% to 85.3%, mainly by teaching dialect vocabulary rather than task logic. AL-QASIDA (AL-QASIDA) shows that models understand dialectal Arabic but are "reluctant" to produce it, drifting back to the standard; most models score below 50% on its ADI2 fidelity metric. IruMozhi (IruMozhi) and MADAR (The MADAR Shared Task) classify which variety a text is in. None of these asks which variety a situation calls for. We reuse their classifiers and fidelity metrics as measurement tools for the generation side.

The context-modeling work supplies our situation schema but stays within one language. Liu et al. (Constructing a Japanese Business Email Corpus Based on Social Situations) encode "Tenor," meaning hierarchy and social distance between writer and reader, over 1,040 scenarios. Japanese BERT recovers superior/subordinate relations above 90% accuracy. Kawano et al. (Analysis of Style-Shifting on Social Media) condition a language model on social-network position. They also show that surprisal-based style-shift measures track human judgment while word-overlap measures do not. We adopt Tenor-style axes for $f$ and treat the second finding as a constraint: we do not evaluate style by counting dialect words. The AAE study also constrains the design. Users there rejected automatic dialect detection and wanted an explicit toggle. Our model therefore produces a default or suggestion that the user can override; it does not switch dialect on its own initiative.

Research Program

The first step is a shared scenario bank of about 400 situations. These are built from Tenor dimensions (hierarchy, intimacy, inside versus outside the organization), audience breadth (private, local public, international), stakes (casual, evaluative, medical or legal), and channel (spoken message, email, public post). The bank is seeded from Liu et al.'s scenario types, the AAE study's scenario categories, and Leemann's three decision scenarios. Each situation is written once in English and adapted by native speakers into the four communities' contexts.

For each scenario, members of four communities give a forced choice between the standard and the dialect plus a strength rating. The communities are U.S. English (MUSE/AAE), German-speaking Switzerland (Standard German/Zurich or Bern Swiss German), Egypt (MSA/Egyptian Arabic), and Tamil Nadu (Literary/Spoken Tamil). We aim for at least five raters per item. Because the AAE study reported no per-scenario numbers, we collect English labels ourselves on the same bank so that source and target are comparable. We recruit outside university pools where possible, since that study's participants skewed educated and online.

We then train $f$ on English labels only, as a head on XLM-R and as a LoRA adapter on mT0-13B. We fit $(w_c, b_c)$ from $k \in \{0, 5, 10, 20, 50\}$ community judgments and evaluate on the remaining items. A secondary generation study takes the chosen variety and asks the model to write the message. It measures whether the output actually stays in that variety, using ADI2 for Arabic and an IruMozhi-trained classifier for Tamil. It also checks whether dialect-feature density stays within the range of human-written references, since the AAE study found model output denser in dialect features than human speech. This links the "when" question to AL-QASIDA's reluctance problem: a correct choice is useless if the model drifts back to the standard while writing.

Baselines

The naive reference point is a constant policy: always standard, and always the community's majority variety. The Swiss data suggest always-dialect is already strong there, so any method must beat the per-community majority rate, not 50%.

The direct-transfer baseline is the same English-trained encoder and head, applied unchanged to each community. This is what the obvious intersection idea would build. It is the most important comparison, because the claim is that it fails exactly where community preference reverses English preference.

The strongest existing method is few-shot in-context prompting of a large model (GPT-4o and Llama-3.1). The prompt contains the same $k$ community judgments, given as scenario–choice pairs, and the model is asked to choose for the new scenario. This follows the few-shot approach shown effective in JSI and WüNLP and AL-QASIDA. No reported result was found on Lacuna for situation-conditioned dialect choice. The nearest reported figures are on adjacent tasks: GPT-4 rising from 76.0% to 85.3% accuracy on DIALECT-COPA comprehension with ten examples, and most models below 50% ADI2 fidelity in AL-QASIDA. Neither is a target number for our task. A third baseline is zero-shot prompting with a written sociolinguistic description of the community, with no labeled examples.

For fairness, all methods see exactly the same $k$ judgments, drawn from the same five random seeds. Prompt templates and the factorized model's regularization each get the same tuning budget of 20 configurations, selected on a held-out English split. The upper bound is a model trained on the full community labels.

Ablations isolate the key ingredient. One adapts only $b_c$, which tests whether a single threshold shift suffices. One replaces $f$ with an untuned multilingual encoder, which tests whether English preference training contributes anything beyond generic semantics. One replaces $f$ with hand-coded situation-axis labels, which tests whether learned features beat the schema itself.

Beating the strongest baseline means two things. First, higher agreement with the community majority and lower divergence from the empirical preference distribution at $k \le 10$, averaged over seeds with non-overlapping bootstrap intervals. Second, specifically on "reversal items," where a community's majority differs from English's.

Why not just…?

Why not just give users a toggle and skip prediction? The AAE study supports this, and we agree the user's explicit choice must always win. However, a toggle still needs a default. It also does not cover drafting for someone else, such as an email to a Zurich colleague or a public post. And every interface still has a policy, implicit in post-training, that AL-QASIDA shows leans toward the standard. If raters in all four communities say they would set the toggle themselves and ignore suggestions, that weakens the deployment case. The measurement result would still stand: how badly do current models impose English prestige ordering on other communities?

Why not just few-shot prompt a large model with local examples? This is the real threat, given how well ten examples work for dialect comprehension. If prompting matches the factorized model at equal $k$, we will report that and conclude that a frozen large model already separates situation from community preference. The benchmark and the reversal-item analysis would remain the contribution. Our reason to expect a gap is that in-context examples have so far taught vocabulary, and smaller models such as mT0 and Aya degraded with longer contexts (JSI and WüNLP). The matched-$k$ curve settles it directly.

Why not just describe the community's norms in the prompt? This is cheap and may capture the default, for instance "Swiss German is used in almost all spoken settings." It is unlikely to capture graded exceptions like the surgeon scenario unless the description happens to mention them. The zero-shot description baseline tests this. If it closes most of the gap, that means the community profile is largely knowable from text, which is itself informative.

Why not just detect the user's dialect and mirror it? The users in the AAE study rejected this. Shoemark et al. and Leemann et al. show that dialect speakers themselves move to the standard in public or high-stakes settings, so mirroring would be wrong in exactly the formal cases our scenario bank isolates.

Decisive Checks

The central test is the reversal-item analysis. If direct English transfer scores below the per-community majority baseline on reversal items, the transfer-of-polarity failure is real. This is what we expect for Swiss German in hiring and travel scenarios, given the 15.3% and 9.7% standard preference. If the factorized model with $k=10$ then recovers those items while keeping English-level accuracy on agreement items, the split into a shared situation model and a local profile is supported. If the only-$b_c$ ablation does as well as the full profile, the community difference is a single threshold, and the simpler story should be preferred. If the untuned encoder does as well as the English-trained $f$, then English preference data contributes nothing transferable, and the cross-lingual claim fails.

On the generation side, the check is whether a correctly chosen dialect survives writing. This means ADI2 and IruMozhi-classifier rates for the full message, not just the first sentence, and feature density within the human reference range.