rshahbaz/pragmatic-code-switch-blindspot

Pragmatic Blind Spots Under Roman Urdu Framing

Question 1: the blind spot

The central blind spot in multilingual evaluation is that models frequently lose pragmatic and social meaning under Roman Urdu framing, even when they understand the individual words. I write in Roman Urdu myself, and I regularly see language models misunderstand the intended meaning in everyday exchanges. While standard multilingual benchmarks evaluate formal Perso-Arabic Urdu script or sanitized translations, millions of speakers use Latin-script Roman Urdu for daily communication. In this informal setting, models struggle with subtle social cues, such as softening a disagreement with an elder or recognizing when a polite offer is not meant to be taken literally. Natural language processing research has largely overlooked these dynamics by treating code-switching and non-standard spelling as machine translation problems rather than challenges in social reasoning.

An early diagnostic evaluation I ran on MiniCPM5-2B across 12 hand-constructed Pakistani scenarios (48 prompts) highlights this gap. In that test, framing situations in Roman Urdu rather than English dropped the model's score by an average of roughly two points on an eight-point rubric. Error analysis pointed to different kinds of breakdowns. In scenarios like respect_register and indirect_no, which fell from 8/8 in English to 1/8 under Roman Urdu framing, the difficulty was pragmatic: the model processed the vocabulary but failed to grasp the social context. In other cases, such as the khala failure where the model confused a maternal aunt with a Middle Eastern dish, a basic vocabulary gap caused the breakdown. When follow-up matched reruns showed that an earlier moderator effect I had observed between isolated code-switches and full narrative framing was actually an artifact of uneven English baselines, I publicly withdrew the claim. On the four cleanly rewritten scenarios, the language gap averaged 2.75 points across two coders. Given the small sample and wide uncertainty interval, this result is an early diagnostic signal rather than a definitive baseline, but it illustrates how models can produce fluent text while still misinterpreting everyday social dynamics in informal registers.

Design

The dataset contains 48 prompts built from 12 scenarios in a 2x2 design crossing language of the situation with whether social context is stated explicitly.

	Context implicit	Context explicit
English	implicit_english	explicit_english
Roman Urdu	roman_urdu_mixed	explicit_context_added

Within a scenario, the Task: line is identical across all four conditions and the Context: line is identical across both explicit conditions. Each prompt therefore differs from its neighbours on exactly one dimension.

Model: openbmb/MiniCPM5-2B, revision 12a3808a956f869c767195e9266b59c4d21d92e2, run with greedy decoding, do_sample=False, max_new_tokens=180, one response per prompt. Environment: transformers 5.17.0, torch 2.11.0+cu128, Tesla T4.

Responses were scored 0 to 2 on intent inference, social appropriateness, unsupported assumptions, and semantic preservation, for a maximum total of 8. Scoring was done by one coder against a rubric fixed before the run. A second coder then scored all 48 responses independently, blind to condition labels and to the first coder's scores, with items in random order (see below).

Results

Mean total score out of 8:

Condition	Mean
explicit_english	5.67
implicit_english	4.50
explicit_context_added	3.83
roman_urdu_mixed	2.33

Paired differences across the 12 scenarios, with bootstrap 95% confidence intervals:

Contrast	Mean	95% CI
Language, context implicit	-2.17	[-3.83, -0.58]
Language, context explicit	-1.83	[-3.42, -0.17]
Context, English	+1.17	[+0.33, +2.17]
Context, Roman Urdu	+1.50	[+0.50, +2.50]
Interaction, recovery in Roman Urdu relative to English	+0.33	[-1.00, +1.58]

Roman Urdu framing cost the model roughly two points out of eight in both context conditions under the first coder, and both intervals excluded zero. Explicit context helped in both languages by a similar amount. The interaction interval spans zero, so this run does not show that context recovers the Roman Urdu condition any more than it helps English.

Apparent moderator in the first run (withdrawn)

Seven of the twelve scenarios already contained a Roman Urdu quotation in every condition, so the manipulation changed only the surrounding English narration. In the other five, the entire situation was rewritten in Roman Urdu. Splitting the language effect by that distinction:

Scenario group	n	Language effect
Whole situation in Roman Urdu	5	-4.80
Roman Urdu quotation inside English narration	7	-0.29

Per scenario, the fully framed group ran -2, -4, -4, -7, -7. The narration-only group ran 0, -4, 0, +1, +2, -1, 0.

A Roman Urdu phrase dropped into an English sentence appeared to cost almost nothing, while rewriting the frame cost nearly five points out of eight. The two groups received different amounts of manipulation, though, because the narration-only group's English condition already contained Roman Urdu. A follow-up run tested that directly and the split did not hold up. See the matched rerun below.

Matched rerun with a fully English baseline

The split above has an obvious confound. In the seven narration-only scenarios the English condition was not fully English: it carried a Roman Urdu word or quotation in every condition, so switching to Roman Urdu changed less than it did in the other five. The flat -0.29 could mean that code-switching is cheap, or only that the manipulation was small.

Those seven scenarios were rewritten so the English condition contains no Roman Urdu at all, for example "their maternal aunt is coming over" in place of "their khala is coming over". The Roman Urdu condition, the Context line and the Task line are unchanged. Everything else matches the first run except the output budget, raised from 180 to 1,024 tokens after a 180-token pass left 20 of 28 responses with no final answer at all, and the shortfall fell mostly on Roman Urdu prompts. Files: eval_rerun.jsonl, results_rerun.jsonl, rerun_scores.jsonl, rerun_analysis.py, rerun_results.txt. The unusable 180-token pass is kept as results_rerun_180.jsonl.

Both coders scored this run blind, with condition labels removed and items shuffled. Four responses still produced no final answer within 1,024 tokens and both coders were instructed to score those 0, so they are excluded from the agreement figures below. Pooled quadratic-weighted kappa on the remaining 24 items is 0.72, with 0.88 on intent inference, 0.89 on semantic preservation, 0.70 on unsupported assumptions and 0.30 on social appropriateness.

Three scenarios contain one of those truncated responses, which leaves their baseline meaningless. On the four scenarios where every condition produced an answer, mean of both coders:

Scenario	English implicit	Roman Urdu implicit
kinship_context	8.0	0.5
favor_indirect	8.0	7.0
family_obligation	8.0	7.5
ambiguity	8.0	6.0

The language effect on those four is -2.75 [-5.88, -0.75], against -0.29 for the same scenarios in the first run. Under coder 2 alone it is -2.25 [-6.00, +0.00], so the interval touches zero.

The moderator claim does not survive this. What looked like a property of code-switching now looks like an artefact of how much each scenario's English condition had been changed. With a genuine language contrast, these scenarios lose points too. The remaining honest statement is narrower: Roman Urdu framing costs this model several points wherever the contrast is real, and the first run's split measured the size of the manipulation rather than a feature of mixed-language input.

This is four scenarios on one model, so it settles nothing on its own. It is enough to retract the moderator reading rather than to replace it with a new one.

Representative failure

On kinship_context, every condition scored 0 or 1 out of 8. The explicit English prompt named the relationship as socially relevant and the model still produced:

The khala is a type of food, often associated with Middle Eastern or North African cuisine. The person says their khala is coming over, so they are inviting family. They need to leave early, so they are leaving to make room for the khala.

Khala is the Urdu word for maternal aunt. Stating that the family relationship matters could not help, because the model did not recognise that the term referred to a person at all. This is a lexical gap surfacing as a pragmatic failure, and it explains why the explicit-context condition produced no recovery on that scenario.

Other scenarios failed differently. respect_register and indirect_no both scored 8/8 in English and 1/8 under Roman Urdu framing, with no vocabulary obstacle involved. Those look like a loss of pragmatic inference rather than a missing word.

Second coder

A second coder, fluent in Urdu and Roman Urdu, scored all 48 responses against the same rubric. Items were shuffled and relabelled, condition names were removed, and the second coder saw only the prompt and the model's final answer. Scores are in results_coder2.jsonl and every figure below is reproduced by agreement.py, with output in agreement_results.txt.

Agreement, quadratic-weighted Cohen's kappa, n=48:

Dimension	Kappa	Exact agreement
Intent inference	0.41	52%
Social appropriateness	0.14	46%
Unsupported assumptions	0.69	62%
Semantic preservation	0.63	60%
Pooled	0.54

Totals correlate at 0.77. The low kappa on social appropriateness mostly reflects the second coder giving 1 on 46 of 48 items, so there was almost no variance to agree on. Intent inference is the dimension where the coders diverged most in substance.

Main contrasts under each coder, bootstrap 95% intervals:

Contrast	Coder 1	Coder 2	Mean of both
Language, context implicit	-2.17 [-3.83, -0.58]	-1.58 [-2.92, -0.25]	-1.88 [-3.29, -0.50]
Language, context explicit	-1.83 [-3.42, -0.17]	-1.17 [-2.75, +0.33]	-1.50 [-2.88, -0.12]
Context, English	+1.17 [+0.33, +2.17]	+0.58 [-0.33, +1.50]	+0.88 [+0.21, +1.54]
Context, Roman Urdu	+1.50 [+0.50, +2.50]	+1.00 [+0.17, +2.08]	+1.25 [+0.54, +2.08]
Interaction	+0.33 [-1.00, +1.58]	+0.42 [-0.83, +1.75]	+0.38 [-0.42, +1.12]
Whole situation in Roman Urdu, n=5	-4.80	-3.80	-4.30 [-5.40, -3.10]
Roman Urdu quote in English narration, n=7	-0.29	0.00	-0.14 [-1.29, +0.86]

The second coder scored more leniently and every effect is smaller, but the direction holds on every contrast. Two intervals cross zero under the second coder alone. Both coders reproduced the first run's split, with almost the whole language effect falling in the five fully framed scenarios. The matched rerun below shows that split should not be read as a code-switching moderator, since it tracked how much each scenario's English condition had been changed. Dropping social appropriateness from the total leaves the main contrasts unchanged.

Per-dimension breakdown

Condition	Intent	Social appropriateness	Unsupported assumptions	Semantic preservation
implicit_english	1.33	0.92	1.33	0.92
roman_urdu_mixed	1.00	0.58	0.42	0.33
explicit_english	1.58	1.42	1.42	1.25
explicit_context_added	1.33	0.92	1.00	0.58

The largest drops under Roman Urdu framing are in unsupported assumptions, from 1.33 to 0.42, and semantic preservation, from 0.92 to 0.33. Intent inference holds up comparatively well. The model often identifies roughly what the speaker wants and then invents surrounding detail or loses the practical implication.

Question 3: path forward

Addressing these failures requires evaluating and training models on communicative intent rather than surface word overlap. My proposed approach starts with a contrastive pragmatic alignment corpus built around specific speech acts. Each scenario will contain four parallel representations: explicit English, indirect English, Roman Urdu framing, and natural code-switched text. For every item, structured pragmatic annotations will define the literal content, the speaker's illocutionary intent, interpersonal constraints such as age or kinship hierarchy, calibrated uncertainty, and acceptable responses.

Training with this resource will follow two stages. First, supervised fine-tuning across the four parallel conditions will teach models to connect different surface realizations to the same underlying pragmatic meaning. Second, preference optimization will reward answers that preserve intent and reflect the appropriate social context. The preference dataset will include hard negatives: responses that are grammatically fluent and literally correct, but nevertheless miss the intended politeness or indirect refusal.

Before large-scale alignment, a practical and low-cost first step is a targeted vocabulary and kinship-term audit. As the khala example showed, a missing word can look like a failure in social reasoning, and each problem requires a different solution. Finally, evaluation must test generalization on held-out expressions and social scenarios to ensure models are not simply memorizing specific phrasing. This diagnostic approach helps determine where multilingual understanding breaks down across scripts and everyday registers, providing a clearer view of model performance.

Limitations

One model, 12 scenarios, 48 prompts, one response per prompt. No estimate of within-item variance.

The first coder was unblinded and knew the hypothesis. The second coder was blind to condition labels, but language is visible in the prompt itself, so no coder can be fully blind to that factor. Agreement is moderate overall and weak on social appropriateness, so per-dimension results carry less weight than totals.

The two scenario groups in the first run received different amounts of manipulation, so its pooled language effect is not a clean estimate of a single quantity. The matched rerun addresses that for the seven affected scenarios, but rests on four usable scenarios after truncation losses.

The rerun used a 1,024-token budget while the first run used 180, so the two runs are not directly comparable on absolute scores. Only the within-run contrasts are.

This is a diagnostic evaluation on hand-constructed items. It is not a benchmark of Urdu, of Pakistani culture, or of code-switching in general. Social practice varies by family, region, age, and individual.