hasanakhras/arabized-loanword-inflection

Blind spot: small LLMs can't inflect Arabized English nouns the way Arabic speakers do

Author: Hasan Mohammed Akhras · Fatima Fellowship technical challenge, Fall 2026

TL;DR

Small open models inflect Arabized English nouns (داتاست، سيرفر، لابتوب) much worse than native Arabic nouns from the same domain. In MSA, Gemma-3-4B gets 38% of native-noun inflections right but only 12% for loanwords. Instead of inflecting the loanword, Gemma usually replaces it with an Arabic translation (منيو → قائمة), and Qwen3-4B usually just repeats the singular. In Syrian colloquial, both models are weak even on native words, and loanwords make it worse.

1. Motivation: where this comes from

I'm a Syrian AI student studying in Makkah. With my classmates, in the lab and in group chats, we almost never say "مجموعة بيانات". We say داتاست, and when there are two of them we say داتاستين without thinking about it. The same goes for سيرفر، سكربت، لينك، فولدر and many others. Nobody taught us a rule for this. It is simply how people talk.

When I used LLMs in Arabic for technical work, I noticed they handle these words badly. They give forms nobody says, or they avoid the word completely. I checked the literature: Arabic morphology benchmarks such as YallaMorph only test native MSA vocabulary, and code-switching work mostly looks at speech recognition and translation. So this part of everyday Arabic is not really measured. I wanted to see if the problem is real and systematic, or just my impression.

This matters because a lot of Arabic speakers, especially students and developers, write and talk to models in this mixed way. A model that can't say "داتاستين" or "سيرفرات" naturally, or that "corrects" the word into a translation, does not really speak the Arabic these users speak.

2. What is tested

Given an Arabized English noun (e.g. لابتوب, laptop), can the model produce four inflections the way speakers use them, in two registers?

Inflection	Syrian colloquial example	MSA example
Dual	لابتوبين	لابتوبان
Plural	لابتوبات	لابتوبات
Definite (ال)	اللابتوب	اللابتوب
1st-person possessive	لابتوبي / اللابتوب تبعي	لابتوبي

Control (native baseline): the same four inflections for 14 native Arabic nouns from the same domain (ملف، جهاز، رابط، تطبيق، حساب، ...). If a model handles the native words but fails on loanwords, the failure is specific to loanwords and is not a general weakness in Arabic morphology.

3. Data (data/)

• 31 Arabized English nouns that I use in daily technical/academic speech (dataset, section, server, script, dashboard, password, link, menu, application, code, ...), plus a few older, established loans (بنك، فيلم، كرت، باص).
• 14 native Arabic nouns as the baseline.
• For every word × register × inflection I wrote the form(s) speakers actually use. Several forms are allowed per cell (e.g. plural of كود = اكواد / كودات; plural of سيكشن = سكاشن / سيكشنات).
• 359 scorable reference cells in total.
• reference_table.xlsx: the annotation sheet (with my confidence flags and notes). reference_long.csv: the same data, one row per cell.

How I built the reference. I picked nouns I actually use, mostly from programming, studying and social media. For each cell I wrote what people say, not what a grammar book would derive. When more than one form is used, I kept all of them (for example the broken plurals سكاشن، اكواد، شباتر next to the sound plurals). I marked my confidence for every word; I was less sure about some MSA forms (منيو، بوط، موبايل) and wrote that in the notes.

A few decisions:

• Possessive in Syrian: people say both لابتوبي and اللابتوب تبعي, so both are accepted.
• Old vs new loans: I added older, established loans (بنك، فيلم، كرت، باص) next to new tech words, to see if models do better on words they have seen for a long time.
• Removed words: ميت (meet) is written like the Arabic word ميّت (dead), so I removed it, and I removed تيرمينال and ستيكر because I wasn't satisfied with their forms.
• Baseline words come from the same domain, and some are the Arabic equivalents of loanwords in the list (رابط/لينك، تطبيق/ابليكيشن), so the comparison is as fair as possible.
• Verbs were left out of this version (see Future work).

4. Setup

• Models: Qwen/Qwen3-4B and google/gemma-3-4b-it (instruction-tuned, run in Colab, greedy decoding, max 64 new tokens, Qwen thinking mode off).
• Prompts: 3 paraphrases per cell. Syrian prompts are written in Syrian colloquial to pull the answer toward the dialect, and MSA prompts are in MSA. 1,077 prompts per model.
• Scoring: an answer is correct if any accepted form appears in it as a whole word, after removing diacritics and normalizing alef/hamza and final yaa.
• Main analysis uses the two direct-question paraphrases (718 answers per model). The third, fill-in-the-blank paraphrase turned out to be badly designed: models filled the blank with arbitrary words (e.g. "هذا ليس بوستك، بل لصّ"). It is reported separately in results/accuracy_by_paraphrase.csv and excluded from the headline numbers.
• Confidence intervals: 95% bootstrap over words (4,000 resamples).

Code: code/loanword_eval.ipynb.

5. Results

Model	Register	Native nouns	Loanwords	Gap (95% CI)
Gemma-3-4B	MSA	37.5%	12.1%	25.4 pts (16.2 – 35.5)
Gemma-3-4B	Syrian	19.6%	8.5%	11.2 pts (5.0 – 17.2)
Qwen3-4B	MSA	29.5%	17.5%	12.0 pts (3.0 – 21.0)
Qwen3-4B	Syrian	8.0%	6.9%	1.2 pts (−2.7 – 5.0), n.s.

In 3 of 4 settings the models are significantly worse on loanwords than on native nouns of the same domain. Per-inflection accuracy is in results/accuracy_by_slot.csv.

How the models fail (automatic categories, results/error_categories_auto.csv)

	Correct	Repeats the singular	Different word / translation	Wrong inflection
Gemma, loanwords, MSA	12%	7%	72%	5%
Gemma, native, MSA	38%	12%	43%	8%
Qwen, loanwords, MSA	18%	57%	14%	12%
Qwen, loanwords, Syrian	7%	82%	9%	2%

"Different word" mixes two things: Arabic translations (منيو → قائمة, سيكشن → قسم, داتاست → بيانات), which dominate in MSA, and unrelated or nonsense words (سيكشن → «بلاش», داتاست → «بزعم»), which are frequent in the Syrian prompts.

Examples from the outputs (all from the two direct paraphrases):

Word	Asked for	Model output	Accepted
منيو	MSA dual	قائمة (Gemma)	منيوان
داشبورد	MSA dual	لوحة (Gemma)	داشبوردان
موبايل	MSA possessive	هاتفي (Gemma)	موبايلي
داتاست	Syrian dual	داتاست (Qwen)	داتاستين
سكربت	MSA possessive	سكربتة (Qwen)	سكربتي
باص	Syrian dual	باصة (Gemma)	باصين
تست	MSA plural	تستون (Gemma)	تستات
سيرفر	MSA dual	سيرفان (Gemma)	سيرفران
شابتر	MSA plural	شابترات (Qwen)	شباتر

6. What I think is going on

Gemma treats the loanword as something to fix, not to inflect. When asked for the dual of منيو it answers قائمة; for داشبورد it answers لوحة. It gives a different word in 72% of loanword cases in MSA, against 43% for native words. My guess is that most formal Arabic text the model learned from avoids these words, so in "Arabic mode" it replaces them with the "proper" word instead of using them the way people do.

Qwen's problem is partly general. It repeats the singular for loanwords (57% in MSA) but also for native words (48%). So part of its failure is not following the instruction at all, which is why its gap is smaller. Still, in MSA it is clearly worse on loanwords.

When the models do try to inflect, the errors show they don't know what a loanword is. They add a feminine ـة (سكربتة، باصة، كرتة), use the human masculine plural ـون (تستون), or invent a dual (سيرفان). They also prefer the default sound plural (شابترات) over the broken plural people actually use (شباتر). This fits my main point: the models know some general Arabic patterns, but not the usage of these specific words.

Syrian colloquial is a blind spot in itself. Even on native words Gemma is at 20% and Qwen at 8%, and many Syrian answers are unrelated words. So for dialect speakers the problem is double: the dialect, and then the loanwords on top of it.

7. Proposed remediation

The main errors are avoidance (translating) and not inflecting at all, more than wrong rules. So the fix should teach the model that these words are normal Arabic words that take normal Arabic morphology.

1. Cheap first step: examples in the prompt. Give a few correctly inflected loanwords (داتاست → داتاستين / داتاستات) before the question. LexNeo-Bench found that adding linguistic knowledge to the prompt helps a lot with borrowing in Luxembourgish, so this is worth testing first.
2. Small fine-tuning dataset (LoRA). Build sentences in Syrian colloquial and MSA where loanwords appear inflected the way speakers use them, including broken plurals and the تبعي possessive. Most of these forms can come from a reference table like this one, checked by more native speakers.
3. Evaluation. Split by word: train on about 20 loanwords and test on the other 11, which the model never saw. If accuracy goes up on the unseen words while native-word accuracy stays the same, the model learned the pattern and did not just memorize words.

I didn't have time to run these in this submission; they are the next step.

8. Limitations

• One annotator (me) for the reference forms; Syrian forms reflect my own (Aleppine) usage and have not been checked across regions.

• Small sample: 31 loanwords, 14 native nouns, 2 models.

• Error categories are assigned automatically by simple rules and were not hand-verified.

• One of three prompt paraphrases was flawed and excluded.

• Both models are weak on Syrian colloquial in general, so the Syrian results mix a dialect weakness with the loanword effect.

• MSA duals: the oblique ـين form was accepted for some words but not all. Accepting it everywhere changes MSA accuracy by less than 1 point.

• The baseline has only 14 words, so its confidence intervals are wide.

9. Future work

• Verbs. While building the verb list I noticed something interesting: many English verbs are not conjugated directly in Syrian Arabic. We say عمل ديليت / منعمل ديليت / اعمل ديليت with the light verb عمل, while other verbs are fully Arabized and conjugated (كنسل، شيّك، كمّت). A good test is whether a model knows when to conjugate and when to use عمل. I want to build this set next.
• More annotators from different Syrian regions, and other dialects.
• An Arabic-specialized model (e.g. Falcon-H1-Arabic) to compare with the general models.
• Run the remediation in Section 7.
• Replace the fill-in-the-blank prompt with a better designed one.

Related work

• Reda et al., YallaMorph (2026), arXiv:2609.10153: LLM morphological generation for MSA, native lemmas only. No dialects, loanwords or code-switching.
• AraLingBench (2025), arXiv:2511.14295: Arabic linguistic competence, multiple choice.
• A Survey of Code-switched Arabic NLP (COLING 2025), arXiv:2501.13419: notes morphological code-switching but no generation benchmark.
• Al-Jarf (2024), Exploring Rule-based and Idiosyncratic Loanword Plural Forms in Arabic, IJLS 4(1): loanword plurals in speakers' usage (zero, broken, sound, English and hybrid plurals).
• Mushait & Al-Athwary (2020), Plural and Gender Inflection of English Loanwords in Colloquial Saudi Arabic.
• Kniaź & Zawrotna (2021), Embedded English verbs in Arabic-English code-switching in Egypt, IJB.
• LexNeo-Bench (2026), arXiv:2605.21227: LLMs and lexical borrowing in Luxembourgish; injecting linguistic knowledge into prompts helps.

Repository layout

data/reference_table.xlsx      annotation sheet (loanwords + native baseline)
data/reference_long.csv        one row per word × register × inflection
results/model_outputs_scored.csv   every prompt, model answer, score, auto error category
results/main_results_gap.csv   native vs loanword accuracy + bootstrap CIs
results/accuracy_by_slot.csv   accuracy per inflection
results/error_categories_auto.csv
results/accuracy_by_paraphrase.csv
code/loanword_eval.ipynb       Colab notebook that produced everything
figures/native_vs_loanword.png