Ali Nirheche

Research

Projects on how speakers and computational models learn the patterns of language.

I study how speakers generalize the patterns of their language, especially patterns with variation and exceptions, and what kind of model best describes what they learn. I combine computational models of learning, experiments with native speakers, and corpora that I build, mostly on Moroccan Arabic, my native language.

Dissertation · Nonconcatenative morphology · Analogical and constraint-based models

Analogy or grammar? Generalizing the Moroccan Arabic plural

My dissertation, co-chaired by Gaja Jarosz and Michael Becker, asks how speakers and computational models generalize inconsistent, nonconcatenative patterns, using the Moroccan Arabic plural system as a case study. Plurals are formed either with a suffix (tomobil → tomobilat ‘cars’) or by changing the shape of the word (ktab → ktub ‘books’), and which option a word takes is only partly predictable. This makes the system an ideal test case for comparing analogical models, which generalize from similar stored words, with grammars of weighted constraints.

Read moreShow less

To study the system, I built the Moroccan Arabic Plurals Corpus and traced its unpredictability to a few historical changes.

Analogy. I showed that analogical models predict plurals best when they compare words by their vocalized template, the prosodic structure of the word with its vowels filled in, rather than by an abstract consonant–vowel template, as had been argued for Modern Standard Arabic.

Grammar. I have given constraint-based analyses of the broken plural, in Stratal Optimality Theory in my MA thesis and in parallel Optimality Theory in a handbook chapter. I also developed a learner for Maximum Entropy (MaxEnt) grammars that addresses the grammar–lexicon balancing problem.

Frequency. I work with Gaja Jarosz on studying how type and token frequency shape speakers’ generalizations in this system.

Papers, talks & corpora

  • in review Not too coarse, not too fine: Morphological generalization over vocalized templates in Moroccan Arabic. Morphology. preprint
  • in press Predicting the Unpredictable: The Case of Moroccan Arabic Broken Plurals. In A. Boudlal, A. Loutfi & A. Noamane (Eds.), The Handbook of Moroccan Arabic Linguistics. Springer.
  • in press A parallel, constraint-based account of Moroccan Arabic iambic broken plurals. In A. Korangy & K. Bensoukas (Eds.), The Handbook of North African Arabic. Springer. slides
  • corpus Moroccan Arabic Plurals Corpus. corpus
  • 2021 Broken plural formation in Moroccan Arabic: a stratal optimality theory account. Master's thesis, Memorial University of Newfoundland. pdf
Large language models · Speech technology · NLP for low-resource languages

Large language models and speech technology

I study how computational models learn language, and I help build resources for languages that current models serve poorly. This work includes speech models for automatic phonetic transcription and an evaluation dataset for machine translation of Moroccan Arabic.

Read moreShow less

Speech models for phonetic transcription. In my work with Virginia Partridge, Joe Pater, Parth Bhangla, and Brandon Prickett, we fine-tuned wav2vec 2.0, a neural speech model, on the TIMIT and Buckeye Corpora to transcribe the pronunciation of varieties of English automatically, so that phonological variation can be studied on a much larger scale.

Language technology for Moroccan Arabic. As a member of the data team at AtlasIA, an open-source community building language technology for Moroccan Darija, I work on linguistic analysis and grammatical rules for AI models. With colleagues at AtlasIA, I co-authored TerjamaBench, a culturally specific benchmark for evaluating translation models on Moroccan Darija.

Papers, talks & datasets

  • 2026 (with Virginia Partridge, Joe Pater, Parth Bhangla & Brandon Prickett) Automated phonetic transcription for varieties of English: wav2vec 2.0 fine-tuned on the Buckeye Corpus. Proceedings of the Annual Meetings on Phonology, 2(1). link
  • 2025 (with Virginia Partridge, Joe Pater, Parth Bhangla & Brandon Prickett) AI-assisted analysis of phonological variation in English. Talk, special session on Deep Phonology, AMP 2025, UC Berkeley. slides
  • 2024 (with Imane Momayiz, Aissam Outchakoucht & Omar Choukrani) TerjamaBench: A culturally specific dataset for evaluating translation models for Moroccan Darija. AtlasIA. link
Phonology · Syntax · Corpora · Experiments · MaxEnt grammars

Variation, exceptionality, and the organization of the lexicon

Many patterns in Moroccan Arabic are variable or have exceptions, in phonology, morphology and syntax. Combining corpora, experiments with native speakers, and probabilistic grammars, I study what speakers know about these patterns, how they extend them to new words, and how the organization of the lexicon explains which words behave as exceptions.

Read moreShow less

Sibilant harmony. In an experimental study, I showed that variable sibilant harmony in Moroccan Arabic is more likely to be triggered the closer the two sibilants are, and I modeled this distance effect with Agreement-by-Correspondence constraints in a MaxEnt grammar.

The definite article assimilation. The definite article l- in Moroccan Arabic assimilates in some [ʒ] initial words and not in others. In a corpus study, I showed that assimilation before [ʒ] depends on the sound that follows it, and in a nonce-word experiment, speakers extended these lexical statistics to new words. I model this variation using a MaxEnt grammar with lexically indexed constraints.

The diglossic Arabic lexicon. With Ali Idrissi, I study how the Moroccan Arabic lexicon is divided into native Moroccan Arabic words and words from Standard Arabic, and how this division explains which words are variable and/or behave as exceptions.

Negation. Moroccan Arabic has two negation markers: ma-…-sh, used with verbs, and mashi, used with other predicates. In my work with Abdelkader El Khaoua, we show, through a large-scale corpus study, that adjectival participles strongly prefer the verbal marker, while plain adjectives prefer mashi but often take the verbal marker, a sign that the verbal marker is spreading to all predicates. We model this variation using the Multiple Grammars framework.

Papers & talks

  • in review (with Abdelkader El-Khaoua) When Adjectives Act Like Verbs: Negation Variation and Change in Moroccan Arabic. Linguistics.
  • 2025 Modeling distance-based variable sibilant harmony in Moroccan Arabic with a MaxEnt grammar. Arabic Linguistics, 1(2), 188–215. link preprint
  • 2025 Variable and exceptional assimilation of the definite article [l-] in Moroccan Arabic. Proceedings of the Annual Meetings on Phonology, 1(1). link
  • 2026 (with Abdelkader El-Khaoua) When Adjectives Act Like Verbs: Negation Variation and Change in Moroccan Arabic. Talk, NWAV 54, Université de Montréal.
  • 2026 (with Abdelkader El-Khaoua) Variable Negation of Adjectival Predicates in Moroccan Arabic: A Large-Scale Corpus Study. Talk, 4th Moroccan Linguistics Forum, Hassan II University, Casablanca. slides
  • 2026 (with Ali Idrissi) Permeability and Stratification in the Moroccan Arabic Diglossic Lexicon. Talk, 39th Annual Symposium on Arabic Linguistics, Indiana University. slides
  • 2024 Variable and exceptional assimilation of the definite article [l–] in Moroccan Arabic. Poster, AMP 2024, Rutgers University. poster
  • 2024 Distance-based sibilant harmony in Moroccan Arabic. Talk, 37th Annual Symposium on Arabic Linguistics, Long Island University. slides
Diachronic phonology · Prosodic morphology · Constraint-based models

Modeling diachronic change: prosodic augmentation in Moroccan Arabic plurals

In my work with Michael Becker, we study a change in progress in Moroccan Arabic plurals, which are gaining an extra final vowel (ktub → ktuba ‘books’). We argue that the vowel is added to satisfy the prosodic constraint NonFinality, a type of repair that had been claimed not to exist.

Read moreShow less

We are now extending this work to the whole plural system and modeling the change as a sequence of grammars over time, using older descriptions of the language alongside contemporary corpus and speaker data.

Papers & talks

  • 2025 (with Michael Becker) NonFinality drives epenthesis in the Moroccan Arabic broken plural. Brill's Journal of Afroasiatic Languages and Linguistics, 17(2), 333–359. link preprint
  • 2025 (with Michael Becker) Prosodic augmentation in the Moroccan Arabic broken plural. Talk, Rencontres du Réseau Français de Phonologie (RFP) 2025, Agadir. slides
  • 2025 (with Michael Becker) NonFinality drives epenthesis in the Moroccan Arabic broken plural. Talk, PhPhoNE 2025, MIT. slides
Phonetics · Syllable structure · Production study

Phonetic evidence for syllable structure

In this work, I used acoustic measures of temporal stability to test competing phonological analyses of syllabification, and showed that this method can decide between them. I use Moroccan Arabic medial geminates as a case study.

Read moreShow less

Papers & posters

  • 2026 Using temporal stability to probe the syllabification of medial geminates: evidence from Moroccan Arabic. Glossa: a journal of general linguistics, 11(1). link
  • 2025 Using temporal stability to probe the syllabification of medial geminates: evidence from Moroccan Arabic. Poster, Rencontres du Réseau Français de Phonologie (RFP) 2025, Agadir. poster poster (French)
Tools

Software and corpora

Tools I have built for constraint-based modeling, and corpora for research on Moroccan Arabic.

  • MaxEntLabWeb app A browser-based tool for fitting Maximum Entropy grammars to categorical and variable data, with support for hidden structure. Launch the app ↗ Source code ↗ Documentation
  • MaxEnt with Hidden Structure in RShiny app An interactive app for generating MaxEnt grammars, with or without hidden structure, built on the HGR model (Staubs 2011). Launch the app ↗ User guide
  • Moroccan Arabic Plurals CorpusCorpus 1,166 singular–plural noun pairs in Moroccan Arabic, derived from the Darija Open Dataset, transcribed in IPA and annotated with gloss, pattern, and plural type. Access the corpus ↗ Details