Research
I work on the syntax of case, agreement and alignment in Iranian languages, and on how language models learn grammatical structure.
Dissertation
Alignment in Human Language and Large Language Models: Evidence from Iranian Languages
| منداڵەکان دەیانترساندم | |
| mɨndaɫ-ɐk-an | dɐ=jan-tɨɾs-an-d-ɨm |
| kid-def-pl | ind=3pl-fear-v-pfv-1sg |
| ‘The kids were scaring me.’ | |
How a grammar sees it
Two agreement markers, two different arguments
In the Sorani past transitive, the clitic =jan agrees with the agent and the suffix -ɨm with the object. The tree is simplified; dashed arrows mark open questions.
How a language model sees it
The same sentence, in Sorani script
Morphemes
Filled: the agreement markers =jan (یان) and -ɨm (م)
Tokens (GPT-4o tokenizer)
=jan is its own token; -ɨm is merged with perfective -d (outlined)
A tokenizer splits text by how often character sequences occur in its training data, not by morpheme. Here it splits the roots mɨndaɫ and tɨɾs into pieces and joins -ɐk with -an.
My dissertation brings my theoretical and computational work together. It examines split alignment in Iranian languages, where case and agreement patterns change with tense and aspect, and asks whether large language models learn these patterns in the way human grammars organize them. By comparing the generalizations found in Kurdish varieties and related languages with what LLMs represent and produce, it uses alignment as a test case for how much grammatical structure these models acquire.
Research interests
Theoretical linguistics
Alignment and argument structure
My primary research investigates split alignment in Sorani Kurdish varieties such as Ardalani. I explore how case and agreement patterns are conditioned by tense and aspect, arguing for a model where these features are dissociated in the syntactic structure. This includes work on stative passives and what they reveal about the argument structure of the verb.
The syntax–morphology interface
I study how syntactic structures are realized morphologically. Working in Distributed Morphology, I analyze clitic placement, head movement and morpheme ordering as post-syntactic operations, including the timing of Vocabulary Insertion, the factors behind allomorphy, and the internal structure of derived nouns.
Human language technology
Tokenization for LLMs Current
Tokenization decides the units a model learns from, and so shapes everything it represents afterward. I ask how linguistically informed units, such as morphemes, affect what models learn, especially for morphologically complex languages.
Morpheme-based embeddings
In my M.S. work, I designed and implemented a morpheme-based embedding model in PyTorch for low-resource, morphologically complex languages. Building word representations from morphemes improved performance on downstream NLP tasks, with Sorani Kurdish as a case study.
LLMs and language acquisition
I use large language models as cognitive models, asking how their learning trajectories and error patterns can shed light on the stages of first language acquisition in children.
NLP tools for Sorani Kurdish Current
I am building a rule-based morphological segmenter for Sorani Kurdish that checks each proposed split against the corpus, along with a cleaned corpus of Sorani Kurdish collected from Wikipedia.
Projects
-
2020–now
Documentation and Comparative Grammar of Five Iranian Languages
This project studies the syntax, morphology and phonology of five Iranian languages. I contribute elicitation, transcription and analysis for Sorani Kurdish, Balochi, Ossetian and Persian. Project website
-
2025–26
Project ADELANTE
Culturally responsive, place-based educational opportunities for Latinx students in the Borderlands. I carried out sentiment analysis and data analysis.
-
2025–26
Pupillometric Insights: Mood Distinctions in Bilingual Spanish Learning
An experimental study of mood distinctions in bilingual Spanish learning. I wrote the Python code for the project.
-
Summer 2021
The Coeur d’Alene Online Language Resource Center
I prepared audio files for the online resource center, which preserves and provides access to the Coeur d’Alene language. Project website
-
2018–20
Atlas of Kurdish Language
Data gathering, manipulation and tagging for a comprehensive linguistic atlas of Kurdish dialects. Project website