Computational Linguistics Commons^™

A Study Of Bengali Grapheme-To-Phoneme Conversion, Arundhati Sengupta

Dissertations, Theses, and Capstone Projects

Grapheme-to-phoneme (G2P) conversion is a foundational component of speech technology, yet the languages for which it has been carefully studied remain a small fraction of the world's languages. Bengali is spoken by over 230 million people, making it one of the world's most widely used languages, but it remains comparatively understudied in natural language processing, and G2P for Bengali in particular has received less attention.

In this dissertation, we use neural architecture comparison, error analysis, and explicit correction techniques to identify and address the sources of G2P error in Bengali. First, we train and compare four neural G2P architectures, including …

Measuring The Tokenization Premium: A Cost Audit For Underserved Language Communities, Avijit Roy, Proma Roy, Hrishitva Patel

Publications and Research

Large language models are increasingly deployed as general-purpose educational and technical assistance systems, but their basic infrastructure does not treat languages equally. One underexamined source of disparity is tokenization: semantically equivalent content can require substantially different token counts across languages, affecting API cost, latency, and usable context length before a model is even invoked. We introduce the Tokenization Equity Audit (TEA), a reproducible benchmark for measuring tokenization premiums in technical tutoring content. TEA evaluates three widely used tokenizers, GPT-4o’s o200k base, Qwen2.5-7B, and Mistral-7B, on a 120-item Python debugging corpus translated from English into Bengali, Hindi, Arabic, Tamil, and Yoruba. …

The Security Of Llm-Generated Code, Christopher Brian Gonzalez Ayala

Student Theses

The rapid adoption of Large Language Models (LLMs) in software development has transformed coding practices by enabling automated code generation, completion, and optimization. Despite these advantages, concerns persist regarding the security and reliability of LLM-generated code. This study presents a comprehensive evaluation of both the functional correctness and security of code produced by three prominent LLMs as of early 2026. A total of 4,800 code snippets were generated using 100 security-focused programming prompts derived from the OWASP Top 10:2025, translated across eight natural languages and two phrasing styles (literal and natural developer-oriented prompts). To assess performance, a multi-stage experimental framework …

G&P2p: A Multi-Source Approach To Grapheme To Phoneme Conversion, Chun-Yi Peng

Dissertations, Theses, and Capstone Projects

This thesis introduces G&P2P, a multi-source framework for grapheme-to-phoneme (G2P) conversion that integrates side pronunciations from multiple lexical resources. Unlike traditional single-source approaches, G&P2P fuses data from multi-sourced pronunciation dictionaries—including CELEX, PronLex, NETTalk, and WikiPron—through several fusion strategies. The goal is to improve model performance on out-of-vocabulary words through multi-source learning. Experiments were conducted with attentive LSTM, pointer-generator LSTM, and pointer-generator Transformer architectures. Models were trained on combinations of datasets and evaluated using word error rate (WER) across five random seeds.

Results show that fusing expert-curated dictionaries such as CELEX and PronLex consistently improves accuracy, achieving an 11.81-point absolute error …

A Machine Learning Approach To Disentangling Developmental Language Disorder From Typical Development In Russian-Speaking Children, Katsiaryna Aharodnik

Dissertations, Theses, and Capstone Projects

This study investigated a machine learning (ML) approach to identifying Developmental Language Disorder (DLD) in Russian-speaking children using narrative data. ML methods can capture subtle linguistic patterns that distinguish typical and atypical development, which is especially important in cross-linguistic contexts where morphosyntactic variation affects the manifestation of DLD. Diagnosis remains challenging in less-studied languages due to limited knowledge of language-specific deficits and a lack of validated assessment tools. This study evaluated whether ML algorithms can provide a more efficient alternative to traditional screening methods.

Two binary classification studies were conducted using corpus data: 1) classification of narratives told by Russian …

Propasafe-Hybrid: A Text-Based Hybrid Propaganda Detection Tool, Thomas Kimmeth, Avijit Roy, Vivek Sharma

Publications and Research

Propagandistic content increasingly circulates through online news and social media, where readers often encounter it with limited scrutiny, highlighting the need for reliable and fine-grained detection. This paper introduces Propasafe-Hybrid, a sentence-level system that integrates a fine-tuned transformer classifier with LLM-based technique classification to identify, label, and explain specific propaganda strategies. The pipeline generates actionable outputs, including highlighted sentences, technique assignments, and concise rationales, so users can immediately understand why a sentence was flagged and how each label was determined. To control inference cost, Propasafe-Hybrid employs a cost-aware pre-filtering stage that forwards only high-likelihood sentences to LLMs, reducing token usage …

Propasafe-Hybrid: A Text-Based Hybrid Propaganda Detection Tool (Ila 2026 Presentation), Thomas Kimmeth, Avijit Roy, Vivek Sharma

Publications and Research

This presentation introduces Propasafe-Hybrid, a hybrid system for sentence-level propaganda detection that combines offline transformer-based classification with selective large language model (LLM) explainability. The system employs a two-stage pipeline in which a local BERT-based classifier evaluates all input text and filters non-propagandistic content, while only high-confidence candidates are forwarded to an LLM for rhetorical technique labeling and explanation. This design enables cost-aware, privacy-conscious, and scalable analysis by reducing unnecessary reliance on external models.

Propasafe-Hybrid identifies propagandistic techniques such as loaded language, obfuscation, and appeal to fear, and generates concise natural language rationales that make these techniques interpretable to users. By …

Structural Silence: When Ai Infrastructure Fails Speakers Of Underrepresented Languages, Avijit Roy, Proma Roy

Publications and Research

Artificial intelligence tools for education and language support are increasingly framed as scalable responses to access gaps in under-resourced communities. Yet the infrastructure underlying these tools—training corpora, tokenization schemes, evaluation benchmarks, and deployment architectures—encodes a set of assumptions that systematically disadvantages speakers of underrepresented languages before a single model is trained. This paper examines those assumptions through the lens of Bengali, one of the world’s most widely spoken languages with roughly 285 million speakers (Ethnologue, 2025; International Communication and Leadership School, 2026), and the structural barriers that emerge when attempting to build AI-assisted educational tools for Bengali-speaking learners in low-connectivity …

Mind The Gap: Morphological Defectivity And Suffixal Competition In Polish, Szymon E. Zuberek

Dissertations, Theses, and Capstone Projects

This study investigates morphological defectivity and suffixal competition in the genitive singular of Polish masculine inanimate nouns. Drawing on survey-based grammaticality judgments from native speakers, it examines how respondents select between the suffixes -a and -u or reject both as unacceptable, thereby signaling defectivity. Mixed-effects logistic regressions revealed that defectivity was rare overall but patterned systematically by age, education, and region, with additional baseline variability across lexical items. Suffix choice showed a strong preference for -a, modulated by age, education, and region, with additional baseline variability across lexical items. These findings inform our understanding of paradigm structure, sociolinguistic variation, and …

Ctrl + Alt + Inner Speech: A Verbal–Cognitive Scaffold (Vcs) Model Of Pathways To Computational Thinking, Daisuke Akiba

Publications and Research

This theoretical paper introduces the Verbal–Cognitive Scaffold (VCS) Model, a cognitively inclusive framework which proposes the cognitive architectures underlying computational thinking (CT). Moving beyond monolithic theories of cognition (e.g., executive-function and metacognitive control models), the VCS Model posits inner speech (InSp) as the predominant cognitive pathway supporting CT operations in neurotypical populations. Synthesizing interdisciplinary scholarship across cognitive science, computational theory, neurodiversity research, and others, this framework articulates distinct mechanisms through which InSp supports CT. The model specifies four primary pathways linking InSp to CT components: verbal working memory supporting decomposition, symbolic representation facilitating pattern recognition and abstraction, sequential processing enabling …

Alle Or Elle: Automatic Speech Recognition On Louisiana French, Emily Chiu

Dissertations, Theses, and Capstone Projects

Applications of automatic speech recognition largely serve the most commonly spoken languages, but can cause harm through bias when used for speakers of underrepresented language varieties who are not adequately supported. This experiment’s goal is to reveal how state-of-the-art end-to-end ASR systems perform with Louisiana French, a nonstandard variety of French that has suffered a history of state-sanctioned language oppression in Louisiana.

Improving Low-Resource Translation With Finite State Grammars, Nicholas J. Uva

Dissertations, Theses, and Capstone Projects

Scarcity of training data continues to pose a problem for the development of neural machine translation systems for low-resource languages. This study develops a method for the incorporation of linguistic information into the training of neural machine translation models for low-resource languages, using morphological grammars created using finite state transducers. This study explores the benefits, historical background, and effectiveness of this approach. This study incorporates morphological tags into a pre-trained multilingual neural machine translation model using a dual encoder structure. This study finds an improvement in performance in the Irish-English translation scenario. This method offers promising results with low computational …

Controlling Emotional Text To Speech Using Complex Adverbial Phrases, Zainab T. Akande

Dissertations, Theses, and Capstone Projects

This study investigates the usage of adverbial modifiers from audiobook data as a resource for training speech synthesizers with a greater range of speech descriptions. The Tacotron2 text-to-speech (TTS) model was used for the purposes of this study. Utilizing the LibriTTS dataset, the Tacotron2 model is trained under two experimental conditions: one incorporating adverbial modifiers into the input text and the other without. The dataset preprocessing involves embedding descriptions using word embeddings and encoding speaker IDs with machine learning techniques. Additionally, the model architecture includes a prosody encoder inspired by prior research. Evaluation of the trained models involves subjective assessments …

Factors Affecting Pretonic Deletion In English Syllables: A Corpus Study, Whitney Barnett-Rhodes

Dissertations, Theses, and Capstone Projects

One type of variation in pronunciation that is associated with fast and/or conversational speech is massive reduction, in which entire syllables are deleted. We focus on a subtype of massive reduction, pretonic vowel deletion, where the vowel in an unstressed syllable is deleted and the flanking consonants form an onset cluster, V --> φ / C__C, such as the word Columbus reduced to the two-syllable pronunciation [klʌmbəs], resulting in the onset cluster [kl]. The main question asks what predictors are associated with pretonic vowel deletion. Furthermore, we evaluate the Lexicalist Hypothesis and Sonority Projection as explanations of the phenomenon. …

Uncovering The Mimicry Of Online Review Breadth And Depth And Its Subsequent Effect On Consumer Responses, Andrea Pelaez Martinez

Dissertations, Theses, and Capstone Projects

Word-of-mouth (WOM) in marketing occurs when consumers discuss a company's product or service or any consumption experience with their friends, family, and others with whom they have any relationship. With the advent of social media, this phenomenon has expanded rapidly into virtual environments where consumer conversation is enabled through chats, forums, social media posts, and online reviews. In response to this rapid growth of online WOM, academics and practitioners have focused their interest on this phenomenon and its implications on consumers, firms, and society. So far, the evidence of the critical role that online WOM plays in helping consumers make …

Computational Approaches To Linguistic Challenges In Arabic Speech Recognition, Enas Albasiri

Dissertations, Theses, and Capstone Projects

This dissertation aims to document the linguistic features of Arabic that pose challenges to speech and language technologies and advance these technologies by developing state-of-the-art computational tools focusing on automatic speech recognition (ASR), text normalization (TN), and corpus development. TN converts expressions such as numbers, dates, and times—named semiotic classes—from their written to their spoken domain, such as converting ‘$84.00’ to ‘eighty-four dollars’, while inverse text normalization (ITN) converts verbalized text to its written form. This conversion is an essential preprocessing step for text-to-speech (TTS), and post-processing step for ASR. Arabic presents a challenge for TN and ITN because one …

Expanding The Corpus Of Vocalized Hebrew Text: Compiling An Unvocalized Text Corpus And Building An Online Interface For Vocalization Annotation, Rachel Shanblatt Bloch

Dissertations, Theses, and Capstone Projects

Written modern Hebrew presents a unique challenge for training computational models for language processing because modern Hebrew text often lacks vocalization. The lack of available vocalized Hebrew data can lead to ambiguity in training these models and generally hinders work on natural language processing problems. The goal of this project is to contribute to the collection of vocalized Hebrew text by collecting and preprocessing a large corpus of unvocalized Hebrew text and building an online annotation tool. The annotation tool allows people to upload unvocalized Hebrew text, to annotate by adding Hebrew vocalization, and to download comma-separated values files of …

Retórica Intercultural En El Discurso Académico Universitario: Las Funciones Retóricas De La Citación En Los Trabajos De Fin De Máster Escritos En Español Y En Inglés Por Hablantes Nativos Y No Nativos, David Sanchez-Jimenez

Publications and Research

This research derives from the interest in learning the cultural differences in citation practices in the academic genre of Master's thesis of native Spanish (Ee), non-native Filipino writers of Spanish (Fe), native Filipino writers of English (Fi), and American writers of English. A total of thirty-two (32) master´s theses – eight (8) for each group – were analyzed. A quantitative and qualitative methodology was used to study this phenomenon based on the computerized textual analysis of the rhetorical function of citations arranged in typological classification that modified the outline proposed by Petrić in his 2007 article. The results obtained from …

How Do We Learn What We Cannot Say?, Daniel Yakubov

Dissertations, Theses, and Capstone Projects

The contributions of this thesis are two-fold. First, this thesis presents UDTube, an easily usable software developed to perform morphological analysis in a multi-task fashion. This work shows the strong performance of UDTube versus the current state-of-the-art, UDPipe, across eight languages, primarily in the annotation of morphological features. The second contribution of this thesis is a exploration into the study of defectivity. UDTube is used to annotate a large amount of data in Greek and Russian which is ultimately used to investigate the plausibility of Indirect Negative Evidence (INE), a popular approach to the acquisition of morphological defectivity. The reported …

Consonant (De)Gradation In Ingrian?, Andrea M. Harrison

Dissertations, Theses, and Capstone Projects

This paper will present a dual method toward data enrichment for low-resource languages. Using Yoyodyne -- a Fairseq-inspired neural library for small-vocabulary sequence-to-sequence generation -- a morphological generation task was tested across labeled data encompassing multiple stages of enrichment for the low-resource language Ingrian. Due to limitations in the available data for Ingrian, weighted finite-state transducers (WFSTs) were used to generate an expanded vocabulary via HFST's toolkit for Uralic languages, and GiellaLT, a source for FST-driven lexica for low-resource languages. Further stages of experimentation used labeled data from related, higher-resource languages (Finnish, Estonian) to encourage cross-lingual transfer in the interest …

Towards Interpretable Machine Reading Comprehension With Mixed Effects Regression And Exploratory Prompt Analysis, Luca Del Signore

Dissertations, Theses, and Capstone Projects

We investigate the properties of natural language prompts that determine their difficulty in machine reading comprehension tasks. While much work has been done benchmarking language model performance at the task level, there is considerably less literature focused on how individual task items can contribute to interpretable evaluations of natural language understanding. Such work is essential to deepening our understanding of language models and ensuring their responsible use as a key tool in human machine communication. We perform an in depth mixed effects analysis on the behavior of three major generative language models, comparing their performance on a large reading comprehension …

Topics For He But Not For She: Quantifying And Classifying Gender Bias In The Media, Tyler J. Lanni

Dissertations, Theses, and Capstone Projects

In this study, we used computational techniques to analyze the language used in news articles to describe female and male politicians. Our corpus included 370 subtexts for male candidates and 374 subtexts for female candidates, gathered through the New York Times API. We conducted two experiments: an LDA topic analysis to explore the data, and a logistic regression to classify the subtexts as either male or female. Our analysis revealed some noteworthy findings that suggest the possibility of developing a gender bias classifier in the future. However, to create a more robust understanding of bias, additional research and data are …

Neural Network Vs. Rule-Based G2p: A Hybrid Approach To Stress Prediction And Related Vowel Reduction In Bulgarian, Maria Karamihaylova

Dissertations, Theses, and Capstone Projects

An effective grapheme-to-phoneme (G2P) conversion system is a critical element of speech synthesis. Rule-based systems were an early method for G2P conversion. In recent years, machine learning tools have been shown to outperform rule-based approaches in G2P tasks. We investigate neural network sequence-to-sequence modeling for the prediction of syllable stress and resulting vowel reductions in the Bulgarian language. We then develop a hybrid G2P approach which combines manually written grapheme-to-phoneme mapping rules with neural network-enabled syllable stress predictions by inserting stress markers in the predicted stress position of the transcription produced by the rule-based finite-state transducer. Finally, we apply vowel …

Evaluating Neural Networks As Cognitive Models For Learning Quasi-Regularities In Language, Xiaomeng Ma

Dissertations, Theses, and Capstone Projects

Many aspects of language can be categorized as quasi-regular: the relationship between the inputs and outputs is systematic but allows many exceptions. Common domains that contain quasi-regularity include morphological inflection and grapheme-phoneme mapping. How humans process quasi-regularity has been debated for decades. This thesis implemented modern neural network models, transformer models, on two tasks: English past tense inflection and Chinese character naming, to investigate how transformer models perform quasi-regularity tasks. This thesis focuses on investigating to what extent the models' performances can represent human behavior. The results show that the transformers' performance is very similar to human behavior in many …

A Sentiment Analysis Of "Filipinx" On Twitter Using A Multinomial Naïve Bayes Classification Model, Clarisse Taboy

Dissertations, Theses, and Capstone Projects

On social media, the use of “Filipinx” as a gender neutral, inclusive term for “Filipino” tends to generate high user engagement, at times without regard for the original context in which the word appears. This project applies computational methods to collect a large dataset in English/Filipino from Twitter containing “Filipinx”, and to train a Naïve Bayes model to classify tweets into three sentiments: positive, neutral, and negative. My methodology takes inspiration from that of four related studies that similarly conducted sentiment analysis on English/Filipino tweets involving various topics, and whose resulting accuracy scores were compared side-by-side. Conducting sentiment analysis on …

Analysis Of The Inter-Annotators Agreement And Its Effect On The System Of Anti-Asian Hate Crime Detection On Twitter During Covid-19, Amir Toliyat

Dissertations, Theses, and Capstone Projects

Coronavirus disease 2019 (COVID-19) started in Wuhan, China, in late 2019, and after being utterly contagious in Asian countries, it rapidly spread to other countries. This disease caused governments worldwide to declare a public health crisis with severe measures taken to reduce the speed of the spread of the disease. This pandemic affected the lives of millions of people. Many citizens that lost their loved ones and jobs experienced a wide range of emotions, such as disbelief, shock, concerns about health, fear about food supplies, anxiety, and panic. All of the aforementioned phenomena led to the spread of racism and …

Brazilian Portuguese-Russian (Braporus) Corpus: Automatic Transcription And Acoustic Quality Of Elderly Speech During Covid-19 Pandemic, Irina A. Sekerina, Anna Smirnova Henriques, Aleksandra Skorobogatova, Natalia Tyulina, Tatiana V. Kachkovskaia, Svetlana Ruseishvili, Sandra Madureira

Publications and Research

This article presents the Brazilian Portuguese-Russian (BraPoRus) corpus, whose goal is to collect, analyze, and preserve for posterity the spoken heritage Russian still used today in Brazil by approximately 1,500 elderly bilingual heritage Russian–Brazilian Portuguese speakers. Their unique 100-year-old variety of moribund Russian is disappearing because it has not been passed to their descendants born in Brazil. During the COVID-19 pandemic, we remotely collected 170 h of speech samples in heritage Russian from 26 participants (Mage = 75.7 years) in naturalistic settings using Zoom or a phone call. To estimate the quality of collected data, we focus on two methodological …

Towards Explaining Variation In Entrainment, Andreas Weise

Dissertations, Theses, and Capstone Projects

Entrainment refers to the tendency of human speakers to adapt to their interlocutors to become more similar to them. This affects various dimensions and occurs in many contexts, allowing for rich applications in human-computer interaction. However, it is not exhibited by every speaker in every conversation but varies widely across features, speakers, and contexts, hindering broad application. This variation, whose guiding principles are poorly understood even after decades of entrainment research, is the subject of this thesis. We begin with a comprehensive literature review that serves as the foundation of our own work and provides a reference to guide future …

Linguistic Abstractions In Children’S Very Early Utterances, Qihui Xu

Dissertations, Theses, and Capstone Projects

How early do children produce multiword utterances? Do children's early utterances reflect abstract syntactic knowledge or are they the result of data-driven learning? We examine this issue through corpus analysis, computational modeling, and adult simulation experiments. Chapter 1 investigates when children start producing multiword utterances; we use corpora to establish the development of multiword utterances and a probabilistic computational model to account for the quantitative change of early multiword utterances. We find that multiword utterances of different lengths appear early in acquisition and increase together, and the length growth pattern can be viewed as a probabilistic and dynamic process.

Chapter …

A Study Of Entrainment In Speech As A Possible Predictor Of Perceived Trust, Mariana Graterol Fuenmayor

Dissertations, Theses, and Capstone Projects

This thesis explores the possibility of using features of speech as possible predictors of perceived trust. It specifically focuses on entrainment, the tendency of participants in a conversation to unconsciously adapt their manner of talking to become more or less similar to each other. We compile a corpus of interviews conducted in English, with different levels of formality and discussing different topics. With it, we distribute a set of surveys to assess raters’ judgment of the participants, focusing on whether they believe the interviewee is trusted by their conversational partner, and whether they find the interlocutors themselves trustworthy. We also …