arXiv subjects

Cody Carroll

Publications and source records attributed to Cody Carroll.

5 recordsLinked to original sources

arXiv2025

No Text Needed: Forecasting MT Quality and Inequity from Fertility and Metadata

We show that translation quality can be predicted with surprising accuracy \textit{without ever running the translation system itself}. Using only a handful of features, token fertility ratios, token counts, and basic linguistic metadata (language family, script, and region), we can forecast ChrF scores for GPT-4o translations across 203 languages in the FLORES-200 benchmark. Gradient boosting models achieve favorable performance ($R^{2}=0.66$ for XX$\rightarrow$English and $R^{2}=0.72$ for English$\rightarrow$XX). Feature importance analyses reveal that typological factors dominate predictions into English, while fertility plays a larger role for translations into diverse target languages. These findings suggest that translation quality is shaped by both token-level fertility and broader linguistic typology, offering new insights for multilingual evaluation and quality estimation.

cs.CL↗

arXiv2025

The Token Tax: Systematic Bias in Multilingual Tokenization

Tokenization inefficiency imposes structural disadvantages on morphologically complex, low-resource languages, inflating compute resources and depressing accuracy. We evaluate 10 large language models (LLMs) on AfriMMLU (9,000 MCQA items; 5 subjects; 16 African languages) and show that fertility (tokens/word) reliably predicts accuracy. Higher fertility consistently predicts lower accuracy across all models and subjects. We further find that reasoning models (DeepSeek, o1) consistently outperform non-reasoning peers across high and low resource languages in the AfriMMLU dataset, narrowing accuracy gaps observed in prior generations. Finally, translating token inflation to economics, a doubling in tokens results in quadrupled training cost and time, underscoring the token tax faced by many languages. These results motivate morphologically aware tokenization, fair pricing, and multilingual benchmarks for equitable natural language processing (NLP).

cs.CL↗

arXiv2021

Learning delay dynamics for multivariate stochastic processes, with application to the prediction of the growth rate of COVID-19 cases in the United States

Delay differential equations form the underpinning of many complex dynamical systems. The forward problem of solving random differential equations with delay has received increasing attention in recent years. Motivated by the challenge to predict the COVID-19 caseload trajectories for individual states in the U.S., we target here the inverse problem. Given a sample of observed random trajectories obeying an unknown random differential equation model with delay, we use a functional data analysis framework to learn the model parameters that govern the underlying dynamics from the data. We show existence and uniqueness of the analytical solutions of the population delay random differential equation model when one has discrete time delays in the functional concurrent regression model and also for a second scenario where one has a delay continuum or distributed delay. The latter involves a functional linear regression model with history index. The derivative of the process of interest is modeled using the process itself as predictor and also other functional predictors with predictor-specific delayed impacts. This dynamics learning approach is shown to be well suited to model the growth rate of COVID-19 for the states that are part of the U.S., by pooling information from the individual states, using the case process and concurrently observed economic and mobility data as predictors.

math.AP↗

arXiv2021

Latent Deformation Models for Multivariate Functional Data and Time Warping Separability

Multivariate functional data present theoretical and practical complications which are not found in univariate functional data. One of these is a situation where the component functions of multivariate functional data are positive and are subject to mutual time warping. That is, the component processes exhibit a common shape but are subject to systematic phase variation across their domains in addition to subject-specific time warping, where each subject has its own internal clock. This motivates a novel model for multivariate functional data that connects such mutual time warping to a latent deformation-based framework by exploiting a novel time warping separability assumption. This separability assumption allows for meaningful interpretation and dimension reduction. The resulting Latent Deformation Model is shown to be well suited to represent commonly encountered functional vector data. The proposed approach combines a random amplitude factor for each component with population based registration across the components of a multivariate functional data vector and includes a latent population function, which corresponds to a common underlying trajectory. We propose estimators for all components of the model, enabling implementation of the proposed data-based representation for multivariate functional data and downstream analyses such as Fr\'echet regression. Rates of convergence are established when curves are fully observed or observed with measurement error. The usefulness of the model, interpretations, and practical aspects are illustrated in simulations and with application to multivariate human growth curves and multivariate environmental pollution data.

stat.ME↗

arXiv2020

Cross-Component Registration for Multivariate Functional Data, With Application to Growth Curves

Multivariate functional data are becoming ubiquitous with advances in modern technology and are substantially more complex than univariate functional data. We propose and study a novel model for multivariate functional data where the component processes are subject to mutual time warping. That is, the component processes exhibit a similar shape but are subject to systematic phase variation across their time domains. To address this previously unconsidered mode of warping, we propose new registration methodology which is based on a shift-warping model. Our method differs from all existing registration methods for functional data in a fundamental way. Namely, instead of focusing on the traditional approach to warping, where one aims to recover individual-specific registration, we focus on shift registration across the components of a multivariate functional data vector on a population-wide level. Our proposed estimates for these shifts are identifiable, enjoy parametric rates of convergence and often have intuitive physical interpretations, all in contrast to traditional curve-specific registration approaches. We demonstrate the implementation and interpretation of the proposed method by applying our methodology to the Zürich Longitudinal Growth data and study its finite sample properties in simulations.

stat.ME↗