11/06/2026
By Mohamed Elgaar

The Kennedy College of Sciences, Miner School of Computer & Information Sciences, invites you to attend a doctoral dissertation defense by Mohamed Elgaar, titled "Linguistic Knowledge for Steering Learning and Generation Dynamics in LLMs."

Details

  • Date: Friday, Nov. 20, 2026
  • Time: 10 a.m. EST
  • Location:Zoom meeting

Committee members

  • Hadi Amiri (Advisor), Assistant Professor, Computer Science Department, UMass Lowell
  • Hong Yu, Professor, Miner School of Computer and Information Sciences, UMass Lowell
  • Anna Rumshisky, Associate Professor, Miner School of Computer and Information Sciences, UMass Lowell
  • Wei Xu, Associate Professor, College of Computing, Georgia Institute of Technology

Abstract

This dissertation investigates how explicit linguistic knowledge can guide the learning, generation, and evaluation of language models. Linguistic complexity is represented through measurable lexical, syntactic, and discourse properties rather than a single universal notion of difficulty. We study these properties as signals for curriculum design, as targets for controlled generation, and as diagnostic variables for understanding extraction failures in clinical narratives.

For supervised learning, we introduce HuCurL, a curriculum discovery framework that searches over training schedules using prior estimates of sample difficulty. The discovered curricula extend beyond conventional easy-to-hard schedules and transfer across datasets and model sizes. We then develop Ling-CL, which estimates the relevance of linguistic indices from validation loss and combines them into multiview curricula. Its strongest gains occur under evaluation across linguistic difficulty strata, where aggregate task scores can conceal substantial weaknesses.

We extend the investigation to single-pass pretraining through experiments with Pythia models from 14M to 1B parameters and budgets of up to 300B tokens. Linguistically motivated orderings change the time spent in shared training phases and avoid a late spectral spike associated with deteriorating corpus-sample loss in small models. The aggregate advantage is not demonstrated in the larger evaluated comparisons. An exploratory scaling analysis locates much of the loss benefit late in training and distinguishes conditional larger-model projections from validated forecasts. Domain-order controls reproduce much of the effect, showing that linguistic sorting on a heterogeneous corpus also acts as a schedule over data sources. These findings characterize the conditions under which ordering affects learning rather than establish a general improvement to modern LLM pretraining.

For generation, LingGen supports variable attribute subsets through an attribute encoder and a power-law masking strategy that exposes the model to both dense and sparse control configurations. LingConv conditions paraphrase generation on 40 linguistic attributes and uses inference-time quality control to improve adherence while preserving semantic content. An interactive LingConv toolkit makes these controls available through diverse generation, reference-matched paraphrasing, and manual attribute adjustment.

Finally, we study linguistic reliability in medical decision extraction. MedDec provides expert-annotated discharge summaries, and MedDecXtract supports inspection and annotation of extracted decisions. Linguistic analysis reveals systematic extraction blind spots: validation recall falls from 58% to 24% across stopword-proportion strata, and relaxed matching raises overall recall from 48% to 71%, indicating substantial boundary disagreement. MedExACT extends this work into a shared benchmark that combines span and token performance with worst-group robustness. MedDecXtract-XAI adds inspection of supporting sentences and decision-linked explanations. Together, the studies establish explicit linguistic measures as useful tools for steering generation and analyzing learning and reliability, while identifying task, scale, and data-composition limits on their effectiveness.