Publication Date

2025

Document Type

Thesis

Committee Members

Lingwei Chen, Ph.D. (Advisor); Wen Zhang, Ph.D. (Committee Member); Michael Raymer, Ph.D. (Committee Member)

Degree Name

Master of Science (MS)

Abstract

Natural-language inference (NLI) asks whether a hypothesis is entailed by, contradicts, or is neutral with respect to a premise. Modern transformers reach high raw accuracy on benchmarks such as SNLI, MNLI, and ANLI, yet they often rely on brittle lexical shortcuts and provide little insight into their decision process. This thesis shows that counterfactual-augmented knowledge distillation can simultaneously boost robustness and supply faithful, token-level explanations—without scaling model size. Four T5-v1_1 students (60M, 220M, 770M, 3B parameters) are trained under four curricula: (1) standard fine-tuning, (2) fine-tuning with free-text rationales, (3) multi-task distillation with naive counterfactuals, and (4) multi-task distillation with guided keyword counterfactuals. On ANLI Round 1, naive counterfactual training lifts T5-small from 0.408 → 0.671 accuracy (+44% error-rate reduction) and T5-base from 0.492 → 0.707, while guided counterfactuals push those scores to 0.678 and 0.729, respectively, and provide automatically recoverable “edit scripts” that flip the label in 82% of cases. Smaller gains persist for T5-xl (0.736 → 0.820). On a 15k / 900 eSNLI split, counterfactual supervision still reduces relative error by 2–6% despite a stronger baseline. Training cost rises sub-linearly: a T5-base+guided-CF run requires 5.8 GPU-hours versus 15 GPU-hours for a vanilla T5-large that attains lower accuracy. Qualitative analysis confirms that single-token edits correct habitual negation and attribute errors better than attention heat-maps. Overall, the results argue that smart data beats sheer scale: counterfactual-enhanced distillation yields lean NLI models that are both more accurate and transparently explainable, charting a path toward deployable, trustworthy language understanding systems.

Page Count

56

Department or Program

Department of Computer Science and Engineering

Year Degree Awarded

2025

ORCID ID

0009-0003-1723-8816


Share

COinS