Date of Award
Spring 2026
Rights
Access is available to all users
Date Available to Non-EWU Users
2026-06-12
Document Type
Thesis
Degree Name
Master of Science (MS) in Computer Science
Department
Computer Science and Electrical Engineering
First Advisor
Dr. Dan Tappan
Second Advisor
Dr. Dan Li
Third Advisor
Dr. Dale Garraway
Abstract
Large transformer models achieve strong performance on natural language understanding tasks but require hundreds of millions of parameters and extensive pretraining. This thesis investigates whether graph neural networks operating on dependency parse trees can provide more parameter-efficient sentence representations for natural language inference, evaluated on two NLI tasks: entailment classification and semantic textual similarity.
Tree Matching Networks (TMN) adapt Graph Matching Networks to linguistic dependency trees with rich node and edge features, evaluated against a BERT baseline at matched parameter counts on identical training data. Tree Transformer Networks (TTN) extend TMN with transformer-based aggregation and tree-aware positional encodings, with component balance and CLS token ablations examining the aggregation bottleneck. All are evaluated on the SNLI and SemEval benchmarks across both matching and embedding paradigms. Additional experiments replace pooling with a pretrained sentence-transformer aggregator under varying propagation and freeze settings.
In controlled from-scratch comparisons, TMN outperforms the BERT baseline at matched parameter counts and scales monotonically with model size, and is effective on semantic similarity regression as well as entailment. TTN variants fall well short of TMN in matching accuracy; CLS token ablations indicate the gap is not explained by pooling strategy. Pretrained aggregation experiments produce mixed results: the GNN propagation stage reduces embedding performance relative to the pretrained model alone, and cross-graph attention degrades regression performance across GNN-enabled architectures.
Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial-No Derivative Works 4.0 International License.
Recommended Citation
Lunder, Jason P., "Tree-Based Graph Neura Networks for Natural Language Inference: From Structure-Only to Hybrid Architectures" (2026). EWU Masters Thesis Collection. 1012.
https://dc.ewu.edu/theses/1012
