Carleton University Institutional Repository
Not a member yet
20816 research outputs found
Sort by
Knowledge-Enhanced Language Models: Integrating Structured Knowledge Graphs in Large Language Model for Multiple-Choice Question-Answering Tasks
Full text linkLarge language models have a wide range of applications, but they still need to be supplemented for specific domain problems. In addition, large language models sometimes suffer from hallucination problems, producing answers that seem reasonable but are actually unfounded. This research investigates the integration of large language models with knowledge graphs to enhance performance on natural language processing tasks, particularly focusing on multiple-choice question-answering tasks. We propose several novel methods leveraging structured knowledge from knowledge graphs to improve the accuracy and relevance of large language model responses. Our contributions include a method to augment text classification tasks using knowledge graphs, which demonstrates performance improvements in sarcasm detection and news category classification, especially when training data is limited. We also introduce the ConKGP method, which enhances the QA-GNN model with perturbations and a contrastive learning framework, achieving better results on the CommonsenseQA and MedQA-USMLE datasets. Our Triple-Aware Reasoning approach integrates knowledge graphs with large language models by extracting and filtering relevant triples, improving the reasoning capabilities of large language models on the CommonsenseQA and OpenBookQA datasets. The TriRAG method, which innovatively uses triples instead of text to provide external knowledge for the language model, addresses the limitations of traditional RAG methods by using triples to build a knowledge graph, enhancing retrieval accuracy and contextual relevance, particularly in educational question-answering tasks. Finally, our MediTriR method improves large language model performance in medical question-answering tasks by retrieving and converting relevant medical literature into triples, achieving substantial accuracy improvements on MedMCQA and Medical QA datasets. These methods show the effectiveness of combining structured knowledge with large language models, highlighting potential applications in various specialized domains
Do Habitat Suitability Models Reflect Actual Movement? Insights from Wild Turkey Dispersal in Maine
Full text linkHabitat suitability models (HSMs) are widely used to infer landscape connectivity and animal movement paths, but their reliability for predicting movement remains unclear. This study evaluated HSM predictions for wild turkey (Meleagris gallopavo) fall dispersal using MaxEnt models based on citizen science presence data and GPS-tracked movements in Maine, USA. Two models, one using all-season data and another fall-specific, were compared to assess predictive accuracy. While turkeys generally moved through higher-suitability areas, suitability estimates at movement points overlapped extensively with random points. The fall-specific model showed only marginal improvement over the all-season model. Results indicate that HSMs built on presence data alone may inadequately predict movement paths. Connectivity models relying on such HSMs risk substantial uncertainty, underscoring the need to validate connectivity predictions with independent movement data to ensure effective conservation planning
An Injection Locked Oscillator in 0.1um GaAs pHEMT for Local Oscillator-Based Phase Shifting Phased Arrays
Full text linkThis thesis presents the design of a low-power Injection-Locked Oscillator (ILO) in a 0.1μm GaAs pHEMT technology, operating at 11.7GHz for Ku-band Low-Earth Orbit (LEO) satellite communication. Two versions were implemented, targeting different applications. The first version was a low-output power version targeting systems with active mixers. The second version is a high-output power version that consumes more power but is capable of driving passive mixers. Although a tapeout was not possible, the chip was simulated extensively across various process and temperature conditions. The low-power variant has an output power of 1.7dBm with a total DC power consumption of 13.3mW. The high-power variant achieves an output power of 7.7dBm with a total DC power consumption of 22.7mW. The ILO achieved a free-running phase noise of -103dBc/Hz at a 1MHz offset, and an injection lock range of +/-500MHz with an injection signal of -4dBm
Automation & Customization: Rethinking Precast Design with Robotics and Modular Systems
No full textThis thesis reimagines precast concrete foundation manufacturing through modular design, robotic automation, and DFMA (Design for Manufacturing and Assembly). It introduces a kit-of-parts for customizable foundation panels and a catalogue of modular, reusable formwork components—both developed through a DFMA process—to reduce material waste, labor, and production time. A closed-loop feedback system between design and fabrication ensures precision and adaptability. The robotically assembled formwork accommodates both standard and complex geometries, streamlining the digital workflow and architectural flexibility. Situated in the Canadian context, this research proposes a scalable, off-site fabrication strategy addressing labor shortages and housing demand. Panels use bolted double butterfly joints for rapid installation, adaptability, and disassembly—eliminating on-site pouring and single-use hardware. Validated through iterative prototyping and a 1:10 scale foundation, this work advances precast techniques into prefabricated on-site assembly. It frames automation as a tool for ecological efficiency, elevated craft, and mass customization in industrialized construction
