| dc.description.abstract | This dissertation makes an exploratory comparison between two semantics models, Latent Semantic Analysis (LSA) and a newly introduced HiMean model based on the HyGene architecture, in a medical decision-making context. Emphasis is placed on using real-world, human decipherable input to produce rational diagnoses. Base rate information is manipulated as a proxy to expertise or learning in different information environments, and outcomes on decision measures are examined. Model performance in terms of correct probe or query identification, alternative hypothesis generation, probe degradation resilience, probability judgments, and diagnostic capability is evaluated. Multidimensional scaling is also employed to investigate two-dimensional projections of the models’ respective semantic spaces. Experimental outcomes reveal that both the LSA and HiMean models, as well as HiMean variants perform well in a variety of conditions. The models produce performance tradeoffs between each other in terms of accuracy, judgment calibration, and robustness to probe error, though not in diagnostic capability. The models are demonstrated to be capable of utilizing non-trained data and producing identification accuracies up to 80%. Generally, both LSA and HiMean prove to be capable decision architectures with a wide variety of potential applications. Some thought is given to future work dedicated to a multi-agent decision system which capitalizes on the strengths of both models. | en_US |
