The Result Model under Inconsistent Knowledge: Theory and Experiments

Abstract

The result model provides a foundation for precedent-based reasoning, yet real case bases are often inconsistent, with precedents pointing to opposite outcomes. To address this, we augment the result model with the Log-Odds Precedent Aggregator (LOPA) - a Naive-Bayes-style log-odds combiner that treats each applicable precedent as uncertain evidence, learns its reliability from data, and produces both a decision and a confidence score. We evaluate LOPA on the DIAS dataset, comparing it against other extensions of the result model, and a strong machine-learning baseline. Results show that the symbolic and hybrid models perform on par with ML, and in several settings slightly better, while remaining fully interpretable. LOPA is especially robust when many weak precedents compete with a few strong ones, and its calibrated confidence supports coverage-reliability tradeoffs.