Rule extraction from autoencoder‐based connectionist computational models

Abstract

SummaryMapping problems are typical research topics related to natural language learning, and they include not only classification mappings but also nonclassification mappings, such as verbs and their past tenses. Connectionist computational models are one of the most popular approaches for simulating those mapping problems; however, their lack of explanatory ability has prevented them from being further used to understand the language learning process. Therefore, the work of extracting rational rules from a connectionist model is as important as simulating the mapping behaviors. Unfortunately, there is no available technique that can be directly applied in those computational models to simulate nonclassification problems. In this paper, an autoencoder‐based connectionist computational model is proposed to derive a rule extraction method that can construct “If‐Then” rational relations with high fidelity for nonclassification mapping problems. To demonstrate its generalizability, this computational model is extended to a modified version to address a multi‐label classification mapping problem related to cognitive style prediction. Experiments prove this computational model's simulation ability and its explanatory ability on nonclassification problems by comparing its fidelity performances with those of the classical connectionist computational model (multilayer perceptron artificial neural network), and its similar ability on a multi‐label classification problem (Felder‐Silverman learning style classification) by comparing its prediction accuracy with those of other rule induction techniques.