How to keep the HG weights non-negative: the truncated Perceptron reweighing rule

Abstract

The literature on error-driven learning in Harmonic Grammar (HG) has adopted the Perceptron reweighing rule. Yet, this rule is not suited to HG, as it fails at ensuring non-negative weights. A variant is thus considered which truncates the updates at zero, keeping the weights non-negative. Convergence guarantees and error bounds for the original Perceptron are shown to extend to its truncated variant.