Modeling Syntactic Knowledge With Neuro-Symbolic Computation

Abstract

To overcome the limitations of prevailing NLP methods, a Hybrid-Architecture Symbolic Parser and Neural Lexicon system is proposed to detect structural ambiguity by producing as many syntactic representations as there are interpretations for an utterance. HASPNeL comprises a symbolic AI, feature-unification parser, a lexicon generated using manual classification and machine learning, and a neural network encoder which tags each lexical item in a synthetic corpus and estimates likelihoods for each utterance’s interpretation with respect to the corpus. Language variation is accounted for by lexical adjustments in feature specifications and minimal parameter settings. Contrary to pure probabilistic system, HASPNeL’s neuro-symbolic architecture will perform grammaticality judgements of utterances that do not correspond to rankings of probabilistic systems; have a greater degree of system stability as it is not susceptible to perturbations in the training data; detect lexical and structural ambiguity by producing all possible grammatical representations regardless of their presence in the training data; eliminate the effects of diminishing returns, as it does not require massive amounts of annotated data, unavailable for underrepresented languages; avoid overparameterization and potential overfitting; test current syntactic theory by implementing a Minimalist grammar formalism; and model human language competence by satisfying conditions of learnability, evolvability, and universality.