Kolmogorov-Arnold network for word-level explainable meaning representation
Boris A. Galitsky · Preprints.org · 2024
We leverage the explainability feature of KAN network and build an explainable language model where certain neurons encode individual words and neuron activation is fully interpretable in terms of the basis of a word. To do that, we propose a continuous word2vec model where a meaning of a word is expressed by a continuous profile of distances of this word from the words in the basis of words which is interpolated. As a result, the whole KAN network can be interpreted as a sequential procedure with word expressions. We then proceed from words to logic programs and develop a clause learning technique based on KAN. A logic program construction process from facts now become fully interpretable. We follow the differentiable inductive logic programming technique, representing a logic program as a matrix learned by KAN. Hence, we obtain an efficient and fully interpretable rule learning approach.