Composition as Nonlinear Combination in Semantic Space: A Computational Characterization of Compound Processing

Abstract

Most Chinese words are compounds formed through the combination of meaningful characters. Yet, due to compositional complexity, it is poorly understood how this combinatorial process affects the access to the whole-word meaning. In the present study, we turned to the recent development in compositional distributional semantics, and employed a deep neural network to learn the less-than-systematic relationship between the constituent characters and the compound words. Based on the compositional representations derived from the computational model, we investigated the combinatorial process in terms of the degree of overlap between the compositional and the lexicalized representations as well as the degree of distinctness of the compositional representation. Analyses of lexical decision and eye-tracking data revealed the effects of both compositional attributes over and above the effects of constituent character features and compound features, indicating an active engagement of the combinatorial process in compound processing. Moreover, with the increase of compound frequency, and thus the increased likelihood that the holistic route prevails, these compositional effects appeared to be dampened. These findings, therefore, provided a computational characterization for the dual-route framework, which sheds light on the universal process of compound comprehension across different languages.