Semantic norms and structural relations in Chinese semantic composition : a dataset and analysis

Abstract

Word meanings can combine, and they can be represented in many different ways. In our study, we present an exploration of the relationship between linguistic structural relations and semantic norm features in the semantic composition process of Chinese phrasal structures. To achieve this goal, we constructed a dataset containing 180 compositional structures, covering six types of linguistic structural relations. Subsequently, we recruited human raters and utilized Large Language Models (LLMs) to rate the relevance of these 180 compositional structures on the dimensions of Binder-style semantic norm features. We found that the prediction of LLMs are generally close to human ratings. Using this dataset, we further investigated whether the semantic norm feature representations of compositional structures can be approximated through composition operations using the two constituents. The results show that the semantic composition process of Chinese compositional structures is similar to that of English, being influenced by a combination of linguistic relations and semantic norm features, and exhibits non-linear compositional patterns. However, unlike English, Chinese compositional structures tend to be better approximated when their semantic norm features are represented in a lower-dimensional space. This not only results in a closer approximation to human ratings but also reveals the interactions among features. Although such a pattern may reflect the paratactic nature of Chinese, further empirical validation is needed to validate this hypothesis.