The current relationship modeling paradigm, grounded in the observational i.i.d assumption, fundamentally misaligns with our causal knowledge understanding due to two key oversights: 1) the unobservable relations, which lead to undetectable hierarchical levels of knowledge, driving the need for model generalizability; 2) the cognitive relative timings, which crucially support our structural knowledge comprehension, resulting in inherent biases within the present Observation-Oriented paradigm. Adopting a novel Relation-Oriented perspective, this paper proposes a new framework to unify the various confusions surrounding causality learning, ranging from traditional causal inference to modern language models. Also, relation-indexed representation learning (RIRL) is raised as a baseline implementation method of the proposed new paradigm, alongside comprehensive experiments demonstrating its efficacy in autonomously identifying dynamical effects in relationship learning.
View on arXiv