| Sumario: | Language models represent word meanings as vectors in a multidimensional space. Building on this property, this study offers a geometric perspective on parallelism in classical Chinese poetry, complementing traditional symbolic interpretations. To automatically detect parallelism in poetic verse, the authors trained a BERT-based classifier on a dataset of over 140,000 regulated poems (lüshi 律詩), achieving performance on par with state-of-the-art generative models such as GPT-4.1 and DeepSeek R1. Unlike general-purpose models, the custom classifier yields unique insights into how poetic meaning is encoded geometrically. The analysis shows that parallel lines exhibit alignment in the model's attention patterns: the 'key' vectors of corresponding characters point in the same direction, while this alignment disappears in non-parallel lines. This finding is interpreted through Peter Gärdenfors's theory of cognitive semantics, which posits that humans make sense of the world by organizing experience into distinct conceptual regions. The authors argue that parallelism functions as a bridging mechanism that temporarily unites these disparate domains of meaning, suggesting a deeper, geometric order that underlies language itself.
|
|---|