Language models represent word meanings as vectors in a multidimensional space. This study leverages this property to explore one of poetry’s most elegant structures: parallelism. While this literary device appears across cultures, its prominence in classical Chinese poetry remains unparalleled, a feature often attributed to the unique nature of the Chinese writing system, where uninflected, monosyllabic units, represented visually as square-shaped characters, allow for a strict, position-by-position matching of poetic lines. In a classic couplet by the Tang poet Wang Wei 王維 (699–761), for instance, the line "The bright moon shines between the pines" 明月松間照 is mirrored by "A clear spring flows over the stones" 清泉石上流, where "bright" corresponds to "clear," "moon" to "spring," and so on, creating a syntactic and semantic symmetry that invites the reader to linger between the lines. The very terms used to describe this phenomenon, both in the West (from the Greek parallelos, "side by side") and in China (duizhang 對仗, evoking symmetrical honor guards), are rich with spatial imagery. Yet, over centuries of literary analysis, this physicality has often been sublimated into metaphysical or symbolic interpretations. Artificial intelligence offers a unique opportunity to revive these dormant spatial connotations, allowing an investigation of parallelism not just as a symbolic device, but as a tangible geometric phenomenon.

The project pursued a dual objective. The first was to establish an empirical baseline for the automatic detection of parallelism by benchmarking the performance of state-of-the-art, general-purpose language models (LLMs). While models like GPT-4.1 and DeepSeek R1 prove effective for this classification task, their application comes with caveats. As general-purpose tools not specifically trained for this type of literary analysis, their inner workings remain largely opaque. Furthermore, their use can lead to specific errors, including hallucinations that misalign concepts by ignoring positional constraints, incorrect word segmentation that compares single characters to multi-character phrases, and inconsistent tokenization that disrupts the fundamental one-to-one correspondence crucial for Chinese parallelism. The second, and central, objective was therefore to construct an interpretable, task-specific model to move beyond simple classification and investigate the geometric mechanisms that encode parallel structures.

The creation of the bespoke classifier began with the compilation of a dataset of over 142,000 regulated poems (lüshi 律詩). An initial heuristic—labeling inner couplets as parallel and outer ones as non-parallel, which is one of the central features of regulated poetry—was found to introduce significant noise, as many poets consistently employed parallel couplets throughout the entire poem, while others completely avoided them, thus retaining the compositional freedom characteristic of pre-Tang poetry. To refine the training data, a filtering technique based on network analysis was developed. A symbolic network was constructed where each node represented a Chinese character, and an edge connected any two characters that appeared in corresponding positions within nominally parallel couplets. The application of the Louvain method for community detection revealed eight robust communities of characters, representing statistically significant semantic and grammatical groupings (e.g., verbs, numbers, geographical names). This character network served as a filter to programmatically prune the dataset and remove the false positive (inner but non-parallel) and false negative (outer but parallel) couplets. The resulting SikuBERT-based classifier, fine-tuned on this cleaned data, achieved an F1 score of 0.78, a performance comparable to the benchmarked LLMs but with the crucial advantage of architectural transparency.

Analysis of the classifier’s internal mechanisms revealed a distinct geometric signature for parallelism that can be found in the model’s attention mechanism. Attention allows the model to weigh the relevance of words to each other through a system of "queries" and "keys." We hypothesized that for a couplet to be classified as parallel, the corresponding characters in each line must provide similar responses to the queries of the classification token (CLS). The results confirmed this hypothesis with high statistical significance. In parallel couplets, the "key" vectors generated by corresponding characters point in the same direction, creating a pattern of geometric isomorphism that is absent in non-parallel lines. This alignment is quantitatively stark: the average Pearson's correlation coefficient between the attention distributions of the two lines reaches 0.60 for parallel couplets, while for non-parallel couplets, the correlation is only 0.03.

Insofar as the structures learned by artificial neural networks can offer insights into cognitive processes happening in the brain, our findings carry potential implications for understanding the human experience of poetry. In particular, our discovery finds a theoretical grounding in Peter Gärdenfors's theory of cognitive semantics, which posits that the human mind organizes concepts not only as symbolic propositions but also within multidimensional conceptual spaces partitioned into distinct, convex regions around prototypes (Voronoi tessellation). We propose that parallelism functions as a cognitive bridging mechanism that temporarily unites disparate regions of this conceptual space. For example, when a poet pairs "sound" with "color," the parallel structure coerces these concepts into a shared dynamic role, reshaping their default representations from static properties into active, flowing event vectors. The AI model succeeds by detecting the resulting geometric congruence; the human reader experiences this same congruence as a synesthetic, embodied insight into the nature of the world. In this discovered geometry, poetry realizes its potential not merely to imitate appearances, as Plato would have it, but to guide the mind in appreciating the latent symmetries that structure meaning itself.