From Character to Poem: Nested Contexts and Scalar Limits of Parallelism Detection in Classical Chinese Poetry

1. Introduction: The Poetics of Constraints

Computational Literary Studies (CLS) confronts a mismatch between machine learning’s demand for discrete, invariant “ground truths” and the contextual, interpretive nature of literary criticism (Long & So, 2016; Piper, 2020). While skeptics argue that digital tools sever texts from their contexts (Bode, 2018, pp. 17–35; Da, 2019) and lack human-like understanding (Bender & Koller, 2020), the counter-assumption that maximizing context invariably improves performance ignores the noise inherent in large-scale data processing, the varying amounts and types of computation required by different contexts, and, most importantly, the fundamental question of what “context” actually is and whether the answer is itself context-specific. This article addresses these tensions by operationalizing “context” as a tunable hyperparameter within a classification task and by exploring the question: what is the appropriate unit of analysis for modeling a literary feature that is itself scale-dependent?

One must first delineate the computational complexity of the task at hand. I focus on parallelism (dui’yu 對語, duizhang 對仗, or dui’ou 對偶), one of the defining features of the Chinese regulated verse (lüshi 律詩) (Cai, 1989; Plaks, 1988). In the standard eight-line (“octave”) format, the poem is structurally divided into four couplets: the Head (shoulian 首聯), the Jaw (hanlian 頷聯), the Neck (jinglian 頸聯), and the Tail (weilian 尾聯). Prosodic law since the Tang (618–907) dictates that the “inner couplets” (the Jaw and the Neck) must exhibit strict parallelism, demanding a precise, one-to-one correspondence between words in identical positions across two lines, in addition to a consistent rhyme scheme and tonal arrangement (Watson, 1971, 109–113). This structural principle can be observed in Bai Juyi’s 白居易 (772–846) poem Grass on the Ancient Plain: Farewell to a Friend 賦得古原草送別:

In the third, parallel couplet, the initial adjective yuan 遠 (“distant”) maps to qing 晴 (“bright”), and the noun fang 芳 (“fragrance”) maps to cui 翠 (“greenery”); both pairs describe sensual qualities. The central verb qin 侵 (“encroaches”) mirrors jie 接 (“meets”), simultaneously contrasting and linking nature with culture. Finally, the object phrase gu dao 古道 (“ancient road”) is counterpoised by huang cheng 荒城 (“ruined town”). Each element in the first verse finds its counterpart in the second, forming what Yu-kung Kao describes as “a complete musical line, [where] both exposition and response form a self-contained cycle” (Kao, 1986, p. 337).

However, this simplicity is deceptive. While parallelism may appear to be a lexical game of retrieving words from a dictionary, it is fundamentally a context-dependent operation. As linguists have long noted, Chinese characters lack fixed morphological markers; their grammatical function is fluid and determined by their syntactic environment rather than inherent word class (Chao, 1968, p. 501; Pulleyblank, 1995, pp. 10–13). Consider the character cui 翠 in the sixth line. In isolation, it is ambiguous: it can function as an adjective (“green” or “emerald”) or as a noun (“greenery,” “jadeite,” or even “kingfisher feathers”). One cannot simply look at the character itself to determine its meaning; we may need to verify the opposing character in the parallel line. Because cui 翠 is paired with fang 芳, which functions here more readily as a noun (“fragrance”), cui is biased toward a nominal reading (“greenery”). Similarly, in the second couplet (lines 3 and 4), the character chun 春 could be a temporal noun (“springtime”), an attributive adjective (“vernal”), or even a temporal adverb (“during spring”); its classification as an adjective modifying feng 風 (“wind”) is supported by its parallel counterpart ye 野 (“wild”), which modifies huo 火 (“fire”). While ordinary reading is “linear and forward,” the reading of parallel structure “diverts the reader’s attention to the side, demanding that he pay attention to the corresponding lateral line” (Kao, 1986, p. 367); “parallelism always recycles the syntactic flow backwards onto itself” (Kao and Mei, 1971, p. 57). The verification of parallelism is not a static lookup but a mutual resolution of ambiguity.

This complexity is further compounded by the fact that the requirement for parallelism is itself structurally bounded (Fuller, 2017, pp. 31–33). For example, a poem of four rigidly parallel couplets is traditionally disparaged as “stiff”; the outer couplets are typically exempt (Lee et al., 2018). Moreover, our assumption that the second line provides a disambiguating signal for the first line or vice versa is itself dependent on our prior knowledge (and the resulting cognitive expectation) that the couplet is intended to be parallel. “Even when it survives independently, a couplet is always part of a complete poem… a couplet describing a ‘far’ scene will often be followed by a couplet describing a ‘near’ scene; a couplet on parting ‘here’ will ask a couplet speculating on the traveler’s state of mind ‘there.’” (Owen, 1985, pp. 91, 100). Judging the parallelism of a couplet, therefore, might require looking beyond the local relationships to the entire piece.

2. The Hierarchy of Meaning: Defining the Units of Analysis

My objective is to determine the semantic parallel status of a given couplet. To this end, I operationalize the “unit of analysis” by defining three distinct scopes of context, each representing a different hypothesis about the most effective scale for recovering the parallelism signal (Figures 1, 2, and 3). In all cases, the input sequence provided to the model includes poetic content and special tokens like [CLS] (classification) and [SEP] (separator), as defined in the BERT architecture (Devlin et al., 2019).

Figure 1

Figure 2

Figure 3

3. Methodology and Experimental Design

To empirically test the impact of different context windows, I aggregated the raw text of over 140,000 poems spanning the Tang, Song, Yuan, Ming, and Qing periods from public repositories, enforcing a regex filter to exclude non-Chinese characters, removing duplicates via full-string hashing, and filtering the corpus to isolate pentasyllabic octave poems. To establish a “ground truth,” I employed a SikuBERT-based classifier previously fine-tuned for this philological task (Kurzynski et al., 2024; 2025). This “teacher” model generated binary labels (parallel/not parallel) for every couplet in the dataset, thereby encoding the feature of parallelism at a fixed, couplet-level scale. Any poem where the prediction confidence fell below 0.8 for any constituent couplet was discarded. This “silver standard” approach (“silver” because not error-free) allowed me to generate the large-scale labeled data necessary for multi-level training.

The filtered data was then structured into a tripartite hierarchy to support different context windows, encompassing four distinct datasets. For character-level training, I used a community detection method described in an earlier study (Kurzynski et al., 2025). Specifically, I drew an edge between any two characters that occurred at the same positions in inner (nominally parallel) couplets in the dataset and then applied the Louvain method for community detection (Blondel et al., 2008) to discover groups of frequently matching characters. In the next step, I selected only those couplets where either all character pairs shared a community (perfect parallelism) or none did (complete lack of parallelism), and from these strictly filtered couplets, I extracted individual character pairs to serve as positive and negative examples. For couplet-level training, I aggregated full couplet strings with their corresponding binary labels. For poem-level tasks, I created two distinct datasets: a sequence labeling dataset (four labels per poem) and a binary classification dataset (one label per poem). Finally, I fine-tuned separate instances of a SikuBERT-based model (Wang et al., 2022) for each unit of analysis. The base experiment for all models consisted of a single training epoch; the Poem-1 and Poem-4 models were also trained for a second epoch for comparison. Where possible, both train and test splits (9,000 and 1,000 examples, respectively) were balanced to provide an equal number of binary labels for each category (see Limitations). The performance of all models was then compared on the unified task of determining couplet parallelism (as labeled by the “teacher” model) using three inference strategies:

1. Direct Evaluation: The Couplet Model was evaluated directly on a test set of couplets.

2. Bottom-Up Aggregation:

   • • I used the Character Model to generate an aggregated prediction for the couplet, which was determined by a majority vote: the couplet was classified as parallel only if at least three of the character pairs were predicted as parallel.

   • • I used the Couplet Model to predict the status of the entire poem. If the two inner couplets are predicted as parallel, the entire poem was classified as “regulated”.

3. Top-Down Inference: I used the Poem Models to infer predictions for the specific couplet:

   • • Via Poem-4: The full poem is fed to the model and the labels for each couplet are extracted.

   • • Via Poem-1: If the model classifies the poem as “regulated,” I infer that both inner couplets are parallel. This inferred status is then compared against the ground truth.

By comparing the accuracy of these methods on the exact same test set of couplets, we can measure the performance loss under deliberate misalignment. The entire pipeline was repeated for 100 trials, re-sampling new train/test splits for each run.

4. Results and Discussion: The Goldilocks Hypothesis of Context Alignment

The results are summarized in Table 1 and Figure 4. The following discussion explores the performance of each model as a decoder of the teacher’s signal, considering the trade-offs between accuracy and other forms of analytical insight.

Figure 4

5. Implications for Digital Humanities Benchmarking

The above results suggest that the ability to computationally recover a given feature depends on modeling alignment and as such gesture towards a regulative logic: to analyze a literary phenomenon, we proceed as if it were encoded at a particular scale and then seek a model architecture aligned with this postulated structure. A classifier designed to detect irony in dialogue may require a different contextual window from one designed to track the evolution of narrative voice. It should be noted that such a regulative argument remains agnostic as to the actual ontological status of the analyzed feature and pertains only to classifiers sensu stricto, i.e., task-specific models with a restricted output space, rather than general-purpose models only employed for classification tasks, in which case the output space is much larger (the model’s entire vocabulary). This perspective transforms the classifier from a passive discoverer of textual truth into an active participant in the hermeneutic process, testing not only the text itself but also the very viability of the scalar assumptions built into its architecture.

The results also demonstrate that digital humanities (DH) benchmarking should be a multidimensional evaluation of a model’s fitness for different scholarly purposes, one that weighs not only accuracy but also the distinction between interpretability and explainability (Rudin, 2019). In this project, the Character → Couplet classification offers structural interpretability: because the final decision is a simple aggregation of five independent character-pair judgments, we can transparently trace a couplet’s classification back to the specific predictions made for each pair. In contrast, the Poem-1 classification is much harder to interpret; we cannot determine whether the model’s validation stems from a genuine detection of parallel syntax, a recognition of the author’s style, or a simple reliance on the poem’s rhyme scheme. That said, the aligned models are more amenable to explainability, allowing for the use of post-hoc techniques to generate approximations (explanations) of their behavior.

In particular, the large-context models such as Couplet or Poem-1/Poem-4 offer geometric insights into how literary phenomena are represented internally by the Transformer models. Analysis of the Poem-1 model confirms an earlier result (Kurzynski et al., 2025): parallel couplets (unlike non-parallel ones) exhibit geometric alignment across the two lines, where corresponding positions are attended to in patterned ways, and the “key” vectors for matched characters point in similar directions in the model’s learned space. In the present study, Poem-1 learns that the inner couplets determine the “regulatedness” of the poem and consequently directs almost all its attention to them (Figure 5). Within these inner parallel couplets, the model distributes attention isomorphically across heads: higher attention on the first character of the first line, for instance, frequently corresponds to higher attention on the first character of the second line, and so on. Such insights into vector poetics carry cognitive implications, insofar as artificial models offer computational analogies to the human brain (Gärdenfors, 2000), but are by necessity unavailable to small-context models.

Figure 5

6. Conclusion

The “contextual turn” in natural language processing (NLP), exemplified by the groundbreaking discovery of contextual word embeddings (see, e.g., Peters et al., 2018), has provided the humanities with powerful analytical tools, yet it also invites a re-evaluation of our methodological frameworks. The present multi-scale analysis points to an important trade-off at the core of computational literary study: while aligning the model’s scale with the feature’s scale might be necessary for maximizing performance and enabling post-hoc explainability, a deliberately misaligned model, despite its lower accuracy, can offer greater interpretability by making its decision points transparent. A distinctive contribution of humanistic benchmarking to literary studies, therefore, might lie less in a monolithic pursuit of accuracy than in articulating, and empirically testing, how different scales of modeling serve different scholarly aims, including prediction, explanation, and interpretation. In this way, the very act of designing a computational model transforms the question of “context” from a silent assumption of disciplinary intimacy into an explicit and contestable methodological choice.

7. Limitations

Several limitations of this project should be acknowledged. First, the established pipeline risks a form of circularity: training student models on teacher-generated labels can propagate teacher biases. The paper does not claim to define a final human gold standard of “regulatedness,” even though the “teacher” model had been previously evaluated on human-labeled data; instead, it claims to benchmark model behavior under nested contexts for a historically theorized feature. In other words, the accuracy of the Couplet Model reflects primarily its ability to distill the teacher’s labeling rule. Moreover, this study operationalizes (Moretti, 2013) parallelism as a binary feature (parallel/not parallel) for the sake of experimental control, which is a necessary simplification that elides the nuanced reality of both poetic practice and critical judgment. Sinological scholarship has long recognized parallelism as a spectrum, encompassing “loose” (kuan dui 寬對) or “flowing” (liushui dui 流水對) parallelism, and even human readers may disagree on borderline cases (Cai, 2008; Yu, 1987).

Furthermore, this project targets semantic parallelism and does not encode tonal templates (ping ze 平仄) or rhyme categories. Since tonal constraints are central to regulated verse, incorporating phonological features could plausibly improve poem-level discrimination by penalizing semantically “plausible” but prosodically invalid poems and providing a more disambiguating signal. At the same time, some of the attention resources otherwise focused on the inner couplets would then be reallocated to other lines to process the ending tokens of all four couplets. This might require yet another modification of classifier architecture or additional computational resources.

Finally, a note on the training process is in order. Due to data sparsity, the Character model’s training proved unstable. This has been mitigated by random seed adjustments in cases where instability has been detected, but the problem suggests that the task of classifying two isolated characters might be ill-posed for a BERT-based architecture, originally designed for longer sequences. Conversely, octave poems with a parallel final couplet are rare in the corpus, which makes class balance difficult to achieve. Consequently, the Poem-4 model learns a simple positional heuristic rather than the actual poetic patterns, achieving deceptively high accuracy simply by learning to predict “0” (not parallel) for the last couplet. This phenomenon, however, further demonstrates the importance of aligning literary features, modeling architectures, and training data.

Additional Files

The datasets accompanying this manuscript can be found at https://doi.org/10.5281/zenodo.18214284.

The GitHub repository is available at https://github.com/mcjkurz/parallelism-benchmark.

Competing Interests

The author has no competing interests to declare.

Author Contributions

Maciej Kurzynski — Conceptualization, Data Curation, Formal Analysis, Investigation, Methodology, Software, Visualization, Writing – Original Draft, Writing – Review & Editing.