Fig. 1 illustrates the masking tendencies of different strategies for different legal terms based on their TF and DF values. Existing augmentation methods often rely on TF-IDF to identify key legal terms. However, case-specific legal terms, e.g., “cancellation” and “inheritance,” tend to appear in only a narrow range of cases, thereby exhibiting low TF-IDF values. Consequently, they are unintentionally treated as unimportant and may be masked during augmentation, degrading classification performance. This mismatch highlights the need for a domain-aware weighting scheme that can distinguish between genuinely irrelevant terms and legally decisive yet sparsely distributed expressions. We contend that masking terms with low TF-IDF values is not necessarily the same as masking unimportant terms in the legal domain.

The x-axis represents document frequency (DF), whereas IDF decreases as DF increases (IDF=log⁡NDF); therefore, the right-hand side of the plot indicates tokens with smaller IDF values, corresponding to more common terms within the corpus. Each marker in Fig. 1 corresponds to the TF-DF coordinate of an individual token, illustrating token-level masking outcomes rather than predefined regions: circles (•) indicating their likelihood to be masked by TF-IDF, squares (◼) denote tokens masked only by the proposed TF-DF strategy, triangles (▴) represent tokens masked by both methods, and diamonds (⧫) mark tokens preserved by both. This clarification distinguishes token-specific masking behavior from the broader conceptual zones described in the rationale, ensuring consistent interpretation of Fig. 1.

The term “Cancellation,” located in the lower-left region of the TF-DF landscape (close to the x-axis, indicating both low term frequency and low document frequency), is essential for identifying cases involving administrative disposition cancellations, and facilitates their distinction from other administrative or tax-related matters.

Fig. 1 reveals a critical drawback of the TF-IDF strategy—many domain-specific and case-specific terms, e.g., “Cancellation,” “Inheritance,” and “Public Official” are located in the lower-left area (low TF and low DF), indicating their likelihood to be masked by TF-IDF. Although they appear infrequently across the corpus, these terms are essential for determining the legal context of specific cases. In contrast, the proposed TF-DF-based masking exhibits improved sensitivity to such legal relevance by preserving these terms. This figure supports the rationale behind the proposed TF-DF-based masking strategy and highlights its difference from existing TF-IDF-based preservation, demonstrating its ability to differentiate legally significant terms from generic ones. This figure substantiates the rationale of the proposed TF-DF approach and highlights its ability to distinguish legally salient tokens from generic ones, ensuring that data augmentation does not distort the core legal reasoning in case texts.

Data preparation in this study follows a systematic multi-stage pipeline designed to transform raw judicial texts into a standardized format suitable for classification, as illustrated in Fig. 2. The dataset is derived from South Korean court rulings, originally provided by the Korean Ministry of Government Legislation via Law Open Data (https://open.law.go.kr, accessed on 18 November 2025 ), and consists of 87,160 legal cases spanning more than seven decades, from 13 January 1952 to 29 February 2024. This long temporal coverage ensures that the dataset captures the evolution of judicial language and legal reasoning in Korea, providing a valuable resource for both historical and contemporary analyses.

Table 1 presents the number of cases for each class along with descriptive statistics, including quartiles, means, and Std of the token lengths per case. In this study, we denote the Intellectual Property Law (IP Property) for brevity. This category encompasses legal disputes concerning patents, trademarks, copyrights, and design rights. A substantial class imbalance is observed: while the largest class (Civil Law) contains 39,830 cases, the smallest category has fewer than 1273 cases, reflecting the uneven distribution of case types in real-world judicial practice and making it challenging for models to learn minority categories effectively.

A primary challenge in preparing this dataset lies in the accurate separation of judicial decisions from legal reasoning, particularly in cases involving complex or multi-layered arguments. To address this, case facts, claims, and judicial decisions were extracted selectively using custom Python scripts, which systematically remove irrelevant metadata such as judge names, court divisions, and procedural annotations. This ensures that the processed text focuses on substantive legal content. The process begins with data extraction from court archives and relevant sources, followed by preprocessing steps that filter out extraneous information and personal identifiers to maintain textual consistency. The cleaned documents are then organized and assigned to their corresponding classification categories, resulting in a corpus that is both structured and legally interpretable.

Given the hybrid nature of the South Korean legal system, which integrates statutory law from civil law traditions with precedent-based reasoning from common law, the dataset requires meticulous structuring [24]. Legal phrase identification is performed to accurately segment case decisions into linguistically meaningful units. This is particularly important in Korean, an agglutinative language where grammatical particles and suffixes convey critical semantic cues. To ensure ethical compliance, personally identifiable information (e.g., party names, addresses, and references to individuals) is systematically removed, thereby safeguarding privacy while preserving the essential content required for downstream NLP tasks.

Unlike existing datasets [25], which primarily focus on criminal and civil law, the proposed dataset encompasses a broader spectrum of legal domains, including family law, intellectual property law, taxation, and administrative law. This broader coverage not only improves the representativeness of the dataset but also enables the training of models capable of handling the diversity of real-world legal cases. Table 2 summarizes the outputs at each stage of the dataset refinement pipeline.

The proposed augmentation method aims to preserve legally important expressions, while introducing corpus-aware variability via probabilistic masking. Instead of relying on uniform or randomly applied masking, we leverage corpus-level statistics to determine the tokens that should be replaced by [MASK] symbols. The full masking algorithm is described in Algorithm 1. The algorithm begins by creating an empty container 𝒟′ to store the augmented corpus, thereby establishing a dedicated repository for masked documents that will be generated in the subsequent steps (Line 1). This initialization step, though seemingly simple, is crucial in ensuring that the augmented data are systematically collected and preserved in a manner that is completely separable from the original corpus, thus preventing unintended data leakage or overwriting.

Next, every legal document in the corpus is tokenized using the KLUE/BERT WordPiece tokenizer, which is designed to segment words into subword units suitable for transformer-based language models (Line 2). This tokenization is not merely a mechanical preprocessing step but an essential foundation for frequency-based analysis, since subword segmentation captures rare and morphologically complex legal terms more effectively than word-level tokenization. Once tokenized, the document frequency df(ω) of each token ω is computed at the corpus level (Line 3). Here, tf(ω) denotes the number of occurrences of ω within a document, and df(ω) the number of documents containing ω. These corpus-level statistics play a pivotal role in guiding subsequent masking decisions, because they reveal how widely distributed each token is across documents rather than just within a single text.

For each individual document d∈𝒟, the algorithm then computes the term frequency tf(ω) for all tokens and assigns an importance score to each token based on a TF-DF formulation: (1)w(ω)=tf(ω)⋅log⁡(1+df(ω)).

The logarithmic smoothing prevents excessively large df values from dominating the weighting scheme, thereby avoiding a situation where frequent but uninformative tokens (e.g., procedural markers such as “submitted,” “record,” or “hearing”) overwhelm more discriminative but less frequent terms. It yields smoother scaling across large corpora and reduces sensitivity to corpus size, maintaining consistent importance estimation across datasets. Adding 1 inside the logarithm ensures numerical stability by avoiding undefined values when df(ω)=0. This design mirrors the stabilizing role of the logarithmic component in TF-IDF while reversing its intent—to emphasize legally meaningful expressions that recur across multiple cases rather than penalizing them.

To ensure comparability and stable scaling across tokens, the scores are subsequently normalized using min–max scaling: (2)w~(ω)=w(ω)−min(w)max(w)−min(w) +ϵ,where ϵ is a small constant introduced for numerical stability. This normalization compresses all weights into the interval [0,1], thus allowing them to be directly interpreted as probabilistic scaling factors for masking (Lines 4–7). Subsequently, for each document d, a copy d′ is created to preserve the original text (Line 8). This duplication ensures that the original legal record remains intact for reference and evaluation, while all augmentation operations are confined to the copy. For each token ω in d′, the algorithm samples a random value r∼𝒰(0,1) and applies the masking rule: if r≤α,w~(ω), where α is a user-defined masking intensity parameter, then the token is replaced with [MASK] (Lines 9–13). This stochastic mechanism introduces controlled randomness into the augmentation process. By linking the masking probability directly to w~(ω), the algorithm biases masking toward less-informative and frequently occurring tokens, while simultaneously lowering the likelihood of masking legally significant expressions such as “trust Asset” or “unjust dismissal.”

A notable advantage of the stochastic masking strategy is its ability to reduce redundancy in the augmented corpus. Deterministic masking would consistently replace the same tokens across documents, resulting in a less diverse dataset [15]. In contrast, the stochastic rule introduces variability between augmented instances, enriching the training distribution with multiple plausible variants of the same document. In this study, we empirically set the masking intensity parameter α=0.2, following prior works [17,20,26] that demonstrated its effectiveness in balancing coverage and diversity in stochastic masking and a sensitivity analysis was performed to validate the stability of this choice.

Once all tokens in a document have been processed under this stochastic masking regime, the resulting augmented document d′ is appended to the container 𝒟′ (Line 14). This process repeats for every document in the input corpus, gradually populating 𝒟′ with augmented versions that maintain the core semantic and legal reasoning of the originals while discarding extraneous information. After all documents have been processed, the fully constructed augmented corpus 𝒟′ is returned as output (Lines 15–16). This corpus serves as the foundation for downstream training, offering a richer and more balanced dataset for classification tasks.

Through stochastic masking guided by TF-DF weighting, the proposed method suppresses peripheral terms such as dates or procedural phrases while preserving legally decisive expressions, thereby operationalizing the algorithmic rationale described above in a concrete and systematic manner. Fig. 3 illustrates the overall workflow of this masking process in greater detail, highlighting how corpus-level statistics are used to determine token salience and guide the replacement of low-importance terms during augmentation.

Furthermore, Table 3 presents illustrative examples of masked outputs, demonstrating that high-frequency but legally irrelevant expressions—such as dates, specific object names, procedural details, or other peripheral descriptors frequently appearing in judicial texts—are effectively suppressed, whereas pivotal legal terms remain intact, thereby preserving the core semantic structure required for correct legal interpretation. These observations highlight the importance of maintaining semantic fidelity in legal text augmentation, as even minor hallucinations can compromise downstream classification and retrieval tasks by subtly altering the factual framing or doctrinal meaning of a case.

Table 4 shows that while AEDA introduces only minor structural noise, DALE fundamentally alters the factual and legal nature of the case itself. Such distortions illustrate why generative augmentation methods are unsuitable for high-fidelity legal datasets, where even subtle lexical substitutions can invert judicial meaning.

For evaluation, the dataset is split into three subsets: 60% for training, 20% for validation, and 20% for testing. To mitigate the imbalance problem, balanced augmentation [27] is applied, adjusting each category to match the size of the largest class rather than simply oversampling smaller ones. Balanced augmentation was applied after generating augmented samples through the TF-DF masking procedure. For each category, the number of cases was adjusted to match the largest class by adding non-redundant masked documents, thereby balancing the training distribution without simple duplication. In addition, balanced augmentation was applied solely to equalize the number of samples across legal categories, without modifying the textual content of any document. This setting inherently isolates the contribution of each augmentation strategy. All augmentations were evaluated under identical conditions using the same balanced dataset and hyperparameter settings. Therefore, any observed performance differences arise solely from the augmentation strategy itself rather than from variations in training data or optimization.

To ensure fair comparison, we evaluated four configurations. The ‘No Augmentation’ baseline denotes training on the original unaltered dataset without any augmentation, serving as a reference for evaluating the contribution of each augmentation method. And a POS Deletion baseline that randomly removes part-of-speech–based tokens, a TF-IDF Masking baseline that masks tokens according to their inverse document frequency, and the proposed TF-DF Masking method. This strategy ensures that underrepresented categories are not disproportionately overlooked. Fig. 4 illustrates the distributions before and after augmentation.

Transformer-based models fine-tuned for legal case classification are considered. The models are trained using an AdamW optimizer with hyperparameters β1=0.9; β2=0.999; weight decay = 0.01; initial learning rate = 1×10−5; adjusted via a linear scheduler; the number of epoch = 20; and batch size = 16. Each experiment is repeated 10 times to ensure robust and reliable evaluation. Given the four basic statistics, true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN), for classification tasks, we evaluate model performance by calculating accuracy and F1 score, which are two widely used metrics in text classification tasks. Accuracy measures the proportion of correctly classified cases relative to all legal categories and is defined as follows: (3)Accuracy=TP+TNTP+TN+FP+FN,where TP, TN, FP, and FN denote true positives, true negatives, false positives, and false negatives, respectively.

The F1 score, which represents the harmonic mean of precision (TPTP+FP) and recall (TPTP+FN), is defined as (4)F1=2×Precision×RecallPrecision+Recall.

This metric provides a balanced evaluation of classification performance by combining both precision and recall.

We evaluated the effectiveness of the proposed TF-DF–based masking augmentation method using two representative models, BERT [1] and LegalBERT [2]. Tables 5 and 6 summarize the results in terms of accuracy and macro F1. Paired t-tests conducted on accuracy across ten runs confirmed statistically significant improvements for all models (p<0.01). Specifically, LegalBERT (p≈0.007), and BERT (p<0.0001) all showed meaningful accuracy gains following TF-DF masking. For LegalBERT, the proposed method achieved an accuracy of 0.8551, outperforming TF-IDF (0.8384), POS-based deletion (0.7775), and no augmentation (0.7436). Similarly, the macro F1 score reached 0.8453, showing a clear improvement over TF-IDF (0.7916), POS (0.6913), and no augmentation (0.5952). For BERT, the proposed method also gave the best performance, with an accuracy of 0.9704 and a macro F1 score of 0.9640. The proposed augmentation outperformed TF-IDF and POS-based methods across both models.

The macro F1 improvements observed in Table 6 are primarily attributable to category-level gains highlighted in Table 7. Most notably, the Administrative Law category showed substantial improvement with the proposed method (0.9013 ± 0.0100) compared to TF-IDF (0.8171 ± 0.0030). Given that Administrative Law represents a large portion of the dataset, this gain had an outsized effect on the overall macro F1. In addition, the Family Law category—characterized by cultural specificity and nuanced linguistic expressions—benefited from TF-DF masking, improving from 0.7795 ± 0.0390 (TF-IDF) to 0.8990 ± 0.0076. This demonstrates that the proposed method effectively preserves contextually critical tokens (e.g., kinship or familial roles) that are often decisive in classification but may be indiscriminately masked under TF-IDF. Similarly, the Taxation category also exhibited meaningful gains (0.8936 ± 0.0149 → 0.9384 ± 0.0047).

Table 8 presents a qualitative comparison of classification results obtained using different masking strategies. This comparison illustrates how the proposed strategy enhances classification reliability by selectively masking peripheral expressions while retaining critical legal terms, thus maintaining semantic fidelity in the augmented corpus. Each block of three rows corresponds to one case: the original judgment text, the version augmented with the proposed TF-DF masking, and the version augmented with TF-IDF masking. In the “Sentence” column, tokens surrounded by brackets (e.g., [word]) indicate the terms that would have been masked during augmentation according to each strategy. The last column reports the ground-truth label and the prediction produced by LegalBERT when trained with the corresponding input. As shown in Cases 1, 3, the proposed masking strategy preserves legally decisive terms such as “seizure invalid”, “gift tax”, enabling the model to predict the correct category. In contrast, TF-IDF masking often removes these essential expressions, which leads to semantic drift and incorrect predictions (e.g., misclassifying taxation as civil law or family law). When TF-IDF masking eliminates key tax-related legal terms such as ‘tax’, ‘seizure’, and ‘trust Asset’, the decisive linguistic markers necessary to situate the dispute within the domain of taxation law are lost. With the fiscal and administrative context removed, the residual sentence can be interpreted merely as a conflict concerning property ownership or possession. Consequently, the model fails to recognize the case as a taxation dispute and instead misclassifies it as a matter falling within the scope of general civil law, particularly property rights disputes.

By contrast, the proposed masking strategy preserves critical legal expressions such as ‘seizure’, ‘trust Asset’, and ‘invalid’, thereby maintaining the fiscal and administrative character of the dispute. Even though some peripheral expressions are masked, the presence of these domain-specific tokens enables the model to correctly identify the case as involving the validity of a tax levy, specifically the annulment of a taxation disposition. Taken together, the substantial improvement in Administrative Law, along with the enhanced handling of culturally specific Family Law cases and the robust recognition of domain-critical terms in Taxation, collectively explain the consistent macro F1 gains of the proposed method across both LegalBERT and BERT.

To assess parameter robustness, we varied the masking intensity parameter α across four levels: 0.05, 0.1, 0.2, and 0.3. As shown in Fig. 5, the F1 scores remained highly consistent across settings, demonstrating that the proposed TF-DF masking is insensitive to small perturbations in α. Both BERT and LegalBERT achieved their highest performance at α=0.2, suggesting that a moderate masking intensity provides an optimal balance between lexical diversity and semantic retention.