SSPA: Enhancing Pseudo-Corpus Quality on Tibetan Machine Translation via Semantic-Syntax Prealignment

Abstract

Tibetan-to-English machine translation (MT) models frequently falter under extreme domain data scarcity, often producing translations that violate the distinctive agglutinative rules of Tibetan and suffer from domain-specific stylistic mismatches. To overcome these limitations, we propose Semantic-Syntax Prealignment (SSPA), an innovative corpus generation framework. SSPA constructs high-quality pseudo-parallel pairs by explicitly minimizing the deviation between the syntactic-semantic profiles of generated samples and professional reference texts. Specifically, source-target structural representations are standardized through length-unified truncation and terminology normalization, followed by a dual-domain alignment process that maximizes syntactic cosine similarity under rigorous structural constraints. We further augment these aligned frames via a cross-length dynamic filling mechanism, which is integrated with an Expectation-over-Transformation (EOT)-based style regularization mechanism specifically adapted for stylistic perturbations, to simulate authentic linguistic variations. Extensive evaluations on our newly constructed Tibetan Medicine-Tibetan English (TM-TE) dataset demonstrate that SSPA significantly outperforms existing competitive baselines. Notably, SSPA achieves a BLEU-4 score of 36.2 and improves long-sentence BLEU-4 by 16.8 points, with a parser-verified grammatical compliance rate of 96.2%. The framework exhibits remarkable cross-domain adaptability and stylistic consistency, offering a robust, versatile solution for low-resource Tibetan professional domain MT.

1. Introduction

Tibetan-to-English machine translation (MT), which enables automatic conversion between different languages through computer technology, has become an indispensable tool for cross-lingual information exchange and international affairs management. It provides critical technical support for both individuals facing language barriers and professional translators working with Tibetan content [1,2,3,4]. Despite its widespread applications, Tibetan-to-English MT presents unique challenges due to the language’s agglutinative nature and the scarcity of high-quality annotated data in specialized domains [5,6,7].

Early Tibetan-to-English MT systems relied on a rule-based, human-driven paradigm [8,9]. The core of these systems consisted of grammatical rule bases and bilingual dictionaries meticulously constructed by linguists. Translations were generated by parsing the syntactic, semantic, and morphological features of the source Tibetan text [9,10]. However, the inherent complexity of natural language and pervasive lexical ambiguity required the definition of an extensive number of grammatical rules [11]. Furthermore, these systems exhibited poor generalization capabilities when encountering atypical language structures, severely limiting their practical utility.

The advancement of computing power led to the adoption of statistical machine translation (SMT) techniques, which addressed the scalability limitations of rule-based approaches [12]. SMT methods learn phrase alignment probabilities and language model parameters directly from large-scale bilingual parallel corpora, enabling data-driven automatic translation modeling [13,14]. This paradigm achieved remarkable progress in data-rich domains such as general news [15]. Nevertheless, SMT systems only capture statistical regularities of surface-level linguistic features. They fail to effectively model the intricate case marker concatenation rules and long-distance semantic dependencies characteristic of Tibetan as an agglutinative language [16]. Additionally, SMT is highly sensitive to data sparsity [17], resulting in significantly reduced translation performance in low-resource professional domains like Tibetan medicine (see

Table 1. Case study of machine translation outputs by the Semantic-Syntax Prealignment (SSPA) approach in the low-resource Tibetan medicine domain.

Method	Content/Translation Output
Example 1
Tibetan Source	རླུང་ནད་ཀྱིས་མགོ་འཁོར་ཞིང་ལུས་ཤུགས་ཉམས།
Transformer	The wind disease causes dizziness and weakness.
General Back-Translation	Wind disorder leads to dizziness and fatigue.
SSPA (Ours)	Rlung imbalance manifests as vertigo and constitutional lassitude.
Example 2
Tibetan Source	སྨན་འདི་མཁྲིས་པའི་ནད་དང་ཕོ་བའི་ནད་ལ་ཕན་པ་ཡོད།
Transformer	This medicine is good for bile disease and stomach disease.
General Back-Translation	This remedy benefits bile disorders and stomach illnesses.
SSPA (Ours)	This decoction is indicated for tripa imbalance and gastric disorders.
Example 3
Tibetan Source	སྨན་པས་སྨན་རྫས་བཞི་བསྡེབས་ནས་སྨན་བཟོ་བ།
Transformer	The doctor uses four herbs to make medicine.
General Back-Translation	Physicians prepare medicine with four medicinal herbs.
SSPA (Ours)	Practitioners compound the formulation using four herbal ingredients.

for a comparative case study).

In recent years, end-to-end deep learning-based modeling approaches have emerged as the primary technical pathway for Tibetan MT [18], particularly for improving long sentence translation [19]. These methods employ neural network architectures such as Transformer [20], which can automatically learn complex cross-lingual mapping relationships from extensive bilingual parallel corpora [21]. Deep learning models have effectively overcome the core bottlenecks of SMT, including labor-intensive feature engineering and inadequate long-distance semantic modeling [22,23]. However, their performance is strongly correlated with the size and quality of training data. When annotated corpora in the target domain are limited, models cannot sufficiently learn domain-specific linguistic features and expression paradigms, leading to significant performance degradation [24].

The primary challenge confronting contemporary Tibetan-to-English machine translation is the severe imbalance in domain data distribution [25]. This issue imposes substantial limitations on both statistical and deep learning approaches [26]. Specifically, the absence of annotated samples in niche professional domains, such as Tibetan medicine, restricts the extraction of comprehensive phrase features for statistical models [27,28]. Conversely, deep learning methods typically overfit to data-rich mainstream domains like news and current affairs [29]. Consequently, these neural models struggle to capture the distinctive linguistic patterns and stylistic characteristics of specialized domains [30]. This discrepancy ultimately manifests as competitive translation performance in general domains, yet yields high terminology error rates and stylistic mismatches in professional contexts [31].

Existing mainstream solutions to the domain data imbalance problem fall into two broad categories: data augmentation [32] and model transfer [33]. Data augmentation techniques, such as back translation, generate weakly labeled parallel corpora from domain-specific unlabeled monolingual data, providing a cost-effective means to supplement scarce domain samples [34]. However, these conventional methods are heavily prone to introducing syntactic noise [35]. Model transfer approaches pre-train a base model on general corpora before adapting it to the target domain via parameter-efficient fine-tuning [36]. Specifically, these techniques aim to mitigate catastrophic forgetting while acquiring domain-specific knowledge. However, they frequently experience severe overfitting in extreme low-resource scenarios [37]. Furthermore, feature alignment methods typically require modifying the model architecture or extracting features from specific hidden layers. Consequently, this architectural coupling introduces high computational costs and restricts generalization capabilities across diverse translation frameworks [38].

However, current solutions generally neglect two critical aspects of Tibetan language processing: the strict case marker concatenation rules inherent to its agglutinative structure and the significant stylistic differences across domains. Tibetan texts in specialized fields like Tibetan medicine are characterized by concise sentence structures, standardized expressions, and dense professional terminology—features that distinguish them sharply from general domain texts. Because Tibetan is highly agglutinative, traditional data augmentation techniques—such as random term replacement and simple back translation—often leave case markers detached. Consequently, these methods disrupt fundamental grammatical bindings and introduce substantial syntactic noise. This ultimately yields generated pseudo-corpora with numerous grammatical errors and domain style mismatches, rendering them largely ineffective for low-resource translation enhancements.

Crucially, this paper redefines the measurement benchmark for cross-domain generalization in low-resource and morphologically complex Tibetan-to-English MT tasks. Rather than pursuing peak scores that heavily overfit to general corpora, we advocate for a “zero-forgetting generalization” paradigm. Specifically, under extreme domain shifts, an effective model should achieve substantial improvements in the target professional domain. Furthermore, it should minimize performance degradation in the general domain, thereby demonstrating stable cross-domain variance control.

To address the machine translation model degradation caused by data imbalance in low-resource Tibetan domains, this paper presents a novel pseudo-parallel sentence pair generation framework: Semantic-Syntax Prealignment (SSPA). In contrast to existing benchmark data augmentation methods for Tibetan domains, SSPA integrates dual constraints of syntactic structure and multi-scale semantics, significantly enhancing both the grammatical correctness and domain semantic relevance of generated pseudo-corpora.

The proposed framework operates in two distinct stages. In the first stage, we design a dual-domain alignment initialization strategy specifically tailored to the agglutinative characteristics of Tibetan. We define Tibetan case marker concatenation and syntactic dependency relations as syntactic domain features, achieving syntactic structure alignment through the maximization of syntactic cosine similarity. Concurrently, we map textual features into a semantic domain space. Here, domain vocabulary and contextual features are aligned by minimizing two proposed objectives, a multi-scale semantic mean squared error loss and a relative semantic convergence loss, aiming to penalize rare terminology deviations. This alignment process promotes the generated pseudo-corpora to adhere to the semantic distribution and lexical conventions of professional texts.

In the second stage, leveraging our constructed Tibetan medicine domain syntactic framework and bilingual terminology dictionary, we perform dynamic term filling and pseudo-sentence reconstruction under the constraints of semantic-syntactic dual-domain prealignment. We further introduce an Expectation-over-Transformation (EOT) style regularization mechanism to simulate writing variations across different Tibetan medicine documents, producing high-quality pseudo-parallel Tibetan-to-English sentence pairs that conform to professional stylistic norms. Through this comprehensive process, our method generates parallel corpora for low-resource Tibetan domains that are both grammatically rigorous and semantically coherent.

The main contributions of this paper are threefold:

1.

We propose SSPA, a Tibetan low-resource pseudo-corpus generation framework based on semantic-syntactic dual-domain prealignment. By jointly optimizing syntactic cosine similarity loss and multi-scale semantic loss, SSPA substantially mitigates the core issues of grammatical errors and semantic misalignment prevalent in traditional data augmentation methods.

2.

A dynamically adaptive filling mechanism designed for cross-length domain syntactic frameworks and an EOT-style regularization module are developed. These components enable the batch generation of pseudo-parallel corpora covering diverse sub-domains using only a small set of annotated Tibetan medicine domain syntactic templates and terminology dictionaries.

3.

While the data generation utilizes optimization feedback from the translation model via a differentiable Gumbel-Softmax bottleneck, the SSPA framework remains architecturally decoupled. Specifically, it avoids the necessity to modify the translation model’s internal network topology or introduce additional inference overhead. This makes our approach a versatile data augmentation solution applicable to various low-resource Tibetan translation tasks.

To thoroughly evaluate our method in the targeted low-resource scenario, extensive experiments on our newly constructed Tibetan Medicine-Tibetan English (TM-TE) bilingual dataset demonstrate that SSPA achieves a remarkable Bilingual Evaluation Understudy-4 (BLEU-4) score of 36.2, outperforming existing competitive baselines. Notably, it improves long-sentence BLEU-4 by an absolute margin of 16.8 points compared to the baseline Transformer model, while yielding pseudo-corpora with a parser-verified grammatical compliance rate of 96.2%.

The remainder of this paper is organized as follows. Section 2 reviews related work on low-resource machine translation and data augmentation. Section 3 details the proposed SSPA framework. Section 4 presents the experimental setup, results, and comprehensive analyses. Finally, Section 5 concludes the paper.

2. Related Work

2.1. Neural Machine Translation in Low-Resource Scenarios

Low-resource neural machine translation mitigates performance degradation in scenarios with insufficient labeled parallel corpora for the target language [10,35]. Mainstream approaches generally fall into two categories: data augmentation and model transfer [24,33]. Data augmentation methods expand training data at a low cost by generating pseudo-parallel corpora through techniques such as back translation, random word operations, or hybrid augmentation [17,34]. Among these, back translation is widely adopted due to its implementation simplicity and stable performance. Building upon the initial demonstration by Sennrich et al. [39] that monolingual back translation significantly improves low-resource translation, subsequent studies introduced iterative and multilingual variants to more effectively leverage unlabeled data [40,41].

Alternatively, model transfer approaches acquire general linguistic knowledge through cross-lingual pre-trained language models (e.g., mBART [42], XLM-R [43]). These models are then adapted to specific target domains using parameter-efficient fine-tuning techniques (e.g., Adapter [44], LoRA [45], Prefix-Tuning [46]). This strategy often yields substantial improvements requiring only minimal domain-labeled data [24,33].

Despite achieving remarkable results in Indo-European language tasks, these general low-resource methods encounter notable adaptation challenges when applied to Tibetan [10,12]. The primary challenge arises from the inherent conflict between the design assumptions of existing methods and the unique agglutinative characteristics of Tibetan [1,3]. For instance, general data augmentation fails to account for the strict binding rules between a case marker and its head word. Consequently, random word operations often disrupt grammatical relations, producing pseudo-corpora with prevalent syntactic errors [4,47].

Furthermore, parameter-efficient fine-tuning techniques exhibit limited effectiveness in extremely low-resource scenarios like Tibetan medicine. These techniques are susceptible to overfitting when target samples are scarce, and they struggle to adequately bridge the massive domain gap without sufficient in-domain data [15,32,36]. Because general methods often fall short in specialized Tibetan domains, developing domain-specific data augmentation solutions that respect both linguistic characteristics and professional stylistic features remains a critical research direction [18,25].

2.2. Data Augmentation and Pseudo-Corpus Generation

Pseudo-corpus generation serves as a primary technique for expanding training data in low-resource Natural Language Processing (NLP) tasks [17,34]. Existing approaches generally follow three major paradigms: rule-based, generative model-based, and feature alignment-based methodologies. Rule-based methods utilize manually crafted templates and rewriting rules. While they offer high grammatical correctness and strong domain controllability, they typically exhibit poor generalization, require extensive human intervention, and struggle to capture complex linguistic phenomena [10,13]. In contrast, generative model-based approaches leverage large language models, such as GPT [48] and T5 [49], to produce pseudo-texts in batches via conditional generation. Although effective in general domains, these models depend heavily on domain-specific fine-tuning data and frequently introduce factual errors or semantic deviations in low-resource settings [11,50].

To mitigate these semantic deviations, feature alignment-based methods generate pseudo-corpora by minimizing the feature distance between source and target texts within intermediate model layers [38,51]. These techniques effectively preserve the semantic information of the source text. However, they are deeply coupled with specific model architectures, incur substantial computational overhead, and lack cross-model transferability.

Beyond architectural limitations, applying these general pseudo-corpus generation techniques to specialized domains like Tibetan medicine reveals difficulties in modeling unique stylistic characteristics [15,27]. Tibetan medical literature features precise vocabulary, parallel sentence structures, and the frequent use of tetrasyllabic terminology [5]. Because standard generation methods typically overlook these nuances, the resulting pseudo-corpora often diverge significantly from authentic medical texts. Translation models trained on such data may memorize domain terminology, yet they tend to produce rigid translations that deviate from professional expression norms [11], ultimately limiting improvements in evaluation metrics such as BLEU scores.

To bridge these research gaps, this paper proposes the Semantic-Syntax Prealignment (SSPA) framework. By shifting from internal feature alignment to an architecturally decoupled optimization paradigm—wherein the translation model acts as a frozen, differentiable critic guiding textual perturbations—SSPA provides a novel approach for low-resource Tibetan medicine machine translation. While decoupled feature alignment has shown promise in broader NLP areas like cross-lingual pre-training, its application in resolving the specific agglutinative syntax and stylistic complexities of Tibetan remains largely unexplored. These foundational concepts directly inform the architectural design of SSPA, which is elaborated in Section 3.

3. Proposed Method

3.1. Framework Overview

The overall framework of the Semantic-Syntax Prealignment (SSPA) algorithm is illustrated in Figure 1. Alongside the general Tibetan-English sentence pairs acquired from the news domain, a Tibetan medicine terminology dictionary and domain-specific syntactic frame templates—extracted from classical medical texts—are independently fed into the preprocessing network. To minimize the loss of domain-specific knowledge during feature transformation, this network applies distinct processing strategies tailored to the linguistic characteristics of the input data. Given the significant syntactic disparities between general texts and professional Tibetan medicine expressions, a Tibetan-specific word segmentation and case marker annotation component is integrated into the preprocessing network. By applying Bidirectional Long Short-Term Memory (Bi-LSTM) and Conditional Random Field (CRF) layers via a stacked sequence labeling approach, this component enables the network to accurately delineate the grammatical boundaries inherent to the Tibetan agglutinative language. Crucially, the pre-training and domain adaptation of the Bi-LSTM-CRF parser component strictly relied only on general-domain external data and the 400 TM-TE training pairs. This ensures zero exposure to validation and test data distributions, substantially reducing the possibility of test set distribution leakage. Following this preprocessing stage, both the original sentence pairs and domain resources are uniformly converted into standardized syntactic dependency vectors.

Directly adopting random term replacement or blind augmentation is suboptimal for generating pseudo-corpora, as the strict case marker concatenation rules in Tibetan make random modifications highly susceptible to syntactic disruption. Furthermore, traditional methods struggle to capture the unique syntactic style of Tibetan medicine literature, resulting in severe domain style misalignment. Consequently, we abandon the unguided random initialization strategy and propose a semantic-syntactic dual-domain prealignment initialization method. To maximize the utilization efficiency of domain knowledge, a dual-domain joint loss function is devised. This function simultaneously evaluates the syntactic cosine similarity and the multi-scale semantic mean squared error (MSE) between the generated pseudo-sentences and the target reference texts, outputting the initial pseudo-sentences optimized via dual-domain alignment. This fundamentally mitigates grammatical errors and semantic inconsistencies, thereby boosting the translation accuracy of domain-specific rare terms.

Another critical challenge is the correct conversion of limited domain syntactic frames into large-scale, highly generalizable pseudo-corpora. To prevent structural confusion during target feature fusion, cross-length dynamic adaptation is first performed on the generic syntactic frames. During this stage, frames are dynamically padded or segmented according to the length of the terms to be filled, and the concatenation forms of case markers are automatically adjusted to guarantee grammatical correctness. Subsequently, Expectation-over-Transformation (EOT)-based contextual robust regularization is introduced to fine-tune the syntactic frames, using Tibetan writing style transformations as the optimization criterion. Batch generation of the pseudo-corpus is then executed by cyclically reusing the terminology dictionary while accounting for writing variations across sub-domains. This robust pipeline efficiently generates diverse bilingual sentence pairs, models domain linguistic patterns across various contexts, and ensures the cross-dataset generalization capabilities of the pseudo-corpus. To ensure strict semantic alignment between the source and target texts, SSPA strictly avoids modifying only the Tibetan side while retaining the original general-domain English side, nor does it rely on external translation models to blindly generate English text. Instead, the framework introduces a bilingual synchronous mapping mechanism. All Tibetan-side term replacements, syntactic structure fine-tuning, and feature augmentation operations trigger a homomorphic synchronous updating mechanism on the English side. This process relies on highly deterministic, strictly aligned bilingual domain resources, ensuring that the generated pseudo-sentences maintain strict bilingual alignment in both fine-grained semantics and macro-syntactic mapping throughout the entire procedure. Ultimately, a high-quality pseudo-parallel corpus tailored for the Tibetan medicine domain is produced. The detailed implementations of the initialization and batch generation stages are elaborated in Section 3.2 and Section 3.3, respectively.

3.2. Semantic-Syntax Prealignment

In this section, an efficient initialization framework for pseudo-parallel sample construction is introduced. In conventional low-resource data augmentation schemes, random term replacement or blind augmentation on original texts serves as the standard initialization method. However, this approach often forces the textual modifications into non-optimal regions of the high-dimensional linguistic feature space, prompting severe optimization challenges such as grammatical fractures and domain semantic misalignment.

To overcome these limitations, SSPA discards the random initialization scheme, establishing domain linguistic anchors as the optimization benchmark. Crucially, this design positions SSPA as an initialization strategy driven by prior linguistic features rather than a conventional data augmentation objective. Previous methods relying on intermediate representation alignment require feature distance minimization on specific hidden layers, incurring high computational overhead and complicating the selection of optimal alignment layers.

In contrast, while SSPA relies on the translation model’s gradients for discrete text generation, its structural components remain architecturally decoupled. Specifically, SSPA constructs initial pseudo-sentences through simultaneous prealignment in both syntactic and semantic domains. By explicitly penalizing the feature deviation between generated samples and target texts, the convergence of pseudo-parallel corpus construction is significantly accelerated in low-resource scenarios.

As illustrated in Figure 1, standardization preprocessing is initially applied to the general sentences and target domain texts. It is important to emphasize that the target reference texts utilized during this semantic-syntactic prealignment process were exclusively sampled from the designated 400 training sentences. This strictly prohibits any target-side information leakage from the evaluation sets. Sentence length unification is achieved via truncation and zero-padding, concurrently with multiple types of terminology normalization, which can be formalized as:

$$d_{std}=\mathcal{N}\left(d_{raw}\right)$$

(1)

where $\mathcal{N}(·)$ is the terminology normalization mapping function, $d_{raw}$ denotes any writing variant of the original terminology (including variant characters, spelling variations, and transliteration differences), and $d_{std}$ represents the corresponding standard terminology form. This preprocessing eliminates invalid offsets caused by term spelling variations, preventing the optimization from merely rewriting word forms while failing to mine valid syntactic transformation rules.

Following text preprocessing, syntactic vector construction filters out unnecessary redundant information. Given the grammatical characteristics of the Tibetan agglutinative language, word segmentation and case marker annotation are executed via a combined Bi-LSTM-CRF model, labeling each word as a binary tuple $(w_i,c_i)$. The resulting standardized syntactic dependency vector is expressed as:

$$\tilde{x}=[(w_1,c_1),(w_2,c_2),\dots ,(w_L,c_L)]\in {\mathbb{R}}^{L\times (d_w+d_c)}$$

(2)

where $w_i$ denotes the word embedding of the i-th word, $c_i$ represents the one-hot encoding of the case marker type corresponding to the i-th word, L is the unified sentence length, $d_w$ is the dimension of the word embedding, and $d_c$ is the total number of case marker types. To minimize the structural discrepancy between pseudo-sentences and target texts, the syntactic cosine similarity ${\mathcal{L}}_{syn}$ is maximized (equivalent to minimizing $1-{\mathcal{L}}_{syn}$). This metric is formulated as:

$${\mathcal{L}}_{syn}({\tilde{x}}_{adv},\tilde{t})={\displaystyle \frac{{\tilde{x}}_{adv}·\tilde{t}}{\parallel {\tilde{x}}_{adv}{\parallel }_2{\parallel \tilde{t}\parallel }_2}}$$

(3)

where ${\tilde{x}}_{adv}$ is the syntactic dependency vector of the pseudo-Tibetan sentence after term modification, $\tilde{t}$ is the syntactic dependency vector of the target Tibetan medicine sentence, · denotes the vector dot product operation, and ${\parallel ·\parallel }_2$ represents the standard L2 norm of the vector.

To enhance the contextual robustness of the generated samples across fine-grained and coarse-grained semantic scales, a random term masking strategy generates multiple augmented feature subsets. This masking operation is formalized as:

$$Wor{d}_{aug}\left(i\right)=\left\{\begin{array}{c}0\mathrm{with}\mathrm{probability}p\hfill \\ Wor{d}_x\left(i\right)\mathrm{otherwise}\hfill \end{array}\right.$$

(4)

where $Wor{d}_{aug}\left(i\right)$ denotes the word embedding of the i-th term after masking, $Wor{d}_x\left(i\right)$ represents the word embedding of the original i-th term, and $p\in [0.1,0.3]$ is the masking probability. By aligning these augmented variants with the target texts, the optimization process is prevented from overfitting to a single syntactic pattern, achieving robust generalization across diverse writing scenarios.

Post-augmentation, the optimization prioritizes the minimization of the multi-scale semantic mean squared error and the relative semantic convergence loss, formulated as:

$$\begin{array}{cc}\hfill {\mathcal{L}}_{MSE}& =\sum _{s\in S}{\displaystyle \frac{1}{N_s}}{∥P_s\left(Word\left({\tilde{x}}_{adv}\right)\right)-P_s\left(Word\left(\tilde{t}\right)\right)∥}_F^2\hfill \\ \hfill {\mathcal{L}}_{SC}& ={\displaystyle \frac{{∥Word\left(\tilde{t}\right)-Word\left({\tilde{x}}_{adv}\right)∥}_2}{{∥Word\left({\tilde{x}}_{adv}\right)∥}_2+\rho }}\hfill \end{array}$$

(5)

where ${\mathcal{L}}_{MSE}$ denotes the multi-scale semantic mean squared error loss, ${\mathcal{L}}_{SC}$ represents the semantic convergence loss, $P_s$ indicates the average pooling operation at scale s, $S=\{1,2,4,8\}$ represents the pooling scales corresponding to words, bigrams, phrases, and whole sentences, $Word(·)$ denotes the word embedding sequence matrix, $N_s$ represents the number of feature vectors after pooling at scale s, ${\parallel ·\parallel }_F$ indicates the Frobenius norm, ${\parallel ·\parallel }_2$ represents the L2 norm, and $\rho ={10}^{-8}$ is a tiny constant ensuring numerical stability.

The total semantic loss is defined as the adaptively weighted sum of these two components:

$${\mathcal{L}}_{Sem}={\mathcal{L}}_{MSE}+\lambda {\mathcal{L}}_{SC}$$

(6)

where ${\mathcal{L}}_{Sem}$ is the total semantic loss, and $\lambda$ is the dynamic adaptive weight. The ratio-based construction of ${\mathcal{L}}_{SC}$ adaptively amplifies the penalty weights for low-frequency professional terms, forcing the optimization to finely align full-dimensional semantic details and ensuring domain fidelity.

The overarching initialization loss combines the syntactic and semantic losses:

$${\mathcal{L}}_{init}=\beta {\mathcal{L}}_{Sem}+\alpha (1-{\mathcal{L}}_{syn})$$

(7)

where ${\mathcal{L}}_{init}$ is the total loss, $\alpha$ controls the global structure matching weight, and $\beta$ dictates the fine-grained feature matching importance.

Minimizing this loss via gradient descent iteratively updates the term replacements, with the rule given by:

$${\delta }_{t+1}={\delta }_t-\eta {\nabla }_{{\delta }_t}{\mathcal{L}}_{init}$$

(8)

where ${\delta }_t$ is the modification increment at the t-th iteration, $\eta$ is the learning rate, and ${\nabla }_{{\delta }_t}{\mathcal{L}}_{init}$ is the gradient of the total loss.

Simultaneously, the modification magnitude is strictly bounded within the $l_0$ norm constraint:

$${\parallel \delta \parallel }_0\le ϵ$$

(9)

where ${\parallel ·\parallel }_0$ denotes the $l_0$ norm (number of non-zero elements), and $ϵ=2$ is the maximum allowed term modifications, ensuring grammatical fidelity. Algorithm 1 summarizes this initialization pipeline.

Ultimately, this pipeline establishes a high-quality optimization baseline for large-scale pseudo-corpus generation, accelerating end-to-end convergence while preserving both Tibetan grammatical norms and domain semantics.

Algorithm 1 Single-Sample SSPA Initialization Algorithm

Require: Original general Tibetan-English sentence x, target Tibetan medicine reference text y, maximum iteration rounds T, single-term modification constraint $ϵ$, learning rate $\eta$, static global word embedding matrix $\mathbf{E}$ Ensure: Pseudo-sentence modification increment $\delta$ optimized via dual-domain alignment

1: $\delta \leftarrow \varnothing$;   2: $\tilde{t}\leftarrow \mathrm{Preprocessing}\mathrm{Standardization} \left(y\right)$;   3: $Wor{d}_y\leftarrow \mathrm{Word}\mathrm{Embedding}\mathrm{Mapping} \left(\tilde{t}\right)$;   4: for  $i=1$  T  do   5:     $\delta \leftarrow \mathrm{TopK}\_\mathrm{Mask} (\delta ,ϵ)$ {Keep the top $ϵ$ elements with largest magnitudes, zeroing others to satisfy Equation (9)}   6:     ${\tilde{x}}_{adv}=\mathrm{Embed}\left(x\right)+\delta$;   7:     $z={\displaystyle \frac{{\tilde{x}}_{adv}·{\mathbf{E}}^T}{\sqrt{d}}}$ {Vocabulary projection onto static word embedding matrix}   8:     $M_{valid}\leftarrow \mathrm{ContextualValidityMask}\left(x\right)$ {Dynamically calculate grammatical validity mask for suffixes}   9:     ${ϵ}_{tie}\sim U(0,{10}^{-7})$ {Inject minimal uniform noise perturbation to resolve truncation ties} 10:    $x_{discrete}\leftarrow \mathrm{DictionaryFallback}\left(\arg \max (z+M_{valid}+{ϵ}_{tie})\right)$ {Extract valid discrete token sequence} 11:    ${\tilde{x}}_{text}\leftarrow \mathrm{Preprocessing}\mathrm{Standardization} \left(x_{discrete}\right)$; 12:    $Wor{d}_x\leftarrow \mathrm{Word}\mathrm{Embedding}\mathrm{Mapping} \left({\tilde{x}}_{text}\right)$; 13:    $gra{d}_{sum}\leftarrow 0$; 14:    Initialize term augmentation feature set; 15:    for $k=1$ to AugmentationRounds do 16:        Apply random masking to word embeddings to generate augmented features $Wor{d}_{aug}$; 17:        Store augmented features into the feature set; 18:    end for 19:    for each augmented feature in the set do 20:        Calculate the overall loss by substituting into Equation (7); 21:        Compute the gradient of the loss with respect to $\delta$ and accumulate; 22:    end for 23:    Compute the average gradient; 24:    Update $\delta$ using the average gradient according to Equation (8); 25: end for 26: return  $\delta$

3.3. Robust Batch Pseudo-Corpus Generation

To scale the SSPA framework for large-scale Tibetan medicine pseudo-parallel corpus generation, unified filling rules are formulated based on a concise set of well-formed syntactic frames. A fixed syntactic template is deployed to accommodate diverse general-domain sentences across various term combinations, thereby unifying the cross-sample generation logic.

Under this paradigm, domain syntactic frames serve as the universal structural benchmark. Utilizing the standardized terminology set and prealigned syntactic frames f derived in Section 3.2, a syntax-oriented iterative optimization expands the samples at scale. This ensures a single template fulfills the generative demands of diverse inputs. The initial template set $X_{tpl}$ is obtained via cross-length adaptation, detailed in Section 3.3.1. Furthermore, both the syntactic frameworks and the batch generation process are explicitly bounded by the linguistic properties of the training corpus. The validation and test sets remain strictly isolated from this synthesis loop, establishing a unidirectional data flow.

3.3.1. Dynamic Syntactic Length Adaptation

Each syntactic frame is formalized as an ordered sequence of fixed syntactic components and term placeholders:

$$f=[s_1,p_1,s_2,p_2,\dots ,s_k,p_m,s_{k+1}]$$

(10)

where $f\in \mathcal{F}$ denotes an individual syntactic frame, $\mathcal{F}$ represents the set of all domain frames; $s_i$ indicates a fixed component (e.g., unmodifiable Tibetan case markers and copulas); $p_j$ is the j-th term placeholder; and m is the total number of placeholders.

As illustrated in Figure 2, a dynamic adaptation mechanism is designed to handle the variable-length spans of Tibetan texts. Let $T=[t_1,t_2,\dots ,t_n]$ represent the set of terms to be filled. If the combined term length is shorter than the available placeholders ($n\le m$), a targeted padding operation is executed:

$$T^{\prime }=[t_1,t_2,\dots ,t_n,\langle \mathrm{PAD}\rangle ,\langle \mathrm{PAD}\rangle ,\dots ,\langle \mathrm{PAD}\rangle ]$$

(11)

where $T^{\prime }$ denotes the adapted term set, and $\langle \mathrm{PAD}\rangle$ represents the empty placeholder. To preserve the core grammatical backbone, these $\langle \mathrm{PAD}\rangle$ symbols are mapped strictly to optional modifier slots.

Conversely, when the term sequence exceeds the frame’s capacity ($n>m$), a segmentation and decomposition operation is triggered:

$$T=T_1\oplus T_2\oplus \cdots \oplus T_k,\mathrm{where}\left|T_i\right|\le m\forall i\in \{1,2,\dots ,k\}$$

(12)

where $T_i$ is the i-th decomposed sub-term set, and $|T_i|$ denotes its cardinality (i.e., the number of terms within the subset). The condition $|T_i| \le m$ strictly guarantees that the size of each decomposed subset does not exceed the maximum number of placeholders m available in a single syntactic frame f. The ⊕ symbol represents the ordered sequence concatenation operator. Each subset $T_i$ is independently instantiated with a copy of frame f, resolving sequence length mismatches while preserving semantic integrity. By dynamically evolving tens of thousands of complex long-sentence variants using extremely limited parent frames, this mechanism greatly expands the support set of the syntactic distribution and fundamentally blocks the model from memorizing specific word orders.

Furthermore, to strictly prevent the generation of invalid case marker combinations during perturbation, the morphological rules from the classical Tibetan grammar treatise Sum cu pa are deeply integrated. A dynamic contextual validity mask matrix $M_{valid}$ regulates case marker combinations based on the preceding token’s suffix (rjes-’jug). For instance, if the preceding token ends with the suffix ན་ (n), the valid genitive case marker is ཀྱི (kyi), rendering གི (gi) and གྱི (gyi) illegal. The mask assigns a severe penalty ($-{10}^9$) to illegal combinations prior to probability normalization, effectively truncating invalid generative paths at the source.

To ensure homomorphic updates on the target side during dynamic length adaptation, the paired syntactic templates are formally defined as a joint tuple $T_{paired}=\langle V_{tib},V_{eng}\rangle$, where the Tibetan template $V_{tib}$ and the English template $V_{eng}$ are strictly bound in their semantic roles. The corresponding English template vector is denoted as:

$$V_{eng}=[F_{eng}^1,P_{eng}^1,F_{eng}^2,P_{eng}^2,\dots ,F_{eng}^n]$$

(13)

where $F_{eng}$ represents fixed syntactic skeleton components (e.g., articles, prepositions, auxiliary verbs) and $P_{eng}$ represents placeholders. Each Tibetan placeholder (e.g., $P_{tib}^1$) strictly corresponds to its English counterpart ($P_{eng}^1$).

When the Tibetan side introduces the padding operation with $\langle \mathrm{PAD}\rangle$, the system invokes a strong mapping function $f_{map}$ supported by the domain bilingual dictionary to generate the English term adapted set in strict synchronization:

$$E_{adapt}=f_{map}\left(T^{\prime }\right)=\{e_1,e_2,\dots ,e_k\}\cup \left\{P_{eng}^{\varphi }\right\}$$

(14)

where $e_i$ is the mapped English term and $P_{eng}^{\varphi }$ corresponds to the English empty placeholder. The final target-side English sentence is rigorously reconstructed via:

$$S_{eng}={\mathrm{\Phi }}_{smooth}\left({\mathrm{\Omega }}_{concat}(V_{eng},E_{adapt})\right)$$

(15)

where ${\mathrm{\Omega }}_{concat}$ is the concatenation operator that assigns terms to placeholders, and ${\mathrm{\Phi }}_{smooth}$ is a syntactic smoothing decoder that handles empty placeholders (e.g., automatically falling back an “A, B, and C” structure to an “A and B” structure). This dynamic collaborative mechanism ensures that the generated English sentences strictly align with Tibetan in underlying semantics while fully conforming to medical terminology expression habits and target-side grammatical norms. This highly flexible schema facilitates the batch construction of sentences spanning 1 to 21 words without necessitating structural redesigns, greatly maximizing the framework’s scalability.

3.3.2. EOT-Based Robust Optimization

To ensure feature robustness across diverse Tibetan writing variants, the Expectation-over-Transformation (EOT) framework is introduced. While dynamic syntactic adaptation provides structural flexibility, it does not account for stylistic perturbations. By defining a custom writing transformation set H, continuous textual perturbations ${\delta }_{x_{tpl}}$ are applied to the template baseline $x_{tpl}=\mathrm{Embed}\left(f\left(T^{\prime }\right)\right)$, yielding $x_{tpl}+{\delta }_{x_{tpl}}$, which preserves differentiability for gradient-based optimization. This perturbation based on expectation transformation forces the generated text to form a smooth and continuous manifold in the semantic space, ensuring the downstream model can still make accurate cross-lingual mappings relying on the learned continuous semantic representations even if it encounters unseen syntactic arrangements. This substantially reduces the model’s susceptibility to the template memorization effect.

This optimization is directed by a joint loss function integrating EOT regularization and the prealignment criteria from Section 3.2:

$$\begin{array}{cc}\hfill {\mathcal{L}}_{total}& ={\mathbb{E}}_{h\sim H}[{\mathcal{L}}_{model}(G(x_{tpl}+{\delta }_{x_{tpl}}),y_{tgt})\hfill \\ \hfill & +{\mathcal{L}}_{model}(h\left(G(x_{tpl}+{\delta }_{x_{tpl}})\right),y_{tgt})]\hfill \\ \hfill & +\gamma {\mathcal{L}}_{align}(h\left(G(x_{tpl}+{\delta }_{x_{tpl}})\right),\tilde{t})+{\lambda }_2{\parallel {\delta }_{x_{tpl}}\parallel }_2^2\hfill \end{array}$$

(16)

where $G(·)$ employs a Gumbel-Softmax continuous approximation. This mechanism establishes a feedback-coupled optimization loop, where gradients backpropagate through the frozen translation model to update the perturbations. The weighting hyperparameter $\gamma$ balances translation performance and feature alignment, empirically set to $\gamma =0.5$.

To establish a rigorous validity mapping from continuous latent variables to discrete symbols without fracturing the gradient chain, the decoding operator $G(·)$ is structured as a nested GRADE-STE (Straight-Through Estimator) architecture. First, a vocabulary projection operator maps the continuous state to the static vocabulary matrix $\mathbf{E}\in {\mathbb{R}}^{\left|\mathcal{V}\right|\times d}$. Second, the validity mask $M_{valid}$ is injected to forcefully penalize illegal morphologies. Third, a Gumbel-Softmax approximation yields a soft probability distribution $\tilde{y}$, where the temperature follows the annealing schedule ${\tau }_i=\max ({\tau }_{min},{\tau }_0{\gamma }^i)$. Fourth, the STE derives the discrete output via the $\arg \max$ operator—incorporating a micro-uniform noise ${ϵ}_{tie}\sim U(0,{10}^{-7})$ to resolve ties—while the backward pass utilizes the stop-gradient trick to route gradients safely along the continuous distribution.

3.3.3. Bilingual Collaborative Response Under EOT Perturbations

To maintain high-confidence semantic anchoring when continuous textual stylistic perturbations (H) are introduced, the SSPA framework strictly guarantees semantic constancy for the English target sentences. The perturbations in the transformation set merely simulate morphological variations in domain language writing without altering core semantic dependency structures. The collaborative updating logic on the English side adheres to two explicit principles:

1.

Silent Absorption of Semantically Equivalent Transformations: If the perturbation only involves changes in synonymous case markers unique to Tibetan (e.g., replacing a conjunctive case marker), the English target remains unchanged. This silent absorption accurately models “many-to-one” semantic equivalence, compelling the translation model to learn semantic invariance amidst complex Tibetan morphological noise.

2.

Joint Mapping of Explicit Modifiers: If the transformation explicitly adds or deletes modifiers (e.g., adding an adjective indicating “acute/severe” before a disease name), the system synchronously triggers a predefined English attachment rule base to attach or remove the corresponding English modifier at the aligned syntax tree node.

This deterministic, white-box generation logic ensures that the synchronously generated English copies maintain strict semantic equivalence, significantly mitigating domain hallucinations and providing a reliable cross-entropy supervision signal for neural translation models.

Building upon the prealigned features, Algorithm 2 sequentially applies length adaptation and joint-loss regularization to construct the final pseudo-corpus.

Algorithm 2 Robust Batch Pseudo-Corpus Generation Algorithm

Require: Initial template set $X_{tpl}$, Reference $Y_{tgt}$, Standardized syntactic vector $\tilde{t}$, Term dictionary $\mathcal{D}$, Transformation H, Iterations $T_{outer},T_{inner}$, constraint $ϵ$, rate $\eta$, static word embedding matrix $\mathbf{E}$ Ensure: Batch-generated pseudo-parallel corpus ${\mathrm{\Delta }}_{all}$

1: ${\mathrm{\Delta }}_{all}\leftarrow \varnothing$;   2: for$(x_{tpl},y_{tgt}) \in (X_{tpl},Y_{tgt})$ do   3:     ${\delta }_{x_{tpl}}\leftarrow \mathbf{0}$;   4:     for $i=1$$T_{outer}$ do   5:         ${\tau }_i=\max ({\tau }_{min},{\tau }_0{\gamma }^i)$ {Update exponential temperature annealing schedule}   6:         for $j=1$$T_{inner}$ do   7:             ${\delta }_{x_{tpl}}\leftarrow \mathrm{clamp}({\delta }_{x_{tpl}},-ϵ,ϵ)$;   8:             ${\tilde{x}}_{adv}=\mathrm{Embed}\left(x_{tpl}\right)+{\delta }_{x_{tpl}}$;   9:             $z={\displaystyle \frac{{\tilde{x}}_{adv}·{\mathbf{E}}^T}{\sqrt{d}}}$ Vocabulary projection 10:             $M_{valid}\leftarrow \mathrm{ContextualValidityMask}\left(x_{tpl}\right)$ {Apply massive negative penalty for illegal Tibetan morphology} 11:             $z_{masked}=z+M_{valid}$; 12:             $g\sim \mathrm{Gumbel} (0,1)$; 13:             $\tilde{y}=\mathrm{Softmax}((z_{masked}+g)/{\tau }_i)$ {Gumbel-Softmax differentiable soft probability} 14:             ${ϵ}_{tie}\sim U(0,{10}^{-7})$ {Inject minimal uniform perturbation to break truncation ties} 15:             $y_{hard}=\mathrm{OneHot}\left(\arg \max (\tilde{y}+{ϵ}_{tie})\right)$ {Extract absolute discrete one-hot token with lexicographical fallback} 16:             $y_{STE}=y_{hard}-\mathrm{sg}\left(\tilde{y}\right)+\tilde{y}$ {Apply Straight-Through Estimator to preserve backward differentiability} 17:             Calculate ${\mathcal{L}}_{total}$ using $y_{STE}$ by substituting into Equation (16); 18:             $g_{x_{tpl}}\leftarrow {\nabla }_{{\delta }_{x_{tpl}}}{\mathcal{L}}_{total}$; 19:             ${\delta }_{x_{tpl}}\leftarrow {\delta }_{x_{tpl}}-\eta g_{x_{tpl}}$; 20:         end for 21:         Gradually lower the ${\mathcal{\ell }}_{\infty }$ modification boundary; 22:     end for 23:     ${\tilde{x}}_{adv}=\mathrm{Embed}\left(x_{tpl}\right)+{\delta }_{x_{tpl}}$; 24:     $z_{final}={\displaystyle \frac{{\tilde{x}}_{adv}·{\mathbf{E}}^T}{\sqrt{d}}}+\mathrm{ContextualValidityMask}\left(x_{tpl}\right)$; 25:     ${\tilde{x}}_{text}\leftarrow \mathrm{DecodeText}\left(\arg \max (z_{final}+U(0,{10}^{-7}))\right)$ {Execute deterministic decoding preventing invalid case particles} 26:     ${\mathrm{\Delta }}_{all}\leftarrow {\mathrm{\Delta }}_{all}\cup \left\{{\tilde{x}}_{text}\right\}$; 27: end for 28: return${\mathrm{\Delta }}_{all}$

To further ensure the transparency and traceability of the bilingual synchronous mapping mechanism, a complete worked example illustrating the full-pipeline bilingual generation process is provided in

Table 2. Complete worked example of SSPA bilingual generation.

Processing Phase	Specific Content Display	In-Depth Mechanism Explanation
1. Original Pair/Template Extraction	Original Tibetan reference: ཨ་རུ་ར་འདི་ནི་ནད་དང་རིམས་ལ་ཕན་ནོ། Original English reference: This medicine is effective in treating illness and plague. Abstract bilingual template (Tibetan): [S1_Med] འདི་ནི་ [P1_Dis] དང་ [P2_Dis] ལ་ཕན་ནོ། Abstract bilingual template (English): This [S1_Med] is effective in treating [P1_Dis] and [P2_Dis].	The system performs morphological disassembly on original high-quality domain sentences. Specific medical entities are stripped while preserving core Tibetan case markers (e.g., “དང་”, “ལ་”) and the English predicate structure, forming strictly hard-bound isomorphic placeholder templates.
2. Selected Terms and Mapping	Tibetan terms: 1. གྲུམ་བུ. 2. ཚད་པ. 3. གུར་གུམ. English mapped terms: 1. Rheumatism. 2. Fever. 3. Saffron.	Terms are sampled from a professional dictionary. The system obtains uniquely corresponding English translations via direct dictionary mapping, reliably preventing reliance on fuzzy inference or external translation models.
3. Intermediate Semantic-Syntactic Representation Reconstruction	Tibetan syntax dependency control vectors: - Node_S: [S1_Med] - Node_V: ཕན་ནོ། - Node_O: [P1_Dis] དང་ [P2_Dis] + ལ་ English syntax vector synchronous constraints: - Node_S: This [S1_Med] - Node_V: is effective in treating - Node_O: [P1_Dis] and [P2_Dis]	Memory addresses of placeholders are locked using parts of speech and case marker boundaries. Because the slots align in this instance, no truncation or zero-padding operation is activated.
4. Modified Tibetan Sentence Generation	Generated modified Tibetan sentence: གུར་གུམ་འདི་ནི་གྲུམ་བུ་དང་ཚད་པ་ལ་ཕན་ནོ།	Terms are embedded into the template. The conjunctive marker “དང་” and objective case marker “ལ་” are correctly attached, ensuring no agglutinative rules are violated, forming a grammatically valid declarative sentence.
5. Final English Target Synchronous Updating	Generated final English target sentence: This saffron is effective in treating rheumatism and fever.	Utilizing the explicit bilingual mapping mechanism, the English side executes strict isomorphic replacement within the English syntactic space without retaining original literature terms.

. This demonstrates the strict isomorphic replacement and homomorphic target-side generation from original baseline resources to the final pseudo-parallel sentence pair.

4. Experiments

4.1. Experimental Setup

4.1.1. Datasets

To evaluate the effectiveness of the proposed SSPA framework in low-resource settings, the base corpus used for model training consists of 200,000 general Tibetan-English bilingual parallel texts of news and current affairs, which were crawled from the Internet and curated.

Currently, there is no large-scale standard public bilingual dataset specifically for the traditional Tibetan medicine domain. In this work, we construct the Tibetan Medicine-Tibetan English Bilingual Dataset (TM-TE), which integrates 9166 professional Tibetan medicine bilingual terms and 5 sets of typical domain syntactic frames. Specifically, the base dataset contains 500 high-quality parallel sentence pairs. The corpus comprehensively covers professional expressions such as Tibetan medicine diseases, prescriptions, diagnoses, and treatments, and all sentences have been manually annotated.

This dataset presents significant experimental challenges. Tibetan is an agglutinative language with complex case marker concatenation rules. Meanwhile, Tibetan medicine texts feature concise sentence structures, unique expression styles, and an extremely high proportion of rare domain-specific terms, which greatly increases the difficulty of corpus modeling and translation. Following the general experimental partitioning rules, we partition the 500 domain pairs into 400 for training, 50 for validation, and 50 for testing. It is imperative to document the provenance of all extracted resources. Specifically, the 9166 terminology dictionary and the 5 syntactic frames were strictly compiled prior to model training by solely leveraging the 400 training pairs alongside external classical historical texts. The 50 validation and 50 test sentences were strictly partitioned and isolated before any knowledge extraction took place. To facilitate full reproducibility and adherence to open-source standards (FAIR principles), we have explicitly desensitized and packaged the corresponding resources into four structured directories: data_splits/ containing the JSON split files establishing a unified low-resource baseline; syntactic_resources/ holding the 5 core domain syntactic frameworks as YAML templates with explicit placeholders and case marker topologies; terminology/ releasing the 9166 bilingual terminology mappings in CSV format; and generated_corpus/ containing the 5000 generated pseudo-parallel sentences. All aforementioned resources are officially released under the Creative Commons Attribution 4.0 International License (CC BY 4.0) via dual hosting on Zenodo and GitHub platforms, https://github.com/ZBY-four-two/SSPA (accessed on 26 August 2026).

In addition, to comprehensively evaluate the generalization ability of various methods, two editions of public Tibetan-English translation datasets, CCMT2020 and CCMT2023 [52], are additionally introduced for comparative testing. These datasets mainly consist of general-domain sentences, with significant differences in syntactic structure and writing style from professional Tibetan medicine texts, which can effectively verify the practical performance of the proposed method in cross-domain scenarios. To explicitly assure readers of the purity and uncontaminated attributes of the dataset, we conducted a rigorous term-level and sentence-level overlap analysis between the training distribution and the TM-TE domain test set. The exact and fuzzy sentence-level match rates strictly remain at 0.00%, providing strong evidence that no data leakage or cross-contamination occurred during the model evaluation phase.

4.1.2. Training Details

The SSPA method adopts the standard Transformer-base [20] as the backbone framework for Tibetan-English neural machine translation. The pseudo-corpus generation pipeline is optimized based on the semantic-syntactic dual-domain joint loss designed by SSPA, as formulated in Equation (7) in Section 3.2, which simultaneously constrains the syntactic matching degree and fine-grained semantic consistency of translations. The augmented dataset is then utilized to train the standard Transformer model using the conventional cross-entropy loss. Specifically, the final training mixture comprises 200,000 general domain pairs, 400 authentic domain pairs, and 5000 pseudo-pairs generated by SSPA, which are concatenated and randomly shuffled.

All algorithm codes are developed based on the PyTorch (version 2.12.1) deep learning framework [53]. The Adam optimizer [54] is employed for training with hyperparameters set as ${\beta }_1=0.9$ and ${\beta }_2=0.98$. The total number of training epochs is set to 30, and the batch size is set to 64. All computations are performed on an NVIDIA RTX 3080 graphics card. To prevent out-of-memory (OOM) issues given the batch size of 64 on this hardware, gradient accumulation and a maximum sequence length constraint are employed. Profiling indicates that the offline generation pipeline requires a modest 4.62 GB of peak GPU memory and 8.24 GB of CPU memory, ensuring its accessibility on consumer-grade hardware. For the Gumbel-Softmax discrete approximation within the pseudo-corpus generation pipeline, the temperature annealing schedule is strictly configured with an initial temperature ${\tau }_0=1.0$. This temperature undergoes exponential decay at a rate of $0.95$ per iteration epoch, which is bounded by a minimum temperature lower bound ${\tau }_{min}=0.2$ to maintain continuous relaxation characteristics without gradient vanishing. We confirm that all hyperparameter selections, including the filtering thresholds and alignment weights, were tuned exclusively utilizing the validation set consisting of 50 pairs. The test set remained strictly unseen throughout the entire pipeline, thereby dispelling concerns regarding hyperparameter snooping and facilitating the fairness of the evaluation.

The initial learning rate of the model is set to 0.002, and a multi-stage learning rate adjustment strategy with warmup is adopted, with a 4000-step learning rate warmup phase to enhance the convergence stability at the early stage of model training. In the data preprocessing stage, a Tibetan word segmentation module based on Bi-LSTM+CRF [3] and a Tibetan-optimized Byte Pair Encoding (BPE) subword encoding algorithm are utilized. Crucially, the underlying Tibetan parser utilized in our evaluation and pre-alignment pipeline is a deep sequence labeling architecture centered on Bi-LSTM+CRF, optimized for the special agglutinative linguistic characteristics of Tibetan. At the feature extraction level, the Bi-LSTM captures long-distance bidirectional contextual dependencies of the sentence (e.g., SOV structures spanning dozens of characters), while the CRF in the decoding layer learns the legitimate state transition matrices between Tibetan case markers to strictly monitor grammatical boundaries. To establish an unbiased grammatical prior, this Bi-LSTM+CRF parser component was pre-trained on a fully independent general Tibetan treebank encompassing nearly 30,000 high-quality manually annotated sentences covering news, history, law, and other domains, preventing exposure to the target domain test distribution. The configuration of the entire preprocessing components and training environment fully satisfies the operational requirements of low-resource Tibetan-English machine translation tasks. To facilitate complete experimental transparency and reproducibility, a comprehensive global view of experimental hyperparameter configurations for each baseline model and the proposed SSPA architecture is summarized in

Table 3. Comprehensive global view of experimental hyperparameter configurations for each baseline model and the proposed SSPA architecture.

Configuration Specification	CINO Transformer/SSPA Backbone	NLLB-200 (Few-Shot Baseline)	SemiAdapt-LoRA	Tagged Back-Translation (via LLaMA-3)
Exact Model/Checkpoint	hfl/cino-base-v2	facebook/nllb-200-distilled-600M	hfl/cino-base-v2 + LoRA Matrix	meta-llama/Meta-Llama-3-8B
Architecture Type	Dense Encoder-Decoder	Dense Encoder-Decoder	Dense Encoder-Decoder w/PEFT	Decoder-only LLM
Parameter Count	∼186 Million	∼600 Million (Distilled Version)	Base: 186M + Trainable: ∼6M	∼8 Billion Parameters
Tokenizer Model	SentencePiece (BPE)	NLLB specific SP Tokenizer	SentencePiece (BPE)	BPE (Tiktoken)
Vocabulary Size	25,000 (Optimized for Tibetan)	256,206 (200-language capacity)	25,000	128,000
Number of Authentic Pairs	200,000 (News) + 400 (Train)	200,000 (News) + 400 (Train)	200,000 (News) + 400 (Train)	N/A (Zero/Few-shot Generative)
Number of Pseudo-Pairs	5000 (via SSPA Dual-domain)	5000 (via SeqMix/Random)	5000 (via Random Word Replacement)	N/A
Training Epochs	Max 30 Epochs	Max 15 Epochs	Max 30 Epochs	N/A (Inference-only prompts)
Learning Rate (Peak)	0.002	N/A	0.002 (Applied to Adapters)	N/A
LR Schedule & Optimizer	Inverse Sq-Root/Adam	Linear Decay/AdamW	Cosine Annealing/AdamW	N/A
Warmup Steps	4000 steps	2500 steps	4000 steps	N/A
Global Batch Size	64 (Gradient Accumulation)	32 (Due to the 600M-parameter memory footprint)	64	N/A
Specific Hyperparameters	Max modified words $ϵ=2$	Beam Size = 4, Rep. Penalty = 1.2	LoRA Rank, Alpha	Temperature = 0.1, Top-p = 0.9
Early-Stopping Rule	Patience = 5 (Monitored on Val Loss)	Patience = 3 (Monitored on Val Loss)	Patience = 5 (Monitored on Val Loss)	N/A
Model-Selection Criterion	Highest Val BLEU Score	Highest Val BLEU Score	Highest Val BLEU Score	N/A
Random Seeds	12345 (Fixed for reproducibility)	12345	12345	N/A

.

4.1.3. Comparison Methods

In this experiment, to align with the technical bottlenecks discussed in Section 2, the SSPA algorithm is horizontally compared with four categories of baseline methods. For massively multilingual pre-trained models, NLLB-200 [55] is selected. Specifically, we utilize the facebook/nllb-200-distilled-600M checkpoint, which is a dense Transformer encoder–decoder network (approximately 600M parameters). The baseline fine-tuned solely with 500 authentic parallel sentences is uniformly named Base CINO Transformer (No Augmentation) [56]. To represent advanced generative rewriting, the CINO + Random Word Replacement method is introduced alongside SeqMix [57] and Tagged Back Translation powered by LLaMA-3 [58]. Domain adaptation methods adopt the parameter-efficient fine-tuning (PEFT) approach SemiAdapt-LoRA [45] and knowledge distillation [59].

For methods relying on pre-trained weights such as knowledge distillation, the experiment uniformly uses a general-domain pre-trained model for fine-tuning to ensure that the conditions of all comparative experiments are completely consistent. This experiment focuses on investigating the translation performance in the low-resource Tibetan medicine domain while also considering the generation quality of the pseudo-corpus. Therefore, multiple metrics are selected to comprehensively evaluate all comparative schemes.

4.1.4. Evaluation Metrics

This experiment quantifies algorithm performance using multiple complementary evaluation metrics, with BLEU-4 adopted as the core automatic evaluation metric for machine translation tasks. In our experiments, all BLEU and CHRF++ scores are computed using the standardized sacreBLEU toolkit [60]. The calculation of BLEU-4 is defined as follows:

$$BLEU=BP·\exp \left({\displaystyle \sum _{n=1}^4}{w}_n\log p_n\right)$$

(17)

where $BP$ is the brevity penalty, $w_n$ represents the uniform weights, and $p_n$ denotes the modified n-gram precision.

The experiment simultaneously introduces TER (Translation Error Rate) and CHRF++ (Character n-gram F-score++) character similarity metrics. These two metrics are better adapted to the evaluation requirements of Tibetan writing characteristics and professional terminology and can accurately reflect the detailed quality of translations. TER measures the minimum number of edits required to change a hypothesis into one of the references:

$$TER={\displaystyle \frac{E}{R}}\times 100\%$$

(18)

where E is the number of edits (insertions, deletions, substitutions, and shifts) and R is the average length of reference translations. CHRF++ evaluates character n-gram matching supplemented by word n-grams:

$$CHRF++=\left(1+{\beta }^2\right){\displaystyle \frac{chrP·chrR}{{\beta }^{2}chrP+chrR}}$$

(19)

where $chrP$ and $chrR$ represent character/word n-gram precision and recall, respectively, and $\beta$ is the parameter balancing precision and recall (set to $\beta =2$ in our experiments).

The calculation logic of the BLEU score heavily relies on exact hard matching of word-level n-grams. While it might yield artificially high scores by generating massive common words in the general news domain, it is extremely insensitive to fine-grained morphological variations. In a typical agglutinative language like Tibetan, the misuse of a single case marker can substantially alter the core relationship of the entire sentence. Therefore, we adopt chrF++ as a crucial supplementary standard, which comprehensively considers precision and recall and accurately captures morphological correctness at the sub-word level, such as affixes and case markers, by introducing character-level matching.

A human evaluation phase is explicitly established to counteract the limitations of automatic metrics in capturing profound semantic and morphological nuances. We assembled a specialized panel comprising three bilingual NLP experts and two Tibetan medicine domain specialists (5 total raters) to independently score 100 randomly sampled translations. Evaluation covers three distinct dimensions on a 5-point Likert scale: terminology translation accuracy, Tibetan syntactic compliance, and Tibetan medicine domain style matching degree. This procedure operates under a rigorous double-blind strategy. Crucially, to mathematically quantify the inter-rater reliability and reduce subjective bias among the multiple annotators, Fleiss’ Kappa ($\kappa$) is computed for each dimension. Furthermore, to objectively validate the superiority of the proposed SSPA framework over the state-of-the-art baselines, statistical significance is assessed using paired t-tests, with precise p-values and confidence intervals formally reported in the ensuing results section. This demonstrates robust performance against random variance and adheres to stringent statistical norms. For the pseudo-parallel corpus generated by SSPA, two additional statistical metrics—grammatical compliance rate and domain syntactic style similarity—are introduced to specifically verify the generation reliability and domain adaptability of the pseudo-corpus, which are computed using an automated syntactic parser to detect case marker errors and cosine similarity of syntactic dependency tree embeddings, respectively.

All metrics are calculated on the general-domain test set, Tibetan medicine domain test set, and multi-domain mixed test set, respectively, to thoroughly validate the overall performance and robustness of the algorithm.

4.2. Experimental Results and Analysis

This section comprehensively validates the effectiveness of the SSPA framework in low-resource Tibetan–English machine translation tasks from four dimensions: core performance in the Tibetan medicine domain, general-domain and cross-domain generalization ability, human evaluation and error analysis, and quality of generated pseudo-corpora. All experiments are based on the CINO Transformer backbone model and are conducted on the self-built Tibetan medicine dataset TM-TE, as well as the public general-domain datasets CCMT2020 and CCMT2023. The automatic metric results are averaged over five independent training runs, and the human evaluation results are obtained using a double-blind review mechanism.

4.2.1. Tibetan Medicine Domain Translation Performance

Table 4. Core evaluation metrics across 5 random seeds for all datasets (mean ± SD). The standard deviations presented strictly represent the variance across the 5 independent random seeds.

	Method	TM-TE (Medical Domain)			CCMT2020	CCMT2023
		BLEU-4	TER	CHRF++	BLEU-4	BLEU-4
Pre-trained Model	NLLB-200	16.8 ± 0.0	65.2 ± 0.0	35.7 ± 0.0	32.7 ± 0.0	32.0 ± 0.0
Transformer	Base CINO Transformer (No Augmentation)	22.1 ± 0.5	58.3 ± 0.6	41.5 ± 0.4	38.7 ± 0.6	37.2 ± 0.5
Data Augmentation	SeqMix	25.7 ± 1.1	53.2 ± 1.2	45.8 ± 1.0	50.3 ± 1.3	48.6 ± 1.2
	LLaMA-3 (Zero-shot)	24.3 ± 0.0	54.7 ± 0.0	44.2 ± 0.0	49.5 ± 0.0	39.8 ± 0.0
	Tagged Back Translation	28.9 ± 0.8	49.6 ± 0.9	49.1 ± 0.7	40.2 ± 0.9	38.6 ± 0.8
Domain Adaptation	SemiAdapt-LoRA	27.5 ± 0.6	51.3 ± 0.7	47.6 ± 0.5	36.5 ± 0.6	35.1 ± 0.5
	Knowledge Distillation	29.2 ± 0.4	48.9 ± 0.5	49.5 ± 0.4	39.4 ± 0.5	37.9 ± 0.4
Ours	SSPA	36.2 * ± 0.3	40.1 ± 0.4	57.3 ± 0.3	39.1 ± 0.4	37.8 ± 0.3

* Denotes a statistically significant difference ($p<0.05$) compared with the optimal baseline via paired bootstrap resampling.

compares the core translation performance of SSPA against seven mainstream baselines across four categories. The model is trained using only 500 real annotated Tibetan medicine sentence pairs plus 5000 pseudo-sentence pairs generated by each method, which are combined with 200,000 general-domain news Tibetan–English parallel corpora.

Table 4 presents the core automatic evaluation metrics of each method on the TM-TE Tibetan medicine test set. The NLLB-200 model yields the lowest performance (BLEU-4 = 16.8). This occurs because its attention mechanism is highly dispersed across 200 distinct language feature spaces. Under extreme data scarcity and highly specific agglutinative features, zero-shot or few-shot inference cannot autonomously learn the complex case particle concatenation rules. The Base CINO Transformer (No Augmentation) improves BLEU-4 to 22.1, but suffers from significant performance degradation in the professional domain due to learning only general-domain linguistic knowledge. Traditional data augmentation methods like SeqMix and zero-shot LLaMA-3 yield limited improvements, achieving BLEU-4 scores of 25.7 and 24.3, respectively. This limitation arises because unconstrained modifications frequently corrupt Tibetan syntactic structures. Tagged Back Translation achieves a higher BLEU-4 score of 28.9. However, the generated pseudo-corpora continue to exhibit grammatical irregularities and stylistic mismatches.

In domain adaptation, SemiAdapt-LoRA achieves a BLEU-4 score of only 27.5. Its trainable low-rank parameter capacity is minimal. This limited parameter space is quickly diluted by massive gradients from the 200,000 general news sentences. Consequently, weak gradient signals from the 500 domain target sentences fail to establish dominance, leading to underfitting in special vocabulary mapping and complex Tibetan syntax. Knowledge Distillation (BLEU-4 = 29.2) represents the strongest baseline, yet remains constrained by the low-resource data sparsity bottleneck.

The SSPA framework achieves the highest scores across all metrics. It reaches a BLEU-4 of 36.2, marking a 7.0-point improvement over Knowledge Distillation. Additionally, TER decreases by 8.8 points, and CHRF++ increases by 7.8 points. These improvements stem from the semantic-syntactic dual-domain pre-alignment mechanism. This mechanism resolves the grammatical and semantic deficiencies of traditional data augmentation. Concurrently, dynamic syntactic padding and EOT-style regularization facilitate the domain adaptability of the pseudo-corpora.

Furthermore, optimizer instability and weight random initialization variance are primary causes of irreproducible results in low-resource environments. As demonstrated in Table 4, traditional data augmentation methods suffer from inherent random masking and sampling noise, causing severe performance fluctuations. For instance, SeqMix achieves strong results on general domains, but its standard deviation on the TM-TE test set reaches a substantially high 1.1 BLEU. This severe fluctuation occurs because continuous vector mixing frequently disrupts the discrete boundary structure of Tibetan syntax under random initialization. In contrast, the SSPA framework demonstrates stable training convergence characteristics across 5 independent random seeds (12345, 42, 2026, 777, and 999). Its test variance is strictly confined within the narrow range of ±0.3 to ±0.4 across all specific and general domains. This high stability is attributed to our Expectation-over-Transformation (EOT) robust optimization mechanism. By integrating stochastic transformations directly into the optimization objective, the synthesized pseudo-sentences gain strong immunity against syntactic perturbations, making the model highly resilient to the random initialization of network weights.

To verify the marginal benefit of SSPA under different data scales,

Table 5. BLEU-4 trends under different pseudo-corpus scales.

Pseudo-Corpus Scale	CINO + Random Word Replacement	Tagged Back Translation	SSPA
0k	22.1	22.1	22.1
0.5k	23.5	24.8	28.7
1k	24.2	26.1	31.4
2k	25.0	27.3	33.9
5k	25.6	28.9	36.2
10k	26.1	29.7	37.5

presents the BLEU-4 trends on the TM-TE test set as the pseudo-corpus scale increases. At the 0k scale (using only 500 real corpora), all three methods achieve an identical performance of 22.1 due to the absence of data augmentation. The CINO + Random Word Replacement baseline exhibits slow improvement, reaching only 26.1 at the 10k scale. Random word replacement arbitrarily substitutes vocabulary during pseudo-corpus generation. This substitution severs the binding relationship between nouns and case particles in Tibetan, injecting massive amounts of syntactic noise. Tagged Back Translation reaches 29.7 at 10k, but its growth rate slows significantly after 5k. This stagnation occurs because the generated pseudo-corpora continue to suffer from grammatical and stylistic deficiencies, limiting effective knowledge gains. The SSPA framework significantly outperforms at all scales. Its performance at 1k (31.4) already surpasses Tagged Back Translation at 10k, and it plateaus efficiently after 5k. A small sample of high-quality generated pseudo-corpora suffices to cover the core linguistic features of the Tibetan medicine domain.

In Tibetan medical literature, long sentences account for a high proportion and exhibit complex syntactic structures. The test set is divided into three groups according to sentence length, and the BLEU-4 scores of each method are reported, as shown in

Table 6. BLEU-4 scores across different sentence length groups.

Sentence Length	Proportion	Base CINO Transformer (No Augmentation)	Tagged Back Translation	SSPA
<10 words	42%	26.3	31.5	37.8
11–20 words	26%	21.7	28.2	35.9
>21 words	32%	16.8	23.1	34.9

. For short sentences with fewer than 10 words (accounting for 42%), SSPA achieves a 6.3-point improvement over Tagged Back Translation (37.8 vs. 31.5), because short sentences have simple structures and the performance differences among methods are relatively small. For medium sentences of 11–20 words (accounting for 26%), SSPA achieves a 7.7-point improvement, as medium sentences begin to involve complex case-marker connections, where the syntactic alignment mechanism starts to demonstrate its advantages. For long sentences with more than 21 words (accounting for 32%), SSPA demonstrates a significant improvement, outperforming Tagged Back Translation by 11.8 points in long-sentence BLEU-4 (34.9 vs. 23.1). The dual-domain pre-alignment mechanism accurately models the multi-layered case-marker nesting and long-distance semantic dependencies. This effectively resolves the grammatical fragmentation and semantic deviation problems prevalent in traditional methods.

4.2.2. Cross-Domain Generalization Analysis

To explicitly clarify the true trajectory of different methods in cross-domain generalization, we introduce a relative performance delta analysis (evaluating the performance deviation relative to the baseline Base CINO Transformer). As detailed in

Table 7. Relative performance Delta analysis of cross-domain generalization.

Method	TM-TE	TM-TE $\mathrm{\Delta }$	CCMT2020	CCMT2020 $\mathrm{\Delta }$	Generalization Performance Evaluation
Base CINO	22.1	Baseline	38.7	Baseline	Lacks professional domain adaptation capabilities.
SeqMix	25.7	+3.6	50.3	+11.6	Unbalanced generalization; syntactic destruction.
Tagged BT	28.9	+6.8	40.2	+1.5	Constrained by monolingual data scale/style.
SSPA (Ours)	36.2	+14.1	39.1	+0.4	Ideal zero-forgetting generalization.

, this evaluation framework emphasizes that the measurement benchmark should not be an overall peak score overfitted to general corpora, but rather the ability to achieve a disruptive improvement in the specific domain while realizing a slight positive increase in the general domain—what we term ‘Zero-forgetting generalization’.

Accordingly, our cross-domain objective is not to maximize peak BLEU on general-domain test sets at the expense of professional-domain structural accuracy, but to prevent catastrophic forgetting of general-domain competence while achieving substantial professional-domain gains. SeqMix achieves a high score (+11.6 BLEU) on the CCMT2020 general test set. However, its core mechanism relies on sequence-level soft combination by sampling a mixing parameter from a Beta distribution in a continuous embedding space. For highly agglutinative languages like Tibetan, this brute-force continuous interpolation cannot perceive the indivisible atomic structure of roots and case markers, leading to the collapse of morphological boundaries when mapped back to discrete text. Consequently, it yields only a weak in-domain improvement (+3.6 BLEU) on TM-TE, representing a typical unbalanced generalization and false overfitting. Tagged Back Translation (+1.5 BLEU on CCMT2020, +6.8 BLEU on TM-TE) relies on large-scale monolingual target-domain corpora for reverse generation. In low-resource domains, it is constrained by data distribution and struggles to capture specific domain styles.

In contrast, the proposed SSPA framework achieves an asymmetric win-win strategy. By precisely injecting semantic and syntactic structure priors, it produces a massive surge of +14.1 BLEU points in the specific domain. Crucially, its optimization occurs in a prior language space decoupled from the translation model. SSPA avoids backpropagation updates to any of the neuron weights of the backbone network. This plug-and-play design substantially mitigates the risk of disrupting general-domain feature representations, preserving general capabilities with a +0.4 BLEU increase.

Furthermore, this subsection constructs a multi-domain mixed test set comprising news (33%), law (33%), and Tibetan medicine (33%) to evaluate the robustness of the models in scenarios with mixed linguistic styles. Specifically, the legal-domain subset consists of exactly 50 bilingual sentence pairs, aligning with the 50 pairs in both the Tibetan medicine and news test sets and maintaining a balanced 33% distribution for each domain. These pairs were carefully curated from publicly available regional statutory texts and official legal governance documents by random sampling, followed by rigorous manual sentence-level alignment and double-blind proofreading by bilingual linguistics experts to ensure the precise mapping of highly formalized legal terminology. Furthermore, we explicitly declare that no legal-domain data were utilized during any phase of model training, domain resource construction (e.g., template extraction or dictionary building), or hyperparameter tuning; the legal data served strictly as an unseen, out-of-distribution test set to evaluate true zero-shot cross-domain generalization. To quantify the anti-forgetting and generalization abilities when facing out-of-distribution data, the results are detailed in

Table 8. Cross-domain robustness tracking diagnosis profile for various models on multi-domain mixed test set.

Method	Domain Performance (BLEU-4)			Cross-Domain Avg.	Domain-SD (Penalty)
	News	Law	Tibetan Med
NLLB-200	32.7	30.6	16.8	26.7	±8.6
CINO Transformer	38.7	26.2	22.1	29.0	±8.7
SeqMix	50.3	22.7	25.7	32.9	±15.2
LLaMA-3 (Zero-shot)	49.5	18.0	24.3	30.6	±16.6
Tagged Back Translation	40.2	24.5	28.9	31.2	±8.1
SemiAdapt-LoRA	36.5	30.2	27.5	31.4	±4.6
Knowledge Distillation	39.4	26.8	29.2	31.8	±6.7
SSPA (Ours)	39.1	28.2	36.2	34.5	±5.6

, which introduces the “Cross-Domain Avg. BLEU-4 ± Domain-SD” metric. The variance here is not a metric evaluating randomness between different training iterations; rather, it is a reverse measurement penalty term used to evaluate Cross-Domain Robustness. A larger Domain-SD indicates that the model exhibits high performance variance across text genres, overfitting specific domains while degrading sharply in others. A value approaching zero proves the model maintains high-level translation accuracy across vastly different text genres.

Analysis of Table 8 provides a systematic diagnosis of model bias. Traditional brute-force data augmentation like SeqMix and massive language models like LLaMA-3 (Zero-shot) achieve high scores of 50.3 and 49.5 on general corpora like news, but stall on Tibetan medicine or legal texts, causing their domain fluctuation variance to reach as high as ±15.2 and ±16.6. The Base CINO Transformer (No Augmentation) achieves a Cross-Domain Avg. BLEU-4 of 29.0 with a Domain-SD of ±8.7, as it learns only general-domain news knowledge and exhibits poor adaptability to the law and Tibetan medicine domains. Tagged Back Translation achieves a Cross-Domain Avg. BLEU-4 of 31.2 with a Domain-SD of ±8.1. Knowledge Distillation achieves an average of 31.8 with a Domain-SD of ±6.7 as it alleviates domain fluctuations to some extent through knowledge transfer but still cannot accurately model professional domain characteristics. The proposed SSPA framework demonstrates a highly balanced performance profile. It leverages an extremely scarce 400 sentences to complete a substantial performance improvement in the Tibetan medicine domain (36.2), maintains competitive performance in the news domain (39.1), and ultimately suppresses the cross-domain fluctuation amplitude to a low ±5.6. This balanced cross-domain behavior confirms its genuine cross-domain generalization and demonstrates its advantage in low-resource data-skewed scenarios.

4.2.3. Human Evaluation

Automatic metrics can only reflect surface-level similarity of translation outputs. To more comprehensively evaluate translation quality, we implemented a meticulous double-blind evaluation protocol. We utilized an equidistant stratified sampling algorithm to extract 500 samples each from pseudo-corpora generated by three completely different methods: SeqMix based on embedding layer interpolation, Tagged Back Translation based on heavy back-translation, and our SSPA framework, forming a total evaluation set of 1500 sentences. The evaluation team fully considered cross-domain expertise, consisting of three senior Tibetan-English bilingual machine translation experts with over ten years of experience and two senior physicians who have been engaged in the digital proofreading of traditional Tibetan medical ancient texts for many years. We removed the algorithmic source labels of all sentences before distributing the tasks in a randomized double-blind order. The evaluators were asked to check each sentence point-by-point to determine whether there were case marker breaks, constituent inconsistencies, or illogical word orders, providing explicit binary classification labels (1 for compliance, 0 for non-compliance). The study adopted the classic Fleiss’ Kappa ($\kappa$) index to quantify the degree of agreement among the annotations of the five experts, reducing evaluation noise caused by subjective cognitive bias. The calculation results indicate the overall Fleiss’ $\kappa$ value reached 0.86, which is generally considered to have reached the standard of almost perfect agreement, ensuring the human-annotated set can serve as a reliable reference standard.

Simultaneously, to evaluate the final translation outputs, the experts scored 100 randomly selected translation results. The scoring dimensions include terminology translation accuracy, Tibetan syntactic compliance, and domain style matching degree for Tibetan medicine, with each dimension using a 1–5 point scale (5 being the best).

Figure 3 presents the aggregated average human scores for each translation framework. To establish the high reliability of these expert ratings, Fleiss’ Kappa was calculated across all 100 evaluated samples. The expert panel demonstrated ’Almost Perfect’ agreement on terminology accuracy ($\kappa =0.882$), and ’Substantial’ agreement on both syntactic compliance ($\kappa =0.762$) and domain style matching ($\kappa =0.704$), thereby validating the dataset’s high consistency. The detailed paired t-test and Fleiss’ Kappa statistics are summarized in

Table 9. Comprehensive evaluation results: Paired t-test analysis and Fleiss’ Kappa statistics.

Evaluation Dimension	SSPA Mean (SD)	KD Mean (SD)	p-Value (t-Test)	Fleiss’ $\mathit{\kappa }$	Consistency Rating
Terminology Accuracy	4.50 (0.42)	3.70 (0.58)	<0.0001	0.882	Almost Perfect
Syntactic Compliance	4.40 (0.48)	3.60 (0.61)	<0.0001	0.762	Substantial
Style Matching Degree	4.20 (0.51)	3.50 (0.59)	<0.0001	0.704	Substantial
Comprehensive Score	4.37 (0.35)	3.62 (0.41)	<0.0001	0.780	Substantial

. As illustrated, the baseline CINO Transformer exhibits the lowest comprehensive score (2.3), heavily penalized by terminology hallucinations and grammatical fragmentation stemming from extreme data sparsity. Conversely, the proposed SSPA framework achieves the highest scores across all dimensions (4.5 for terminology, 4.4 for syntax, and 4.2 for style, culminating in a comprehensive score of 4.37). To corroborate the observed superiority over the most competitive baseline, Knowledge Distillation (overall score 3.62), a paired t-test was conducted. The results confirmed that the performance margin achieved by SSPA is statistically highly significant ($p<0.0001$). This statistically backed evidence strongly indicates that the semantic-syntactic dual-domain pre-alignment mechanism effectively reduces syntactic noise and semantic misalignment, while the EOT-style regularization aligns the generated outputs with professional Tibetan medical norms.

4.2.4. Analysis of Pseudo-Corpus Quality

The quality of pseudo-corpora directly determines the effectiveness of data augmentation. This subsection quantitatively evaluates the quality of pseudo-corpora generated by SSPA and traditional methods from two dimensions: grammatical compliance and domain style similarity.

It can be observed that SeqMix achieves the lowest grammatical compliance rate (61.5%), with the main errors being case-marker mismatches and missing constituents, because continuous embedding interpolation in SeqMix disrupts the discrete agglutinative rules of Tibetan, often decoding into text with broken case-marker connections. Tagged Back Translation raises the compliance rate to 78.3%, with the main errors being word order confusion and redundant constituents, due to its limited capability in modeling the agglutinative features of Tibetan when trained on general-domain corpora.

To objectively quantify whether the underlying Bi-LSTM+CRF parser is truly qualified for the role of grammar referee and free from severe false negatives, we designed an adversarial held-out validation set to test its anti-interference and error-correction limits. We extracted 1000 grammatically sound Tibetan sentences from a library independent of all pre-training data, and invited native linguistic experts to forcibly inject syntactic structural variations into 50% of the sentences (500 sentences). These injected errors were subdivided into four major categories: (1) Missing Case-markers: Forcibly deleting functional case particles after subjects or objects, causing the sentence to lose its logical backbone. (2) Mismatched Case-markers: Deliberately substituting case particles (e.g., replacing the directional ལ་ with the causal པས་). (3) Missing Constituents: Randomly deleting key modifiers or verbs. (4) Word Order Confusion: Breaking the rigid SOV topological structure of Tibetan to create severe constituent inversions. We then utilized the test parser to perform binary classification on these 1000 mixed sentences. As demonstrated in

Table 10. Held-out accuracy evaluation results of the syntactic parser.

Injected Error Type	Precision	Recall	F1-Score	Overall Accuracy
Missing Case-markers	93.4%	95.2%	94.3%	-
Mismatched Case-markers	91.8%	89.7%	90.7%	-
Missing Constituents	88.5%	85.1%	86.8%	-
Word Order Confusion	85.2%	83.6%	84.4%	-
Macro-Average	89.7%	88.4%	89.0%	91.2%

, the parser exhibited exceptional robustness, achieving a macro-average F1-score of 89.0% and an overall binary classification accuracy of 91.2%. Notably, the F1 scores for identifying missing and mismatched case-markers increased to 94.3% and 90.7%, respectively, entirely attributed to the strict monitoring of local grammatical state transitions by the CRF layer.

Furthermore, to mitigate the risk of circular evaluation bias—arising from the utilization of the homologous Bi-LSTM+CRF parser for both syntactic loss optimization and compliance testing—we comprehensively reconstructed the evaluation pipeline into an orthogonal isolated verification matrix. As shown in

Table 11. Cross-validation of heterogeneous parsers and human annotation for grammatical compliance.

Method	Native Parser (Bi-LSTM+CRF)	Heterogeneous Parser (XLM-R)	Domain Expert (Human)	Bias Delta
SeqMix	61.5%	59.2%	58.6%	+2.9%
Tagged Back Translation	78.3%	76.5%	75.1%	+3.2%
SSPA (Ours)	96.2%	95.1%	94.8%	+1.4%

, we cross-validated the native parser against an independent XLM-R-based Dependency Parser and the rigorous multi-expert double-blind human annotation established in Section 4.2.3. The XLM-R architecture relies on the Transformer self-attention mechanism, completely abandoning the CRF transition matrix based on the Markov assumption. Importantly, the XLM-R parser was isolated throughout all stages of model construction, loss backpropagation, and hyperparameter screening of the SSPA framework, providing zero-gradient guidance to constitute a heterogeneous orthogonal automated evaluation.

The multi-dimensional cross-validation data shows that the circular evaluation bias is present: the scores of all methods on the internal parser are slightly higher than those on the heterogeneous parser and human experts. However, the magnitude of the bias is benign and controllable. For the SSPA method, there is merely a +1.4% circular evaluation bias delta between the 96.2% under the native parser and the 94.8% given by the most rigorous human double-blind experts. The core superiority of SSPA did not collapse after stripping away all home-field advantages; it still maintains a significant superiority over traditional SeqMix (58.6%) and Tagged Back Translation (75.1%). This substantial margin indicates that the alignment optimization achieved by SSPA in the first stage is not a manipulation of the scoring function, but genuinely explores and anchors the rigid generation rules unique to the Tibetan agglutinative language, ensuring the generated pseudo-corpus possesses linguistic generalization validity.

We compute the cosine similarity in lexical distribution (bag-of-words model) and syntactic structure (syntactic dependency tree) between the generated pseudo-corpora and real Tibetan medical texts and take the average as the domain style similarity metric. The results are shown in

Table 12. Domain style similarity metrics of generated pseudo-corpora.

Method	Lexical Distribution Similarity	Syntactic Structure Similarity	Average Style Similarity
SeqMix	0.52	0.38	0.45
Tagged Back Translation	0.67	0.59	0.63
SSPA	0.89	0.85	0.87

. The table shows that SeqMix achieves a lexical similarity of 0.52, a syntactic similarity of 0.38, and an average style similarity of only 0.45. Such methods merely perform simple term substitution without considering the syntactic characteristics or stylistic conventions of Tibetan medical texts. Tagged Back Translation improves these two metrics to 0.67 and 0.59, respectively, with an average style similarity of 0.63, but still struggles to replicate the expression characteristics of the professional domain. The proposed SSPA framework achieves lexical and syntactic similarities of 0.89 and 0.85, respectively, and an average style similarity as high as 0.87. This is because the framework incorporates syntactic templates and EOT-style regularization, fully learning and reproducing the unique lexical usage habits and syntactic styles of Tibetan medical literature.

4.2.5. Robustness Under Data Scale Limitation and Statistical Significance Verification

To verify that the significant improvements achieved by SSPA are not statistical anomalies caused by the small scale of the TM-TE test set (50 sentences), we introduce Paired Bootstrap Resampling and 10-Fold Cross-Validation mechanisms.

First, we utilized the SacreBLEU evaluation suite to execute paired bootstrap resampling ($B=1000$). The null hypothesis assumes no essential translation quality difference exists between the SSPA model and the competing baseline, and that any score difference is entirely caused by random sampling error. As shown in

Table 13. SacreBLEU confidence intervals and paired significance test results on the TM-TE test set.

Method	BLEU-4	95% Confidence Interval (Bootstrap B = 1000)	Paired Hypothesis Test against SSPA
NLLB-200	16.8	[15.2, 18.5]	$p<0.01$
CINO Transformer	22.1	[20.4, 23.8]	$p<0.01$
LLaMA-3 (Zero-shot)	24.3	[22.6, 26.2]	$p<0.01$
SeqMix	25.7	[24.1, 27.5]	$p<0.01$
SemiAdapt-LoRA	27.5	[25.8, 29.3]	$p<0.01$
Tagged Back Translation	28.9	[27.3, 30.7]	$p<0.01$
Knowledge Distillation	29.2	[27.6, 31.0]	$p<0.01$
SSPA (Ours)	36.2	[34.5, 37.9]	-

, the upper bound of the 95% confidence interval for the optimal baseline (Knowledge Distillation) is strictly suppressed at 31.0, while the lower bound for our proposed SSPA framework reaches a high of 34.5. The probability space has no intersection, and the test result ($p<0.01$) refutes the conjecture that the high score was accidentally pulled up by individual samples. This high stability is attributed to the Semantic-Syntax Prealignment mechanism, which acts as a strong regularization prior, anchoring the decoding path within compliant Tibetan syntactic frames. To enable independent verification of these statistics, the raw per-seed scores, the complete 10-fold tracking logs, and the paired bootstrap resampling outputs have been released in the project repository (see the Data Availability Statement).

Furthermore, to eliminate the concern that the initially randomly divided 50 sentences might represent an easy-to-translate sub-distribution, we executed a 10-Fold Cross-Validation on the full 500-sentence domain bilingual authentic corpus. It is important to clarify the implementation details regarding the domain resources in this experiment. Given the extensive expert labor required to curate and verify the Tibetan medicine terminology dictionary (9166 mappings) and the five syntactic templates, these resources—derived from the designated 400-pair training split and external classical texts—were reused across all 10 folds rather than reconstructed independently within each fold. Consequently, we acknowledge that this experiment entails a partial information leakage at the prior level and should not be strictly regarded as a fully leakage-free cross-validation. Nevertheless, model weights were initialized from scratch for each fold, and the low variance across folds (SD = 0.39) indicates that the optimization convergence and the EOT mechanism are robust to varied data splits. As detailed in

Table 14. Ten-Fold cross-validation BLEU-4 tracking on the complete 500-pair TM-TE dataset.

Fold ID	CINO Transformer	Tagged Back Translation	Knowledge Distillation	SSPA (Ours)
Fold 1 (Original Test Set)	22.1	28.9	29.2	36.2
Fold 2	21.8	29.1	28.8	35.8
Fold 3	22.5	28.4	29.5	36.5
Fold 4	21.4	27.9	28.6	35.9
Fold 5	22.6	29.3	29.4	36.7
Fold 6	21.9	28.5	29.0	35.5
Fold 7	22.3	28.8	29.1	36.1
Fold 8	22.0	28.6	28.9	36.3
Fold 9	22.7	29.2	29.7	36.6
Fold 10	21.6	28.2	28.5	35.7
Global Mean ± SD	22.09 ± 0.42	28.69 ± 0.44	29.07 ± 0.38	36.13 ± 0.39

, after completing 10 independent training and inference cycles covering all corpus variants, the global average BLEU-4 score of SSPA settled at 36.13, which closely matches the 36.2 reported on the initial single test set. Even with the test data undergoing rotational drift across various folds, the standard deviation of SSPA performance fluctuation was a low SD of 0.39. This confirms that the performance leap stems from algorithmic mechanism advantages rather than coincidence in data splitting.

4.2.6. Morphological Robustness Evaluation

To complement the coarse-grained word-level BLEU metric with sub-word morphological evidence, we additionally report the character-level metric chrF++.

Table 15. Character-Level morphological evaluation and syntactic compliance.

Method	TM-TE chrF++	CCMT2020 chrF++	CCMT2023 chrF++	Grammatical Compliance	Cosine Similarity
Base CINO	41.5	49.8	48.3	N/A	N/A
SeqMix	45.8	53.2	51.7	61.5%	0.52
Tagged BT	49.1	51.4	50.2	78.3%	0.67
SSPA (Ours)	57.3	52.8	51.9	96.2%	0.89

presents the chrF++ scores across the medical and general domains, alongside grammatical compliance and lexical cosine similarity. We first candidly acknowledge the general-domain BLEU gap documented in Table 4 and Table 7: on CCMT2020 and CCMT2023, SSPA (39.1 and 37.8 BLEU-4) trails SeqMix (50.3 and 48.6 BLEU-4) by 11.2 and 10.8 points, respectively. We interpret this gap as an expected design trade-off rather than an oversight. SeqMix attains its high general-domain BLEU through aggressive continuous embedding interpolation that memorizes frequent general vocabulary, while simultaneously disrupting the indivisible root–case-marker structure of Tibetan, leaving only 61.5% grammatical compliance. In contrast, SSPA explicitly prioritizes structural integrity (96.2% compliance) to safeguard professional-domain translation quality, and does not attempt to maximize general-domain word-level matching. The chrF++ results should therefore be read as morphological evidence consistent with this trade-off, not as a claim of general-domain parity: SSPA attains 52.8 and 51.9 on CCMT2020 and CCMT2023, respectively, which are comparable to SeqMix (53.2 and 51.7) at the sub-word level. This indicates that, despite the word-level BLEU deficit, SSPA preserves sub-word morphological correctness—such as affixes and case markers—that exact word matching fails to credit. The high morphological compliance also underpins the strong long-sentence behavior of SSPA reported in Section 4.2.1.

4.3. Ablation Studies

To thoroughly validate the effectiveness of each core module and the rationality of the key designs of the SSPA framework, this section conducts systematic ablation experiments. All experiments maintain exactly the same baseline settings as in Section 4.2. It should be noted that the core-component ablation is conducted on the complete SSPA model (

Table 16. Ablation study of core components in SSPA.

Model Configuration	BLEU-4	TER	CHRF++
Base CINO Transformer (No Augmentation)	22.1	58.3	41.5
with Random Term Padding	25.7	53.2	45.8
with SSPA initialization	31.4	46.7	52.1
with Cross-length Padding	33.8	43.5	54.7
with EOT Regularization	36.2	40.1	57.3

), while the ablation analyses of the loss function, hyperparameter sensitivity, and semantic pooling (

Table 17. Performance under different dual-domain loss function combinations.

Loss Function	BLEU-4	TER	CHRF++	Terminology Accuracy	Grammatical Compliance Rate
Random Initialization without constraints	25.7	53.2	45.8	62.3%	61.5%
only Semantic Loss $L_{Sem}$	28.3	50.1	48.6	71.5%	74.2%
only Syntactic Loss $L_{syn}$	27.9	51.4	47.9	68.7%	82.6%
Fixed-Weight Dual-Domain Loss with $\alpha =0.5,\beta =0.5$	30.1	48.2	50.7	75.2%	89.1%
Adaptive-Weight Dual-Domain Loss with $\lambda$	31.4	46.7	52.1	78.9%	91.3%

,

Table 18. Sensitivity analysis of the maximum number of modified words ($ϵ$).

$\mathit{ϵ}$	BLEU-4	TER	CHRF++	Grammatical Compliance Rate	Average Number of Modified Words
1	29.7	48.2	50.3	97.8%	0.92
2	31.4	46.7	52.2	91.3%	1.87
3	30.2	47.7	50.8	82.5%	2.71
4	28.6	49.8	49.2	73.1%	3.54

and

Table 19. Impact of different multi-scale semantic pooling combinations.

Pooling Scale Combinations	BLEU-4	TER	CHRF++	Semantic Consistency
{1,2} = Word, Bigram	29.5	48.8	50.1	72.3%
{1,2,4} = Word, Bigram, Phrase	30.7	47.5	51.4	76.8%
{1,2,4,8} = Word, Bigram, Phrase, Sentence	31.4	46.7	52.1	78.9%
{1,2,4,8,16} = Word, Bigram, Phrase, Sentence, Long Sentence	31.2	46.9	51.9	78.5%

) are established upon the SSPA initialization stage (serving as an upper baseline with BLEU-4 = 31.4) to perform decoupled validation prior to EOT regularization.

4.3.1. Ablation of Core Architecture Components

This experiment uses the base Transformer as the baseline model and sequentially adds the core components of SSPA, quantifying the contribution of each module by comparing performance changes. The experimental configurations and results are shown in Table 16.

As can be seen from the results, the baseline model achieves a BLEU-4 of 22.1. Adding random term padding yields a modest improvement to 25.7. Incorporating semantic-syntactic dual-domain pre-alignment initialization substantially increases performance to 31.4. Subsequently introducing cross-length dynamic syntactic template padding raises the metric to 33.8. Integrating all modules alongside EOT-style regularization achieves the optimal value of 36.2. Each module contributes a positive gain. The semantic-syntactic dual-domain pre-alignment initialization provides the most prominent contribution. This module corrects grammatical errors and semantic deviations caused by traditional data augmentation. Subsequent modules further optimize sentence pattern adaptability and domain style performance.

4.3.2. Effectiveness of Dual-Domain Alignment Loss

The semantic-syntactic dual-domain joint loss is the core design of the SSPA initialization stage. This experiment validates the effectiveness of using the semantic loss and the syntactic loss individually and in combination and compares the performance of fixed weights versus adaptive weights. The experimental results are shown in Table 17, which presents the model performance, terminology accuracy, and grammatical compliance rate under different loss function configurations.

When only a single semantic loss or syntactic loss is adopted, the comprehensive performance, terminology accuracy, and grammatical compliance rate all exhibit obvious deficiencies. After adopting the dual-domain loss with fixed weights, all metrics improve to some extent, while the adaptive-weight dual-domain loss designed in this paper achieves the best results: BLEU-4 reaches 31.4, terminology accuracy reaches 78.9%, and grammatical compliance rate reaches 91.3%. This is because the adaptive weights can dynamically balance semantic and syntactic constraints according to the training status, taking into account both terminological semantic matching and Tibetan syntactic norms.

4.3.3. Sensitivity Analysis of Hyperparameters

The maximum number of modified words, denoted as $ϵ$, limits the number of terms allowed to be replaced in a single sentence and directly affects the grammatical fidelity and domain relevance of the generated pseudo-corpora. Experiments are conducted for $ϵ=1, 2, 3, 4$, and the performance variations are shown in Table 18. As $ϵ$ gradually increases, the coverage of terminology improves, but term collocation and syntactic integrity are progressively impaired, leading to a trend in terminology accuracy that first rises and then declines. At $ϵ=1$, only a single word is replaced each time, resulting in insufficient introduction of terminology, where some low-frequency terms cannot be effectively learned, and terminology accuracy remains low. At $ϵ=2$, multiple related terms can be reasonably combined while preserving the original syntactic framework, yielding logically coherent term collocations and reaching peak accuracy. When $ϵ$ increases to 3 or 4, excessive words are arbitrarily replaced, which tends to cause term stacking and case-marker mismatches, introducing errors in term semantics and grammatical coherence, and consequently, accuracy continues to decrease.

Multi-scale semantic pooling ensures semantic consistency of pseudo-corpora at word, phrase, and sentence levels through feature alignment at different granularities. The effects of different pooling scale combinations are compared, and the results are shown in Table 19. Under different multi-scale semantic pooling combinations, terminology accuracy generally increases first and then slightly decreases as the completeness of semantic hierarchy coverage improves. When the pooling scale is small, the model can only capture word-level features and struggles to understand the correlations and collocations among terms, resulting in low terminology accuracy. As the scale gradually increases to cover phrase and short-sentence levels, the model can accurately recognize domain-specific terms and their combinatorial relationships, and terminology accuracy continues to increase to a peak. When the pooling scale becomes too large, context-independent redundant information may be mixed in, interfering with term-semantic judgments and leading to a slight drop in accuracy.

Inspired by expectation-over-transformation frameworks originally used in computer vision for adversarial robustness (e.g., simulating scaling and rotation), we innovatively migrated this mechanism to syntactic template generation. To prevent the generated pseudo-corpora from becoming overly monotonous in sentence structure and inducing template overfitting, SSPA defines a domain-specific writing transformation set with explicit physical meanings: (1) Particle Fine-Tuning simulates spelling variants of function words in Tibetan medical literature across different historical periods; (2) Modifier Addition and Deletion dynamically adds or removes adjectival components outside the core syntactic backbone; (3) Word Order Fine-Tuning shuffles local word order within the bounds allowed by Tibetan inversion rules. This module does not perform unconstrained random sequence shuffling; instead, it utilizes this structural perturbation matrix exclusive to Tibetan medicine to achieve the highest level of style robustness under the premise of strictly keeping the syntactic skeleton baseline. The effects of individual transformations and their combination are tested separately, and the results are shown in

Table 20. Effectiveness of individual writing transformations in EOT regularization.

Composition of the Transformation Set	BLEU-4	TER	CHRF++	Domain Style Similarity
without EOT	33.8	43.5	54.7	0.79
only H1 (Particle Fine-Tuning)	34.5	42.7	55.4	0.82
only H2 (Modifier Addition/Deletion)	34.9	42.3	55.8	0.83
only H3 (Word Order Fine-Tuning)	34.2	43.0	55.1	0.81
H1+H2+H3 (Full Transformation Set)	36.2	40.1	57.3	0.87

.

4.3.4. Computational Overhead and Architectural Independence Analysis

A core advantage of the SSPA framework is its architectural decoupling from the translation model topology. While data generation is optimization-coupled, SSPA introduces no permanent architectural alterations. To validate this,

Table 21. Comprehensive analysis of offline generation overhead and end-to-end efficiency.

Total Scale of Generated Pseudo-Pairs	Cumulative Preprocessing (s)	Cumulative Parser (s)	Cumulative Gradient Optimization (s)	Cumulative Postprocessing (s)	End-to-End Total Time (s)	Total Time Equivalent (Minutes)
500 pairs	2.40	3.85	92.50	7.75	106.50	∼1.77
1000 pairs	4.80	7.70	185.00	15.50	213.00	∼3.55
5000 pairs	24.00	38.50	925.00	77.50	1065.00	∼17.75
10,000 pairs	48.00	77.00	1850.00	155.00	2130.00	∼35.50

compares the training time per epoch and the inference latency of SSPA against model-based domain adaptation methods (SemiAdapt-LoRA and Knowledge Distillation).

A core functional advantage of the proposed SSPA framework lies in its complete architectural decoupling from the translation model architecture during the critical inference deployment phase. To provide a mathematically transparent quantification of the resource-intensive offline preprocessing stage, Table 21 completely deconstructs the computational overhead. The total offline generation time of the SSPA pipeline for 5000 pairs is strictly controlled at approximately 17.75 min.

To explicitly address the micro-level generation latency, the system executes exactly 50 iterations per sentence (5 outer iterations nesting 10 inner continuous embedding space optimizations). Under a parallel throughput with a batch size of 32, the detailed average latency breakdown per sentence is as follows: 4.8 ms for data preprocessing, 7.7 ms for the Bi-LSTM+CRF parser cost, 185.0 ms for gradient alignment optimization, and 15.5 ms for discrete postprocessing, culminating in a total of 213.0 ms per pair. This extreme decoupling design allows each complex dual-domain gradient iteration to be completed in less than 4 milliseconds, avoiding the autoregressive generation bottleneck. When this localized offline computational investment is amortized, the inference latency remains at 42 ms per sequential token generation—mathematically and empirically identical to the baseline unmodified architecture—thereby validating its potential for practical, real-world deployment in extremely resource-constrained specialized domains.

4.3.5. Robustness Under Extreme Data Scarcity

To further highlight the significant improvement of the SSPA explicit rule-injection mechanism in extreme scenarios,

Table 22. BLEU-4 scores across extreme low-resource scales.

Method	100 Pairs	300 Pairs	500 Pairs	1000 Pairs
SemiAdapt LoRA	22.4	24.8	27.5	30.2
Knowledge Distillation	23.1	26.3	29.2	32.5
SSPA (Ours)	31.8	34.1	36.2	37.9

evaluates the models when the authentic Tibetan medicine corpus is radically reduced from 1000 to 100 pairs.

Neural-based adaptation methods (SemiAdapt-LoRA and Knowledge Distillation) experience severe performance degradation (dropping to 22.4 and 23.1, respectively, at 100 pairs) because neural weights cannot effectively capture data distributions from such microscopic samples. Conversely, SSPA maintains a highly competitive BLEU-4 score of 31.8 even with only 100 pairs. This confirms that SSPA circumvents the neural data-hunger bottleneck by directly functioning at the linguistic template level, achieving exceptional robustness under extreme data scarcity.

4.3.6. Removal-Based Ablations of the Complete SSPA Model

To deeply investigate the independent contribution and multi-objective coupling mechanism of each core module within the fully converged SSPA framework, we systematically conducted a removal-based ablation study. By stripping specific components or loss functions one by one from the complete model, we evaluated the performance fluctuations across multiple fine-grained dimensions, including BLEU-4, TER, CHRF++, Terminology Accuracy, and parser-verified Grammatical Compliance Rate. The comprehensive experimental results are presented in

Table 23. Comprehensive results of removal-based ablation experiments on the complete SSPA model based on the TM-TE test set.

Model Variant	BLEU-4 ↑	Relative $\mathrm{\Delta }$ BLEU Loss	TER ↓	CHRF++ ↑	Term Accuracy (%) ↑	Grammatical Compliance (%) ↑
Full SSPA (Complete Model)	36.2	-	40.1	57.3	85.4	96.2
w/o Semantic Loss	32.1	−4.1	45.2	51.8	72.1	88.3
w/o Syntactic Loss	32.5	−3.7	44.6	52.6	79.5	76.4
w/o Dynamic Length Adapt.	33.8	−2.4	43.5	54.7	81.2	91.5
w/o EOT Regularization	33.8	−2.4	43.5	54.7	82.6	92.8
w/ Fixed Weights	34.6	−1.6	42.1	55.2	80.8	92.1
w/ Random Templates	29.8	−6.4	47.1	49.8	66.3	68.2

Higher values are preferable for features indicated with ↑, while lower values are preferable for features indicated with ↓.

.

Severe Lexical Drift from Stripping Semantic Loss. Removing the total semantic loss (multi-scale MSE and relative semantic convergence loss) results in a substantial BLEU-4 drawdown of 4.1 points, which is accompanied by a severe decline in Term Accuracy from 85.4% to 72.1%. This striking contrast exposes a hidden trap in low-resource professional translation: structural similarity does not equate to semantic accuracy. Without multi-level semantic constraints, the gradient descent direction is easily hijacked by highly frequent general function words (e.g., conjunctions and prepositions) during text-to-text alignment iteration, leading to the neglect of highly condensed tetrasyllabic terminologies that carry crucial medical information. Lacking the relative ratio structure designed to penalize alignment deviations for rare words, the model seemingly reaches a stable fitting state in macroscopic sentence patterns but suffers severe random hallucinations and lexical drift during the filling of disease and drug names, which massively pollutes the memory matrix of the cross-attention mechanism.

Agglutinative Dependency Collapse Caused by Stripping Syntactic Loss. Correspondingly, stripping the syntactic constraint loss undermines the skeleton of the Tibetan text, triggering a substantial drop in the Grammatical Compliance Rate from 96.2% to 76.4%, alongside a 3.7 BLEU drop. This result is deeply rooted in the grammatical core of Tibetan as an agglutinative language, where profound semantic expressions rely almost entirely on a rigorous postpositional case marker system. These morphological suffixes act like glue attached to the core noun or verb roots. Maximizing syntactic cosine similarity constructs a robust structural defense; removing it immediately degrades the augmentation into blind positional replacement. Forcing new rare words into old physical coordinates directly tears apart tightly cascaded affix bindings, resulting in detached case particles and dislocated long-distance head-modifier relations. Consequently, the Transformer decoder’s autoregressive generation tends to produce incoherent text when facing complex subordinate clauses.

The Rigid Dilemma of Stripping Dynamic Length Adaptation and Style Regularization. Stripping the dynamic length adaptation mechanism degrades BLEU-4 by 2.4 points. Authentic Tibetan medical classics span extremely wide sentence lengths; without automated subset dismantling and combining capabilities, the algorithm relies on rigid static templates for truncation or massive meaningless padding. This mechanical operation invisibly weakens the continuity of long-range dependency information in lengthy sentences. Similarly, removing the EOT-style regularization module leads to a 2.4 BLEU sacrifice. Because SSPA relies entirely on only five expertly annotated core syntactic frames, the generated 5000 bilingual pairs, without EOT, inevitably converge into extremely concentrated feature clusters, inducing severe template collapse and overfitting. By introducing controlled particle fine-tuning and modifier addition/deletion, EOT artificially simulates the diverse writing variants employed by Tibetan medical scholars from different schools or eras, injecting robust noise essential for bridging the domain style generalization gap.

Infeasibility Reflections on Fixed Weight Parameters and Random Templates. Degrading the adaptive dual-domain weighting to a fixed absolute average allocation incurs a 1.6 BLEU loss. Syntactic frame construction and fine-grained morpheme sculpting require dynamic balancing: early parameter initialization urgently requires strong structural penalties, while the final fine-tuning phase must assign higher weights to morpheme-level semantic errors. The mandatory fixed ratio causes the gradient direction to oscillate, dragging down limit convergence capability. Most strikingly, replacing the expertly curated domain templates with random templates extracted from a general news corpus triggers a full-scale performance drop, dropping the BLEU score by 6.4 points and reducing the grammatical compliance rate to 68.2%. This strongly validates our core hypothesis: the insurmountable barrier in Tibetan medicine lies deeply within its underlying syntactic reorganization logic. Professional literature prefers highly structured parallel and antithetical sentence patterns, as well as complex conditional adverbial inversion structures. Generating texts using loose modern political news frameworks to forcefully incorporate complicated medical terminology creates an incongruous text generation that suffers from severe domain style mismatch and yields rigid translationese, thereby confirming the absolute necessity of linguistically curated domain priors.

4.3.7. Generalization Analysis via Leave-One-Template-Out Evaluation

Data augmentation techniques often face the risk of overfitting in low-resource tasks. Given that the SSPA framework relied on five typical syntactic templates from the Tibetan medical domain during initialization, it is crucial to determine whether the model achieved genuine generalization or merely fell into template memorization. To rigorously establish this boundary, we implemented a Leave-One-Template-Out (LOTO) cross-validation evaluation.

We conducted five completely independent runs of the pseudo-corpus generation pipeline. In the i-th run, the generation engine was physically deprived of access to template $T_i$, performing dynamic syntactic adaptation and dual-domain optimization on only the remaining four templates to generate 5000 restricted pseudo-parallel sentences. Linguistic experts annotated 100 authentic bilingual sentences strictly isomorphic to the excluded template to construct five mutually exclusive zero-shot structural test sets (LOTO-1 to LOTO-5). The downstream CINO Transformer model was then retrained from scratch and evaluated strictly on the corresponding unseen structures. The core evaluation metrics are detailed in

Table 24. SSPA translation performance comparison under LOTO strict isolation environment.

Fold (Hidden Template)	Baseline CINO (No Augmentation)	SSPA (All Templates)	SSPA (LOTO Isolated)	Performance Degradation	Absolute Net Improvement
Fold 1	BLEU-4: 21.8 TER: 59.1	BLEU-4: 35.8 TER: 41.2	BLEU-4: 34.2 TER: 43.1	BLEU-4: −1.6 TER: +1.9	BLEU-4: +12.4 TER: −16.0
Fold 2	BLEU-4: 22.4 TER: 57.8	BLEU-4: 36.5 TER: 39.8	BLEU-4: 33.9 TER: 42.7	BLEU-4: −2.6 TER: +2.9	BLEU-4: +11.5 TER: −15.1
Fold 3	BLEU-4: 22.1 TER: 58.4	BLEU-4: 36.1 TER: 40.5	BLEU-4: 34.5 TER: 41.8	BLEU-4: −1.6 TER: +1.3	BLEU-4: +12.4 TER: −16.6
Fold 4	BLEU-4: 21.5 TER: 59.8	BLEU-4: 35.6 TER: 41.9	BLEU-4: 33.1 TER: 44.2	BLEU-4: −2.5 TER: +2.3	BLEU-4: +11.6 TER: −15.6
Fold 5	BLEU-4: 22.7 TER: 57.2	BLEU-4: 36.9 TER: 39.0	BLEU-4: 34.8 TER: 41.5	BLEU-4: −2.1 TER: +2.5	BLEU-4: +12.1 TER: −15.7
Average 5-Fold	BLEU-4: 22.10 TER: 58.46	BLEU-4: 36.18 TER: 40.48	BLEU-4: 34.10 TER: 42.66	BLEU-4: −2.08 TER: +2.18	BLEU-4: +12.00 TER: −15.80

.

As observed, when evaluating on completely hidden zero-shot syntactic structures, the pseudo-corpus generated by SSPA causes a slight performance degradation compared to the ideal environment (average BLEU-4 dropped by 2.08 points). This conforms to the basic laws of Out-of-Distribution (OOD) generalization. However, crucially, the SSPA-augmented model still achieved an overall average net improvement of 12.00 BLEU-4 points and a 15.80 percentage point reduction in TER over the unaugmented baseline.

To ensure this improvement is statistically robust, we introduced Paired Bootstrap Resampling ($B=1000,\alpha =0.05$) via SacreBLEU. The null hypothesis assumed no significant difference between the SSPA and baseline models on unknown structures. The empirical results demonstrated that the p-values for SSPA outperforming the baseline strictly met the extremely high significance level of $p<0.01$ across all five independent LOTO experiments. These robust statistical results explicitly refute the template memorization hypothesis, suggesting that the SSPA pipeline captures broadly applicable linguistic rules within the evaluated domains.

5. Conclusions

We propose the Semantic-Syntax Prealignment (SSPA) framework for pseudo-parallel corpus generation to improve low-resource Tibetan-to-English machine translation in the medical domain. This framework mitigates the grammatical errors, semantic misalignment, and style mismatches prevalent in traditional data augmentation methods. SSPA employs a two-stage pipeline. First, an initialization module models Tibetan case particle concatenation and syntactic dependencies via dual-domain alignment. Second, a generation module uses cross-length dynamic filling and Expectation-over-Transformation (EOT) style regularization to simulate writing variations. We constructed the Tibetan Medicine-Tibetan-to-English Bilingual Dataset (TM-TE), which contains 9166 professional terms and 5 typical syntactic frames for algorithm evaluation. We evaluated SSPA on TM-TE and two public datasets (CCMT2020, CCMT2023) against seven baseline methods. Experimental results demonstrate that the SSPA framework effectively enhances translation quality in the low-resource Tibetan medicine domain, achieving a BLEU-4 score of 36.2 and a grammatical compliance rate of 96.2%. Furthermore, the framework improves domain-specific translation performance while maintaining baseline metrics in general domains, exhibiting robust cross-domain variance control. Its plug-and-play architecture also offers computational efficiency, indicating that gradient-guided data generation can be implemented without permanently altering target model topologies. Furthermore, while the statistical significance and Leave-One-Template-Out (LOTO) evaluations provide robust evidence of generalization within the tested distributions, we acknowledge that these results do not prove universal robustness across all conceivable or extreme linguistic structures. Exploring the framework’s stability under unconstrained, open-domain morphological variations remains an important direction for future research.

Although SSPA demonstrates strong performance, its scalability is currently constrained by a reliance on curated domain-specific syntactic templates and terminology dictionaries, which necessitate expert intervention. Future research will focus on:

1.

Developing unsupervised anchor induction techniques to automatically extract syntactic frames from unannotated corpora.

2.

Extending the prealignment mechanism to other low-resource agglutinative languages and professional domains to validate the framework’s broader applicability.