The application of intelligent generation model for international discourse of grand canal culture based on artificial intelligence and BPNN model Authors: Haiyan Gu, Jie Bai, Jun Xie, Cui Zhang Published: 2026 ABSTRACT AbstractChina’s Grand Canal, as an ancient artificial canal with the largest engineering scale and the longest mileage in the world, faces problems such as inconsistent terminology translation, a lack of cultural context, and low efficiency of artificial generation in its international cultural dissemination. In view of the above problems, this study proposes an intelligent generation model for the international discourse of Grand Canal culture based on Backpropagation Neural Network (BPNN). This study has built a special Chinese-English parallel corpus for the Grand Canal culture. The original data contains about 150,000 candidate sentence pairs. After strict cleaning, filtering, and manual verification, 120,000 high-quality parallel sentence pairs are finally obtained for model training and evaluation. On this basis, a three-layer feedforward network architecture is designed; momentum optimization and adaptive learning rate decay strategies are introduced. The proposed model is realized by software simulation. Based on the PyTorch deep learning framework, all experiments are completed on a NVIDIA Tesla V100 graphics processing unit. The model evaluation adopts three automatic indicators (namely Bilingual Evaluation Understudy (BLEU), Recall-Oriented Understudy for Gisting Evaluation (ROUGE), and Translation Edit Rate (TER)), and manual assessment. The experimental results show that the proposed model has a BLEU-4 score of 0.438, a ROUGE-L score of 0.592, and a TER of 0.385. In the manual assessment, this model’s cultural fidelity score is 4.1 (out of 5). The above results have improved by 54.2% and 21.1% compared with Phrase-Based Statistical Machine Translation (PBSMT) on BLEU-4 and Google Translate on ROUGE-L. The research results verify the applicability of the optimized BPNN in the vertical cultural field; concurrently, it provides an explorable technical path for improving the accuracy and efficiency of the international communication of Chinese cultural heritage.Keywords: Grand Canal culture, International discourse system, BPNN, Cross-cultural communication, Intelligent translationSubject terms: Engineering, Mathematics and computing Introduction The Beijing-Hangzhou Grand Canal, as an ancient artificial canal with the longest mileage and the largest engineering scale in the world, still faces severe challenges in its international cultural communication. The existing international discourse system mostly relies on traditional manual translation and static text display; there are deficiencies such as poor contextual adaptability, different terminology translation standards, and a lack of historical allusions1–3. For example, culturally specific items with profound Chinese characteristics, such as “canal transportation”, “sluice dam”, and “water engineering wisdom”, often lose their meanings due to cultural default in cross-language conversion; this makes it difficult to arouse emotional resonance and deep cognition among international audiences. Therefore, how to use cutting-edge technology to break through language barriers and realize the leap of the Grand Canal culture from “physical connection” to “discourse connection” has become a key issue to be solved urgently in the field of cultural heritage protection and international communication. The swift development of Artificial Intelligence (AI), Natural Language Processing (NLP), and Natural Language Generation (NLG) technologies offers a novel solution for cross-cultural communication4–6. Among many neural network architectures, the backpropagation neural network (BPNN), as a classic multi-layer feedforward network, is seldom mentioned in the era of general large models; however, it still shows its unique advantages in small-sample and high-precision tasks in specific vertical fields7–9. BPNN constantly adjusts the network weight and bias through the error back propagation algorithm; besides, it can effectively fit complex semantic logic and stylistic features in Grand Canal cultural texts. Compared with large models such as Long Short-Term Memory (LSTM) networks or Transformers, the optimized BPNN has the characteristics of a relatively simple structure, fast training convergence speed, strong interpretability, and low consumption of computing resources10–12. It is particularly suitable for building a dedicated discourse generation system for the specific cultural field of the Grand Canal; meanwhile, it can realize efficient and intelligent content production under limited computing power conditions. This study aims to explore and construct an intelligent generation framework for Grand Canal culture international discourse based on the BPNN model. Thus, it can solve the issue of low efficiency and cultural distortion in current international communication and verify the model’s performance in all directions. This study expects to prove the application potential of BPNN in the vertical culture field at the technical level; it also devotes itself to improving the accuracy and appeal of the international communication of Grand Canal culture through intelligent discourse generation means. As a result, a replicable and promotable technical paradigm and practical path can be offered for the “going global” of excellent traditional Chinese culture. Literature review In recent years, research on the international communication of Grand Canal culture has shifted from macroscopic strategic discussions to micro-textual empirical analysis. Scholars are generally concerned about how to overcome the phenomenon of “cultural discount” to achieve effective cross-cultural communication13. Dong et al. (2024) pointed out a serious problem of “terminology alienation” in current Grand Canal publicity texts. In addition, traditional manual translation had difficulty unifying international expressions of core concepts such as “canal transportation system”, which led to the cognitive fragmentation of international audiences14. Lin et al. (2026) further found a gap between static museum commentaries and dynamic digital narratives, and a lack of a generative mechanism that could adaptively adjust discourse style based on audience background15. Wang et al. (2024) emphasized that existing communication modes primarily focused on the listing of historical facts. They ignored the Grand Canal’s emotional dimension as living heritage, resulting in a weak appeal in international discourse16. These studies collectively reveal the limitations of traditional communication methods in handling dense cultural information; it is urgent to introduce intelligent technologies to reconstruct the international discourse system and effect a paradigm shift from one-way indoctrination to intelligent interaction. In the field of technology-enabled cultural heritage communication, the NLG model based on deep learning has become a research hotspot. Still, model optimization for specific vertical cultural fields remains challenging17–19. Badawi et al. (2023) compared various neural network architectures and found that transformer models performed well on general translation tasks20. However, their generalization ability was often limited in cultural heritage domains, where technical terms were dense, and training data were relatively scarce. Such models also carried risks of hallucinatory generation. In contrast, owing to its compact structure and fast convergence, BPNN’s feature-fitting ability on small- and medium-sized datasets cannot be ignored. Mienye et al. (2024) showed that the improved BPNN’s semantic consistency by introducing a momentum factor and an adaptive learning rate strategy was better than that of the basic recurrent neural network (RNN) when handling historical texts with strong logical correlations21. Alomar et al. (2025) proposed a hybrid neural architecture; their research demonstrated that simplified feedforward networks combined with attention mechanisms could effectively reduce computational complexity and maintain high translation accuracy in a specific context22. This provided solid theoretical support for the adoption of an optimized BPNN model in this study. The existing literature has made remarkable progress in the application of algorithms and cultural communication strategies; nevertheless, the deep integration of the two and the construction of an end-to-end intelligent generation system specifically for the Grand Canal culture remains a blank. Vishwakarma (2023) clearly pointed out in the latest review that it was necessary to establish a large model in the vertical domain or a special small model including a special corpus, customized algorithms, and multi-dimensional evaluation indicators; this can process the general model’s shortcomings in dealing with cultural nuances23. Most of the existing studies have not disclosed the mathematical derivation process at the model’s bottom in detail, especially the specific parameter updating mechanisms in cultural semantic mapping are unknown. Therefore, this study aims to fill this gap and systematically verify its unique value in improving the generation quality, fidelity, and communication efficiency of international discourse by constructing a special corpus for the Grand Canal culture. Thus, a new academic perspective and technical path can be offered for AI to empower the international communication of cultural heritage. Research model Network architecture and forward propagation mechanism In this study, BPNN is constructed as a three-layer feedforward structure; it is used for the post-optimization task of preliminary translation. The core challenge of Grand Canal culture communication is terminology standardization and cultural fidelity enhancement. This can be expressed as a nonlinear classification and mapping problem, and a three-layer architecture is enough. The three-layer BPNN proposed here is not a complete end-to-end sequence-to-sequence translation system. The complete machine translation task involves many complex links (e.g., variable-length sequence coding, context-dependent modeling, decoding generation, etc.); it usually needs a specially designed architecture (such as the recurrent neural network (RNN) or Transformer) to deal with it. However, in the vertical field of international communication of Grand Canal culture, the core difficulty is not general translation ability, but the following two specific issues: First, terminology standardization. The text of Grand Canal culture contains a large number of culturally specific items with Chinese characteristics, such as “canal transportation”, “sluice dam,” and “water engineering wisdom”. The translation of these terms by existing universal translation systems is often inconsistent, and the same Chinese term may be translated into many different English expressions in various contexts; this leads to the cognitive fragmentation of international readers. Second, the cultural fidelity has improved. In addition to terminology, the text of Grand Canal culture also contains rich historical allusions and emotional colors. A universal translation system tends to translate word-for-word, which results in the loss of historical depth and cultural connotation of the original text in the translation. In view of the above two problems, this study redefines the task as improving the cultural communication effect of the translation through terminology standardization and style optimization based on the existing preliminary translation. This task can be modeled as a nonlinear classification and mapping problem rather than a complete sequence generation issue. Specifically, given a fixed-length word-vector window (including the translation fragments to be optimized and their contextual features), the model needs to output the category labels of standardized terms or the optimized local translation representations. A three-layer BPNN is just suitable for this task because of its simple structure, fast convergence, and strong nonlinear fitting ability. The forward propagation process is a key stage for information to flow from the input to the output end. The input vector is first linearly weighted and summed with the weight matrix; then, it is processed by an activation function to extract deep semantic features, and finally, the prediction results are generated at the output layer. This process not only determines the model’s expressive ability but also is a basis for subsequent error calculation. Its core lies in accurately capturing the complex contextual associations of culturally specific items such as “canal transportation” and “water engineering” through the collaborative effect of multi-layer neurons. Specifically, the word vectors received by the input layer are first transmitted to the hidden layer. Each hidden layer neuron performs a weighted summation of all input features, and this linear transformation process is shown in Equation (1): and represent the j-th neuron’s net input value and bias term in the hidden layer; the latter is used to adjust the activation threshold. xi denotes the i-th neuron’s input value in the input layer, that is, the i-th dimensional feature of the word vector of Grand Canal texts. is a weight coefficient connecting the i-th and j-th neurons in the input and hidden layer, respectively. n refers to the total number of neurons in the input layer, corresponding to a word vector dimension. j means the hidden layer neuron’s index number. In order to endow the model with the ability to handle nonlinear semantic relationships in Grand Canal culture, the hidden layer’s net input must be transformed through a nonlinear activation function. This study selects the Sigmoid function24; it can compress the net input in any real number domain into the interval (0,1), simulating the excitation and inhibition states of neurons. Thereby, it extracts high-order features of abstract concepts such as “millennium ancient charm” and “water conservancy wisdom”. Its calculation logic can be written as Eq. (2): hi denotes the j-th neuron’s final output value in the hidden layer. e is the base of the natural logarithm; represents a net input of the hidden layer. After feature extraction by the hidden layer, the signal hj continues to propagate forward to the output layer. Similar to the hidden layer, the output layer neurons first perform a weighted summation of the signals from all hidden layer nodes to form the output layer’s net input. This step completes the linear mapping from the implicit semantic space to the target vocabulary space, and the specific mathematical expression is as follows: and stand for the k-th neuron’s bias term and net input value in the output layer; refers to the weight coefficient connecting the j-th neuron in the hidden layer and the k-th neuron in the output layer; m represents the total number of neurons in the hidden layer. K represents the output layer neuron’s index number. Finally, to convert the output layer’s net input into a probability distribution with clear physical meaning for generating specific English words, the model uses the Softmax function25 to normalize the net input. This function ensures that the sum of the output probabilities of all candidate words is 1; thus, this model can output the most likely translation of Grand Canal cultural terms in the form of probability, as denoted in Eq. (4): denotes the probability value output by the k-th neuron in the output layer, indicating the possibility of being predicted as the k-th word in the target language vocabulary. L represents the total number of neurons in the output layer, that is, the size of the target language vocabulary. p is the summation index, traversing all output layer neurons. refers to a net input of the output layer. After obtaining the probability distribution of the output layer, the model needs to select specific output words from it. Because the proposed model’s task is local terminology standardization and fragment optimization, rather than complete sentence generation, greedy search is adopted as the decoding strategy. That is, every step selects the word with the highest probability as the output. The computational overhead of greedy search is small, and it has achieved satisfactory results in this task. For long sentence generation scenarios that need global optimization, more complex decoding strategies (such as beam search) can be considered, and this study lists it as future work. Given the probability distribution yk of the output layer, the output word w∗ of greedy search can be expressed as: L represents the size of the target language vocabulary list. The BPNN’s three-layer network structure is presented in Fig. 1: Three-layer network structure of BPNN. Loss function construction and error backpropagation This study constructs a loss function based on cross-entropy to quantify the discrepancy between the Grand Canal cultural discourse generated by the model and the standard reference translations, thereby guiding the optimization direction of network parameters. This function effectively measures the distance between the predicted probability distribution and the true label distribution. A smaller value indicates a more accurate generated translation. After the total error is calculated, the error backpropagation algorithm uses the chain rule in calculus to derive the output layer’s error gradient backward to the hidden and input layers. The core of this process lies in computing the partial derivatives (i.e., gradients) of each weight and bias with respect to the total error, thus clarifying the direction and magnitude of parameter adjustment. This mechanism guarantees that the model continuously reduces prediction bias when handling complex cultural concepts such as the “canal transportation system”. First of all, for a single training sample, the model calculates the cross-entropy loss between a predicted probability yk and a true label tk, as follows: E represents the current single sample’s total error value; tk means the k-th value in the target vector, which is 1 if the correct target word is k and 0 otherwise. In actual training, in order to maintain gradient stability, the average loss of all samples in a mini-batch is usually used as the final optimization objective. This setting helps smooth random noise and makes model convergence more stable26, as indicated in Eq. (7): J(w,b) denotes the average loss of the entire mini-batch data and is the objective function for optimization. N represents the number of samples contained in the mini-batch. Es denotes the s-th sample’s error value. w and b refer to the set of all weights and biases in the model. Based on the average loss, the backpropagation process first calculates an error term for the output layer neurons. This error term represents the loss function’s sensitivity to the output layer’s net input and provides a direct basis for subsequent weight updates. Its simplified calculation form is: represents the k-th neuron’s error gradient term in the output layer. Subsequently, the error term is back-propagated to the hidden layer through a connection weight. The error term of hidden layer neurons depends not only on the error transmitted from the subsequent layer but also on the derivative of its own activation function. This process reflects the error distribution mechanism between layers, as demonstrated in: represents the j-th neuron’s error gradient term in the hidden layer. denotes the weight connecting the hidden and output layers. represents the Sigmoid activation function’s derivative, reflecting the neuron’s sensitivity under the current input. After obtaining the error terms of each layer, the partial derivative of the loss function with respect to the output layer weights can be calculated. This gradient indicates the direction and rate at which the weight should be adjusted to reduce the error. The calculation expression is shown as follows: represents the partial derivative of the average loss weight . hj,s and denote the hidden layer output and output layer error term corresponding to the s-th sample. Similarly, the partial derivative of the loss function to the output layer bias is only related to the error term because the bias’ input is always 1. Its gradient is calculated as follows: refers to the partial derivative of the average loss with respect to the bias . For the connection weights between the input and hidden layers, their gradient calculation depends on the hidden layer’s error term and the input layer’s original input; this reflects the contribution of input features to the final error: means the average loss with respect to the weight ’s partial derivative. xi,s and stand for the input layer input and the hidden layer error term corresponding to the s-th sample. Lastly, the gradient of the hidden layer bias is also determined by the mean of the hidden layer error terms, thus completing the gradient calculation of all trainable parameters, as given in Eq. (13): represents the partial derivative of the average loss with respect to the bias . Parameter update strategy and optimization algorithm After obtaining the gradients of all parameters, the model uses an optimization algorithm to iteratively update the weights and biases to gradually approach the global optimal solution. The traditional stochastic gradient descent (SGD) method tends to fall into local optima or generate oscillations27. Therefore, this study introduces SGD with momentum. The momentum term accelerates convergence in flat regions and effectively suppresses oscillations in canyon regions by accumulating the direction of historical gradients; this makes it particularly suitable for optimizing complex semantic surfaces in Grand Canal cultural texts. In addition, this study adopts a time-step-based adaptive learning rate decay strategy to avoid late convergence difficulties caused by a fixed learning rate. The step size is dynamically adjusted as the number of training epochs increases, ensuring rapid learning in the early training stage and fine-tuning in the later stage. The weight update rule is calculated as: wnew and wold indicate the weight values before and after the update. η represents the learning rate, which controls the current gradient update’s step size. denotes the currently calculated weight gradient. α represents the momentum factor used to retain the historical update direction. is the weight update amount in the previous iteration. The update of the bias term follows the same logic as the weights. The momentum mechanism28 smooths the update path and prevents model instability caused by gradient mutations of individual rare cultural words. Its update is calculated as follows: bnew and bold represent the bias values after and before the update; refers to the offset update amount in the previous iteration. In view of further improving the model’s generalization ability and convergence accuracy, the learning rate η is not fixed, but gradually decays with the increase of the number of training iterations t. This adaptive attenuation mechanism ensures that the model can quickly approach the optimal region with a large step length at the initial stage of training; meanwhile, it can finely search the optimal solution with a small step length at the later stage to avoid oscillation near the minimum value. Its dynamic adjustment is expressed in Eq. (16): ηt denotes the real-time learning rate at the t-th iteration; η0 is the initially set learning rate; λ represents the decay coefficient, controlling the rate at which the learning rate decreases; t refers to the current number of training iterations. The training and parameter update process of BPNN is plotted in Fig. 2: Training and parameter update process of BPNN. Variable-length input and context dependency processing Because the standard BPNN requires that the dimension of the input vector is fixed, it cannot directly process the natural language sequence of variable length. This study adopts the following two strategies to solve this problem. One is the sliding window mechanism. For the text whose length exceeds the window size, overlapping sliding windows are used for segmentation. This study is set to 300-word vector dimensions. The window step size is 150 words, and 50% overlapping area is reserved between adjacent windows to ensure the continuity of context information. Each window is independently input into the network for processing, and the output results are concatenated and restored to complete text according to the window order. The other is position coding and entity marking. In order to make up for the deficiency of sliding window mechanisms in capturing long-distance dependence, position coding and key entity markers are explicitly added to the input features. The position coding adopts the sine and cosine functions to make the model perceive the relative position information of words. Key entity tags extract entities such as time, place, and people from the text by a named entity recognition method; then, their binary features are concatenated into word vectors, so that the model can perceive the semantic association with long-distance entities in a fixed window. Experimental design and performance evaluation Datasets collection In order to ensure the reproducibility and reliability of the study, this part designs a systematic process to build a special Chinese-English parallel corpus for the Grand Canal culture. The whole construction process includes four stages: original data collection, noise filtering and cleaning, terminology standardization, and manual sampling verification. Stage 1: original data collection. The Chinese and English subsets in the public OPUS 100 dataset are selected as the core corpus. Domain-specific keywords such as “Grand Canal”, “canal transportation”, “Beijing-Hangzhou”, “sluice dam”, and “water conservancy” are used to retrieve parallel sentence pairs related to Grand Canal culture from massive data. In addition, the full bilingual text of the China Grand Canal World Heritage Assessment Report, publicly available on the official website of the United Nations Educational, Scientific, and Cultural Organization (UNESCO), is supplemented to enhance the field coverage. About 150,000 original candidate parallel sentence pairs are obtained in this initial collection stage. Stage 2: Noise filtering and cleaning. The original candidate sentence pairs are strictly cleaned, and invalid or low-quality items are eliminated. The following filtering criteria is applied in turn: (a) Delete sentence pairs whose source language or target language is empty or only contains punctuation marks; (b) Delete sentence pairs with the length ratio of source language/target language greater than 3:1 or less than 1:3; such abnormal ratio usually indicates alignment errors; (c) Delete sentence pairs containing HTML tags, unprinted characters or obvious coding errors; (d) De-duplication is performed based on the exact string matching between the source language and the target language. After this stage, there are about 132,000 sentence pairs left. Stage 3: Terminology standardization. Aiming at the inconsistency in the translation of core cultural concepts, such as “canal transportation”, “sluice dam”, and “water engineering wisdom”, a domain glossary covering 50 core terms is compiled manually. The preferred English translation of each term is determined according to authoritative sources. Subsequently, the table lookup method is used to automatically replace the cleaned corpus; this can ensure the consistent use of terms in all sentence pairs. At this stage, the number of sentence pairs is not reduced, but the data quality is markedly improved. Stage 4: Manual sampling verification. In order to verify the alignment quality and semantic fidelity, two bilingual experts with a research background in Grand Canal culture are invited as independent annotators; 1,000 sentence pairs are randomly selected from the standardized corpus for testing. Each sentence pair checks the following three aspects: (a) whether the sentence alignment is correct, that is, whether English has accurately translated the Chinese original; (b) whether there are obvious translation errors; (c) whether the use of terms conforms to the standardized glossary. The consistency of Cohen’s kappa between the two annotators is 0.96, and the overall alignment accuracy is 98.6%. Based on the results of this verification, about 12,000 pairs of questionable sentences that failed the manual test or had insufficient cultural relevance are excluded. Therefore, the final high-quality dataset contains 120,000 Chinese-English parallel sentence pairs. The 120,000 Chinese-English parallel sentence pairs after final verification are randomly divided into the training set (96,000), verification set (12,000), and test set (12,000) according to the ratio of 8:1:1. The training set is used for model parameter learning; the verification set is adopted for hyperparameter tuning and early stopping control; the test set is employed for final performance evaluation. All comparison models are trained and tested under the same data division to ensure fair comparison. Experimental environment The core computing task of this experiment is accelerated by a graphics processing unit (GPU); the deep learning framework integrates the Compute Unified Device Architecture (CUDA) parallel computing library to optimize the efficiency of matrix operations. This environment provides a stable bottom support for the large-scale iterative training of the Grand Canal cultural corpus. The specific system configuration, processor model, storage specifications, and key software versions are outlined in Table 1: Table 1Experimental environment settings.ComponentSpecification/VersionGPUNVIDIA Tesla V100 (32gigabytes (GB) video memory)Central processing unit (CPU)Intel Xeon Gold 6248R (20-core)Random access memory (RAM)256GB Double Data Rate 4 (DDR4) Error Correction Code (ECC)Hard disk2-terabyte (TB) NVMe Solid State Drive (SSD)Operating systemUbuntu 20.04 LTSDeep learning frameworkPyTorch 1.9.0 CUDA 11.1Programming languagePython 3.8.10 Experimental environment settings. Parameters setting The network topology in this study is set as a three-layer architecture, and the input dimensions align with the word vector space. The number of hidden layer nodes is determined through grid search to capture the deep semantic dependencies of the Grand Canal texts. The specific values for network layer scale, regularization coefficients, batch size, and iteration control thresholds are exhibited in Table 2: Table 2Hyperparameter settings.Parameter nameParameter valueInput layer dimension (n)300Number of hidden layer nodes (m)512Output layer dimension (L)25,000Initial learning rate (η0)0.01Momentum factor (α)0.9Decay coefficient (λ)0.001Batch size64Maximum number of iterations100Early stopping patience value5 Hyperparameter settings. Performance evaluation This study selects four baseline models for comparison. (1) Phrase-Based Statistical Machine Translation (PBSMT): Implemented with the Moses toolkit, using default phrase extraction and parameter tuning configurations, and a 5-gram language model trained on the training set. (2) Basic RNN: A single-layer RNN architecture is adopted; its hidden layer dimension is 512, using a cross-entropy loss function and a stochastic gradient descent optimizer with momentum. (3) Transformer: A standard Transformer architecture is used; it sets up 4 encoder layers, 4 decoder layers, and 8 attention heads; the hidden layer dimension is 512, and the feed-forward network dimension is 2048. (4) Google Translate: It is invoked via the official API and uses its default neural machine translation (NMT) engine. All trainable baseline models are trained and tuned on the same training and validation sets under the same hardware environment. All automatic evaluation indicators (such as Bilingual Evaluation Understudy (BLEU), Recall-Oriented Understudy for Gisting Evaluation (ROUGE), and Translation Edit Rate (TER) are calculated on the same test set of 12,000 samples. Among them, BLEU measures the similarity between machine-translated text and human reference text; its core is to calculate the matching accuracy of n-grams and conduct a length penalty for short translations. The value range is [0, 1], and the higher the value, the more similar it is to the reference text. ROUGE measures the coverage of key information in reference texts by machine-generated texts, and its core is the recall-based evaluation. The value also ranges from 0 to 1, and a higher value indicates more complete coverage of key information. TER reflects the proportion of editing operations required by professional translators to revise the output of machine translation. The lower the value, the more mature the translation, and the lower the cost of post-editing. Editing operations include four types: insertion, deletion, substitution, and shift. In order to eliminate the influence of randomness, each experiment is repeated five times, and the mean and standard deviation (SD) of the indicators are reported. For the proposed BPNN, five runs use different random seeds to initialize network weights. To verify whether the performance difference between the proposed model and each baseline model is statistically significant, paired bootstrap resampling is used to calculate the p-value after 10,000 resamples. When p < 0.05, the difference is considered statistically significant. (1) BLEU score comparison (machine translation accuracy). Figure 3; Table 3 present the performance of different models on the BLEU-1 and BLEU-4 indicators. Among them, BLEU-1 reflects word-level matching, and BLEU-4 reflects phrase-level matching. Appendix 1 shows the BLEU-4 scores of each model in independent runs. The results show that the BPNN model proposed by this study reaches 0.567 on BLEU-1 and 0.438 on BLEU-4; they are significantly better than all baseline models. The BLEU-4 score of BPNN is increased by 54.2%, 27.0%, 16.5%, and 12.6% compared with the PBSMT model (0.284), Basic RNN (0.345), Google Translate (0.376), and Transformer (0.389), respectively. Paired bootstrap resampling tests confirm that all improvements are statistically significant, and P values are less than 0.05. This result indicates that translations generated by the BPNN model are closer to human reference translations in word selection and phrase collocation; this effectively solves the rigid translation and word order disorder issues in traditional models. Comparison of BLEU scores. Table 3Comparison of BLEU scores among different models.ModelBLEU-1BLEU-4PBSMT0.412 ± 0.0080.284 ± 0.006Basic RNN0.498 ± 0.0070.345 ± 0.005Transformer0.531 ± 0.0060.389 ± 0.005Google Translate0.521 ± 0.0040.376 ± 0.004BPNN (Ours)0.567 ± 0.0050.438 ± 0.004Note: Values represent the mean ± SD of five independent runs. All differences between BPNN and baseline models are statistically significant (paired bootstrap resampling, p < 0.05). Comparison of BLEU scores among different models. Note: Values represent the mean ± SD of five independent runs. All differences between BPNN and baseline models are statistically significant (paired bootstrap resampling, p < 0.05). (2) ROUGE score comparison. Figure 4; Table 4 depict the performance of different models on the ROUGE indicators. ROUGE-1 measures unigram recall; ROUGE-L measures longest common subsequence coverage, reflecting the integrity of semantic logic chains. The results show that the proposed model achieves 0.612 and 0.592 on ROUGE-1 and ROUGE-L, respectively, both significantly outperforming all baseline models. Notably, the proposed model’s ROUGE-L score is 21.1% and 6.9% higher than that of Google Translate (0.489) and Transformer (0.554). This demonstrates that the BPNN model can fully retain core logical elements (such as time, location, and characters) in Grand Canal historical events when generating long sentences, avoiding severe information loss. The high ROUGE-1 score further proves that the proposed model successfully retrieves most key content words; this guarantees the completeness and information density of international discourse. Comparison of ROUGE scores. Table 4Comparison of ROUGE scores for different models.ModelROUGE-1ROUGE-LPBSMT0.478 ± 0.0070.442 ± 0.006Basic RNN0.545 ± 0.0060.512 ± 0.005Transformer0.581 ± 0.0050.554 ± 0.005Google Translate0.521 ± 0.0040.489 ± 0.004BPNN (Ours)0.612 ± 0.0050.592 ± 0.004Note: Values refer to the mean ± SD of five independent runs. All differences between BPNN and baseline models are statistically significant (p < 0.05). Comparison of ROUGE scores for different models. Note: Values refer to the mean ± SD of five independent runs. All differences between BPNN and baseline models are statistically significant (p < 0.05). (3) TER score comparison. Figure 5; Table 5 show the performance of various models on the translation edit rate (TER) indicator. This indicator reflects the proportion of editing operations required by professional translators to revise machine translation outputs; a lower value indicates more mature translations and lower post-editing costs. Editing operations include shift, substitution, insertion, and deletion. TER score comparison. Table 5Comparison of TER scores of diverse models.ModelTEREditing operation ratio (%)PBSMT0.562 ± 0.00956.2Basic RNN0.468 ± 0.00746.8Transformer0.412 ± 0.00641.2Google Translate0.421 ± 0.00542.1BPNN (Ours)0.385 ± 0.00538.5Note: Values are the mean ± SD of five independent runs. All differences between BPNN and baseline models are statistically significant (p < 0.05). Comparison of TER scores of diverse models. Note: Values are the mean ± SD of five independent runs. All differences between BPNN and baseline models are statistically significant (p < 0.05). The proposed model’s TER score is 0.385; this means that translators only need to adjust approximately 38.5% of the content to meet publication standards. Compared with the PBSMT model (0.562), Basic RNN (0.468), Google Translate (0.421), and Transformer (0.412), the proposed model reduces editing workload by 31.5%, 17.7%, 8.6%, and 6.6%, respectively. This result shows that translations generated by the proposed model have high maturity in grammatical structure and terminology usage; this markedly reduces post-editing costs and demonstrates high practical application value. (4) Manual assessment. In order to further verify the model’s advantages in cultural communication, this study designs a standardized manual assessment process. First, a total of six evaluators were recruited and divided into three groups. Group 1 includes two bilingual scholars specializing in Grand Canal culture, and they have more than five years of Grand Canal cultural heritage research experience. Group 2 consists of two professional translators; both of them hold Level 1 China Accreditation Test for Translators and Interpreters certificates and have more than five years of Chinese-English translation experience. Group 3 is two native English speakers, both with Chinese culture research backgrounds and Hanyu Shuiping Kaoshi (HSK) 6 proficiency. This composition aims to comprehensively reflect evaluation perspectives on domain knowledge, translation professionalism, and target language intuition. Next, 250 generated sentence pairs are randomly sampled from the test set; they are respectively from the proposed five models: BPNN, Google Translate, Transformer, PBSMT, and Basic RNN. Each model has 50 sentence pairs. When presented to evaluators, information about the model source is completely concealed, and the order of sentence pairs is randomly shuffled to ensure blind evaluation. The scoring dimensions and criteria are evaluated using a Likert scale ranging from 1 to 5. A score of 1, 2, 3, 4, and 5 indicates very poor, poor, fair, good, and excellent. Evaluators score the following four dimensions, respectively. First, fluency, namely, whether the translation conforms to English expression conventions. Second, grammatical correctness, whether there are basic grammatical errors (such as subject-verb disagreement and tense mistakes). Third, terminology accuracy, whether the translation of core cultural terms is standardized and consistent, for example, “cao yun” is translated as “canal transportation” and “zha ba” as “sluice dam”. Fourth, cultural fidelity, whether the original historical connotations and emotional tones are preserved. Evaluators receive unified training on scoring criteria and learn specific anchor examples for each score level in advance. To assess the consistency of scores among the six evaluators, the Intraclass Correlation Coefficient (ICC) is calculated. The results show that the ICC values are 0.85 for fluency, 0.89 for grammatical correctness, 0.81 for terminology accuracy, and 0.84 for cultural fidelity, all ranging from 0.75 to 0.90; this indicates that the assessment results have good to excellent consistency. Therefore, the scores reported below for each dimension are the arithmetic means of the six evaluators. Based on the above evaluation process, Figs. 6 and 7 demonstrate the manual assessment results for the language quality dimension and cultural communication dimension, respectively. Manual assessment - language quality dimension. Manual assessment - cultural communication dimension. In Fig. 6, the BPNN model proposed in this study scores 4.4 on fluency, outperforming Google Translate (4.2), Transformer (4.0), Basic RNN (3.6), and PBSMT (3.2). This means that sentences generated by the BPNN model are more rhythmic, more in line with native English reading habits, and effectively avoid Chinglish expressions. In terms of grammatical correctness, the BPNN model scores as high as 4.5, also exceeding all comparison models. This shows that the optimized momentum and learning rate decay strategies effectively eliminate basic grammatical errors (e.g., subject-verb disagreement and tense mistakes); this ensures stable and reliable language output. In Fig. 7, targeting the uniqueness of Grand Canal culture, the BPNN model scores as high as 4.6 on terminology accuracy, significantly outperforming Google Translate (4.1), Transformer (4.2), Basic RNN (3.9), and PBSMT (3.4). This result proves that the BPNN model successfully distinguishes the concept of “canal transportation” from general transportation; this also reflects the advantages of vertical domain corpus training. Most importantly, the BPNN model scores 4.1 on cultural fidelity, markedly exceeding Google Translate (3.5), Transformer (3.6), Basic RNN (3.2), and PBSMT (2.9). This indicates that the BPNN model translates literal meanings and preserves the original historical depth and emotional tone through hidden-layer nonlinear mapping. This can effectively reduce the “cultural discount” phenomenon in cross-cultural communication and achieve the elevation from linguistic conversion to cultural dissemination. Discussion The experimental results profoundly reveal the unique advantages of the special optimized BPNN model in the field of vertical culture. Different from the conclusion pointed out by Bouschery et al. (2023), the universal Transformer model was prone to illusion and poor terminology consistency in low-resource cultural data29. In contrast, the proposed model has achieved a remarkable breakthrough in terminology accuracy with targeted corpus training, effectively avoiding the risk of mistranslation of uncommon cultural concepts. ALMahadin et al. (2025) showed that the traditional RNN had the problem of semantic forgetting when dealing with long-distance historical narrative30. This study markedly improves the ROUGE-L score by introducing momentum factor and adaptive learning rate strategies; this proves the superiority of the improved BPNN in maintaining the logical coherence of long texts. Abdurahman et al. (2024) argued that it failed to capture the subtle feelings of culture using a universal large model31. In view of the challenge, the proposed model’s high score performance in the manual assessment of cultural fidelity offers a strong response. To sum up, this study verifies the robustness of BPNN in the generation of Grand Canal discourse; it also provides a new empirical basis and technical paradigm to handle issues that make it difficult to have both high accuracy and deep cultural connotation in cross-cultural communication of cultural heritage. Conclusion Research contribution This study aims to handle the issues of inconsistent terminology translation, lack of cultural context, and low efficiency of artificial generation in the international communication of Grand Canal culture. Meanwhile, it preliminarily explores an intelligent generation model of international discourse based on BPNN. The main contributions can be summarized as the following three aspects. (1) Data resource construction. This study has built a special Chinese-English parallel corpus for the Grand Canal culture. The original data contains about 150,000 candidate sentence pairs. After strict cleaning, filtering, and manual verification, 120,000 high-quality sentence pairs are finally obtained. The corpus covers many thematic dimensions (e.g., water conservancy engineering technology, historical evolution, folk customs, and intangible cultural heritage). (2) Model verification. This study trains and tests the three-layer BPNN model on the corpus and compares it with other similar models. Experimental results indicate that the proposed model is superior to the above baseline model in terms of terminology accuracy and cultural fidelity. This model’s BLEU-4, ROUGE-L, and TER scores reach 0.438, 0.592, and 0.385; meanwhile, the cultural fidelity score in the manual assessment is 4.1 (out of 5). These results preliminarily show that the optimized BPNN has certain application potential in the task of discourse generation in the vertical culture field. (3) Technical path exploration. Different from the mainstream end-to-end NMT architecture, this study decomposes the international discourse generation task of Grand Canal culture into two sub-problems: terminology standardization and style optimization; it also uses a three-layer BPNN as the post-processing optimization module. This technical path offers a referenceable alternative solution for intelligent content production in low-resource, high-exclusivity cultural fields. Of course, the generalization ability of this path still needs to be verified in more cultural fields. Future works and research limitations It should be pointed out that this study has the following limitations. For example, the automatic evaluation indicator mainly measures the formal similarity between the translation and reference texts; it cannot fully reflect the real effect of cultural communication, such as reader acceptance and emotional resonance. Corpus principally covers official texts, but it is not representative of folk oral legends and dialect expressions. The sliding window mechanism of the three-layer BPNN may lose some long-distance semantic dependence when dealing with long historical narrative. This study is only verified in a vertical field of Grand Canal culture, and the generalization ability of the conclusion is not clear. In view of the above limitations, future work can be promoted in the following directions. First, behavioral indicators (e.g., questionnaire survey and eye tracking) are introduced to evaluate the effect of cultural communication more objectively. Second, this study can expand the coverage of the corpus to include folklore, oral historical materials, and multimodal data. Third, in order to enhance the ability of long-distance dependency modeling, future work can explore the hybrid architecture of BPNN and attention mechanisms. Finally, this framework is applied to the Great Wall culture, Silk Road culture, and other cultural heritage fields to test its generalization ability. Acknowledgements The authors acknowledge the help from their university colleagues. Appendix Appendix see Tab A1 Appendix 1BLEU-4 scores of each model in 5 independent runs.RunBPNNTransformerGoogle TranslateRNNPBSMT10.4420.3910.3780.3480.28620.4350.3860.3740.3420.28130.4410.3930.3800.3490.28840.4360.3870.3750.3440.28350.4380.3880.3730.3430.282Mean value0.4380.3890.3760.3450.284Standard deviation0.0040.0050.0040.0050.006 BLEU-4 scores of each model in 5 independent runs. Author contributions Haiyan Gu: Conceptualization, methodology, software, validation, formal analysis, investigation, resources, data curation, writing—original draft preparationJie Bai: writing—review and editing, visualization, supervision, project administration, funding acquisitionJun Xie: methodology, software, validation, formal analysisCui Zhang: methodology, software, validation, formal analysis. Funding This study was the achievement of Anhui Provincial Philosophy and Social Science Planning Project (Grant No. AHSKY2023D121). Data availability The data supporting the findings of this study are available from the corresponding author upon reasonable request.Processed data supporting the findings of this study are available from the corresponding author Jie Bai upon reasonable request via e-mail 34256276@qq.com. Declarations Competing interests The authors declare no competing interests. Ethics statement This article does not contain any studies with human participants or animals performed by any of the authors. All methods were performed in accordance with relevant guidelines and regulations. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The data supporting the findings of this study are available from the corresponding author upon reasonable request.Processed data supporting the findings of this study are available from the corresponding author Jie Bai upon reasonable request via e-mail 34256276@qq.com.