A method for generating interactive dialogue adaptation for cross-border e-commerce between China and Russia

Abstract

The application provides a Sino-Russian cross-border e-commerce interactive dialogue adaptation generation method, and belongs to the technical field of artificial intelligence and natural language processing, and comprises the following steps: collecting Sino-Russian cross-border e-commerce commodity multi-modal data, performing text, video and image extraction processing, and constructing a multi-modal Sino-Russian cross-border e-commerce cultural corpus containing cross-cultural supervision samples; based on a large language model, visual semantic information is fused, a cultural expert hybrid LoRA module and a self-adaptive cultural routing mechanism are introduced, and a cross-cultural dialogue generation model is constructed; based on a joint loss optimization model parameter, self-adaptive optimization and cultural expert global pruning are realized to achieve efficient training of model resources; the trained cross-cultural dialogue generation model is deployed to an e-commerce platform, commodity information and target cultural demand are input, and localized marketing and customer service content are generated. The application realizes the localized generation of commodity content in different languages and cultural environments, and improves the real-time application capability in the cross-border e-commerce scene.

Description

Technical Field

[0001] This invention relates to a method for generating interactive discourse in Sino-Russian cross-border e-commerce, belonging to the field of artificial intelligence and natural language processing technology. Background Technology

[0002] With the rapid development of global cross-border e-commerce and the increasing popularity of digital marketing models, the scale of commodity transactions and user interactions between different countries and regions continues to expand. Due to significant differences in language expression habits, consumption preferences, marketing styles, and customer service communication methods among consumers from different cultural backgrounds, the same product information often requires localized adjustments in language for different target markets. Therefore, utilizing artificial intelligence technology to automatically generate e-commerce content and adapt language to different cultural environments has become an important technological direction for improving the operational efficiency of cross-border e-commerce and enhancing user experience.

[0003] Currently, the main technical approaches for cross-cultural e-commerce content generation include machine translation-based methods, template rule-based methods, and pre-trained large language models. Some studies utilize machine translation models to convert text between different languages, directly transforming product descriptions in the source language into content in the target language. Other methods use pre-set marketing templates or manual rules to generate fixed-format promotional copy and customer service responses based on product categories. However, these existing methods still have the following shortcomings: First, existing machine translation or text generation methods primarily focus on language-level conversion, lacking in-depth modeling of differences in consumer cultural habits, marketing expressions, and interaction scenarios across different countries. This can easily lead to generated content that meets semantic translation requirements but does not conform to the expression habits of the target market. Second, existing cross-cultural content generation methods typically rely on single-text modal information, failing to fully utilize multimodal e-commerce information such as product images, video descriptions, and customer service interactions, making it difficult for the model to comprehensively understand product characteristics and corresponding marketing scenarios. Third, existing large-scale model fine-tuning methods usually require significant training and inference resources, making it difficult to meet real-time generation requirements in actual cross-border e-commerce server or edge deployment environments.

[0004] Therefore, there is an urgent need for a cross-cultural discourse generation method that can combine multimodal cross-cultural e-commerce data, efficient fine-tuning of hybrid parameters by cultural experts, and resource adaptive optimization mechanisms. Summary of the Invention

[0005] The purpose of this invention is to provide a method for generating interactive discourse in Sino-Russian cross-border e-commerce, which realizes the automated localization of product content generation in different language and cultural environments, and reduces the computational overhead of model deployment while ensuring generation quality, thereby improving the real-time application capabilities in cross-border e-commerce scenarios.

[0006] To achieve the above objectives, the present invention employs the following technical solution: A method for generating interactive discourse in Sino-Russian cross-border e-commerce includes the following steps: Collect cross-border e-commerce data in both Chinese and Russian for the same product, and perform multimodal data extraction and standardization processing to obtain standard Chinese e-commerce data and standard Russian e-commerce data; based on the standard Chinese e-commerce data and standard Russian e-commerce data, construct cross-cultural supervised samples and bilingual autoregressive training samples, and organize and store them according to product categories and interaction scenarios to form a multimodal cross-border e-commerce cultural corpus between China and Russia. A cross-cultural discourse generation model based on hybrid LoRA by cultural experts is constructed, comprising a visual feature extraction model, an adapter, a basic language model, and a hybrid LoRA module by cultural experts. Visual semantic features of image data are extracted using the visual feature extraction model and the adapter, and then fused with multimodal semantics of text data to obtain multimodal cue features. These features are then processed by the basic language model to generate the corresponding discourse output. The hybrid LoRA module by cultural experts is introduced into the Transformer layer of the basic language model. This module includes shared LoRA parameters, expert LoRA parameters, a cultural matching scoring module, and an adaptive cultural routing module. A cross-cultural discourse generation model was trained using a multimodal Sino-Russian cross-border e-commerce cultural corpus. The model parameters were optimized using a multi-task cross-cultural joint loss. Under a fixed LoRA total rank budget, the rank of shared LoRA and the number of experts in each Transformer layer were adaptively adjusted according to the gradient contribution relationship between shared LoRA parameters and expert LoRA parameters. After training, experts were pruned according to their LoRA parameter importance scores to obtain the optimized cross-cultural discourse generation model. After deploying the cross-cultural discourse generation model, the model processes the product text and optional image information in the input Chinese or Russian product information, and constructs cross-cultural generation prompts based on the target language and target cultural market demand to form the model input sample. The model input sample is then input into the deployed model to generate corresponding marketing copy in the target cultural environment.

[0007] Preferably, the multimodal data extraction and standardization process specifically includes: The pre-trained Whisper-large-v3 model is used to extract the audio content from the product introduction video and generate the corresponding text introduction sequence. At the same time, the product introduction video is segmented at fixed time intervals to obtain the image introduction sequence. The text introduction sequence and the image introduction sequence are concatenated to obtain the product introduction data set. Remove HTML tags, web links, emoticons, duplicate text, and whitespace characters from marketing copy and customer service dialogue records. Segment the text according to periods, question marks, exclamation marks, and line breaks. When the length of a single text segment exceeds the preset length, divide it into a text segment by 512 consecutive tokens to obtain the marketing copy text and customer service dialogue text.

[0008] Preferably, the cross-cultural supervision samples include long text cross-cultural supervision samples and short text cross-cultural supervision samples; The long-text cross-cultural supervision sample is constructed as follows: using the product description data group in the standard Chinese e-commerce data as input and the text description sequence in the corresponding Russian product description data group as supervision labels, a long-text cross-cultural supervision sample from Chinese to Russian is constructed; and using the Russian product description data group as input and the Chinese text description sequence as output, a long-text cross-cultural supervision sample from Russian to Chinese is constructed. The method for constructing the short text cross-cultural supervision sample is as follows: key sentences that can express the same marketing intent or customer service semantics in Chinese marketing copy text and customer service dialogue text are extracted by a large language model, and a short text semantic correspondence relationship is established with the corresponding Russian marketing copy text and customer service dialogue text to construct the short text cross-cultural supervision sample. The bilingual autoregressive training sample is constructed as follows: Chinese product description text, marketing copy text, and customer service dialogue text are spliced ​​together in their original order to form a continuous Chinese text sequence. The corresponding Russian text is then used to form a continuous Russian text sequence in the same way. The bilingual autoregressive training sample is constructed using the next token prediction method.

[0009] Preferably, the visual feature extraction model adopts CLIP-ViT-L / 14, the adapter includes two fully connected layers, and the basic language model adopts the pre-trained Qwen3-8B-Instruct; the basic language model is kept frozen during the training of the cross-cultural discourse generation model.

[0010] Preferably, the cultural expert hybrid LoRA module specifically includes: For each Transformer layer of the base language model, multiple levels of LoRA are inserted into the Query and Value. The LoRA update of the l-th layer is represented as: , in, For the pre-trained Qwen3-8B-Instruct model The original matrix parameters of the layer, For the pre-trained Qwen3-8B-Instruct model The update matrix after applying hierarchical LoRA to the layer. For cross Shared LoRA parameters of the layers For the selection from the expert head pool targeting the first The first layer The update matrix corresponding to each expert, i.e., the LoRA parameters of each expert. For the corresponding weights, This represents the total number of experts at the current level. The corresponding LoRA rank number is , The corresponding rank is 1, the first The total rank of the layer is ,and and All are adaptive optimization parameters; The expert head pool consists of P LoRAs with rank 1, and P ≥ K.

[0011] Preferably, the culture matching scoring module specifically includes: For each expert with rank 1 in the expert head pool, compute its low-rank increment representation: , in, Indicates the input number Feature representation before the Trasformer layer, Indicates the first expert in the first pool Parameters corresponding to each expert Indicates the first The expert in the The cultural semantic activation state generated by the layer, i.e., the low-rank incremental representation; Calculate the contribution of each expert to the current input: , Indicates the first The expert in the The contribution of the layer; The expert with the highest contribution is selected as the basic expert of this layer, and the cosine similarity of the update direction between the basic expert and the remaining experts is calculated; the expert with the lowest cosine similarity is selected as the second basic expert, and the expert set composed of the basic experts is used as the expert head pool corresponding to this layer. Iteratively select the expert with the lowest average cosine similarity to the currently selected expert set from the remaining experts as the new basic expert, and update the selected expert set until the required number of experts K for this layer is reached.

[0012] Preferably, the adaptive culture routing module specifically includes: For the z basic experts selected in the l-th layer, trainable gating parameters are introduced for each basic expert through a gating layer, and the fusion weights are obtained by Softmax normalization. During the training of the cross-cultural discourse generation model, the gating parameters of experts from different cultures in each layer are optimized and the contribution ratio of experts is adjusted.

[0013] Preferably, the multi-task cross-cultural joint loss is as follows: for long text cross-cultural supervised samples and short text cross-cultural supervised samples, the loss during training is the cross-entropy loss between the model output text and the real target cultural text; for bilingual autoregressive training samples, the autoregressive loss is calculated by the next token prediction method; the two types of losses are weighted and summed to obtain the joint training loss Loss; during the training process, bilingual autoregressive training samples are first used for model fine-tuning training, and then cultural supervised samples are used for training. The rank and number of experts sharing LoRA in each Transformer layer are adaptively adjusted as follows: Set the total LoRA rank budget for each Transformer layer to R, and R = r + K, initialize r = K = R / 2; During the training of the cross-cultural discourse generation model, the average gradient contribution corresponding to the shared LoRA parameters and the average gradient contribution corresponding to the expert LoRA parameters were statistically analyzed respectively: , , in, Indicates the first Layers share the average gradient contribution of LoRA to the current training task. This represents the average gradient contribution of the LoRA of the experts at this layer. denoted as the multi-task cross-cultural joint training loss, which is the weighted sum of the cross-entropy loss corresponding to the long text cross-cultural supervised samples and the short text cross-cultural supervised samples, and the autoregressive loss corresponding to the bilingual autoregressive training samples; when When an expert's rank is converted to a shared rank, When a shared rank is converted into the number of experts, a rank transfer operation is performed only once if the gradient contribution relationship remains consistent over three consecutive training cycles. The adaptive adjustment process is stopped when the cross-cultural discourse generation model has reached 80% of the training iterations.

[0014] Preferably, the pruning of experts based on their LoRA parameter importance scores specifically includes: For the trained cross-cultural discourse generation model, calculate the importance score of the k-th expert in layer l: , in, This represents the importance score of the k-th expert in the l-th layer using LoRA; Set the expert pruning ratio parameters according to the actual computing resources and inference speed requirements during deployment. ; Rank all experts in the Transformer layers globally according to their importance scores, and select the experts with the lowest importance from the entire expert set. The proportional experts are deleted; the deleted experts no longer participate in the subsequent reasoning process, and the gating weights of the remaining experts are renormalized. The normalized gating weights are then used to reason about the remaining experts. The expert pruning ratio parameter is set as follows: Pre-measure the inference latency and memory usage of a single expert on the target deployment platform. Determine the ratio of the number of experts to be retained to the total number of experts based on the target inference speed requirements. Set 1 as the difference between the ratio of the number of LoRA experts to be retained and the total number of LoRA experts. .

[0015] Preferably, after deploying the cross-cultural discourse generation model, the product text and optional image information are processed for the input Chinese or Russian product information, and cross-cultural generation prompts are constructed in combination with the target language and target cultural market demand to form the model input sample; the model input sample is input into the deployed model to generate cross-cultural marketing copy, product introduction text or customer service response content in the target cultural environment; For the input Chinese or Russian product information, the product text and optional image information are processed, and cross-cultural generated prompts are constructed based on the target language and target cultural market demands to form the model input sample, specifically including: Obtain the Chinese or Russian product information to be converted, wherein the product information includes product text description and / or product introduction video and / or product image information; when the input product information includes a product introduction video, perform text extraction and image extraction on the product introduction video to obtain the corresponding product introduction text information and product image information; when the input product information only contains text information, directly obtain the product text description. The cross-cultural generated prompts are used to specify the source language, target language, and target cultural market demand, and are combined with product text descriptions, product introduction text information, and product image information to construct input samples.

[0016] The advantages of this invention are: By constructing a multimodal Sino-Russian cross-border e-commerce cultural corpus, integrating multi-source information such as product introduction videos, product texts, marketing copy, and customer service dialogues, the model can learn the differences in cultural expressions under different language environments, thereby improving the matching degree between the generated content and the language habits and marketing styles of the target market.

[0017] By using a cross-cultural discourse generation model that combines LoRA with cultural experts, visual semantic information and textual semantic information are jointly modeled. The model learns general cultural knowledge and personalized cultural expression patterns by using shared LoRA and expert LoRA respectively, thereby improving the discourse generation capabilities of the model in different product categories and cross-cultural interaction scenarios.

[0018] The adaptive expert optimization training method can dynamically adjust the capacity ratio of shared LoRA and expert LoRA under a fixed total LoRA rank budget, and combine it with a resource-constrained cultural expert global pruning method to reduce the computational overhead of ineffective experts, thereby achieving efficient inference in resource-constrained environments.

[0019] By deploying a cross-cultural discourse generation model that blends trained and pruned cultural experts with LoRA onto a cross-border e-commerce server, and combining cross-cultural generation prompts with a cultural expert adaptive routing mechanism, marketing copy, product descriptions, and customer service responses can be generated in real time based on the target language and target cultural market needs, thereby improving the operational efficiency and user interaction experience of cross-border e-commerce.

[0020] This invention enables automated localization of product content generation in different language and cultural environments, while reducing model deployment computational overhead and improving real-time application capabilities in cross-border e-commerce scenarios, while ensuring generation quality. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0022] Figure 1 This is a flowchart of the overall technical roadmap.

[0023] Figure 2 A structural diagram of a cross-cultural discourse generation model for cultural experts using LoRA.

[0024] Figure 3 Optimize the training flowchart for adaptive experts.

[0025] Figure 4 A diagram illustrating the effects of cross-cultural discourse generation under different combinations of modal information.

[0026] Figure 5 Analysis of the distribution and generation capabilities of cultural experts in different cultural contexts.

[0027] Figure 6 Analysis of model performance and resource consumption under different expert optimization strategies. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Example 1 A method for adapting and generating interactive discourse in Sino-Russian cross-border e-commerce is proposed. This method first collects product introduction videos, marketing copy, and customer service dialogue data from Sino-Russian cross-border e-commerce, constructing a multimodal Sino-Russian cross-border e-commerce cultural corpus containing multimodal information and cross-cultural supervised samples. Then, a cross-cultural discourse generation model based on hybrid LoRA (Local Expert Relationship Analysis) is constructed. This model integrates product text and image semantic information, introduces a collaborative modeling mechanism of shared cultural knowledge and expert cultural knowledge, and utilizes an adaptive cultural expert structure and an adaptive cultural routing module to adjust discourse expression under different cultural scenarios. Furthermore, an adaptive expert optimization training method based on cultural complexity awareness is proposed. This method optimizes model parameters through multi-task cross-cultural joint loss and dynamically adjusts the model's representational capacity based on a shared-expert capacity adaptive optimization method. Simultaneously, a resource-constrained global pruning method for cultural experts is combined to reduce the computational cost of model deployment. Finally, the trained hybrid LoRA cross-cultural discourse generation model is deployed to a cross-border e-commerce application environment, generating localized marketing copy, product introduction text, and customer service responses based on input product information and target cultural needs. This invention can improve the cultural adaptability of content generation in Sino-Russian cross-border e-commerce, while reducing the resource consumption for model training and deployment, and achieving efficient cross-cultural e-commerce interactive discourse generation.

[0030] like Figure 1 As shown, the specific steps include: S1: Collect cross-border e-commerce data in both Chinese and Russian for the same product, and perform multimodal data extraction and standardization processing to obtain standard Chinese e-commerce data and standard Russian e-commerce data; based on the standard Chinese e-commerce data and standard Russian e-commerce data, construct cross-cultural supervision samples and bilingual autoregressive training samples, and organize and store them according to product categories and interaction scenarios to form a multimodal cross-border e-commerce cultural corpus in China and Russia. S2: Construct a cross-cultural discourse generation model based on hybrid LoRA by cultural experts, including a visual feature extraction model, an adapter, a basic language model, and a hybrid LoRA module by cultural experts; extract visual semantic features from image data through the visual feature extraction model and the adapter, fuse them with multimodal semantics of text data to obtain multimodal cue features, and then generate corresponding discourse output results through the basic language model; introduce a hybrid LoRA module by cultural experts into the Transformer layer of the basic language model, the hybrid LoRA module by cultural experts includes shared LoRA parameters, expert LoRA parameters, a cultural matching scoring module, and an adaptive cultural routing module; S3: The cross-cultural discourse generation model is trained using a multimodal Sino-Russian cross-border e-commerce cultural corpus, and the model parameters are optimized through multi-task cross-cultural joint loss. Under a fixed LoRA total rank budget, the rank of shared LoRA and the number of experts in each Transformer layer are adaptively adjusted according to the gradient contribution relationship between shared LoRA parameters and expert LoRA parameters. After training, experts are pruned according to the importance score of expert LoRA parameters to obtain the optimized cross-cultural discourse generation model. S4: After deploying the cross-cultural discourse generation model, process the product text and optional image information for the input Chinese or Russian product information, and construct cross-cultural generation prompts based on the target language and target cultural market demand to form the model input sample; input the model input sample into the deployed model to generate the corresponding marketing copy in the target cultural environment.

[0031] As a refinement of the above embodiments, in step S1, due to significant differences in language expression habits and marketing styles across countries and regions, marketing copy for the same product often has different levels of user acceptance and marketing effectiveness in different cultural contexts. Therefore, directly using machine translation to convert Chinese marketing content into Russian, or Russian marketing content into Chinese, makes it difficult to achieve localized adaptation of discourse in the target market, easily leading to problems such as marketing semantic deviation and decreased user acceptance. Therefore, this invention first constructs a multimodal cultural adaptation dataset that conforms to the characteristics of cross-cultural marketing between China and Russia.

[0032] Specifically, the following steps are included: S1-1 Raw Cross-Border E-commerce Data Collection: Raw cross-border e-commerce data is collected from the company's internal cross-border e-commerce product management system and the product databases of partner companies. Specifically, for products sold simultaneously in the Chinese and Russian markets, corresponding bilingual (Chinese and Russian) product introduction videos, marketing copy, and customer service dialogue records are collected, thus obtaining raw Chinese cross-border e-commerce data. and original Russian cross-border e-commerce data .

[0033] S1-2 Multimodal E-commerce Data Extraction: Based on the raw Chinese cross-border e-commerce data collected in S1-1 and original Russian cross-border e-commerce data We perform unified data extraction and standardized processing on product introduction videos, marketing copy, and customer service dialogue records.

[0034] Specifically, for The product introduction video in the video is first extracted using a pre-trained Whisper-large-v3 model, and then a corresponding text introduction sequence is generated. Simultaneously, frames of the product introduction video are extracted at fixed time intervals, and an image is saved every 2 seconds to obtain a sequence of Chinese image introductions. And splice together the product description data set. ; for The marketing copy and customer service dialogue records were processed, removing HTML tags, web links, emoticons, duplicate text, and whitespace characters. Sentences were then segmented using periods, question marks, exclamation marks, and line breaks. When a single text segment exceeded a preset length, it was divided into a single text fragment using 512 consecutive tokens, resulting in the marketing copy text. and customer service dialogue text ; Therefore, original Chinese cross-border e-commerce data Standard Chinese e-commerce data was obtained after processing. Similarly, the same method was used to analyze the original Russian cross-border e-commerce data. Data extraction and standardization were performed to obtain standard Russian e-commerce data. ,in, These are the Russian-language product description data groups, marketing copy texts, and customer service dialogue texts.

[0035] S1-3 Construction of Cross-Cultural Supervisory Labels: Based on Standard Chinese E-commerce Data Obtained in S1-2 and standard Russian e-commerce data Furthermore, long-text cross-cultural supervised samples, short-text cross-cultural supervised samples, and bilingual autoregressive training samples were constructed respectively. Specifically: Construction of a Long-Text Cross-Cultural Supervision Sample: Targeting Product Description Data Sets in Standard Chinese E-commerce Data Product description data group For input, use the text description sequence corresponding to the Russian product description data group. As a supervisory label, a cross-cultural supervisory sample of long texts from Chinese to Russian was constructed. Similarly, product description data sets in Russian. Using Russian as input and a sequence of Chinese text descriptions as output, a long-text cross-cultural supervised sample is constructed from Russian to Chinese. .

[0036] Construction of a Short-Text Cross-Cultural Supervision Sample: Marketing Copywriting Texts for Chinese and customer service dialogue text The pre-trained large language model Whisper-large-v3 is used to extract key sentences that express the same marketing intent or customer service semantics, and these are combined with the corresponding Russian marketing copy text. and customer service dialogue text Establish semantic correspondences between short texts and construct supervised samples of short texts. ,in, This indicates key Chinese sentences. This indicates the corresponding key phrases in Russian. For example, the Chinese marketing copy "Limited-time flash sale, only 100 pieces left" corresponds to the Russian marketing copy: "Мгновеннаяскидка,количествотовараограничено."

[0037] Bilingual Autoregressive Training Sample Construction: To enhance the model's bilingual language modeling ability, standard Chinese e-commerce data was used. and standard Russian e-commerce data Each of these was used to construct autoregressive training samples. Specifically, Chinese product description text, marketing copy text, and customer service dialogue text were concatenated in their original order to form a continuous Chinese text sequence, and the corresponding Russian text was used in the same way to form a continuous Russian text sequence. The bilingual autoregressive training data was then constructed using the next-to-the-next-to-toe prediction method. ,in This represents the sequence of the first t tokens of the current input. For the (t+1)th token, and .

[0038] Construction of the S1-4 Multimodal Sino-Russian Cross-border E-commerce Cultural Corpus: Based on the data collection, extraction, and supervised sample construction methods described in S1-1 to S1-3, further data on Sino-Russian cross-border e-commerce under different product categories and different e-commerce interaction scenarios are collected and stored according to product numbers, forming a multimodal Sino-Russian cross-border e-commerce cultural corpus for various products. ,in, This represents the cross-cultural e-commerce data corresponding to the i-th group of products, and .

[0039] As a refinement of the above embodiments, step S2 addresses the challenge that, due to the different discourse organization methods in product descriptions, marketing expressions, and customer interactions across different cultural backgrounds, directly applicable large language models struggle to automatically adjust their generation strategies according to the target cultural environment. Therefore, this invention constructs a cross-cultural discourse generation model based on a hybrid LoRA framework using cultural experts. Through multimodal semantic fusion and efficient fine-tuning of cultural expert parameters, it achieves adaptive generation of target cultural discourse content from input product information.

[0040] The network architecture of this model is as follows: Figure 2 As shown, the specific steps include: S2-1 Multimodal Cross-cultural Basic Module Construction: Based on a pre-trained large language model, a basic model for cross-cultural discourse generation is constructed, and a visual semantic encoding module is introduced to process product image information.

[0041] Specifically, Qwen3-8B-Instruct is used as the base language model, and CLIP-ViT-L / 14 is used as the visual feature extraction model. First, the pre-trained CLIP-ViT-L / 14 receives image data. As input, and , To introduce the sequence of Russian images, we obtain the visual features of the images. Because of the modal differences between the CLIP visual feature space and the large language model text feature space, it is not possible to directly input the visual feature space into the language model for cross-cultural discourse generation.

[0042] express The range of values ​​for is and The union of two image sequences: the images input to the CLIP visual encoder may come from frame sequences of either a Chinese or Russian product introduction video. Images from both languages ​​share the same CLIP pre-trained visual feature extractor, and are uniformly mapped to obtain the image visual features. .

[0043] Therefore, the present invention further designs an adapter to incorporate image visual features. The semantic representation is mapped to a space consistent with the Qwen3-8B-Instruct input space. The adapter consists of two fully connected layers, representing image visual features. Visual semantic features are obtained through the adapter. ; then and text data Concatenation is used as a multimodal cue feature. ,in Subsequently, multimodal cue features The pre-trained Qwen3-8B-Instruct model performs cross-modal semantic understanding through the Transformer encoding layer and generates corresponding discourse outputs based on the target culture's needs.

[0044] In addition, to preserve the existing general language capabilities of the pre-trained large language model and reduce the parameter update cost during cross-cultural adaptation, the pre-trained Qwen3-8B-Instruct model is kept frozen during training and adapted by inserting an additional cultural expert hybrid LoRA module.

[0045] Therefore, the trainable parameters of this model include the parameters of the pre-trained CLIP-ViT-L / 14 model. Adapter parameters Cultural experts mix LoRA module parameters .

[0046] S2-2 Cultural Expert Hybrid LoRA Module Design: For the pre-trained Qwen3-8B-Instruct model, the number of Transformer layers is... Then, for each layer of Transformer, insert multi-level LoRA on the Query and Value, that is: for the first layer... Layer, LoRA update representation is: , in, and These are the pre-trained Qwen3-8B-Instruct models. The original matrix parameters of the layer and the update matrix after applying hierarchical LoRA. For cross The LoRA parameters of the layers are shared across all layers and optimized together. For the selection from the expert pool targeting the first The first layer The update matrix corresponding to each LoRA expert, i.e., the LoRA parameters of the expert. The corresponding weights are given, and the total number of experts in this layer is... .in, The corresponding LoRA rank number is ,and The corresponding rank is 1, therefore, in the th The total rank of the layer is ,and and These are all adaptive optimization parameters.

[0047] In addition, the expert head pool includes a total of There are LoRAs, each with a rank of 1, and it is required that... , preset And it is expected that performance enhancement can increase However, this also increases computational cost. During fine-tuning training, each layer adaptively selects from the expert head pool. Each LoRA constitutes an adaptive expert update, and the selection method is shown in S2-3.

[0048] S2-3 Adaptive Cultural Expert Structure Design: In order to realize the dynamic combination of cultural experts in different cross-cultural scenarios, this invention constructs a cultural matching scoring module to dynamically select and combine LoRA experts in the expert head pool.

[0049] Specifically, for each LoRA expert with rank=1 in the expert head pool, the corresponding low-rank increment representation is calculated: , in, Indicates the input number Feature representation before the Trasformer layer, Indicates the first expert in the first pool Parameters corresponding to each expert Indicates the first The expert in the The cultural semantic activation state generated by the layer is represented by the low-rank increment.

[0050] Then, the contribution of each LoRA expert to the current input is calculated first. , Indicates the first The expert in the The contribution of each layer is ultimately determined by all Contribution sequence corresponding to each LoRA expert ,in The larger the value, the more significant the cultural semantic correction effect can be produced by the corresponding LoRA expert.

[0051] choose The largest LoRA expert serves as the base expert for this layer; to avoid multiple LoRA experts learning the same discourse expression patterns, the cosine similarity of the update direction between candidate experts and selected experts is further calculated: , in, This represents the low-rank increment corresponding to the selected basic expert. Let be the cosine similarity between the i-th expert and the basic experts. Indicates matrix transpose; Next, the expert with the lowest cosine similarity to the base experts is selected from the remaining experts and added as the second base expert, becoming a member of the corresponding expert head pool for this layer. Only the first base expert is selected directly based on contribution. Afterwards, the expert with the lowest average cosine similarity to the currently selected expert set is selected from the remaining experts as a new base expert. This ensures significant differences in parameter update directions among experts from different cultures, preventing multiple experts from learning the same discourse expression patterns. Only one new expert is selected at a time, and the currently selected expert set is updated until the required number of experts for this layer is reached. .

[0052] Finally, the adaptive cultural routing module of S2-4 dynamically combines the existing basic experts in the expert head pool.

[0053] S2-4 Adaptive Culture Routing Module Design: A culture routing fusion module based on trainable gating parameters is designed, targeting the culture routing fusion module already selected in the l-th layer of S2-3. A basic expert Trainable gating parameters are introduced for each basic expert through a gating layer. This parameter is then used to obtain the normalized fusion weights via Softmax. , among which, among which ,and During model training, the gating parameters of different cultural experts at each layer are automatically optimized, enabling the model to adjust the proportion of expert contributions based on different product categories, language environments, and user needs.

[0054] Finally, the first Full LoRA incremental update for layer Transformer: When the first The expert head pool of the layer reached hour, This is the complete update in S2-2.

[0055] As a refinement of the above embodiments, step S3, in order to enable the cross-cultural discourse generation model of cultural experts hybrid LoRA to automatically adjust the model's expressive ability according to different cross-cultural scenarios and avoid the problems of wasted computational resources in simple cultural scenarios and insufficient adaptation to complex cultural scenarios caused by a fixed number of experts and a fixed LoRA rank, proposes an adaptive expert optimization training method based on cultural complexity awareness. During model training, the multimodal Sino-Russian cross-border e-commerce cultural corpus constructed in S1 is used. Joint training is performed, and the shared LoRA rank parameter in each Transformer layer is dynamically optimized based on the cultural complexity of the input samples. and the number of experts The adaptive expert optimization training process is as follows: Figure 3 As shown, the specific steps include: S3-1 Multi-Task Cross-Cultural Joint Training Loss Construction: A Multimodal Sino-Russian Cross-Border E-commerce Cultural Corpus Based on S3 Joint training tasks are constructed according to different discourse generation needs.

[0056] Specifically, for cross-cultural monitoring samples of long and short texts. , as well as During training, the loss is calculated as the cross-entropy loss between the model's output text and the real target cultural text; for bilingual autoregressive training samples... The autoregressive loss is calculated using the next token prediction method; the joint training loss is obtained by weighted summation of these two types of losses. .

[0057] Furthermore, bilingual autoregressive training samples were first used during the training process. The model undergoes fine-tuning training, followed by training with culturally supervised samples. The fine-tuning parameters include the pre-trained CLIP-ViT-L / 14 model parameters. Adapter parameters Cultural experts mix LoRA module parameters And initialize it.

[0058] S3-2 Shared-Expert Capacity Adaptive Optimization: In order to achieve a dynamic balance between shared cultural knowledge and personalized cultural knowledge under a fixed parameter budget, this invention proposes a shared-expert capacity adaptive optimization method based on gradient contribution awareness.

[0059] Specifically, the total LoRA rank budget for each Transformer layer is pre-set to be... ,and ,in, This indicates the number of rank values ​​corresponding to shared LoRA. This indicates the number of experts currently active in the layer, and is initialized... .

[0060] During model training, the shared LoRA parameters are counted separately. And expert LoRA parameters Corresponding gradient response: , , in, Indicates the first Layers share the average gradient contribution of LoRA to the current training task. This represents the average gradient contribution of the LoRA of the experts at this layer as a whole; denoted as the multi-task cross-cultural joint training loss, which is the weighted sum of the cross-entropy loss corresponding to the long text cross-cultural supervised samples and the short text cross-cultural supervised samples, and the autoregressive loss corresponding to the bilingual autoregressive training samples.

[0061] Subsequently, the rank assignment is dynamically adjusted based on the gradient contribution relationship between shared LoRA and expert LoRA: when In this case, it indicates that the current layer loss optimization mainly relies on cross-cultural shared semantic representations, therefore the number of experts is reduced, and an expert rank is converted into a shared rank: , ;when When this occurs, it indicates that the current layer requires more capacity to express cultural differences, therefore reducing the shared representation capacity and converting a shared rank into the number of experts: , .

[0062] Meanwhile, to avoid model oscillations caused by frequent structural adjustments during training, a rank transfer operation is performed only once when the gradient contribution relationship remains consistent for three consecutive training cycles. Ultimately, while keeping the total LoRA parameter budget R constant, different Transformer layers adaptively adjust the ratio of shared knowledge capacity to cultural expert capacity according to actual cross-cultural generation needs.

[0063] Therefore, while dynamically adjusting the capacity ratio of shared LoRA and expert LoRA, a joint training loss based on S3-1 is constructed. Model optimization is performed. When the model training reaches 80% of the preset number of iterations, the shared-expert capacity adjustment process is stopped, and training continues under adaptive settings to finally obtain the trained cultural expert hybrid LoRA cross-cultural discourse generation model.

[0064] S3-3 Resource-Constrained Global Pruning of Cultural Experts: To reduce the computational overhead of the model during actual deployment after training and to adjust the model size according to different inference speed requirements, this invention further designs a cultural expert pruning method based on global importance ranking. Since shared LoRAs are used to learn the universal language expression capabilities required in different cultural scenarios, directly deleting shared LoRAs may reduce the overall cross-cultural transferability of the model. Therefore, pruning optimization is only performed on expert LoRAs used to supplement the expression of cultural differences.

[0065] Specifically, for the cross-cultural discourse generation model using hybrid LoRA for cultural experts after training in S3-2, the importance scores of the expert LoRA parameters in all Transformer layers were statistically analyzed. For the... The first in the layer Several experts calculated its importance: , in, The LoRA importance score of the k-th expert in the l-th layer is represented by the product of the gating weight and the expert parameter norm. The larger the value, the more important the expert is to the cross-cultural discourse generation task.

[0066] Furthermore, based on the actual computing resources and inference speed requirements during deployment, expert pruning ratio parameters are set. ,in This indicates the percentage of experts that need to be removed. For example, when you want to reduce the computational overhead of the expert module by approximately 20%, set this... Subsequently, all cultural experts in the Transformer layers were scored according to their importance. Perform a global ranking and select the least important experts from the entire expert set. The proportion expert made the deletion, that is: ,in, This indicates the set of experts that need to be deleted. Indicates before calculation Large values. Ultimately, the deleted experts no longer participate in the subsequent reasoning process; and the gating weights of the remaining experts are renormalized. ,use Reasoning is applied to the remaining experts to ensure that the pruned expert group still maintains normal discourse generation capabilities.

[0067] Set the expert pruning ratio parameters according to the actual computing resources and inference speed requirements during deployment. , , This indicates the proportion of experts that need to be deleted. Specifically, the inference latency and memory usage of a single LoRA expert on the target deployment platform are measured in advance. Based on the target inference speed requirements, the ratio of the number of LoRA experts to be retained to the total number of all LoRA experts is determined. The difference between 1 and the ratio of the number of LoRA experts to be retained to the total number of all LoRA experts is defined as... .

[0068] As a refinement of the above embodiments, step S4 is to realize the application of the trained cross-cultural discourse generation model in actual cross-border e-commerce scenarios. This invention deploys the cross-cultural discourse generation model of cultural expert hybrid LoRA trained and pruned in S3 into the cross-border e-commerce marketing system. By inputting multimodal information of goods and cultural needs of the target market, it realizes the real-time generation of Sino-Russian cross-cultural e-commerce content, specifically including the following steps: S4-1 Cross-cultural Discourse Generation Model Deployment: Deploy the trained and pruned cross-cultural discourse generation model using LoRA hybrid language with cultural experts to the enterprise's cross-border e-commerce server; the deployed model includes pre-trained Qwen3-8B-Instruct basic language model parameters and fine-tuned CLIP-ViT-L / 14 model parameters. Adapter parameters And cultural experts mix LoRA module parameters .

[0069] Among them, the cultural expert mixed LoRA module parameters Including shared LoRA parameters The expert LoRA parameter set corresponding to each Transformer layer and the gating weights of the normalized residual experts .

[0070] S4-2 Product Multimodal Information Input Processing: First, obtain the Chinese or Russian product information to be converted. The product information includes at least a product text description. Furthermore, it may include product introduction videos or product image information. When the input contains a product introduction video, the multimodal e-commerce data extraction method described in S1 is used to extract text and images from the video to obtain the corresponding product introduction text information. and product image information When the input contains only text information, directly retrieve the product text description. .

[0071] Subsequently, corresponding cross-cultural generation prompts were constructed based on actual conversion needs. ,in, Used to specify the source language, target language, and target cultural market demands, and combined with product text descriptions. Optional product description text information and product image information Together they form the model input samples .

[0072] For example, when converting Chinese product marketing content into Russian marketing language, the prompt could be: "Please convert the following Chinese product descriptions into marketing copy that conforms to the expression habits of Russian consumers, keeping the core product information unchanged, and adopting a promotional style suitable for the Russian market." When converting Russian product descriptions into Chinese marketing language, the prompt could be: "Please convert the following Russian product descriptions into marketing copy that conforms to the reading habits of Chinese consumers, highlighting the product's selling points and adopting localized marketing expressions."

[0073] S4-3 Cross-cultural discourse generation and reasoning: Input the model constructed in S4-2 into the sample. Inputting into the cultural expert hybrid LoRA cross-cultural discourse generation model deployed in S4-1, marketing copy corresponding to the target cultural environment is generated through the model's internal multimodal semantic understanding and cultural expert adaptive routing mechanism. It also serves as localized language content in cross-border e-commerce platforms for product release, marketing promotion, and customer service scenarios.

[0074] Experimental Analysis To verify the effectiveness of the proposed Sino-Russian cross-border e-commerce interactive discourse adaptation and generation method, a simulation experimental environment for Sino-Russian cross-border e-commerce content generation was constructed. The experiment collected Chinese and Russian product descriptions, marketing copy, and customer service interaction data under different product categories, and constructed a multimodal Sino-Russian cross-border e-commerce cultural corpus based on the method of this invention. On this basis, experiments were conducted to verify the multimodal data construction method, the cross-cultural discourse generation model using a hybrid LoRA approach with cultural experts, and the adaptive expert optimization training method based on cultural complexity awareness.

[0075] (1) Validity analysis of the multimodal Sino-Russian cross-border e-commerce cultural corpus Figure 4 This paper presents a comparative study of cross-cultural discourse generation performance under different input modalities. Experiments were conducted using text-only input, combined text and image input, video information extraction input, and training with the multimodal Sino-Russian cross-border e-commerce cultural corpus constructed in this invention. When using only text information, the model primarily relies on product names and descriptions for language conversion, making it difficult to fully understand product visual features and marketing scenarios. Adding image information allows the model to further identify visual semantics such as product appearance and usage scenarios, improving the completeness of the generated content. Further combining video text extraction information allows the model to acquire richer product descriptions and interactive semantics. This invention, by constructing a multimodal Sino-Russian cross-border e-commerce cultural corpus and jointly modeling product images, product description text, and marketing interaction information, enables the model to more accurately understand the cultural context of products and generate cross-cultural discourse content that conforms to the expression habits of the target market.

[0076] (2) Analysis of the structural validity of the hybrid LoRA model of cultural experts Figure 5 This study showcases the expert activation patterns and generative capabilities of the hybrid LoRA module for cultural experts in different cultural scenarios. The experiments selected Chinese marketing scenarios, Russian marketing scenarios, and bilingual customer service interaction scenarios as test subjects, and statistically analyzed the activation weight distribution of expert LoRA in different Transformer layers.

[0077] It can be observed that the expert activation patterns differ significantly across different cultural scenarios. Some experts make a high contribution to marketing tasks in Russia, while others are more suitable for Chinese market expression and customer service interaction tasks. Compared to a fixed LoRA structure, the cultural expert hybrid LoRA model proposed in this invention can dynamically select expert combinations based on input cultural needs through an adaptive cultural expert structure and an adaptive cultural routing module. This enables the model to possess both cross-cultural universal expression capabilities and target market personalized expression capabilities.

[0078] (3) Analysis of the effect of adaptive expert optimization based on cultural complexity perception Figure 6 The changes in model performance and computational resource consumption under different expert optimization strategies are demonstrated. Experiments compare the methods with a fixed number of experts, a fixed LoRArank method, and the cultural complexity-aware adaptive expert optimization training method proposed in this invention.

[0079] Fixed expert count methods employ the same computational scale across all cultural scenarios, leading to wasted computational resources in simple cultural scenarios and insufficient expressive power in complex cultural scenarios. While fixed LoRArank methods maintain model structural stability, they cannot adjust the ratio of shared to personalized knowledge according to different task requirements. This invention employs a shared-expert capacity adaptive optimization method, dynamically adjusting the rank parameter of the shared LoRA and the number of experts based on the gradient contribution relationship between the shared LoRA parameters and the expert LoRA parameters. Combined with a resource-constrained global pruning method for cultural experts, low-importance experts are removed. This reduces the computational overhead of the expert module while maintaining good cross-cultural discourse generation performance, achieving a dynamic balance between model performance and deployment resources.

[0080] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for adapting and generating interactive discourse in Sino-Russian cross-border e-commerce, characterized in that, Includes the following steps: Collect cross-border e-commerce data in both Chinese and Russian for the same product, and perform multimodal data extraction and standardization processing to obtain standard Chinese e-commerce data and standard Russian e-commerce data; based on the standard Chinese e-commerce data and standard Russian e-commerce data, construct cross-cultural supervised samples and bilingual autoregressive training samples, and organize and store them according to product categories and interaction scenarios to form a multimodal cross-border e-commerce cultural corpus between China and Russia. A cross-cultural discourse generation model based on hybrid LoRA by cultural experts is constructed, comprising a visual feature extraction model, an adapter, a basic language model, and a hybrid LoRA module by cultural experts. Visual semantic features of image data are extracted using the visual feature extraction model and the adapter, and then fused with multimodal semantics of text data to obtain multimodal cue features. These features are then processed by the basic language model to generate the corresponding discourse output. The hybrid LoRA module by cultural experts is introduced into the Transformer layer of the basic language model. This module includes shared LoRA parameters, expert LoRA parameters, a cultural matching scoring module, and an adaptive cultural routing module. A cross-cultural discourse generation model was trained using a multimodal Sino-Russian cross-border e-commerce cultural corpus. The model parameters were optimized using a multi-task cross-cultural joint loss. Under a fixed LoRA total rank budget, the rank of shared LoRA and the number of experts in each Transformer layer were adaptively adjusted according to the gradient contribution relationship between shared LoRA parameters and expert LoRA parameters. After training, experts were pruned according to their LoRA parameter importance scores to obtain the optimized cross-cultural discourse generation model. After deploying the cross-cultural discourse generation model, the model processes the product text and optional image information in the input Chinese or Russian product information, and constructs cross-cultural generation prompts based on the target language and target cultural market demand to form the model input sample. The model input sample is then input into the deployed model to generate corresponding marketing copy in the target cultural environment.

2. The method for generating interactive discourse in Sino-Russian cross-border e-commerce according to claim 1, characterized in that, The multimodal data extraction and standardization process specifically includes: The pre-trained Whisper-large-v3 model is used to extract the audio content from the product introduction video and generate the corresponding text introduction sequence. At the same time, the product introduction video is segmented at fixed time intervals to obtain the image introduction sequence. The text introduction sequence and the image introduction sequence are concatenated to obtain the product introduction data set. Remove HTML tags, web links, emoticons, duplicate text, and whitespace characters from marketing copy and customer service dialogue records. Segment the text according to periods, question marks, exclamation marks, and line breaks. When the length of a single text segment exceeds the preset length, divide it into a text segment by 512 consecutive tokens to obtain the marketing copy text and customer service dialogue text.

3. The method for generating interactive discourse in Sino-Russian cross-border e-commerce according to claim 1, characterized in that, The cross-cultural supervision samples include long-text cross-cultural supervision samples and short-text cross-cultural supervision samples; The long-text cross-cultural supervision sample is constructed as follows: using the product description data group in the standard Chinese e-commerce data as input and the text description sequence in the corresponding Russian product description data group as supervision labels, a long-text cross-cultural supervision sample from Chinese to Russian is constructed; and using the Russian product description data group as input and the Chinese text description sequence as output, a long-text cross-cultural supervision sample from Russian to Chinese is constructed. The method for constructing the short text cross-cultural supervision sample is as follows: key sentences that can express the same marketing intent or customer service semantics in Chinese marketing copy text and customer service dialogue text are extracted by a large language model, and a short text semantic correspondence relationship is established with the corresponding Russian marketing copy text and customer service dialogue text to construct the short text cross-cultural supervision sample. The bilingual autoregressive training sample is constructed as follows: Chinese product description text, marketing copy text, and customer service dialogue text are spliced ​​together in their original order to form a continuous Chinese text sequence. The corresponding Russian text is then used to form a continuous Russian text sequence in the same way. The bilingual autoregressive training sample is constructed using the next token prediction method.

4. The method for generating interactive discourse in Sino-Russian cross-border e-commerce according to claim 1, characterized in that, The visual feature extraction model uses CLIP-ViT-L / 14, the adapter includes two fully connected layers, and the base language model uses the pre-trained Qwen3-8B-Instruct; the base language model is kept frozen during the training of the cross-cultural discourse generation model.

5. The method for generating interactive discourse in Sino-Russian cross-border e-commerce according to claim 4, characterized in that, The cultural expert hybrid LoRA module specifically includes: For each Transformer layer of the base language model, multiple levels of LoRA are inserted into the Query and Value. The LoRA update of the l-th layer is represented as: , in, For the pre-trained Qwen3-8B-Instruct model The original matrix parameters of the layer, For the pre-trained Qwen3-8B-Instruct model The update matrix after applying hierarchical LoRA to the layer. For cross Shared LoRA parameters of the layers For the selection from the expert head pool targeting the first The first layer The update matrix corresponding to each expert, i.e., the LoRA parameters of each expert. For the corresponding weights, This represents the total number of experts at the current level. The corresponding LoRA rank number is , The corresponding rank is 1, the first The total rank of the layer is ,and and All are adaptive optimization parameters; The expert head pool consists of P LoRAs with rank 1, and P ≥ K.

6. The method for generating interactive discourse in Sino-Russian cross-border e-commerce according to claim 5, characterized in that, The cultural matching scoring module specifically includes: For each expert with rank 1 in the expert head pool, compute its low-rank increment representation: , in, Indicates the input number Feature representation before the Transformer layer Indicates the first expert in the first pool Parameters corresponding to each expert Indicates the first The expert in the The cultural semantic activation state generated by the layer, i.e., the low-rank incremental representation; Calculate the contribution of each LoRA expert to the current input: , Indicates the first The expert in the The contribution of the layer; The expert with the highest contribution is selected as the basic expert of this layer, and the cosine similarity of the update direction between the basic expert and the remaining experts is calculated; the expert with the lowest cosine similarity is selected as the second basic expert, and the expert set composed of the basic experts is used as the expert head pool corresponding to this layer. Iteratively select the expert with the lowest average cosine similarity to the currently selected expert set from the remaining experts as the new basic expert, and update the selected expert set until the required number of experts K for this layer is reached.

7. The method for generating interactive discourse in Sino-Russian cross-border e-commerce according to claim 6, characterized in that, The adaptive culture routing module specifically includes: For the z basic experts selected in the l-th layer, trainable gating parameters are introduced for each basic expert through a gating layer, and the fusion weights are obtained by Softmax normalization. During the training of the cross-cultural discourse generation model, the gating parameters of experts from different cultures in each layer are optimized and the contribution ratio of experts is adjusted.

8. The method for generating interactive discourse in Sino-Russian cross-border e-commerce according to claim 1, characterized in that, The multi-task cross-cultural joint loss is as follows: For long-text cross-cultural supervised samples and short-text cross-cultural supervised samples, the loss during training is the cross-entropy loss between the model output text and the real target cultural text; for bilingual autoregressive training samples, the autoregressive loss is calculated by the next token prediction method; the joint training loss Loss is obtained by weighted summation of the two types of losses; during the training process, bilingual autoregressive training samples are first used for model fine-tuning training, and then cultural supervised samples are used for training. The rank and number of experts sharing LoRA in each Transformer layer are adaptively adjusted as follows: Set the total LoRA rank budget for each Transformer layer to R, and R = r + K, initialize r = K = R / 2; During the training of the cross-cultural discourse generation model, the average gradient contribution corresponding to the shared LoRA parameters and the average gradient contribution corresponding to the expert LoRA parameters were statistically analyzed respectively: , , in, Indicates the first Layers share the average gradient contribution of LoRA to the current training task. This represents the average gradient contribution of the LoRA of the experts at this layer. denoted as the multi-task cross-cultural joint training loss, which is the weighted sum of the cross-entropy loss corresponding to the long text cross-cultural supervised samples and the short text cross-cultural supervised samples, and the autoregressive loss corresponding to the bilingual autoregressive training samples; when When an expert's rank is converted to a shared rank, When a shared rank is converted into the number of experts, a rank transfer operation is performed only once if the gradient contribution relationship remains consistent over three consecutive training cycles. The adaptive adjustment process is stopped when the cross-cultural discourse generation model has reached 80% of the training iterations.

9. The method for generating interactive discourse in Sino-Russian cross-border e-commerce according to claim 8, characterized in that, The process of pruning experts based on their LoRA parameter importance scores specifically includes: For the trained cross-cultural discourse generation model, calculate the importance score of the k-th expert in layer l: , in, This represents the importance score of the k-th expert in the l-th layer using LoRA; Set the expert pruning ratio parameters according to the actual computing resources and inference speed requirements during deployment. ; Rank all experts in the Transformer layers globally according to their importance scores, and select the experts with the lowest importance from the entire expert set. The proportional experts are deleted; the deleted experts no longer participate in the subsequent reasoning process, and the gating weights of the remaining experts are renormalized. The normalized gating weights are then used to reason about the remaining experts. The expert pruning ratio parameter is set as follows: Pre-measure the inference latency and memory usage of a single expert on the target deployment platform. Determine the ratio of the number of experts to be retained to the total number of experts based on the target inference speed requirements. Set 1 as the difference between the ratio of the number of LoRA experts to be retained and the total number of LoRA experts. .

10. The method for generating interactive discourse in Sino-Russian cross-border e-commerce according to claim 1, characterized in that, After deploying the cross-cultural discourse generation model, the product text and optional image information are processed for the input Chinese or Russian product information. Cross-cultural generation prompts are constructed by combining the target language and target cultural market demand to form the model input sample. The model input sample is then input into the deployed model to generate cross-cultural marketing copy, product description text or customer service response content in the target cultural environment. For the input Chinese or Russian product information, the product text and optional image information are processed, and cross-cultural generated prompts are constructed based on the target language and target cultural market demands to form the model input sample, specifically including: Obtain the Chinese or Russian product information to be converted, wherein the product information includes product text description and / or product introduction video and / or product image information; when the input product information includes a product introduction video, perform text extraction and image extraction on the product introduction video to obtain the corresponding product introduction text information and product image information; when the input product information only contains text information, directly obtain the product text description. The cross-cultural generated prompts are used to specify the source language, target language, and target cultural market demand, and are combined with product text descriptions, product introduction text information, and product image information to construct input samples.

Images