An artificial intelligence-based web text translation system

Abstract

The application relates to the technical field of data processing, and provides a web text translation system based on artificial intelligence. A feature extraction module generates genre, style vector, cultural marker sequence and rhythm feature vector through a deep neural network, solves the problem that existing technologies cannot accurately capture unique properties of web texts, and provides fine-grained basis for translation. A dynamic translation module dynamically adjusts decoding parameters, interruption timing and translation method selection, realizes style restoration, cultural adaptation and rhythm matching, and overcomes the defects of traditional translation style distortion, cultural element processing stiffness and rhythm misplacement. A feedback optimization module generates and incrementally updates knowledge vectors, so that the system can continuously learn new language phenomena and user needs, solves the poor adaptability of static systems, and the three work together to improve the accuracy, adaptability and efficiency of web text translation.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to an artificial intelligence-based online text translation system. Background Technology

[0002] With the surge in demand for the global dissemination of online literature, machine translation technology has gradually become a core support for the overseas expansion of online literature. Currently, mainstream translation technology has evolved from rule-based and statistical methods to the neural machine translation stage, achieving end-to-end translation of general texts through pre-trained language models, and making significant progress in sentence fluency and semantic accuracy. For the online literature field, some technical solutions attempt to improve the professionalism of translation by fine-tuning models through domain corpora or building specialized terminology databases, thereby achieving automated processing of online literature translation.

[0003] Although machine translation has been applied in the online literature field, it still suffers from the following key shortcomings: Existing technologies largely focus on literal semantic conversion, failing to capture the unique genre characteristics and stylistic attributes of online literature, resulting in generic translations that lose the stylistic distinctiveness of the original work; online literature contains a large number of internet slang, subcultural memes, and cultural metaphors, and existing technologies, through literal translation or simple substitution, cannot accurately transfer cultural connotations, easily causing comprehension barriers for the target audience; furthermore, online literature language is rapidly evolving, and users have diverse preferences for translation styles, but the model parameters and terminology databases of existing technologies are mostly static, making it difficult to achieve rapid iterative optimization through efficient feedback mechanisms. Existing technologies have not solved the technical problem of how to dynamically adjust and iterate translation strategies based on the fine-grained characteristics of the online literature domain. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this application provides an artificial intelligence-based online text translation system, which includes: a feature extraction module, a dynamic translation module, and a feedback optimization module;

[0005] The feature extraction module is used to process online text data through a deep neural network to output the genre vector and style vector of the online text, combine it with a preset online text knowledge graph to output a cultural marker sequence, and determine the dialogue density, emotional fluctuation variance and plot density in the online text data to generate a rhythm feature vector.

[0006] The dynamic translation module is used to encode the web article text data, genre vector, style vector and rhythm feature vector as conditional signals through a self-attention encoder. During the decoding process, the interruption timing is determined based on the cultural marker sequence, and the equivalent expression library is queried to select the web article translation method to dynamically output the web article translation text. At the same time, the temperature parameters and sampling strategy of the decoder are controlled according to the style vector, and the bundle search parameters are dynamically adjusted based on the rhythm feature vector.

[0007] The feedback optimization module is used to receive and process user feedback data and web crawling data to generate knowledge vectors. It trains the knowledge vectors based on efficient parameter fine-tuning technology to incrementally update the equivalent expression library and web article knowledge graph.

[0008] As an optional implementation, the genre vector and style vector of the output web article include:

[0009] The online text data is preprocessed, and the preprocessed online text data is feature-encoded using a deep neural network to generate a hidden layer feature sequence.

[0010] The hidden feature sequence is used to focus on the genre semantics of online articles to output a genre feature matrix, and the hidden feature sequence is used to capture the language expression patterns of online articles to output a style feature matrix.

[0011] Global pooling is performed on the genre feature matrix and style feature matrix respectively to output the genre vector and style vector of the online article.

[0012] As an optional implementation, the output cultural marker sequence includes:

[0013] Load a pre-defined online literature knowledge graph, which includes cultural elements and classification attributes of the online literature field;

[0014] The hidden layer feature sequence encoded by the deep neural network is semantically similar to the cultural element features in the online article knowledge graph to identify the matching cultural element features in the online article text data.

[0015] Based on the classification attributes of the successfully matched cultural elements in the online article knowledge graph, type labeling is performed, and a cultural label sequence is output.

[0016] As an optional implementation, the generation of the rhythm feature vector includes:

[0017] The online text data is segmented into paragraphs, and the proportion of dialogue sentences in each paragraph is calculated to form the dialogue density.

[0018] The sentiment of each dialogue sentence is analyzed and quantified into a sentiment value. The variance of the sentiment value is calculated through a sliding window to form the sentiment fluctuation variance.

[0019] Identify plot events in online text data, calculate the number of plot events and the number of sentences between them in each paragraph, and form plot density;

[0020] The system processes dialogue density, emotion fluctuation variance, and plot density using a long short-term memory network to output a temporal feature matrix. Based on the temporal feature matrix, the rhythm intensity of each natural segment is calculated, and the rhythm intensity is weighted and summed using a self-attention mechanism to generate a rhythm feature vector.

[0021] As an optional implementation, the dynamically output translated online text includes:

[0022] The text encoding sequence of online article data is hierarchically fused with genre vectors, style vectors, and rhythm feature vectors to form a joint encoding vector;

[0023] During the interruption, the equivalent expression library is queried based on the classification attributes of the cultural marker sequence to obtain candidate translations and calculate the semantic similarity between the candidate translations and the joint encoding vector, and a subset of translations is selected.

[0024] The length and sentence complexity of each translation in the translation subset are matched with the rhythm intensity of the corresponding natural paragraph in the rhythm feature vector to select the web text translation method based on the matching fitness.

[0025] After the interruption ends, the decoder starts translation based on the selected web text translation method to output the translated web text. It also uses a self-attention mechanism to associate the translated web text before and after the interruption, and adjusts the sentence structure of the translated web text to smooth the translated web text.

[0026] As an optional implementation, determining the interruption timing based on the cultural marker sequence includes:

[0027] Based on the cross-cultural differences of cultural elements in the online literature knowledge graph, the cultural marker sequence is prioritized to form a marker priority.

[0028] After the self-attention encoder outputs the text encoding sequence, the relationship between the decoding position of the self-attention encoder and the position of the cultural tag sequence in the online text data is tracked in real time. When the self-attention encoder processes the online text data that has been classified as high-priority tags, it is determined to be an interruption opportunity.

[0029] After an interruption is triggered based on the timing of the interruption, the decoding step size is dynamically determined by querying the number of online translations of the cultural marker sequence in the equivalent representation library, according to the complexity of the cultural marker sequence.

[0030] As an optional implementation, the temperature parameters and sampling strategy of the control decoder include:

[0031] Feature analysis is performed on the style vector to extract multiple feature components that represent the language expression patterns of online articles. The feature components include expression certainty, language formality, and emotional expression intensity.

[0032] The temperature parameters of the decoder are determined based on the expressive determinism, and the sampling strategy of the decoder is determined by comprehensively considering the formality of the language and the intensity of emotional expression.

[0033] The determined temperature parameters and sampling strategy are input into the decoder to regulate the sentence structure of the decoder during the generation of online article translation text.

[0034] As an optional implementation, the dynamically adjusted beam search parameters include:

[0035] The rhythm feature vector is decomposed temporally to extract the rhythm intensity, rhythm change rate and rhythm type of each natural segment;

[0036] The bundle width is adjusted based on the rhythm intensity to search the bundle width, and the filtering threshold of the translation subset is dynamically set in combination with the rhythm change rate. The decoder termination condition is also adjusted according to the rhythm type.

[0037] As an optional implementation, the generation of knowledge vectors includes:

[0038] It receives user feedback data and web crawling data, preprocesses the user feedback data and web crawling data, uses an attention mechanism to generate feedback feature vectors from the user feedback data, and performs semantic alignment between the web crawling data and the online article knowledge graph to generate crawling feature vectors.

[0039] The semantic similarity between the feedback feature vector and the crawled feature vector is calculated to adjust the fusion ratio, and time decay weights are assigned to the feedback feature vector and the crawled feature vector obtained in different time windows to fuse and generate a heterogeneous feature matrix that includes user feedback intent and new network knowledge.

[0040] Generate knowledge vectors from heterogeneous feature matrices using deep neural networks.

[0041] As an optional implementation, the incremental update equivalent expression library includes:

[0042] Calculate the domain relevance between the knowledge vector and the translation in the equivalent representation library to obtain the domain knowledge vector;

[0043] Based on domain knowledge vector parsing, new translation methods are generated, and a candidate set of translation methods is generated. The priority of the candidate set of translation methods is comprehensively ranked according to accuracy, timeliness and style adaptability in order to select alternative translation methods.

[0044] The selected alternative translations are added to the equivalent expression library, and an initial weight is assigned to each alternative translation. When an alternative translation is selected, the initial weight of the alternative translation is adjusted in real time based on user feedback data of the online article translation text.

[0045] Semantic matching is performed between domain knowledge vectors and online text knowledge graphs to verify the consistency of cultural element features and classification attributes of candidate translations. When there is a discrepancy, incremental updates of the online text knowledge graph are triggered.

[0046] Compared with existing technologies, the beneficial effects of this application are as follows: The feature extraction module generates genre, style vectors, cultural marker sequences, and rhythm feature vectors through deep neural networks, solving the problem that existing technologies are unable to accurately capture the unique attributes of online texts, and providing fine-grained basis for translation; The dynamic translation module dynamically adjusts decoding parameters, interruption timing, and translation method selection to achieve style restoration, cultural adaptation, and rhythm matching, overcoming the defects of traditional translation style distortion, stiff handling of cultural elements, and rhythm misalignment; The feedback optimization module enables the system to continuously learn new language phenomena and user needs through knowledge vector generation and incremental update mechanisms, solving the problem of poor adaptability of static systems. The three modules work together to improve the accuracy, adaptability, and efficiency of online text translation. Attached Figure Description

[0047] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0048] Figure 1 A system flowchart of an artificial intelligence-based online text translation system provided for embodiments of this application;

[0049] Figure 2 An output cultural marker sequence diagram of an AI-based online text translation system provided in this application embodiment;

[0050] Figure 3 This application provides an embodiment of an AI-based online text translation system that dynamically outputs translated online text. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of this application more apparent and understandable, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0052] Example

[0053] like Figure 1 The diagram shown is a system flowchart of an AI-based online text translation system provided in this application embodiment. The system includes a feature extraction module, a dynamic translation module, and a feedback optimization module.

[0054] The feature extraction module is used to process online text data through deep neural networks to output genre vectors and style vectors of online texts. It combines a preset online text knowledge graph to output a cultural marker sequence and determines the dialogue density, emotional fluctuation variance and plot density in the online text data to generate rhythm feature vectors.

[0055] Specifically, the output of the genre vector and style vector for online articles includes:

[0056] The online text data is preprocessed, and the preprocessed online text data is feature-encoded using a deep neural network to generate a hidden layer feature sequence.

[0057] The hidden feature sequence is used to focus on the genre semantics of online articles to output a genre feature matrix, and the hidden feature sequence is used to capture the language expression patterns of online articles to output a style feature matrix.

[0058] Global pooling is performed on the genre feature matrix and style feature matrix respectively to output the genre vector and style vector of the online article.

[0059] Online text data often contains redundant information and non-standard expressions, which can interfere with deep neural networks' capture of core semantics. Therefore, preprocessing is necessary to unify the format and enhance effective information. First, regular expressions are used to match and remove non-text content such as chapter titles. Second, a Unicode encoding filter is used to perform basic cleaning of the online text data, removing garbled characters, emoticons, and special characters. Then, a bidirectional maximum matching word segmentation algorithm is used to segment the online text data. The word segmentation dictionary needs to be pre-imported with terms specific to the online literature field, such as "quick transmigration," "system," and "cultivation," to ensure that special terms are not split. The online text data is then divided into the smallest semantic units. Finally, the word segmentation results are converted into a text sequence that the deep neural network can recognize, preparing for subsequent feature encoding. This process removes irrelevant information interference, ensures the semantic integrity and format uniformity of the text sequence, and allows the deep neural network to focus on the core content of the online text for feature encoding.

[0060] The preprocessed text sequence is a discrete word index, which needs to be transformed into continuous semantic features through a deep neural network in order to further extract the implied genre and style information. A Transformer-based deep neural network processes the preprocessed text sequence. This deep neural network uses a multi-layer encoder, each layer consisting of a multi-head self-attention mechanism and a feedforward neural network. After the preprocessed text sequence is input into the deep neural network, the word index is first transformed into word vectors through an embedding layer, and then positional encoding is added to reflect the sequential relationship of words. The multi-head self-attention mechanism captures the contextual semantics of the text by calculating the association weights between different words, such as the association between cultivation, spiritual power, and realm. The feedforward neural network performs a non-linear transformation on the attention output to enhance feature expression, thereby encoding the semantic information in the text sequence layer by layer. The contextual associations and word meanings of the text sequence are transformed into feature values, ultimately generating a hidden feature sequence containing deep semantics and contextual information. This transforms the discrete text sequence into a continuous hidden feature sequence, comprehensively capturing the semantic associations in the online text data and providing a rich semantic foundation for subsequent separation of genre and style features.

[0061] The hidden layer feature sequence contains multiple semantic information such as genre and style. It is necessary to achieve feature separation through targeted extraction to avoid mutual interference between different attribute features and ensure that the subsequently generated vectors can accurately represent their respective attributes. Two parallel feature extraction branches are set at the top layer of the Transformer encoder, including a genre extraction branch and a style extraction branch. In the genre extraction branch, a preset genre vocabulary is loaded, such as sects and tribulations in the fantasy genre, or interstellar and mecha in the science fiction genre. The words in the genre vocabulary are converted into genre word vectors through the embedding layer. The similarity between the hidden layer feature sequence and the genre word vectors is calculated to generate genre attention weights, so that the deep neural network focuses on the genre semantics of the online article. Local genre features are extracted through convolutional layers to output a genre feature matrix.

[0062] In the style extraction branch, a syntactic analyzer identifies sentence structures in the text sequence, such as exclamatory and rhetorical questions, and labels sentiment-related words, such as witty sentences in a humorous style and rigorous expressions in a serious style, to obtain sentence feature vectors and sentiment word vectors. Then, the similarity between the hidden layer feature sequence and the sentence feature vector and sentiment word vector is calculated to generate style attention weights, which guide the deep neural network to capture the language expression patterns of online articles. Style features are extracted through convolutional layers to output a style feature matrix. This achieves effective decoupling of genre features and style features, allowing the two feature matrices to focus on their respective core attributes, avoiding feature confusion between different genres within the same style, and improving the targeting of the feature matrix.

[0063] The genre feature matrix and style feature matrix are two-dimensional structures with sequence dependencies. However, the dynamic translation module requires fixed-dimensional vectors as conditional signals. Global pooling is used to achieve dimensionality compression and feature aggregation. For the genre feature matrix, a strategy combining max pooling and average pooling is employed. Max pooling extracts the most significant genre features from the matrix, such as the features related to realm breakthroughs in the cultivation genre. Average pooling preserves the overall genre feature distribution. The two are concatenated and mapped to a genre vector through a fully connected layer. The same operation is applied to the style feature matrix, focusing on core style features such as formality and emotional intensity, and outputting a style vector. This ensures that both vectors reflect the core information of their respective feature matrices. Thus, while preserving core features, the high-dimensional feature matrix is ​​compressed into fixed-dimensional vectors. This reduces data dimensionality, facilitating downstream module processing, and ensures that the vectors accurately represent genre and style attributes. The generated genre and style vectors serve as conditional signals for the dynamic translation module, directly influencing the decoder's temperature parameters and sampling strategies, providing a basis for achieving style-adaptive translation.

[0064] Specifically, such as Figure 2 As shown, the output cultural marker sequence includes:

[0065] Load a pre-defined online literature knowledge graph, which includes cultural elements and classification attributes of the online literature field;

[0066] The hidden layer feature sequence encoded by the deep neural network is semantically similar to the cultural element features in the online article knowledge graph to identify the matching cultural element features in the online article text data.

[0067] Based on the classification attributes of the successfully matched cultural elements in the online article knowledge graph, type labeling is performed, and a cultural label sequence is output.

[0068] It needs to be explained that the initial construction method of the online literature knowledge graph is to crawl online literature texts from multiple genres and authoritative cultural interpretation materials. Through named entity recognition and phrase extraction technology, core cultural elements such as popular online slang, subcultural memes, cultural metaphors, and domain terms are extracted to form an original element library. Then, based on the semantic attributes and usage scenarios of the elements, they are stored as three-level nodes. Finally, through co-occurrence analysis and semantic reasoning, the relationships between elements are labeled, and combined with cross-cultural comparison materials, attributes such as cross-cultural differences and translation difficulty are added to the elements to form the initial online literature knowledge graph, which is stored in a graph database.

[0069] Online literature text data contains a large number of domain-specific cultural elements, such as internet memes and subcultural symbols. The translation of these elements requires combining their specific meanings and classifications. The pre-set online literature knowledge graph provides structured knowledge support for their identification and marking. During the processing, the pre-set online literature knowledge graph is first loaded. The pre-set online literature knowledge graph is stored in a graph structure and contains three levels of nodes. The first-level node is the major cultural category, such as internet catchphrases and traditional cultural allusions. The second-level node is the sub-type, such as homophone memes and abbreviation memes under internet catchphrases. The third-level node is the specific cultural element, such as yyds and jue jue zi, etc. Each third-level node is associated with the cultural element characteristics and classification attributes. For example, the attribute of yyds is an exaggerated praise type of internet meme. The online literature knowledge graph is loaded through a graph database. This online literature knowledge graph contains common cultural elements in the online literature field. Each cultural element has its unique cultural element characteristics and clear classification attributes, providing a reference standard for the subsequent identification and marking of cultural element characteristics; providing a standardized knowledge reference for the identification of cultural element characteristics, ensuring that there is a unified basis for the identification and marking process, and avoiding deviations caused by subjective judgment.

[0070] The hidden layer feature sequence contains the deep semantics of the text. By matching with the cultural element characteristics in the online literature knowledge graph, the cultural elements in the text can be accurately located, avoiding missed or misjudged cases caused by relying solely on literal matching. The hidden layer feature sequence encoded by the deep neural network is segmented into subsequences by sentence. Each subsequence corresponds to the semantic characteristics of a sentence. Calculate the cosine similarity between each subsequence and the cultural element characteristics of all third-level nodes in the online literature knowledge graph as the semantic similarity between the two. Compare the semantic similarity with the similarity threshold to determine whether the subsequence contains matching cultural element characteristics, so as to find the part of the hidden layer feature sequence that matches the cultural elements in the online literature knowledge graph, and thus identify the cultural element characteristics in the online literature text data; thereby improving the accuracy of cultural element identification through deep semantic matching, and at the same time, the marking of suspected elements enhances the comprehensiveness of identification, avoiding the omission of emerging cultural elements.

[0071] Cultural elements with different classification attributes require different translation strategies. For example, traditional cultural allusions need to retain their cultural connotations, while internet memes need to retain their transmissibility. The type marking can clarify the processing direction of each element. For the successfully matched cultural element characteristics, query their corresponding second-level classification attributes in the online literature knowledge graph, and mark them in the format of element position and classification attribute. The element position is determined by the start and end indexes of the subsequence in the online literature text data. Arrange all the markings in the text order to form a cultural marking sequence; thus making the types of cultural elements clearly distinguishable, providing a clear processing basis for the dynamic translation module, ensuring that the translation strategy is adapted to the element type. The cultural marking sequence is directly input into the dynamic translation module, providing a reference for the interruption timing judgment and translation method selection, and affecting the rationality of the translation process and the quality of the translation.

[0072] Specifically, generating rhythmic feature vectors includes:

[0073] The online text data is segmented into paragraphs, and the proportion of dialogue sentences in each paragraph is calculated to form the dialogue density.

[0074] The sentiment of each dialogue sentence is analyzed and quantified into a sentiment value. The variance of the sentiment value is calculated through a sliding window to form the sentiment fluctuation variance.

[0075] Identify plot events in online text data, calculate the number of plot events and the number of sentences between them in each paragraph, and form plot density;

[0076] The system processes dialogue density, emotion fluctuation variance, and plot density using a long short-term memory network to output a temporal feature matrix. Based on the temporal feature matrix, the rhythm intensity of each natural segment is calculated, and the rhythm intensity is weighted and summed using a self-attention mechanism to generate a rhythm feature vector.

[0077] Dialogue is a crucial element of online novel rhythm, and its density reflects the intensity of interaction within paragraphs. For instance, high-density dialogue often corresponds to tense plots, requiring quantitative calculations to provide foundational data for rhythm analysis. This process involves segmenting the online novel text data using line breaks to obtain natural paragraphs, defining the scope of each paragraph, identifying dialogue sentences through rule matching, and counting the number of dialogue sentences and the total number of sentences within each paragraph. The total number of sentences is marked with periods, question marks, and exclamation marks at the end. The ratio of the number of dialogue sentences to the total number of sentences is calculated to determine the dialogue density of each paragraph. This transforms the distribution characteristics of dialogue into quantifiable indicators, objectively reflecting the interactive attributes of paragraphs and providing foundational data for rhythm feature extraction.

[0078] The intensity of emotional fluctuations affects readers' perception of rhythm. For example, rapid emotional changes correspond to a tense rhythm. Quantifying emotional fluctuations through variance can enrich the dimensions of rhythmic features. Using BERT-based fine-tuning, Chinese sentiment analysis is performed on each dialogue sentence to output a sentiment value in the range [-1, 1]. A negative sentiment value indicates a negative tone, and a positive value indicates a positive tone. A sliding window is set, and the variance of the sentiment value within each window is calculated to obtain the variance of the emotional fluctuations corresponding to each window. This quantifies the intensity of emotional changes, so that rhythmic features not only include structural information but also incorporate emotional factors, thus more comprehensively reflecting the rhythm of online novels.

[0079] Plot events are the core of driving the story forward, and their distribution density directly determines the pace of a paragraph. For example, a high density of events corresponds to a fast pace, and quantitative calculations are needed to reflect the tightness of the plot progression. Based on existing trigger word extraction models, key plot events in online novel text data are identified, such as the outbreak of battle and the revelation of identities. Then, the number of plot events in each paragraph is counted, and the number of sentences between adjacent plot events is calculated. The ratio of the number of events to the number of sentences between them is used as the plot density of that paragraph. This quantifies the distribution characteristics of plot events, accurately reflects the tightness of the plot in a paragraph, and provides a key indicator for rhythm analysis.

[0080] Dialogue density, sentiment variance, and plot density are discrete rhythm indicators that need to be integrated into a unified rhythm feature vector through temporal modeling and aggregation operations to provide overall rhythm information for the dynamic translation module. Dialogue density, sentiment variance, and plot density are arranged into a temporal sequence according to the order of natural paragraphs. This temporal sequence is then input into a Long Short-Term Memory (LSTM) network. The LSM network uses a gating mechanism to capture temporal dependencies in the sequence; for example, an increase in plot density is accompanied by an increase in sentiment variance. This results in a temporal feature matrix, which is then used to calculate the rhythm intensity of each natural paragraph to reflect the strength of the paragraph's rhythm. Finally, a self-attention mechanism is used to weight the rhythm intensity, with paragraphs exhibiting drastic rhythm changes receiving higher weights. The weighted sum is then used to output the rhythm feature vector. Integrating these discrete rhythm indicators into a unified vector comprehensively and selectively reflects the rhythmic characteristics of the online text, meeting the rhythm information requirements of the dynamic translation module. The rhythm feature vector provides a basis for adjusting the beam search parameters in the dynamic translation module, ensuring that the rhythm of the translated text remains consistent with the original text.

[0081] The dynamic translation module uses web article text data, genre vectors, style vectors, and rhythm feature vectors as conditional signals, which are then encoded by a self-attention encoder. During the decoding process, the module determines the interruption timing based on the cultural marker sequence and queries the equivalent expression library to select the web article translation method, so as to dynamically output the translated web article text. At the same time, it controls the temperature parameters and sampling strategy of the decoder based on the style vector and dynamically adjusts the bundle search parameters based on the rhythm feature vector.

[0082] It needs to be explained that the initial construction method of the equivalent expression library is to collect high-quality bilingual corpora of online articles, such as published translations and translations by professional translators, and to screen texts covering multiple schools and styles to ensure that they contain rich terminology, cultural elements, and sentence structure translations. Then, the parallel corpora are bilingually aligned to extract the correspondence between the original text terms or phrases and the translations, forming initial translation pairs. Synonymous translations are merged through semantic similarity calculation. Finally, the translations are stored according to cultural elements, and each translation is labeled with its applicable scenarios. Based on the frequency of the translation in the parallel corpora and human scoring, an initial weight is assigned, with high-frequency and high-quality translations having higher weights for priority retrieval in subsequent queries.

[0083] Furthermore, determining the timing of the interruption based on the cultural marker sequence includes:

[0084] Based on the cross-cultural differences of cultural elements in the online literature knowledge graph, the cultural marker sequence is prioritized to form a marker priority.

[0085] After the self-attention encoder outputs the text encoding sequence, the relationship between the decoding position of the self-attention encoder and the position of the cultural tag sequence in the online text data is tracked in real time. When the self-attention encoder processes the online text data that has been classified as high-priority tags, it is determined to be an interruption opportunity.

[0086] After an interruption is triggered based on the timing of the interruption, the decoding step size is dynamically determined by querying the number of online translations of the cultural marker sequence in the equivalent representation library, according to the complexity of the cultural marker sequence.

[0087] The difficulty of cross-language translation varies significantly among different cultural elements. For example, cultivation-related terms are more difficult to translate accurately than everyday conversations. Prioritization is necessary to ensure that high-difficulty elements receive focused processing, avoiding uneven translation quality due to uniform treatment. The cross-cultural difference degree of each cultural element is retrieved from the online literature knowledge graph. This cross-cultural difference degree is pre-set based on the historical translation error rate and cultural specificity of the cultural element in different languages. For each marker in the cultural marker sequence, the cross-cultural difference degree is compared with a difference threshold to classify the markers into three priorities. For example, cultivation-related terms have high cross-cultural difference due to the lack of direct correspondences in other cultures and are classified as high priority. Traditional festivals such as the Spring Festival, although culturally different, have common translations and are classified as medium priority. Everyday words such as eating have no significant differences and are classified as low priority. If a marker contains multiple cultural elements, such as cultivators celebrating the Spring Festival, the cultivation element with the highest cross-cultural difference degree is designated as high priority. This differentiated prioritization clarifies the allocation of translation resources, ensuring that high-difficulty cultural elements receive more refined processing and reducing the risk of translation errors from the source.

[0088] The self-attention decoder needs to pause the regular translation process before processing high-priority cultural elements to allow time for querying equivalent expression libraries and selecting translations. If the corresponding position is missed and then interrupted, the translation of cultural elements will be out of context. When the self-attention encoder outputs the text encoding sequence, it simultaneously records the mapping relationship between each encoding position and the original paragraph and sentence. When the self-attention encoder outputs the text encoding sequence, it simultaneously records the natural paragraph and sentence index of the web text data corresponding to each encoding position. The position indexer compares the decoder's current decoding position with the starting position of the high-priority marker in the cultural marker sequence in real time. For example, when the decoder decodes a natural paragraph, if the decoding position corresponds to the starting position of a high-priority marker, it is determined to be an interruption opportunity, and the decoding process is paused. If the high-priority marker contains multiple consecutive sentences of text, such as a dense paragraph of internet slang in a long dialogue, the starting position of the entire marker interval is used as the interruption trigger point to trigger the interruption. In this way, precise position tracking ensures that the interruption timing is completely synchronized with the processing rhythm of high-priority cultural elements, avoiding contextual fragmentation caused by premature interruption or insufficient translation preparation caused by delayed interruption.

[0089] The complexity of different cultural markers varies significantly, requiring different translation selection times. A fixed decoding step size can lead to inefficient processing of simple elements or insufficient preparation of translations for complex elements. After an interruption is triggered, the complexity of the current high-priority markers is analyzed, such as the number of cultural elements contained in the marker and the degree of semantic relevance. At the same time, the number of corresponding web text translations for that marker is queried in the equivalent expression library. The more web text translations there are, the more time is needed for selection. For example, a single cultivation term marker has low complexity and few web text translations, so the decoding step size can be set to a shorter one. However, when a combination of cultivation allusions and internet memes is marked with high complexity and many web text translations, the decoding step size needs to be extended. If the number of web text translations for that marker in the equivalent expression library is zero, it is a completely new cultural element, and the step size is automatically extended to start a temporary translation generation process. The dynamic step size mechanism ensures that the interruption duration is precisely matched with the processing difficulty of the cultural markers, maximizing translation efficiency while ensuring the quality of translation selection and avoiding resource waste. A reasonable decoding step size provides sufficient time for the translation selection process in the dynamically output web text translation, ensuring that the final selected translation is both accurate and context-appropriate.

[0090] Specifically, such as Figure 3 As shown, the dynamically output translated text of online articles includes:

[0091] The text encoding sequence of online article data is hierarchically fused with genre vectors, style vectors, and rhythm feature vectors to form a joint encoding vector;

[0092] During the interruption, the equivalent expression library is queried based on the classification attributes of the cultural marker sequence to obtain candidate translations and calculate the semantic similarity between the candidate translations and the joint encoding vector, and a subset of translations is selected.

[0093] The length and sentence complexity of each translation in the translation subset are matched with the rhythm intensity of the corresponding natural paragraph in the rhythm feature vector to select the web text translation method based on the matching fitness.

[0094] After the interruption ends, the decoder starts translation based on the selected web text translation method to output the translated web text. It also uses a self-attention mechanism to associate the translated web text before and after the interruption, and adjusts the sentence structure of the translated web text to smooth the translated web text.

[0095] The text encoding sequence only contains surface semantics, while genre vectors, style vectors, and rhythm feature vectors are key factors determining translation strategies. These need to be integrated into a unified signal through fusion to ensure that the chosen translation method simultaneously adapts to both text content and stylistic attributes. First, the text encoding sequence output by the self-attention encoder is fused with the genre vector. Vector superposition strengthens the influence of genre features such as fantasy and science fiction on the encoding. Next, the fused result is fused with the style vector, using an attention mechanism to highlight the weight of style features such as humor and seriousness in the encoding. Finally, it is fused with the rhythm feature vector, using hierarchical processing to enhance the speed... Rhythm feature vectors, such as rhythm and slow rhythm, are incorporated into the encoding. For example, after the text encoding of the fantasy genre is fused with the genre vector, the encoding weight of terms such as spiritual power and realm will be strengthened. Humorous text will highlight the encoding features of light-hearted expression, ultimately forming a joint encoding vector. Hierarchical fusion not only preserves the independent features of each vector, but also realizes the organic connection of cross-dimensional information, so that the joint encoding vector can comprehensively reflect the integrated translation needs of content, style and rhythm. The joint encoding vector provides multi-dimensional evaluation criteria for the selection of translation subsets, ensuring that the candidate translations are highly matched with the original online text data in terms of semantics, style and rhythm.

[0096] The equivalent expression library contains a large number of translations, and not all translations are suitable for the genre and style of the current text. It is necessary to narrow down the scope by categorizing attributes and filtering semantic similarity to improve the accuracy of translation selection. During the interruption, based on the categorization attributes of cultural marker sequences, the corresponding translations are queried in the equivalent expression library. After filtering out irrelevant translations, candidate translations are finally obtained. For the retrieved candidate translations, they are converted into feature vectors by an encoder, and then semantic similarity is calculated with the joint encoding vector. Candidate translations with high similarity are retained to form a translation subset. If the translation subset contains high-quality translations marked by the feedback optimization module, they will be retained first and their ranking will be automatically improved. Through categorized retrieval and semantic filtering, candidate translations that are highly suitable for the features of the current text can be quickly identified from the massive number of translations, which greatly improves the efficiency and accuracy of translation selection.

[0097] The same cultural elements require different translations in paragraphs with different rhythms. For example, fast-paced battles require short sentences, while slow-paced descriptions require long sentences. Semantic matching alone cannot meet the rhythm adaptation requirements. This paper analyzes the length and sentence complexity of each translation in the subset of translation methods. Length refers to the number of words in a sentence, and sentence complexity refers to whether it contains clauses and the number of modifiers. The rhythm intensity of the corresponding natural paragraph is extracted from the rhythm feature vector. For example, battle scenes have high rhythm intensity, while descriptive scenes have low rhythm intensity. The length and sentence complexity of the translation are then correlated with the rhythm intensity. The system compares different translation methods, such as matching fast-paced paragraphs with short sentences and simple sentence structures, and slow-paced paragraphs with long sentences and complex sentence structures. The system scores the translations based on their suitability; for example, short sentence translations receive higher scores when matched with high-paced text. The translation with the highest suitability is then selected. If a translation is semantically suitable but rhythmically incompatible, the sentence structure is adjusted, for example, by splitting long sentences to fit the fast pace. Through this assessment of the suitability between formal features and rhythmic intensity, the system ensures that the selected translation is not only semantically accurate but also maintains consistency with the original text in terms of rhythm, thus improving the reading fluency of the translation.

[0098] Interruptions cause temporary pauses in the translation process. Directly splicing the translations before and after the interruption can result in abrupt sentence structure and logical breaks, requiring smoothing to ensure coherence. After the interruption ends, the decoder initiates translation using the selected web text translation method, generating a translated fragment of the current paragraph based on the context to output the translated text. A self-attention mechanism analyzes the sentence structure features and conjunction usage habits of the pre-interruption translation, including active and passive sentences. The sentence structure of the current translated fragment is adjusted based on these features and conjunction usage habits; for example, if the pre-interruption translation used more active sentences, the current translation will also primarily use active sentences. By adding transitional words such as "however" and "in addition" at the connection points before and after the interruption, the contextual logic becomes more coherent. If the interruption involves dialogue, the consistency of the character's tone is maintained. For example, if a character is accustomed to using colloquial expressions, this feature must be carried through in the translation. Through the association and sentence structure adjustment of the self-attention mechanism, the translation gaps caused by the interruption are completely eliminated, making the entire online article translation text an organic whole in terms of semantic logic and language style, improving the reader experience. The smoothed online article translation text provides high-quality output for the entire dynamic translation process, and its sentence structure and style features also provide reference samples for the user feedback collection of the feedback optimization module.

[0099] Specifically, the temperature parameters and sampling strategy for controlling the decoder include:

[0100] Feature analysis is performed on the style vector to extract multiple feature components that represent the language expression patterns of online articles. The feature components include expression certainty, language formality, and emotional expression intensity.

[0101] The temperature parameters of the decoder are determined based on the expressive determinism, and the sampling strategy of the decoder is determined by comprehensively considering the formality of the language and the intensity of emotional expression.

[0102] The determined temperature parameters and sampling strategy are input into the decoder to regulate the sentence structure of the decoder during the generation of online article translation text.

[0103] Style vectors are a comprehensive representation of the overall language style of online texts. Without extracting specific feature components, they cannot be directly used for precise control of decoder parameters. Certainty of expression, formality of language, and intensity of emotional expression are core dimensions determining the flexibility, rigor, and emotional tone of translated sentences, requiring separate analysis for targeted adjustment. A feature parsing network is used to analyze style vectors. This network contains three parallel fully connected layers, each corresponding to the extraction of feature components representing the language expression patterns of online texts. For certainty of expression, the network identifies feature dimensions related to words such as affirmation and necessity in the style vector to analyze the degree of certainty in the text. For formality of language, it focuses on features related to written and colloquial language to distinguish whether the text is academic or conversational. For intensity of emotional expression, it determines the intensity of the text's emotion by associating the feature weights of emotion words in the style vector. During the parsing process, the accuracy of each feature component is calibrated by comparing it with a pre-set style template vector. This transforms the abstract style vector into concrete, quantifiable feature components, providing a clear control basis for determining subsequent temperature parameters and sampling strategies, avoiding ambiguity in style control.

[0104] The decoder's temperature parameter controls the randomness of the output, with low temperatures corresponding to deterministic output and high temperatures to diverse output. This must match the determinism of the text's expression. The sampling strategy determines the range and tendency of vocabulary selection, adapting to both formality and emotional intensity. When the expression is highly deterministic, such as in popular science articles, a low temperature parameter is set to limit the decoder's random selection, ensuring stable sentence structure and vocabulary—that is, prioritizing common and clear expressions. When the expression is less deterministic, such as in imaginative descriptions in fictional articles, the temperature parameter is increased to enhance output diversity, allowing for synonym substitution and sentence variations. Regarding the sampling strategy, if... For texts with high formality and low emotional intensity, such as academic online novels, a restrictive sampling strategy is used, selecting only from a high-frequency and rigorous vocabulary pool. For texts with low formality and high emotional intensity, such as youth campus online novels, an extended sampling strategy is used, allowing the inclusion of colloquial and emotionally rich vocabulary. If both characteristics are mixed, such as strong emotional expression interspersed in formal narration, the proportion of vocabulary in the sampling strategy is dynamically adjusted. The matching of temperature parameters with the determinism of expression ensures a balance between accuracy and diversity in the translation. The comprehensive adaptation of the sampling strategy to the formality and emotional intensity of the language makes the vocabulary selection and expression style of the translation highly consistent with the original text.

[0105] The core function of the decoder is to generate the sentence structure of the translation. This requires the synergistic effect of temperature parameters and sampling strategies to ensure that the generated sentence structure conforms to the expressive certainty of the original text while also adapting to the formality and emotional intensity of the language. After receiving the temperature parameters, the decoder adjusts the probability distribution of word predictions to achieve regulation. Specifically, at low temperatures, the weight of high-probability words is amplified, while at high temperatures, the probabilities of each word are balanced. Simultaneously, the sampling strategy influences sentence construction by filtering the vocabulary pool and setting selection priorities. The vocabulary pool filtering includes restrictive strategies to filter colloquial words, and the selection priorities include expansion strategies to prioritize words with emotional connotations. For example, for texts with high formality and weak emotional intensity, the sampling strategy may be more favorable at low temperatures. After several iterations, the decoder generates long sentences with complete structure and standard vocabulary. For texts with low formality and strong emotion, under high temperature parameters, it generates diverse and lively sentence structures such as short sentences and exclamatory sentences. During the generation process, the attention mechanism is used to associate with the context to ensure the coherence of the sentence structure. The coordinated control of temperature parameters and sampling strategies enables the sentence structure generated by the decoder to accurately match the original text in terms of certainty, formality, and emotional expression, avoiding translation jargon or style distortion. The controlled sentence structure is a core component of the online text translation, and its quality directly affects the reader's acceptance of the translation. It also provides evaluation samples of the style restoration effect for the feedback optimization module.

[0106] Specifically, dynamically adjusting the beam search parameters includes:

[0107] The rhythm feature vector is decomposed temporally to extract the rhythm intensity, rhythm change rate and rhythm type of each natural segment;

[0108] The bundle width is adjusted based on the rhythm intensity to search the bundle width, and the filtering threshold of the translation subset is dynamically set in combination with the rhythm change rate. The decoder termination condition is also adjusted according to the rhythm type.

[0109] The rhythm feature vector is a comprehensive representation of the overall rhythm of an online novel. Adjusting the bundle search parameters requires targeting the specific rhythm characteristics of each paragraph, including temporal speed, degree of change, and rhythm type. Therefore, it is necessary to obtain paragraph-level rhythm indicators through temporal decomposition. The rhythm feature vector is decomposed into paragraph sub-vectors along paragraph boundaries using a temporal segmentation algorithm. Each paragraph sub-vector corresponds to the rhythm information of a paragraph. Rhythm intensity, rhythm change rate, and rhythm type are extracted from the paragraph sub-vectors. Rhythm intensity includes high intensity in combat paragraphs and low intensity in environmental descriptions; rhythm change rate includes the degree of abrupt change from slow to fast rhythm; and rhythm type includes dialogue-intensive rhythm, plot-driven rhythm, and emotional fluctuation rhythm. During extraction, the classification accuracy of each indicator is calibrated by comparing it with a preset rhythm feature template. For continuous rhythm changes across paragraphs, the trend information is retained as an auxiliary feature. This transforms the global rhythm feature vector into specific paragraph-level rhythm indicators, allowing the adjustment of the bundle search parameters to be precise down to each paragraph, avoiding the coarseness of overall rhythm control.

[0110] The beamwidth controls the number of candidate translations; a wide beamwidth corresponds to a diverse range of candidates, while a narrow beamwidth corresponds to a more concise selection. It needs to be matched to the rhythm intensity. For example, a fast rhythm requires rapid screening, while a slow rhythm allows for comparison of multiple candidates. The screening threshold determines the retention criteria for candidate translations and needs to be adjusted according to the rhythm change rate. For example, when there is a sudden change in rhythm, the threshold needs to be widened to include more translations. The decoder termination condition affects sentence length and needs to be adapted to the rhythm type. For example, a conversational rhythm requires short sentence termination, while a descriptive rhythm can terminate with long sentences. When the rhythm intensity is high, a narrow beamwidth is used to reduce the number of candidates and speed up translation; when the rhythm intensity is low, a wide beamwidth is used to increase candidate diversity. If the rhythm change rate is high, the screening threshold for the subset of translations is lowered to allow the inclusion of more marginal candidate translations, such as those not frequently used. Using expressions that fit the sudden rhythm, if the rhythm change rate is low, the screening threshold is increased to streamline candidates; for rhythm types, short sentence termination conditions are set for dialogue-intensive rhythms, medium-length termination conditions are set for plot-driven rhythms, and emotional fluctuation rhythms are dynamically adjusted according to emotional intensity, for example, slightly longer sentences are allowed to fully express strong emotions; the matching of bundle width and rhythm intensity balances translation efficiency and candidate diversity, and the adjustment of the screening threshold with the rhythm change rate ensures the adaptability of the translation method when the rhythm changes abruptly, and the adaptation of termination conditions and rhythm type makes the length of the translated sentence consistent with the tightness or relaxation of the original rhythm. The adjusted bundle search parameters directly affect the translation method selection and sentence generation process, and their rationality determines the degree of synchronization between the translated rhythm and the original text, affecting the reader's reading experience.

[0111] The feedback optimization module receives and processes user feedback data and web crawling data to generate knowledge vectors. It trains the knowledge vectors based on efficient parameter fine-tuning technology to incrementally update the equivalent expression library and web article knowledge graph.

[0112] Furthermore, generating knowledge vectors includes:

[0113] It receives user feedback data and web crawling data, preprocesses the user feedback data and web crawling data, uses an attention mechanism to generate feedback feature vectors from the user feedback data, and performs semantic alignment between the web crawling data and the online article knowledge graph to generate crawling feature vectors.

[0114] The semantic similarity between the feedback feature vector and the crawled feature vector is calculated to adjust the fusion ratio, and time decay weights are assigned to the feedback feature vector and the crawled feature vector obtained in different time windows to fuse and generate a heterogeneous feature matrix that includes user feedback intent and new network knowledge.

[0115] Generate knowledge vectors from heterogeneous feature matrices using deep neural networks.

[0116] User feedback data and web crawled data are the core basis for system optimization. However, the raw data is disorganized and needs to be preprocessed into structured feature vectors before it can be used for subsequent fusion. User feedback needs to highlight key opinions, and web crawled data needs to be linked to the existing knowledge system to ensure relevance. After receiving user feedback data and web crawled data, preprocessing is performed first. User feedback data includes suggestions for modifying the translation and rating comments, while web crawled data includes the latest online article vocabulary, popular translations, and cultural element analyses. For user feedback data, irrelevant and redundant information, such as emoticons and repetitive expressions, is removed. Then, an attention mechanism is used to focus on the core feedback content, such as... Inaccurate translations of certain cultivation terminology are transformed into feedback feature vectors that focus on the user's true intent, highlighting the key points of user concern. For web crawling data, semantic matching algorithms are used to associate it with cultural element features and classification attributes in the online literature knowledge graph. For example, newly crawled internet memes are aligned with the categories of popular internet slang in the online literature knowledge graph. Information irrelevant to the online literature domain is filtered out before generating crawling feature vectors to ensure they are consistent with the existing knowledge system. Preprocessing removes noisy data, attention mechanisms enhance the core information of user feedback, and semantic alignment ensures the domain relevance of web crawling data, providing a high-quality input foundation for subsequent feature fusion.

[0117] Feedback feature vectors and crawled feature vectors may have overlapping or conflicting information. For example, the old translations provided by users may differ from the new translations crawled. Semantic similarity needs to be calculated to adjust the fusion ratio. At the same time, new data is more timely than old data and needs to be given higher weight to reflect dynamic changes. The semantic similarity between two feature vectors is calculated using cosine similarity. If the similarity is high, a balanced fusion ratio is used; if the similarity is low, the ratio is adjusted according to the importance of information. Time decay weights are assigned to the two feature vectors in different time windows, with weights assigned according to time proximity, giving higher weights to recent data and lower weights to older data. Then, the two are merged into a heterogeneous feature matrix by matrix concatenation, while retaining their respective feature dimensions to distinguish user intent and new network knowledge. Semantic similarity adjustment avoids the forced fusion of irrelevant information, and time decay weights ensure the priority of new information, so that the heterogeneous feature matrix can not only integrate the value of the two types of data, but also reflect timeliness and relevance.

[0118] Heterogeneous feature matrices contain multi-dimensional and multi-source information with complex structures. Deep neural networks are needed to extract deep-level correlation features and transform them into compact and representative knowledge vectors for subsequent updates to the equivalent representation library and online literature knowledge graph. A deep neural network containing convolutional and fully connected layers processes the heterogeneous feature matrix. Convolutional layers extract local features, such as the correspondence between keywords in user feedback and crawled translations, while fully connected layers fuse global features, such as overall feedback trends and the evolution of online translations, ultimately generating knowledge vectors. During the training of the deep neural network, optimization is performed to ensure that the knowledge vectors accurately predict user acceptance of new translations, effectively linking user needs with new online knowledge. The non-linear mapping capability of deep neural networks compresses complex heterogeneous information into concise knowledge vectors, preserving key features while reducing the complexity of subsequent processing, making the update process more efficient. The generated knowledge vectors serve as a bridge for incremental updates to the equivalent representation library and online literature knowledge graph, and their representational ability directly determines the accuracy and applicability of the updated content.

[0119] Specifically, the incremental update equivalent expression library includes:

[0120] Calculate the domain relevance between the knowledge vector and the translation in the equivalent representation library to obtain the domain knowledge vector;

[0121] Based on domain knowledge vector parsing, new translation methods are generated, and a candidate set of translation methods is generated. The priority of the candidate set of translation methods is comprehensively ranked according to accuracy, timeliness and style adaptability in order to select alternative translation methods.

[0122] The selected alternative translations are added to the equivalent expression library, and an initial weight is assigned to each alternative translation. When an alternative translation is selected, the initial weight of the alternative translation is adjusted in real time based on user feedback data of the online article translation text.

[0123] Semantic matching is performed between domain knowledge vectors and online text knowledge graphs to verify the consistency of cultural element features and classification attributes of candidate translations. When there is a discrepancy, incremental updates of the online text knowledge graph are triggered.

[0124] Knowledge vectors contain information unrelated to online text translation, such as user feedback on the interface and crawled content from non-online text domains. It is necessary to filter out domain features related to translations in the equivalent expression library to ensure that updated content focuses on online text translation scenarios. Feature vectors of existing translations in the equivalent expression library are extracted, including cultural element features and style attributes. Based on mutual information calculation, the domain relevance between knowledge vectors and these translation feature vectors is analyzed, and feature vectors with domain relevance below a relevant threshold are filtered out. Feature vectors closely related to online text domain translations are retained to form domain knowledge vectors. Domain knowledge vectors eliminate irrelevant information interference, making subsequent translation parsing and updates more focused on the core needs of online text translation and avoiding waste of system resources.

[0125] Domain knowledge vectors contain new translation information, such as user-suggested optimized translations and popular new translations online. Specific translations need to be extracted from these vectors. However, the quality of these new translations varies, requiring multi-dimensional evaluation to select high-quality translations for inclusion in the equivalent expression library. Sequence decoding algorithms are used to extract specific new translations from the domain knowledge vectors, such as new translations of a cultivation term or popular translations of an internet meme, forming a candidate set of translations. Each translation in this set is evaluated from three dimensions: accuracy, timeliness, and style adaptability. Accuracy indicates whether it accurately conveys the original semantics of the online text data; timeliness indicates whether it conforms to current language habits; and style adaptability indicates whether it matches the style of different online text genres. After weighted scoring, translations are ranked, and those with higher scores are selected as candidate translations. The parsing process transforms abstract vector information into specific translations, and multi-dimensional ranking ensures the high quality of candidate translations, providing reliable content for updating the equivalent expression library.

[0126] After candidate translations are added to the equivalent expression library, their applicability is reflected through weights. Initial weights are set based on screening scores, but user feedback in actual use better reflects their true effectiveness. Therefore, initial weights need to be dynamically adjusted to optimize subsequent translation selection. Candidate translations are categorized by domain and stored in the equivalent expression library. Initial weights are assigned based on the screening scores; for example, translations with higher scores have higher initial weights. When a candidate translation is selected by the decoder and a translation is generated, user feedback on the translation is collected, such as accuracy or awkwardness. Positive feedback increases its weight, while negative feedback decreases it, allowing the weights to reflect the actual usage effect of the translation in real time. The weight range is limited to 0-1 to avoid extreme values. Initial weights ensure that high-quality translations are used first, and the dynamic adjustment mechanism links the weights of translations with actual effects, allowing the equivalent expression library to self-optimize and gradually accumulate high-value translations.

[0127] The alternative translations contain new cultural elements or classifications, such as translations of emerging internet memes. If these translations are inconsistent with the cultural element features and classification attributes in the existing online literature knowledge graph, it will lead to knowledge conflicts in the system. The online literature knowledge graph needs to be verified and updated to maintain consistency. Semantic matching of domain knowledge vectors with the cultural element features and classification attributes in the online literature knowledge graph is performed to check whether the cultural element features and classification attributes associated with the alternative translations are consistent with the online literature knowledge graph. If inconsistencies are found, such as the new internet meme involved in the alternative translations not having a corresponding classification in the online literature knowledge graph, an incremental update of the online literature knowledge graph is triggered. The new cultural element features are added to the online literature knowledge graph and labeled with the correct classification attributes to ensure that the online literature knowledge graph can accurately cover new cultural information. Consistency verification avoids internal conflicts in the system's knowledge. The incremental update of the online literature knowledge graph ensures that it can absorb new cultural elements in a timely manner, providing more comprehensive knowledge support for subsequent translations. The updated online literature knowledge graph will then act on the feature extraction module, enabling it to more accurately identify new cultural elements and generate cultural marker sequences to form a closed loop.

Claims

1. An artificial intelligence-based web text translation system, characterized by, The application relates to a web text translation method and device. The feature extraction module is used for processing web text data through a deep neural network to output a genre vector and a style vector of the web text, combining a preset web text knowledge graph to output a cultural marker sequence, and determining a dialogue density, an emotional fluctuation variance and a plot density in the web text data to generate a rhythm feature vector. The dynamic translation module is used for taking the web text data, the genre vector, the style vector and the rhythm feature vector as conditional signals, encoding through a self-attention encoder, judging a discontinuation opportunity according to the cultural marker sequence in a decoding process, querying an equivalent expression library to select a web text translation method, and dynamically outputting a web text translation text, while controlling a temperature parameter and a sampling strategy of a decoder according to the style vector and dynamically adjusting beam search parameters based on the rhythm feature vector. The feedback optimization module is used for receiving and processing user feedback data and network crawling data to generate a knowledge vector, training the knowledge vector based on a parameter efficient fine-tuning technology to incrementally update the equivalent expression library and the web text knowledge graph. The output of the genre vector and the style vector of the web text comprises:

2. The web content translation system based on artificial intelligence of claim 1, wherein, The web text data is preprocessed, and the preprocessed web text data is feature-encoded through a deep neural network to generate a hidden layer feature sequence. The genre semantic of the web text is focused from the hidden layer feature sequence to output a genre feature matrix, and the language expression mode of the web text is captured from the hidden layer feature sequence to output a style feature matrix. Global pooling operations are respectively performed on the genre feature matrix and the style feature matrix to output the genre vector and the style vector of the web text. The output of the cultural marker sequence comprises:

3. The web content translation system based on artificial intelligence of claim 2, wherein, A preset web text knowledge graph is loaded, and the web text knowledge graph comprises cultural element features and classification attributes of web text fields. The semantic similarity of the hidden layer feature sequence encoded through the deep neural network and the cultural element features in the web text knowledge graph is matched to identify the matched cultural element features in the web text data. The classification attributes of the matched cultural element features in the web text knowledge graph are used for type marking to output the cultural marker sequence. The generation of the rhythm feature vector comprises:

4. The web translation system based on artificial intelligence of claim 3, wherein, The web text data is segmented according to natural paragraphs, the proportion of dialogue sentences in the total number of sentences in each natural paragraph is calculated to form a dialogue density. The emotion of each dialogue sentence is analyzed and quantified as an emotional value, and the variance of the emotional value is calculated through a sliding window to form an emotional fluctuation variance. The plot events in the web text data are identified, the number and interval sentence number of the plot events in each natural paragraph are calculated to form a plot density. The dialogue density, the emotional fluctuation variance and the plot density are processed through a long short-term memory network to output a time sequence feature matrix, the rhythm intensity of each natural paragraph is calculated according to the time sequence feature matrix, and the rhythm intensity is weighted and summed through a self-attention mechanism to generate the rhythm feature vector. The dynamic output of the web text translation text comprises:

5. The web content translation system based on artificial intelligence of claim 4, wherein, The text encoding sequence of the web text data is hierarchically fused with the genre vector, the style vector and the rhythm feature vector to form a joint encoding vector. ​ During the interruption, the classification attribute query equivalent expression library based on the cultural marker sequence is obtained, candidate translations are calculated, and the semantic similarity between the candidate translations and the joint coding vector is calculated to screen out a translation subset; The length and sentence complexity of each translation in the translation subset are matched with the rhythm intensity of the corresponding natural paragraph in the rhythm feature vector to select the web translation according to the matching fitness; After the interruption, the decoder starts translation based on the selected web translation to output the web translation text, and through the self-attention mechanism, the web translation text before and after the interruption is associated to adjust the sentence structure of the web translation text to smooth the web translation text.

6. The web translation system based on artificial intelligence of claim 5, wherein, The interruption timing is determined according to the cultural marker sequence, which includes: According to the cross-cultural difference of the cultural element characteristics in the web knowledge graph, the priority of the cultural marker sequence is divided to form a marker priority; After the self-attention encoder outputs the text coding sequence, the relationship between the decoding position of the self-attention encoder and the position of the cultural marker sequence in the web text data is tracked in real time, and when the self-attention encoder processes the web text data divided into high-priority markers, it is determined as an interruption timing; Based on the interruption timing, according to the complexity of the cultural marker sequence, the number of web translations of the cultural marker sequence in the equivalent expression library is queried to dynamically determine the decoding step of the interruption.

7. The web translation system based on artificial intelligence of claim 6, wherein, The temperature parameter and sampling strategy of the decoder include: Perform feature analysis on the style vector to extract multiple feature components representing the language expression mode of the web, including expression certainty, language formality, and emotional expression intensity; Determine the temperature parameter of the decoder based on the expression certainty, and determine the sampling strategy of the decoder based on the language formality and emotional expression intensity; Input the determined temperature parameter and sampling strategy into the decoder to regulate the sentence structure of the web translation text generated by the decoder.

8. The web translation system based on artificial intelligence of claim 7, wherein, The dynamic adjustment of the beam search parameter includes: Time sequence decomposition is performed on the rhythm feature vector to extract the rhythm intensity, rhythm change rate and rhythm type of each natural paragraph; Based on the rhythm intensity, the beam width of the beam search parameter is adjusted, the screening threshold of the translation subset is dynamically set combined with the rhythm change rate, and the decoder termination condition is adjusted according to the rhythm type.

9. The web translation system based on artificial intelligence of claim 8, wherein, The generation of the knowledge vector includes: Receive user feedback data and web crawling data, and preprocess the user feedback data and web crawling data, generate a feedback feature vector through an attention mechanism, and perform semantic alignment between the web crawling data and the web knowledge graph to generate a crawling feature vector; Calculate the semantic similarity between the feedback feature vector and the crawling feature vector to adjust the fusion ratio, and assign time decay weights to the feedback feature vectors and crawling feature vectors obtained in different time windows to fuse to generate a heterogeneous feature matrix including user feedback intentions and network new knowledge; The heterogeneous feature matrix is generated into a knowledge vector through a deep neural network.

10. The web content translation system based on artificial intelligence of claim 9, wherein, The incremental update of the equivalent expression library includes: Calculate the domain relevance of the knowledge vector and the translation in the equivalent expression library to obtain a domain knowledge vector; Based on the field knowledge vector, new translation methods are analyzed to generate a candidate set of translation methods, and the priority of the candidate set of translation methods is comprehensively sorted according to accuracy, timeliness and style adaptability to screen the alternative translation methods; The screened alternative translation methods are added to the equivalent expression library, and each alternative translation method is assigned an initial weight, and when the alternative translation method is selected, the initial weight of the alternative translation method is adjusted in real time according to the user feedback data of the network text translation text; Through semantic matching of the field knowledge vector and the network knowledge graph, the consistency of the cultural element characteristics and the classification attributes of the alternative translation method is checked, and when the consistency is not consistent, the incremental update of the network knowledge graph is triggered.

Images