An artificial intelligence-based semantic analysis and recognition method

Abstract

This invention discloses an artificial intelligence-based semantic analysis and recognition method, specifically relating to the field of semantic analysis, to address the problem of inaccurate identification and disambiguation of culture-loaded words in existing cross-language text processing, leading to frequent semantic shifts and inconsistent translation results. The method involves decomposing the text structure and parsing cultural metaphors in cross-language comparative corpora to identify culture-loaded words with specific cultural meanings. It further obtains their semantic vector sets in the source and target cultural contexts, calculates the semantic translation distance between them to generate translation paths, and then trains a contextual disambiguation model by combining the translation paths with the context. This enables the model to output the final semantic probability distribution of culture-loaded words in the target culture when the input text is given. Based on this probability distribution, semantic mapping prompts are generated during cross-language semantic conversion, thereby achieving accurate cross-cultural semantic parsing and prompting of culture-loaded words.

Description

Technical Field

[0001] This invention relates to the field of semantic analysis technology, and more specifically, to a semantic analysis and recognition method based on artificial intelligence. Background Technology

[0002] In cross-language information processing and intelligent translation scenarios, as globalization deepens, the differences in text semantics under different cultural backgrounds are gradually emerging. In particular, some words with strong cultural dependence are prone to ambiguity and deviation in the process of cross-cultural communication. These words often carry specific social history, regional customs or group metaphors and are called culturally loaded words.

[0003] In news reporting, text translation, cross-border business contracts, and social media content understanding, relying solely on literal meanings in the translation or semantic analysis of culture-loaded words often leads to misinterpretations and even the risk of miscommunication. For example, the symbolic difference between the Chinese word "dragon" and its English counterpart lies in their different semantic connotations in everyday and specific contexts. Existing cross-linguistic processing methods typically focus on lexical alignment or statistical probabilities, but struggle to accurately capture the polysemy and semantic shifts of culture-loaded words in source and target cultural contexts. This results in problems such as inappropriate semantic conversion, miscommunication, or cultural conflicts in cross-cultural communication scenarios. Especially in applications requiring high accuracy, such as multilingual version comparison of legal documents, bilingual analysis of international business negotiation materials, cross-cultural public opinion analysis, and intelligent customer service response systems, the inability to effectively identify and disambiguate the semantic connotations of culture-loaded words will prevent the provision of stable and reliable cross-cultural semantic conversion results.

[0004] Therefore, there is an urgent need for a method that can combine comparative corpora, contextual information, and cross-cultural semantic mapping relationships to accurately identify culture-loaded words and establish semantic shift paths from the source culture to the target culture. By using a model-based approach to perform contextual disambiguation, it can provide accurate prompts and explanations in the process of cross-language semantic conversion, thereby improving the adaptability and reliability of intelligent translation and semantic analysis in cross-cultural communication. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide a semantic analysis and recognition method based on artificial intelligence to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A semantic analysis and recognition method based on artificial intelligence includes the following steps:

[0008] S1. Perform text structure decomposition and cultural metaphor analysis on cross-linguistic comparative corpora to identify culturally loaded words with specific cultural meanings in the comparative corpora.

[0009] S2. Obtain the first semantic vector set of culture-loaded words in the source cultural context and the second semantic vector set in the target cultural context;

[0010] S3. Calculate the semantic escaping distance between the first semantic vector set and the second semantic vector set, and generate the word escaping path from the source culture to the target culture;

[0011] S4. Based on the word-loaded translation path, the context disambiguation model is trained by combining the context in the cross-linguistic comparative corpus.

[0012] S5. Input the cultural load word context and corresponding escape path of the source language text to be analyzed into the context disambiguation model, and output the final semantic probability distribution of the cultural load word in the target culture.

[0013] S6. Based on the final semantic probability distribution, provide semantic mapping prompts during cross-language semantic conversion.

[0014] In a preferred embodiment, step S1, which involves performing text structure decomposition and cultural metaphor analysis on the cross-linguistic comparative corpus to identify culturally loaded words with specific cultural meanings in the comparative corpus, specifically includes:

[0015] Perform word segmentation and part-of-speech tagging operations on the cross-linguistic comparative corpus, dividing the text into lexical units and recording the part-of-speech categories;

[0016] The system calls up a pre-defined dictionary of culturally loaded words, compares the vocabulary units with the culturally loaded word entries in the dictionary, and marks the vocabulary units that match the entries as culturally loaded words.

[0017] In cross-linguistic comparative corpora, the context window of the sentence in which the culture-loaded words are located is extracted, and the function of the culture-loaded words in the context is analyzed by combining syntactic dependency relations and semantic role labeling. The function includes symbolic nouns, action phrases and fixed semantic expression vocabulary.

[0018] Record contextual fragments of culturally loaded words in cross-linguistic comparative corpus texts.

[0019] In a preferred embodiment, the culturally loaded word dictionary is specifically established as follows:

[0020] Based on the language categories of the comparative corpus, vocabulary and phrases containing cultural metaphors, symbolic usages and fixed expressions are collected from the cultural corpus of the corresponding language categories;

[0021] The cultural corpus is divided into a source cultural corpus and a target cultural corpus, including historical documents, literary works, media reports, and social texts;

[0022] The collected words and phrases were manually screened and semantically annotated, and culturally loaded word entries were created according to language category, part of speech category, and semantic annotation.

[0023] Culture-loaded word entries are stored as structured data to form a culture-loaded word dictionary for culture-loaded word identification.

[0024] In a preferred embodiment, step S2, obtaining the first semantic vector set of culture-loaded words in the source cultural context and the second semantic vector set in the target cultural context, specifically includes:

[0025] Read the context fragments of culture-loaded words and input the fragments into a semantic encoding model trained on the source culture corpus and a semantic encoding model trained on the target culture corpus, respectively.

[0026] Based on the output of the semantic encoding model, the semantic vector representations of the culturally loaded words in the source cultural context are obtained and merged into the first semantic vector set, and the semantic vector representations in the target cultural context are merged into the second semantic vector set;

[0027] During the merging process, the first semantic vector set and the second semantic vector set are respectively associated with the corresponding cultural load words.

[0028] In a preferred embodiment, step S3, calculating the semantic escaping distance between the first semantic vector set and the second semantic vector set to generate the load word escaping path from the source culture to the target culture, specifically includes:

[0029] Map the first semantic vector set and the second semantic vector set to a unified semantic vector space;

[0030] In the unified vector space, a similarity measure is performed on each pair of semantic vector sets of culture-loaded words, and the similarity value is recorded and stored in the escape comparison result table;

[0031] The similarity values ​​recorded in the escape comparison results table are weighted based on the credibility of the corpus source to generate an overall cross-cultural semantic difference index.

[0032] Based on the semantic difference index, output the escaping path of culturally loaded words from the source culture to the target culture, and establish an escaping path table to record the correspondence between the first semantic vector set of culturally loaded words and the escaping path;

[0033] The escape path includes the direction and magnitude of the change from the source culture semantic vector to the target culture semantic vector.

[0034] In a preferred embodiment, step S4, training the context disambiguation model based on the load word's escape path and in conjunction with the context in the cross-linguistic comparative corpus, specifically includes:

[0035] A training sample set for constructing a context disambiguation model is provided. The training samples include context fragments of culture-loaded words in the source cultural corpus, escape paths, and semantic annotations of the loaded words in the target language.

[0036] A dual-channel neural network architecture is established. The first channel uses an attention mechanism to encode the source language context, and the second channel extracts features from the escape path.

[0037] Design a cross-channel feature fusion layer to fuse context encoding features and escape path features to generate comprehensive semantic features;

[0038] The fully connected layer maps the comprehensive semantic features to the semantic space corresponding to the target cultural corpus, and the output layer outputs the probability distribution of the corresponding semantic annotations.

[0039] The model is supervised learning by using the semantic annotations of the loaded words in the target language as supervision signals, and the model parameters are updated iteratively.

[0040] In a preferred embodiment, step S5, which involves inputting the context of the culture-loaded words in the source language text to be analyzed and their corresponding escape paths into the context disambiguation model, and outputting the final semantic probability distribution of the culture-loaded words in the target culture, specifically includes:

[0041] The text to be analyzed is preprocessed to extract culturally loaded words and their corresponding contextual fragments.

[0042] Obtain the semantic vector representation of the cultural load word context fragments of the text to be analyzed in the source cultural context, and load the corresponding escape path in the escape path table based on the semantic vector representation;

[0043] The context fragments of the culture-loaded words in the text to be analyzed and their corresponding escape paths are input into a trained context disambiguation model, which outputs the final semantic probability distribution of the culture-loaded words in the target culture.

[0044] In a preferred embodiment, step S6, which involves providing semantic mapping prompts during cross-language semantic conversion based on the final semantic probability distribution, specifically includes:

[0045] The semantic annotation with the highest proportion in the final semantic probability distribution is detected. When the semantic annotation does not belong to the literal semantic meaning of the cultural load word, it is judged as having semantic shift.

[0046] When a semantic offset is determined, a semantic mapping prompt is generated. The prompt includes the position index of the culturally loaded word in the text, the semantic annotation in the target cultural context, and the corresponding probability value.

[0047] The technical effects and advantages of the semantic analysis and recognition method based on artificial intelligence proposed in this invention are as follows:

[0048] By identifying culture-loaded words in cross-linguistic comparative corpora and performing semantic parsing in conjunction with context, this invention can establish precise semantic translation paths between the source and target cultures, effectively addressing the semantic shift problem caused by existing methods that rely solely on literal meaning or statistical alignment. By constructing a contextual disambiguation model and introducing joint inputs of context vectors and translation paths, this invention can dynamically determine the polysemy of culture-loaded words and output a semantic probability distribution consistent with the target cultural context, significantly improving the accuracy and stability of cross-cultural semantic recognition. Furthermore, the translation path table established in this invention provides a searchable reference for semantic mapping in new texts, enhancing the model's adaptability to unknown corpus environments.

[0049] In practical applications, this invention can be widely used in scenarios such as cross-cultural information retrieval, cross-language intelligent translation, bilingual legal document comparison, and international business document analysis, effectively reducing the risk of semantic misjudgment and information distortion, and enhancing the credibility and reliability of cross-cultural communication. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of a semantic analysis and recognition method based on artificial intelligence according to the present invention. Detailed Implementation

[0051] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0052] Example 1

[0053] Figure 1 This invention provides a semantic analysis and recognition method based on artificial intelligence, which includes the following steps:

[0054] S1. Perform text structure decomposition and cultural metaphor analysis on cross-linguistic comparative corpora to identify culturally loaded words with specific cultural meanings in the comparative corpora.

[0055] S2. Obtain the first semantic vector set of culture-loaded words in the source cultural context and the second semantic vector set in the target cultural context;

[0056] S3. Calculate the semantic escaping distance between the first semantic vector set and the second semantic vector set, and generate the word escaping path from the source culture to the target culture;

[0057] S4. Based on the word-loaded translation path, the context disambiguation model is trained by combining the context in the cross-linguistic comparative corpus.

[0058] S5. Input the cultural load word context and corresponding escape path of the source language text to be analyzed into the context disambiguation model, and output the final semantic probability distribution of the cultural load word in the target culture.

[0059] S6. Based on the final semantic probability distribution, provide semantic mapping prompts during cross-language semantic conversion.

[0060] In step S1, the cross-linguistic comparative corpus is subjected to text structure decomposition and cultural metaphor analysis to identify culturally loaded words with specific cultural meanings in the comparative corpus.

[0061] The cultural corpus is differentiated based on the language categories of the reference texts to ensure a strict distinction between the source and target cultural contexts during processing. The cultural corpus should be representative and comprehensive to identify words and phrases with stable cultural meanings. In constructing the source cultural corpus, multiple text sources are selected, including historical documents, literary works, media reports, and social media texts. Historical documents possess strong cultural sedimentation characteristics, reflecting long-standing cultural metaphors, such as fixed expressions formed by traditional festivals, typical figures, or historical events. Literary works often contain numerous symbolic uses, such as metonymy, allegory, or poetic expression; this type of corpus can reveal deep-seated metaphors commonly used in specific cultures. Media reports reflect common expressions in social reality, such as metaphorical vocabulary in political language and economic news. Social media texts represent the contemporary public context, containing emerging metaphorical expressions and popular online phrases, supplementing aspects not covered by traditional corpora. The target cultural corpus is collected using the same methods as the source cultural corpus, also including four categories: historical documents, literary works, media reports, and social media texts. In practice, the selection criteria for the corpus should be specified. For example, the time span should cover the last ten to a hundred years to ensure that both contemporary and traditional expressions are included; the text size should reach several million words to ensure sufficient coverage; and the text type should cover both formal and informal contexts to ensure that the semantic application scenarios of the loaded words are fully reflected. During the collection process, cultural metaphors, symbolic usages, and fixed expressions should be the core selection targets. For example, metaphorical expressions such as "dragon" and "drinking tea" should be extracted from the Chinese corpus, and corresponding words should be extracted from the English corpus.

[0062] Vocabulary and phrases undergo manual screening and semantic annotation to ensure the accuracy and representativeness of the final culturally loaded entries. During implementation, each vocabulary item is reviewed by language experts, cultural scholars, or personnel with cross-linguistic research experience. The manual screening criteria include three aspects: First, confirming whether the word has a culturally specific meaning. For example, "dragon" in Chinese often symbolizes auspiciousness and imperial power, with a significant cultural difference in meaning compared to other languages, thus meeting the screening criteria. Conversely, if a word only has a literal meaning and has no cross-cultural difference, such as "table" or "chair," it is not a culturally loaded word and should be removed. Second, confirming whether the word has a high usage frequency in the corpus. Low-frequency and unrepresentative expressions should not be included to avoid bias caused by corpus imbalance. Finally, confirming the polysemy and stability of the word in context. If a word has the same meaning in different cultural environments, it does not need to be recorded as a culturally loaded entry. After manual screening, the semantic annotation stage begins. Semantic annotation consists of three dimensions: language category, part-of-speech category, and semantic annotation. Language categories are used to clarify whether an entry belongs to the source culture or the target culture. For example, "dragon" in Chinese corpora is labeled as a source culture entry, while "dragon" in English corpora is labeled as a target culture entry. Part-of-speech categories are used to distinguish between nouns, verbs, phrases, or fixed expressions. For example, "drinking tea" is labeled as a fixed expression, while "dragon" should be labeled as a noun. Semantic annotations require a detailed description of the cultural meaning, including both literal and figurative or symbolic meanings. For example, "dragon" in Chinese is metaphorically annotated as "auspicious, authoritative, and a national symbol," while "drinking tea" is metaphorically annotated as "being summoned by the police." This annotation method creates a clear classification and semantic description for each culturally loaded word entry.

[0063] After semantic annotation is completed, all culture-loaded word entries are stored as structured data, establishing a systematic dictionary of culture-loaded words. During implementation, the field design for data storage is first determined. Each entry should contain at least the following information: the word itself, its language category, part-of-speech category, semantic annotation content, typical contextual examples, and corpus source information. For example, the entry "dragon" could contain the following fields: word = "dragon"; language category = Chinese; part-of-speech category = noun; semantic annotation = "auspicious, authoritative, national symbol"; contextual example = "descendants of the dragon"; corpus source = "historical documents, literary works". To ensure data retrieval and consistency, standardized classification labels are used for semantic annotation, such as "symbolic noun", "metaphorical phrase", and "fixed expression". Contextual examples should retain at least one complete sentence of context to facilitate the extraction of contextual features during model training, and should be stored using a table structure or key-value pair structure.

[0064] For cross-linguistic comparative corpora, word segmentation and part-of-speech tagging are performed to ensure that the text can be broken down into clear lexical units. Cross-linguistic comparative corpora must be accurate mutual translations that include cultural metaphors and symbolic understanding. When processing source and target language texts, word segmentation is performed according to the linguistic rules of each language. For example, in Chinese corpora, continuous sequences of Chinese characters are segmented according to dictionary rules and statistical probabilities, while in English corpora, word segmentation is based directly on spaces and punctuation. After word segmentation, each lexical unit is tagged with a part-of-speech tag, classifying it into basic grammatical categories such as nouns, verbs, adjectives, and adverbs. Part-of-speech tagging must maintain a consistent classification system to ensure alignment between source and target language results. During the tagging process, the part-of-speech information for each lexical unit is explicitly saved and stored in a structured form, such as forming a pairing record of "lexical unit - part-of-speech category".

[0065] After obtaining the word segmentation and part-of-speech tagging results, each lexical unit is compared one by one with a culturally loaded vocabulary dictionary. The culturally loaded vocabulary dictionary contains manually selected and semantically annotated entries, covering common cultural metaphors, symbolic expressions, and fixed phrases in both the source and target cultures. During the comparison process, string matching or lemmatization comparison should be used to ensure that different forms of words correctly correspond to dictionary entries. When a lexical unit is detected as a perfect match with a dictionary entry, it should be marked as a culturally loaded word.

[0066] After tagging culture-loaded words, their contextual information is analyzed. Specifically, the complete sentence containing the culture-loaded word is extracted, and the scope of the context window is defined accordingly. The size of the context window can be set to a fixed length, such as including a certain number of words to the left and right of the target word, or naturally divided by the sentence boundaries. After extracting the context window, the syntactic dependencies within the window are analyzed to clarify the dependency paths between the culture-loaded word and other words, such as subject-verb, verb-object, and attributive-head relationships. By combining syntactic dependency and semantic role analysis, the functional attributes of the culture-loaded word in the sentence can be accurately determined. For example, "dragon" should be identified as a symbolic noun in the sentence "He is a dragon," while in "dragon boat race," it is part of a fixed semantic expression. After analysis, the results should be recorded as structured information of "culture-loaded word - context window - functional type," and a correspondence should be established with the original corpus location. The functional type needs to be limited to a preset range, such as symbolic nouns, action phrases, and fixed semantic expressions, to ensure consistency in classification.

[0067] In step S2, the first semantic vector set of the culture-loaded words in the source cultural context and the second semantic vector set in the target cultural context are obtained.

[0068] The process involves reading contextual fragments of culture-loaded words and then inputting them into two semantic encoding models trained on different cultural corpora. One model is trained on a source cultural corpus, and the other on a target cultural corpus. During training, both models learn the semantic distribution patterns of words within their respective cultural contexts. Therefore, when given the same contextual fragment, they will output semantic representations with cultural characteristics. In practice, it is crucial to ensure a consistent input format; that is, the contextual fragments must undergo the same word segmentation, part-of-speech tagging, and dependency relation annotation processes to guarantee consistent data received by both models.

[0069] After the context fragments are input into two semantic encoding models, the models perform multi-layer semantic parsing of the input text based on the language features and corpus distribution patterns learned during training, and finally output vector representations containing the semantic information of culture-loaded words. In models trained on source cultural corpora, the output semantic vectors can reflect the common meanings and contextual features of culture-loaded words in the source cultural context. To ensure consistency in the processing, after the semantic encoding models output, the contextual semantic vectors containing culture-loaded words need to be uniformly extracted, and the parts directly related to the culture-loaded words are extracted separately as the semantic vector representations of the culture-loaded words. Subsequently, all vectors extracted from the source cultural contexts are merged into the first semantic vector set, and all vectors extracted from the target cultural contexts are merged into the second semantic vector set. The merging operation needs to ensure consistency of order and traceability of source, that is, each vector must be accompanied by a corpus source identifier and a context fragment index. It should be clear that the same culture-loaded word does not necessarily have only one semantic vector representation. The specific semantic vector representation to be selected needs to be determined in conjunction with its context, which is the reason for choosing to input the context of the culture-loaded word into the semantic encoding model.

[0070] In step S3, the semantic escaping distance between the first semantic vector set and the second semantic vector set is calculated to generate the load word escaping path from the source culture to the target culture.

[0071] The first and second semantic vector sets are mapped to a unified semantic vector space to ensure that semantic representations from different cultural contexts can be compared within the same coordinate system. Specifically, this involves establishing consistent dimensional representations and coordinate references for the two vector sets. Through vector normalization, scale alignment, and feature dimension matching, the semantic vectors of the source and target cultures remain comparable in the same space. For example, when the semantic encoding model of the source culture outputs a multi-layered feature representation, dimensionality reduction and projection operations are used to map it to a pre-defined set of semantic dimensions. The semantic vectors of the target culture also undergo the same processing steps to ensure that the mapped vectors have the same length and structure.

[0072] After obtaining two semantic vector sets in a unified semantic vector space, the semantic vectors of the source culture and the target culture are compared one by one, their similarity is calculated, and the results are recorded. In practice, using culture-loaded words as indices, each semantic vector in the source culture context is paired with all semantic vectors in the target culture context, and a similarity measurement is performed between each pair of vectors. The method of similarity measurement must remain consistent throughout the data processing; numerical similarity results can be obtained by measuring the distance or angle between vectors. During execution, it is necessary to ensure that all comparison results have a unique record identifier. All similarity values ​​are stored in a structured escape comparison result table, which stores the number of comparison pairs and their corresponding similarity values ​​in a row-column format, and includes contextual fragment identification information for both the source and target cultures.

[0073] After obtaining the translation comparison results table, the similarity results are further weighted and statistically analyzed to generate an overall cross-cultural semantic difference index. The weighting is based on the credibility level of the corpus source, and this level must be clearly defined in advance. For example, historical documents and authoritative publications are given higher weights, news media and official reports are given medium weights, and social network texts are given lower weights. This method ensures that the statistical results are closer to the semantic patterns of authoritative contexts. In implementation, the culturally loaded word should be used as an index, and all similarity results for that word should be weighted, summed, and normalized to form a comprehensive value, which serves as the semantic difference index of that culturally loaded word in the cross-cultural context. During the statistical process, it is necessary to ensure that every similarity value participates in the calculation, but its contribution ratio is determined by the corpus source.

[0074] A threshold is used to determine whether to generate escape paths based on the obtained cross-cultural semantic difference index. The threshold is defined before implementation; for example, if the difference index is below a certain value, it indicates that the source and target cultures have similar understandings of the culture-loaded word, and an escape path is not needed. If the difference index is above this value, it indicates significant semantic differences, and a corresponding escape path must be generated. The threshold can be set based on statistical patterns from a large corpus, for example, in the range of 0.2 or 0.3, ensuring a balance between sensitivity and stability. When it is confirmed that an escape path needs to be generated, the relationship between the source and target cultural semantic vectors of the culture-loaded word should be clearly recorded. This relationship includes two dimensions: direction and magnitude. Direction describes the direction of semantic migration in a unified semantic vector space, such as from positive to negative symbolism. Magnitude describes the magnitude of semantic shift, i.e., the distance or angle between vectors. All generated escape paths must be stored in an escape path table, indexed by the culture-loaded word, recording the binding relationship between its first semantic vector set and the corresponding escape path.

[0075] In S4, a context disambiguation model is trained based on the load word's escape path and in conjunction with the context in the cross-linguistic comparative corpus.

[0076] A training corpus sample set is constructed to provide complete data support for the subsequent training of the context disambiguation model. The basic unit of the training sample is a culturally loaded word instance. Each instance must contain three types of content: first, the context fragment of the culturally loaded word in the source culture corpus; second, the semantic path of the loaded word from the source culture to the target culture; and third, the semantic annotation in the target language environment. The context fragment is selected with the complete sentence as the smallest unit, extending forward or backward as necessary to ensure accurate semantic positioning of the culturally loaded word. The semantic path is obtained from the semantic path table established in the previous step, which records the correspondence between the semantic vectors of the source culture and the semantic vectors of the target culture, and can be obtained through retrieval. The semantic annotation part is the result of manual or expert annotation, usually using a structured tagging method, such as marking a culturally loaded word as a symbolic meaning, metaphorical meaning, or fixed semantic expression. Each training sample must completely contain the above three types of information and maintain the correspondence.

[0077] After preparing the training sample set, the core network architecture of the context disambiguation model was constructed. This architecture employs a dual-channel design to ensure that contextual information and escape path information can be modeled independently and fused in subsequent stages. The first channel specifically processes source language context fragments, using an attention mechanism to weighted encode the words in the context. The attention mechanism highlights contextual words with a higher correlation to culturally loaded words, such as verbs or adjectives with syntactic dependencies, in order to generate a deep semantic representation of the context. The second channel specifically extracts features from the escape path input. The escape path is described by the difference between the source and target cultural semantic vectors. During the input process, it needs to be structured and parsed to encode the direction and magnitude of change, forming a numerical vector that can be processed by the network. The two channels operate independently. The first channel outputs the context-encoded feature sequence, and the second channel outputs the escape path feature vector. This architecture design ensures that the model can capture both contextual semantics and cross-cultural semantic transfer information.

[0078] After dual-channel feature extraction, a cross-channel feature fusion layer is set in the network structure to achieve joint modeling of context features and escape path features. The design of the fusion layer must ensure sufficient interaction between the two types of features without losing hierarchy. In specific implementation, the fusion layer takes the context feature sequence and escape path feature vector as input, uses an attention weight mechanism to calculate the correlation between them, and assigns different weight values ​​to highlight the parts of the escape path that play a key role in the semantic interpretation of the context. For example, when a context fragment points to a metaphorical usage, the fusion layer automatically increases the proportion of the escape path features corresponding to that metaphor, so that the comprehensive semantic features can more accurately reflect the semantic transfer rules. The fusion result is a set of comprehensive vectors, which structurally contain both the semantic dependencies of the context and the differences in cross-cultural escape paths. Through this fusion method, the model can make full use of multi-dimensional information in the subsequent prediction process, improving the accuracy of disambiguation judgment.

[0079] After generating comprehensive semantic features, a fully connected layer is needed for semantic space mapping, gradually converging the features to the semantic space corresponding to the target cultural corpus. The role of the fully connected layer is to compress and nonlinearly transform the high-dimensional comprehensive semantic features, mapping them step by step to a spatial structure consistent with the target semantic annotation labels. After mapping, the model's output layer generates a set of probability distributions reflecting the multiple possible semantic meanings of the culture-loaded word in the target cultural context and their probabilities. To ensure the model can correctly learn this mapping relationship, a supervision signal is introduced during training. The supervision signal comes from manually annotated target cultural semantic annotations in the training samples, which serve as the standard answer and are compared with the probability distribution output by the model. The model updates parameters iteratively during training. In each iteration, the difference between the predicted distribution and the true annotation is calculated, and the network weights are adjusted using the backpropagation algorithm to narrow the gap between the two. After multiple iterations, the model can gradually and stably learn the cross-cultural semantic correspondence of culture-loaded words in different contexts, ultimately achieving reliable prediction from context and escape paths to the target semantic annotation.

[0080] In step S5, the context of the culturally loaded words in the source language text to be analyzed and their corresponding escape paths are input into the context disambiguation model, and the final semantic probability distribution of the culturally loaded words in the target culture is output, specifically including:

[0081] The text to be analyzed undergoes preprocessing to extract culture-loaded words and their corresponding contextual fragments. The method of extraction is consistent with the steps described above, namely, after extracting the contextual fragments of culture-loaded words through comparison with a culture-loaded word dictionary, these fragments must be input into a semantic encoding model trained on a source cultural corpus to generate semantic vector representations. This encoding process transforms the text fragment into a fixed-length numerical vector, which represents the semantic features of the fragment in a unified semantic space. After encoding, a search is performed in the escape path table. The search is based on the correspondence between the semantic vector and the source cultural semantic vectors stored in the escape path table. By comparing the vectors, the closest entry is automatically located, and the corresponding escape path is loaded. The escape path contains information about the direction and magnitude from the source cultural semantic vector to the target cultural semantic vector. Throughout this process, consistency in the search operation must be ensured; that is, the same escape path is always retrieved for the same culture-loaded word in the same context (or when the semantic vector similarity reaches a preset synonym threshold, such as 90%) to avoid ambiguity in the results.

[0082] The obtained context fragments and corresponding escape paths are fed as joint input into a pre-trained context disambiguation model. The model employs a dual-channel architecture: the first channel encodes the context fragments, and the second channel extracts features from the escape paths. These two channels dynamically interact at the feature fusion layer to generate comprehensive semantic features. Subsequently, the comprehensive semantic features are non-linearly transformed and dimension-reduced by the fully connected layer, finally generating a set of probability values ​​for semantic labels within the target cultural context at the output layer. The output is a probability distribution because culture-loaded words often have multiple possible interpretations in the target culture. The model needs to provide a relative confidence level among all candidate semantic annotations, and the probability distribution is the quantification of this confidence level. The probability distribution is generated by performing a normalization operation on the comprehensive semantic features, ensuring that the sum of the probability values ​​of all candidate labels is 1, intuitively reflecting the relative likelihood of different semantic interpretations. For example, for the culture-loaded word "drink tea," the output might be "drinking tea: 0.35, receiving conversation: 0.60, other: 0.05," where the values ​​represent the probability distribution.

[0083] In step S6, semantic mapping prompts are provided during cross-language semantic conversion based on the final semantic probability distribution.

[0084] The semantic probability distribution output by the context disambiguation model is tested. This distribution is calculated by the model from multiple candidate semantic annotations for culture-loaded words within the target cultural context. Each semantic annotation in the distribution carries a corresponding probability value, and the sum of all probability values ​​is one. During testing, the semantic annotation with the highest probability value should be explicitly used as the primary reference result. To determine whether this primary semantic annotation is consistent with the literal meaning of the culture-loaded word, a literal semantic set in the culture-loaded word dictionary needs to be established beforehand and compared during the testing phase. A semantic shift occurs when the semantic annotation with the highest probability value is not included in the literal semantic set. For example, if "drinking tea" is annotated as "holding a meeting" in the target cultural context, and this result has a higher probability value than other semantic annotations, then since this annotation is inconsistent with the literal meaning of "drinking tea," the culture-loaded word is considered to have undergone a semantic shift.

[0085] After determining semantic shifts, semantic mapping prompts must be generated immediately to provide guidance to users or the system during cross-cultural semantic translation. These prompts should contain three core elements. The first is the location index of the culture-loaded word in the text being analyzed. This index indicates the specific location of the shifted word in the text and is typically recorded using character or word numbers to ensure accurate positioning. The second is the semantic annotation in the target cultural context output by the context disambiguation model. This annotation represents the interpretation of the determined semantic shift and visually demonstrates the potential special semantic meaning of the loaded word in the target culture. The third is the probability value corresponding to this semantic annotation. This value quantifies the relative confidence of the annotation among all candidate interpretations. For example, if the word "drink tea" in the text is detected as "hold a meeting," and the probability value of this annotation is 0.65, the prompt would include the location index of the word "drink tea," the semantic annotation "hold a meeting," and the probability value "0.65." These prompts should be generated in a consistent format for easy use in subsequent text annotation, translation assistance, or other semantic processing stages. In this way, we can ensure that semantic shifts are not only detected, but also output in a clear and structured form, providing actionable data for semantic understanding in cross-cultural scenarios.

[0086] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0087] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0088] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0089] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0090] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0091] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0092] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0093] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0094] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0095] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. 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 semantic analysis and recognition method based on artificial intelligence, characterized in that, Includes the following steps: S1. Perform text structure decomposition and cultural metaphor analysis on cross-linguistic comparative corpora to identify culturally loaded words with specific cultural meanings in the comparative corpora. S2. Obtain the first semantic vector set of culture-loaded words in the source cultural context and the second semantic vector set in the target cultural context; S3. Calculate the cross-cultural semantic difference index between the first semantic vector set and the second semantic vector set, and generate the word-loaded translation path from the source culture to the target culture; S4. Based on the word-loaded translation path, the context disambiguation model is trained by combining the context in the cross-linguistic comparative corpus. S5. Input the cultural load word context and corresponding escape path of the source language text to be analyzed into the context disambiguation model, and output the final semantic probability distribution of the cultural load word in the target culture. S6. Based on the final semantic probability distribution, provide semantic mapping prompts during cross-language semantic conversion; In step S3, calculating the cross-cultural semantic difference index between the first semantic vector set and the second semantic vector set, and generating the word-loaded translation path from the source culture to the target culture, specifically includes: Map the first semantic vector set and the second semantic vector set to a unified semantic vector space; In the unified vector space, a similarity measure is performed on each pair of semantic vector sets of culture-loaded words, and the similarity value is recorded and stored in the escape comparison result table; The similarity values ​​recorded in the escape comparison results table are weighted based on the credibility of the corpus source to generate an overall cross-cultural semantic difference index. Based on the semantic difference index, output the escaping path of culturally loaded words from the source culture to the target culture, and establish an escaping path table to record the correspondence between the first semantic vector set of culturally loaded words and the escaping path; The escape path includes the direction and magnitude of the change from the source cultural semantic vector to the target cultural semantic vector; In S4, the training of the context disambiguation model based on the load word's escape path and in conjunction with the context in the cross-linguistic comparative corpus specifically includes: A training sample set for constructing a context disambiguation model is provided. The training samples include context fragments of culture-loaded words in the source cultural corpus, escape paths, and semantic annotations of the loaded words in the target language. A dual-channel neural network architecture is established. The first channel uses an attention mechanism to encode the source language context, and the second channel extracts features from the escape path. Design a cross-channel feature fusion layer to fuse context encoding features and escape path features to generate comprehensive semantic features; The fully connected layer maps the comprehensive semantic features to the semantic space corresponding to the target cultural corpus, and the output layer outputs the probability distribution of the corresponding semantic annotations. The model is supervised learning by using the semantic annotations of the loaded words in the target language as supervision signals, and the model parameters are updated iteratively.

2. The semantic analysis and recognition method based on artificial intelligence according to claim 1, characterized in that, In step S1, the text structure decomposition and cultural metaphor analysis of the cross-linguistic comparative corpus are performed to identify culturally loaded words with specific cultural meanings in the comparative corpus, including: Perform word segmentation and part-of-speech tagging operations on the cross-linguistic comparative corpus, dividing the text into lexical units and recording the part-of-speech categories; The system calls up a pre-defined dictionary of culturally loaded words, compares the vocabulary units with the culturally loaded word entries in the dictionary, and marks the vocabulary units that match the entries as culturally loaded words. In cross-linguistic comparative corpora, the context window of the sentence in which the culture-loaded words are located is extracted, and the function of the culture-loaded words in the context is analyzed by combining syntactic dependency relations and semantic role labeling. The function includes symbolic nouns, action phrases and fixed semantic expression vocabulary. Record contextual fragments of culturally loaded words in cross-linguistic comparative corpus texts.

3. The semantic analysis and recognition method based on artificial intelligence according to claim 2, characterized in that, The specific method for establishing the culturally loaded word dictionary is as follows: Based on the language categories of the comparative corpus, vocabulary and phrases containing cultural metaphors, symbolic usages and fixed expressions are collected from the cultural corpus of the corresponding language categories; The cultural corpus is divided into a source cultural corpus and a target cultural corpus, including historical documents, literary works, media reports, and social texts; The collected words and phrases were manually screened and semantically annotated, and culturally loaded word entries were created according to language category, part of speech category, and semantic annotation. Culture-loaded word entries are stored as structured data to form a culture-loaded word dictionary for culture-loaded word identification.

4. The semantic analysis and recognition method based on artificial intelligence according to claim 1, characterized in that, In step S2, obtaining the first semantic vector set of culture-loaded words in the source cultural context and the second semantic vector set in the target cultural context specifically includes: Read the context fragments of culture-loaded words and input the fragments into a semantic encoding model trained on the source culture corpus and a semantic encoding model trained on the target culture corpus, respectively. Based on the output of the semantic encoding model, the semantic vector representations of the culturally loaded words in the source cultural context are obtained and merged into the first semantic vector set, and the semantic vector representations in the target cultural context are merged into the second semantic vector set; During the merging process, the first semantic vector set and the second semantic vector set are respectively associated with the corresponding cultural load words.

5. The semantic analysis and recognition method based on artificial intelligence according to claim 1, characterized in that, In step S5, the context of the culturally loaded words in the source language text to be analyzed and their corresponding escape paths are input into the context disambiguation model, and the final semantic probability distribution of the culturally loaded words in the target culture is output, specifically including: The text to be analyzed is preprocessed to extract culturally loaded words and their corresponding contextual fragments. Obtain the semantic vector representation of the cultural load word context fragments of the text to be analyzed in the source cultural context, and load the corresponding escape path in the escape path table based on the semantic vector representation; The context fragments of the culture-loaded words in the text to be analyzed and their corresponding escape paths are input into a trained context disambiguation model, which outputs the final semantic probability distribution of the culture-loaded words in the target culture.

6. The semantic analysis and recognition method based on artificial intelligence according to claim 1, characterized in that, In step S6, the semantic mapping prompts during cross-language semantic conversion based on the final semantic probability distribution specifically include: The semantic annotation with the highest proportion in the final semantic probability distribution is detected. When the semantic annotation does not belong to the literal semantic meaning of the cultural load word, it is judged as having semantic shift. When a semantic offset is determined, a semantic mapping prompt is generated. The prompt includes the position index of the culturally loaded word in the text, the semantic annotation in the target cultural context, and the corresponding probability value.

Images