An AIGC content style checking method for an automobile marketing scene

Abstract

This invention relates to the field of text processing technology and discloses an AIGC content style verification method for automotive marketing scenarios. The method obtains the initial word vector sequence of the automotive marketing text to be verified, identifies domain evaluation words belonging to a pre-built automotive marketing style knowledge base, calculates the difference between the initial word vector and the semantically neutral vector of the identified domain evaluation words, filters out semantic offsets to obtain a semantically neutralized vector, then concatenates this semantically neutralized vector with the style vector to generate a style-decoupled word vector, and replaces the corresponding initial word vectors in the initial word vector sequence with the style-decoupled word vector to obtain a reconstructed word vector sequence. The style verification result is generated based on the reconstructed word vector sequence. This invention eliminates the interference of literal semantics on style determination by directionally separating and reconstructing the semantic and style information of words within the vector space, thus improving the accuracy of style verification for automotive marketing texts.

Description

Technical Field

[0001] This invention relates to the field of text processing technology, and in particular to an AIGC content style verification method for automotive marketing scenarios. Background Technology

[0002] With the widespread application of AIGC technology in the automotive marketing field, style verification of generated text to ensure its conformity with brand tone, vehicle positioning, and marketing scenario requirements has become a crucial step in the content production process. One of the core tasks of style verification is to accurately determine the style attributes of words in the text, such as identifying the true style connotations of words like "violence," "taming," and "lightness" in the context of automotive marketing.

[0003] However, in existing text style analysis methods, the semantic and stylistic information of words are entangled in the vector space, encoded together in the same vector representation, and cannot be effectively separated. This entanglement leads to serious style misjudgment problems in the specific field of automotive marketing. For example, in the performance car marketing text "This car accelerates too violently," the word "violent" simultaneously carries aggressive, negative semantic information and powerful stylistic information in the general vector space. Because existing methods cannot decouple these two types of information from the vector representation, they often incorrectly judge the overall style of the text as negative or inappropriate based on the dominant negative semantic component, thus marking a qualified or even excellent marketing copy as unqualified.

[0004] The root cause of these problems lies in the fact that existing methods treat word vectors as an indivisible whole for feature extraction and style comparison. When style analysis is required, they perform global operations on word vectors, lacking a technique to selectively separate and process semantic and stylistic components within the vector space. This results in style evaluation results being constantly influenced by the literal semantics of words, causing the evaluation benchmark to drift.

[0005] Furthermore, even when attempting to extract style shift components from entangled vectors using difference operations, direct difference operations face reliability issues. When semantics and style are highly entangled, information from different dimensions in the vector space permeates each other, and the shift vector obtained by direct difference may be meaningless noise rather than a valid component that truly reflects style differences. Existing technologies lack mechanisms to verify the validity of such operational results, making it impossible to determine whether the extracted style components are reliable, leading to uncertainty in the decoupling results.

[0006] In summary, in the field of automotive marketing text style analysis, there is an urgent need for a method that can reliably decouple semantic and style information of words in vector space and verify the validity of the decoupling results, in order to solve the problem of style misjudgment caused by information entanglement in existing technologies. Summary of the Invention

[0007] This invention provides an AIGC content style verification method for automotive marketing scenarios. Its main purpose is to solve the problem that in the style analysis of automotive marketing texts, the semantic information and style information of words are entangled in the vector space, and existing methods cannot effectively separate the two, leading to style misjudgment. At the same time, the style shift component obtained by directly performing differential operations on the entangled vectors lacks reliability verification, affecting the accuracy of the decoupling results.

[0008] To achieve the above objectives, this invention provides an AIGC content style verification method for automotive marketing scenarios, comprising: Obtain the car marketing text to be verified, and obtain the initial word vector sequence of the car marketing text to be verified; Identify domain evaluation terms belonging to a pre-built automotive marketing style knowledge base in the automotive marketing text to be verified. The automotive marketing style knowledge base pre-records semantic neutral vectors and style vectors for each domain evaluation term. For each identified domain evaluation word, the initial word vector corresponding to the domain evaluation word in the initial word vector sequence is obtained. The semantic offset is filtered out by calculating the difference between the initial word vector and the semantic neutral vector to form a semantic neutral vector. The semantic neutral vector is then concatenated with the style vector to generate the style decoupled word vector of the domain evaluation word. The corresponding initial word vectors in the initial word vector sequence are replaced with the style-decoupled word vectors to obtain the reconstructed word vector sequence. Based on the reconstructed word vector sequence, the style verification result of the automotive marketing text to be verified is generated.

[0009] Optionally, obtaining the initial word vector sequence of the automotive marketing text to be verified includes: The car marketing text to be verified is segmented into words to obtain a word sequence; By querying a pre-built word vector mapping table, each word in the word sequence is mapped to an initial word vector. All the obtained initial word vectors are arranged in the order of the word sequence to form the initial word vector sequence.

[0010] Optionally, identifying domain evaluation terms belonging to a pre-built automotive marketing style knowledge base in the automotive marketing text to be verified includes: The automotive marketing text to be verified is subjected to part-of-speech tagging to obtain a part-of-speech tagging sequence; Based on the part-of-speech tagging sequence, words belonging to the preset part-of-speech range are extracted from the automotive marketing text to be verified, forming a candidate style word set; Each word in the candidate style word set is matched with the domain evaluation words recorded in the automotive marketing style knowledge base, and the successfully matched words are taken as the identified domain evaluation words.

[0011] Optionally, the automotive marketing style knowledge base pre-records semantic neutral vectors and style vectors for each domain evaluation term, obtained through the following steps: Multiple candidate words are obtained from the automotive marketing corpus, and the general word vector and style tag of each candidate word are obtained; Aggregate the general word vectors of the candidate words with the same style tag, and use the center vector obtained by aggregation as the style vector corresponding to the style tag; For each candidate word, the projected component of the style vector is extracted from its general word vector, and the residual vector obtained after extraction is used as the semantic neutral vector of the candidate word.

[0012] Optionally, for each identified domain evaluation word, obtaining the initial word vector corresponding to that domain evaluation word in the initial word vector sequence includes: Obtain the location information of the domain evaluation terms in the automotive marketing text to be verified; Based on the location information, an initial word vector corresponding to the location information is found from the initial word vector sequence, and the found initial word vector is used as the initial word vector corresponding to the domain evaluation word.

[0013] Optionally, the step of calculating the difference between the initial word vector and the semantically neutral vector, and filtering out the resulting semantic offset to form a semantically neutralized vector, includes: Obtain the semantic neutral vector, and subtract the initial word vector from the semantic neutral vector element by element to obtain the semantic offset vector; Calculate the similarity between the semantic offset vector and a pre-acquired set of benchmark vectors, which consists of the semantic offset vectors of other words that have the same style label as the domain evaluation word but have different semantics; When all the calculated similarities are lower than the preset similarity threshold, the semantic offset vector is determined to be valid, and subsequent filtering steps are performed; otherwise, the value of the semantic neutral vector is adjusted and the process returns to the step of recalculating the semantic offset vector.

[0014] Optionally, concatenating the semantically neutralized vector with the style vector to generate style-decoupled word vectors for the domain evaluation words includes: Obtain the dimension number of the semantically neutralized vector, and sequentially append all the element values ​​of the style vector to the end of the semantically neutralized vector. Use the vector formed after appending as the style decoupling word vector.

[0015] Optionally, replacing the corresponding initial word vectors in the initial word vector sequence with the style-decoupled word vectors to obtain the reconstructed word vector sequence includes: Obtain the position index of the identified domain evaluation word in the initial word vector sequence; Keeping all initial word vectors in the initial word vector sequence unchanged except for the position index, the initial word vector corresponding to the position index is replaced with the style decoupling word vector, and the entire vector sequence formed after the replacement operation is used as the reconstructed word vector sequence.

[0016] Optionally, generating the style verification result of the automotive marketing text to be verified based on the reconstructed word vector sequence includes: Summing up all word vectors in the reconstructed word vector sequence, and dividing the summed vector by the total number of word vectors in the reconstructed word vector sequence, yields the final style feature vector of the automotive marketing text to be verified. Calculate the matching degree between the final style feature vector and the pre-acquired target style baseline vector, and generate the style verification result based on the matching degree.

[0017] Optionally, the steps for obtaining the target style reference vector are as follows: Obtain the target style parameters of the automotive marketing text to be verified, wherein the target style parameters include at least one of the target brand tone, target vehicle type, and target marketing scenario; From a pre-built style benchmark library, retrieve the final style feature vector of the benchmark text that matches the target style parameters, and use the retrieved final style feature vector as the target style benchmark vector.

[0018] The present invention has the following beneficial effects: 1. This invention achieves effective decoupling of semantic and style information by reconstructing word vectors within a vector space. Specifically, it calculates the difference between the initial word vector and the semantically neutral vector, filters out the resulting semantic offset to form a semantically neutral vector, and then concatenates the semantically neutral vector with the style vector. The style-decoupled word vectors generated through this process have had their semantic components neutralized, no longer carrying literal semantic biases that could interfere with style determination, while style information is explicitly encoded in a specific dimension of the vector. Therefore, when evaluating the style of the entire text, the evaluation results are no longer affected by the original literal semantics of the words—for example, the word "violence" will no longer lead to the incorrect determination of a performance car marketing text as having inappropriate style due to its negative semantic meaning. This fundamentally solves the problem of style evaluation benchmark drift caused by information entanglement in existing methods, significantly improving the accuracy of style verification for automotive marketing texts.

[0019] 2. In the calculation of the semantic offset vector, this invention introduces a set of benchmark vectors generated from words of similar style but different semantics. A closed-loop mechanism of verification-feedback-adjustment is established by determining the similarity between the semantic offset vector and the benchmark vectors. When the similarity between the semantic offset vector and the benchmark vector is lower than a preset threshold, it indicates that the current decoupling result is valid, and subsequent filtering steps can continue; otherwise, the value of the semantically neutral vector is adjusted and recalculated. This mechanism effectively avoids the problem that direct difference operations may produce meaningless noise vectors when semantics and style are highly entangled, which may be mistakenly filtered out as style components, ensuring the controllability of the decoupling process and the reliability of the decoupling result.

[0020] 3. This invention obtains target style parameters that include at least one of the following: target brand tone, target vehicle type, and target marketing scenario. It then retrieves the style feature vectors of benchmark texts matching these parameters from a pre-built style benchmark library as the target style benchmark vector. This allows the style verification comparison benchmark to dynamically adapt to the specific requirements of different brands, vehicle types, and marketing scenarios, without needing to build a separate verification system for each style requirement. This enhances the versatility and flexibility of this method in actual automotive marketing operations, meeting diverse and multi-layered style verification needs. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating an AIGC content style verification method for automotive marketing scenarios, provided as an embodiment of the present invention. The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0022] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0023] This application provides an AIGC content style verification method for automotive marketing scenarios. The execution entity of this AIGC content style verification method for automotive marketing scenarios includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application embodiment: a server, a terminal, etc. In other words, the AIGC content style verification method for automotive marketing scenarios can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0024] Reference Figure 1 The diagram shown is a flowchart illustrating an AIGC content style verification method for automotive marketing scenarios according to an embodiment of the present invention. In this embodiment, the AIGC content style verification method for automotive marketing scenarios includes: S1. Obtain the car marketing text to be verified, and obtain the initial word vector sequence of the car marketing text to be verified; In this embodiment of the application, obtaining the initial word vector sequence of the car marketing text to be verified includes: The car marketing text to be verified is segmented into words to obtain a word sequence; By querying a pre-built word vector mapping table, each word in the word sequence is mapped to an initial word vector. All the obtained initial word vectors are arranged in the order of the word sequence to form the initial word vector sequence.

[0025] The automotive marketing text to be verified is a piece of automotive marketing copy generated by the AIGC system that requires style review, such as social media tweets or product descriptions; word segmentation refers to using computational linguistics methods to divide the text into the smallest linguistic units with independent semantics, each unit being called a word; a word sequence is an ordered list formed by arranging these words in the order they appear in the original text; a word vector mapping table is a data structure that is pre-trained and persisted in storage, internally using tokens as keys and fixed-dimensional real vectors as values ​​to form a mapping relationship from the token space to the vector space; initial word vectors refer to the original dense vector representation obtained by looking up a table without any subsequent decoupling or reconstruction operations; the initial word vector sequence is an ordered vector group formed by storing all initial word vectors in the corresponding order of the word sequence.

[0026] First, a text to be validated is obtained through an external interface, and a string object is generated and stored in memory. A specific marketing copy, "This SUV's handling is light and agile," is used as an example to guide the subsequent steps. The system calls HanLP, a Chinese word segmentation toolkit based on Hidden Markov Models and Conditional Random Fields, to perform full segmentation and part-of-speech tagging on the string, resulting in a word sequence. The built-in dictionary of the word segmentation tool has been supplemented with automotive terminology to ensure that professional terms such as "SUV" are not missegmented. The word segmentation output is a word sequence of length five: W=[w1,w2,w3,w4,w5], where w1 is "this," w2 is "SUV," w3 is "handling," w4 is "light and agile," and w5 is "agile." Here, wi represents the i-th word in the word sequence, where i is an integer from 1 to 5.

[0027] Next, the system loads the offline-constructed word vector mapping table. This mapping table is derived from the Skipgram word vector model, which is jointly trained using publicly available automotive vertical media corpora and general encyclopedia corpora, with an output dimension d=300. During model training, the context window width is 5 words, the negative sampling number is 10, the initial learning rate is 0.025 and decreases linearly with the training process, and the word vectors after training convergence are stored in a mapping table file in binary format, where each word character occupies one line of record, containing the word's UTF8 string and 300 consecutive 32-bit single-precision floating-point numbers. During loading, the system uses the mmap mechanism to map the file to the virtual address space and constructs a hash index. The hash function used is FNV1a, and the hash table size is 1.5 times the training vocabulary capacity |V|=1,200,000. Collisions are resolved using chaining, making the average time complexity of each table lookup operation close to O(1).

[0028] For each word wi in the word sequence W, the system performs a lookup operation xi = Lookup(wi, M), where M represents the loaded word vector mapping table, and xi ∈ R300 is a 300-dimensional real vector, called the initial word vector. If a word is not found in the mapping table during the query, the system will generate a temporary vector using random initialization based on a normal distribution N(0, 0.01) to ensure the consistency of the sequence dimensions. Each retrieved vector is stored in a contiguous array in memory, and the element values ​​of all vectors follow the IEEE 754 single-precision floating-point format and are dimensionless.

[0029] Finally, the system pushes all initial word vectors into a two-dimensional array (5, 300) created using the Python scientific computing library NumPy, in word sequence order. This array is denoted as the initial word vector sequence X = [x1, x2, x3, x4, x5]. In this example, the first three elements of the initial word vector x3 corresponding to the word "manipulation" are [0.112, 0.348, -0.561], and the first three elements of x4 corresponding to the word "lightness" are [-0.203, 0.076, 0.414]. These specific values ​​will be used for semantic offset difference operations and validity checks in subsequent step S3, ensuring complete consistency of the data throughout the entire embodiment.

[0030] This step transforms unstructured natural language marketing text into a numerical matrix with a unified dimension that retains contextual semantic information. This eliminates the processing difficulties caused by varying text lengths and diverse expressions, providing a standardized input format for subsequent precise decoupling of semantics and style in the vector space, making word-level style separation possible.

[0031] S2. Identify domain evaluation words in the automotive marketing text to be verified that belong to the pre-built automotive marketing style knowledge base. The automotive marketing style knowledge base pre-records a semantic neutral vector and a style vector for each domain evaluation word. In this embodiment of the application, identifying domain evaluation terms belonging to a pre-built automotive marketing style knowledge base in the automotive marketing text to be verified includes: The automotive marketing text to be verified is subjected to part-of-speech tagging to obtain a part-of-speech tagging sequence; Based on the part-of-speech tagging sequence, words belonging to the preset part-of-speech range are extracted from the automotive marketing text to be verified, forming a candidate style word set; Each word in the candidate style word set is matched with the domain evaluation words recorded in the automotive marketing style knowledge base, and the successfully matched words are taken as the identified domain evaluation words.

[0032] In this embodiment of the application, the automotive marketing style knowledge base pre-records semantic neutral vectors and style vectors for each domain evaluation term, which are obtained through the following steps: Multiple candidate words are obtained from the automotive marketing corpus, and the general word vector and style tag of each candidate word are obtained; Aggregate the general word vectors of the candidate words with the same style tag, and use the center vector obtained by aggregation as the style vector corresponding to the style tag; For each candidate word, the projected component of the style vector is extracted from its general word vector, and the residual vector obtained after extraction is used as the semantic neutral vector of the candidate word.

[0033] The automotive marketing style knowledge base is a domain-specific database built offline and persisted on storage media. It stores several words that significantly contribute to the overall style expression in automotive marketing scenarios. Each word is recorded in the database as a data record containing a lexical symbol, a semantically neutral vector, and a style vector. Domain evaluation words refer to specific words in the text that are included in this knowledge base and therefore require semantic and style decoupling processing of their word vectors. Part-of-speech tagging is the process of assigning a part-of-speech category label to each word in a word sequence. The preset part-of-speech range is a set of part-of-speech tags pre-selected based on linguistic features and application requirements. The candidate style word set is a temporary word set composed of all words belonging to the specified part-of-speech range selected from the text. A style tag is an identifier manually or semi-automatically labeled with the typical style type carried by a word in the automotive marketing context. A general word vector is a conventional word embedding representation trained from a large-scale general corpus, without specific domain-specific style priors. Aggregation refers to the operation of finding the central tendency of multiple vectors belonging to the same style tag in the vector space. The projection component refers to the part of the vector that is orthogonally projected onto the direction of another vector. The residual vector is the part remaining after subtracting the projection component from the original vector.

[0034] This step consists of two parts: a domain word recognition submodule and a knowledge base construction submodule. Both run on the same server node and share the word vector mapping table resources in memory.

[0035] The domain word recognition submodule receives the word sequence W from S1, as well as the part-of-speech tagging sequence synchronously output by HanLP during word segmentation. HanLP uses the PKU level-two tagging set, containing 42 tags. The preset part-of-speech range selected in this embodiment includes six categories: adjectives, nouns, verbs, stative words, distinguishing words, and idioms, corresponding to the tags a, n, v, z, b, and i. The system iterates through each word in W, filtering out conjunctions, prepositions, auxiliary words, and punctuation marks. Words that meet the part-of-speech criteria are added one by one to a temporary list data structure, forming a candidate style word set. Taking the example text "This SUV is light and agile" from S1 as an example, the part-of-speech tagging sequence is [this / r, style / q, SUV / nx, control / v, light / a, agile / a]. After part-of-speech filtering, SUV, control, light, and agile are retained as candidate style words. Here, nx is a custom automotive proper noun tag, which has been incorporated into the noun range for unified processing in the system configuration file.

[0036] The system then uses each term in the candidate set as a query key to match it in the automotive marketing style knowledge base. This knowledge base is organized using a red-black tree structure, where the key is a term string and the value is a structure containing a 300-dimensional semantically neutral vector and several style vectors. The keys in the red-black tree are arranged in UTF-8 lexicographical order, and the average time complexity of the lookup operation is O(logN), where N is the total number of terms in the knowledge base. The matching operation uses strict string equality comparison, distinguishing between uppercase and lowercase letters. After matching, "light" and "agile" in the candidate set are matched and become the identified domain evaluation terms; while "SUV" and "handling" are not included in the knowledge base and are treated as ordinary words, retaining their original vectors without decoupling according to the subsequent S4 rules.

[0037] The knowledge base construction submodule was built offline during the system initialization phase, and the results were loaded as read-only files for use in the online phase. Before construction began, developers collected approximately ten million words of text from review articles, forum posts, and official press releases from three major automotive media outlets over the past five years. Using the word segmentation tool and the same Skipgram model training process from S1, a word vector model for the automotive field was obtained, from which 20,000 words with a frequency exceeding 500 times were extracted as a candidate word set. Simultaneously, several annotators with over three years of experience in automotive media were invited to assign at least one style tag to each candidate word. The style tag system includes six categories: dynamic and exciting, elegant and refined, rugged off-road, intelligent and technological, family-friendly, and business-like. The annotation process employed a cross-validation model, with each set of tags independently assigned by two annotators. Inconsistent entries were arbitrated by a third senior annotator, ultimately resulting in a table showing the correspondence between the general word vector and its style tag for each candidate word.

[0038] For each style tag, the system collects the generic word vectors of all candidate words labeled with that tag. Let the vector set belonging to a certain tag contain k vectors, each a 300-dimensional tuple of real components. The system sequentially calculates the arithmetic mean of each dimension by adding the k values ​​and dividing by k; the result is dimensionless. After performing this operation on all 300 dimensions, a 300-dimensional center vector is obtained, which is the style vector corresponding to that style tag. In this way, each of the six style tags obtains a style vector.

[0039] For a single candidate word, after obtaining its general word vector and the style vector of its corresponding tag, the system performs projection stripping. First, the projection components of the candidate word's general word vector in the style vector direction are calculated. This operation is performed dimension by dimension. The calculation process involves multiplying the component values ​​of corresponding dimensions of the two vectors one by one, then summing all 300-dimensional products to obtain a scalar value. This scalar value is then multiplied by the component value of the style vector in each dimension to obtain the projection vector. Next, the value of each dimension of the candidate word's general word vector is subtracted from the value of the corresponding dimension of the projection vector. The difference forms a 300-dimensional residual vector, which is the semantic neutral vector of the candidate word. Taking the candidate word "light" as an example, its style tag is "elegant and refined," and its general word vector is a tuple of 300 components. The first three components of the neutral vector obtained after the above stripping have example values ​​of 0.087, -0.312, and 0.201. These neutral vector values ​​will be used in step S3 to calculate the semantic offset, ensuring the data continuity of the embodiment. The semantic neutral vectors of all candidate words are generated one by one in this manner and stored in the knowledge base file together with the style vectors.

[0040] This step precisely separates the words that truly express style from a large number of functional words in the text to be verified, significantly compressing the search space for subsequent decoupling operations and avoiding the invalid processing of non-keywords. The semantically neutral vectors and style vectors pre-built in the knowledge base are obtained through collaboration between large-scale domain corpora and manual annotation. Their stripping construction strategy decomposes the general semantics and style tendency of each word into two orthogonal vector representations, providing a high-quality reference basis for the evidence-based removal of semantic biases in step S3, directly alleviating the problem in the background technology of being unable to independently determine style components due to semantic and style entanglement.

[0041] S3. For each identified domain evaluation word, obtain the initial word vector corresponding to the domain evaluation word in the initial word vector sequence, and filter out the semantic offset by calculating the difference between the initial word vector and the semantic neutral vector to form a semantic neutral vector. Then, concatenate the semantic neutral vector with the style vector to generate the style decoupled word vector of the domain evaluation word. In this embodiment of the application, the step of obtaining the initial word vector corresponding to each identified domain evaluation word in the initial word vector sequence includes: Obtain the location information of the domain evaluation terms in the automotive marketing text to be verified; Based on the location information, an initial word vector corresponding to the location information is found from the initial word vector sequence, and the found initial word vector is used as the initial word vector corresponding to the domain evaluation word.

[0042] In this embodiment of the application, the step of calculating the difference between the initial word vector and the semantically neutral vector, and filtering out the resulting semantic offset to form a semantically neutralized vector, includes: Obtain the semantic neutral vector, and subtract the initial word vector from the semantic neutral vector element by element to obtain the semantic offset vector; Calculate the similarity between the semantic offset vector and a pre-acquired set of benchmark vectors, which consists of the semantic offset vectors of other words that have the same style label as the domain evaluation word but have different semantics; When all the calculated similarities are lower than the preset similarity threshold, the semantic offset vector is determined to be valid, and subsequent filtering steps are performed; otherwise, the value of the semantic neutral vector is adjusted and the process returns to the step of recalculating the semantic offset vector.

[0043] In this embodiment of the application, the step of concatenating the semantically neutralized vector with the style vector to generate a style-decoupled word vector for the domain evaluation word includes: Obtain the dimension number of the semantically neutralized vector, and sequentially append all the element values ​​of the style vector to the end of the semantically neutralized vector. Use the vector formed after appending as the style decoupling word vector.

[0044] Location information refers to the index of a word in the word sequence formed after word segmentation of the original text. The semantic offset vector is a difference tuple obtained by subtracting the initial word vector from the semantic neutral vector along each dimension; it carries the degree and direction of a word's deviation from its neutral semantics in the current context. Element-by-element subtraction refers to the operation of subtracting the values ​​of two elements with the same position index from two numerical sequences of equal length. The benchmark vector is a set of pre-calculated and stored semantic offset vectors from other words that share the same style tag as the current domain evaluation word but have different semantic content. Similarity is a scalar value used to measure the degree of consistency in direction between two numerical sequences of equal length; here, cosine similarity is used. The preset similarity threshold is a pre-set and persistent judgment boundary value in the system configuration. The semantic neutralization vector is the numerical sequence obtained after removing the semantically offset parts determined to be valid from the initial word vector; its semantic color is close to a neutral state. Concatenation refers to the operation of joining two independent numerical sequences end-to-end into a longer sequence, maintaining a fixed order. Style-decoupled word vectors are numerical sequences generated through the above-mentioned differential, verification, and concatenation processes. The first half carries a neutralized semantic description, while the second half explicitly carries the style characteristics of the word in the context of automotive marketing.

[0045] This step is executed on the server's central processing unit by the vector decoupling computation module, using single-precision floating-point operations throughout. The computation process is encapsulated within a transactional processing function to ensure the atomicity and consistency of intermediate states. The input data comes from the initial word vector sequence output by S1 and the domain evaluation word recognition results delivered by S2.

[0046] The system first iterates through the list of domain evaluation words identified by S2, locating the position of each word sequentially. Taking the domain evaluation word "lightweight" confirmed in S2 as an example, the system backtracks to the word sequence W generated by S1, where W is arranged as: this, SUV, handling, lightweight, agile. After linear scanning or hash-assisted search, the position number of the word "lightweight" is found to be 4. The position numbers of each word start counting from 1. Based on this position number, the system directly retrieves the 4th 300-dimensional array from the initial word vector sequence X by index. This array is the initial word vector for "lightweight". The storage structure of X is a contiguous block of memory stored in row-major order, with a one-to-one mapping between row indices and word positions. The retrieval operation is completed directly through pointer offset.

[0047] The system then proceeds to the semantic offset calculation and closed-loop verification subprocess. It reads the lightweight semantic neutral vector from the knowledge base file. This vector is an offline product of the S2 knowledge base construction phase, with its first three component values ​​being 0.087, -0.312, and 0.201. Simultaneously, it reads the lightweight initial word vector, whose first three component values, given in S1, are -0.203, 0.076, and 0.414. The system performs element-wise subtraction, subtracting the first-dimensional component value of the initial word vector from the first-dimensional component value of the neutral vector, subtracting the second-dimensional component value from the second-dimensional component value, and so on, up to the 300th dimension. The differences obtained from the subtraction operations are arranged in the same order to form a new 300-dimensional array, which is the semantic offset vector. The calculation demonstration uses the first dimension as an example: -0.203 minus 0.087 equals -0.290; the second dimension: 0.076 minus -0.312 equals 0.388; and the third dimension: 0.414 minus 0.201 equals 0.213. This semantic offset vector fully preserves all the fluctuations of lightness relative to its neutral state in the original context. It is dimensionless, and the value of each component varies between -1 and 1.

[0048] To determine whether the aforementioned semantic offset vectors are signals that effectively carry style differences rather than random noise, the system invokes a validity verification mechanism. The system first loads a set of verification benchmark vectors from the verification benchmark cache. This cache is filled during system initialization. Specifically, for each style tag in the knowledge base, several words matching that tag but with different literal meanings are selected, and their semantic offset vectors are pre-calculated using the element-by-element subtraction method described above and persistently saved. The number of verification benchmark vectors is set to 5, selected based on a semantic similarity of less than 0.3 with the current word, while maintaining a completely identical style tag. For the style tag "lightweight," examples of verification benchmark words under the same style tag "elegant and refined" include "slender," "streamlined," "refined," "silky," and "flowing." The system calculates the cosine similarity between the semantic offset vector of "lightweight" and these 5 verification benchmark vectors one by one. The cosine similarity is calculated by multiplying the corresponding dimension components of the two vectors to obtain the sum of the products, then multiplying the square roots of the sum of the squares of each component of the two vectors to obtain the denominator, and finally taking the quotient. Five similarity scalar values ​​are obtained in one calculation, with values ​​ranging from zero to one.

[0049] The system compares these five similarity values ​​with a preset similarity threshold stored in the configuration file. This threshold is empirically set to 0.25. When all five similarity values ​​are strictly less than 0.25, the current semantic offset vector is considered valid, and it does not show significant unidirectional similarity to the offset patterns of other words in the same style, indicating that the offset originates from the semantic specificity of the word itself rather than general noise from the style channel. In the actual calculations performed by Lightweight, the five similarity values ​​were 0.12, 0.08, 0.19, 0.15, and 0.10, all below 0.25, and were therefore considered valid. If any similarity value is greater than or equal to 0.25, the system will determine that the semantic offset vector contains common noise that should not be present, and the value of the semantic neutral vector needs to be adjusted. The adjustment strategy involves adding a random perturbation of 0.005 to each component value of the semantically neutral vector. The perturbation values ​​follow a normal distribution with a mean of zero and a standard deviation of 0.005. The semantic offset vector is then recalculated using the adjusted neutral vector, and similarity comparisons are performed again until all similarities fall below a threshold or the number of adjustments reaches a preset limit of 20. If the limit is reached and the problem persists, the decoupling of the word is abandoned, and the initial word vector is used directly in subsequent steps without modification.

[0050] After validity verification, the system performs semantic offset filtering, which removes the semantic offset vector from the initial word vector. This operation is also performed dimension by dimension, subtracting the corresponding dimension component of the semantic offset vector from the component of each dimension of the initial word vector. After this step, the first three component values ​​of the resulting semantically neutralized vector are 0.087, -0.312, and 0.201, which exactly match the semantically neutralized vector component values ​​provided by S2. This indicates that the semantic offset part has been completely removed, leaving only the neutralized semantic core.

[0051] The system reads the style vector recorded by Lightness from the S2 knowledge base. This style vector also consists of 300 single-precision floating-point numbers, corresponding to the center vector of the tag "elegant and refined". Next, the dimension of the semantically neutralized vector is obtained, which is 300 in this case. All elements of the style vector are sequentially appended to the end of the semantically neutralized vector. The result of this append operation is an array with a total dimension of 600, where the first 300 elements come from the semantically neutralized vector and the last 300 elements come from the style vector. This 600-dimensional array is the style decoupling word vector for Lightness. Similarly, the other domain evaluation word, Agility, identified by S2, is processed in parallel following the same complete process.

[0052] This step achieves explicit separation and reliable recombination of semantic and stylistic information at the individual word level. By detecting and filtering out the semantically floating parts that cause interference in the initial word vectors, and then extending and integrating the domain-preset standard style components in a clearly encoded form, the word "lightness" no longer disrupts the overall style judgment due to its potential semantic bias, and its stylistic value in automotive marketing texts can be directly measured. After introducing a closed-loop verification mechanism, each semantic shift extraction undergoes corroboration verification with synonyms of the same style, avoiding the risk of noise and style misjudgment that may be introduced by direct difference in the background technology, thus significantly improving the reliability of the decoupled output.

[0053] S4. Replace the corresponding initial word vectors in the initial word vector sequence with the style-decoupled word vectors to obtain the reconstructed word vector sequence; In this embodiment of the application, replacing the corresponding initial word vectors in the initial word vector sequence with the style-decoupled word vectors to obtain the reconstructed word vector sequence includes: Obtain the position index of the identified domain evaluation word in the initial word vector sequence; Keeping all initial word vectors in the initial word vector sequence unchanged except for the position index, the initial word vector corresponding to the position index is replaced with the style decoupling word vector, and the entire vector sequence formed after the replacement operation is used as the reconstructed word vector sequence.

[0054] The position index refers to the storage sequence number of the domain evaluation word in the initial word vector sequence. This sequence number starts from zero and corresponds one-to-one with the position of each word in the sequence. The replacement operation refers to the process of removing the original data at a specified storage location in the sequence and filling that location with newly generated data. The remaining initial word vectors refer to the vectors corresponding to ordinary words in the initial word vector sequence that were not identified as domain evaluation words by S2. These vectors maintain their original values ​​obtained from the word vector mapping table without any changes. The reconstructed word vector sequence is a new set of vectors used to describe the entire text to be verified, formed by replacing some positions with style-decoupled word vectors and retaining the original initial word vectors in the remaining positions. The vectors within it are arranged in the original word order.

[0055] This step is performed by the sequence reconstruction module in the server memory using a combination of in-situ updates and dimension alignment. No additional persistent writes are generated throughout the step; the result handle is passed to the next processing stage only after the data structure transformation is completed in memory.

[0056] The system first obtains the positions of the two domain evaluation words in the word sequence from the recognition results of S2. For the example text "SUV handling is light and agile," S2 has confirmed "light" and "agile" as domain evaluation words. According to the position acquisition logic in S3, the position number of "light" in the word sequence is 4, corresponding to a position index of 3 starting from zero. The position number of "agile" is 5, corresponding to a position index of 4. These two index values ​​are stored in an integer array for subsequent use.

[0057] The system then reads in the initial word vector sequence X generated by S1. X exists in memory as a two-dimensional array with a shape of 5 rows and 300 columns. Each row is an initial word vector, and the row number is the position index. The specific content of X follows the example data given in S1, where the 0th row is the 300-dimensional vector corresponding to this model, the 1st row is the 300-dimensional vector corresponding to SUV, the 2nd row is the 300-dimensional vector corresponding to handling, the 3rd row is the 300-dimensional initial word vector corresponding to lightness, and the 4th row is the 300-dimensional initial word vector corresponding to agility.

[0058] The system then reads in two style-decoupled word vectors generated by S3: a 600-dimensional vector corresponding to "lightness" and a 600-dimensional vector corresponding to "agility." The first 300 components of the "lightness" style-decoupled word vector are the semantically neutral vector component values ​​given in S3, such as 0.087, -0.312, and 0.201, while the last 300 components are the corresponding component values ​​of the "elegant and refined" style vector. The "agility" style-decoupled word vector is constructed similarly, also with 600 dimensions.

[0059] The system begins the replacement operation. Because the style-decoupled word vectors are 600-dimensional, while the unreplaced ordinary word vectors in the initial word vector sequence are still 300-dimensional, the replaced sequence will exhibit inconsistent dimensions. To address this, the system performs dimension unification processing on the entire replaced sequence. Specifically, the system first iterates through the position index array, replacing the row corresponding to index 3 with a lightweight 600-dimensional vector, and the row corresponding to index 4 with a flexible 600-dimensional vector. After the replacement, rows 0 to 2 in memory remain 300-dimensional, while rows 3 and 4 become 600-dimensional.

[0060] The system then performs dimensional expansion on the vectors corresponding to non-domain evaluation words to align them with the dimensionality of the style-decoupled word vectors. The expansion method is zero-padding, where the value 0.0 is filled into the 301st to 600th component positions of each 300-dimensional vector. The basis for this zero-padding operation is that these ordinary words do not participate in semantic and style decoupling; their style information is not explicitly extracted, and a zero value represents their neutral, non-contributing state in the style dimension. After dimensional expansion, the corresponding vector becomes 600-dimensional, with the first 300 dimensions representing the component values ​​of the original word vectors, and the last 300 dimensions all being zero. "SUV" and "handling" are processed similarly.

[0061] The system arranges all vectors in their original word order to construct a new two-dimensional array with 5 rows and 600 columns. This array is the reconstructed word vector sequence. Each row in the reconstructed word vector sequence corresponds to a word in the original text, and the row order is completely consistent with the word order in the original text, ensuring that the structural information of the text is completely preserved in the vector space.

[0062] Taking the third row of the reconstructed sequence as an example, its first three components have values ​​of 0.087, -0.312, and 0.201, while the 301st to 600th components are the values ​​of each dimension of the elegant and refined style vector. The 0th row is the original semantic vector of this style followed by 300 zero values. The entire sequence is stored contiguously in memory in row-major order, with each row occupying 600 single-precision floating-point numbers, for a total of 3000 floating-point numbers.

[0063] This step, while maintaining the word order and overall structure of the text, completes a partial reconstruction of the semantic and stylistic representations of keywords in the vector space. The substitution operation replaces words carrying style information with semantically neutral and explicitly style-decoupled representations, while ordinary words retain their original semantic expression and are zeroed out in the style dimension. This constructs a vectorized text model with hierarchical semantic and stylistic representations throughout the text. This reconstruction process makes the contribution of each word clearer and more controllable when subsequently aggregating style features across the entire text. Semantic and stylistic components no longer permeate or interfere with each other, directly providing a structural solution to the style evaluation benchmark drift problem caused by semantic entanglement in the background techniques.

[0064] S5. Based on the reconstructed word vector sequence, generate the style verification result of the car marketing text to be verified.

[0065] In this embodiment of the application, generating the style verification result of the car marketing text to be verified based on the reconstructed word vector sequence includes: Summing up all word vectors in the reconstructed word vector sequence, and dividing the summed vector by the total number of word vectors in the reconstructed word vector sequence, yields the final style feature vector of the automotive marketing text to be verified. Calculate the matching degree between the final style feature vector and the pre-acquired target style baseline vector, and generate the style verification result based on the matching degree.

[0066] In this embodiment of the application, the steps for obtaining the target style baseline vector are as follows: Obtain the target style parameters of the automotive marketing text to be verified, wherein the target style parameters include at least one of the target brand tone, target vehicle type, and target marketing scenario; From a pre-built style benchmark library, retrieve the final style feature vector of the benchmark text that matches the target style parameters, and use the retrieved final style feature vector as the target style benchmark vector.

[0067] Summing all word vectors refers to the operation of accumulating the vector representation of each word in the reconstructed word vector sequence dimension by dimension, resulting in a single cumulative vector with the same dimension as the individual word vectors. The total number of word vectors refers to the number of words in the word sequence constituting the text to be verified, which is equal to the number of rows in the reconstructed word vector sequence. Division refers to an element-wise scalar division operation, dividing each dimension component value of the cumulative vector by the total number. The final style feature vector is a single vector representation of the overall style tendency of the entire text, obtained through the above summation and averaging calculations. Target style parameters are the set of attributes specifying the desired text style when the user submits the verification task; their values ​​come from a pre-defined standardized style description system. Target brand tone refers to the stable style and temperament type formed by a car brand in long-term market communication. Target vehicle category refers to the functional and morphological sub-category of the vehicle. Target marketing scenario refers to the specific channels and communication contexts for marketing content delivery. The style benchmark library is a pre-built structured database loaded into memory during system initialization, storing the mapping relationship between multiple sets of standard style parameters and their corresponding benchmark feature vectors. Matching degree refers to the directional consistency between two vectors in a multidimensional space; it is a dimensionless scalar value between zero and one. Style verification result is the system's final output judgment on whether the text to be verified conforms to the target style requirements.

[0068] This step is completed by the style verification decision module on the server's central processing unit. The operation process is divided into four sequential stages: average pooling, benchmark retrieval, matching degree calculation, and conclusion generation. Single-precision floating-point operations are used throughout the process.

[0069] During the average pooling phase, the system receives the reconstructed word vector sequence output by S4 as input. This sequence has a shape of 5 rows and 600 columns, with the 5 rows corresponding to the 5 words "light and agile handling of this SUV". The system creates a 600-dimensional all-zero accumulation array, sequentially reading the 600 component values ​​from each row and accumulating them one by one into the corresponding position in the array according to their dimension index. After accumulation, a 600-dimensional summation vector is obtained, where each component value of the vector is the sum of the values ​​of the 5 original word vectors in that dimension.

[0070] The system then obtains the total number of word vectors, a value determined by the length of the word sequence recorded in the S1 segmentation stage, which is 5 in this example. The system performs a scalar division by 5 on each dimension component of the summed vector. The result of this division forms a new 600-dimensional vector, which is the final style feature vector of the text to be verified. The physical meaning of the averaging operation is to normalize the text-level style features to the level of contribution of individual words, eliminating the incomparability caused by different text lengths.

[0071] The first 300 dimensions of the final style feature vector carry the average semantic information of the entire text after semantic neutralization, while the last 300 dimensions carry the explicit style information injected by the domain evaluation words. Taking the words "light" and "agile" as examples, the last 300 dimensions of their style components all originate from the same style tag, "elegant and refined." After averaging with the vectors of the other three words, which are all zero in this dimension, the values ​​of the components in the last 300 dimensions that are aligned with the "elegant and refined" style vector are approximately 40% of the corresponding style vector component values.

[0072] During the benchmark retrieval phase, the system first obtains the target style parameters of the text to be validated. These parameters are passed in as a JSON field attached to an external API call, containing three optional key-value pairs. Taking the example text "SUV handling is light and agile" as an example, the passed parameters are: brand tone = elegant and refined, vehicle type = urban SUV, and marketing scenario = social media tweet. The system concatenates these three parameter values ​​into a composite key string, which is then used for precise searching in the style benchmark library.

[0073] The style benchmark library was built offline during the system initialization phase. The builders selected one hundred marketing texts from each major brand, sub-category, and usage scenario, all manually reviewed and deemed style exemplars. Following the same processing flow as the S1 to S5 front-ends, the final style feature vector for each text was pre-calculated, and the style parameters and vectors corresponding to the text were stored as key-value pairs in a hash index table. The hash function used was the same FNV1a algorithm as the S1 word vector mapping table, and the hash table capacity was twice the total number of preset parameter combinations. During retrieval, the system used the target style parameter combination key as the query condition, located the corresponding entry in the hash table, and retrieved its stored 600-dimensional vector as the target style benchmark vector. This benchmark vector represents the standard position of the text corresponding to that parameter combination in the vector space under ideal conditions.

[0074] In the matching degree calculation stage, the system calculates the cosine similarity between the final style feature vector and the target style reference vector. The calculation process involves first multiplying the corresponding dimensions of the two vectors, summing all 600 products to obtain the numerator, then calculating the square root of the sum of the squares of all components of each vector, multiplying the two square roots to obtain the denominator, and finally dividing the numerator by the denominator. This quotient is the matching degree, ranging from zero to one. The closer the value is to one, the more consistent the directions of the two vectors are, and the closer the text style is to the target requirement. In the example, the last 300 dimensions of the final style feature vector highly match the last 300 dimensions of the target style reference vector, while the first 300 dimensions of the neutral semantic part are also relatively close to the semantics of the reference text. The calculated matching degree example value is 0.82.

[0075] During the conclusion generation phase, the system compares the calculated matching score with a preset matching threshold in the system configuration file. This threshold is set to 0.65 based on the stringency of the business scenario. If the matching score is greater than or equal to 0.65, the system generates a style-compliant verification result and returns it to the caller along with a boolean value of true and the matching score value. If the matching score is less than 0.65, the system generates a style-incompatible verification result and returns a boolean value of false and the matching score value. The example text has a matching score of 0.82, which is greater than 0.65. The system outputs a style-compliant verification result, indicating that the copywriting is consistent with the predetermined goals in terms of brand tone, vehicle type, and marketing scenario.

[0076] This step unifies the overall integrity and scenario adaptability of style verification by condensing the semantic and stylistic information of the entire text into a unified feature vector and retrieving corresponding standard references from a benchmark library based on flexibly configurable target style parameters. The use of average pooling to construct text-level vectors ensures that verification conclusions are no longer constrained by extreme values ​​of individual words, resulting in statistically stable output. The multi-dimensional combination retrieval mechanism of target style parameters allows the same verification system to simultaneously serve the differentiated style requirements of different brands, vehicle models, and marketing channels, effectively solving the problem in the background technology where existing methods, lacking dynamic adaptation capabilities, must repeatedly build independent systems for each style requirement.

[0077] This step is the final stage of the overall methodology. Its inputs are the reconstructed word vector sequence output from S4 and the externally provided target style parameters. After three processing steps—pooling, retrieval, and matching—it outputs the style verification conclusion. Thus, from the original text acquisition in S1, to keyword recognition in S2, to word-level decoupling and reconstruction in S3, to sequence replacement in S4, and finally to full-text evaluation in S5, these five steps form a complete data processing pipeline from text to judgment. Each step is tightly connected through intermediate data structures such as vector sequences and positional indices, collectively solving the technical challenge of misjudging AIGC text style due to semantic and style entanglement in automotive marketing scenarios.

[0078] In the several embodiments provided by this invention, it should be understood that the disclosed method can be implemented in other ways.

[0079] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0080] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, and technology that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0081] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for validating the style of AIGC content in automotive marketing scenarios, characterized in that, The method includes: Obtain the car marketing text to be verified, and obtain the initial word vector sequence of the car marketing text to be verified; Identify domain evaluation terms belonging to a pre-built automotive marketing style knowledge base in the automotive marketing text to be verified. The automotive marketing style knowledge base pre-records semantic neutral vectors and style vectors for each domain evaluation term. For each identified domain evaluation word, the initial word vector corresponding to the domain evaluation word in the initial word vector sequence is obtained. The semantic offset is filtered out by calculating the difference between the initial word vector and the semantic neutral vector to form a semantic neutral vector. The semantic neutral vector is then concatenated with the style vector to generate the style decoupled word vector of the domain evaluation word. The corresponding initial word vectors in the initial word vector sequence are replaced with the style-decoupled word vectors to obtain the reconstructed word vector sequence. Based on the reconstructed word vector sequence, the style verification result of the automotive marketing text to be verified is generated.

2. The AIGC content style verification method for automotive marketing scenarios as described in claim 1, characterized in that, The process of obtaining the initial word vector sequence of the automotive marketing text to be verified includes: The car marketing text to be verified is segmented into words to obtain a word sequence; By querying a pre-built word vector mapping table, each word in the word sequence is mapped to an initial word vector. All the obtained initial word vectors are arranged in the order of the word sequence to form the initial word vector sequence.

3. The AIGC content style verification method for automotive marketing scenarios as described in claim 1, characterized in that, The process of identifying domain evaluation terms belonging to a pre-built automotive marketing style knowledge base in the automotive marketing text to be verified includes: The automotive marketing text to be verified is subjected to part-of-speech tagging to obtain a part-of-speech tagging sequence; Based on the part-of-speech tagging sequence, words belonging to the preset part-of-speech range are extracted from the automotive marketing text to be verified, forming a candidate style word set; Each word in the candidate style word set is matched with the domain evaluation words recorded in the automotive marketing style knowledge base, and the successfully matched words are taken as the identified domain evaluation words.

4. The AIGC content style verification method for automotive marketing scenarios as described in claim 1, characterized in that, The automotive marketing style knowledge base pre-records semantic neutral vectors and style vectors for each domain evaluation term, which are obtained through the following steps: Multiple candidate words are obtained from the automotive marketing corpus, and the general word vector and style tag of each candidate word are obtained; Aggregate the general word vectors of the candidate words with the same style tag, and use the center vector obtained by aggregation as the style vector corresponding to the style tag; For each candidate word, the projected component of the style vector is extracted from its general word vector, and the residual vector obtained after extraction is used as the semantic neutral vector of the candidate word.

5. The AIGC content style verification method for automotive marketing scenarios as described in claim 1, characterized in that, For each identified domain evaluation word, obtaining the initial word vector corresponding to that domain evaluation word in the initial word vector sequence includes: Obtain the location information of the domain evaluation terms in the automotive marketing text to be verified; Based on the location information, an initial word vector corresponding to the location information is found from the initial word vector sequence, and the found initial word vector is used as the initial word vector corresponding to the domain evaluation word.

6. The AIGC content style verification method for automotive marketing scenarios as described in claim 1, characterized in that, The step of calculating the difference between the initial word vector and the semantically neutral vector, and filtering out the resulting semantic offset to form a semantically neutral vector, includes: Obtain the semantic neutral vector, and subtract the initial word vector from the semantic neutral vector element by element to obtain the semantic offset vector; Calculate the similarity between the semantic offset vector and a pre-acquired set of benchmark vectors, which consists of the semantic offset vectors of other words that have the same style label as the domain evaluation word but have different semantics; When all the calculated similarities are lower than the preset similarity threshold, the semantic offset vector is determined to be valid, and subsequent filtering steps are performed; otherwise, the value of the semantic neutral vector is adjusted and the process returns to the step of recalculating the semantic offset vector.

7. The AIGC content style verification method for automotive marketing scenarios as described in claim 1, characterized in that, The step of concatenating the semantically neutralized vector with the style vector to generate style-decoupled word vectors for the domain evaluation words includes: Obtain the dimension number of the semantically neutralized vector, and sequentially append all the element values ​​of the style vector to the end of the semantically neutralized vector. Use the vector formed after appending as the style decoupling word vector.

8. The AIGC content style verification method for automotive marketing scenarios as described in claim 1, characterized in that, The step of replacing the corresponding initial word vectors in the initial word vector sequence with the style-decoupled word vectors to obtain the reconstructed word vector sequence includes: Obtain the position index of the identified domain evaluation word in the initial word vector sequence; Keeping all initial word vectors in the initial word vector sequence unchanged except for the position index, the initial word vector corresponding to the position index is replaced with the style decoupling word vector, and the entire vector sequence formed after the replacement operation is used as the reconstructed word vector sequence.

9. A method for verifying the style of AIGC content in an automotive marketing scenario as described in any one of claims 1 to 8, characterized in that, The step of generating the style verification result of the automotive marketing text to be verified based on the reconstructed word vector sequence includes: Summing up all word vectors in the reconstructed word vector sequence, and dividing the summed vector by the total number of word vectors in the reconstructed word vector sequence, yields the final style feature vector of the automotive marketing text to be verified. Calculate the matching degree between the final style feature vector and the pre-acquired target style baseline vector, and generate the style verification result based on the matching degree.

10. The AIGC content style verification method for automotive marketing scenarios as described in claim 9, characterized in that, The steps for obtaining the target style baseline vector are as follows: Obtain the target style parameters of the automotive marketing text to be verified, wherein the target style parameters include at least one of the target brand tone, target vehicle type, and target marketing scenario; From a pre-built style benchmark library, retrieve the final style feature vector of the benchmark text that matches the target style parameters, and use the retrieved final style feature vector as the target style benchmark vector.

Images