Machine learning-based financial marketing copy generation and auditing system

Abstract

The application discloses a machine learning-based financial marketing script automatic generation and auditing system, comprising: a data processing module for acquiring and preprocessing financial marketing task data; an input construction module for constructing script input representation based on a standardized task data set; a script generation module for generating a candidate marketing script set through a script generation model; an auditing feature module for semantic analysis and risk feature extraction processing; an auditing judgment module for semantic matching and risk judgment processing; a script constraint module for script constraint and content correction processing; and a result storage module for correlating and packaging a target marketing script set and a structured auditing result set and writing into an auditing result database. The application is based on machine learning and semantic constraint modeling, realizes integrated processing of financial marketing script automatic generation, auditing and correction, and has the advantages of strong compliance, high accuracy and high automation degree.

Description

Technical Field

[0001] This invention relates to the field of intelligent text processing technology, and in particular to a machine learning-based automatic generation and review system for financial marketing copy. Background Technology

[0002] In financial marketing, the generation and review of marketing copy have long relied on human experience or simple automated tools. Current technology commonly involves directly outputting marketing copy through template filling, rule concatenation, or general text generation models, followed by compliance checks by humans or systems based on keywords and rule bases. While these solutions reduce labor costs to some extent, the generation and review stages are usually independent, lacking unified semantic expression and processing logic, making it difficult to form a cohesive automated process.

[0003] Meanwhile, existing technologies still have significant limitations in financial scenarios. The copywriting generation stage often focuses on fluency and marketing expression, with insufficient modeling of constraints related to financial product attributes, risk factors, and disclosure structures, easily leading to semantic biases or compliance risks. The review stage largely remains at the level of rule matching or text comparison, lacking the comprehensive ability to judge risk semantics, expression structure, and overall consistency. This makes it difficult to accurately locate and controllably correct risk content, resulting in review results not directly impacting copywriting optimization. Overall, the level of intelligence and compliance reliability needs improvement.

[0004] Therefore, how to provide a machine learning-based automatic generation and review system for financial marketing copy is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an automatic generation and review system for financial marketing copy based on machine learning. This invention, based on machine learning and semantic constraint modeling, realizes the integrated processing of automatic generation, review and correction of financial marketing copy, and has the advantages of strong compliance, high accuracy and high degree of automation.

[0006] The machine learning-based automatic generation and review system for financial marketing copy according to an embodiment of the present invention includes:

[0007] The data processing module is used to acquire financial marketing task data and preprocess it to obtain a standardized task dataset;

[0008] The input building module is used to construct and fuse product feature vectors, audience feature vectors, and marketing goal vectors based on a standardized task dataset to obtain the copy input representation;

[0009] The copy generation module is used to generate a set of candidate marketing copy by copying, filtering and probabilistically fusing the input copy representation through a copy generation model.

[0010] The review feature module is used to perform semantic parsing and risk feature extraction on the candidate marketing copy set to obtain the copy review feature set.

[0011] The review and judgment module is used to perform semantic matching and risk assessment on the document review feature set and output a structured review result set.

[0012] The copywriting constraint module is used to perform copywriting constraints and content correction on the candidate marketing copy set based on the structured review result set, and generate the target marketing copy set;

[0013] The results storage module is used to associate and encapsulate the target marketing copy set and the structured review result set, and write them to the review result database.

[0014] Optionally, modules can be integrated using the following methods:

[0015] Acquire financial marketing task data and preprocess it to obtain a standardized task dataset;

[0016] Based on a standardized task dataset, product feature vectors, audience feature vectors, and marketing goal vectors are constructed and integrated to obtain a copywriting input representation.

[0017] The text input representation is fed into the text generation model, and a set of candidate marketing copy is generated through semantic constraint copy filtering and back-modulation probability fusion.

[0018] Semantic analysis and risk feature extraction were performed on the candidate marketing copy set to obtain a copy review feature set for financial scenarios;

[0019] Perform semantic matching and risk assessment on the copy review feature set, and output a structured review result set that corresponds one-to-one with the candidate marketing copy set;

[0020] Based on the structured review result set, the copywriting is constrained and the content is revised to generate a target marketing copy set;

[0021] The target marketing copy set and the corresponding structured audit result set are associated, encapsulated, and written into the audit result database.

[0022] Optionally, the financial marketing task data includes financial product data, marketing activity data, and target audience data, and the preprocessing includes data format standardization, missing data processing, abnormal data filtering, data encoding, and feature normalization processing.

[0023] Optionally, obtaining the text input representation includes the following specific steps:

[0024] Financial product data, target audience data, and marketing campaign data are extracted from the standardized task dataset and organized according to the field structure requirements to obtain product feature field sequence, audience feature field sequence, and marketing target field sequence.

[0025] Data encoding is performed on each field in the product feature field sequence, audience feature field sequence, and marketing objective field sequence, and the encoding results are normalized to obtain the corresponding normalized feature sequence.

[0026] Based on the field weights corresponding to the fields in the normalized feature sequence of each field, the normalized features of each field are weighted and fused to generate product feature vector, audience feature vector and marketing target vector.

[0027] The product feature vector, audience feature vector, and marketing objective vector are concatenated, aligned in dimensions, and mapped to a unified feature space to obtain the copy input representation.

[0028] Optionally, the generation of the candidate marketing copy set includes the following specific steps:

[0029] The text input representation is converted into a set of input sequence representations, and then input into the encoder of the text generation model to perform sequence feature encoding processing, so as to obtain the encoder output sequence representation and encoder state;

[0030] The copy generation model is constructed using a pointer generation network structure, and improvements are made to the pointer generation network structure. The improvements include: introducing a pointer selection substructure based on financial semantic constraints into the pointer mechanism, performing constraint screening on the position of the input sequence before calculating the replication probability, and introducing a review feature reverse modulation substructure into the generation probability control unit, performing reverse modulation on the fusion weight in the fusion stage of generation probability and replication probability.

[0031] Based on the current decoding state of the decoder and the output sequence representation of the encoder, an attention mechanism is used to calculate the attention weights corresponding to each input sequence position, thus obtaining the attention weight sequence.

[0032] The encoder output sequence representation, attention weight sequence, and input pointer selection substructures of financial product semantic labels, risk semantic types, and text structure semantic roles obtained by parsing the text input representation are used to constrain and filter the input sequence positions to obtain a set of input sequence positions that can participate in replication. Then, the attention weight sequence is constrained and normalized based on the set of input sequence positions that can participate in replication to obtain a constrained attention weight sequence.

[0033] In the pointer mechanism, the replication probability distribution is calculated based on the constrained attention weight sequence, and the encoder output sequence representation is weighted and converged based on the constrained attention weight sequence to obtain the context representation;

[0034] In the decoder, the generation probability distribution is calculated based on the context representation, encoder state, and current decoding state of the decoder. The generation probability control unit calculates the relative weight relationship between the generation probability distribution and the replication probability distribution and outputs the fusion weight.

[0035] Semantic constraint features are obtained from the text input representation parsing, and the semantic constraint features are input into the review features to obtain the modulation coefficients by the reverse modulation substructure. The modulation coefficients are then used to reverse modulate the fusion weights to obtain the modulation fusion weights.

[0036] The generation probability distribution and the replication probability distribution are fused and calculated based on the modulation fusion weight to obtain the output probability distribution. Then, stepwise decoding is performed based on the output probability distribution to generate a set of candidate marketing copy. The output probability of the candidate terms in the output probability distribution is obtained by weighting the generation probability and replication probability of the candidate terms according to the modulation fusion weight.

[0037] Optionally, obtaining the text review feature set includes the following specific steps:

[0038] Each candidate marketing copy in the candidate marketing copy set is subjected to text normalization processing to obtain a standardized candidate copy sequence;

[0039] The standardized candidate copy sequence is segmented to obtain a copy fragment set consisting of multiple copy fragments. Semantic features are extracted from each copy fragment in the copy fragment set and mapped to a unified semantic space to obtain a copy fragment-level semantic representation.

[0040] The global semantic features of the standardized candidate copy sequence are extracted and mapped to a unified semantic space to obtain the overall semantic representation of the copy. The copy fragment-level semantic representation is indexed and associated with the overall semantic representation of the copy to obtain the semantic representation set of candidate marketing copy.

[0041] Based on the semantic representation set, risk element extraction is performed on the risk-related semantic content in the candidate marketing copy to obtain a risk element set. Each risk element in the risk element set is then field-coded according to risk category, risk object, risk triggering condition and risk disclosure method to generate a risk element feature vector set.

[0042] The risk element feature vector set is combined and modeled according to the co-occurrence relationship of risk categories, the correlation relationship of risk triggering conditions, and the structural relationship of risk disclosure methods to obtain the risk pattern feature vector that represents the risk expression structure of candidate marketing copy.

[0043] By binding the semantic representation of copy fragments, the semantic representation of copy as a whole, the set of risk element feature vectors, and the feature vector of risk patterns to the same candidate marketing copy identifier, and by structuring and encapsulating the various features after binding according to feature type, a copy review feature set for financial scenarios is obtained.

[0044] Optionally, the output of the structured audit result set includes the following specific steps:

[0045] Based on the candidate marketing copy identifiers, the semantic representations of copy fragments, the semantic representations of copy as a whole, the risk element feature vector set, and the risk pattern feature vector set in the copy review feature set are called to obtain the feature items to be reviewed that correspond one-to-one with the candidate marketing copy set.

[0046] Semantic matching calculations are performed on the text fragment-level semantic representation and the text overall-level semantic representation of each feature item to be reviewed, respectively, to obtain the fragment-level semantic matching score sequence and the overall-level semantic matching score, and the semantic tag field is determined based on the semantic matching score;

[0047] Based on the risk element feature vector set, perform risk element consistency judgment processing on the feature items to be reviewed, generate risk element judgment results, and generate risk identification fields based on the risk element judgment results;

[0048] Based on the risk pattern feature vector, risk pattern determination processing is performed on the feature items to be reviewed, risk pattern determination results are generated, and constraint type fields are determined based on the risk pattern determination results.

[0049] Based on the index association relationship between the semantic representation of text fragments and the semantic representation of text as a whole, the text fragments associated with the risk identification field and the constraint type field are located, and the corresponding text fragment index information and text sequence position range information are extracted to generate the constraint position field.

[0050] The segment-level semantic matching score sequence, the overall-level semantic matching score, the risk element judgment result and the risk pattern judgment result are mapped to the judgment quantity, and the score fusion calculation is performed under a unified judgment scale to obtain the fusion score;

[0051] The fusion score is normalized and mapped to obtain the audit confidence field. The output is a structured audit result set including the risk identifier field, semantic label field, constraint type field, constraint location field, and audit confidence field.

[0052] Optionally, the generation of the target marketing copy set includes the following specific steps:

[0053] The candidate marketing copy set and the corresponding structured review result set are called by field according to the candidate marketing copy identifier, and constraint processing entries corresponding to each candidate marketing copy are constructed.

[0054] Based on the constraint processing items, compliance constraint verification processing is performed on the candidate marketing copy, the compliance constraint items corresponding to the constraint type field in the candidate marketing copy are identified, and the scope of risk semantic content that needs to be constrained is determined by combining the risk identifier field and the semantic tag field.

[0055] Based on the constraint location field, risk fragment location processing is performed on the risk semantic content that needs to be constrained, and the corresponding text fragment index information and text sequence location range information are extracted to obtain the set of constrained text fragments;

[0056] Based on the constraint type field, risk identifier field, and audit confidence field, the constraint triggering strategy matching process is performed on the set of constrained text fragments to determine the constraint triggering strategy identifier corresponding to each constrained text fragment.

[0057] Based on the constraint trigger strategy identifier, the content of each constrained copy fragment is modified to generate a modified copy fragment. The modified copy fragment is then backfilled into the copy sequence position range information corresponding to the original candidate marketing copy according to the copy fragment index information to obtain the target marketing copy. Finally, all target marketing copy are collected to generate a target marketing copy set.

[0058] The beneficial effects of this invention are:

[0059] This invention integrates the generation, review, and revision processes of financial marketing copy into a single model, achieving a closed-loop process from task data input to compliant copy output. Compared to existing technologies where generation and review are separate processes, this invention introduces information such as financial semantic constraints, risk semantic types, and copy structure roles at the copy generation stage. This allows the generation model to balance marketing expression with financial compliance requirements when outputting candidate marketing copy, reducing the probability of risky content from the outset. Simultaneously, a unified semantic representation and feature organization method provides a consistent data foundation for subsequent review and constraint processing, improving the overall process's coherence and stability.

[0060] During the review phase, this invention constructs a multi-layered semantic matching and risk assessment mechanism based on fragment-level and overall-level semantic representations, risk element characteristics, and risk pattern characteristics. This mechanism not only identifies the presence of risks in the text but also performs structured analysis of risk expression from dimensions such as risk category, triggering conditions, and disclosure structure. This approach overcomes the limitations of traditional keyword-based or simple rule-based matching, resulting in higher semantic accuracy and business interpretability in the review results. It can also precisely locate specific text fragments that trigger risks or constraints, providing a clear basis for subsequent revisions.

[0061] Furthermore, this invention achieves effective feedback of the review results to the generated copy by directly applying the structured review results to the copy constraint and content modification process. By matching the constraint type, risk indicator, and confidence level with the corresponding constraint trigger strategy, targeted modifications are made to the constrained copy fragments, thereby improving the compliance and consistency of the copy while preserving the original marketing intent and language style. This mechanism avoids the problem of requiring extensive manual modifications after review in traditional technologies, significantly improving the intelligence, processing efficiency, and practical usability of automatic generation and review of financial marketing copy, and has good engineering application value and promotion prospects. Attached Figure Description

[0062] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0063] Figure 1 This is a flowchart of the method for automatically generating and reviewing financial marketing copy based on machine learning proposed in this invention.

[0064] Figure 2 This is a flowchart of the copy generation model processing of the machine learning-based automatic generation and review system for financial marketing copy proposed in this invention.

[0065] Figure 3 This is a flowchart of the text review feature processing of the machine learning-based automatic generation and review system for financial marketing copy proposed in this invention. Detailed Implementation

[0066] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0067] refer to Figure 1-3 A machine learning-based system for automatically generating and reviewing financial marketing copy, including:

[0068] The data processing module is used to acquire financial marketing task data and preprocess it to obtain a standardized task dataset;

[0069] The input building module is used to construct and fuse product feature vectors, audience feature vectors, and marketing goal vectors based on a standardized task dataset to obtain the copy input representation;

[0070] The copy generation module is used to generate a set of candidate marketing copy by copying, filtering and probabilistically fusing the input copy representation through a copy generation model.

[0071] The review feature module is used to perform semantic parsing and risk feature extraction on the candidate marketing copy set to obtain the copy review feature set.

[0072] The review and judgment module is used to perform semantic matching and risk assessment on the document review feature set and output a structured review result set.

[0073] The copywriting constraint module is used to perform copywriting constraints and content correction on the candidate marketing copy set based on the structured review result set, and generate the target marketing copy set;

[0074] The results storage module is used to associate and encapsulate the target marketing copy set and the structured review result set, and write them to the review result database.

[0075] In this embodiment, the modules are interconnected using the following method:

[0076] Acquire financial marketing task data and preprocess it to obtain a standardized task dataset;

[0077] Based on a standardized task dataset, product feature vectors, audience feature vectors, and marketing goal vectors are constructed and integrated to obtain a copywriting input representation.

[0078] The text input representation is fed into the text generation model, and a set of candidate marketing copy is generated through semantic constraint copy filtering and back-modulation probability fusion.

[0079] Semantic analysis and risk feature extraction were performed on the candidate marketing copy set to obtain a copy review feature set for financial scenarios;

[0080] Perform semantic matching and risk assessment on the copy review feature set, and output a structured review result set that corresponds one-to-one with the candidate marketing copy set;

[0081] Based on the structured review result set, the copywriting is constrained and the content is revised to generate a target marketing copy set;

[0082] The target marketing copy set and the corresponding structured audit result set are associated, encapsulated, and written into the audit result database.

[0083] In this embodiment, the financial marketing task data includes financial product data, marketing activity data, and target audience data. The financial product data includes product type information, yield or fee information, product term or rule information, and product risk level information. The marketing activity data includes marketing activity type information, activity time information, activity channel information, activity theme information, and activity rule information. The target audience data includes audience basic attribute information, audience behavioral characteristic information, audience preference characteristic information, and audience risk tolerance information. The preprocessing includes data format unification, missing data processing, abnormal data filtering, data encoding, and feature normalization processing.

[0084] In this embodiment, obtaining the text input representation includes the following specific steps:

[0085] Financial product data, target audience data, and marketing campaign data are extracted from the standardized task dataset and organized according to the field structure requirements to obtain product feature field sequence, audience feature field sequence, and marketing target field sequence.

[0086] Data encoding is performed on each field in the product feature field sequence, audience feature field sequence, and marketing objective field sequence, and the encoding results are normalized to obtain the corresponding normalized feature sequence.

[0087] Based on the field weights corresponding to the fields in the normalized feature sequence of each field, the normalized features of each field are weighted and fused to generate product feature vector, audience feature vector and marketing target vector. The field weights are used to characterize the importance of each field to the task of generating financial marketing copy.

[0088] The product feature vector, audience feature vector, and marketing objective vector are concatenated, aligned in dimensions, and mapped to a unified feature space to obtain the copy input representation.

[0089] In this embodiment, the generation of the candidate marketing copy set includes the following specific steps:

[0090] The text input representation is converted into a set of input sequence representations, and then input into the encoder of the text generation model to perform sequence feature encoding processing, so as to obtain the encoder output sequence representation and encoder state;

[0091] The copy generation model is constructed using a pointer generation network structure, and improvements are made to the pointer generation network structure. The improvements include: introducing a pointer selection substructure based on financial semantic constraints into the pointer mechanism, performing constraint screening on the position of the input sequence before calculating the replication probability, and introducing a review feature reverse modulation substructure into the generation probability control unit, performing reverse modulation on the fusion weight in the fusion stage of generation probability and replication probability.

[0092] The sequence feature encoding process includes: performing vector mapping processing on each sequence element in the input sequence representation set according to the arrangement order of each sequence element in the input sequence, mapping each sequence element to a sequence feature representation of a unified dimension; during the sequence mapping process, introducing sequence position information to perform position-related encoding on each sequence feature representation to represent the relative order relationship of different sequence elements in the input sequence; based on the feature association relationship between adjacent sequence elements and the entire sequence, performing context association modeling processing on each sequence feature representation to integrate the context feature information of the preceding and following sequence elements; after completing the context association modeling, performing layer-by-layer update and state convergence processing on each sequence feature representation, outputting an encoder output sequence representation that represents the overall semantic features of the input sequence, and simultaneously generating an encoder state that represents the overall encoding state of the input sequence;

[0093] Based on the current decoding state of the decoder and the output sequence representation of the encoder, an attention mechanism is used to calculate the attention weights corresponding to each input sequence position, thus obtaining the attention weight sequence.

[0094] The attention weight sequence is obtained by: inputting the current decoding state of the decoder and the sequence feature representations corresponding to each input sequence position in the encoder output sequence representation into the same feature mapping processing unit to obtain an intermediate association representation that characterizes the feature matching relationship between the decoding state and the input sequence position; based on the intermediate association representation, calculating an association score that characterizes the degree of correlation between the decoding state and each input sequence position through a weighted summation method, wherein the magnitude of the association score reflects the intensity of attention of the decoding state to the corresponding input sequence position; performing normalization processing on the association scores corresponding to each input sequence position to convert each association score into a weight value that satisfies the probability distribution constraint; arranging the normalized weight values ​​according to the order of the input sequence positions to form an attention weight sequence that characterizes the degree of attention of the decoding state to each input sequence position;

[0095] The encoder output sequence representation, attention weight sequence, and input pointer selection substructures of financial product semantic labels, risk semantic types, and text structure semantic roles obtained by parsing the text input representation are used to constrain and filter the input sequence positions to obtain a set of input sequence positions that can participate in replication. Then, the attention weight sequence is constrained and normalized based on the set of input sequence positions that can participate in replication to obtain a constrained attention weight sequence.

[0096] The semantic tags for financial products include semantic identification information used to characterize the core attributes of financial products. The semantic identification information includes product category identifiers, return or fee rate attribute identifiers, term or rule attribute identifiers, and product risk level identifiers.

[0097] The risk semantic types include risk semantic identification information used to characterize the semantic attributes of risk-related content in financial marketing copy. The risk semantic identification information includes risk semantics of uncertainty of returns, risk semantics of market volatility, risk semantics of liquidity, risk semantics of compliance restrictions, and risk semantics of the possibility of loss.

[0098] The semantic roles of the copywriting structure include structural semantic identification information used to characterize the functional positioning of different text fragments in the overall copywriting structure of financial marketing copywriting. The structural semantic identification information includes semantic roles for product introduction, benefit description, risk warning, applicable object description, and compliance statement.

[0099] The constraint filtering includes: in the pointer selection substructure, based on the sequence feature representations corresponding to each input sequence position in the encoder output sequence representation, aligning and mapping the financial product semantic tags, risk semantic types, and text structure semantic roles associated with each input sequence position to form position semantic constraint features; matching and determining the semantic constraint features of each input sequence position with the generation semantic requirements corresponding to the current decoding state according to semantic matching rules, identifying input sequence positions that meet semantic consistency constraints and risk compliance constraints; marking input sequence positions that pass the semantic matching determination as valid copy positions, and marking input sequence positions that fail the matching determination as invalid copy positions, forming a set of input sequence positions that can participate in copying;

[0100] The semantic matching rules are constructed based on product compliance requirements, risk disclosure standards, and copywriting structure constraints in financial marketing scenarios. They include: product risk consistency rules based on the correspondence between financial product semantic tags and risk semantic types, used to limit the range of risk semantic types allowed for different financial product semantic tags; risk structure constraint rules based on the constraint relationship between risk semantic types and copywriting structure semantic roles, used to limit the legitimate semantic role positions of risk-related semantic content within the copywriting structure; and product structure adaptation rules based on the adaptation relationship between financial product semantic tags and copywriting structure semantic roles, used to limit the types of copywriting structure semantic roles that different financial product semantic tags can correspond to during the copywriting generation process.

[0101] The constraint normalization includes: in the pointer selection substructure, based on the set of input sequence positions that can participate in replication, performing position constraint processing on the attention weight sequence, suppressing the attention weights corresponding to input sequence positions marked as invalid replication positions to zero values; re-normalizing the attention weights corresponding to the retained valid replication positions, so that the attention weights corresponding to each valid replication position numerically satisfy the sum constraint condition; arranging the normalized attention weights according to the order of the input sequence positions to obtain a constraint attention weight sequence distributed only within the set of input sequence positions that can participate in replication;

[0102] In the pointer mechanism, the replication probability distribution is calculated based on the constrained attention weight sequence, and the encoder output sequence representation is weighted and converged based on the constrained attention weight sequence to obtain the context representation;

[0103] The calculation of the replication probability distribution includes: in the pointer mechanism, using the constraint attention weight sequence as the probability weight basis for each input sequence position to participate in the replication behavior, for the sequence feature representation corresponding to each input sequence position in the encoder output sequence representation, the corresponding constraint attention weight is associated and mapped with the sequence feature representation of the input sequence position; for input sequence positions with the same term identifier, their corresponding constraint attention weights are accumulated and aggregated to obtain the cumulative weight value of each term under the replication path; the cumulative weight values ​​corresponding to all terms that can participate in replication are normalized to convert the cumulative weight value of each term into a value that satisfies the probability distribution constraint, forming a replication probability distribution used to characterize the probability of each replicable term being selected under the replication path;

[0104] In the decoder, the generation probability distribution is calculated based on the context representation, encoder state, and current decoding state of the decoder. The generation probability control unit calculates the relative weight relationship between the generation probability distribution and the replication probability distribution and outputs the fusion weight.

[0105] The calculation of the generation probability distribution includes: in the decoder, the context representation, encoder state, and current decoder state are concatenated to form a joint decoding feature that characterizes the generation conditions at the current decoding moment; a linear mapping process is performed on the joint decoding feature to map it to a generation feature space with the same dimension as the target vocabulary space, resulting in generation feature values ​​for each candidate word in the target vocabulary space; based on the generation feature values ​​corresponding to each candidate word, a generation score value is calculated for each candidate word in the target vocabulary space, so that each candidate word corresponds to a score result that characterizes its probability of being generated; a normalization process is performed on the generation score values ​​of all candidate words to convert the generation score values ​​of each candidate word into generation probability values ​​that satisfy the probability distribution constraints, forming a generation probability distribution that characterizes the probability of each candidate word being output through the generation path;

[0106] The calculation and output of the fusion weight includes: the generation probability control unit acquiring the current decoding state, context representation, and copy probability distribution information of the decoder at the current decoding time, where the distribution feature information is used to characterize the overall attention of the copy path to the input sequence at the current decoding time; performing joint feature mapping on the current decoding state and context representation to obtain generation path features that characterize the semantic strength of the generation path at the current decoding time; performing joint input processing on the generation path features and the distribution features of the copy probability distribution, and performing comparative modeling on the two to evaluate the relative fit of output terms using the generation path or the copy path at the current decoding time; calculating a weight value to characterize the relative proportion of the generation path and the copy path based on the comparative modeling results, and performing constraint processing on the weight value to ensure that its value meets the range requirements for probability fusion calculation; and outputting the constrained weight value as the fusion weight.

[0107] Semantic constraint features are obtained from the text input representation parsing, and the semantic constraint features are input into the review features to obtain the modulation coefficients by the reverse modulation substructure. The modulation coefficients are then used to reverse modulate the fusion weights to obtain the modulation fusion weights.

[0108] The modulation coefficients are obtained by: extracting semantic constraint features related to financial product attributes, risk constraints, and text structure constraints from the text input representation; grouping the semantic constraint features according to constraint categories to form constraint feature subsets that respectively characterize the strength of product constraints, risk constraints, and structural constraints; performing numerical standardization on the feature values ​​in each constraint feature subset to ensure that the feature values ​​under different constraint categories are within a uniform scale range; calculating the corresponding constraint strength index based on each standardized constraint feature subset to quantify the degree of restriction of different types of semantic constraints on text generation in the current generation task; weighting and integrating the constraint strength indices to obtain a comprehensive constraint strength value that characterizes the overall semantic constraint strength; and calculating the relative impact of the generation path and the replication path on the adjustment required at the current decoding time based on the comprehensive constraint strength value, and mapping the relative impact to modulation coefficients.

[0109] The reverse modulation calculation includes: in the review feature reverse modulation substructure, obtaining the fusion weight output by the generation probability control unit and the modulation coefficient calculated from the semantic constraint features; using the fusion weight as the basic weight value and the modulation coefficient as the modulation control quantity, performing reverse modulation processing on the fusion weight according to the semantic constraint strength direction represented by the modulation coefficient, wherein, when the comprehensive constraint strength value is not less than the constraint strength threshold, the weight value of the corresponding generation path proportion in the fusion weight is suppressed and adjusted, and the weight value of the copy path proportion is correspondingly enhanced; when the comprehensive constraint strength value is less than the constraint strength threshold, the weight value of the corresponding generation path proportion in the fusion weight is enhanced and adjusted, and the weight value of the copy path proportion is correspondingly weakened; the modulation coefficient threshold is a threshold used to distinguish the adjustment direction of the generation path proportion; after the adjustment is completed, performing numerical normalization and range constraint processing on the modulated fusion weight, so that the generation path proportion and the copy path proportion meet the value requirements of the probability fusion calculation, and obtaining the modulated fusion weight;

[0110] The generation probability distribution and the replication probability distribution are fused and calculated based on the modulation fusion weight to obtain the output probability distribution. Then, stepwise decoding is performed based on the output probability distribution to generate a set of candidate marketing copy. The output probability of the candidate terms in the output probability distribution is obtained by weighting and summing the generation probability and replication probability of the candidate terms according to the modulation fusion weight.

[0111] The stepwise decoding includes: at the current decoding moment, using the output probability distribution as the basis for candidate term selection, determining the output term at the current decoding moment based on the probability value corresponding to each candidate term in the output probability distribution, and recording the output term as part of the generated sequence; inputting the current decoded output term into the decoder, updating the current decoding state of the decoder, and entering the next decoding moment in combination with the updated current decoding state; repeatedly performing candidate term selection, decoding state update, and generated sequence expansion processing based on the updated output probability distribution; when the length of the generated sequence reaches the limited length condition, determining that the decoding termination condition is met; and collecting the complete term sequences corresponding to each decoding path obtained when the decoding termination condition is met to generate a candidate marketing copy set including multiple generated results.

[0112] In this embodiment, obtaining the text review feature set includes the following specific steps:

[0113] Each candidate marketing copy in the candidate marketing copy set is subjected to text normalization processing to obtain a standardized candidate copy sequence. The text normalization processing includes unifying the character format in the candidate marketing copy, cleaning abnormal symbols and irrelevant punctuation, standardizing the mapping of numbers and unit expressions, and unifying the text encoding and word segmentation boundaries.

[0114] The standardized candidate copy sequence is segmented to obtain a copy fragment set consisting of multiple copy fragments. Semantic features are extracted from each copy fragment in the copy fragment set and mapped to a unified semantic space to obtain a copy fragment-level semantic representation.

[0115] Global semantic features of standardized candidate copy sequences are extracted and mapped to a unified semantic space to obtain a holistic semantic representation of the copy. The copy fragment-level semantic representation is indexed and associated with the holistic semantic representation of the copy to obtain a set of semantic representations of candidate marketing copy. The index association includes binding the copy fragment-level semantic representation with the corresponding holistic semantic representation of the copy based on the positional relationship of the copy fragment in the original copy sequence.

[0116] Based on the semantic representation set, risk element extraction processing is performed on the risk-related semantic content in the candidate marketing copy to obtain a risk element set. Each risk element in the risk element set is then coded into fields according to risk category, risk object, risk triggering condition, and risk disclosure method to generate a risk element feature vector set. The field coding includes mapping each risk element to a corresponding field identifier according to risk category, risk object, risk triggering condition, and risk disclosure method, and performing unified formatting and vectorization combination on each field identifier.

[0117] The risk element feature vector set is combined and modeled according to the co-occurrence relationship of risk categories, the correlation relationship of risk triggering conditions, and the structural relationship of risk disclosure methods to obtain the risk pattern feature vector that represents the risk expression structure of candidate marketing copy.

[0118] The combined modeling includes: statistically analyzing the risk element feature vector set based on the co-occurrence relationship of risk categories; grouping and associating the risk element feature vectors corresponding to risk categories that appear simultaneously in the same candidate marketing copy to form a risk element combination reflecting the co-occurrence characteristics of risk categories; based on the association relationship of risk triggering conditions, performing association connection processing on the risk element feature vectors in the risk element combination according to the dependency and sequence relationship between triggering conditions to construct a risk element connection structure reflecting the logical relationship of risk triggering; and on this basis, according to the structural relationship of risk disclosure methods, and according to the disclosure order and structural position of risk elements in candidate marketing copy, structurally integrating the risk element connection structure to generate a risk pattern feature vector.

[0119] By binding the semantic representation of copy fragments, the semantic representation of copy as a whole, the set of risk element feature vectors, and the feature vector of risk patterns to the same candidate marketing copy identifier, and by structuring and encapsulating the various features after binding according to feature type, a copy review feature set for financial scenarios is obtained.

[0120] In this embodiment, the output of the structured audit result set includes the following specific steps:

[0121] Based on the candidate marketing copy identifiers, the semantic representations of copy fragments, the semantic representations of copy as a whole, the risk element feature vector set, and the risk pattern feature vector set in the copy review feature set are called to obtain the feature items to be reviewed that correspond one-to-one with the candidate marketing copy set.

[0122] Semantic matching calculations are performed on the text fragment-level semantic representation and the text overall-level semantic representation of each feature item to be reviewed, respectively, to obtain the fragment-level semantic matching score sequence and the overall-level semantic matching score, and the semantic tag field is determined based on the semantic matching score;

[0123] The semantic matching calculation includes: performing feature alignment processing on the semantic feature vectors of each segment in the semantic representation of the text fragments and the semantic representation of the text as a whole; calculating the semantic similarity value between the semantic feature vectors of each text fragment and the semantic feature vector of the text as a whole; forming a segment-level semantic matching score sequence; and simultaneously, quantifying the degree of overall semantic consistency based on the global semantic feature distribution within the text as a whole-level semantic representation to obtain the overall-level semantic matching score.

[0124] The determination of the semantic tag field includes: performing unified scale mapping and numerical normalization on the fragment-level semantic matching score sequence and the overall-level semantic matching score; based on the normalized semantic matching score results, labeling the semantic consistency status corresponding to the feature item to be reviewed, and generating the semantic tag field.

[0125] Based on the risk element feature vector set, perform risk element consistency judgment processing on the feature items to be reviewed, generate risk element judgment results, and generate risk identification fields based on the risk element judgment results;

[0126] The risk element consistency determination process includes: based on the risk element feature vector set, performing consistency comparison and legality verification on each risk element feature vector according to the field combination relationship of risk category, risk object, risk triggering condition and risk disclosure method; marking field combinations that pass the consistency verification as consistent risk elements, and marking field combinations that fail the consistency verification as abnormal risk elements; and generating risk element determination results by summarizing the determination results of consistent risk elements and abnormal risk elements.

[0127] The generation of the risk identification field includes: summarizing and analyzing the judgment status of each risk element in the feature item to be reviewed based on the risk element judgment results; when there is a combination of fields marked as abnormal risk elements, the corresponding feature item to be reviewed is marked as having a risk status and a corresponding risk identification field is generated; when all risk elements are marked as consistent risk elements, the corresponding feature item to be reviewed is marked as having no risk status and a corresponding risk identification field is generated.

[0128] Based on the risk pattern feature vector, risk pattern determination processing is performed on the feature items to be reviewed, risk pattern determination results are generated, and constraint type fields are determined based on the risk pattern determination results.

[0129] The risk pattern determination process includes: extracting pattern feature information representing the co-occurrence relationship of risk categories, the correlation relationship of risk triggering conditions, and the structural relationship of risk disclosure methods from the risk pattern feature vector of the feature items to be reviewed; matching and comparing the pattern feature information with the risk expression structure patterns allowed in financial marketing compliance scenarios item by item to identify whether the combination of risk categories meets the co-occurrence requirements, whether the risk triggering conditions meet the logical correlation constraints, and whether the arrangement of risk disclosure methods in the text structure meets the disclosure structure specifications; and comprehensively judging the above matching and comparison results to generate a risk pattern determination result that represents whether the risk expression structure of the candidate marketing text meets the compliance requirements.

[0130] The determination of the constraint type field includes: based on the risk mode determination result, identifying the specific risk expression structure type that triggers the risk mode determination, and classifying the risk expression structure type; determining the constraint type identifier to be applied to the candidate marketing copy according to the constraint attribute corresponding to the classification identifier, wherein the constraint type identifier is used to characterize the restriction method corresponding to the risk expression structure in the review process; and writing the constraint type identifier into the structured review result as the constraint type field.

[0131] Based on the index association relationship between the semantic representation of text fragments and the semantic representation of text as a whole, the text fragments associated with the risk identification field and the constraint type field are located, and the corresponding text fragment index information and text sequence position range information are extracted to generate the constraint position field.

[0132] The segment-level semantic matching score sequence, the overall-level semantic matching score, the risk element judgment result and the risk pattern judgment result are mapped to the judgment quantity, and the score fusion calculation is performed under a unified judgment scale to obtain the fusion score;

[0133] The fusion score is obtained by: performing sequence aggregation processing on multiple fragment-level semantic matching scores to generate a fragment-level semantic consistency score; performing weighted fusion calculation on the fragment-level semantic consistency score and the overall-level semantic matching score after performing unified scale mapping to obtain a semantic consistency fusion score; based on the risk element judgment results, statistically analyzing the number and distribution of abnormal risk elements in the feature items to be reviewed to obtain the judgment quantity at the risk element level; simultaneously, based on the risk pattern judgment results, summarizing the abnormalities in the risk expression structure to obtain the judgment quantity at the risk pattern level; and performing unified scale mapping and summarization calculation on the judgment quantities at the risk element level and the judgment quantities at the risk pattern level to obtain a unified risk judgment quantification value; and performing weighted fusion calculation on the semantic consistency fusion score and the risk judgment quantification value to obtain a fusion score.

[0134] The fusion score is normalized and mapped to obtain the audit confidence field. The output is a structured audit result set including the risk identifier field, semantic label field, constraint type field, constraint location field, and audit confidence field.

[0135] In this embodiment, the generation of the target marketing copy set includes the following specific steps:

[0136] The candidate marketing copy set and the corresponding structured review result set are called by field according to the candidate marketing copy identifier, and constraint processing entries corresponding to each candidate marketing copy are constructed.

[0137] Based on the constraint processing items, compliance constraint verification processing is performed on the candidate marketing copy, the compliance constraint items corresponding to the constraint type field in the candidate marketing copy are identified, and the scope of risk semantic content that needs to be constrained is determined by combining the risk identifier field and the semantic tag field.

[0138] Based on the constraint location field, risk fragment location processing is performed on the risk semantic content that needs to be constrained, and the corresponding text fragment index information and text sequence location range information are extracted to obtain the set of constrained text fragments;

[0139] Based on the constraint type field, risk identifier field, and audit confidence field, the constraint triggering strategy matching process is performed on the set of constrained text fragments to determine the constraint triggering strategy identifier corresponding to each constrained text fragment.

[0140] The constraint trigger strategy matching process includes: determining the constraint processing category corresponding to each constrained text fragment based on the constraint type field; identifying the risk status attribute of the constrained text fragment based on the risk identifier field; classifying the confidence level of the constrained text fragment based on the review confidence level field; inputting the constraint processing category, risk status attribute, and confidence level as matching conditions into the constraint trigger strategy set to perform matching retrieval processing, filtering out constraint trigger strategies that match the condition combination of each constrained text fragment, and determining a corresponding constraint trigger strategy identifier for each constrained text fragment;

[0141] The constraint triggering strategy identifier includes a constraint processing action identifier, a correction strength identifier, an applicable risk type identifier, and a processing priority identifier.

[0142] Based on the constraint trigger strategy identifier, the content of each constrained copy fragment is modified to generate a modified copy fragment. The modified copy fragment is then backfilled into the copy sequence position range information corresponding to the original candidate marketing copy according to the copy fragment index information to obtain the target marketing copy. Finally, all target marketing copy are collected to generate a target marketing copy set.

[0143] Example 1:

[0144] To verify the feasibility of this invention in practice, it was applied to an actual online financial product marketing scenario conducted by a financial institution. This institution regularly pushes marketing copy for wealth management, funds, and insurance to different audiences through multiple channels. The copy must simultaneously meet requirements for differentiated product expression, audience matching, and strict compliance and risk disclosure. Under the existing work model, marketing personnel mainly rely on manual experience or template-based tools to write the copy, which is then reviewed line by line by compliance personnel. This results in low generation efficiency, severe copy homogenization, omissions in risk descriptions, and concentrated review pressure, making it difficult to achieve rapid and stable output during peak marketing periods.

[0145] In this scenario, the system of this invention first accesses the existing marketing data environment of the financial institution, automatically acquiring financial product information, marketing activity information, and target audience-related data. It then performs unified preprocessing on data from different sources to form a standardized task dataset with a consistent structure. The system subsequently constructs a copywriting input representation based on this data, fusing product attributes, audience characteristics, and marketing objectives within the same feature space. This allows the subsequent copywriting generation process to simultaneously perceive product constraints, audience differences, and marketing appeals. During the copywriting generation stage, the system introduces a semantic constraint-based copy filtering mechanism to ensure the copy accurately references key product information. Simultaneously, it combines a back-modulation probabilistic fusion mechanism to adjust the generation path for high-risk or strongly constrained content, thereby maintaining natural language while avoiding non-compliant expressions.

[0146] The generated candidate marketing copy automatically enters the review feature processing flow. The system standardizes the copy and extracts semantic features at both the overall and fragment levels to further identify risk-related expressions. By constructing risk element features and risk pattern features, the system can comprehensively analyze each candidate copy from multiple perspectives, including semantic consistency, completeness of risk elements, and rationality of risk disclosure structure. Subsequently, the system performs semantic matching and risk assessment on the review features, automatically outputting structured review results containing information such as risk identifiers, constraint types, constraint locations, and review confidence levels.

[0147] In actual operation, when the risk descriptions in the copy are insufficient or the structure does not conform to the specifications, the system can accurately locate the relevant statements and automatically trigger corresponding constraint policies based on the review results to revise or reorganize the copy content. The revised copy, while maintaining the original marketing intent, provides more complete risk warnings and a clearer structure, reducing the need for repeated manual modifications. Related application records show that in the financial institution's daily marketing process, the overall processing time for copy generation and review has been significantly shortened, the frequency of manual review intervention has been significantly reduced, and the compliance department's feedback on the consistency of risk in the copy has become more stable. This embodiment demonstrates that the present invention effectively solves the problem of balancing efficiency in financial marketing copy generation with compliance review, exhibiting good applicability and practical value in real business environments.

[0148] Table 1. Performance Comparison of the Invention and Traditional Financial Marketing Copywriting Generation and Review Methods

[0149] Indicator Categories Traditional rules / manual methods Method of the present invention Time to generate a single piece of copy (s) 3.20 2.45 Review time for a single piece of copy (s) 5.60 3.10 Overall copywriting processing time (s) 8.80 5.55 Risk factor missed detection rate (%) 6.8 4.1 Accuracy rate of risk location (%) 86.5 93.2 Semantic consistency score (0–1) 0.81 0.89 Compliance correction rework rate (%) 18.4 12.7

[0150] As can be clearly seen from Table 1, the method of the present invention is superior to the traditional method in many indicators.

[0151] In terms of the time required to generate a single piece of text, traditional methods average 3.20 seconds, while the method of this invention reduces this to 2.45 seconds. This improvement does not stem from simple model acceleration, but rather from the introduction of a semantically constrained copy filtering mechanism in the text generation stage. This mechanism allows the model to prioritize the use of existing high-confidence input information during the generation process, reducing the number of times invalid generation paths are explored, thereby reducing the overall decoding time.

[0152] Regarding document review time, the traditional method takes 5.60 seconds, while the method of this invention takes only 3.10 seconds, a significant reduction. This improvement mainly stems from the structured modeling of risk elements and risk patterns in the review feature module, which gives risk-related content a clear organizational form in the semantic space. This avoids the problem of repeated scanning of the entire text by manual or rule-based systems, making the review and judgment process more focused and efficient.

[0153] The overall processing time for the text was reduced from 8.80 seconds to 5.55 seconds, demonstrating the synergistic optimization effect of the generation and review stages. Because this invention embeds semantic constraints and risk modulation mechanisms in the generation stage, the overall quality of candidate texts entering the review stage is higher, thereby reducing the complex judgment burden of the review stage. This kind of sequential processing is not available in traditional serial processing methods.

[0154] The risk element omission rate decreased from 6.8% to 4.1%. This is because the present invention adopts a risk element fieldization and consistency judgment mechanism, which not only identifies risk keywords, but also comprehensively considers the matching relationship between risk objects, triggering conditions and disclosure methods, thereby reducing the probability of risk omission from a structural level.

[0155] In terms of risk location accuracy, the method of this invention achieves 93.2%, which is higher than the 86.5% of the traditional method. This improvement comes from the index association design of fragment-level semantic representation and global-level semantic representation, which enables the system to accurately locate risk-related fragments while maintaining global semantic understanding, rather than relying on fuzzy rule matching.

[0156] The semantic consistency score improved from 0.81 to 0.89, indicating that the generated copy has improved in both overall coherence and local semantic fit. This is directly related to the reverse modulation probabilistic fusion mechanism, which can suppress unstable generation paths under high constraints, making the output content more aligned with the predetermined semantic goals.

[0157] The compliance correction rework rate decreased from 18.4% to 12.7%, demonstrating the invention's significant advantage in first-time pass rate. The fundamental reason lies in the fact that the constraint triggering strategy is not a simple deletion or replacement, but rather a differentiated correction based on risk status and confidence level, thereby reducing secondary risks caused by excessive modifications.

[0158] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A machine learning-based automatic generation and review system for financial marketing copy, characterized in that: include: The data processing module is used to acquire financial marketing task data and preprocess it to obtain a standardized task dataset; The input building module is used to construct and fuse product feature vectors, audience feature vectors, and marketing goal vectors based on a standardized task dataset to obtain the copy input representation; The copy generation module is used to generate a set of candidate marketing copy by copying, filtering and probabilistically fusing the input copy representation through a copy generation model. The review feature module is used to perform semantic parsing and risk feature extraction on the candidate marketing copy set to obtain the copy review feature set. The review and judgment module is used to perform semantic matching and risk assessment on the document review feature set and output a structured review result set. The copywriting constraint module is used to perform copywriting constraints and content correction on the candidate marketing copy set based on the structured review result set, and generate the target marketing copy set; The results storage module is used to associate and encapsulate the target marketing copy set and the structured review result set, and write them to the review result database.

2. The machine learning-based automatic generation and review system for financial marketing copy as described in claim 1, characterized in that, The modules are connected in the following way: Acquire financial marketing task data and preprocess it to obtain a standardized task dataset; Based on a standardized task dataset, product feature vectors, audience feature vectors, and marketing goal vectors are constructed and integrated to obtain a copywriting input representation. The text input representation is fed into the text generation model, and a set of candidate marketing copy is generated through semantic constraint copy filtering and back-modulation probability fusion. Semantic analysis and risk feature extraction were performed on the candidate marketing copy set to obtain a copy review feature set for financial scenarios; Perform semantic matching and risk assessment on the copy review feature set, and output a structured review result set that corresponds one-to-one with the candidate marketing copy set; Based on the structured review result set, the copywriting is constrained and the content is revised to generate a target marketing copy set; The target marketing copy set and the corresponding structured audit result set are associated, encapsulated, and written into the audit result database.

3. The machine learning-based automatic generation and review system for financial marketing copy as described in claim 2, characterized in that, The financial marketing task data includes financial product data, marketing activity data, and target audience data. The preprocessing includes data format standardization, missing data processing, abnormal data filtering, data encoding, and feature normalization.

4. The machine learning-based automatic generation and review system for financial marketing copy as described in claim 2, characterized in that, The process of obtaining the text input representation includes the following specific steps: Financial product data, target audience data, and marketing campaign data are extracted from the standardized task dataset and organized according to the field structure requirements to obtain product feature field sequence, audience feature field sequence, and marketing target field sequence. Data encoding is performed on each field in the product feature field sequence, audience feature field sequence, and marketing objective field sequence, and the encoding results are normalized to obtain the corresponding normalized feature sequence. Based on the field weights corresponding to the fields in the normalized feature sequence of each field, the normalized features of each field are weighted and fused to generate product feature vector, audience feature vector and marketing target vector. The product feature vector, audience feature vector, and marketing objective vector are concatenated, aligned in dimensions, and mapped to a unified feature space to obtain the copy input representation.

5. The machine learning-based automatic generation and review system for financial marketing copy as described in claim 2, characterized in that, The generation of the candidate marketing copy set includes the following specific steps: The text input representation is converted into a set of input sequence representations, and then input into the encoder of the text generation model to perform sequence feature encoding processing, so as to obtain the encoder output sequence representation and encoder state; The copy generation model is constructed using a pointer generation network structure, and improvements are made to the pointer generation network structure. The improvements include: introducing a pointer selection substructure based on financial semantic constraints into the pointer mechanism, performing constraint screening on the position of the input sequence before calculating the replication probability, and introducing a review feature reverse modulation substructure into the generation probability control unit, performing reverse modulation on the fusion weight in the fusion stage of generation probability and replication probability. Based on the current decoding state of the decoder and the output sequence representation of the encoder, an attention mechanism is used to calculate the attention weights corresponding to each input sequence position, thus obtaining the attention weight sequence. The encoder output sequence representation, attention weight sequence, and input pointer selection substructures of financial product semantic labels, risk semantic types, and text structure semantic roles obtained by parsing the text input representation are used to constrain and filter the input sequence positions to obtain a set of input sequence positions that can participate in replication. Then, the attention weight sequence is constrained and normalized based on the set of input sequence positions that can participate in replication to obtain a constrained attention weight sequence. In the pointer mechanism, the replication probability distribution is calculated based on the constrained attention weight sequence, and the encoder output sequence representation is weighted and converged based on the constrained attention weight sequence to obtain the context representation; In the decoder, the generation probability distribution is calculated based on the context representation, encoder state, and current decoding state of the decoder. The generation probability control unit calculates the relative weight relationship between the generation probability distribution and the replication probability distribution and outputs the fusion weight. Semantic constraint features are obtained from the text input representation parsing, and the semantic constraint features are input into the review features to obtain the modulation coefficients by the reverse modulation substructure. The modulation coefficients are then used to reverse modulate the fusion weights to obtain the modulation fusion weights. The generation probability distribution and the replication probability distribution are fused and calculated based on the modulation fusion weight to obtain the output probability distribution. Then, stepwise decoding is performed based on the output probability distribution to generate a set of candidate marketing copy. The output probability of the candidate terms in the output probability distribution is obtained by weighting the generation probability and replication probability of the candidate terms according to the modulation fusion weight.

6. The machine learning-based automatic generation and review system for financial marketing copy as described in claim 2, characterized in that, The acquisition of the text review feature set includes the following specific steps: Each candidate marketing copy in the candidate marketing copy set is subjected to text normalization processing to obtain a standardized candidate copy sequence; The standardized candidate copy sequence is segmented to obtain a copy fragment set consisting of multiple copy fragments. Semantic features are extracted from each copy fragment in the copy fragment set and mapped to a unified semantic space to obtain a copy fragment-level semantic representation. The global semantic features of the standardized candidate copy sequence are extracted and mapped to a unified semantic space to obtain the overall semantic representation of the copy. The copy fragment-level semantic representation is indexed and associated with the overall semantic representation of the copy to obtain the semantic representation set of candidate marketing copy. Based on the semantic representation set, risk element extraction is performed on the risk-related semantic content in the candidate marketing copy to obtain a risk element set. Each risk element in the risk element set is then field-coded according to risk category, risk object, risk triggering condition and risk disclosure method to generate a risk element feature vector set. The risk element feature vector set is combined and modeled according to the co-occurrence relationship of risk categories, the correlation relationship of risk triggering conditions, and the structural relationship of risk disclosure methods to obtain the risk pattern feature vector that represents the risk expression structure of candidate marketing copy. By binding the semantic representation of copy fragments, the semantic representation of copy as a whole, the set of risk element feature vectors, and the feature vector of risk patterns to the same candidate marketing copy identifier, and by structuring and encapsulating the various features after binding according to feature type, a copy review feature set for financial scenarios is obtained.

7. The machine learning-based automatic generation and review system for financial marketing copy as described in claim 2, characterized in that, The output of the structured audit result set includes the following specific steps: Based on the candidate marketing copy identifiers, the semantic representations of copy fragments, the semantic representations of copy as a whole, the risk element feature vector set, and the risk pattern feature vector set in the copy review feature set are called to obtain the feature items to be reviewed that correspond one-to-one with the candidate marketing copy set. Semantic matching calculations are performed on the text fragment-level semantic representation and the text overall-level semantic representation of each feature item to be reviewed, respectively, to obtain the fragment-level semantic matching score sequence and the overall-level semantic matching score, and the semantic tag field is determined based on the semantic matching score; Based on the risk element feature vector set, perform risk element consistency judgment processing on the feature items to be reviewed, generate risk element judgment results, and generate risk identification fields based on the risk element judgment results; Based on the risk pattern feature vector, risk pattern determination processing is performed on the feature items to be reviewed, risk pattern determination results are generated, and constraint type fields are determined based on the risk pattern determination results. Based on the index association relationship between the semantic representation of text fragments and the semantic representation of text as a whole, the text fragments associated with the risk identification field and the constraint type field are located, and the corresponding text fragment index information and text sequence position range information are extracted to generate the constraint position field. The segment-level semantic matching score sequence, the overall-level semantic matching score, the risk element judgment result and the risk pattern judgment result are mapped to the judgment quantity, and the score fusion calculation is performed under a unified judgment scale to obtain the fusion score; The fusion score is normalized and mapped to obtain the audit confidence field. The output is a structured audit result set including the risk identifier field, semantic label field, constraint type field, constraint location field, and audit confidence field.

8. The machine learning-based automatic generation and review system for financial marketing copy as described in claim 2, characterized in that, The generation of the target marketing copy set includes the following specific steps: The candidate marketing copy set and the corresponding structured review result set are called by field according to the candidate marketing copy identifier, and constraint processing entries corresponding to each candidate marketing copy are constructed. Based on the constraint processing items, compliance constraint verification processing is performed on the candidate marketing copy, the compliance constraint items corresponding to the constraint type field in the candidate marketing copy are identified, and the scope of risk semantic content that needs to be constrained is determined by combining the risk identifier field and the semantic tag field. Based on the constraint location field, risk fragment location processing is performed on the risk semantic content that needs to be constrained, and the corresponding text fragment index information and text sequence location range information are extracted to obtain the set of constrained text fragments; Based on the constraint type field, risk identifier field, and audit confidence field, the constraint triggering strategy matching process is performed on the set of constrained text fragments to determine the constraint triggering strategy identifier corresponding to each constrained text fragment. Based on the constraint trigger strategy identifier, the content of each constrained copy fragment is modified to generate a modified copy fragment. The modified copy fragment is then backfilled into the copy sequence position range information corresponding to the original candidate marketing copy according to the copy fragment index information to obtain the target marketing copy. Finally, all target marketing copy are collected to generate a target marketing copy set.

Images