Security transaction instruction intelligent analysis method and system based on large model

Abstract

The invention relates to the technical field of financial science and technology, and discloses a security transaction instruction intelligent analysis method and system based on a large model, and the method comprises the steps: carrying out the compliance screening of an instruction text flow and an instruction feature chart, and obtaining an initial text flow and an initial feature chart; analyzing a transaction feature factor corresponding to the initial feature chart, extracting a space-time structure feature vector from the initial feature chart, and analyzing a structure description text corresponding to the space-time structure feature vector to generate a chart analysis text corresponding to the initial feature chart; calculating segmentation entropy loss corresponding to the initial segmented text, and performing segmentation iterative optimization on the initial text stream to obtain a target segmented text; generating an instruction target text corresponding to the security transaction instruction, calculating a structured likelihood between instruction tags, and determining an association aggregation tag in the instruction tags; and analyzing the transaction intention corresponding to the security transaction instruction, and generating an analysis result corresponding to the security transaction instruction. According to the invention, the accuracy of security transaction instruction analysis can be improved.

Description

Technical Field

[0001] This invention relates to a method and system for intelligent parsing of securities trading orders based on large models, belonging to the field of financial technology. Background Technology

[0002] In the financial field, traders often use colloquial and non-standardized instructions to improve communication efficiency. Although such trading texts contain core elements such as trading targets, operational directions, and risk control conditions, they present analytical challenges. First, the lack of grammatical structure leads to the dispersion of key information, making it difficult for traditional Natural Language Processing (NLP) techniques to locate elements through syntax tree parsing. Second, the dense superposition of elements can easily lead to semantic ambiguity. For example, in the statement "buy 50 lots of Ping An Bank and China Merchants Bank," the attribution of "50 lots of each" requires deep semantic reasoning.

[0003] Existing technologies mostly use rule matching or traditional machine learning models for semantic parsing. Rule matching relies on manually written grammar templates, which are difficult to cover the dynamically changing trading terms and expressions in the financial field. For example, the instructions of emerging "quantitative hedging arbitrage" strategies are prone to parsing failure due to rule lag. They can only complete the semantic parsing of content, but it is difficult to achieve the semantic parsing of associations. They cannot accurately restore the relationship between elements and instructions, resulting in logical gaps in the information presented in structured tables, which in turn reduces the accuracy of securities trading instruction parsing. Summary of the Invention

[0004] This invention provides a method and system for intelligent parsing of securities trading orders based on a large model, the main purpose of which is to improve the accuracy of securities trading order parsing.

[0005] To achieve the above objectives, the present invention provides a method for intelligent parsing of securities trading orders based on a large model, comprising: Receive a securities trading instruction to be processed, perform multimodal information splitting on the securities trading instruction to obtain an instruction text stream and an instruction feature chart, perform compliance screening on the instruction text stream and the instruction feature chart to obtain an initial text stream and an initial feature chart; Obtain the trading element library of the preset securities model, analyze the trading feature factors corresponding to the initial feature chart, combine the trading element library and the trading feature factors, extract the spatiotemporal structure feature vector from the initial feature chart, analyze the structural description text corresponding to the spatiotemporal structure feature vector, and generate the chart parsing text corresponding to the initial feature chart. The initial text stream is segmented to obtain initial segmented text. The segmentation entropy loss corresponding to the initial segmented text is calculated. Based on the segmentation entropy loss, the initial text stream is segmented iteratively optimized to obtain the target segmented text. By combining the chart parsing text and the target segmented text, the instruction target text corresponding to the securities trading instruction is generated, the instruction tags corresponding to the instruction target text are extracted, the structured likelihood between the instruction tags is calculated, and the associated aggregation tags in the instruction tags are determined based on the structured likelihood. Based on the associated aggregation tags, the trading intent corresponding to the securities trading instruction is analyzed. Combining the trading intent and the structured likelihood, the parsing result corresponding to the securities trading instruction is generated.

[0006] Optionally, the step of performing multimodal information decomposition processing on the securities trading instruction to obtain the instruction text stream and instruction feature chart includes: The securities trading instructions are subjected to voiceprint feature separation to obtain voiceprint feature vectors and denoised speech streams; The denoised speech stream is converted into speech-to-text to obtain structured instruction text; Based on a preset set of business logic rules, the structured instruction text is subjected to text filtering to obtain an instruction text stream; The visual association features corresponding to the instruction text stream are extracted, and the instruction feature chart corresponding to the securities trading instruction is analyzed by combining the voiceprint feature vector and the visual association features.

[0007] Optionally, the compliance screening of the instruction text stream and the instruction feature chart to obtain the initial text stream and the initial feature chart includes: The instruction text stream is subjected to risk labeling processing to obtain a labeled text stream; The labeled text stream is then de-identified to obtain a de-identified text stream; The desensitized text stream is subjected to rule adaptation processing to obtain an initial text stream; Based on the initial text stream, the initial feature chart is subjected to collaborative verification processing to obtain the initial feature chart.

[0008] Optionally, the step of combining the transaction element library and the transaction feature factors to extract the spatiotemporal structure feature vector from the initial feature chart includes: A knowledge graph is constructed from the transaction element database to obtain the transaction element graph; The transaction feature factors are encoded to obtain a feature factor vector; Identify the transaction element nodes in the transaction element graph and analyze the correlation strength between the feature factor vector and the transaction element nodes; Based on the correlation strength, the associated element nodes corresponding to the transaction feature factors are selected from the transaction element nodes; Based on the associated element nodes, extract the graph topology information corresponding to the transaction feature factors from the transaction element graph; Extract the time series information of the graph from the initial feature graph, and perform feature aggregation processing on the time series information of the graph and the topological structure information of the graph to obtain the spatiotemporal structure feature vector.

[0009] Optionally, calculating the segmentation entropy loss corresponding to the initial segmented text includes: Perform syntactic-semantic mapping on the initial segmented text to obtain a set of mapping relationships; Based on the set of mapping relationships, analyze the syntactic and semantic co-features corresponding to the initial segmented text; Based on the syntactic-semantic co-features, the initial segmented text is reasonably classified to obtain reasonable and unreasonable segmented fragments. The number of segments corresponding to the reasonable segmentation segments and the unreasonable segmentation segments is counted to obtain the number of reasonable segments and the number of unreasonable segments. The segmentation entropy loss corresponding to the initial segmented text is calculated by combining the number of reasonable segments, the syntactic-semantic co-features, and the number of unreasonable segments.

[0010] Optionally, the segmentation entropy loss corresponding to the initial segmented text is calculated by combining the number of reasonable segments, the syntactic-semantic co-features, and the number of unreasonable segments, including: Based on the aforementioned syntactic-semantic co-features, calculate the reasonable segmentation rate for each text segment in the initial segmented text; Based on the number of reasonable segments and the number of unreasonable segments, calculate the imbalance ratio corresponding to the initial segmented text; Based on the imbalance ratio, set a reasonable weight factor corresponding to the initial segmented text; Combining the reasonable weighting factor, the reasonable segmentation rate, the reasonable number of segments, and the unreasonable number of segments, the segmentation entropy loss corresponding to the initial segmented text is calculated using the following formula: ; Where A represents the segmentation entropy loss corresponding to the initial segmented text. This indicates that the initially segmented text is a valid segment. This indicates that the initially segmented text is an irrational segment, where x represents the number of irrational segments and y represents the number of irrational segments. Indicates the reasonable weighting factor. This represents the reasonable segmentation rate corresponding to the i-th text segment in the initial segmented text, where i represents the sequence number of the initial segmented text.

[0011] Optionally, calculating the reasonable segmentation rate for each text segment in the initial segmented text based on the syntactic-semantic co-features includes: Heterogeneous graph construction is performed on the syntactic-semantic co-functional features to obtain a syntactic-semantic joint graph structure; The syntax-semantic joint graph structure is subjected to convolutional fusion processing to obtain a collaborative feature matrix; The cooperative feature matrix is ​​subjected to adversarial constraint processing to obtain the constraint feature matrix; The constraint feature matrix is ​​subjected to spatiotemporal fusion processing to obtain fused constraint features; Analyze the feature granularity of the fusion constraint features and calculate the attention coefficient corresponding to the feature granularity; By combining the attention coefficient and the fusion constraint features, the reasonable segmentation rate corresponding to each text segment in the initial segmented text is calculated.

[0012] Optionally, calculating the structured likelihood between the instruction tags includes: Calculate the tag flow probability between the instruction tags and determine the probability mapping value corresponding to the instruction tag; Combining the tag flow probability and the probability mapping value, the structured likelihood between the instruction tags is calculated using the following formula: ; Where D represents the structured likelihood between instruction labels. This represents the tag transition probability when the state of the e-th instruction tag transitions to the state of the (e+1)-th instruction tag. This represents the e-th instruction label. This represents the (e+1)th instruction label. R represents the probability mapping value of the e-th instruction tag, where e represents the sequence number of the instruction tag and R represents the number of instruction tags.

[0013] Optionally, analyzing the trading intent corresponding to the securities trading instruction based on the associated aggregation tags includes: Analyze the tag semantics corresponding to the associated aggregated tags, and construct a semantic association network corresponding to the associated aggregated tags based on the tag semantics; The semantic association network is subjected to hierarchical clustering to obtain a hierarchical classification network; Extract the intent-driven features from the hierarchical classification network, perform intent matching on the core features of the network, and obtain the initial interaction intent; Calculate the coupling weight coefficients of the intent-dominant feature in the hierarchical classification network; Based on the coupling weight coefficient, the initial interaction intent is filtered to obtain the trading intent corresponding to the securities trading instruction.

[0014] To address the aforementioned problems, this invention also provides a securities trading instruction intelligent parsing system based on a large model, the system comprising: The instruction initial processing module is used to receive securities trading instructions to be processed, perform multimodal information splitting processing on the securities trading instructions to obtain instruction text stream and instruction feature chart, and perform compliance screening on the instruction text stream and instruction feature chart to obtain initial text stream and initial feature chart. The chart text analysis module is used to obtain the trading element library of the preset securities model, parse the trading feature factors corresponding to the initial feature chart, combine the trading element library and the trading feature factors, extract the spatiotemporal structure feature vector from the initial feature chart, analyze the structural description text corresponding to the spatiotemporal structure feature vector, and generate the chart parsing text corresponding to the initial feature chart. The text segmentation processing module is used to perform text segmentation processing on the initial text stream to obtain initial segmented text, calculate the segmentation entropy loss corresponding to the initial segmented text, and perform segmentation iterative optimization on the initial text stream based on the segmentation entropy loss to obtain the target segmented text; The instruction tag analysis module is used to combine the chart parsing text and the target segmented text to generate the instruction target text corresponding to the securities trading instruction, extract the instruction tags corresponding to the instruction target text, calculate the structured likelihood between the instruction tags, and determine the associated aggregation tags in the instruction tags based on the structured likelihood. The instruction parsing module is used to analyze the trading intent corresponding to the securities trading instruction based on the associated aggregation tags, and generate the parsing result corresponding to the securities trading instruction by combining the trading intent and the structured likelihood.

[0015] Compared to the problems described in the background technology, this invention, by performing multimodal information decomposition processing on the securities trading instructions, can break down complex trading instructions into information of different dimensions, providing a data foundation for subsequent accurate compliance screening and effectively improving the accuracy and efficiency of securities trading instruction compliance detection. This invention, by analyzing the trading feature factors corresponding to the initial feature chart, can transform the complex information presented by the chart into quantifiable and analyzable key factors, making the chart information more targeted and operable, facilitating the discovery of hidden trading patterns and trends behind the chart. Furthermore, this invention, by performing text segmentation processing on the initial text stream to obtain initial segmented text, can quickly and initially divide the original text, providing a foundation for subsequent refined processing and calculating the... The segmentation entropy loss corresponding to the initial segmented text can reveal the quality of the initial segmented text, facilitating subsequent iterative optimization. Furthermore, this invention generates the target text for the securities trading instruction by combining the chart parsing text and the target segmented text, thereby obtaining the core content of the securities trading instruction. By extracting the instruction tags corresponding to the target text, key semantic identifiers of the instruction can be obtained, providing a basis for subsequent calculation of the structured likelihood between the instruction tags. Finally, this invention analyzes the trading intent corresponding to the securities trading instruction based on the associated aggregation tags, thereby obtaining the detailed trading purpose of the securities trading instruction, facilitating the generation of accurate instruction parsing results. Therefore, the intelligent securities trading instruction parsing method and system based on a large model provided by this invention can improve the accuracy of securities trading instruction parsing. Attached Figure Description

[0016] Figure 1 A flowchart illustrating a method for intelligent parsing of securities trading instructions based on a large model, provided in an embodiment of the present invention; Figure 2 A schematic diagram of the segmentation and iterative optimization process in the intelligent parsing method for securities trading instructions based on a large model provided by the present invention; Figure 3 This is a schematic diagram of the modules for implementing the intelligent parsing system for securities trading instructions based on a large model, provided as an embodiment of the present invention.

[0017] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] This application provides a method for intelligent parsing of securities trading orders based on a large model. The executing entity of this method 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: a server, a terminal, etc. In other words, the method for intelligent parsing of securities trading orders based on a large model 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.

[0020] Example 1: Reference Figure 1 The diagram shown is a flowchart illustrating a securities trading order intelligent parsing method based on a large model, according to an embodiment of the present invention. In this embodiment, the securities trading order intelligent parsing method based on a large model includes: S1. Receive the securities trading instruction to be processed, perform multimodal information splitting processing on the securities trading instruction to obtain the instruction text stream and instruction feature chart, perform compliance screening on the instruction text stream and the instruction feature chart to obtain the initial text stream and initial feature chart.

[0021] This invention decomposes complex securities trading instructions into information of different dimensions by performing multimodal information decomposition processing, providing a data foundation for subsequent accurate compliance screening and effectively improving the accuracy and efficiency of securities trading instruction compliance detection.

[0022] The securities trading instruction refers to an instruction submitted by an investor or institution that contains various types of trading information and has not yet undergone compliance checks and processing. For example, an investor may issue an instruction through trading software to buy a certain number of shares of a stock and set a specific price. The instruction text stream refers to a continuous sequence of text information extracted from the securities trading instruction, which covers the core descriptive content of the trading instruction, such as the trading direction (buy or sell), trading price conditions (limit price, market price, etc.), and trading time requirements. The instruction feature chart refers to the result of visually displaying the key data and potential features in the securities trading instruction in the form of charts.

[0023] As an embodiment of the present invention, the step of performing multimodal information decomposition processing on the securities trading instruction to obtain an instruction text stream and an instruction feature chart includes: The securities trading instructions are subjected to voiceprint feature separation to obtain voiceprint feature vectors and denoised speech streams; The denoised speech stream is converted into speech-to-text to obtain structured instruction text; Based on a preset set of business logic rules, the structured instruction text is subjected to text filtering to obtain an instruction text stream; The visual association features corresponding to the instruction text stream are extracted, and the instruction feature chart corresponding to the securities trading instruction is analyzed by combining the voiceprint feature vector and the visual association features.

[0024] Wherein, the voiceprint feature vector and the denoised speech stream are respectively the biometric representation and denoised speech signal of the securities trading instruction; the structured instruction text is the textualized result of the denoised speech stream; the preset business logic rule set is the business compliance criterion used to filter the instruction text, and the instruction text stream is the compliant text of the structured instruction text after filtering based on the rule set: for example, the rule set includes rules such as "trading time must be between 9:30 and 15:00" and "single transaction amount ≤ 1 million yuan", and after filtering, the structured instruction text removes illegal content such as "ordering outside of trading hours" to obtain a compliant instruction text stream; the visual association feature is the corresponding visual transaction information feature of the instruction text stream.

[0025] Furthermore, the securities trading instructions can be subjected to voiceprint feature separation using voiceprint recognition and signal denoising algorithms to obtain voiceprint feature vectors and denoised speech streams. For example, the Mel Frequency Cepstral Coefficient (MFCC) algorithm can be used to extract voiceprint features, and wavelet denoising algorithms can be used to remove environmental noise. The denoised speech stream can be converted into structured text using Automatic Speech Recognition (ASR) technology to obtain structured instruction text. For instance, an end-to-end model based on deep learning (such as Transformer) can be used to convert the speech "Buy 100 shares of Tencent stock at market price" into structured text. Based on a preset business logic rule set, the structured text can be matched using a rule engine or regular expression. The instruction text undergoes text filtering to obtain an instruction text stream, such as removing instructions containing illegal trading targets or those with incorrect formats. Visual association features corresponding to the instruction text stream can be extracted using computer vision (CV) algorithms and natural language visual association models, such as identifying stock candlestick chart patterns mentioned in the instruction and extracting key technical indicators. Combining the voiceprint feature vector and the visual association features, a multimodal fusion analysis model is used to analyze the instruction feature chart corresponding to the securities trading instruction. For example, a neural network can be used to fuse voiceprint sentiment features, text trading information, and visual technical indicators to generate a visual chart containing trading intent, user identity credibility, and market trends.

[0026] This invention obtains an initial text stream and an initial feature chart by performing compliance screening on the instruction text stream and the instruction feature chart. It can accurately identify transaction instruction information that meets compliance requirements, exclude parts with potential violation risks, and ensure that the instructions entering the subsequent processing process are compliant. The initial text stream and the initial feature chart are the text stream and feature chart obtained after removing non-compliant content with the assistance of the instruction text stream and the instruction feature chart, respectively.

[0027] As an embodiment of the present invention, the compliance screening of the instruction text stream and the instruction feature chart to obtain an initial text stream and an initial feature chart includes: The instruction text stream is subjected to risk labeling processing to obtain a labeled text stream; The labeled text stream is then de-identified to obtain a de-identified text stream; The desensitized text stream is subjected to rule adaptation processing to obtain an initial text stream; Based on the initial text stream, the initial feature chart is subjected to collaborative verification processing to obtain the initial feature chart.

[0028] The labeled text stream is the output of the instruction text stream after semantic enhancement by a generative large model and pre-annotation by a risk classification model, supplementing implicit intent and risk information, and marking potential violation keywords or risk levels. The desensitized text stream is the result of the labeled text stream after hiding sensitive fields (such as account numbers and customer identity information) through data anonymization, generalization and other techniques, such as replacing the account number in the labeled text containing the customer's account with virtual characters. The initial text stream is the final text output that conforms to the current compliance rules after the desensitized text stream has undergone rule adaptation processing and is compared and filtered with a compliance rule library that dynamically adjusts weights according to the real-time market environment. If new market regulations restrict the trading of a certain type of stock, the desensitized text stream containing that type of stock will be removed, and the compliant part will be retained to form the initial text stream.

[0029] Furthermore, the instruction text stream can be risk-labeled using a risk classification model based on natural language processing to obtain a labeled text stream. For example, the BERT model can be used to identify risk keywords such as "abnormally high price" and "new stock" in the instruction "buy a new stock at an abnormally high price" and label them as "high-risk transaction". The labeled text stream can be desensitized using a data anonymization algorithm to obtain a desensitized text stream. For example, the k-anonymization algorithm can be used to replace the customer's ID number in the instruction with common characters. The desensitized text stream can be adapted to rules using a dynamic rule engine to obtain an initial text stream. For example, when the market fluctuates, the rule engine automatically activates the "restrict high-risk stock trading" rule to remove instructions involving related stocks. Based on the initial text stream, the initial feature chart can be collaboratively verified using a multimodal contrastive learning model to obtain an initial feature chart. For example, the transaction amount displayed in the chart is compared with the amount described in the text stream. If the chart shows a transaction amount of 10 million yuan, while the text stream describes it as 1 million yuan, the chart is corrected or marked as abnormal.

[0030] S2. Obtain the trading element library of the preset securities model, parse the trading feature factors corresponding to the initial feature chart, combine the trading element library and the trading feature factors, extract the spatiotemporal structure feature vector from the initial feature chart, analyze the structural description text corresponding to the spatiotemporal structure feature vector, and generate the chart parsing text corresponding to the initial feature chart.

[0031] This invention, by analyzing the trading feature factors corresponding to the initial feature chart, can transform the complex information presented by the chart into quantifiable and analyzable key factors, making the chart information more targeted and operable, and facilitating the discovery of hidden trading patterns and trends behind the chart. The trading element library refers to a collection of key information covering various aspects of securities trading stored in a pre-set securities model. For example, the trading element library records the opening price, closing price, highest price, and lowest price of a stock within a specific time period. The trading feature factors refer to key indicators extracted from the initial feature chart that reflect the core characteristics and changing trends of securities trading, such as price trend factors, trading volume fluctuation factors, and time cycle factors. For example, by statistically analyzing the increase or decrease in trading volume within a unit of time, the trading volume fluctuation factor can be obtained to analyze market trading activity. Optionally, the analysis of the trading feature factors corresponding to the initial feature chart can be implemented using feature extraction algorithms in machine learning, such as using principal component analysis (PCA) to reduce the dimensionality of multi-dimensional data in the chart and extract the most representative trading feature factors.

[0032] This invention, by combining the transaction element library and the transaction feature factors, extracts a spatiotemporal structure feature vector from the initial feature chart. This enables the integration and quantification of transaction information in both time and space dimensions, constructing a structured data representation that reflects the dynamic changes in securities transactions. This facilitates subsequent analysis and processing. The spatiotemporal structure feature vector is feature data in the initial feature chart that integrates the time-series information and spatial dimension information of securities transactions, represented in vector form, reflecting the evolutionary pattern of securities transactions in the spatiotemporal dimension.

[0033] As an embodiment of the present invention, the step of extracting a spatiotemporal structure feature vector from the initial feature chart by combining the transaction element library and the transaction feature factors includes: A knowledge graph is constructed from the transaction element database to obtain the transaction element graph; The transaction feature factors are encoded to obtain a feature factor vector; Identify the transaction element nodes in the transaction element graph and analyze the correlation strength between the feature factor vector and the transaction element nodes; Based on the correlation strength, the associated element nodes corresponding to the transaction feature factors are selected from the transaction element nodes; Based on the associated element nodes, extract the graph topology information corresponding to the transaction feature factors from the transaction element graph; Extract the time series information of the graph from the initial feature graph, and perform feature aggregation processing on the time series information of the graph and the topological structure information of the graph to obtain the spatiotemporal structure feature vector.

[0034] The transaction element graph is a structured graphical representation of the transaction element database formed using knowledge graph construction technology. For example, it presents elements such as price, volume, and trading time in stock trading as nodes and edges. The feature factor vector is a numerical vector obtained by encoding the transaction feature factors; for example, encoding stock price fluctuation trends as a two-dimensional vector [0.8, 0.2]. The transaction element node is the basic building block in the transaction element graph, corresponding to the transaction element, such as the "opening price" node in the graph. The correlation strength is a measure of the correlation between the feature factor vector and the transaction element node, such as through cosine similarity. The correlation between the stock price volatility factor vector and the "trading volume" node is calculated. The correlated element node is selected from the trading element nodes that has a strong correlation with the trading characteristic factor, such as the "turnover rate" and "price-to-earnings ratio" nodes that have a significant impact on stock price volatility. Based on the correlated element node, the local graph topology information corresponding to the trading characteristic factor is extracted from the trading element graph, that is, the connection relationship of these nodes in the graph, such as the edge connection information between the "turnover rate" and "trading volume" nodes. The chart time series information is the sequence data that changes over time in the initial feature chart, such as the daily opening price and closing price time series in the K-line chart.

[0035] Furthermore, a knowledge graph can be constructed from the transaction element database using graph database modeling techniques to obtain a transaction element graph; the transaction feature factors can be encoded using an autoencoder network to obtain feature factor vectors, such as using a variational autoencoder (VAE) to map the transaction feature factors to a low-dimensional latent space; transaction element nodes in the transaction element graph can be identified using entity recognition algorithms, such as the BERT-based named entity recognition (NER) algorithm; the association strength between the feature factor vectors and the transaction element nodes can be analyzed using an attention mechanism with weighted calculations, such as a multi-head attention mechanism that calculates and fuses the association weights of features in different dimensions; based on the association strength... The threshold is set to filter the associated element nodes corresponding to the transaction feature factors from the transaction element nodes. Based on the associated element nodes, the graph topology information corresponding to the transaction feature factors can be extracted from the transaction element graph using a subgraph extraction algorithm, such as a subgraph extraction algorithm based on depth-first search (DFS). The graph temporal information in the initial feature graph can be extracted using a time series decomposition method. The spatiotemporal graph convolutional network can be used to perform feature aggregation processing on the graph temporal information and the graph topology information to obtain a spatiotemporal structure feature vector, such as using ST-GCN (Spatiotemporal Graph Convolutional Network) to learn features in both the time dimension and the graph structure dimension simultaneously.

[0036] This invention generates chart parsing text corresponding to the initial feature chart by analyzing the structural description text corresponding to the spatiotemporal structural feature vector. This can transform the abstract feature vector into an understandable natural language description, allowing for a more intuitive understanding of the trading information conveyed by the chart. The structural description text refers to the text generated by natural language processing technology based on the spatiotemporal structural feature vector, and the chart parsing text is the text content that explains the features and trends of the securities trading chart.

[0037] S3. Perform text segmentation processing on the initial text stream to obtain initial segmented text, calculate the segmentation entropy loss corresponding to the initial segmented text, and perform segmentation iterative optimization on the initial text stream based on the segmentation entropy loss to obtain the target segmented text.

[0038] This invention performs text segmentation on the initial text stream to obtain initial segmented text, which can quickly divide the original text into preliminary segments, providing a foundation for subsequent fine-grained processing. By calculating the segmentation entropy loss corresponding to the initial segmented text, the quality of the initial segmented text can be understood, thus facilitating subsequent iterative optimization processing.

[0039] Furthermore, the initial text stream can be segmented using an instruction segmenter to obtain the initial segmented text. The instruction segmenter is a pre-trained model or tool used to segment a continuous text stream according to semantics, syntax, or specific rules. It has multiple built-in segmentation strategies and algorithms and can flexibly adjust the segmentation method according to the characteristics of the text. For example, for a news text stream, the instruction segmenter can segment it into independent sentence or paragraph units based on information such as sentence structure, paragraph division, and keywords. The instruction segmenter can be constructed using a sequence-to-sequence model based on the Transformer architecture.

[0040] As an embodiment of the present invention, calculating the segmentation entropy loss corresponding to the initial segmented text includes: Perform syntactic-semantic mapping on the initial segmented text to obtain a set of mapping relationships; Based on the set of mapping relationships, analyze the syntactic and semantic co-features corresponding to the initial segmented text; Based on the syntactic-semantic co-features, the initial segmented text is reasonably classified to obtain reasonable and unreasonable segmented fragments. The number of segments corresponding to the reasonable segmentation segments and the unreasonable segmentation segments is counted to obtain the number of reasonable segments and the number of unreasonable segments. The segmentation entropy loss corresponding to the initial segmented text is calculated by combining the number of reasonable segments, the syntactic-semantic co-features, and the number of unreasonable segments.

[0041] The mapping relationship set is the set of structure-meaning correspondences obtained after the initial segmented text undergoes syntactic-semantic mapping. For example, after mapping, the syntactic structure of "beautiful flowers" is an attributive plus a headword, and the semantic role is the correspondence between modifiers and entities. The syntactic-semantic co-feature is the matching and association feature between the syntactic structure and semantic role of the initial segmented text. The reasonable segment and the unreasonable segment are the segmentation units that conform to or do not conform to the syntactic-semantic consistency standard after the initial segmented text is reasonably classified based on the syntactic-semantic co-feature. For example, "reading books" has syntactic and semantic consistency and belongs to the reasonable segment, while "reading the sky" has semantic contradiction and is an unreasonable segment.

[0042] Furthermore, NLP toolkits (such as spaCy and AllenNLP) can be used to perform syntactic-semantic mapping on the initial segmented text to obtain a set of mapping relationships. For example, spaCy can be used for dependency parsing, and AllenNLP can be used for semantic role labeling to map "cat chasing butterfly" to "syntactic structure: subject (cat) + predicate (chasing) + object (butterfly)" and "semantic roles: agent (cat) + action (chasing) + patient (butterfly)". Based on the set of mapping relationships, rule matching and feature extraction algorithms are used to analyze the syntactic-semantic co-features of the initial segmented text. For example, it is determined whether the arguments in the syntactic structure meet the selection constraints of semantic roles. If "cat drinks stone" appears, the semantic requirement of "drinking" is liquid, which does not match "stone", resulting in low co-features. Based on the syntactic-semantic co-features, a support vector machine (SVM) classification model is used to classify the initial segmented text reasonably to obtain reasonable and unreasonable segmented segments.

[0043] Furthermore, as an optional embodiment of the present invention, the segmentation entropy loss corresponding to the initial segmented text is calculated by combining the number of reasonable segments, the syntactic-semantic co-features, and the number of unreasonable segments, including: Based on the aforementioned syntactic-semantic co-features, calculate the reasonable segmentation rate for each text segment in the initial segmented text; Based on the number of reasonable segments and the number of unreasonable segments, calculate the imbalance ratio corresponding to the initial segmented text; Based on the imbalance ratio, set a reasonable weight factor corresponding to the initial segmented text; Combining the reasonable weighting factor, the reasonable segmentation rate, the reasonable number of segments, and the unreasonable number of segments, the segmentation entropy loss corresponding to the initial segmented text is calculated using the following formula: ; Where A represents the segmentation entropy loss corresponding to the initial segmented text. This indicates that the initially segmented text is a valid segment. This indicates that the initially segmented text is an irrational segment, where x represents the number of irrational segments and y represents the number of irrational segments. Indicates the reasonable weighting factor. This represents the reasonable segmentation rate corresponding to the i-th text segment in the initial segmented text, where i represents the sequence number of the initial segmented text.

[0044] The imbalance ratio is a quantification of the degree of imbalance between the number of reasonable and unreasonable segmented segments corresponding to the initial segmented text. It is calculated by the maximum and minimum ratio of the number of the two types of segments to reflect the imbalance state of the sample distribution. The reasonable weight factor is an adjustment parameter dynamically set for the initial segmented text to balance the difference in the number of reasonable and unreasonable segmented segments and to increase the minority class loss weight. Furthermore, based on the imbalance ratio, the basic weight can be mapped using an exponential function to set the reasonable weight factor corresponding to the initial segmented text.

[0045] Furthermore, as an optional embodiment of the present invention, the step of calculating the reasonable segmentation rate corresponding to each text segment in the initial segmented text based on the syntactic-semantic co-features includes: Heterogeneous graph construction is performed on the syntactic-semantic co-functional features to obtain a syntactic-semantic joint graph structure; The syntax-semantic joint graph structure is subjected to convolutional fusion processing to obtain a collaborative feature matrix; The cooperative feature matrix is ​​subjected to adversarial constraint processing to obtain the constraint feature matrix; The constraint feature matrix is ​​subjected to spatiotemporal fusion processing to obtain fused constraint features; Analyze the feature granularity of the fusion constraint features and calculate the attention coefficient corresponding to the feature granularity; By combining the attention coefficient and the fusion constraint features, the reasonable segmentation rate corresponding to each text segment in the initial segmented text is calculated.

[0046] Wherein, the syntax-semantic joint graph structure is a structured graph model representation of the syntax-semantic collaborative features; the collaborative feature matrix is ​​the graph convolution feature aggregation result of the syntax-semantic joint graph structure; the constraint feature matrix is ​​a robust optimized feature representation of the collaborative feature matrix after adversarial constraint processing; the fusion constraint feature is the spatiotemporal dimension information fusion output of the constraint feature matrix; the feature granularity is the multi-scale feature abstraction level of the fusion constraint feature; and the attention coefficient represents the dynamic importance weighting parameter corresponding to the feature granularity.

[0047] Furthermore, the syntactic-semantic co-feature extraction and heterogeneous graph construction module can be used to perform heterogeneous graph construction processing on the syntactic-semantic co-features to obtain a syntactic-semantic joint graph structure; the syntactic-semantic joint graph structure can be processed by convolutional fusion using a graph convolutional neural network (GCN) to obtain a co-feature matrix; the co-feature matrix can be subjected to adversarial constraint processing using an adversarial training module (such as an adversarial loss function or discriminator) to obtain a constraint feature matrix; the constraint feature matrix can be subjected to spatiotemporal fusion processing using a spatiotemporal fusion network (such as a temporal-spatial feature fusion model combining LSTM / GRU and CNN) to obtain fused constraint features; the feature granularity of the fused constraint features can be analyzed using a multi-scale feature analysis module, such as a pyramid pooling module; the attention coefficients corresponding to the feature granularity can be calculated using an attention mechanism, such as calculating the similarity between feature granularity vectors using scaled dot-product attention, and then generating attention coefficients through softmax activation; finally, the attention coefficients and the fused constraint features can be combined to calculate the reasonable segmentation rate corresponding to each text segment in the initial segmented text using a fully connected function.

[0048] This invention improves the accuracy of semantic boundary recognition in text segmentation and reduces redundant segmentation errors by iteratively optimizing the initial text stream based on the segmentation entropy loss. The target segmented text is the optimal segmentation result obtained by iteratively adjusting the text segment boundaries based on the segmentation entropy loss to minimize the entropy value of the segmentation result (i.e., maximizing semantic consistency within segmentation units and maximizing semantic differences between units). Further, to intuitively demonstrate the iterative optimization process of the initial text stream segmentation, please refer to the following... Figure 2 This diagram illustrates the segmentation iterative optimization process in the intelligent parsing method for securities trading instructions based on a large model provided by this invention. This process is based on segmentation entropy loss. On one hand, it performs backpropagation to update the parameters of the instruction segmenter, and on the other hand, it adjusts the segmentation strategy threshold while simultaneously performing cross-modal loss fusion. Then, it uses the instruction segmenter after the above operations to re-segment the text, finally generating optimized segmented text. Through multi-path optimization (parameter update, threshold adjustment, and cross-modal fusion), iterative improvement of text segmentation is achieved.

[0049] S4. Combining the chart parsing text and the target segmentation text, generate the instruction target text corresponding to the securities trading instruction, extract the instruction tags corresponding to the instruction target text, calculate the structured likelihood between the instruction tags, and determine the associated aggregation tags in the instruction tags based on the structured likelihood.

[0050] This invention generates the target text of the securities trading instruction by combining the parsed text of the chart and the target segmented text, thereby obtaining the core content of the securities trading instruction. By extracting the instruction tags corresponding to the instruction target text, the key semantic identifiers of the instruction can be obtained, providing a basis for the subsequent calculation and processing of the structured likelihood between the instruction tags.

[0051] The target text of the instruction is the instruction text of the securities trading instruction obtained by fusing the chart parsing text and the target segmentation text. The instruction tag is the key semantic identifier corresponding to the target text of the instruction. Furthermore, by combining the chart parsing text and the target segmentation text, the target text of the instruction corresponding to the securities trading instruction can be generated through semantic fusion and structural adaptation algorithms. If the chart parsing text is "A stock's K-line chart shows a golden cross of moving averages, and the trading volume increases by 20% month-on-month," and the target segmentation text is "Place a market order to buy 1,000 shares of this stock," the target text of the instruction is generated by integrating chart trends, trading actions, and other information through semantic fusion and structural adaptation algorithms: "Based on the chart signal of a golden cross of stock moving averages and a 20% increase in trading volume, place a market order to buy 1,000 shares [stock code]." The instruction tag corresponding to the target text of the instruction can be extracted using a tag extraction tool compiled in JAVA.

[0052] This invention calculates the structured likelihood between the instruction tags to understand the degree of correlation between them, providing a quantitative basis for determining the associated aggregation tags in the subsequent instruction tags. The structured likelihood represents the probability of association and the degree of structured rationality of the instruction tags under the securities trading business logic.

[0053] As an embodiment of the present invention, calculating the structured likelihood between the instruction tags includes: Calculate the tag flow probability between the instruction tags and determine the probability mapping value corresponding to the instruction tag; Combining the tag flow probability and the probability mapping value, the structured likelihood between the instruction tags is calculated using the following formula: ; Where D represents the structured likelihood between instruction labels. This represents the tag transition probability when the state of the e-th instruction tag transitions to the state of the (e+1)-th instruction tag. This represents the e-th instruction label. This represents the (e+1)th instruction label. R represents the probability mapping value of the e-th instruction tag, where e represents the sequence number of the instruction tag and R represents the number of instruction tags.

[0054] Wherein, the tag transition probability represents the transition probability between the instruction tags (i.e., the conditional probability of transitioning from one tag to another); the probability mapping value represents the probability value corresponding to the instruction tag (such as the single tag edge probability). Furthermore, the tag transition probability between the instruction tags can be calculated using a Bayesian graph model, such as a Hidden Markov Model. For example, by using a Hidden Markov Model to count the number of times one tag 1 transitions to another tag 2, and then dividing by the total number of times tag 1 is used, the tag transition probability can be obtained; the occurrence frequency of the corresponding tag can be calculated based on the occurrence frequency, which is the corresponding probability mapping value.

[0055] This invention determines the associated aggregation tags in the instruction tags based on the structured likelihood, thereby filtering out tags with tight logic and conforming to securities trading business rules, thus improving the efficiency and accuracy of subsequent instruction parsing. The associated aggregation tags are combinations of tags with business logic aggregation selected from the instruction tags through structured likelihood. Furthermore, when the structured likelihood is greater than a preset likelihood, the associated aggregation tags in the instruction tags are determined. The preset likelihood can be 0.8, or it can be set according to the actual application scenario.

[0056] S5. Based on the associated aggregation tags, analyze the trading intent corresponding to the securities trading instruction, and combine the trading intent and the structured likelihood to generate the parsing result corresponding to the securities trading instruction.

[0057] This invention analyzes the trading intent corresponding to the securities trading instruction based on the associated aggregation tags, thereby obtaining the detailed trading purpose of the securities trading instruction, so as to generate accurate instruction parsing results in the future. The trading intent is the core trading purpose corresponding to the securities trading instruction.

[0058] As an embodiment of the present invention, the step of analyzing the trading intent corresponding to the securities trading instruction based on the associated aggregation tag includes: Analyze the tag semantics corresponding to the associated aggregated tags, and construct a semantic association network corresponding to the associated aggregated tags based on the tag semantics; The semantic association network is subjected to hierarchical clustering to obtain a hierarchical classification network; Extract the intent-driven features from the hierarchical classification network, perform intent matching on the core features of the network, and obtain the initial interaction intent; Calculate the coupling weight coefficients of the intent-dominant feature in the hierarchical classification network; Based on the coupling weight coefficient, the initial interaction intent is filtered to obtain the trading intent corresponding to the securities trading instruction.

[0059] Wherein, the semantic association network is the semantic association structured network corresponding to the associated aggregation label; the hierarchical classification network is the hierarchical structure network after hierarchical clustering of the semantic association network; the intent-driven feature is the core semantic feature in the hierarchical classification network; the initial interaction intent is the preliminary intent judgment obtained by intent matching of the network core features; and the coupling weight coefficient is the semantic and structural association weight coefficient of the intent-driven feature in the hierarchical classification network.

[0060] Furthermore, semantic analysis can be used to analyze the semantics of the associated aggregated tags; based on the tag semantics, a semantic association network corresponding to the associated aggregated tags can be constructed using a graph construction tool; a hierarchical clustering algorithm can be used to perform hierarchical clustering on the semantic association network to obtain a hierarchical classification network; a feature extraction model can be used to extract the intent-dominant features in the hierarchical classification network, such as a graph neural network (GAT) based on an attention mechanism, which identifies core semantic features by calculating the attention weights of each node; an intent matching algorithm can be used to perform intent matching on the core features of the network to obtain the initial interaction intent, such as constructing a preset intent template library and calculating semantic similarity to map network features to standard intent categories such as "portfolio adjustment" and "risk hedging"; when the coupling weight coefficient is lower than a preset coefficient, the initial interaction intent is filtered to obtain the trading intent corresponding to the securities trading instruction.

[0061] Furthermore, as an optional embodiment of the present invention, calculating the coupling weight coefficients of the intent-dominant feature in the hierarchical classification network includes: Calculate the network centrality of the intent-dominant feature in the hierarchical classification network; The depth of the feature hierarchy of the intent-driven features in the hierarchical classification network is statistically analyzed. Combining the network centrality and the feature level depth, the coupling weight coefficient of the intent-driven feature in the hierarchical classification network is calculated using the following formula: ; Where L represents the coupling weight coefficient of the intention-dominant feature in the hierarchical classification network. Indicating the dominant feature of intent The characteristic of network centrality, Indicating the dominant feature of intent Business relevance of features Indicating the dominant feature of intent Feature hierarchy depth This represents the hierarchical attenuation coefficient.

[0062] Wherein, network centrality represents the core position or connection importance of the intent-dominant feature in the hierarchical classification network; feature level depth represents the topological level position of the intent-dominant feature in the hierarchical classification network or the path length from the root node; business relevance represents the strength of the correlation between the intent-dominant feature and specific business needs; and the level decay coefficient represents a quantification factor that indicates the decrease in feature influence as the network level increases.

[0063] Furthermore, the network centrality of the intent-dominant feature in the hierarchical classification network can be calculated using betweenness centrality or degree centrality algorithms; the feature level depth of the intent-dominant feature in the hierarchical classification network can be calculated by counting the number of path edges from feature nodes to the root node or by the level index; the business correlation degree can be obtained by analyzing business process dependencies, regulatory rule associations, or expert experience scores; and the level decay coefficient can be obtained by preset experience values ​​or by training and optimization based on historical data.

[0064] This invention generates parsing results for securities trading instructions by combining the trading intent and the structured likelihood, thereby obtaining execution strategy suggestions for the securities trading instructions and improving the risk control accuracy of securities trading. Furthermore, by combining the trading intent and the structured likelihood, the parsing results for the securities trading instructions are generated by first clarifying the core purpose of the instruction (such as risk hedging or value investing) based on the trading intent, then checking the completeness of instruction elements (such as mandatory fields like price and quantity) and format compliance based on the structured likelihood; finally, combining the intent determination with the compliance check results, an parsing result is generated that includes a description of the trading purpose, compliance conclusions, and operational suggestions, for example, "The instruction intent is risk aversion and stop-loss, the structured likelihood is 0.92, and it is recommended to immediately execute a sell operation."

[0065] Compared to the problems described in the background technology, this invention, by performing multimodal information decomposition processing on the securities trading instructions, can break down complex trading instructions into information of different dimensions, providing a data foundation for subsequent accurate compliance screening and effectively improving the accuracy and efficiency of securities trading instruction compliance detection. This invention, by analyzing the trading feature factors corresponding to the initial feature chart, can transform the complex information presented by the chart into quantifiable and analyzable key factors, making the chart information more targeted and operable, facilitating the discovery of hidden trading patterns and trends behind the chart. Furthermore, this invention, by performing text segmentation processing on the initial text stream to obtain initial segmented text, can quickly and initially divide the original text, providing a foundation for subsequent refined processing and calculating the... The segmentation entropy loss corresponding to the initial segmented text can reveal the quality of the initial segmented text, facilitating subsequent iterative optimization. Furthermore, this invention generates the target text for the securities trading instruction by combining the chart parsing text and the target segmented text, thereby obtaining the core content of the securities trading instruction. By extracting the instruction tags corresponding to the target text, key semantic identifiers of the instruction can be obtained, providing a basis for subsequent calculation of the structured likelihood between the instruction tags. Finally, this invention analyzes the trading intent corresponding to the securities trading instruction based on the associated aggregation tags, thereby obtaining the detailed trading purpose of the securities trading instruction, facilitating the generation of accurate instruction parsing results. Therefore, the intelligent securities trading instruction parsing method and system based on a large model provided by this invention can improve the accuracy of securities trading instruction parsing.

[0066] Example 2: like Figure 3 The diagram shown is a functional block diagram of the intelligent securities trading instruction parsing system based on a large model according to the present invention.

[0067] The large-model-based intelligent securities trading instruction parsing system 300 described in this invention can be installed in an electronic device. Depending on the functions implemented, the large-model-based intelligent securities trading instruction parsing system may include an instruction initialization processing module 301, a chart text analysis module 302, a text segmentation processing module 303, an instruction tag analysis module 304, and an instruction parsing module 305. The module described in this invention can also be referred to as a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and are stored in the memory of the electronic device.

[0068] In this embodiment of the invention, the functions of each module / unit are as follows: The instruction initial processing module 301 is used to receive the securities trading instruction to be processed, perform multimodal information splitting processing on the securities trading instruction to obtain the instruction text stream and instruction feature chart, and perform compliance screening on the instruction text stream and instruction feature chart to obtain the initial text stream and initial feature chart. The chart text analysis module 302 is used to obtain the trading element library of the preset securities model, parse the trading feature factors corresponding to the initial feature chart, combine the trading element library and the trading feature factors, extract the spatiotemporal structure feature vector from the initial feature chart, analyze the structural description text corresponding to the spatiotemporal structure feature vector, and generate the chart parsing text corresponding to the initial feature chart. The text segmentation processing module 303 is used to perform text segmentation processing on the initial text stream to obtain initial segmented text, calculate the segmentation entropy loss corresponding to the initial segmented text, and perform segmentation iterative optimization on the initial text stream based on the segmentation entropy loss to obtain the target segmented text; The instruction tag analysis module 304 is used to combine the chart parsing text and the target segmentation text to generate the instruction target text corresponding to the securities trading instruction, extract the instruction tags corresponding to the instruction target text, calculate the structured likelihood between the instruction tags, and determine the associated aggregation tags in the instruction tags based on the structured likelihood. The instruction parsing module 305 is used to analyze the trading intent corresponding to the securities trading instruction based on the associated aggregation tag, and generate the parsing result corresponding to the securities trading instruction by combining the trading intent and the structured likelihood.

[0069] In detail, the modules in the intelligent securities trading instruction parsing system 200 based on a large model described in this embodiment of the invention employ the same methods as described above. Figure 1 The method uses the same technical means as the intelligent parsing method for securities trading orders based on large models described in the article, and can produce the same technical effect, so it will not be repeated here.

[0070] 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.

[0071] 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 intelligent analysis of securities transaction instructions based on a large model, characterized in that, The method comprises: receiving a securities transaction instruction to be processed, performing multi-modal information splitting processing on the securities transaction instruction to obtain an instruction text stream and an instruction feature graph, performing compliance screening on the instruction text stream and the instruction feature graph to obtain an initial text stream and an initial feature graph; obtaining a transaction element library of a preset securities big model, analyzing a transaction characteristic factor corresponding to the initial feature graph, combining the transaction element library and the transaction characteristic factor, extracting a space-time structure feature vector from the initial feature graph, analyzing a structure description text corresponding to the space-time structure feature vector, and generating a graph analysis text corresponding to the initial feature graph; performing text segmentation processing on the initial text stream to obtain an initial segmented text, calculating a segmentation entropy loss corresponding to the initial segmented text, performing segmentation iteration optimization on the initial text stream based on the segmentation entropy loss, and obtaining a target segmented text; combining the graph analysis text and the target segmented text to generate an instruction target text corresponding to the securities transaction instruction, extracting an instruction label corresponding to the instruction target text, calculating a structured likelihood between the instruction labels, and determining an associated aggregation label in the instruction labels based on the structured likelihood; based on the associated aggregation label, analyzing a transaction intention corresponding to the securities transaction instruction, combining the transaction intention and the structured likelihood, and generating an analysis result corresponding to the securities transaction instruction.

2. The large model-based security transaction instruction intelligent parsing method of claim 1, wherein, The multi-modal information splitting processing on the securities transaction instruction to obtain an instruction text stream and an instruction feature graph comprises: performing voiceprint feature separation on the securities transaction instruction to obtain a voiceprint feature vector and a noise-reduced audio stream; performing speech text conversion on the noise-reduced audio stream to obtain a structured instruction text; performing text filtering processing on the structured instruction text based on a preset business logic rule set to obtain an instruction text stream; extracting a visual association feature corresponding to the instruction text stream, combining the voiceprint feature vector and the visual association feature, and analyzing an instruction feature graph corresponding to the securities transaction instruction. 3.The large model-based security transaction instruction intelligent parsing method of claim 1, wherein, The compliance screening on the instruction text stream and the instruction feature graph to obtain an initial text stream and an initial feature graph comprises: performing risk labeling processing on the instruction text stream to obtain a labeled text stream; performing desensitization processing on the labeled text stream to obtain a desensitized text stream; performing rule adaptation processing on the desensitized text stream to obtain an initial text stream; based on the initial text stream, performing collaborative verification processing on the initial feature graph to obtain an initial feature graph. 4.The large model-based security transaction instruction intelligent parsing method of claim 1, wherein, The combination of the transaction element library and the transaction characteristic factor to extract a space-time structure feature vector from the initial feature graph comprises: performing knowledge graph construction on the transaction element library to obtain a transaction element graph; performing encoding processing on the transaction characteristic factor to obtain a feature factor vector; identifying a transaction element node in the transaction element graph, analyzing the association strength between the feature factor vector and the transaction element node; Filtering out the associated element node corresponding to the transaction characteristic factor from the transaction element node based on the association strength; Extracting the graph topology structure information corresponding to the transaction characteristic factor from the transaction element graph according to the associated element node; Extracting graph timing information in the initial feature graph, and performing feature aggregation processing on the graph timing information and the graph topology structure information to obtain a spatiotemporal structure feature vector. 5.The large model-based security transaction instruction intelligent parsing method of claim 1, wherein, The calculation of the segmentation entropy loss corresponding to the initial segmentation text comprises: Performing syntax-semantics mapping on the initial segmentation text to obtain a mapping relationship set; Based on the mapping relationship set, analyzing the syntax-semantics collaborative feature corresponding to the initial segmentation text; Based on the syntax-semantics collaborative feature, classifying the initial segmentation text reasonably to obtain reasonable segmentation segments and unreasonable segmentation segments; Counting the number of segments corresponding to the reasonable segmentation segments and the unreasonable segmentation segments to obtain a reasonable segment number and an unreasonable segment number; Combining the reasonable segment number, the syntax-semantics collaborative feature, and the unreasonable segment number, calculating the segmentation entropy loss corresponding to the initial segmentation text. 6.The method of claim 5, wherein, Combining the reasonable segment number, the syntax-semantics collaborative feature, and the unreasonable segment number, calculating the segmentation entropy loss corresponding to the initial segmentation text comprises: Based on the syntax-semantics collaborative feature, calculating a reasonable segmentation rate corresponding to each text segment in the initial segmentation text; Combining the reasonable segment number and the unreasonable segment number, calculating an imbalance ratio of the initial segmentation text; Based on the imbalance ratio, setting a reasonable weight factor corresponding to the initial segmentation text; Combining the reasonable weight factor, the reasonable segmentation rate, the reasonable segment number, and the unreasonable segment number, calculating the segmentation entropy loss corresponding to the initial segmentation text by the following formula: ； wherein A represents a segmentation entropy loss corresponding to the initial segmented text, represents that the initial segmented text is a reasonable segmentation fragment, represents that the initial segmented text is an unreasonable segmentation fragment, x represents a reasonable fragment number, and y represents an unreasonable fragment number, represents a reasonable weight factor, represents a reasonable segmentation rate corresponding to the i-th text fragment in the initial segmented text, and i represents a sequence number corresponding to the initial segmented text.

7. The large model-based security transaction instruction intelligent parsing method of claim 6, wherein, The calculation of the reasonable segmentation rate corresponding to each text segment in the initial segmentation text based on the syntax-semantics collaborative feature comprises: Performing heterogeneous graph construction processing on the syntax-semantics collaborative feature to obtain a syntax-semantics joint graph structure; Performing convolution fusion processing on the syntax-semantics joint graph structure to obtain a collaborative feature matrix; Performing adversarial constraint processing on the collaborative feature matrix to obtain a constraint feature matrix; Performing spatiotemporal fusion processing on the constraint feature matrix to obtain a fusion constraint feature; Analyzing the feature granularity of the fusion constraint feature, and calculating an attention coefficient corresponding to the feature granularity; Combining the attention coefficient and the fusion constraint feature, calculating the reasonable segmentation rate corresponding to each text segment in the initial segmentation text. 8.The large model-based security transaction instruction intelligent parsing method of claim 1, wherein, The calculation of the structured likelihood between the instruction labels comprises: Calculating a label flow probability between the instruction labels, and determining a probability mapping value corresponding to the instruction labels; Combining the label flow probability and the probability mapping value, calculating the structured likelihood between the instruction labels by the following formula: ； wherein D denotes a structured likelihood between instruction tags, denotes a tag flow probability of a state transition from an e-th instruction tag to an e+1-th instruction tag, denotes an e-th instruction tag, denotes an e+1-th instruction tag, denotes a probability mapping value of an e-th instruction tag, e denotes a serial number of an instruction tag, and R denotes a number of instruction tags. 9.The large model-based security transaction instruction intelligent parsing method of claim 1, wherein, The analysis of the transaction intention corresponding to the securities transaction instruction based on the associated aggregated label comprises: Analyze the label semantics corresponding to the associated aggregated label, and construct a semantic association network corresponding to the associated aggregated label based on the label semantics; Perform hierarchical clustering processing on the semantic association network to obtain a hierarchical classification network; Extract the intention dominant features in the hierarchical classification network, and perform intention matching on the network core features to obtain an initial interaction intention; Calculate the coupling weight coefficient of the intention dominant features in the hierarchical classification network; Based on the coupling weight coefficient, the initial interaction intention is screened to obtain the transaction intention corresponding to the securities transaction instruction.

10. A large model-based security transaction instruction intelligent analysis system, characterized in that, The system comprises: An instruction initial processing module is configured to receive a securities transaction instruction to be processed, perform multi-modal information splitting processing on the securities transaction instruction to obtain an instruction text stream and an instruction feature graph, and perform compliance screening on the instruction text stream and the instruction feature graph to obtain an initial text stream and an initial feature graph. A graph text analysis module is configured to obtain a transaction element library of a preset securities large model, analyze transaction characteristic factors corresponding to the initial feature graph, extract a spatio-temporal structure feature vector from the initial feature graph by combining the transaction element library and the transaction characteristic factors, analyze structure description text corresponding to the spatio-temporal structure feature vector, and generate graph analysis text corresponding to the initial feature graph. A text segmentation processing module is configured to perform text segmentation processing on the initial text stream to obtain an initial segmented text, calculate a segmentation entropy loss corresponding to the initial segmented text, perform segmentation iteration optimization on the initial text stream based on the segmentation entropy loss, and obtain a target segmented text. An instruction label analysis module is configured to generate instruction target text corresponding to the securities transaction instruction by combining the graph analysis text and the target segmented text, extract instruction labels corresponding to the instruction target text, calculate a structured likelihood between the instruction labels, and determine associated aggregated labels in the instruction labels based on the structured likelihood. An instruction analysis module is configured to analyze a transaction intention corresponding to the securities transaction instruction based on the associated aggregated labels, and generate an analysis result corresponding to the securities transaction instruction by combining the transaction intention and the structured likelihood.

Images