Financial case fund analysis report generation method and system based on semantic graph large model

Abstract

The invention provides a financial case fund analysis report generation method and system based on a semantic graph large model, and belongs to the technical field of natural language processing, and the method comprises the steps: S1, obtaining a to-be-analyzed text which is financial case data; s2, performing semantic analysis on a to-be-analyzed text according to the entity recognition model and the financial field entity dictionary, and constructing a semantic graph; s3, determining an abstract of the to-be-analyzed text; and S4, generating an analysis report according to the key entity of the semantic graph and the determined abstract. The method has the beneficial effects that by introducing a semantic graph technology and combining a large language model, the ability of understanding financial case texts is enhanced, and key information in fund circulation can be analyzed more accurately; through a report generation mechanism driven by a large model, the structure and content of the report can be dynamically adjusted based on a semantic graph analysis result, deep analysis is realized, and a specialized report better meeting financial case requirements is provided.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing technology, and in particular to a method and system for generating a financial case fund analysis report based on a semantic graph large model. Background Art

[0002] With the rapid development of financial markets and information technology, various new types of financial crimes are increasing, which has put forward higher professional requirements for multiple units including procuratorates, court systems, judicial administrative agencies and financial regulatory agencies.

[0003] When handling financial cases under traditional models, semantic analysis typically relies on natural language processing (NLP) technologies, such as rule-based systems or traditional machine learning models. These methods often struggle to capture deep semantic relationships when processing complex financial texts. Text descriptions in financial cases often contain a large number of professional terms, abbreviations, and complex sentence structures, which traditional NLP technologies struggle to process effectively. Existing technologies lack the ability to deeply understand contextual semantics and are unable to accurately identify key entities and their relationships in capital flows. Traditional machine learning models rely on large amounts of labeled data, but the high cost of labeling financial case data leads to insufficient model generalization capabilities.

[0004] Existing report generation modules are often based on templated report generation tools, resulting in fixed report content and a lack of flexibility and in-depth analytical capabilities. Templated report generation tools typically utilize predefined report structures and content frameworks, making it difficult to dynamically adjust content based on specific case requirements. This makes it difficult to meet user demands for personalized reports, and templated approaches may not provide sufficient flexibility to meet specific user requirements. Templated report generation tools typically only produce surface-level analytical results and lack in-depth insights. Templated report generation tools often employ a one-size-fits-all design and lack targeted optimization for the specific needs of the financial sector. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention provides a method for generating a financial case fund analysis report based on a semantic graph large model; on the other hand, it also provides a financial case fund analysis report generation system based on a semantic graph large model.

[0006] The technical problem solved by the present invention can be achieved by adopting the following technical solutions:

[0007] A method for generating a financial case fund analysis report based on a semantic graph large model, comprising:

[0008] Step S1, obtaining a text to be analyzed, wherein the text to be analyzed is financial case data;

[0009] Step S2: semantically parse the text to be analyzed based on an entity recognition model and a financial entity dictionary to construct a semantic graph; wherein the entity recognition model includes a bidirectional long short-term memory layer and a conditional random field layer, the bidirectional long short-term memory layer is used to identify key entities, and the conditional random field layer is used for sequence labeling;

[0010] Step S3, determining the summary of the text to be analyzed;

[0011] Step S4: generating an analysis report based on the key entities of the semantic graph and the determined summary.

[0012] Preferably, step S2 includes:

[0013] Step S21, extracting key entities from the text to be analyzed, and performing enhancement processing on the extracted key entities to obtain enhanced key entities;

[0014] Step S22, inputting the enhanced key entity into the entity recognition model to identify the key entity;

[0015] Step S23: performing transformation based on the identified key entities and key entity relationships to construct the semantic graph.

[0016] Preferably, the step S21 includes:

[0017] Step S211: for each extracted key entity, determine whether the key entity belongs to the financial domain entity dictionary; if the key entity belongs to the financial domain entity dictionary, execute step S22; otherwise, execute step S23;

[0018] Step S212: taking the sum of the vector representations of the key entity and the entities in the financial domain entity dictionary to which it belongs as the enhanced key entity;

[0019] Step S213: Using the key entity as the enhanced key entity.

[0020] Preferably, in the conditional random field layer, the conditional probability of the label sequence is:

[0021]

[0022] Where: P(Y|X) represents the probability of the labeled sequence Y under the condition of the input text sequence X of the text to be analyzed; Represents the label y t-1 Move to label y t The state transition probability of y is t, t represents the position index of label y in the annotation sequence Y, t = 1, ..., n, n represents the length of the annotation sequence Y; φ(yt ,h t ) represents the label y t Based on the hidden state sequence h t The local score of , φ represents the linear transformation function.

[0023] Preferably, step S3 includes:

[0024] Step S31, for each sentence in the text to be analyzed, performing an importance evaluation on the sentence to obtain a total importance score of each sentence, and screening candidate sentences for the abstract according to the total importance score;

[0025] Step S32, semantically encoding the candidate sentence based on a generative summary generation model to generate a probability of the candidate sentence being selected as a summary;

[0026] Step S33, performing weighted fusion on the total importance score of each sentence and the probability of the candidate sentence being selected as the summary to obtain the total score of all candidate sentences;

[0027] Step S34: Determine the candidate sentence with the highest total score as the summary of the text to be analyzed.

[0028] Preferably, the total importance score of the sentence is:

[0029]

[0030] Among them, Score TF-IDF (s) represents the total importance score of sentence s; tfidf(w,D) represents the term frequency-inverse document frequency index of keyword w in the text to be analyzed.

[0031] Preferably, the probability of the candidate sentence being selected as the summary is:

[0032] P(s∈summary)=σ(W e ·h s +b e )

[0033] Among them, P(s∈summary) represents the probability that sentence s is selected as the summary; σ represents the sigmoid activation function; W e represents weight; h s represents the encoding representation of sentence s; b e represents the bias term.

[0034] Preferably, in the generative summary generation model, the generation process of the summary sequence is:

[0035]

[0036] Among them, P(y|s) represents the probability of the summary sequence y under the condition of sentence s; P(y t |y <t ,h s ) represents the generation probability, h s represents the encoded representation of sentence s; t represents the position index in the summary sequence y, t = 1, ..., T, T represents the length of the summary sequence y.

[0037] Preferably, in step S4, the key entities of the semantic graph and the determined summary are processed according to a multi-layer self-attention codec to generate the analysis report.

[0038] On the other hand, a financial case fund analysis report generation system based on a semantic graph large model is also provided, which is used to implement the above-mentioned financial case fund analysis report generation method based on a semantic graph large model, including:

[0039] A data acquisition module is used to acquire the text to be analyzed, wherein the text to be analyzed is financial case data;

[0040] A semantic graph construction module, connected to the data acquisition module, is used to perform semantic parsing on the text to be analyzed based on an entity recognition model and a financial entity dictionary to construct a semantic graph; wherein the entity recognition model includes a bidirectional long short-term memory layer and a conditional random field layer, the bidirectional long short-term memory layer is used to identify key entities, and the conditional random field layer is used for sequence labeling;

[0041] A summary generation module, connected to the semantic graph construction module, for determining a summary of the text to be analyzed;

[0042] A report generation module is connected to the semantic graph construction module and the summary generation module respectively, and is used to generate an analysis report according to the key entities of the semantic graph and the determined summary.

[0043] The advantages or beneficial effects of the technical solution of the present invention are:

[0044] This invention introduces semantic graph technology and combines it with a large language model to enhance the ability to understand financial case texts, and can more accurately parse key information in capital flows. Through a report generation mechanism driven by a large model, the report structure and content can be dynamically adjusted based on the semantic graph analysis results, achieving in-depth analysis and providing professional reports that better meet the needs of financial cases. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a flow chart of a method for generating a financial case fund analysis report based on a semantic graph large model in a preferred embodiment of the present invention;

[0046] Figure 2 A schematic diagram of the process of constructing a semantic graph in a preferred embodiment of the present invention;

[0047] Figure 3 A schematic diagram of a key entity enhancement process in a preferred embodiment of the present invention;

[0048] Figure 4 A schematic diagram of a process flow for generating a summary in a preferred embodiment of the present invention;

[0049] Figure 5 This is a structural diagram of a financial case fund analysis report generation system based on a semantic graph large model in a preferred embodiment of the present invention;

[0050] Figure 6 This is a diagram of the overall architecture of the system in a preferred embodiment of the present invention;

[0051] Figure 7 A schematic diagram of the process of entity recognition in a preferred embodiment of the present invention;

[0052] Figure 8 Schematic diagram of the report generation process in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0053] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0054] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0055] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0056] In a preferred embodiment of the present invention, based on the above-mentioned problems existing in the prior art, a method and system for generating a financial case fund analysis report based on a semantic graph large model (SGLM) is provided. By deeply integrating semantic graph modeling, domain pre-training, dynamic reasoning and adaptive generation technology, the limitations of the existing financial case fund analysis report generation technology are solved, and the problems of insufficient semantic understanding of traditional NLP technology and lack of in-depth analysis in templated report generation are overcome, so as to improve the efficiency and accuracy of writing financial case analysis reports.

[0057] Among them, such as Figure 1As shown, the method includes:

[0058] Step S1, obtaining the text to be analyzed, which is financial case data;

[0059] Step S2: semantically parse the text to be analyzed based on the entity recognition model and the financial entity dictionary to construct a semantic graph; wherein the entity recognition model includes a bidirectional long short-term memory layer and a conditional random field layer, the bidirectional long short-term memory layer is used to identify key entities, and the conditional random field layer is used for sequence labeling;

[0060] Step S3, determining the summary of the text to be analyzed;

[0061] Step S4: Generate an analysis report based on the key entities of the semantic graph and the determined summary.

[0062] Specifically, existing natural language processing (NLP) methods struggle to effectively capture the deep semantic relationships within financial case text, leading to inaccurate identification of key entities and their relationships. This invention, by introducing semantic graph technology and combining it with a large model, enhances the understanding of financial case text, enabling more accurate analysis of key information in capital flows and improving semantic parsing capabilities.

[0063] Existing templated report generation tools lack dynamic adjustment capabilities and are unable to generate personalized analysis reports based on case requirements. The embodiments of the present invention, through a large-scale model-driven report generation mechanism, can dynamically adjust the report structure and content based on semantic graph analysis results, enabling in-depth analysis and providing specialized reports that better meet the needs of financial cases, thereby improving the flexibility and professionalism of report generation.

[0064] As a preferred embodiment, wherein Figure 2 As shown, step S2 includes:

[0065] Step S21, extracting key entities from the text to be analyzed, and enhancing the extracted key entities to obtain enhanced key entities;

[0066] Among them, such as Figure 3 As shown, step S21 includes:

[0067] Step S211: for each extracted key entity, determine whether the key entity belongs to the financial domain entity dictionary. If the key entity belongs to the financial domain entity dictionary, execute step S22; otherwise, execute step S23;

[0068] Step S212: taking the sum of the key entity and the vector representation of the key entity as the enhanced key entity;

[0069] Step S213: taking the key entity as the enhanced key entity.

[0070] Step S22: inputting the enhanced key entity into the entity recognition model to identify the key entity;

[0071] Step S23: transform the identified key entities and key entity relationships to construct a semantic graph.

[0072] Specifically, in this embodiment, semantic graph big model technology is used to enhance the ability to deeply analyze the text information of financial cases.

[0073] First, using financial domain knowledge, key entities are extracted, such as the parties involved, transaction objects, accounts, capital flows, etc., and the extracted key entities are enhanced.

[0074] For example, the entity enhancement process may adopt the following formula:

[0075]

[0076] in:

[0077] D represents the entity dictionary in the financial field;

[0078] I(x t ∈w) represents the indicator function; when the key entity x t When the entity w belongs to the entity dictionary in the financial field, I(x t ∈w) takes 1; when the key entity x t When the entity w does not belong to the entity dictionary in the financial field, I(x t ∈w) takes 0;

[0079] v w Represents the key entity x t The vector representation of entity w in the financial entity dictionary;

[0080] x′ t Represents the enhanced key entity.

[0081] The semantic graph is constructed by parsing case descriptions, transaction records and relevant laws and regulations, using an improved BiLSTM (bidirectional long short-term memory network)-CRF (conditional random field) model architecture.

[0082] The Bidirectional Long Short-Term Memory (BiLS™) network is a recurrent neural network with hidden states in both forward and backward directions, which can capture sentence structure and dependencies in text. It effectively models contextual information and, through bidirectional computation, better understands the relationship between the current label and the surrounding context.

[0083] In this embodiment, the enhanced key entity x′ t It is used as the input of the bidirectional long short-term memory network (BiLSTM) model to improve the recognition accuracy of key entities (such as parties involved in the case and capital flows).

[0084] For the input text sequence X=(x1,x2,…,x n ), n represents the length of the input text sequence, and the bidirectional LSTM model performs forward and backward encoding respectively.

[0085] Hidden state sequence output by the forward LSTM layer for:

[0086]

[0087] in, Represents the hidden state sequence output by the forward LSTM layer at time t-1;

[0088] Hidden state sequence output by the backward LSTM layer for:

[0089]

[0090] in, Represents the hidden state sequence output by the backward LSTM layer at time t+1;

[0091] Finally, the hidden state sequences in the two directions are concatenated to obtain:

[0092]

[0093] Among them, h t Represents the hidden state sequence output by the bidirectional LSTM model.

[0094] Conditional Random Fields (CRF) is a discriminative probabilistic undirected graphical model used for sequence labeling in natural language processing. It considers interactions between labels and uses a global inference algorithm to optimize the labeling results for the entire sequence. CRF models are able to capture contextual constraints and the transition patterns between labels in sequence labeling tasks.

[0095] The CRF layer defines a global score function. n ) is:

[0096]

[0097] Where: P(Y|X) represents the probability of the labeled sequence Y under the condition of the input text sequence X of the text to be analyzed;

[0098] Represents the label y t-1 Move to label y t The state transition probability of , t represents the position index of label y in the annotation sequence Y, t = 1, ..., n, n represents the length of the annotation sequence Y;

[0099] φ(y t ,h t ) represents the label y t Based on the hidden state sequence h t The local score of can be obtained by linear transformation; φ represents the linear transformation function.

[0100] In this embodiment, the hidden state sequence h is output by Bi-LSTM. t The local score φ(y t ,h t ) and the state transition probability output by the CRF layer Weighting can improve the accuracy of the model in identifying key entities and improve the robustness of the model.

[0101] By converting key entity relationships into semantic graphs and storing, calculating and reasoning in a graph structure, it can capture deep semantic relationships that are difficult to identify with traditional NLP methods.

[0102] Traditional NLP methods struggle to effectively process the complex semantics of financial cases. This embodiment of the present invention utilizes a large, pre-trained model for intelligent semantic parsing and reasoning, enhancing case text understanding. This model, fine-tuned using proprietary financial corpus, accurately identifies key factors in capital flows.

[0103] Unlike traditional templated report generation tools, the adaptive report generation mechanism of this invention uses a dynamic content generation strategy based on semantic graphs and large-scale model inference results. It can dynamically adjust the content and structure of the report according to the specific circumstances of the case. This not only improves the flexibility of the report but also better meets personalized needs.

[0104] In this embodiment, the report content is dynamically adjusted based on case characteristics, covering modules such as basic case information, transaction structure analysis, capital flow path, and risk assessment. Report modules are dynamically combined based on user needs, allowing for flexible adjustment of analysis depth and personalized report content.

[0105] As a preferred embodiment, wherein Figure 4 As shown, step S3 includes:

[0106] Step S31: For each sentence in the text to be analyzed, the importance of the sentence is evaluated to obtain a total importance score of each sentence, so as to screen candidate sentences for the summary according to the total importance score;

[0107] Among them, in the process of screening candidate sentences, first, the importance of the keywords in the sentence is evaluated according to the word frequency-inverse document frequency to obtain the importance evaluation score of each keyword; then, the importance evaluation scores of all keywords in the sentence are summarized to obtain the total importance score of each sentence.

[0108] Step S32: semantically encode the candidate sentence based on a generative summary generation model to generate a probability that the candidate sentence is selected as a summary; the generative summary generation model adopts the BERT model;

[0109] Step S33: performing a weighted fusion of the total importance score of each sentence and the probability of the candidate sentence being selected as the summary to obtain the total score of all candidate sentences;

[0110] Step S34: Determine the candidate sentence with the highest total score as the summary of the text to be analyzed.

[0111] Specifically, traditional natural language processing (NLP) techniques are used to extract summary information from basic case information. For example, techniques such as term frequency–inverse document frequency (TF-IDF) and Latent Dirichlet Allocation (LDA) are used to extract keywords and analyze topics, thereby generating brief descriptions. However, this approach is only suitable for extracting basic information and is difficult to meet the needs of generating specialized content.

[0112] In this embodiment, in the TF-IDF word frequency calculation, for a word w in a given document D, its word frequency tf(w,D) is defined as:

[0113]

[0114] In a corpus containing N documents, the inverse document frequency index idf(w) of word w is:

[0115]

[0116] Here, df(w) represents the number of documents containing word w.

[0117] The TF-IDF score tfidf(w,D) of word w in document D is:

[0118] tfidf(w,D)=tf(w,D)×idf(w)

[0119] Pair of sentences The total importance score of the sentence can be defined TF-IDF (s) is:

[0120]

[0121] The Score TF-IDF The (s) score reflects the importance of keywords in a sentence and is used to preliminarily screen candidate abstract sentences.

[0122] BERT is a Transformer-based pre-trained language model that can extract contextual representations of text. It learns rich linguistic knowledge through unsupervised pre-training on a large-scale corpus and can be trained for downstream tasks through fine-tuning tasks.

[0123] In the process of summarization, BERT, as a pre-trained language model, can perform high-quality semantic encoding on the input sentence. The BERT encoding function is denoted as E(·):

[0124] h s =E(s)

[0125] Among them, h s Represents the semantic encoding representation of sentence s.

[0126] In the extractive approach, a linear layer can be used to score each sentence:

[0127] P(s∈summary)=σ(W e ·h s +b e )

[0128] in:

[0129] W e and b e are all trainable parameters, W e represents the weight, b e represents the bias term;

[0130] σ represents the sigmoid activation function;

[0131] P(s∈summary) represents the probability that sentence s is selected as the summary.

[0132] Using generative summary, the summary sequence y=(y1,y2,…,y T ) can be described as:

[0133]

[0134] Among them, P(y|s) represents the probability of the summary sequence y under the condition of sentence s; P(y t |y <t ,h s ) represents the generation probability, h s represents the encoded representation of sentence s; t represents the position index in the summary sequence y, t = 1, ..., T, T represents the length of the summary sequence y.

[0135] Generation probability P(y t |y <t ,h s ) is usually modeled by a Transformer Decoder, relying on previously generated words w and BERT-encoded sentence representation h s .

[0136] This example combines TF-IDF and BERT summary generation technology to intelligently summarize the case description, avoid redundant information, and improve reading efficiency. In order to take into account both keyword weights and deep semantic information, a weighted fusion strategy can be adopted. The final score of sentence s is defined as:

[0137] Score final (s)=α·Score TF-IDF (S)+(1-α)·P(s∈summary)

[0138] Among them: α represents the weight parameter, α∈[0,1σ;

[0139] Score final (s) represents the total score of the candidate sentence.

[0140] By calculating the Score for all candidate sentences final (s), the sentences with the highest scores can be selected to form the final summary, thus realizing intelligent summary generation, which avoids redundant information and improves reading efficiency.

[0141] As a preferred embodiment, in step S4, the key entities of the semantic graph and the determined summary are processed according to the multi-layer self-attention codec to generate an analysis report.

[0142] Specifically, traditional fund analysis reports are mostly manually written or generated based on templates. This approach is not only inefficient but also prone to missing key information. Typically, report generation uses predefined templates and incorporates data query results to populate basic case information and transaction data summaries. Key content also requires manual supplementation, resulting in a low level of automation overall.

[0143] In this example, the analysis report is generated by integrating key entities, summaries, and key evidence corpora. The core of the report generation consists of a multi-layer Transformer decoder. The first layer maps the input sequence z to a higher dimension and performs nonlinear activation. The second layer then maps it back to the original dimension:

[0144] FFN(z)=W2σ(W1z+b1)+b2

[0145] Where z represents the input sequence of the network, including key entities, summaries, and key evidence corpus; W1 and W2 represent the weight parameters of the feedforward neural network FFN; b1 and b2 represent the bias terms of the feedforward neural network FFN; σ represents the ReLU activation function.

[0146] FFN is one of the core components of the Transformer model. It consists of two layers of linear transformation and activation function. Its role is to perform nonlinear transformation on the representation of each position, enhancing the expressive power of the model.

[0147] In each layer, layer normalization is performed after residual connection with the original input sequence to ensure stable gradient flow while maintaining the advantages brought by the depth of the model.

[0148] Residual connections and layer normalization are used within the layer:

[0149]

[0150] Among them, LayerNorm represents layer normalization; MultiHead represents attention head; represents the intermediate result after the first layer calculation, that is, the feature representation after multi-head self-attention, residual connection and layer normalization; z′ represents the intermediate result after the second layer calculation, that is, the feature representation after FFN, residual connection and layer normalization;

[0151] MultiHead is a core component of the self-attention mechanism in the Transformer model. Its core idea is to split the input features into multiple "subspaces" (i.e., multiple attention heads), calculate the attention weights for each subspace in parallel, and finally concatenate the outputs of all heads and map them back to the original dimension. This model can capture the associations between words in the input sequence from different perspectives.

[0152] Through multi-layer self-attention and FFN calculation with causal masks, the context-dependent hidden representation H is obtained:

[0153] H=(h1,h2,…,hn)

[0154] By using the last hidden state, linear projection and softmax(·) function are used to obtain the generation probability of the next word, thereby realizing autoregressive text generation.

[0155] More specifically, it can also include financial compliance analysis. Based on knowledge graph reasoning and combined with relevant laws and regulations, compliance analysis is automatically generated to assist judicial institutions and financial regulatory authorities in decision-making.

[0156] Specifically, human-machine collaborative optimization primarily involves two components: an expert feedback loop and dynamic updates to the domain knowledge base. The expert feedback loop allows users to annotate and correct entities, relationships, and conclusions in generated reports (e.g., marking incorrect funding paths). This feedback data is then used for iterative model training. Dynamic updates to the domain knowledge base automatically incorporate frequently revised content (e.g., terminology related to new criminal methods) into the domain dictionary and knowledge graph.

[0157] On the other hand, a financial case fund analysis report generation system based on a semantic graph large model is also provided, which is used to implement the financial case fund analysis report generation method based on a semantic graph large model as described above. Figure 5 Shown, including:

[0158] Data acquisition module 1, used to acquire the text to be analyzed, which is financial case data;

[0159] Semantic graph construction module 2, connected to data acquisition module 1, is used to perform semantic analysis on the text to be analyzed based on the entity recognition model and the financial entity dictionary to construct a semantic graph. The entity recognition model includes a bidirectional long short-term memory layer and a conditional random field layer. The bidirectional long short-term memory layer is used to identify key entities, and the conditional random field layer is used for sequence labeling.

[0160] Summary generation module 3, connected to semantic graph construction module 3, is used to determine the summary of the text to be analyzed;

[0161] The report generation module 4 is connected to the semantic graph construction module 2 and the summary generation module 3 respectively, and is used to generate an analysis report according to the key entities of the semantic graph and the determined summary.

[0162] In order to improve efficiency and ensure the accuracy of financial case handling, the embodiments of the present invention use large model technology to improve the accuracy and efficiency of case analysis and report generation, providing strong support for the efficient and accurate handling of financial cases.

[0163] This embodiment adopts a large model modeling technology based on domain-enhanced semantic graphs. The semantic graph construction module 2 extracts key entities (such as parties involved in the case, transaction objects, accounts, capital flows, etc.) and their relationships by parsing case descriptions, transaction records and relevant laws and regulations, and stores them in the form of a graph structure. The semantic graph construction method of bidirectional LSTM-CRF plus domain entity dictionary is adopted, combined with the dynamic reasoning ability of the large model to achieve deep semantic analysis of financial cases. The improved bidirectional LSTM-CRF model integrates the financial entity dictionary to improve the recognition accuracy of key entities. The entity relationship is stored through the semantic graph, supporting the graph neural network to reason about the capital path.

[0164] This example utilizes a domain-adaptive large-scale model fine-tuning framework. A pre-trained large-scale model performs deep semantic analysis of financial case texts and is fine-tuned using specialized financial corpus to accurately identify specialized terminology, abbreviations, and complex sentence structures. The pre-trained model, based on the Transformer architecture, achieves precise domain semantic understanding through fine-tuning with specialized financial corpus and combined training with semantic graph embedding. Domain-adaptive pre-training is performed using unstructured corpus such as financial case judgments and regulatory documents, enabling the model to grasp specialized expression logic.

[0165] This embodiment utilizes a dynamically configurable report generation engine. Based on semantic graphs and large-scale model inference results, it dynamically generates case-specific fund analysis reports. Modular content generation technology, based on semantic graph inference results, supports personalized report structure and deep adaptive adjustment. Automatically matching content based on user roles, intelligent compliance inference, and automatically generating violation point analysis.

[0166] like Figure 6 As shown, the method for generating a capital analysis report in this embodiment mainly includes the following steps. There is a clear sequence between the steps and they cannot be interchanged at will. However, some steps (such as expert feedback updates) can exist as asynchronous supplementary processes:

[0167] Data collection and preprocessing stage

[0168] Step 1, data input:

[0169] Collect original financial case data from various sources;

[0170] Step 2, data preprocessing:

[0171] Initial processing of the collected raw financial case data involves cleaning it to remove inaccurate or irrelevant information, converting it into a structured format for easier processing, and converting unstructured data into a structured form. The final output is a series of standardized, structured text and data information, laying a solid foundation for subsequent in-depth analysis.

[0172] Deep semantic analysis and semantic graph construction stage

[0173] Step 3, semantic analysis:

[0174] The pre-processed data is then fed into the deep semantic parsing stage, where advanced natural language processing techniques are used, particularly through large pre-trained models to encode the text.

[0175] Step 4, entity recognition:

[0176] In conjunction with deep semantic parsing, the system effectively extracts key information using an improved bidirectional LSTM-CRF model and a domain-specific entity dictionary. This process not only identifies important entities within a document but also reveals their relationships. The final result is a preliminary parsing report containing detailed entity details and their interrelationships.

[0177] like Figure 7 As shown in Figure 2, the entity recognition process specifically includes the following steps:

[0178] Step 41, case text input;

[0179] Step 42, entity recognition;

[0180] Step 43: Identify fund flows based on the domain entity dictionary;

[0181] Step 44: Identify the parties involved in the case based on the improved BiLSTM-CRF model;

[0182] Step 45, relationship extraction;

[0183] Step 46: construct semantic edges, including transfer, control, and relationship;

[0184] Step 47, semantic storage;

[0185] Step 48, graph structure example;

[0186] Step 5, semantic graph construction:

[0187] After obtaining preliminary parsing results, a semantic graph is constructed, which stores the entity-relationship graph structure. This step follows pre-defined node and edge rules to convert the parsed entities and their relationships into a graphical representation and store it in a graph database. This graphical representation makes complex relationships and patterns more intuitive and easy to understand, while also facilitating subsequent analysis and querying. In this way, the system generates structured semantic graph data, providing strong support for in-depth analysis.

[0188] Step 6, graph calculation: perform calculations using a graph structure;

[0189] Graph neural network reasoning analysis stage

[0190] Step 7, graph neural network capital path reasoning:

[0191] Based on the constructed semantic graph data, the system uses graph neural network (GNN) algorithms for reasoning and analysis. GNNs are particularly adept at discovering hidden patterns and unusual paths within graphs, which is crucial for identifying potential risks in capital flows. By identifying and analyzing these patterns, the system can calculate corresponding risk indicators, providing detailed reasoning results and comprehensive risk assessment data on capital flows.

[0192] Report dynamic generation stage

[0193] Step 8, Dynamic Report Generation Engine:

[0194] After integrating the reasoning results from step 7, the report is dynamically generated. Based on the user's role, the system selects the appropriate report module and dynamically assembles the report content, covering basic case information, capital flows, risk warnings, and legal basis. A concise and clear text summary and insightful graphical analysis are also generated. This comprehensive analytical report provides valuable reference for decision makers.

[0195] like Figure 8 As shown in the figure, the report generation process specifically includes the following steps:

[0196] Step 81, obtaining semantic graph analysis results;

[0197] Step 82: Determine the user role. User roles include default, regulatory agency, and procuratorate. If the user role is the default role, execute steps 8311-8312; if the user role is the regulatory agency, execute steps 8321-8322; if the user role is the procuratorate, execute steps 8331-8332;

[0198] Step 8311, basic analysis;

[0199] Step 8312, visualization of the capital path;

[0200] Step 8321, compliance advice;

[0201] Step 8322, citing relevant laws and regulations;

[0202] Step 8331, legal basis;

[0203] Step 8332, automatically associate relevant laws and regulations;

[0204] Step 84, knowledge graph query;

[0205] Step 85, historical similar cases;

[0206] Step 86, matching legal terms;

[0207] Step 87, dynamic content assembly;

[0208] Step 88, the report template may be configured;

[0209] Step 89: Output analysis report;

[0210] After the analysis report is generated, it can be used to assist decision-making, such as the general model risk assessment in step 91, the regulatory agency's compliance recommendations in step 92, and the legal basis of the procuratorate in step 93;

[0211] User interaction and feedback update phase

[0212] Step 10, user interface interaction;

[0213] Step 11: Expert feedback annotation;

[0214] Step 12: Update the domain knowledge base;

[0215] Step 13: Fine-tune the framework of the large model.

[0216] The system not only provides users with detailed analysis reports but also encourages feedback. Every user's browsing, editing, and feedback is treated as a valuable data resource, used to correct system errors, update the domain knowledge base, and adjust large model parameters. This approach achieves closed-loop learning, continuously optimizing system performance and ensuring it always operates at its optimal state.

[0217] This invention significantly improves the efficiency, accuracy, and professionalism of generating financial case fund analysis reports through the deep integration of the semantic graph model and the dynamic generation engine. The specific technical effects are as follows:

[0218] (1) Enhanced semantic parsing capabilities. By combining an improved bidirectional LSTM-CRF model with a financial entity dictionary, the accuracy of identifying key entities such as parties involved in the case and transaction objects is improved, effectively resolving missed detection issues caused by specialized terminology and complex sentence structures. By injecting structured features of semantic graphs into the large model fine-tuning framework, the ability to understand contextual associations is enhanced, the model's reasoning accuracy for the capital chain is improved, and the false positive rate is reduced.

[0219] (2) Improve report generation efficiency and quality. The time required to generate a single report is reduced to less than 1 minute, and batch case processing is supported, significantly improving generation efficiency. Dynamic module assembly technology can automatically match content modules based on user roles, and the compliance intelligent reasoning module can automatically associate relevant regulatory provisions, establishing a full-dimensional violation analysis framework. Combined with the BERT-based case information intelligent summary system, semantic distillation technology is used to improve the purity of information extraction, achieving precise focusing of legal elements and redundant filtering.

[0220] (3) Improve the system's self-evolution capabilities. Through the expert feedback closed-loop mechanism, the model adaptation cycle for new criminal method terminology has been shortened from several months required for traditional manual updates to several days, improving response speed and significantly shortening the model iteration cycle. It supports the dynamic expansion of the domain knowledge base, and high-frequency correction content (such as "NFT money laundering path") is automatically incorporated into the knowledge graph, improving the system's recognition accuracy of new crime patterns.

[0221] The above are only preferred embodiments of the present invention and do not limit the implementation mode and protection scope of the present invention. For those skilled in the art, it should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the contents of this specification and illustrations should be included in the protection scope of the present invention.

Claims

1. A method for generating a financial case fund analysis report based on a semantic graph large model, characterized in that: include: Step S1, obtaining a text to be analyzed, wherein the text to be analyzed is financial case data; Step S2: semantically parse the text to be analyzed based on an entity recognition model and a financial entity dictionary to construct a semantic graph; wherein the entity recognition model includes a bidirectional long short-term memory layer and a conditional random field layer, the bidirectional long short-term memory layer is used to identify key entities, and the conditional random field layer is used for sequence labeling; Step S3, determining the summary of the text to be analyzed; Step S4: generating an analysis report based on the key entities of the semantic graph and the determined summary.

2. The method for generating a financial case fund analysis report based on a semantic graph large model according to claim 1 is characterized in that: The step S2 comprises: Step S21, extracting key entities from the text to be analyzed, and performing enhancement processing on the extracted key entities to obtain enhanced key entities; Step S22, inputting the enhanced key entity into the entity recognition model to identify the key entity; Step S23: performing transformation based on the identified key entities and key entity relationships to construct the semantic graph.

3. The method for generating a financial case fund analysis report based on a semantic graph large model according to claim 2 is characterized in that: The step S21 includes: Step S211: for each extracted key entity, determine whether the key entity belongs to the financial domain entity dictionary; if the key entity belongs to the financial domain entity dictionary, execute step S22; otherwise, execute step S23; Step S212: taking the sum of the vector representations of the key entity and the entities in the financial domain entity dictionary to which it belongs as the enhanced key entity; Step S213: Using the key entity as the enhanced key entity.

4. The method for generating a financial case fund analysis report based on a semantic graph large model according to claim 1 is characterized in that: In the conditional random field layer, the conditional probability of the label sequence is: Where: P(Y|X) represents the probability of the labeled sequence Y under the condition of the input text sequence X of the text to be analyzed; Represents the label y t-1 Move to label y t The state transition probability of y is t, t represents the position index of label y in the annotation sequence Y, t = 1, ..., n, n represents the length of the annotation sequence Y; φ(y t ,h t ) represents the label y t Based on the hidden state sequence h t The local score of , φ represents the linear transformation function.

5. The method for generating a financial case fund analysis report based on a semantic graph large model according to claim 1 is characterized in that: The step S3 comprises: Step S31, for each sentence in the text to be analyzed, performing an importance evaluation on the sentence to obtain a total importance score of each sentence, and screening candidate sentences for the abstract according to the total importance score; Step S32, semantically encoding the candidate sentence based on a generative summary generation model to generate a probability of the candidate sentence being selected as a summary; Step S33, performing weighted fusion on the total importance score of each sentence and the probability of the candidate sentence being selected as the summary to obtain the total score of all candidate sentences; Step S34: Determine the candidate sentence with the highest total score as the summary of the text to be analyzed.

6. The method for generating a financial case fund analysis report based on a semantic graph large model according to claim 5 is characterized in that: The total importance score of the sentence is: Among them, Score TF-IDF (s) represents the total importance score of sentence s; tfidf(w,D) represents the term frequency-inverse document frequency index of keyword w in the text to be analyzed.

7. The method for generating a financial case fund analysis report based on a semantic graph large model according to claim 5 is characterized in that: The probability that the candidate sentence is selected as the summary is: P(s∈summary)=σ(W e ·h s +b e ) Among them, P(s∈summary) represents the probability that sentence s is selected as the summary; σ represents the sigmoid activation function; W e represents weight; h s represents the encoding representation of sentence s; b e represents the bias term.

8. The method for generating a financial case fund analysis report based on a semantic graph large model according to claim 5 is characterized in that: In the generative summary generation model, the generation process of the summary sequence is: Among them, P(y|s) represents the probability of the summary sequence y under the condition of sentence s; P(y t |y <t ,h s ) represents the generation probability, h s represents the encoded representation of sentence s; t represents the position index in the summary sequence y, t = 1, ..., T, T represents the length of the summary sequence y.

9. The method for generating a financial case fund analysis report based on a semantic graph large model according to claim 1 is characterized in that: In step S4, the key entities of the semantic graph and the determined summary are processed according to a multi-layer self-attention codec to generate the analysis report.

10. A financial case fund analysis report generation system based on a semantic graph large model, characterized in that: A method for generating a financial case fund analysis report based on a semantic graph large model according to any one of claims 1 to 9, comprising: A data acquisition module is used to acquire the text to be analyzed, wherein the text to be analyzed is financial case data; A semantic graph construction module, connected to the data acquisition module, is used to perform semantic parsing on the text to be analyzed based on an entity recognition model and a financial entity dictionary to construct a semantic graph; wherein the entity recognition model includes a bidirectional long short-term memory layer and a conditional random field layer, the bidirectional long short-term memory layer is used to identify key entities, and the conditional random field layer is used for sequence labeling; A summary generation module, connected to the semantic graph construction module, for determining a summary of the text to be analyzed; A report generation module is connected to the semantic graph construction module and the summary generation module respectively, and is used to generate an analysis report according to the key entities of the semantic graph and the determined summary.

Images