A Method and System for Financial Behavior Identification Based on Rule-Injected Unsupervised Neural Networks

Abstract

This invention discloses a method and system for identifying financial behavior based on a rule-injected unsupervised neural network. According to the rule paradigm for identifying financial behavior facts, a corresponding fact-finding rule paradigm and rule-mapping injection labels are constructed. Unsupervised neural network information extraction technology is used to extract fact-finding elements, which are then associated with relevant parties. The unsupervised extraction results and the fact-finding rule paradigm are combined to output rule-injected fact-finding results for each relevant party. Simultaneously, a neural network feature extractor and a multi-label classifier are used to obtain neural network-based fact-finding results. Finally, combining the two fact-finding results, a rule-injection arbitration mechanism is employed to generate the final fact-finding result. Compared with existing technologies, this invention simultaneously combines the high accuracy of neural networks and the high interpretability of rule-based reasoning; the fusion of these two aspects results in higher accuracy in the analysis of financial behavior facts.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and knowledge graph technology, specifically to a method and system for judicial determination of financial behavior based on rule-injected unsupervised neural networks. Background Technology

[0002] Financial transactions differ from general transactions, often involving multiple individuals and situations where one person may be guilty of multiple offenses. As these processes become increasingly complex over time, using simple machine learning and deep learning multi-label classifiers for fact-finding suffers from low accuracy and poor interpretability. Currently, most AI models for fact-finding reasoning are end-to-end models (i.e., neural network models), and no similar technology has yet been found for injecting rules into unsupervised models into fact-finding reasoning architectures.

[0003] Currently, most artificial intelligence models for fact-finding reasoning are end-to-end models (i.e., neural network models).

[0004] Most models and methods only extract behavioral feature elements to build fact-finding models, without considering the relationship between the actors and these behavioral elements. Financial transactions often involve multiple actors, each with their own corresponding behavioral elements. When making fact-finding decisions for each actor, these behavioral elements need to be considered individually and thoroughly. Modeling methods using mainstream approaches often have the following shortcomings for intelligent fact-finding in the judicial field of financial transactions:

[0005] 1) The fact-finding results are poorly interpretable; for example, the use of machine learning or deep learning models may result in fact-finding results that are unreadable to humans.

[0006] 2) It can only identify entities and relatively simple relationships, and cannot accurately match elements with relevant people. It cannot meet the needs in scenarios involving multiple people's actions or multiple crimes committed by one person, and its auxiliary function for intelligent fact-finding is insufficient.

[0007] 3) The methods for fact-finding are primarily geared towards the broad criminal field, lacking specificity in the financial sector. Therefore, they are difficult to apply with high precision and practicality in the financial sector.

[0008] How to design an intelligent method and system for fact-finding in the financial behavior and judicial fields that overcomes and optimizes the above-mentioned defects is the technical problem that this invention urgently needs to solve. Summary of the Invention

[0009] To address the technical issues mentioned above, this invention proposes a method and system for determining the facts of financial activities based on rule-injected unsupervised neural networks. Under the rule-injected fact-determination reasoning framework of the unsupervised neural network model, the method utilizes a rule-injected arbitration mechanism for intelligent analysis to achieve the determination of facts in the judicial field of financial activities.

[0010] This invention is achieved using the following technical solution:

[0011] The present invention provides a method for identifying financial behavior based on a rule-injected unsupervised neural network, the method comprising the following specific steps:

[0012] Step 1: Obtain the factual text of the financial transaction to be determined;

[0013] Step 2: Generate injection-based fact-finding results and neural network-based fact-finding results based on unsupervised neural network models;

[0014] Step 3: Based on the combination of the rule-injected fact-finding results and the neural network-based fact-finding results, the rule-injected arbitration mechanism is used for intelligent analysis to obtain the final fact-finding results.

[0015] This invention discloses a financial behavior recognition system based on a rule-injected unsupervised neural network. The system includes an acquisition module,

[0016] The generation module and rule injection into the arbitration fact-finding module, wherein:

[0017] The acquisition module is used to acquire the text of the financial behavior facts to be determined.

[0018] The generation module is used to generate injection-based fact-finding results and neural network-based fact-finding results based on unsupervised neural network models.

[0019] The rule-injected arbitration fact-finding module is used to perform intelligent analysis based on the combination of the rule-injected fact-finding results and the neural network-based fact-finding results, and to obtain the final fact-finding results.

[0020] Compared with the prior art, the present invention can achieve the following positive technical effects:

[0021] 1) It can simultaneously combine the high accuracy of neural networks and the high interpretability of rule-based reasoning. By integrating the two and using a rule-injection arbitration mechanism for intelligent analysis, the analysis results on the determination of facts of financial behavior 2 have a higher accuracy.

[0022] 2) It has better model interpretability than other pure neural network reasoning methods. Attached Figure Description

[0023] Figure 1 is a flowchart of the overall process of a financial behavior identification method based on rule-injected unsupervised neural network according to the present invention;

[0024] Figure 2 is a block diagram of a financial behavior recognition system based on a rule-injected unsupervised neural network according to the present invention;

[0025] Figure 3 is a schematic diagram of the specific implementation process of the present invention;

[0026] Figure 4 shows an example of an unsupervised neural network behavior feature extraction model;

[0027] Figure 5 is a schematic diagram of an example of a fact-finding rule paradigm;

[0028] Figure 6 This is a schematic diagram of Example 2 of the fact-finding rule paradigm. Detailed Implementation

[0029] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. All other embodiments of the technical solutions obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0030] Figure 1 shows a flowchart of a financial behavior identification method based on a rule-injected unsupervised neural network according to the present invention. The process specifically includes the following steps:

[0031] Step S101: Obtain the factual text of the financial transaction to be determined;

[0032] Step S102: Construct a fact-finding rule paradigm and rule mapping injection labels based on the financial behavior fact text. Inject the fact-finding rule paradigm into the behavior feature extraction model of the unsupervised neural network to extract fact-finding elements that are associated with each relevant person. Input the fact-finding elements into the fact-finding rule paradigm to obtain the rule-injected fact-finding results of each relevant person and output the rule-injected fact-finding results of each relevant person.

[0033] The rule-mapping injection labels and corresponding fact-finding rule paradigms are designed based on fact-finding rules. The injection labels are mainly used to inject into the unsupervised neural network extraction model to better extract relevant fact-finding elements. Simultaneously, the extracted fact-finding elements can be filled into the fact-finding rule paradigm for fact-finding, making the entire process interconnected. Associating fact-finding elements with relevant individuals simultaneously combines the high accuracy of neural networks with the high interpretability of rule-based reasoning; the fusion of these two approaches leads to higher accuracy.

[0034] Pre-trained models such as BERT, Roberta, and Nezha are used to encode the input financial behavior fact text to be determined. Then, the pooling layer of the pre-trained model is taken as the text feature. The text feature is fed into a fully connected layer classifier to perform multi-label classification of the crime, and a neural network-based fact determination result is obtained.

[0035] Step S103: Use a neural network feature extractor and a multi-label classifier on the rule-injected fact-finding results of each relevant person to obtain neural network fact-finding results;

[0036] Unsupervised neural network models utilize models pre-trained using transfer learning methods. These models are pre-trained on a large number of upstream tasks that are similar to downstream tasks, enabling them to achieve good results on downstream tasks.

[0037] Step 4: Based on the rule-injected fact-finding results of each relevant party and the neural network-based fact-finding results used as the baseline fact-finding results, perform intelligent analysis of the fact-finding results: When the financial crime involved in the neural network-based fact-finding results or the rule-injected fact-finding results are not empty, use the rule-injected fact-finding results to arbitrate the baseline fact-finding results. Finally, only the crimes that have appeared in the injected fact-finding results are retained in the baseline fact-finding results, and the hit fact-finding rule paradigm is returned as a highly interpretable fact-finding analysis result.

[0038] In summary, the rule-injection arbitration mechanism in this step specifically trains a neural network feature extractor and a multi-label classifier based on general behavior. The neural network-based fact-finding result is used as the baseline fact-finding result. Simultaneously, a rule-injection neural network model and a fact-finding rule paradigm are used to obtain the rule-injection-based fact-finding result. Finally, combining the neural network-based fact-finding result and the rule-injection-based fact-finding result to obtain the final fact-finding result allows for a more accurate and clearly interpretable fact-finding process.

[0039] Figure 2 shows a financial behavior identification system based on a rule-injected unsupervised neural network. The system includes an acquisition module 1, a generation module 2, and a rule-injected arbitration fact-finding module 3. Wherein:

[0040] Module 1 is used to acquire the text of the financial behavior facts to be determined.

[0041] Generation module 2 is used to generate injection-based fact-finding results and neural network-based fact-finding results based on unsupervised neural network models;

[0042] The rule-injected arbitration fact-finding module 3 is used to perform intelligent analysis based on the combination of the rule-injected fact-finding results and the neural network-based fact-finding results, and to obtain the final fact-finding results.

[0043] The generation module constructs fact-finding rules and rule-mapping injection tags based on the financial behavior fact text, injects the fact-finding rule paradigm into the behavior feature extraction model of the unsupervised neural network, extracts the fact-finding elements that are respectively associated with each relevant person, inputs the fact-finding elements into the fact-finding rule paradigm to obtain the rule-injected fact-finding results of each relevant person, and outputs the rule-injected fact-finding results of each relevant person.

[0044] Furthermore, a neural network feature extractor and a multi-label classifier are used to obtain neural network-based fact-finding results from the rule-injected fact-finding results of each relevant person.

[0045] The rule-injected arbitration fact-finding module uses the neural network-based fact-finding result as the baseline fact-finding result. When the neural network-based fact-finding result involves a financial crime or the rule-injected fact-finding result of each relevant person is not empty, the rule-injected fact-finding result is used to arbitrate the baseline fact-finding result, retaining the rule-injected fact-finding result and the crimes that have occurred.

[0046] Specifically, the implementation process of rule injection in unsupervised neural network models includes:

[0047] ① The rule mapping injection tag is appended to the end of the text of the financial behavior fact to be determined as input;

[0048] ② Use a transfer learning-based information extraction model to extract the labeled text;

[0049] ③ Patch the rule mapping into the labels and the corresponding label text to the end of the input financial behavior fact text;

[0050] ④ Use a transfer learning-based information extraction model to extract the subject corresponding to the labeled text, i.e., the relevant person;

[0051] ⑤ Aggregate tags around relevant individuals, and each relevant individual will eventually receive a tuple extraction result.

Claims

1. A method for identifying financial behavior based on rule-injected unsupervised neural networks, characterized in that, The method includes the following steps: Step 1: Obtain the factual text of the financial transaction to be determined; Step 2: Generate rule-injected fact-finding results and neural network-based fact-finding results; including: The rule-injected fact-finding results are as follows: a fact-finding rule paradigm and a rule-mapping injection label are constructed based on the financial behavior fact text. The fact-finding rule paradigm and the rule-mapping injection label are injected into the behavior feature extraction model of an unsupervised neural network. Fact-finding elements corresponding to each relevant person are extracted based on the rule-mapping injection label. The fact-finding elements are input into the fact-finding rule paradigm to obtain the rule-injected fact-finding results of each relevant person. The rule-injected fact-finding results of each relevant person are then output. The neural network-based fact-finding result: The neural network feature extractor and multi-label classifier are used on the rule-injected fact-finding results of each relevant person to obtain the neural network-based fact-finding result; Step 3: Based on the combination of the rule-injected fact-finding results and the neural network-based fact-finding results, intelligent analysis is performed using the rule-injected arbitration mechanism to obtain the final fact-finding results. The neural network-based fact-finding results are used as the baseline fact-finding results. When the neural network-based fact-finding results involve financial behavior determination results or the rule-injected fact-finding results of the relevant parties are not empty, the rule-injected fact-finding results are used to arbitrate the baseline fact-finding results, retaining the determination results that have appeared in the rule-injected fact-finding results.

2. A financial behavior recognition system based on rule-injected unsupervised neural networks, characterized in that, The system includes an acquisition module, a generation module, and a rule injection arbitration fact-finding module, wherein: The acquisition module is used to acquire the text of the financial behavior facts to be determined. The generation module is used to generate rule-injected fact-finding results and neural network-based fact-finding results based on an unsupervised neural network model. This includes constructing a fact-finding rule paradigm and rule-mapping injection tags based on the financial behavior fact text; injecting the fact-finding rule paradigm and the rule-mapping injection tags into an unsupervised neural network behavior feature extraction model; extracting fact-finding elements corresponding to each relevant person based on the rule-mapping injection tags; inputting the fact-finding elements into the fact-finding rule paradigm to obtain the rule-injected fact-finding results for each relevant person; and outputting the rule-injected fact-finding results for each relevant person. The rule-injected arbitration fact-finding module is used to perform intelligent analysis based on the combination of the rule-injected fact-finding result and the neural network fact-finding result, and to obtain the final fact-finding result. The neural network fact-finding result is used as the baseline fact-finding result. When the neural network fact-finding result involves the determination of financial behavior or the rule-injected fact-finding result of each relevant party is not empty, the rule-injected fact-finding result is used to arbitrate the baseline fact-finding result, and the determination results that have appeared in the rule-injected fact-finding result are retained.

Images