A financial fraud prevention knowledge promotion method, device, equipment and storage medium

Abstract

The application relates to the technical field of financial propaganda, and in particular to a financial fraud-prevention knowledge propaganda method, device and equipment and a storage medium. The method comprises the following steps: constructing and training a language processing model to obtain a financial dialogue model; a digital virtual human database is constructed; the digital virtual human database comprises a plurality of virtual humans appearing in historical real fraud and fraud cases; a financial question input by a user is received, the financial question is input into the financial dialogue model to obtain propaganda information to be replied; a corresponding virtual human and a virtual scene are determined according to the financial question, and the virtual scene and the virtual human are loaded and displayed; and the virtual human is driven to output the propaganda information through voice. Therefore, the application can effectively solve the problems of poor effect and low efficiency of existing financial propaganda activities.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of financial propaganda, and in particular to a financial fraud prevention knowledge propaganda method, device, equipment and storage medium. BACKGROUND

[0002] In the field of bank finance, various financial fraud and fraud behaviors often occur, causing losses and risks to banks and customers. In order to improve the risk awareness and prevention awareness of customers, banks need to carry out financial propaganda activities to introduce the methods and means of financial fraud and fraud to the public and improve the risk prevention ability of customers.

[0003] At present, traditional bank financial propaganda activities are mostly in the form of propaganda materials and videos, and the effect is limited. Therefore, improving the effect and efficiency of bank financial propaganda is a problem to be solved. SUMMARY

[0004] Therefore, the embodiments of the present application provide a financial fraud prevention knowledge propaganda method, device, equipment and storage medium, which can effectively solve the problems of poor effect and low efficiency of existing financial propaganda activities.

[0005] In a first aspect, the embodiments of the present application provide a financial fraud prevention knowledge propaganda method, comprising:

[0006] building and training a language processing model to obtain a financial dialogue model;

[0007] building a digital virtual human database; the digital virtual human database includes a plurality of virtual humans appearing in historical real fraud and fraud cases;

[0008] receiving a financial question input by a user, inputting the financial question into the financial dialogue model to obtain propaganda information to be replied;

[0009] determining a corresponding virtual human and a virtual scene according to the financial question, and loading and displaying the virtual scene and the virtual human;

[0010] driving the virtual human to output the propaganda information by voice.

[0011] In some embodiments, the method further comprises:

[0012] receiving a user instruction and / or a set instruction, and using the financial dialogue model to obtain a corresponding fraud and fraud case from the financial fraud and fraud database;

[0013] obtaining a corresponding virtual human according to the scene role in the fraud and fraud case, and obtaining a corresponding virtual scene according to the scene information in the fraud and fraud case;

[0014] Load the virtual scene and the virtual person, and drive each virtual person to explain the case according to the scene script in the fraud and fraud case.

[0015] In some embodiments, the financial fraud and fraud database is constructed by the following method:

[0016] Analyzing historical real fraud and fraud cases to obtain scene scripts, scene information, and each scene role; the scene script describes the speech of each scene role according to the scene order;

[0017] The scene script, scene information, and each scene role constitute a fraud and fraud case, and are stored in a fraud and fraud case data table.

[0018] In some embodiments, the digital virtual person database is constructed, including:

[0019] A plurality of virtual voices are constructed by using a speech synthesis technology;

[0020] A plurality of virtual images are constructed by constructing the character models commonly appearing in the historical real fraud and fraud cases;

[0021] A plurality of virtual actions are constructed by constructing virtual person actions according to the action characteristics of the characters appearing in the historical real fraud and fraud cases;

[0022] The virtual voices, virtual images, and virtual actions are combined to obtain a plurality of virtual persons.

[0023] In some embodiments, the virtual scene is obtained by the following method:

[0024] The corresponding scene data is obtained by analyzing the historical real fraud and fraud cases;

[0025] The corresponding virtual scene is constructed by using two-dimensional and / or three-dimensional modeling technology according to the scene data.

[0026] In some embodiments, the virtual person and the virtual scene are stored in combination with corresponding set keywords;

[0027] The corresponding virtual person and virtual scene are determined according to the financial problem, including:

[0028] The keywords in the financial problem are extracted by using a keyword extraction technology, and the corresponding virtual person and virtual scene are matched according to the keywords.

[0029] In some embodiments, the language processing model includes a transformer encoder and an autoregressive decoder.

[0030] The transformer encoder comprises N encoding layers, N>1, and the encoding layers comprise a multi-head self-attention mechanism and a feedforward neural network.

[0031] The autoregressive decoder comprises N decoding layers, and the decoding layers comprise a multi-head self-attention mechanism, a multi-head attention mechanism and a feedforward neural network.

[0032] In a second aspect, the embodiments of the present application provide a financial fraud prevention knowledge promotion device, comprising:

[0033] A virtual person database construction module is configured to construct a digital virtual person database, and the digital virtual person database comprises a plurality of virtual persons appearing in historical real fraud and fraud cases.

[0034] A financial dialogue module is configured to receive a financial question input by a user, input the financial question into a financial dialogue model to obtain promotion information to be replied, and obtain the promotion information to be replied from a language processing model constructed by training.

[0035] A scene display loading module is configured to determine a corresponding virtual person and a virtual scene according to the financial question, and load and display the virtual scene and the virtual person.

[0036] A promotion driving module is configured to drive the virtual person to output the promotion information through voice.

[0037] In a third aspect, the embodiments of the present application provide a terminal device, comprising a processor and a memory, wherein the memory stores a computer program, and the processor is configured to execute the computer program to implement the financial fraud prevention knowledge promotion method provided in the first aspect of the present application.

[0038] In a fourth aspect, the embodiments of the present application provide a readable storage medium, which stores a computer program, and the computer program is executed on a processor to implement the financial fraud prevention knowledge promotion method provided in the first aspect of the present application.

[0039] The embodiments of the present application have the following beneficial effects:

[0040] The embodiments of the present application construct a digital virtual person database to obtain a plurality of virtual persons, and then realize intelligent dialogue by combining the virtual persons with a language processing model.

[0041] The virtual person can realize interactive communication with the user, can effectively improve the interest and acceptance speed of the user, and can further improve the efficiency and effect of bank financial propaganda.

[0042] The financial dialogue model adopts a voice interaction mode, the customer can obtain answers through simple voice questions, avoids complicated operations, and improves customer experience. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some of the embodiments of the present application, and therefore should not be regarded as limiting the scope, and for those skilled in the art, other related drawings can also be obtained without creative labor.

[0044] Figure 1 A flowchart of a financial fraud prevention knowledge promotion method according to an embodiment of the present application is shown;

[0045] Figure 2 A schematic diagram of a scene script in a fraud and fraud case of a financial fraud prevention knowledge promotion method according to an embodiment of the present application is shown;

[0046] Figure 3 Another flowchart of a financial fraud prevention knowledge promotion method according to an embodiment of the present application is shown;

[0047] Figure 4 A structural schematic diagram of a financial fraud prevention knowledge promotion device according to an embodiment of the present application is shown.

[0048] Main component symbol explanation:

[0049] 410 - virtual person database construction module; 420 - financial dialogue module; 430 - scene display loading module; 440 - promotion driving module. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments.

[0051] The components of the embodiments of the present application generally described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of the present application.

[0052] Hereinafter, the terms "include", "have", and their conjugates, which are used in various embodiments of the present application, merely indicate the presence of the features, numbers, steps, operations, elements, components, or combinations thereof, and do not preclude the presence or addition of one or more other features, numbers, steps, operations, elements, components, or combinations thereof. In addition, the terms "first", "second", "third", and the like, are used only to distinguish the description, and cannot be understood as indicating or implying a relative importance.

[0053] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which various embodiments of the present application belong. The terms (such as terms defined in a generally used dictionary) will be interpreted to have the same meaning as the contextual meaning in the relevant technical field and will not be interpreted to have an idealized or overly formal meaning unless clearly defined in various embodiments of the present application.

[0054] Some embodiments of the present application will be described in detail below with reference to the accompanying drawings. The following embodiments and features in the embodiments can be combined with each other without conflict.

[0055] Most existing financial propaganda activities use paper materials or videos and the like to make propaganda. However, these propaganda forms can only make the audience passively accept, cannot actively choose the knowledge needed, most of the audience are not interested, there is no interaction and no sense of participation, and the content of the propaganda is also not remembered, which ultimately leads to poor propaganda effect and low propaganda efficiency. Therefore, the present application proposes a financial fraud prevention knowledge propaganda method, device, equipment and storage medium, which can effectively solve the problems of poor effect and low efficiency of existing financial propaganda activities.

[0056] The financial fraud prevention knowledge propaganda method will be described below in conjunction with some specific embodiments.

[0057] Figure 1 A flowchart of the financial fraud prevention knowledge propaganda method of the embodiments of the present application is shown. Exemplarily, the financial fraud prevention knowledge propaganda method includes the following steps:

[0058] S10, a language processing model is constructed and trained to obtain a financial dialogue model.

[0059] In step S10, a language processing model is constructed, a financial propaganda question and answer knowledge training set is constructed, and the constructed language processing model is trained to obtain a financial dialogue model. The financial propaganda question and answer knowledge training set is made by collecting financial propaganda question and answer knowledge.

[0060] The language processing model includes a transformer encoder and an autoregressive decoder. The transformer encoder includes multiple attention mechanisms for encoding the input sequence. Specifically, the transformer encoder includes N encoding layers, N > 1, which include a multi-head self-attention mechanism and a feed-forward neural network. The multi-head self-attention mechanism is used to enable the language processing model to focus on the relationships between different positions in the input sequence when encoding, thereby better understanding the entire sequence.

[0061] The autoregressive decoder includes N decoding layers, which include a multi-head self-attention mechanism, a multi-head attention mechanism, and a feed-forward neural network. The autoregressive decoder is used to generate the next word or token, with the input being the text sequence generated by the transformer encoder.

[0062] In addition to implementing financial propaganda knowledge question answering, the financial dialogue model can also extract keywords from the input text. After receiving the input text sequence, the financial dialogue model needs to preprocess the text sequence. Preprocessing data includes data cleaning, tokenization, and conversion to id, etc. operations, so that the language processing model can better understand and process data, thereby improving the performance and effect of the language processing model.

[0063] The training process of the language processing model is as follows:

[0064] 1) Preprocess the data.

[0065] It can be understood that the financial propaganda question and answer knowledge training set also needs to be preprocessed during the training process. Preprocessing data includes data cleaning, tokenization, and conversion to id, etc. operations.

[0066] (1) Data cleaning mainly involves text cleaning of dialogue data, including removing noise and useless information. For example, removing special characters, punctuation marks, HTML tags, and emoticons, etc. The specific operation is as follows:

[0067] Remove special characters: delete special characters in the text, such as punctuation marks, such as periods, commas, exclamation marks, question marks, and colons, etc. Since these characters do not affect the meaning of the text, they can be removed. For example, the example text: Hello, world! What's new? Cleaned text: Hello world, what's new?

[0068] Remove numbers: according to the task requirements, you can choose to remove numbers or keep numbers. For some tasks, such as sentiment analysis, numbers may not have important significance and can be deleted. Example text: I have 3 cats. Cleaned text: I have a cat.

[0069] Remove extra spaces: merge multiple consecutive spaces into one space to ensure that the spaces in the text are used correctly and consistently. Example text: This is a test. Cleaned text: This is a test.

[0070] Further, data cleaning also includes performing text normalization operations such as converting uppercase letters to lowercase letters, expanding abbreviations, and sorting text, etc. The specific operations are as follows:

[0071] Convert to lowercase: Convert all letters to lowercase to help eliminate differences in word case and enable the model to better understand and generalize.

[0072] Remove stop words: Stop words are words that frequently appear in text but usually do not carry important meaning, such as prepositions, articles, etc. Depending on task requirements, these stop words can be removed to reduce noise and data dimensionality. Example text: I want to go to the store. Cleaned text: go to the store.

[0073] Process abbreviations: Expand abbreviations to full names for better understanding of the text. For example, expand "can't" to "cannot".

[0074] Remove URLs and HTML tags: For text containing URL links or HTML tags, they can be removed or replaced with special markers to avoid interference with the model.

[0075] Emoticon processing: Depending on task requirements, emoticons can be removed or converted to special markers to handle emotional or semantic expression. Example text: I'm so happy! Cleaned text: I'm so happy [smile].

[0076] Process repeated characters: For example, reduce multiple consecutive question marks or exclamation marks to one to reduce repeated information in the text. Example text: I'm so excited!!! Cleaned text: I'm so excited!

[0077] Filter sensitive information: Depending on the specific application scenario, sensitive information needs to be filtered to ensure that the output of the model meets privacy and security requirements, such as card numbers and names. Example text: My ID number is 123456789. Cleaned text: My ID number is [sensitive information].

[0078] (2) Tokenization mainly uses the current common analysis mode, and the commonly used tokenization technology in current technology can also use regular matching tokenization or use BPE (Byte Pair Encoding) tokenization. By repeatedly merging the most frequently occurring characters or character combinations to construct a vocabulary, it can capture common words and word fragments while preserving information about rare words and word combinations.

[0079] The following is a brief introduction to the tokenization process based on BPE technology:

[0080] Building Initial Vocabulary: Divide the text data into individual characters or combinations of characters (e.g., letters, numbers, punctuation marks) and calculate their frequencies.

[0081] Iterative Merging: Select the most frequently occurring characters or combinations of characters and merge them into a new sub-word. This merging process is iterated multiple times, with each iteration selecting a new most frequent combination to merge until a pre-set vocabulary size is reached or a stopping condition (e.g., number of iterations) is met.

[0082] Generating Vocabulary: The results of the iterative merging process are used as the final vocabulary. Each entry in the vocabulary represents a sub-word or complete word.

[0083] Tokenization: Use the generated vocabulary to tokenize the input text. This means replacing word or word fragments in the text with their corresponding entries in the vocabulary.

[0084] As an example, let's consider the following dialogue to illustrate the vocabulary construction process:

[0085] Original Text: Xiaoming said, "I want to apply for a bank card." Xiaohong replied, "Okay, let's go tomorrow."

[0086] Preprocessed Text: Xiaoming said I want to apply for a bank card Xiaohong replied Okay let's go tomorrow.

[0087] Tokenized Result: ["Xiaoming", "said", "I", "want", "to", "apply", "for", "a", "bank", "card", "Xiaohong", "replied", "Okay", "let's", "go", "tomorrow"].

[0088] Building Vocabulary: Sort the tokenized results by frequency and select the most common words to form the vocabulary. For example, select the top 10 words as the vocabulary, numbered as follows:

[0089] Vocabulary: {"Xiaoming": 0, "said": 1, "I": 2, "want": 3, "to": 4, "apply": 5, "for": 6, "a": 7, "bank": 8, "card": 9}.

[0090] Adding Special Markers: Add some special markers to the vocabulary, such as a start marker ( <start>), end tag ( <end>) and unknown word token ( <unk>).

[0091] Final vocabulary:{" <start> ":0," <end> ":1," <unk>{"2":"Xiaoming", "3":"said", "4":"I", "5":"want", "6":"to", "7":"apply", "8":"for", "9":"Xiaohong", "10":"answer", "11":"OK", "12":"we", "13":"tomorrow", "14":"to", "15":"go"}

[0092] In the tokenization phase, the question and answer are segmented into individual words or subwords to ensure that the keywords can correspond to the vocabulary in the vocabulary table. For the question and answer in the example:

[0093] Question: What card does Xiaoming want to apply for?

[0094] Answer: Xiaoming wants to apply for a credit card.

[0095] The tokenization result is:

[0096] Question: ["Xiaoming", "want", "to", "apply", "for", "what", "card"]

[0097] Answer: ["Xiaoming", "want", "to", "apply", "for", "credit", "card"]

[0098] Next, in the indexing phase, the tokenized words are converted into the corresponding index in the vocabulary table. Using the vocabulary table number, each word is converted into the corresponding integer index. For the question and answer in the example:

[0099] Question: [3, 6, 7, 13].

[0100] Answer: [3, 6, 7, 8].

[0101] When using the trained financial dialogue model, the answer can be given according to the integer index.

[0102] Unsupervised pre-training:

[0103] The language processing model uses unsupervised pre-training method, that is, training on large-scale unlabeled text data set. The pre-training process contains two stages: Masked Language Modeling (MLM) and Next Sentence Prediction (NSP). In the MLM task, the language processing model is used to predict which word should be in the mask position. In the NSP task, the language processing model is used to judge whether two sentences are adjacent or separated to learn the relationship between sentences.

[0104] Supervised fine-tuning:

[0105] After completing unsupervised pre-training, the language processing model usually needs to be further supervised fine-tuning to adapt to specific tasks and financial propaganda question and answer knowledge training sets. For example, in the text generation task, the pre-trained model can be used to generate new text sequences; in the question and answer task, the pre-trained model can be used to answer questions, etc. At this time, the model needs to be fine-tuned on the labeled data set to optimize its performance and effect.

[0106] The language processing model adopts in-context learning for supervised fine-tuning, that is, guiding the language processing model and giving an example to teach the language processing model what content it should output. For example, the language processing model outputs translation content, then the language processing model should be given the following input:

[0107] The user inputs to the language processing model: Please translate the following Chinese into English: Apple => apple; Do you think the language processing model is a useful tool? =>.

[0108] After the above preprocessing and a series of training, a financial dialogue model is obtained, and the financial dialogue model is used to realize financial anti-fraud knowledge question and answer and keyword extraction.

[0109] S20, constructing a digital virtual human database; the digital virtual human database includes a plurality of virtual humans appearing in historical real fraud and fraud cases.

[0110] In an embodiment, in order to quickly build a propaganda scene, the digital virtual human database is constructed in step S20, comprising:

[0111] 1) A plurality of virtual voices are constructed by using a speech synthesis technology.

[0112] The virtual voice mainly includes standard male voice, standard female voice, child voice and male emotional voice, etc.

[0113] 2) A plurality of virtual images are obtained by constructing the character models commonly appearing in historical real fraud and fraud cases.

[0114] The appearance of the virtual human image mainly includes police image, ordinary person image, old person image and middle-aged person image, etc. The character model can be a two-dimensional model, or a three-dimensional model constructed by using a three-dimensional modeling technology.

[0115] 3) A plurality of virtual actions are obtained by constructing virtual human actions according to the action characteristics of the characters appearing in historical real fraud and fraud cases. The virtual actions construct a plurality of body actions and facial expression actions according to the characteristics of the specific virtual human image.

[0116] 4) Combining virtual sound, virtual image and virtual action to obtain a plurality of virtual people. Combining virtual action with corresponding virtual image and virtual sound to obtain a specific virtual person.

[0117] In theory, various virtual images, virtual sounds and virtual actions can be randomly combined to obtain different virtual people, which facilitates rapid generation of case scenarios. Generally, appropriate virtual sound and virtual action are selected according to the characteristics of the virtual person to meet the characteristics of the real person.

[0118] S30, receiving a financial question input by a user, and inputting the financial question into a financial dialogue model to obtain propaganda information to be replied.

[0119] The user can input the financial question by voice or in text. If voice input is used, the embodiment needs to convert the voice into text. For example, the voice financial question obtained is converted into text by using the voice-to-text technology of Kedou Xunfei.

[0120] The financial dialogue model is trained according to a financial propaganda question and answer knowledge training set. The financial dialogue model can give appropriate propaganda information according to the financial question.

[0121] S40, determining a corresponding virtual person and virtual scene according to the financial question, and loading and displaying the virtual scene and the virtual person.

[0122] In an embodiment, the virtual person and the virtual scene are stored in combination with corresponding set keywords. According to the keywords, the storage path of the corresponding virtual person and virtual scene can be obtained.

[0123] In step S40, the corresponding virtual person and virtual scene are determined according to the financial question, which includes:

[0124] The keywords in the financial question are extracted by using a keyword extraction technology, and the corresponding virtual person and virtual scene are matched according to the keywords. The trained financial dialogue model can also be used for text keyword extraction.

[0125] The method of text keyword extraction by the financial dialogue model mainly includes the following steps:

[0126] Data preprocessing: data cleaning and preprocessing are performed on the text to be analyzed, irrelevant information and noise are removed, and keywords are extracted.

[0127] Keyword extraction: automatically identify keywords in the text using a language processing model and sort them.

[0128] Keyword filtering: filter according to the importance of the keywords, retain the number of keywords while ensuring the representativeness and expressiveness of the keywords.

[0129] Keyword display: The extracted keywords are presented in a visual form to help users better understand the text topic.

[0130] The storage path of the virtual person and the virtual scene retrieved according to the keywords extracted by the language processing model is retrieved to obtain the virtual person and the virtual scene.

[0131] In an embodiment, the virtual scene is obtained by the following method:

[0132] The corresponding scene data is obtained by analyzing historical real fraud and fraud cases.

[0133] The corresponding virtual scene is constructed by two-dimensional or three-dimensional modeling technology according to the scene data. The scene data includes the sites appearing in the historical real fraud and fraud cases, such as house buildings, street buildings, greenery, and police rooms, etc.

[0134] S50, driving the virtual person to perform voice output propaganda information.

[0135] The propaganda information is converted into voice by using the text-to-speech technology of the company, and the virtual action played by the virtual person is controlled, giving a voice and a full display effect.

[0136] In an embodiment, the method further comprises: constructing a financial fraud and fraud behavior database. The financial fraud and fraud behavior database stores historical real fraud and fraud cases.

[0137] Specifically, the financial fraud and fraud database is constructed by the following method:

[0138] The scene script, scene information, and each scene role are obtained by analyzing historical real fraud and fraud cases; the dialogues of each scene role are described in the scene script according to the scene order. Exemplarily, the scene script is as shown in Figure 2 .

[0139] The scene script, scene information, and each scene role constitute a fraud and fraud case, and are stored in a fraud and fraud case data table.

[0140] The historical real fraud and fraud cases include: virtual network loan fraud, such as impersonating a company credit APP, inducing customers to download a fake APP, obtaining personal information, and fabricating service fees, security deposits, etc. to attempt to defraud customers' property. There are also cases of impersonating well-known financial companies, such as impersonating the official customer service of a certain financial company, accurately stating the customer's personal information, and reminding the customer to deposit a security deposit, and activating the account to normally loan interest, inducing the fraud to clear the debt, and transferring the property. There are also cases of impersonating public security organs for fraud, such as impersonating the police and requiring the customer to transfer the funds online to the police's safe account. In this embodiment, the roles and scenes in the above cases are identified, and the corresponding virtual person and virtual scene can be matched according to the corresponding identification. For example, the police and the fraudster are identified in the form of character number. When using this case, the corresponding virtual person can be obtained according to the character number.

[0141] Further, in addition to the question-and-answer financial knowledge propaganda, the method further includes using a polling method to propagate historical real fraud and fraud cases, or receiving user voice instructions and propagating historical real fraud and fraud cases according to keywords extracted therefrom.

[0142] Specifically, the user instruction or the set instruction uses a financial dialogue model to obtain the corresponding fraud and fraud case from the financial fraud and fraud database. The above-mentioned user instruction includes: 1) user voice input information, using a financial dialogue model to extract keywords according to the text sequence corresponding to the voice information, and determining the fraud and fraud case according to the keywords. 2) The instruction of selecting the fraud and fraud case is input by the interface button. The set instruction includes obtaining the fraud and fraud case according to the set order for polling type playback. The corresponding virtual person is obtained according to the scene role in the fraud and fraud case, and the corresponding virtual scene is obtained according to the scene information in the fraud and fraud case. For example, the scene role in the fraud and fraud case is identified by a keyword, for example, a police officer, and the virtual person resource is also saved by a keyword. Finally, the corresponding virtual person of the police officer can be obtained according to the keyword in the set path. Load and display the virtual scene and the virtual person, and drive each virtual person to explain the case according to the scene script in the fraud and fraud case.

[0143] Further, in order to facilitate the expansion of fraud and fraud cases, the present application also abstracts a general function for reading fraud and fraud cases to obtain scene scripts, scene information, and scene roles. Then, a function for matching scene information and scene roles is abstracted to quickly obtain corresponding virtual scene and virtual person model resources.

[0144] The financial fraud prevention knowledge propaganda method of the present application will be introduced in combination with a specific example as shown in Figure 3 The following steps are included:

[0145] S110, a language processing model is constructed and trained to obtain a financial dialogue model;

[0146] According to historical real fraud and fraud cases, a digital virtual person database is constructed to obtain a plurality of virtual persons;

[0147] According to historical real fraud and fraud cases, a financial fraud and fraud database is constructed to obtain a plurality of fraud and fraud cases;

[0148] According to historical real fraud and fraud cases, a corresponding scene model is constructed to obtain a plurality of virtual scenes.

[0149] S120, the user selects a question and answer type propaganda mode, and proposes a financial question through a voice mode; the voice form financial question is obtained, and a voice to text technology is used to convert the voice data into a text form financial question.

[0150] S130, the text form financial question is input into the financial dialogue model to obtain propaganda information to be replied.

[0151] S140, the financial dialogue model is used to extract keywords from the text form financial question. According to the keywords, virtual scenes and virtual persons are matched and loaded to display the virtual scenes and virtual persons.

[0152] S150, a text to voice technology is used to convert the propaganda information into a voice form propaganda information, and the speech virtual action of the virtual person is played synchronously.

[0153] S160, if the user switches to a polling propaganda mode, or there is no voice interaction for a long time, the polling propaganda mode is automatically switched.

[0154] S170, according to a set instruction, a fraud and fraud case is obtained from the financial fraud and fraud database, a corresponding virtual person is obtained according to a scene role in the fraud and fraud case, and a corresponding virtual scene is obtained according to scene information in the fraud and fraud case. Then, the virtual scene and the virtual person are loaded and displayed, and each virtual person is driven to explain the case according to the scene script in the fraud and fraud case.

[0155] The image and voice of the digital virtual person of the application can effectively improve the acceptance and observation interest of the customer, and further improve the efficiency and effect of the bank financial propaganda.

[0156] The application adopts a voice interaction mode, the user can obtain answers through simple voice questions, avoid complicated operations, and improve customer experience.

[0157] The application also constructs a financial fraud and fraud database, obtains fraud and fraud cases in a polling mode, and explains the fraud and fraud cases through virtual persons, which vividly and vividly propagates financial fraud knowledge.

[0158] The present application can continuously update the financial fraud and fraud behavior database according to market changes and customer needs, ensuring that the model always has the latest financial knowledge and information.

[0159] The present application can quickly expand fraud and fraud cases by abstracting general functions for reading fraud and fraud cases and functions for matching scene information and scene roles, without the need to change the code again.

[0160] Figure 4 A structural schematic diagram of a financial fraud prevention knowledge propaganda device according to an embodiment of the present application is shown. The exemplary financial fraud prevention knowledge propaganda device includes a virtual person database construction module 410, a financial dialogue module 420, a scene display loading module 430, and a propaganda driving module 440.

[0161] The virtual person database construction module is configured to construct a digital virtual person database, and the digital virtual person database includes a plurality of virtual persons appearing in historical real fraud and fraud cases.

[0162] The financial dialogue module is configured to receive a financial question input by a user, input the financial question into a financial dialogue model to obtain propaganda information to be replied, and obtain the propaganda information to be replied by a language processing model constructed by training.

[0163] The scene display loading module is configured to determine a corresponding virtual person and a virtual scene according to the financial question, and load and display the virtual scene and the virtual person.

[0164] The propaganda driving module is configured to drive the virtual person to output the propaganda information by voice.

[0165] It can be understood that the device of the present embodiment corresponds to the financial fraud prevention knowledge propaganda method of the above-mentioned embodiment, and the optional items in the above-mentioned embodiment are also applicable to the present embodiment, so the description is not repeated here.

[0166] The present application also provides a terminal device, which includes a processor and a memory, wherein the memory stores a computer program, and the processor runs the computer program, so that the terminal device executes the functions of the financial fraud prevention knowledge propaganda method or each module of the financial fraud prevention knowledge propaganda device.

[0167] The processor can be an integrated circuit chip with a processing capability of signals. The processor can be a general processor, including a central processing unit (CPU), a graphics processing unit (GPU), and a network processor (NP), a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, a discrete gate or transistor logic device, a discrete hardware component, at least one of the above. The general processor can be a microprocessor or the processor can be any conventional processor, etc., which can implement or execute the disclosed methods, steps and logic block diagrams in the embodiments of the present application.

[0168] The memory can be, but is not limited to, a random access memory (RAM), a read only memory (ROM), a programmable read only memory (PROM), an erasable programmable read only memory (EPROM), an electrically erasable programmable read only memory (EEPROM) and the like. The memory is used to store a computer program. After receiving an execution instruction, the processor can execute the computer program accordingly.

[0169] The present application also provides a readable storage medium for storing the computer program used in the terminal device.

[0170] It should be understood that all the functional modules or units in the embodiments of the present application can be integrated or can exist alone, and two or more functional modules or units can be integrated to form an independent part.

[0171] In addition, each functional module or unit in the embodiments of the present application can be integrated together to form an independent part, or each functional module can exist alone, or two or more functional modules can be integrated to form an independent part.

[0172] If the functions are realized in the form of software function modules and sold or used as an independent product, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application can be embodied in the form of a software product, and the computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a smart phone, a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0173] The above describes only the specific implementation of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered within the protection scope of the present application.< / unk> < / end> < / start> < / unk> < / end> < / start>

Claims

1. A method for disseminating financial fraud prevention knowledge, characterized in that, include: A language processing model is constructed and trained to obtain a financial dialogue model; wherein, training the language processing model includes: using an unlabeled text dataset, training the language processing model using an unsupervised pre-training method, and using a labeled dataset to perform supervised fine-tuning of the unsupervised pre-trained language processing model using a context learning approach. Construct a digital virtual human database; the digital virtual human database includes various virtual humans that appeared in historical real fraud and scam cases; Receive financial questions input by users, input the financial questions into the financial dialogue model to obtain the promotional information that needs to be responded to; Based on the financial problem, determine the corresponding virtual person and virtual scene, and load and display the virtual scene and the virtual person; Drive the virtual human to output the promotional information via voice; The method further includes: The system receives user instructions and / or set instructions, uses the financial dialogue model to retrieve corresponding fraud and deception cases from the financial fraud and deception database; retrieves corresponding virtual characters based on the scenario roles in the fraud and deception cases, and retrieves corresponding virtual scenes based on the scenario information in the fraud and deception cases; loads and displays the virtual scenes and the virtual characters, and drives each virtual character to explain the case according to the scenario script in the fraud and deception cases.

2. The method for disseminating financial fraud prevention knowledge according to claim 1, characterized in that, The financial fraud and deception database was constructed using the following method: By analyzing real historical fraud and scam cases, scenario scripts, scene information, and roles in each scenario are obtained; the scenario scripts describe the dialogue of each role in each scenario according to the sequence of the scenarios. The scenario script, scenario information, and each scenario role constitute a fraud and scam case, which is then stored in a fraud and scam case data table.

3. The method for disseminating financial fraud prevention knowledge according to claim 1, characterized in that, The construction of the digital virtual human database includes: Various virtual voices were created using speech synthesis technology; Models of characters frequently appearing in the aforementioned historical fraud and scam cases were constructed to obtain multiple virtual characters; Based on the characteristics of human movements in real historical fraud and scam cases, virtual human movements are constructed to obtain multiple sets of virtual movements. By combining the virtual voice, the virtual image, and the virtual actions, a variety of virtual humans can be obtained.

4. The method for disseminating financial fraud prevention knowledge according to claim 1, characterized in that, The virtual scene was obtained using the following method: Analyze the aforementioned historical real fraud and scam cases to obtain corresponding scenario data; The corresponding virtual scene is constructed using two-dimensional and / or three-dimensional modeling techniques based on the scene data.

5. The method for disseminating financial fraud prevention knowledge according to claim 1, characterized in that, The virtual person and the virtual scene are both stored in conjunction with corresponding set keywords; The step of determining the corresponding virtual person and virtual scenario based on the financial problem includes: Keyword extraction technology is used to extract keywords from the financial question, and the corresponding virtual person and virtual scene are matched based on the keywords.

6. The method for disseminating financial fraud prevention knowledge according to any one of claims 1 to 5, characterized in that, The language processing model includes a transformer encoder and an autoregressive decoder; The converter encoder includes N coding layers, where N > 1, and the coding layers include a multi-head self-attention mechanism and a feedforward neural network; The autoregressive decoder includes N decoding layers, which include: a multi-head self-attention mechanism, a multi-head attention mechanism, and a feedforward neural network.

7. A financial fraud prevention knowledge dissemination device, characterized in that, include: The virtual human database construction module is used to build a digital virtual human database; The digital virtual human database includes various virtual humans that appeared in historical real fraud and scam cases; the training of the language processing model includes: using an unlabeled text dataset, training the language processing model using an unsupervised pre-training method, and using a labeled dataset to perform supervised fine-tuning of the unsupervised pre-trained language processing model using a context learning approach. The financial dialogue module is used to receive financial questions input by users, input the financial questions into the financial dialogue model to obtain the promotional information that needs to be responded to; the financial dialogue model is obtained by training and constructing a language processing model; The scene display loading module is used to determine the corresponding virtual person and virtual scene based on the financial question, and to load and display the virtual scene and the virtual person; The advocacy-driven module is used to drive the virtual human to output the advocacy information via voice. The advocacy-driven module is also used to: receive user instructions and / or pre-defined instructions to retrieve corresponding fraud and deception cases from the financial fraud and deception database using the financial dialogue model; retrieve corresponding virtual human characters based on the scenario roles in the fraud and deception cases; retrieve corresponding virtual scenarios based on the scenario information in the fraud and deception cases; load and display the virtual scenarios and the virtual human characters; and drive each virtual human character to explain the case according to the scenario script in the fraud and deception cases.

8. A terminal device, characterized in that, The terminal device includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the financial fraud prevention knowledge dissemination method according to any one of claims 1-6.

9. A readable storage medium, characterized in that, It stores a computer program, which, when executed on a processor, implements the financial fraud prevention knowledge dissemination method according to any one of claims 1-6.

Images