Early warning method and device for offline transaction of business network and program product

Abstract

The invention discloses an early warning method and device for offline transaction of a business outlet and a program product, relates to the field of artificial intelligence and is applied to the field of financial science and technology, and the method comprises the steps: collecting the input information of a customer, converting the input information into text data through a deep learning model, and obtaining a confidence coefficient; a first transaction feature is extracted from the text data, and the first transaction feature at least comprises a first transaction behavior and a first transaction amount; obtaining a to-be-executed instruction of the target object through the application program interface, and extracting a second transaction feature from the to-be-executed instruction; and comparing the first transaction behavior with the second transaction behavior, comparing the first transaction amount with the second transaction amount, and triggering transaction early warning according to a first comparison result and the confidence coefficient. According to the method and the device, the problem that in the related technology, when a business network teller processes a customer transaction, the transaction is mistakenly executed due to errors such as a transaction amount error or a keyboard mistakenly touched possibly caused by manual processing is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, and is applied to the field of financial technology. Specifically, the present application relates to a method and device for early warning of offline transactions in business outlets, and a program product. BACKGROUND

[0002] The counter transaction of commercial banks has long relied on manual operation. In particular, when dealing with cash deposit and withdrawal, transfer and remittance, etc., the teller needs to operate according to the customer's oral or written instructions. In this traditional mode, the operation accuracy depends entirely on the individual ability and concentration of the teller, which is easily affected by external factors, such as mishearing of numbers due to dialect differences, fatigue caused by long hours of work, or keyboard mis-touch of the operation interface, etc., thereby causing amount errors. Once the error occurs and is not corrected immediately, it will cause the customer to lose money, trigger customer dissatisfaction and even complaints, and damage the bank's reputation. At the same time, the post-transaction, correction or account adjustment not only consumes additional time and human resources, but also may attract the attention of the bank's internal audit and regulatory agencies, increasing compliance costs.

[0003] In view of the problem that in the related art, when a teller in a business outlet handles a customer transaction, due to manual processing, errors such as transaction amount errors or keyboard mis-touches may occur, resulting in incorrect transaction execution, no effective solution has been proposed so far. SUMMARY

[0004] The main purpose of the present application is to provide a method and device for early warning of offline transactions in business outlets, and a program product, to solve the problem that in the related art, when a teller in a business outlet handles a customer transaction, due to manual processing, errors such as transaction amount errors or keyboard mis-touches may occur, resulting in incorrect transaction execution.

[0005] In order to achieve the above-mentioned purpose, according to one aspect of the present application, a method for early warning of offline transactions in business outlets is provided, which comprises: detecting that a customer performs a target transaction at a service window in a business outlet, collecting input information of the customer, and converting the input information into text data through a deep learning model to obtain a confidence level; extracting a first transaction feature from the text data, wherein the first transaction feature at least includes a first transaction behavior and a first transaction amount; obtaining a to-be-executed instruction of a target object through an application programming interface, and extracting a second transaction feature from the to-be-executed instruction, wherein the second transaction feature at least includes a second transaction behavior and a second transaction amount, and the target object is a staff member in the business outlet who handles the target transaction; comparing the first transaction behavior with the second transaction behavior, comparing the first transaction amount with the second transaction amount, and triggering a transaction early warning according to a first comparison result and the confidence level.

[0006] Further, the business window is deployed with a microphone array to collect the input information of the customer, and convert the input information into text data through a deep learning model to obtain a confidence level, including: adopting an adaptive beamforming algorithm to control the microphone array to collect audio stream data and timestamp information of the audio stream data; adopting a preset algorithm to process the audio stream data to obtain processed audio stream data, wherein the preset algorithm at least includes one of the following: spectral subtraction, a speech enhancement algorithm based on a deep learning model; extracting an audio feature vector from the processed audio stream data through a deep learning model; converting the audio feature vector into text type data through a speech conversion model to obtain the text data and the confidence level output by the model.

[0007] Further, extracting an audio feature vector from the processed audio stream data through a deep learning model includes: cutting the processed audio stream data to obtain multiple segments of audio data; extracting acoustic features from the multiple segments of audio data based on the timestamp information to obtain a time sequence feature sequence; extracting local acoustic features from the time sequence feature sequence through a convolutional neural network to obtain a first feature vector; inputting the first feature vector into a transformer model to extract context-dependent information through the transformer model to obtain the audio feature vector.

[0008] Further, extracting a first transaction feature from the text data includes: performing text preprocessing on the text data to obtain processed text data, wherein the text preprocessing at least includes one of the following: removing redundant characters, punctuation symbol standardization, full-width or half-width conversion, speech correction processing, and text standardization; analyzing the processed text data based on a large language model to identify a customer transaction intent and a first amount feature, wherein the large language model is a model trained using a financial domain knowledge base; performing standardization processing on the first amount feature through the large language model to obtain a second amount feature; fusing the customer transaction intent and the second amount feature through the large language model, and extracting the first transaction behavior, the first transaction amount, and a transaction amount type from the fused data.

[0009] Further, after the processed text data is parsed based on the large language model to identify the customer transaction intention and the first amount feature, the method further comprises: identifying, based on the large language model, a third transaction feature of the target transaction from the processed text data, wherein the third transaction feature at least includes one of the following: transaction account information, transaction party information, and bank information; after obtaining the to-be-executed instruction of the target object through the application programming interface, the method further comprises: extracting a fourth transaction feature from the to-be-executed instruction, wherein the fourth transaction feature at least includes one of the following: transaction account information, transaction party information, and bank information; comparing each feature in the third transaction feature with the corresponding feature in the fourth transaction feature, and triggering a transaction warning according to a second comparison result.

[0010] Further, triggering a transaction warning according to the first comparison result and the confidence level comprises: determining that a first condition is met when the first comparison result indicates that the first transaction behavior and the second transaction behavior are different; determining that a second condition is met when a difference between the first transaction amount and the second transaction amount is greater than a first difference threshold, wherein the first difference threshold is calculated based on the first transaction amount; determining that a third condition is met when the confidence level is lower than a preset confidence threshold; and triggering a transaction warning when the first condition, the second condition, or the third condition is met.

[0011] Further, triggering a transaction warning when the first condition, the second condition, or the third condition is met comprises: pausing execution of the to-be-executed instruction for a preset time period and prompting the target object to adjust the to-be-executed instruction through voice broadcast when the first condition is met; displaying a pop-up window in a device operated by the target object to prompt the target object to adjust the to-be-executed instruction when the second condition is met and the difference is less than or equal to a second difference threshold, wherein the second difference threshold is calculated based on the first transaction amount, and the second difference threshold is greater than the first difference threshold; and pausing execution of the target transaction and sending transaction information of the target transaction to a target business system when the second condition is met and the difference is greater than the second difference threshold, wherein risk information existing in the target transaction is identified through the target business system.

[0012] Further, the business window is deployed with a display terminal, and before the first transaction feature is extracted from the text data, the method further comprises: detecting that the display terminal receives a customer instruction, and collecting the text data input by the customer through the display terminal and collecting timestamp information of the text data.

[0013] In order to achieve the above-mentioned object, according to another aspect of the present application, a transaction early warning device for an offline transaction of a business site is provided, which comprises: a first acquisition unit configured to acquire input information of a customer in the case that a target transaction of the customer is detected at a service window of the business site, and convert the input information into text data through a deep learning model to obtain a confidence level; a first extraction unit configured to extract first transaction features from the text data, wherein the first transaction features at least include a first transaction behavior and a first transaction amount; a second extraction unit configured to obtain a to-be-executed instruction of a target object through an application programming interface, and extract second transaction features from the to-be-executed instruction, wherein the second transaction features at least include a second transaction behavior and a second transaction amount, and the target object is a staff member of the business site who handles the target transaction; and a comparison unit configured to compare the first transaction behavior with the second transaction behavior, compare the first transaction amount with the second transaction amount, and trigger a transaction early warning according to a first comparison result and the confidence level.

[0014] Further, the service window is provided with a microphone array, and the first acquisition unit comprises: an acquisition subunit configured to control the microphone array to acquire audio stream data and timestamp information of the audio stream data by using an adaptive beamforming algorithm; a first processing subunit configured to process the audio stream data by using a preset algorithm to obtain processed audio stream data, wherein the preset algorithm at least includes one of the following: a spectral subtraction method, and a speech enhancement algorithm based on a deep learning model; a first extraction subunit configured to extract an audio feature vector from the processed audio stream data by using a deep learning model; and a conversion subunit configured to convert the audio feature vector into text type data by using a speech conversion model to obtain the text data output by the model and the confidence level.

[0015] Further, the extraction subunit comprises: a cutting module configured to cut the processed audio stream data to obtain multiple pieces of audio data; a first extraction module configured to extract acoustic features from the multiple pieces of audio data based on the timestamp information to obtain a time sequence feature sequence; a second extraction module configured to extract local acoustic features from the time sequence feature sequence by using a convolutional neural network to obtain a first feature vector; and a third extraction module configured to input the first feature vector into a transformer model, extract context-dependent information by using the transformer model, and obtain the audio feature vector.

[0016] Further, the first extraction unit comprises: a second processing subunit, configured to perform text preprocessing on the text data to obtain processed text data, wherein the text preprocessing comprises at least one of the following: removing redundant characters, standardizing punctuation marks, converting full-width or half-width, speech correction processing, and text standardization; an analysis subunit, configured to analyze the processed text data based on a large language model to identify a customer transaction intention and a first amount feature, wherein the large language model is a model trained based on a financial domain knowledge base; a third processing subunit, configured to standardize the first amount feature based on the large language model to obtain a second amount feature; and a second extraction subunit, configured to fuse the customer transaction intention and the second amount feature based on the large language model, and extract the first transaction behavior, the first transaction amount, and a transaction amount type from the fused data.

[0017] Further, the first extraction unit further comprises: an identification subunit, configured to identify a third transaction feature of the target transaction based on the large language model after analyzing the processed text data based on the large language model to identify a customer transaction intention and a first amount feature, wherein the third transaction feature comprises at least one of the following: transaction account information, transaction party information, and bank information; the device further comprises: a third extraction subunit, configured to extract a fourth transaction feature from a to-be-executed instruction of a target object after obtaining the to-be-executed instruction through an application programming interface, wherein the fourth transaction feature comprises at least one of the following: transaction account information, transaction party information, and bank information; and a comparison subunit, configured to compare each feature in the third transaction feature with a corresponding feature in the fourth transaction feature, and trigger a transaction warning based on a second comparison result.

[0018] Further, the comparison unit comprises: a first determination subunit, configured to determine that a first condition is met when the first comparison result indicates that the first transaction behavior and the second transaction behavior are different; a second determination subunit, configured to determine that a second condition is met when the first comparison result indicates that a difference between the first transaction amount and the second transaction amount is greater than a first difference threshold value, wherein the first difference threshold value is calculated based on the first transaction amount; a third determination subunit, configured to determine that a third condition is met when the confidence level is lower than a preset confidence threshold value; and a warning subunit, configured to trigger a transaction warning when the first condition, the second condition, or the third condition is met.

[0019] Further, the early warning subunit comprises: a first processing module, configured to, in the case that the first condition is met, suspend execution of the to-be-executed instruction within a preset time period, and prompt the target object to adjust the to-be-executed instruction in the form of voice broadcast; a second processing module, configured to, in the case that the second condition is met and the difference is less than or equal to a second difference threshold, display a pop-up window in the device operated by the target object to prompt the target object to adjust the to-be-executed instruction, wherein the second difference threshold is calculated based on the first transaction amount, and the second difference threshold is greater than the first difference threshold; and a third processing module, configured to, in the case that the second condition is met and the difference is greater than the second difference threshold, suspend execution of the target transaction, and send transaction information of the target transaction to a target business system, wherein the target business system is used to identify risk information existing in the target transaction.

[0020] Further, the business window is deployed with a display terminal, and the device further comprises a second acquisition unit, configured to, before extracting the first transaction feature from the text data, acquire the text data input by the customer through the display terminal when it is detected that the display terminal receives a customer instruction, and acquire timestamp information of the text data.

[0021] To achieve the above object, according to an aspect of the present application, there is provided a computer program product comprising a computer program which, when executed by a processor, implements any of the above-mentioned early warning methods for offline transactions in business outlets. The computer program, when executed by the processor, implements the steps of the early warning method for offline transactions in business outlets described in the embodiments of the present application.

[0022] To achieve the above object, according to an aspect of the present application, there is provided a computer-readable storage medium comprising stored computer instructions, wherein the computer instructions, when executed by a processor, implement any of the above-mentioned early warning methods for offline transactions in business outlets.

[0023] To achieve the above object, according to an aspect of the present application, there is provided an electronic device comprising one or more processors and a memory, the memory being configured to store one or more programs, wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to implement any of the above-mentioned early warning methods for offline transactions in business outlets.

[0024] In the embodiment of the present application, when a customer conducts a target transaction at a service window in a service site, the input information of the customer is collected, and the input information is converted into text data by a deep learning model to obtain a confidence level; first transaction features are extracted from the text data, wherein the first transaction features at least include a first transaction behavior and a first transaction amount; a target object is obtained by an application programming interface, and second transaction features are extracted from the to-be-executed instruction, wherein the second transaction features at least include a second transaction behavior and a second transaction amount, and the target object is a staff member in the service site who handles the target transaction; the first transaction behavior is compared with the second transaction behavior, the first transaction amount is compared with the second transaction amount, and a transaction early warning is triggered according to a first comparison result and the confidence level, thereby solving the technical problem that a transaction error occurs due to a mistake in transaction amount or a keyboard error when a teller in a service site handles a customer transaction.

[0025] The input information captured in real time is processed by the deep learning model, the input information when a customer conducts a target transaction can be captured in real time, the input information is converted into clear text data, and the confidence level of the recognition result is calculated, thereby realizing an effective conversion from unstructured information to structured data and achieving the purpose of accurate information analysis.

[0026] Meanwhile, the first transaction behavior is compared with the second transaction behavior, the first transaction amount is verified with the second transaction amount, a transaction early warning is triggered according to a first comparison result and in combination with the recognition confidence level, so that the system can discover and warn any inconsistency or potential error in time before a transaction is executed, thereby realizing active prevention and control of operational risks. This series of steps collectively constitute a teller counter transaction and amount intelligent verification technology based on natural language processing, which greatly improves the automation level and accuracy of the transaction process, reduces the fund risk caused by human errors, and optimizes the quality of bank teller counter services and operational efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0027] The accompanying drawings, which form a part of the present application, are intended to provide further understanding of the present application, and the illustrative embodiments of the present application and their description serve the purpose of explaining the present application. The accompanying drawings in accordance with embodiments of the present application are as follows:

[0028] Figure 1 is a hardware structure block diagram of a computer terminal (or mobile device) for implementing a warning method for offline transactions in a service site according to the embodiment 1 of the present application;

[0029] Figure 2 is a flowchart of an optional warning method for offline transactions in a service site according to the embodiment 1 of the present application;

[0030] Figure 3 is an optional intelligent verification flowchart of a counter transaction based on a natural language processing technology according to Embodiment 1 of the present application;

[0031] Figure 4 is an optional transaction information comparison flowchart in the intelligent verification flowchart of the counter transaction according to Embodiment 1 of the present application;

[0032] Figure 5 is a schematic diagram of an early warning device for an offline transaction of a business site according to Embodiment 2 of the present application;

[0033] Figure 6 is a schematic diagram of an early warning electronic device for an offline transaction of a business site according to Embodiment 3 of the present application. DETAILED DESCRIPTION

[0034] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0035] It should be noted that the processing method, device, storage medium and electronic device in the present application file and the methods and devices thereof can be used in the field of financial technology to improve the accuracy of the counter in executing transaction business operations during the process of handling customer transactions. The processing method, device, storage medium and electronic device of the present application file can also be used in any field other than the field of financial technology, and the application field of the processing method, device, storage medium and electronic device of the present application file is not limited.

[0036] It should be noted that the information (including but not limited to user device information, user personal information, collected data, used data, generated data, processed data, etc.) and data (including but not limited to data for analysis, stored data, displayed data, collected information, used information, generated information, processed information, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of related data comply with relevant laws, regulations and standards of relevant countries and regions, necessary security measures are taken, public order and good customs are not violated, and appropriate operation portals are provided for users to choose authorization or refusal. For example, the system and the interface between the related users or institutions provide appropriate operation portals for users to choose to agree or refuse the automatic decision result; if the user chooses to refuse, the expert decision process is entered.

[0037] Embodiment 1

[0038] According to the embodiment of the present application, a method for early warning of offline transactions of business outlets is also provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that here.

[0039] The method provided by the embodiment 1 of the present application can be executed in a mobile terminal, a computer terminal or similar computing device. Figure 1 is a hardware structure block diagram of a computer terminal (or mobile device) for implementing the early warning method of offline transactions of business outlets according to the embodiment 1 of the present application. As shown in Figure 1 , the computer terminal 10 (or mobile device) can include one or more (in the figure, 102a, 102b, …, 102n are used to show) processors 102 (the processor 102 can include but not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 104 for storing data, and a transmission device 106 for communication function. In addition, it can also include a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which can be included as one of the ports of the BUS bus), a network interface, a power supply and / or a camera. Those skilled in the art can understand that Figure 1 The structure shown is only schematic, which does not limit the structure of the above-mentioned electronic device. For example, the computer terminal 10 can also include more or less components than those shown in Figure 1 , or have a different configuration from Figure 1 .

[0040] It should be noted that the one or more processors 102 and / or other data processing circuits described above can be referred to as "data processing circuits" herein. The data processing circuit can be embodied in whole or in part as software, hardware, firmware or any other combination. In addition, the data processing circuit can be a single independent processing module, or any one of the other elements combined into the computer terminal 10 (or mobile device) in whole or in part. As referred to in the embodiments of the present application, the data processing circuit as a kind of processor control (for example, the selection of the variable resistance terminal path connected with the interface).

[0041] The memory 104 can be used to store software programs of application software and modules, such as program instructions / data storage devices corresponding to the pre-warning method for offline transactions of business outlets in the embodiments of the present application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, that is, implements the pre-warning method for offline transactions of business outlets as described above. The memory 104 can include a high-speed random access memory, and can also include a non-volatile memory, such as one or more magnetic storage devices, flash memories, or other non-volatile solid-state memories. In some examples, the memory 104 can further include memories remotely arranged with respect to the processor 102, which can be connected to the computer terminal 10 through a network. Examples of the above-mentioned network include but are not limited to the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0042] The transmission device 106 is used to receive or send data via a network. Specific examples of the above-mentioned network can include a wireless network provided by a communication provider of the computer terminal 10. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to be able to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency (Radio Frequency, RF) module, which is used to communicate with the Internet in a wireless manner.

[0043] The display can be, for example, a touch screen type liquid crystal display (LCD), which can enable a user to interact with the user interface of the computer terminal 10 (or a mobile device).

[0044] Under the above-mentioned operating environment, the present application provides a pre-warning method for offline transactions of business outlets as shown in Figure 2 Figure 2 is a flowchart of an optional pre-warning method for offline transactions of business outlets according to Embodiment 1 of the present application.

[0045] In step S201, when the customer performs a target transaction at the service window of the business outlet, the input information of the customer is collected, and the input information is converted into text data through a deep learning model to obtain a confidence degree.

[0046] In this embodiment 1, when the target transaction is performed at the service window of the business outlet, the input information of the customer is first captured by the deployed collection device. This information is then sent to the deep learning model for analysis, and the model converts the customer input into structured text data. In this conversion process, the deep learning model will attach a confidence score to each recognition result, which quantifies the reliability of the model output result and serves as an important basis for evaluating the accuracy of the transaction information in subsequent steps.​

[0047] Step S202, extracting a first transaction feature from the text data, wherein the first transaction feature at least includes a first transaction behavior and a first transaction amount.

[0048] In this embodiment 1, a deep semantic analysis is performed on the received text data, aiming to extract key transaction information therefrom. The first transaction feature extracted from the text data mainly includes two core components: the first transaction behavior and the first transaction amount. The first transaction behavior refers to the type of transaction that the customer intends to perform, covering specific operations such as deposit, withdrawal, transfer or payment, etc.; while the first transaction amount is the specific value of the customer's transaction, which needs to be accurately identified and converted into a standardized numerical format for subsequent amount checking.

[0049] Step S203, obtaining a to-be-executed instruction of a target object through an application programming interface, and extracting a second transaction feature from the to-be-executed instruction, wherein the second transaction feature at least includes a second transaction behavior and a second transaction amount, and the target object is a staff member in a business outlet who handles target transactions.

[0050] In this embodiment 1, the business processing system used by the staff member in the business outlet is accessed in real time through an application programming interface (API) to obtain the to-be-executed transaction instruction. This instruction set is regarded as the to-be-executed instruction, which contains specific transaction details. From the to-be-executed instruction, the second transaction feature is extracted, which includes the basic attributes of the transaction: the second transaction behavior and the second transaction amount. The second transaction behavior corresponds to the transaction type selected by the staff member, such as deposit, withdrawal, etc.; while the second transaction amount is the specific transaction value input by the staff member in the system.

[0051] Step S204, comparing the first transaction behavior with the second transaction behavior, comparing the first transaction amount with the second transaction amount, and triggering a transaction warning according to the first comparison result and a confidence level.

[0052] In this embodiment 1, in order to ensure the accuracy of the transaction, a dual verification process of transaction behavior and transaction amount needs to be performed. First, the first transaction behavior expressed by the customer is compared with the second transaction behavior input by the staff member to ensure that they match. Second, the first transaction amount, i.e. the exact transaction value indicated by the customer, is accurately compared with the second transaction amount, i.e. the amount input by the staff member in the system. The comparison result is defined as the first comparison result, which is used to evaluate the consistency of the transaction information.

[0053] On the basis of the first comparison result, in combination with the confidence index obtained from previous speech recognition, whether to trigger a transaction warning is determined according to a preset rule, so as to effectively prevent errors before transaction execution and ensure the safety and accuracy of financial operations or financial transactions.

[0054] Optionally, in the transaction pre-warning method of the business outlet offline transaction provided in Embodiment 1 of the present application, a microphone array is deployed at the service window to collect the input information of the customer, and the input information is converted into text data through a deep learning model to obtain the confidence index, including: using an adaptive beamforming algorithm to control the microphone array to collect audio stream data and timestamp information of the audio stream data; using a preset algorithm to process the audio stream data to obtain processed audio stream data, wherein the preset algorithm at least includes one of the following: spectral subtraction, a speech enhancement algorithm based on a deep learning model; extracting an audio feature vector from the processed audio stream data through a deep learning model; converting the audio feature vector into text type data through a speech conversion model to obtain text data output by the model and the confidence index.

[0055] In Embodiment 1, in order to realize real-time and accurate speech information collection and processing, a microphone array is first installed at the service window. The design of the microphone array adopts an adaptive beamforming algorithm. Subsequently, the directivity of the microphone array is continuously monitored and adjusted to focus on collecting sounds from the customer, while recording the timestamp information of the sound signals.

[0056] Then, the collected raw audio stream data is preprocessed. This process involves applying spectral subtraction or a speech enhancement algorithm based on a deep learning model, aiming to reduce the impact of background noise and improve the clarity of the speech signal. By removing or weakening non-speech signal components such as device operation sound and environmental noise, the quality of the processed audio stream data is improved, creating favorable conditions for subsequent speech recognition.

[0057] Secondly, a deep learning model is used to extract features from the processed audio stream data. This model is usually based on the Conformer architecture, which is a neural network model used in Automatic Speech Recognition (ASR) systems. It combines the advantages of Convolutional Neural Networks (CNN) and Transformers, enabling better Sequence-to-Sequence modeling, effectively capturing local features and long-range dependencies in audio signals, forming an audio feature vector as input for the next step of text generation.

[0058] Finally, the audio feature vector is converted into text data through a specially designed speech conversion model. This model adds the grammatical structure of the language and the specific context of the bank transaction to the training set during the training process to train the model to accurately transcribe the audio stream content into text. According to the accuracy evaluation result of the speech conversion result of the model, a quantitative confidence index is output to evaluate the reliability of the transcription.

[0059] In an alternative embodiment, the microphone array captures customer voice instructions in a counter scene directionally, providing high-quality input for the subsequent step of converting voice to text. Exemplarily, a multi-channel directional microphone array is integrated at the counter workstation, including a standard USB directional microphone (120° pickup angle) fixed on the customer side of the counter. Real-time adaptive beamforming algorithms (e.g., Minimum Variance Distortionless Response (MVDR) and Generalized Sidelobe Canceler (GSC)) are used to dynamically track the direction of customer voice and generate a spatial filtering beam, significantly suppressing typical counter interference noise (e.g., keyboard tapping sound of the teller, equipment running noise (printer, currency counting machine), adjacent counter conversation, and hall environment background noise). The standard USB directional microphone is plug-and-play connected to the counter PC, and the audio stream is directly transmitted to the deep learning model. The standard USB directional microphone output interface supports mainstream voice recognition software development kits (Software Development Kit, SDK), and outputs real-time audio stream with timestamp.

[0060] Through the above steps, the efficient conversion from voice input to text data is achieved, and the confidence level of the recognition result can be quantified. This process not only can process customer voice in real time, but also can ensure the accuracy and reliability of the transcription, laying a solid foundation for subsequent transaction intent analysis and risk verification.

[0061] Optionally, in the transaction pre-warning method of the business site offline transaction provided in Embodiment 1 of the present application, the audio feature vector is extracted from the processed audio stream data through a deep learning model, including: cutting the processed audio stream data to obtain multiple audio data; extracting acoustic features from the multiple audio data based on timestamp information to obtain a time sequence feature sequence; extracting local acoustic features from the time sequence feature sequence through a convolutional neural network to obtain a first feature vector; inputting the first feature vector into a transformer model to extract context-dependent information through the transformer model to obtain the audio feature vector.

[0062] In this embodiment 1, in order to accurately identify the voice instructions of the customer in the bank counter environment, the pre-processed audio stream data is first cut. This process follows fixed sampling rate and frame length parameters, and the continuous audio stream is divided into a series of independent audio segments, aiming to prepare suitable data format for subsequent feature extraction, and to ensure that each segment can represent a relatively complete and independent voice unit.

[0063] Then, based on the timestamp information of each audio segment, acoustic features are extracted from the multiple audio data segments obtained by cutting using audio signal processing technology. The use of timestamps helps to accurately identify the order and position of each audio segment in the actual conversation, which is crucial for understanding the context. The extraction of acoustic features, including spectral analysis and energy calculation, generates a sequence of time sequence features representing the characteristics of the audio segment.

[0064] Among them, the representation of the characteristics of the audio segment can include: 80-dimensional Mel filter bank energy FBank, that is, a set of triangular filters distributed on the Mel frequency scale, which segments the frequency spectrum of the sound wave signal to obtain the output energy of each filter, that is, the FBank feature, wherein 80 dimensions represent that 80 such filters are used to obtain 80 energy values, thereby forming a vector describing the characteristics of the voice signal; 3-dimensional pitch feature usually contains fundamental frequency (F0), zero crossing rate of fundamental frequency and stability of fundamental frequency, which is very helpful for distinguishing different phonemes in speech, especially for identifying numbers and other words in languages with tone changes (such as Chinese); self-supervised feature, that is, learning meaningful feature representation from original audio signal through self-supervised learning without manually labeled data.

[0065] Secondly, the time sequence feature sequence is processed by a convolutional neural network to identify and extract local acoustic features, and a first feature vector is obtained. CNN is selected because of its excellent performance on images and audio signals, which can effectively capture local patterns in audio segments, such as phoneme and plosive features, which are crucial for speech recognition and differentiation, especially in cases where dialects or accents differ greatly. For example, through a multi-layer convolutional neural network (for example, depthwise separable convolution), local acoustic patterns (such as phoneme segments, plosive features) are efficiently extracted, capturing short-time invariance features of speech signals (such as consonant transient, vowel formant structure), and being robust to noise and channel differences.

[0066] Finally, the first feature vector is fed into a Transformer model, which leverages its powerful context modeling capabilities to further extract global dependency information from the audio features, generating the final audio feature vector. Transformers are renowned for their self-attention mechanism, which enables them to handle dependencies in long sequences of data, making them particularly effective in understanding coarticulation phenomena, intonation variations, and diversity in numerical expressions in customer speech. This ensures that the output feature vector comprehensively reflects the essence of the speech content. For example, using a Transformer encoder (multi-head self-attention mechanism and feed-forward network) to model long-range acoustic context dependencies addresses common issues in customer speech, such as numerical coarticulation (e.g., converting "three million five" to "35000"), intonation variations, and ambiguous pronunciation (e.g., distinguishing between "four" and "ten").

[0067] Through the above steps, the conversion from raw audio stream data to structured audio feature vectors is achieved, providing high-precision, semantically rich inputs for subsequent natural language processing, thereby improving the accuracy and robustness of transaction intent recognition, reducing transaction risks caused by speech recognition errors, and enhancing the security and efficiency of bank counter transactions.

[0068] Optionally, in the transaction pre-warning method of the business site offline transaction provided in Embodiment 1 of the present application, the first transaction feature is extracted from the text data, including: performing text preprocessing on the text data to obtain processed text data, wherein the text preprocessing at least includes one of the following: removing redundant characters, standardizing punctuation, converting full-width or half-width, speech correction processing, and text standardization; analyzing the processed text data based on a large language model to identify customer transaction intent and a first amount feature, wherein the large language model is a model trained using a financial domain knowledge base; performing standardization processing on the first amount feature through the large language model to obtain a second amount feature; and fusing the customer transaction intent and the second amount feature through the large language model, and extracting a first transaction behavior, a first transaction amount, and a transaction amount type from the fused data.

[0069] In Embodiment 1, in order to accurately analyze the text information provided by the customer in the bank counter transaction, the original text data is first preprocessed. This stage covers removing redundant characters, standardizing punctuation, unifying full-width or half-width character formats, and basic speech correction processing, ensuring the cleanliness and consistency of the text data. The preprocessing also includes text standardization, which converts colloquial expressions into standardized bank terminology, preparing for subsequent in-depth analysis.

[0070] Then, the pre-processed text data is deeply analyzed based on a pre-trained large language model. The model is trained based on a financial domain knowledge base, and its task is to identify and interpret the customer's transaction intent while accurately extracting the transaction-related amount features. This process involves the model's understanding of the syntax and semantics of the text data, ensuring that it can accurately map the customer's colloquial or non-standard expressions to the standard descriptions of bank transactions, such as parsing "store some money" as a "deposit" operation and converting "five thousand five" to the specific value "5500".

[0071] Secondly, the large language model further standardizes the extracted amount features to obtain second amount features. This standardization step aims to eliminate the diversity of amount expressions, such as converting "one million two thousand" to "12000" and "one hundred and five yuan" to "105.00", ensuring the consistency of the amount data format for subsequent accurate comparison and verification.

[0072] Finally, the customer transaction intent and the second amount features are fused by the large language model, and the first transaction behavior, the first transaction amount, and its amount type are accurately extracted from the fused data. The first transaction behavior can be a standardized financial service code or financial service name, and the amount type is associated with the transaction type, such as deposit principal, transfer amount, etc. The model uses its deep understanding in the financial field to accurately determine the transaction behavior (such as deposit, withdrawal, transfer) and the nature of the amount (such as principal, commission), while ensuring the correctness of the transaction amount value.

[0073] Through the above steps, the conversion from unstructured text data to structured transaction instructions is achieved, which can accurately and timely parse the customer's transaction intent and amount, while considering the diversity and complexity of amount expressions, ensuring the accuracy and consistency of information interpretation.

[0074] Optionally, in the transaction pre-warning method provided in Embodiment 1 of the present application, after the processed text data is analyzed based on the large language model to identify the customer's transaction intent and the first amount feature, the above method further includes: identifying the third transaction feature of the target transaction based on the large language model, wherein the third transaction feature at least includes one of the following: transaction account information, transaction party information, and opening bank information; after obtaining the to-be-executed instruction of the target object through the application programming interface, the above method further includes: extracting the fourth transaction feature from the to-be-executed instruction, wherein the fourth transaction feature at least includes one of the following: transaction account information, transaction party information, and opening bank information; comparing each feature in the third transaction feature with the corresponding feature in the fourth transaction feature, and triggering a transaction warning based on the second comparison result.

[0075] In this embodiment 1, in order to ensure the accuracy of the account and object information in the bank counter transaction, first, a large language model deeply trained by the financial field knowledge base is used to identify the preprocessed text data. This identification process aims to extract key third transaction features from the information provided by the customer, including but not limited to the last four digits of the transaction account number, the name of the transaction counterpart, and the opening bank information of the transaction account, aiming to ensure the completeness and accuracy of the transaction information.

[0076] Then, through the application program interface, the business instructions to be executed by the staff are obtained in real time from the bank counter transaction system. This instruction usually contains standardized transaction codes and detailed transaction parameters, including the fourth transaction features mentioned above, namely transaction account information, transaction party information, and opening bank information, ensuring direct comparison and reference with the customer's transaction intention.

[0077] Secondly, the third transaction features extracted from the customer text data are accurately compared with the fourth transaction features obtained directly from the clerk's to-be-executed instructions to generate a second comparison result. This comparison process covers all key information points such as account number, transaction object, and bank name, aiming to verify the consistency of transaction details and prevent transaction risks caused by information input errors.

[0078] Finally, based on the second comparison result, if inconsistencies are detected between the third transaction features and the fourth transaction features, a transaction warning mechanism is triggered. The triggering of the warning depends on the severity of the inconsistent information, such as minor differences in account numbers that may trigger a slight warning, while complete inconsistencies in transaction parties or opening banks may cause the transaction to be suspended until it is confirmed to be correct, ensuring the safety and compliance of the transaction process.

[0079] Through the above steps, the real-time verification of the account and object information in the customer's transaction instructions and the clerk's operation instructions is achieved, which can effectively identify and intercept potential transaction errors, improve the transparency and control of the transaction process, and also strengthen the bank's risk control ability and the safety of customer funds. The application of this technology significantly improves the front-end prevention and control effect of operational risk, avoids subsequent disputes and losses, and optimizes the processing efficiency and customer experience of bank counter business.

[0080] Optionally, in the transaction pre-warning method for offline transactions of business outlets provided in Embodiment 1 of the present application, the transaction pre-warning is triggered according to the first comparison result and the confidence level, including: when the first comparison result indicates that the first transaction behavior and the second transaction behavior are different, it is determined that the first condition is met; when the first comparison result indicates that the difference between the first transaction amount and the second transaction amount is greater than the first difference threshold, it is determined that the second condition is met, wherein the first difference threshold is calculated based on the first transaction amount; when the confidence level is lower than the preset confidence threshold, it is determined that the third condition is met; and the transaction pre-warning is triggered when the first condition, the second condition or the third condition is met.

[0081] In Embodiment 1, in order to identify and handle the possible risks in bank counter transactions in time, first, the transaction data obtained in real time is analyzed in depth. This analysis process includes comparing the transaction behavior expressed by the customer (the first transaction behavior) with the transaction operation entered by the clerk (the second transaction behavior), and determining whether the two are consistent through the first comparison result. If it is found that they are not consistent, it is immediately determined that the first condition is met, which means that there is a potential error or mismatch in the transaction behavior.

[0082] Then, the accuracy of the transaction amount is further evaluated, and the difference between the first transaction amount and the second transaction amount is analyzed through the first comparison result. If the difference exceeds the first difference threshold calculated based on the first transaction amount, it is considered that the second condition is met. The first difference threshold is dynamically set, which can flexibly adjust the triggering standard of pre-warning according to the size of the transaction amount, enhancing the flexibility and robustness of the system.

[0083] Secondly, considering the uncertainty of voice recognition, when the received confidence score is lower than the preset confidence threshold, it is also determined that the third condition is met. The confidence threshold is set based on statistical analysis of a large amount of voice data, which reflects the confidence level of the system on the conversion result of the customer's voice. A lower confidence level suggests that additional verification or manual review may be needed to ensure the accuracy of the amount.

[0084] Finally, when any of the above conditions is met, i.e. the first condition, the second condition or the third condition, the transaction pre-warning is triggered. The triggering of the pre-warning means that there is a risk point in the transaction, which needs to be further checked or the operation needs to be suspended to avoid potential loss of funds or other risks. The start of the pre-warning mechanism will immediately notify the clerk, prompting him to check the accuracy of the transaction information and make corrections or confirmations if necessary.

[0085] Through the above steps, real-time monitoring and verification of key information (transaction behavior and transaction amount) in bank counter transactions, as well as evaluation of the recognition result confidence, are achieved. This mechanism can effectively intercept confusion of transaction behavior, errors in amount entry, and uncertainty in the recognition process, ensuring the accuracy and security of transactions, reducing bank operational risks from the source, and improving the immediacy and effectiveness of risk prevention and control.

[0086] Optionally, in the transaction pre-warning method of the business site offline transaction provided in Embodiment 1 of the present application, the transaction pre-warning is triggered in the case where the first condition, the second condition or the third condition is met, including: in the case where the first condition is met, the execution of the to-be-executed instruction is suspended within a preset time period, and the target object is prompted to adjust the to-be-executed instruction in the form of voice broadcast; in the case where the second condition is met and the difference is less than or equal to the second difference threshold, a pop-up window is displayed in the device operated by the target object to prompt the target object to adjust the to-be-executed instruction, wherein the second difference threshold is calculated based on the first transaction amount, and the second difference threshold is greater than the first difference threshold; in the case where the second condition is met and the difference is greater than the second difference threshold, the execution of the target transaction is suspended, and the transaction information of the target transaction is sent to the target business system, wherein the risk information existing in the target transaction is identified through the target business system.

[0087] In Embodiment 1, in order to effectively manage the operational risk in bank counter transactions, the matching of transaction behavior is first monitored and judged. When it is detected that the transaction behavior expressed by the customer is inconsistent with the transaction behavior entered by the teller, it is determined that the first condition is met, and then the execution process of the to-be-executed transaction instruction is automatically suspended within a preset time period. This suspension operation aims to provide sufficient time for the teller to re-examine and adjust the transaction behavior, avoiding the risk of fund loss caused by incorrect transaction direction.

[0088] Then, the accuracy of the transaction amount is verified. If it is found that the amount difference exceeds the first difference threshold (for example, 5% of the transaction amount) set based on the size of the transaction amount, but is lower than the second difference threshold (for example, 50% of the transaction amount) set more leniently, the second condition is met. At this time, a pop-up window can be triggered on the device operated by the teller to directly prompt the teller that the amount entry may have errors, but given that the difference is within an acceptable range, the teller is allowed to manually confirm or adjust the transaction amount after verification to balance risk prevention and control and transaction efficiency.

[0089] Secondly, in the case of further expansion of the difference in transaction amount, exceeding the second difference threshold, more stringent control measures will be taken. This not only means suspending the execution of the target transaction, but also automatically sending detailed transaction information to the bank's risk control business system. Based on the received information, the risk control system further analyzes the potential risks in the transaction, such as dishonest behavior, operational errors, etc., and through comprehensive assessment, develops appropriate risk handling strategies to ensure the safety and compliance of the transaction.

[0090] Finally, through the above warning and control mechanism, according to the satisfaction of different conditions, the ladder response measures are taken from voice broadcast prompt, pop-up warning to transaction suspension and risk information reporting, thus forming a multi-level risk prevention and control system. The establishment of this system aims to monitor the accuracy of transaction behavior and amount in real time, quickly respond to potential risks, reduce customer financial losses caused by bank operational errors, and at the same time provide timely feedback and correction opportunities for tellers, improve the efficiency and quality of transaction processing.

[0091] Through the above steps, the proactive management and timely intervention of operational risks in the bank counter transaction scene are achieved, which significantly enhances the safety and customer satisfaction of bank transactions, reduces the risks caused by operational errors, and improves the risk prevention and control ability and business operation efficiency of the bank.

[0092] Optionally, in the pre-warning method for offline transactions in the business site provided in Embodiment 1 of the present application, a display terminal is deployed at the business window, and before extracting the first transaction feature from the text data, the above method further comprises: when detecting that the display terminal receives the client instruction, collecting the text data input by the client through the display terminal, and collecting the timestamp information of the text data.

[0093] In this embodiment 1, in order to ensure effective communication with the client, a display terminal can be deployed in the counter operation area. The display terminal is not only limited to presenting information, but also integrates text input functions such as touch screen keyboard or handwriting board, allowing deaf and mute clients to directly input their transaction needs and service requests, ensuring that all clients can communicate transaction details in a way suitable for themselves, such as account information, transaction type and specific transaction amount.

[0094] Then, the state of the display terminal is continuously monitored, and when it is detected that the client is inputting information through the display terminal, i.e. typing on the keyboard or handwriting input, the data collection function is immediately activated. This means that the display terminal will capture the text data input by the client, whether it is transaction type, amount, or account information, etc., to ensure the complete record of the original client instruction.

[0095] Secondly, in addition to the collection of text data, the timestamp information of each data input needs to be recorded. The acquisition of the timestamp is crucial for subsequent processing, as it can accurately identify the generation time of the customer's instruction, provide a time reference for real-time comparison and sequential execution of transactions, and ensure the accuracy and timeliness of transaction processing.

[0096] Finally, the collected text data along with its timestamp information will undergo further natural language processing and transaction behavior identification. This real-time transmission mechanism ensures that customer instructions can quickly reach the processing link, without delaying the execution of transactions, while also providing the necessary information foundation for subsequent risk warning and verification processes.

[0097] Through the above steps, even in the face of deaf-mute customers, the scheme can collect their handwritten or keyboard input text data through the display terminal, ensuring accurate collection and processing of transaction information. The application of this technology not only promotes barrier-free communication, but also maintains the accuracy and security of all customer transactions, while accurately capturing customer transaction request information, improving transaction instruction entry efficiency and reducing human error, providing detailed and accurate customer instruction data for subsequent transaction information verification, and improving the intelligent level of bank counter service and customer satisfaction.

[0098] Optionally, in Embodiment 1, Figure 3 is an optional intelligent verification process flowchart for counter transactions based on natural language processing technology provided by Embodiment 1 of the present application. As Figure 3 shown, first, the transaction information input by the user's voice or directly by text is received by the system, which converts the voice instruction into processable text data in real time through automatic speech recognition (ASR) technology, or directly processes the text information submitted by the user. Then, these text data are parsed by the natural language processing (NLP) module to accurately extract the customer's transaction behavior and transaction amount features, providing key information for subsequent verification steps. Secondly, through the application programming interface (API) provided by the bank counter transaction system, the transaction operations and amount data entered by the teller are synchronized in real time and compared with the NLP processing results. Finally, based on the comparison results, the consistency of the transaction behavior and the amount is evaluated, and once a mismatch or anomaly is found, the early warning mechanism is triggered immediately to notify the teller to check in time, thereby achieving proactive control of transaction risks. Through this series of steps, Figure 3 the complete process of intelligent verification technology for counter transactions from information collection to risk warning is demonstrated, ensuring the accuracy and security of transactions.

[0099] Optionally, in Embodiment 1, Figure 4 is an optional transaction information comparison process flowchart in the intelligent verification process for counter transactions provided by Embodiment 1 of the present application. As Figure 4As shown, first, the natural language processing (NLP) module analyzes the transaction information provided by the customer, whether it is text converted from voice recognition or directly inputted text, and extracts the key transaction behavior and amount data. Then, through the application program interface (API) provided by the financial institution, the system obtains the operation details entered by the teller in the transaction system in real time, including the transaction type and amount. Second, the system accurately compares the customer transaction information parsed by NLP with the teller-entered data obtained through the API interface, aiming to verify the consistency of the transaction content. Finally, according to the comparison result, if the two information are completely consistent, the transaction is allowed to continue execution; otherwise, if differences are found, the transaction warning will be triggered immediately, prompting the teller of possible operation errors, effectively preventing financial risks. This flowchart clearly shows how the intelligent verification technology ensures the accuracy of transaction information through technical means before the transaction is executed, thereby improving the safety and efficiency of bank counter services.

[0100] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown.

[0101] In summary, the pre-warning method for offline transactions in business outlets provided by the embodiments of the present application collects the input information of the customer when the customer conducts the target transaction at the service window in the business outlet, and converts the input information into text data through a deep learning model to obtain a confidence level; extracts first transaction features from the text data, wherein the first transaction features at least include a first transaction behavior and a first transaction amount; obtains a to-be-executed instruction of a target object through an application program interface, and extracts second transaction features from the to-be-executed instruction, wherein the second transaction features at least include a second transaction behavior and a second transaction amount, and the target object is a staff member in the business outlet who handles the target transaction; compares the first transaction behavior with the second transaction behavior, compares the first transaction amount with the second transaction amount, and triggers a transaction warning according to the first comparison result and the confidence level, solving the problem in the related art that the transaction is executed incorrectly due to errors such as transaction amount errors or keyboard mis-touches when the teller in the business outlet handles the transaction of the customer.

[0102] By processing the input information captured in real time through a deep learning model, the input information of the customer when conducting the target transaction can be captured in real time, converted into clear text data, and the confidence level of the recognition result can be calculated, realizing effective conversion from unstructured information to structured data, and further achieving the purpose of accurate information analysis.

[0103] Meanwhile, by comparing the extracted first transaction behavior with the second transaction behavior, verifying the first transaction amount with the second transaction amount, triggering a transaction warning according to the first comparison result and in combination with the recognition confidence, the system can discover and alert any inconsistency or potential error in time before the transaction is executed, thus realizing the active prevention and control of operational risks. This series of steps collectively constitute the intelligent verification technology for counter transactions and amounts based on natural language processing, greatly improving the automation level and accuracy of the transaction process, reducing the fund risks caused by human errors, and optimizing the counter service quality and operation efficiency of the bank.

[0104] Embodiment 2

[0105] The embodiment of the present application also provides a business site offline transaction early warning device. It should be noted that the business site offline transaction early warning device of the embodiment of the present application can be used to execute the business site offline transaction early warning method provided by the embodiment of the present application. The business site offline transaction early warning device provided by the embodiment of the present application is introduced as follows.

[0106] According to the embodiment of the present application, a device for implementing the business site offline transaction early warning method is also provided, Figure 5 is a schematic diagram of the business site offline transaction early warning device provided by the embodiment 2 of the present application. As shown in the figure, Figure 5 the device comprises a first acquisition unit 501, a first extraction unit 502, a second extraction unit 503 and a comparison unit 504.

[0107] Specifically, the first acquisition unit 501 is configured to acquire input information of a customer when the customer performs a target transaction at a service window of a business site, and convert the input information into text data through a deep learning model to obtain a confidence.

[0108] The first extraction unit 502 is configured to extract first transaction features from the text data, wherein the first transaction features at least include a first transaction behavior and a first transaction amount.

[0109] The second extraction unit 503 is configured to obtain a to-be-executed instruction of a target object through an application programming interface, and extract second transaction features from the to-be-executed instruction, wherein the second transaction features at least include a second transaction behavior and a second transaction amount, and the target object is a staff member who handles the target transaction in the business site.

[0110] The comparison unit 504 is configured to compare the first transaction behavior with the second transaction behavior, compare the first transaction amount with the second transaction amount, and trigger a transaction warning according to the first comparison result and the confidence.

[0111] The business outlet offline transaction early warning device provided by the embodiment of the application, when the customer conducts the target transaction at the business window in the business outlet, the first acquisition unit 501 acquires the input information of the customer, and converts the input information into text data through a deep learning model to obtain a confidence level; the first extraction unit 502 extracts first transaction features from the text data, wherein the first transaction features at least include: a first transaction behavior and a first transaction amount; the second extraction unit 503 obtains a to-be-executed instruction of a target object through an application programming interface, and extracts second transaction features from the to-be-executed instruction, wherein the second transaction features at least include: a second transaction behavior and a second transaction amount, and the target object is a staff member in the business outlet who handles the target transaction; the comparison unit 504 compares the first transaction behavior with the second transaction behavior, compares the first transaction amount with the second transaction amount, and triggers a transaction early warning according to the first comparison result and the confidence level, thereby solving the problem that in the related art, when a teller in a business outlet handles a customer transaction, due to manual processing, errors such as transaction amount errors or keyboard mis-touches may occur, resulting in incorrect transaction execution.

[0112] The input information captured in real time is processed through the deep learning model, the input information of the customer when conducting the target transaction can be captured in real time, the input information is converted into clear text data, and the confidence level of the recognition result is calculated, thereby achieving effective conversion from unstructured information to structured data, and further achieving the purpose of accurate information analysis.

[0113] Meanwhile, the first transaction behavior is compared with the second transaction behavior, the first transaction amount is verified with the second transaction amount, the transaction early warning is triggered according to the first comparison result and in combination with the recognition confidence level, so that the system can discover and alert any inconsistency or potential error in time before the transaction is executed, thereby achieving active prevention and control of operational risks. This series of steps jointly constitute a teller transaction and amount intelligent verification technology based on natural language processing, greatly improving the automation level and accuracy of the transaction process, reducing the fund risk caused by human errors, and optimizing the quality of bank teller services and operational efficiency.

[0114] Optionally, in the transaction pre-warning device for offline transactions in business outlets provided in Embodiment 2 of the present application, the above-mentioned service window is deployed with a microphone array, the first acquisition unit 501 comprises: an acquisition subunit configured to control the microphone array to acquire audio stream data and timestamp information of the audio stream data by using an adaptive beamforming algorithm; a first processing subunit configured to process the audio stream data by using a preset algorithm to obtain processed audio stream data, wherein the preset algorithm at least comprises one of the following: a spectral subtraction method and a speech enhancement algorithm based on a deep learning model; a first extraction subunit configured to extract an audio feature vector from the processed audio stream data by using a deep learning model; and a conversion subunit configured to convert the audio feature vector into text type data by using a speech conversion model to obtain model output text data and a confidence level.

[0115] Optionally, in the transaction pre-warning device for offline transactions in business outlets provided in Embodiment 2 of the present application, the above-mentioned extraction subunit comprises: a cutting module configured to cut the processed audio stream data to obtain multiple pieces of audio data; a first extraction module configured to extract acoustic features from the multiple pieces of audio data based on the timestamp information to obtain a time sequence feature sequence; a second extraction module configured to extract local acoustic features from the time sequence feature sequence by using a convolutional neural network to obtain a first feature vector; and a third extraction module configured to input the first feature vector into a transformer model to extract context-dependent information by using the transformer model to obtain the audio feature vector.

[0116] Optionally, in the transaction pre-warning device for offline transactions in business outlets provided in Embodiment 2 of the present application, the first extraction unit 502 comprises: a second processing subunit configured to perform text preprocessing on the text data to obtain processed text data, wherein the text preprocessing at least comprises one of the following: removing redundant characters, standardizing punctuation marks, converting full-width or half-width, speech correction processing, and text standardization; an analysis subunit configured to analyze the processed text data based on a large language model to identify a customer transaction intention and a first amount feature, wherein the large language model is a model trained by using a financial domain knowledge base; a third processing subunit configured to perform standardization processing on the first amount feature by using the large language model to obtain a second amount feature; and a second extraction subunit configured to fuse the customer transaction intention and the second amount feature by using the large language model, and extract a first transaction behavior, a first transaction amount, and a transaction amount type from the fused data.

[0117] Optionally, in the transaction pre-warning device for offline transactions in business outlets provided in Embodiment 2 of the present application, the first extraction unit 502 further includes an identification subunit configured to, after identifying the customer transaction intention and the first amount feature based on the analysis of the processed text data by the large language model, identify, based on the large language model, a third transaction feature of the target transaction from the processed text data, where the third transaction feature includes at least one of the following: transaction account information, transaction party information, and bank information; the device further includes a third extraction subunit configured to, after obtaining the to-be-executed instruction of the target object through the application programming interface, extract a fourth transaction feature from the to-be-executed instruction, where the fourth transaction feature includes at least one of the following: transaction account information, transaction party information, and bank information; and a comparison subunit configured to compare each feature in the third transaction feature with a corresponding feature in the fourth transaction feature, and trigger the transaction pre-warning according to the second comparison result.

[0118] Optionally, in the transaction pre-warning device for offline transactions in business outlets provided in Embodiment 2 of the present application, the comparison unit 504 includes a first determination subunit configured to determine that the first condition is met when the first comparison result indicates that the first transaction behavior and the second transaction behavior are different; a second determination subunit configured to determine that the second condition is met when the first comparison result indicates that the difference between the first transaction amount and the second transaction amount is greater than the first difference threshold value, where the first difference threshold value is calculated based on the first transaction amount; a third determination subunit configured to determine that the third condition is met when the confidence level is lower than a preset confidence threshold value; and a pre-warning subunit configured to trigger the transaction pre-warning when the first condition, the second condition, or the third condition is met.

[0119] Optionally, in the transaction pre-warning device for offline transactions in business outlets provided in Embodiment 2 of the present application, the pre-warning subunit includes a first processing module configured to, when the first condition is met, suspend the execution of the to-be-executed instruction within a preset time period, and prompt the target object to adjust the to-be-executed instruction in the form of voice broadcast; a second processing module configured to, when the second condition is met and the difference is less than or equal to a second difference threshold value, display a pop-up window in the device operated by the target object to prompt the target object to adjust the to-be-executed instruction, where the second difference threshold value is calculated based on the first transaction amount, and the second difference threshold value is greater than the first difference threshold value; and a third processing module configured to, when the second condition is met and the difference is greater than the second difference threshold value, suspend the execution of the target transaction, and send the transaction information of the target transaction to the target business system, where the risk information existing in the target transaction is identified through the target business system.

[0120] Optionally, in the early warning device for offline transactions at business outlets provided in Embodiment 2 of this application, the aforementioned business window is equipped with a display terminal, and the device further includes: a second acquisition unit, used to acquire text data input by the customer and timestamp information of the acquired text data through the display terminal when the display terminal receives a customer instruction before extracting the first transaction feature from the text data.

[0121] It should be noted that the first acquisition unit 501, the first extraction unit 502, the second extraction unit 503, and the comparison unit 504 mentioned above correspond to steps S201 to S204 in Embodiment 1. The two modules and the corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in Embodiment 1. It should be noted that the above modules or units can be hardware or software components stored in memory (e.g., memory 104) and processed by one or more processors (e.g., processors 102a, 102b, ..., 102n). The above modules can also be part of the device and run in the computer terminal 10 provided in Embodiment 1.

[0122] Example 3

[0123] Embodiments of this application may provide an electronic device. Figure 6 This is a schematic diagram of an early warning electronic device for offline transactions at business outlets, provided according to Embodiment 3 of this application. Figure 6 As shown, the electronic device may include: one or more ( Figure 6 (Only one is shown) Processor 602, memory 604, memory controller, and peripheral interface, wherein the peripheral interface is connected to the radio frequency module, audio module and display.

[0124] The memory can be used to store software programs and modules, such as the program instructions / modules corresponding to the methods and apparatus in the embodiments of this application. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby implementing the above-described methods. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0125] The processor can call information and application programs stored in the memory through the transmission device to perform the following steps: when a customer conducts a target transaction is detected at a service window in a service site, collecting input information of the customer, and converting the input information into text data through a deep learning model to obtain a confidence level; extracting first transaction features from the text data, wherein the first transaction features at least include: a first transaction behavior and a first transaction amount; obtaining a to-be-executed instruction of a target object through an application program interface, and extracting second transaction features from the to-be-executed instruction, wherein the second transaction features at least include: a second transaction behavior and a second transaction amount, and the target object is a staff member who handles the target transaction in the service site; comparing the first transaction behavior with the second transaction behavior, comparing the first transaction amount with the second transaction amount, and triggering a transaction warning according to the first comparison result and the confidence level.

[0126] The processor can call information and application programs stored in the memory through the transmission device to perform the following steps: the service window is deployed with a microphone array, and the input information of the customer is collected and converted into text data through a deep learning model to obtain a confidence level, including: controlling the microphone array to collect audio stream data and timestamp information of the audio stream data by using an adaptive beamforming algorithm; processing the audio stream data by using a preset algorithm to obtain processed audio stream data, wherein the preset algorithm at least includes one of the following: a spectral subtraction method, a speech enhancement algorithm based on a deep learning model; extracting an audio feature vector from the processed audio stream data through a deep learning model; converting the audio feature vector into text type data through a speech conversion model to obtain text data output by the model and the confidence level.

[0127] The processor can call information and application programs stored in the memory through the transmission device to perform the following steps: extracting an audio feature vector from the processed audio stream data through a deep learning model, including: cutting the processed audio stream data to obtain multiple pieces of audio data; extracting acoustic features from the multiple pieces of audio data based on the timestamp information to obtain a time sequence feature sequence; extracting local acoustic features from the time sequence feature sequence through a convolutional neural network to obtain a first feature vector; inputting the first feature vector into a transformer model, extracting context-dependent information through the transformer model to obtain the audio feature vector.

[0128] The processor can call information and application programs stored in the memory through the transmission device to perform the following steps: extracting first transaction features from the text data, including: performing text preprocessing on the text data to obtain processed text data, wherein the text preprocessing at least includes one of the following: removing redundant characters, punctuation symbol standardization, full-width or half-width conversion, speech correction processing, and text standardization; analyzing the processed text data based on a large language model to identify a customer transaction intent and a first amount feature, wherein the large language model is a model trained using a financial domain knowledge base; performing standardization processing on the first amount feature through the large language model to obtain a second amount feature; fusing the customer transaction intent and the second amount feature through the large language model, and extracting a first transaction behavior, a first transaction amount, and a transaction amount type from the fused data.

[0129] The processor can call information and application programs stored in the memory through the transmission device to perform the following steps: after analyzing the processed text data based on a large language model to identify a customer transaction intent and a first amount feature, the method further includes: identifying a third transaction feature of a target transaction based on the large language model, wherein the third transaction feature at least includes one of the following: transaction account information, transaction party information, and bank information; after obtaining the to-be-executed instruction of the target object through the application program interface, the method further includes: extracting a fourth transaction feature from the to-be-executed instruction, wherein the fourth transaction feature at least includes one of the following: transaction account information, transaction party information, and bank information; comparing each feature in the third transaction feature with the corresponding feature in the fourth transaction feature, and triggering a transaction warning according to the second comparison result.

[0130] The processor can call information and application programs stored in the memory through the transmission device to perform the following steps: triggering a transaction warning according to the first comparison result and the confidence, including: determining that the first condition is met when the first comparison result indicates that the first transaction behavior and the second transaction behavior are different; determining that the second condition is met when the difference between the first transaction amount and the second transaction amount is greater than a first difference threshold, wherein the first difference threshold is calculated based on the first transaction amount; determining that the third condition is met when the confidence is lower than a preset confidence threshold; and triggering a transaction warning when the first condition, the second condition, or the third condition is met.

[0131] The processor can access information and applications stored in the memory via a transmission device to execute the following steps: Triggering a transaction alert when a first, second, or third condition is met, including: If the first condition is met, pausing the execution of the pending instruction within a preset time period and prompting the target object to adjust the pending instruction via voice broadcast; If the second condition is met and the difference is less than or equal to a second difference threshold, displaying a pop-up window on the device operated by the target object to prompt the target object to adjust the pending instruction, wherein the second difference threshold is calculated based on the first transaction amount and is greater than the first difference threshold; If the second condition is met and the difference is greater than the second difference threshold, pausing the execution of the target transaction and sending the transaction information of the target transaction to the target business system, wherein the target business system identifies risk information present in the target transaction.

[0132] The processor can call the information and application stored in the memory through the transmission device to perform the following steps: The business window is equipped with a display terminal. Before extracting the first transaction feature from the text data, the above method further includes: when the display terminal detects that it has received a customer instruction, collecting the text data input by the customer through the display terminal, as well as the timestamp information of the collected text data.

[0133] This application provides an early warning method for offline transactions at a branch. When a customer is detected conducting a target transaction at the branch's service window, the method collects the customer's input information and converts it into text data using a deep learning model to obtain a confidence level. A first transaction feature is extracted from the text data, which includes at least a first transaction behavior and a first transaction amount. An execution instruction for the target object is obtained through an application programming interface (API), and a second transaction feature is extracted from the instruction. This second transaction feature includes at least a second transaction behavior and a second transaction amount. The target object is the staff member at the branch handling the target transaction. The first and second transaction behaviors are compared, as are the first and second transaction amounts. Based on the first comparison result and the confidence level, a transaction early warning is triggered. This method solves the technical problem of incorrect transaction execution caused by errors such as incorrect transaction amounts or accidental keystrokes during customer transactions handled by branch tellers.

[0134] By processing the real-time captured input information through a deep learning model, it is possible to capture the input information of customers when conducting target transactions in real time, convert it into clear text data, and calculate the confidence level of the recognition results. This achieves an effective transformation from unstructured information to structured data, and further achieves the goal of accurate information parsing.

[0135] Simultaneously, by comparing the extracted first and second transaction behaviors and verifying the first and second transaction amounts, and triggering transaction alerts based on the first comparison result and the identification confidence level, the system can promptly detect and warn of any inconsistencies or potential errors before transaction execution, thus achieving proactive control of operational risks. This series of steps collectively constitutes the over-the-counter transaction and intelligent amount verification technology based on natural language processing, greatly improving the automation and accuracy of the transaction process, reducing financial risks caused by human error, and optimizing the quality of bank over-the-counter services and operational efficiency.

[0136] Those skilled in the art will understand that Figure 6 The structure shown is for illustrative purposes only. Electronic devices can also be smartphones, tablets, handheld computers, mobile internet devices (MIDs), PADs, and other terminal devices. Figure 6 This does not limit the structure of the aforementioned electronic device. For example, electronic devices may also include components that are more... Figure 6 The more or fewer components shown (such as network interfaces, display devices, etc.), or having the same Figure 6 The different configurations shown.

[0137] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0138] Example 4

[0139] Embodiments of this application also provide a storage medium. Optionally, in this embodiment, the storage medium can be used to store the program code executed by the early warning method for offline transactions at business outlets provided in Embodiment 1.

[0140] Optionally, in this embodiment, the storage medium may be located in any computer terminal in a group of computer terminals in a computer network, or in any mobile terminal in a group of mobile terminals.

[0141] This application also provides a computer program product, which, when executed on a data processing device, is suitable for performing early warning method steps for offline transactions at business outlets.

[0142] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0143] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0144] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of units or modules may be electrical or other forms.

[0145] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0146] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0147] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0148] The above are merely preferred embodiments of this application. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A pre-warning method for off-line transactions of a business site, characterized in that, The method comprises the following steps: In the case that the business window in the business site detects that the customer conducts a target transaction, the input information of the customer is collected, and the input information is converted into text data through a deep learning model to obtain a confidence level; First transaction features are extracted from the text data, wherein the first transaction features at least include a first transaction behavior and a first transaction amount; A target object's to-be-executed instruction is obtained through an application programming interface, and second transaction features are extracted from the to-be-executed instruction, wherein the second transaction features at least include a second transaction behavior and a second transaction amount, and the target object is a staff member in the business site who handles the target transaction; The first transaction behavior and the second transaction behavior are compared, the first transaction amount and the second transaction amount are compared, and a transaction early warning is triggered according to a first comparison result and the confidence level.

2. The method of claim 1, wherein, The business window is provided with a microphone array, the input information of the customer is collected, and the input information is converted into text data through a deep learning model to obtain a confidence level, comprising: An adaptive beamforming algorithm is used to control the microphone array to collect audio stream data and timestamp information of the audio stream data; A preset algorithm is used to process the audio stream data to obtain processed audio stream data, wherein the preset algorithm at least includes one of the following: spectral subtraction, a speech enhancement algorithm based on a deep learning model; An audio feature vector is extracted from the processed audio stream data through a deep learning model; The audio feature vector is converted into text type data through a speech conversion model to obtain the text data output by the model and the confidence level.

3. The method of claim 2, wherein, An audio feature vector is extracted from the processed audio stream data through a deep learning model, comprising: The processed audio stream data is cut to obtain multiple segments of audio data; Acoustic features are extracted from the multiple segments of audio data based on the timestamp information to obtain a time sequence feature sequence; Local acoustic features are extracted from the time sequence feature sequence through a convolutional neural network to obtain a first feature vector; The first feature vector is input into a transformer model to extract context-dependent information through the transformer model to obtain the audio feature vector.

4. The method of claim 1, wherein, First transaction features are extracted from the text data, comprising: Text preprocessing is performed on the text data to obtain processed text data, wherein the text preprocessing at least includes one of the following: removing redundant characters, standardizing punctuation marks, converting full-width or half-width, speech correction processing, and text standardization; The processed text data is parsed based on a large language model to identify a customer transaction intent and a first amount feature, wherein the large language model is a model trained using a financial domain knowledge base; The first amount feature is standardized through the large language model to obtain a second amount feature; The customer transaction intent and the second amount feature are fused through the large language model, and the first transaction behavior, the first transaction amount, and a transaction amount type are extracted from the fused data.

5. The method of claim 4, wherein, After the large language model is used to analyze the processed text data to identify the customer transaction intention and the first amount feature, the method further includes: Based on the large language model, the third transaction feature of the target transaction is identified by identifying the processed text data, wherein the third transaction feature at least includes one of the following: transaction account information, transaction party information, and bank information; After obtaining the to-be-executed instruction of the target object through the application programming interface, the method further includes: The fourth transaction feature is extracted from the to-be-executed instruction, wherein the fourth transaction feature at least includes one of the following: transaction account information, transaction party information, and bank information; The third transaction feature and the corresponding feature in the fourth transaction feature are compared, and a transaction warning is triggered according to the second comparison result.

6. The method of claim 1, wherein, According to the first comparison result and the confidence, a transaction warning is triggered, including: When the first comparison result indicates that the first transaction behavior and the second transaction behavior are different, a first condition is determined to be met; When the first comparison result indicates that the difference between the first transaction amount and the second transaction amount is greater than a first difference threshold, a second condition is determined to be met, wherein the first difference threshold is calculated based on the first transaction amount; When the confidence is lower than a preset confidence threshold, a third condition is determined to be met; When the first condition, the second condition, or the third condition is met, a transaction warning is triggered.

7. The method of claim 6, wherein, When the first condition, the second condition, or the third condition is met, a transaction warning is triggered, including: When the first condition is met, the execution of the to-be-executed instruction is suspended within a preset time period, and the target object is prompted to adjust the to-be-executed instruction in the form of voice broadcast; When the second condition is met and the difference is less than or equal to a second difference threshold, a pop-up window is displayed in the device operated by the target object to prompt the target object to adjust the to-be-executed instruction, wherein the second difference threshold is calculated based on the first transaction amount, and the second difference threshold is greater than the first difference threshold; When the second condition is met and the difference is greater than the second difference threshold, the execution of the target transaction is suspended, and the transaction information of the target transaction is sent to a target business system, wherein the target business system is used to identify risk information existing in the target transaction.

8. The method of claim 1, wherein, The business window is deployed with a display terminal, and before the first transaction feature is extracted from the text data, the method further includes: When it is detected that the display terminal receives a customer instruction, the text data input by the customer is collected through the display terminal, and the timestamp information of the text data is collected.

9. A pre-warning device for off-line transactions of a business site, characterized by, Including: A first collection unit is configured to collect input information of a customer when a target transaction is performed by the customer at a business window of a business site, and convert the input information into text data through a deep learning model to obtain a confidence. The first extraction unit is configured to extract a first transaction feature from the text data, wherein the first transaction feature at least includes a first transaction behavior and a first transaction amount; The second extraction unit is configured to acquire a to-be-executed instruction of a target object through an application program interface, and extract a second transaction feature from the to-be-executed instruction, wherein the second transaction feature at least includes a second transaction behavior and a second transaction amount, and the target object is a staff member in the business outlet who handles the target transaction; The comparison unit is configured to compare the first transaction behavior with the second transaction behavior, compare the first transaction amount with the second transaction amount, and trigger a transaction early warning according to a first comparison result and the confidence level.

10. A computer program product comprising computer instructions, characterized in that, The computer instructions, when executed by the processor, implement the steps of the early warning method for the offline transaction of the business outlet according to any one of claims 1 to 8.

Images