Multi-mode arrearage risk portrait-based payment reminding talking skill selection method and device

Abstract

The invention discloses a payment reminding talking skill selection method and device based on a multi-mode arrearage risk portrait, and belongs to the technical field of data mining and risk management. The method comprises the steps that multi-source user data are collected, structured data are obtained regularly in a daily partition synchronization mode, and unstructured data are transmitted in a rapid query and collection mode; the data is cleaned, structural features and multi-modal interaction features are extracted, and the multi-modal interaction features comprise emotion recognition features and intention recognition features; training a risk level quantification model based on artificial intelligence by using the clean data set and the multi-modal interaction data set, and outputting a centesimal risk score; and dividing the users into high, medium and low risk levels according to the risk scores, and automatically matching corresponding collection verbal skill processes and strategies. According to the invention, efficient, accurate and automatic assessment of the user risk is realized, and the pertinence and adaptability of the collection strategy are significantly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data mining and risk management, in particular to a collection call strategy selection method and device based on a multi-modal arrear risk portrait. BACKGROUND

[0002] With the explosive growth of mobile communication user scale, a huge number of users generate huge monthly phone bills, and user arrears have become a major financial risk faced by communication operators. The traditional collection method relying on manual customer service is not only inefficient and costly, but also difficult to cover comprehensively, and there is a problem of complaints or compliance risks caused by improper communication when facing tens of millions of users.

[0003] Under this background, intelligent collection systems integrating automatic speech recognition (ASR), natural language processing (NLP) and speech synthesis (TTS) and other technologies have emerged. Such systems can process standardized collection tasks 7x24 hours without interruption, significantly reducing the basic operating cost. However, existing systems still face serious challenges in actual complex scenarios: First, the strategy is single and lacks precision. Existing systems mostly use fixed templates, which cannot be dynamically adjusted according to individual differences such as user arrears reasons, credit history, and consumption ability. The same strategy is used for temporary forgetful premium customers and habitual arrears users, which not only has limited effect, but also may damage customer relationship.

[0004] Second, the interaction is rigid and the understanding ability is weak. In real conversations, the system cannot accurately identify the deep intention and emotion analysis of the user's complex, ambiguous or strongly emotional expressions. It is difficult to distinguish the authenticity of the user's promise, leading to ineffective dialogue guidance and missing the collection opportunity.

[0005] Third, the closed loop is missing and the optimization relies on manual intervention. The system generally lacks data-driven self-optimization capability. After the collection is completed, the effect evaluation and strategy adjustment still need a lot of manual intervention, and the system cannot automatically learn successful experience from massive interaction data, resulting in poor adaptability and long iteration cycle.

[0006] Therefore, how to develop an intelligent collection system that can deeply understand user intent, dynamically execute personalized strategies, and have self-learning ability has become a key technical problem to be solved in the industry. SUMMARY

[0007] The main purpose of the present application is to provide a collection call strategy selection method and device based on a multi-modal arrear risk portrait, which aims to overcome the core defects of no perception, no distinction and no closed loop in the prior art by constructing a multi-modal user arrear risk portrait, to realize the dynamic, precise and automated closed loop of the collection strategy, thereby improving the collection efficiency and optimizing the user experience.

[0008] To achieve the above object, the application provides a method for selecting a collection call based on a multi-modal overdue payment risk portrait, which comprises the following steps: S1: collecting multi-source user data required by the China Unicom system; wherein the multi-source user data comprises basic billing records, payment flow records and user package usage data in structured data, and historical call records and historical collection message data in unstructured data; S2: preprocessing the multi-source user data, extracting the overdue payment frequency feature of the structured data, and performing speech signal framing and windowing processing, text cleaning and intent recognition on the unstructured data to generate user emotion features and repayment intention features; S3: constructing a dynamic risk quantification model based on the overdue payment frequency, emotion features and repayment intention features, mapping the repayment intention confidence by using the truncated normal distribution and generating a repayment intention value, combining the emotion weight for weighted summation and then converting by using the sigmoid function to calculate the user risk score; S4: dividing the risk score into high / medium / low risk levels according to a preset risk threshold, and calling the corresponding speech strategy template based on the risk level, wherein the high-risk users adopt the strong speech of penalty details and shutdown countdown, the low-risk users adopt the preferential induction speech combined with the package usage habit, and the medium-risk users adopt the mixed strategy speech.

[0009] Further, the S1 comprises: synchronizing the structured data to the data processing platform in a Hive daily partition synchronization manner at T+1 day; storing the unstructured data in a Hbase+MinIO object storage manner, querying the target user's unstructured data to the data processing platform through ES, and collecting the target user's unstructured data.

[0010] Further, the step of preprocessing the multi-source user data, extracting the overdue payment frequency feature of the structured data, and performing speech signal framing and windowing processing, text cleaning and intent recognition on the unstructured data to generate user emotion features and repayment intention features comprises: performing dirty data cleaning and repeated data elimination processing on the basic billing records and payment flow records, and counting the overdue payment frequency of the user in the last 12 months according to the processed basic billing records and payment flow records; extracting the audio tracks of the two channels of the historical voice call records, and preliminarily removing the noise in the historical voice call records; counting the reply rate of the user of the historical collection message, and cleaning the reply text of the user; wherein the two channels comprise a left channel and a right channel, the left channel is a customer service, and the right channel is a user; the text cleaning comprises removing irrelevant characters, unifying coding, standardizing, and removing stop words.

[0011] Further, the step of constructing a dynamic risk quantification model based on the number of overdue payments, emotional features, and repayment intention features, mapping the repayment intention confidence through truncated normal distribution and generating a repayment intention value, and then calculating the user risk score through sigmoid conversion after weighted summation combined with emotional weight includes: converting the user's historical voice call record into a text message, identifying the user's emotion, and extracting the user's emotional negative value; manually labeling the text message and inputting it into the nlu module for fine-tuning processing, and optimizing the nlu module through multiple iterations to classify the user's intention.

[0012] Further, the step of dividing the risk score into high / medium / low risk levels according to the preset risk threshold and calling the corresponding dialogue strategy template based on the risk level, wherein high-risk users use strong dialogue strategies with penalty details and shutdown countdown, low-risk users use preferential inducement dialogue strategies combined with package usage habits, and medium-risk users use mixed strategy dialogue strategies includes: logarithmic transformation of the number of overdue payments and compression of the range; define the distribution parameters of the classification of the user's intention, then use the truncated normal distribution function to construct the probability distribution of the distribution parameters to obtain the value of the classification of the intention mapped between 0-1; input the historical voice call record into the SenseVoice-Small model to obtain a probability dictionary, and obtain the emotional negative value between 0-1 through weighted summation of the emotional weight and the probability dictionary and sigmoid conversion to calculate the relevant features; bring the calculated relevant features into the risk level quantification model to generate a risk score, and take truncation processing on the risk score, and convert the risk score into a percentage form to obtain a percentage risk score; When calling the outbound nlp interface, collect the risk score and features of the corresponding user, and select the corresponding dialogue strategy; use a one-way transformer model to fine-tune the dialogue strategy to obtain a relatively optimal dialogue strategy.

[0013] The application also provides a device for selecting a collection speech based on a multi-modal overdue risk portrait, which comprises: a multi-source data acquisition and synchronization module, which is used for acquiring multi-source user data from a communication system, including structured data and unstructured data, and synchronizing the structured data to a data processing platform in time by using a daily partition synchronization technology, and transmitting the unstructured data to the data processing platform by using a fast query and collection technology, so as to establish a complete and real-time accessible data source basis; a structured data cleaning and feature extraction module, which is used for cleaning dirty data and removing duplicate data from the basic bill and payment stream records, and extracting user behavior features from the cleaned structured data by using a statistical modeling technology, including the number of overdue payment times in the past 12 months, so as to obtain a standardized and analyzable clean data set; a multi-modal unstructured data feature extraction module, which is used for extracting features from voice call records and collection message records, and converting voice signals into text and simultaneously identifying user emotions by using a multi-task voice processing model, and extracting user reply rates and intention values by using a text cleaning technology and an intention recognition model, so as to obtain a multi-modal interaction data set containing negative emotion values and repayment intention values; a risk level quantization model training module, which is used for training a risk level quantization model by using the clean data set and the multi-modal interaction data set, and fusing overdue times, repayment intention values, negative emotion values and reply rates into a risk score by using a weighted linear model technology combined with logarithmic transformation and normalization processing, and outputting a percentage risk score by using truncation processing and percentage conversion; the risk score is classified into three categories of high, medium and low risks by using a threshold division technology, and corresponding collection speech processes and operation strategies are automatically selected according to the risk categories by using a strategy mapping algorithm.

[0014] Further, the multi-modal unstructured data feature extraction module is specifically used for: converting the user's historical voice call records into text messages, simultaneously identifying the user's emotions, and extracting negative emotion values of the user; manually labeling the text messages and inputting them into an nlu module for fine-tuning processing, so as to optimize the nlu module by multiple iterations, and realize classification of the user's intention.

[0015] Further, the multi-source data acquisition and synchronization module is specifically used for: synchronizing the structured data to the data processing platform in time by using a Hive daily partition synchronization mode T+1 day; storing the unstructured data in batches by using a Hbase+MinI0 object storage mode, and collecting the unstructured data of the target user to the data processing platform by using ES fast query.

[0016] Further, the structured data cleaning and feature extraction module is specifically configured to: clean dirty data and remove duplicate data from the basic bill records and payment transaction records, and count the number of payment overdue times of a user in the past 12 months according to the processed basic bill records and payment transaction records; extract the audio tracks of two channels of the historical voice call records, and preliminarily remove the noise in the historical voice call records; count the reply rate of the user of the historical payment reminder message, and clean the reply text of the user; wherein the two channels include a left channel and a right channel, the left channel is a customer service, and the right channel is a user; the text cleaning includes removing irrelevant characters, unifying coding, standardizing, and removing stop words.

[0017] Further, the multi-modal unstructured data feature extraction module is specifically configured to: convert the historical voice call records of the user into text messages, identify the emotion of the user, and extract the negative emotion value of the user; manually label the text messages and input the nlu module for fine-tuning processing, to optimize the nlu module through multiple iterations, and realize classification of the intention of the user.

[0018] Further, the risk level quantification model training module is specifically configured to: perform logarithmic transformation on the number of payment overdue times and compress the range; define the distribution parameters of the classification of the intention of the user, and then construct the probability distribution of the distribution parameters by using a truncated normal distribution function, to obtain a value between 0 and 1 that the classification of the intention is mapped to; input the historical voice call records into a SenseVoice-Small model, to obtain a probability dictionary, and obtain a negative emotion value between 0 and 1 by using weighted summation of the emotion weight and the probability dictionary and sigmoid conversion, to calculate a related feature; wherein the related feature includes: the data form after logarithmic transformation and compression of the range of the number of payment overdue times, the value between 0 and 1 that the classification of the intention is mapped to, the negative emotion value between 0 and 1, and the reply rate of the user of the historical payment reminder message; bring the calculated related features into a risk level quantification model, to generate a risk score, and perform truncation processing on the risk score, and convert the risk score into a percentage form, to obtain a risk score in percentage form; when calling an outbound nlp interface, collect the risk score and the features of the corresponding user, and select a corresponding dialogue strategy; fine-tune the dialogue strategy by using a one-way transformer model, to obtain a relatively optimal dialogue strategy.

[0019] Compared with the prior art, the technical scheme provided by the application has at least the following beneficial effects: 1. Multi-modal data fusion: map the voice acoustic features (MFCC), text semantic vectors (BERT) and payment behavior embedding to a unified space, and fully utilize the multi-source user data of an operator to construct a user portrait.

[0020] 2. Personalized Payment Collection Engine: This engine selects different scripts and strategies based on user profiles for payment collection. It simplifies the process for low-risk users, preventing excessive collection efforts. Simultaneously, it reduces resource consumption and improves user experience. For high-risk users, it provides targeted collection efforts, effectively increasing collection conversion rates.

[0021] 3. Enhanced Interaction Capabilities: Anticipate user emotions based on historical call and SMS records, and adopt targeted communication strategies to reduce customer complaints.

[0022] This invention presents a method and system for selecting collection scripts based on multimodal debt risk profiling, effectively overcoming the core shortcomings of existing technologies such as single strategy, rigid interaction, and lack of optimization loop. This method achieves accurate perception of users' repayment intentions and emotional states. Through quantitative risk assessment and intelligent matching of script strategies, it significantly improves the accuracy and adaptability of collection operations, enhancing both collection efficiency and user experience, and strengthening the system's decision-making capabilities and overall operational efficiency in complex scenarios. Attached Figure Description

[0023] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating a method for selecting collection scripts based on a multimodal delinquency risk profile, provided in an embodiment of the present invention; Figure 2 This invention provides an architecture diagram of a collection script selection system based on a multimodal delinquency risk profile. Figure 3 This is a structural diagram of the SensenceVoice-Small model provided in an embodiment of the present invention; Figure 4 An intent recognition flowchart provided for embodiments of the present invention; Figure 5 This is a schematic diagram of a collection script selection device based on a multimodal overdue payment risk profile, provided in an embodiment of the present invention. Detailed Implementation

[0024] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] In the following, the technical solutions in the embodiments of the present application will be described clearly and completely with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work should belong to the protection scope of the present application.

[0026] The core idea of the embodiments of the present application is that by constructing a multi-modal dynamic portrait system integrating voice, text and behavior data, acoustic features, semantic vectors and payment behavior are embedded and mapped to a unified representation space, realizing accurate quantification of user arrear risk and emotional intention. Based on this portrait, the system can adaptively match different users with different risk levels and emotional states with differentiated dialogue strategies, thereby transforming the traditional standardized collection process into an intelligent interactive closed loop that deeply perceives user state, dynamically decides and continuously optimizes, ultimately improving the efficiency of collecting money while ensuring the synergistic optimization of user experience and system operation benefits.

[0027] A multi-modal arrear risk portrait-based collection dialogue selection method and system according to an embodiment of the present application will be described below with reference to the accompanying drawings.

[0028] Embodiment 1 The present embodiment provides a multi-modal arrear risk portrait-based collection dialogue selection method. As shown in the figure, the method comprises the following steps: Figure 1 S1: Collecting multi-source user data required by the China Unicom system; wherein the multi-source user data includes basic billing records, payment stream records and user package usage data in structured data, and historical call records and historical collection SMS data in unstructured data.

[0029] This step is the data basis for subsequent user risk portrait and collection strategy optimization, aiming to realize the standardized access and efficient management of massive heterogeneous data. First, a multi-source data collection system is established. The multi-source user data includes structured data and unstructured data, wherein the structured data covers basic billing records, payment stream records, user package usage data and other business data with a clear schema; the unstructured data includes historical voice call records, historical collection SMS and other original data rich in user interaction information.

[0030] ​Further, the synchronization and collection of the two types of data are realized by using differentiated technical solutions. For structured data, the Hive daily partition synchronization mechanism is used to synchronize the data to the data processing platform in a T+1 day cycle, ensuring data timeliness and batch processing efficiency. For unstructured data, batch storage is performed by combining HBase distributed database and MinIO object storage, and an index engine is built using Elasticsearch to realize fast retrieval and targeted collection of unstructured data of target users, effectively supporting subsequent feature extraction and multi-modal analysis.

[0031] Specifically, in the data access process, the legality and compliance of the data source and the reliability of the data quality need to be ensured. By establishing a data bloodline tracking mechanism, the source, conversion process, and usage path of the data are recorded, providing an auditable data foundation for subsequent model training and strategy decision-making. This multi-level, heterogeneous unified data access system lays a solid data foundation for building an accurate user risk portrait.

[0032] S2: Preprocess multi-source user data, extract the number of overdue payment features of structured data, and perform speech signal framing and windowing processing, text cleaning, and intent recognition on unstructured data to generate user emotion features and repayment intention features.

[0033] This step is a key preprocessing link for building a user risk portrait, aiming to convert raw data into standardized features for model training through systematic data cleaning and feature engineering. First, perform data quality governance on structured data. For basic billing records and payment stream records, implement a dirty data cleaning process, including imputation of missing values, detection and correction of outliers, and identification and removal of duplicate records based on business rules and data correlation, to ensure data record completeness and uniqueness.

[0034] Further, user behavior features are constructed based on cleaned structured data. Specifically, by time series aggregation and business rule judgment, the number of overdue payments of users in the last 12 months is calculated, which will be one of the core indicators for evaluating user repayment intention and credit risk. In the feature calculation process, the overdue judgment standard needs to be clearly defined, such as using the difference between the billing date and the actual payment date exceeding the agreed grace period as the overdue judgment condition, to ensure business consistency and interpretability of feature calculation.

[0035] Specifically, for unstructured data, a multi-modal preprocessing procedure is implemented. For historical voice call records, first, audio track separation processing is performed to extract two independent audio channels, left and right, where the left channel corresponds to the customer voice and the right channel corresponds to the user voice, providing structured audio input for subsequent voice emotion analysis and intent recognition. At the same time, audio noise reduction algorithms based on deep learning, such as spectral subtraction or Wave-U-Net-based models, are used to preliminarily filter out background noise and line interference, improving the signal-to-noise ratio and analyzability of the voice data.

[0036] Specifically, for historical collection reminder message data, multi-dimensional interaction feature extraction is performed. First, the message reply rate of the user in the historical collection reminder scene is counted, which reflects the user's response willingness and participation in the collection reminder behavior. Further, the user's reply text is deeply cleaned and standardized, including: removing irrelevant characters such as URL links and special symbols; converting different encoding formats of text to UTF-8 standard encoding; implementing text normalization operations, including conversion between traditional and simplified Chinese, uniform case; and removing stop words based on a combination of domain dictionaries and general dictionaries, retaining key words with substantive semantic content to lay the foundation for subsequent text semantic analysis.

[0037] Through the above systematic data cleaning and feature extraction process, a clean dataset containing high-quality structured features and preliminarily processed unstructured data is finally output, providing reliable data input for subsequent multi-modal user portrait construction and risk model training.

[0038] S3: Based on the number of payment overdue times, emotional features and repayment intention features, a dynamic risk quantification model is constructed, the repayment intention confidence is mapped through truncated normal distribution to generate a repayment intention value, and the emotional weight is weighted and summed after sigmoid conversion to calculate the user risk score.

[0039] This step is the core link to achieve accurate user state perception, aiming to extract quantitative features with business value from voice and text data through advanced multi-modal analysis technology. First, a parallel processing channel for voice data is established. For separated user-side voice data, an end-to-end speech recognition model is used to convert audio signals into text messages; at the same time, a deep learning-based emotion recognition model is used to identify the user's emotional state from the prosody, pitch, energy and other acoustic features of the voice.

[0040] Further, a quantitative representation system of emotional features is constructed. In a preferred embodiment, the voice data is processed by a pre-trained emotion recognition model (such as SenseVoice-Small) to obtain a probability distribution containing multiple emotion categories, and then emotion weights are defined based on business needs, the probability distribution is weighted and summed, and a sigmoid function is converted to quantify complex emotional states into a sentiment negative value between 0 and 1. The larger the value, the stronger the negative emotions exhibited by the user in the historical interaction, providing an important behavioral signal for subsequent risk assessment.

[0041] Specifically, in the aspect of intent recognition, a domain-adaptive natural language understanding framework is established. First, based on the dialogue text of the historical collection scene, the business experts manually annotate according to the pre-defined intent classification system to form a high-quality labeled data set. Then the labeled data is input into the pre-trained natural language understanding module for fine-tuning, and through multiple rounds of iterative optimization, the model can accurately identify the deep intent expressed by the user in the collection interaction.

[0042] Specifically, in a preferred embodiment, the intent classification system is designed in advance from the business dimension, including but not limited to: positive repayment intention, request for extension, expression of dissatisfaction, consultation of questions, refusal to communicate, etc. The intent classification results of the text message are normalized to map them to an intention value between 0 and 1, which reflects the strength of the user's willingness to perform the repayment obligation.

[0043] Through the above-mentioned deep feature extraction of all unstructured data, a multi-modal interaction data set containing sentiment negative values and repayment intention values is finally formed, and these features will improve the user risk portrait from two key dimensions of emotional state and repayment willingness, providing multi-angle data support for subsequent risk quantification modeling.

[0044] S4: According to the pre-set risk threshold, the risk score is divided into high / medium / low risk levels, and based on the risk level, the corresponding dialogue strategy template is called, wherein the high-risk user adopts the strong dialogue of default fine details and shutdown countdown, the low-risk user adopts the preferential inducement dialogue combined with the habit of using the package, and the medium-risk user adopts the mixed strategy dialogue.

[0045] This step is the core link to realize the accurate quantification of user risk, aiming to integrate multi-dimensional user features into a unified interpretable risk assessment index through systematic feature engineering and statistical modeling technology. First, the original features in the cleaned data set are standardized for preprocessing. For the key behavior feature of overdue payment times, the log transformation method is used to reduce the data skewness, and the range compression algorithm is used to normalize it to a specific interval, eliminating the influence of feature dimension and retaining its discrimination ability in risk identification. By combining the quantitative risk assessment results with dynamic dialogue strategies, precise collection intervention is achieved. First, a mapping system of risk levels and collection strategies is established. Based on the percentage risk score, the user group is accurately divided into high, medium, and low risk levels through pre-set thresholds. Among them, the high risk level corresponds to strong collection intervention demand, the medium risk level needs to balance efficiency and user experience, and the low risk level focuses on friendly reminders and relationship maintenance.

[0046] Further, a probabilistic mapping system of intention classification results is established. Based on the intention classification results shown by users in historical interactions, the distribution parameters of each type of intention are defined, and the corresponding probability distribution model is constructed using the truncated normal distribution function. By this distribution function, discrete intention classification results are converted into continuous intention values between 0 and 1, realizing the conversion from qualitative judgment to quantitative evaluation, and providing stable and reliable willingness feature input for the model. A strategy-based outbound execution mechanism is constructed. When calling the natural language processing interface of the intelligent outbound system, the risk score and multi-dimensional features of the user are transmitted as context parameters in real time, and the system automatically selects the matching dialogue process according to the pre-set strategy mapping rules. For example, for high-risk users, the intensive collection dialogue emphasizing the consequences of default is enabled, for medium-risk users, the standard dialogue guiding negotiation is used, and for low-risk users, the simple dialogue of gentle reminders is used.

[0047] Specifically, in terms of emotion feature quantification, a deep learning-based sentiment analysis model is used for fine processing. The pre-processed user-side historical voice call records are input into the SenseVoice-Small pre-training model to obtain a probability dictionary containing multiple emotion categories. Based on the weight coefficients of each emotion category defined by business experts, the weighted sum of the probability dictionary and the sigmoid function conversion are used to calculate the sentiment negative value between 0 and 1, which comprehensively reflects the emotional state tendency of the user in historical interactions. In terms of dialogue optimization, a unidirectional Transformer-based generative model is used for dynamic dialogue adjustment. Historical successful collection cases are used to construct training corpus, and the model is trained through supervised fine-tuning to learn the optimal dialogue expression in different risk scenarios. This model can dynamically generate text content that better fits the current dialogue context within the pre-set dialogue framework based on real-time user feedback, realizing real-time optimization of dialogue strategy.

[0048] Specifically, in a preferred embodiment, a weighted linear combination model is constructed as the core algorithm for risk level quantification. The processed multi-dimensional features, including the normalized number of payment overdue times, continuous intention values, negative sentiment values, and historical reminder message reply rates, are linearly combined according to predefined weight coefficients to generate an initial risk score. To ensure the reasonableness of the score range, the initial score is truncated to eliminate the influence of extreme values, and finally the score is converted to a percentage form through linear transformation to form the final risk score. A pre-constructed script strategy library containing various collection scenarios is established, and a mapping relationship between the strategy and the risk features is established. For example, for high-risk users with strong negative emotions, the system automatically selects a script branch containing empathetic expressions and solutions; for medium-risk users with clear repayment intentions, the system preferentially uses confirmatory scripts with simplified processes. Through this multi-dimensional strategy matching mechanism, the accuracy and adaptability of the collection intervention are ensured.

[0049] Through the above systematic modeling process, a percentage risk score with clear business meaning and cross-user comparability is finally output, which comprehensively reflects the user's repayment intention, historical behavior pattern, and emotional attitude, providing a scientific and reliable decision basis for subsequent collection strategy formulation. Finally, a closed-loop intelligent collection decision system is formed, which improves the collection efficiency while effectively controlling the operational risk and maintaining the long-term healthy development of user relationships.

[0050] The method for selecting a collection script based on a multi-modal overdue risk portrait according to the embodiments of the present application realizes accurate matching and dynamic optimization of collection strategies based on quantified risk scores, effectively solving the core defects of single strategy, rigid interaction, and lack of optimization closed loop in the prior art. This method realizes full-process automation from data collection, feature extraction, risk quantification to strategy generation, significantly improving the accuracy of collection operations and the system's self-adaptation ability, improving the user experience while improving the efficiency of collecting money, and enhancing the decision-making effectiveness and operational benefits of the intelligent collection system in complex business scenarios.

[0051] Embodiment 2 The embodiments of the present application also provide a collection script selection system based on a multi-modal overdue risk portrait, as shown in Figure 2 The system comprises: The collection script selection system based on a multi-modal overdue risk portrait aims to construct a multi-modal collection script selection system through five modules of multi-source data collection, data preprocessing, multi-modal data fusion, risk level quantification model construction, and script flow strategy selection. The system can be realized through the following sub-steps: S101: Multi-source data collection.

[0052] Specifically, the data collection synchronizes the basic billing records, payment stream records, and user package usage data in the structured data of the Unicom users to the data processing platform in a Hive daily partition synchronization manner, T+1 day data, under the premise of complying with national laws and regulations and regulatory policies, ensuring information security and user privacy; the historical call records and historical collection reminder message data in the unstructured data are stored in batches in a Hbase+MinIO object storage manner, and the target user's unstructured data is collected to the data processing platform through ES fast query. At this point, the multi-source user data required by the system has been obtained.

[0053] S102: Data preprocessing.

[0054] Specifically, this stage aims to clean up structured and unstructured data and extract features from structured data. First, the dirty data cleaning and duplicate data removal are performed on the basic billing and payment stream records, and then the user's payment overdue times in the past 12 months are counted according to the above two data. For voice call record data, since the system adopts two channels for recording, the left channel is for customer service and the right channel is for users, so only the left and right channel audio tracks need to be extracted to distinguish customer service and user voice, and noise is removed at the same time. For historical collection reminder message, the user's reply rate is counted, and the user's reply text is cleaned, including removing irrelevant characters, unifying encoding, standardizing, and removing stop words.

[0055] S103: User intent / emotion recognition.

[0056] Specifically, this step aims to extract features from historical call records and historical collection reminder message records, i.e., voice and text unstructured data. First, the user voice call information is converted to text information by the model to facilitate subsequent analysis of user intent, and the model supports multi-task, which supports recognizing user emotion while completing the voice-to-text task. At this point, we have converted multi-modal data sources into text data and extracted user sentiment negative values as one of the features. A portion of the text data is manually labeled and fed into the nlu module for fine-tuning to meet the needs of the customized collection scenario. The model can be iteratively optimized to classify user intent, i.e., the repayment intention value as a user feature. The specific implementation steps are as follows: S1031: Voice-to-text / emotion recognition multi-task.

[0057] Specifically, for historical call record data, vad detection is first performed to identify when there is a voice in the audio (start and end point detection), filter invalid silence and background noise, which is usually used as a pre-step for voice recognition, and the detection results are merged and sorted to generate multiple batches of data to improve the utilization of computing resources.

[0058] Further, the effective speech segment is sent to the asr speech recognition module after being extracted by the vad, while completing the two tasks of speech to text and emotion recognition. The main process of speech signal feature extraction is: pre-emphasis, framing, windowing, fast fourier transform, mel filtering, and de-meaning, so as to convert the speech signal into fbank features. Feature extraction focuses on the spectral energy information related to human ear perception, and greatly reduces the dimension and improves the calculation efficiency. Then the features are sent to the transformer-encoder structure for multi-task recognition. The specific feature extraction process is as follows: S10311: pre-emphasis.

[0059] Specifically, it is realized by a first-order high-pass filter, and its formula is: (1) The effect of pre-emphasis makes the high-frequency characteristics in the speech signal more obvious, thereby improving the intelligibility of the speech.

[0060] S10312: framing.

[0061] Specifically, the speech signal is not stable in a macroscopic sense, but has short-time stationarity in a microscopic sense, so it needs to be divided into short segments, and each short segment is a frame. In this scheme, the sampling frequency of the audio is 16 kHz, and the length of each frame is 25 ms, so there are sampling points in each frame. Because windowing operation is needed later, in order to avoid the loss of frame edge information, there needs to be an overlap part when framing. The time difference between the starting positions of adjacent two frames is called frame shift. In this scheme, the moving step length between frames is 10 ms, so there are sampling points in each frame shift.

[0062] S10313: windowing.

[0063] Further, the framing in the previous step is equivalent to intercepting the speech segment with a rectangular window: the signals within the frame are retained, and the signals outside the frame are zeroed. The sudden truncation in the time domain will cause the side lobes to spread infinitely in the frequency domain, and the energy will spread from the main frequency to irrelevant frequency bands, causing false peaks in the frequency spectrum and resulting in spectral leakage. The core function of windowing is to suppress spectral leakage. In this scheme, the Hamming window is selected to smooth the frame edge signals by attenuation, so that the truncation boundary becomes continuous, thereby reducing the energy of the frequency domain side lobes. After windowing, the originally aperiodic speech signal will exhibit some characteristics of periodic functions, and the specific formula is as follows: )(2) Where n is the index of the sampling point in the window (starting from 0); N is the total length of the window (frame length); 0.54 and 0.46 are optimization coefficients (minimum side lobe energy).

[0064] S10314: fast fourier transform.

[0065] Specifically, since the characteristics of the signal are usually difficult to see in the time domain, the energy distribution in the frequency spectrum is usually obtained by converting each frame of time domain features to the frequency domain through fast Fourier transform (see Equation 3), and the energy information is extracted by taking the modulus square of the frequency spectrum to obtain the amplitude spectrum of the speech signal (see Equation 4).

[0066] (3) (4) S10315: Mel filter.

[0067] The purpose of this step is to convert the linear frequency into the Mel scale to simulate human auditory perception. In this scheme, 80 triangular filters are used (see Equation 5) to weight and sum the amplitude spectrum to obtain the FBank features (see Equation 6): (5) (6) S10316: LFR (Lower Frame Rate).

[0068] This step will splice multiple consecutive segments into one segment.

[0069] S10317: Feature normalization.

[0070] Specifically, the cepstral coefficients of the audio signal are normalized to improve the comparability and robustness of the features. Thus, the speech signal feature extraction is completed.

[0071] S10318: Feed into model - multi-task recognition.

[0072] Specifically, in this scheme, the SenceVoice-Small open source model is used. In order to support multiple speech understanding functions, the model takes four embeddings of speech features as the input of the encoder, which includes ASR (speech to text) and SER (emotion recognition) required by this scheme. The emotion recognition categories include: angry, happy, neutral, sad, and unknown. The speech feature extraction vector completed in steps 1-7 is fed into the SAN-M memory self-attention network. In the training stage, SER uses the cross-entropy loss function, and ASR uses the CTC loss function (the complete model structure is shown in Figure 3 ).

[0073] S1032: User repayment intention labeling, model fine-tuning.

[0074] S10321: Data labeling.

[0075] Specifically, at the initial stage of the implementation of this scheme, the intention recognition general model needs to be fine-tuned in the customized scene to achieve the ideal recognition effect, so the user call text content obtained needs to be labeled first, and divided into high, medium and low three grades, see Table 1 for specific examples.

[0076] Table 1: User intention labeling examples

[0077] S10322: Model fine-tuning.

[0078] The scheme first fine-tunes the model with labeled data to meet the user intention recognition task in the customized collection scene. The complete recognition steps are: S103221, using SpacyTokenizer for word segmentation.

[0079] S103222, using the word vector pre-training model chinese-roberta-wwm-ext to obtain dense features. The model is a model trained on generalized corpus. The words segmented in the previous step are converted into dense feature matrices carrying various feature information through the pre-training model, but the feature matrix at this time cannot meet the specific scene we need, and needs to be further optimized and adjusted in the specific scene, that is, fine-tune the model with labeled data; At the same time, use multi-hot encoding of n-grams to obtain sparse features.

[0080] S103223, using transformer + Conditional Random Field (CRF) for intention recognition and entity extraction. After adjusting, aligning and splicing the feature latitude through the feedforward neural network (FFW), the sparse features and dense features are sent into the transformer and CRF to output the intention classification and entity, respectively. The loss function adopts dot-product loss point product loss, including intention, entity and mask three parts of the loss function (as shown in Figure 4 ).

[0081] S104: Risk level quantification model construction.

[0082] Specifically, the feature extraction required for user portrait construction has been completed, including overdue times, repayment intention, emotion category, and average response rate. The model is shown in formula 7, and the specific meanings of the variables are shown in Table 2: (7) Table 2: Technical definition of each variable

[0083] Specifically, the formula design principle and implementation details: S1041: Log transformation of overdue times to compress the range.

[0084] Example: .

[0085] S1042: Repayment intention normalization.

[0086] Specifically, the repayment intention in this scheme is finally outputted with a repayment intention category (high / medium / low) and a confidence by fine-tuning the intent recognition model. In order to map the repayment intention into a continuous value between 0 and 1, first define the distribution parameters, see Table 3: Table 3

[0087] Then use truncated normal distribution to construct the probability distribution (see formula 8): (8) Where, ; ; F(x) is the cumulative distribution function (CDF) of the standard normal distribution; [a, b]: truncated range; ; ; Substitute the confidence output by the intent recognition model; output: quantile point parameter value x corresponding to the probability, which is the value of the repayment intention category mapped to between 0 and 1; example: the intent recognition model outputs the repayment intention category as high, with a confidence of 0.8, and the input formula gives the repayment intention value as .

[0088] S1042: Emotion negative value normalization.

[0089] The weight distribution of the impact of emotions on the collection risk is shown in Table 4: Table 4: Emotion classification weight distribution

[0090] Specifically, the output of SenseVoice-Small is a probability dictionary, for example, prob_example = {'angry':0.75,'sad': 0.15, 'neutral': 0.1}, the purpose is to convert the probability dictionary into a negative value between 0 and 1. The way adopted by this scheme is to combine the emotion weight and the probability dictionary to do weighted sum and then sigmoid conversion (see formula 9): (9) Where, is the probability output of the i-th emotion category, is the weight of the i-th emotion category.

[0091] Furthermore, after feature processing, the data is substituted into the risk level quantification model and truncated, then converted to a percentage system, as shown in Formula 10: (10) S105: Choice of communication strategies.

[0092] Specifically, the risk scores are first divided into three categories—high, medium, and low—based on the score ranges in Tables 5, 6, and 7. Then, corresponding collection scripts and strategies are selected according to different users. Table 5: Low-risk users (0-35 points)

[0093] Table 6: Medium-risk users (35-70 points)

[0094] Table 7: High-risk users (70-100 points)

[0095] This invention discloses a collection script selection system based on multimodal debt risk profiling, which enables personalized matching and optimization of collection strategies. This system effectively addresses the technical shortcomings of traditional collection systems, such as simplistic strategies and rigid interactions, significantly improving the accuracy and adaptability of collection operations. While increasing conversion rates, it also optimizes user experience and enhances the overall system performance in complex business scenarios.

[0096] Example 3 This invention also provides a collection script selection device 10 based on a multimodal overdue payment risk profile, such as... Figure 5 As shown, the device includes: The multi-source data acquisition and synchronization module 100 is used to collect multi-source user data from the China Unicom system, including structured and unstructured data. It uses daily partition synchronization technology to synchronize structured data to the data processing platform on a regular basis, and uses fast query and aggregation technology to transmit unstructured data to the data processing platform, so as to establish a complete and real-time accessible data source foundation.

[0097] Specifically, structured data is synchronized to the data processing platform daily via Hive partitions, scheduled for T+1 days. Unstructured data is stored in batches using HBase + Mini0 object storage, and unstructured data of target users is quickly queried and aggregated to the data processing platform via Elasticsearch. The structured data includes basic billing records, payment transaction records, and user package usage information, while the unstructured data includes historical voice call records and historical payment reminder SMS messages.

[0098] The structured data cleaning and feature extraction module 200 is used for dirty data cleaning and duplicate data elimination of the basic bill and payment transaction records, and extracts user behavior features from the cleaned structured data using statistical modeling techniques, including the number of payment overdue times in the past 12 months, to obtain a standardized and analyzable clean data set.

[0099] Specifically, the dirty data cleaning and duplicate data elimination processing are performed on the basic bill records and payment transaction records, and the number of payment overdue times of the user in the past 12 months is counted according to the processed basic bill records and payment transaction records; the audio tracks of two channels of the historical voice call records are extracted, and the noise in the historical voice call records is preliminarily removed; the reply rate of the user of the historical collection reminder message is counted, and the reply text of the user is text cleaned; wherein the two channels include a left channel and a right channel, the left channel is a customer service, and the right channel is a user; the text cleaning includes removing irrelevant characters, unifying coding, standardizing, and removing stop words.

[0100] The multi-modal unstructured data feature extraction module 300 is used for feature extraction of voice call records and collection reminder message records, and converts voice signals to text and simultaneously identifies user emotions using a multi-task voice processing model, and extracts user reply rate and intent value using text cleaning technology and an intent recognition model, to obtain a multi-modal interaction data set containing emotional negative value and repayment intent value.

[0101] Specifically, the historical voice call records of the user are converted into text messages, and the emotions of the user are identified, and the emotional negative value of the user is extracted; the text messages are manually labeled and input into the nlu module for fine-tuning processing, and the nlu module is optimized through multiple iterations to classify the user's intent.

[0102] The risk level quantification model training module 400 is used for training a risk level quantification model using the clean data set and the multi-modal interaction data set, and fusing the overdue times, repayment intent value, emotional negative value, and reply rate into a risk score using a weighted linear model technique combined with logarithmic transformation and normalization processing, and outputting a percentage risk score through truncation processing and percentage conversion; the risk score is classified into three categories of high, medium, and low risk through threshold division technology, and the corresponding collection script process and operation strategy are automatically selected according to the risk category using a strategy mapping algorithm.

[0103] Specifically, the number of overdue payment times is logarithmically transformed and the range is compressed; the distribution parameters of the classification of the user's intention are defined, and then the probability distribution of the distribution parameters is constructed by using a truncated normal distribution function to obtain a value between 0 and 1 that the classification of the intention is mapped to; the historical voice call record is input into a SenseVoice-Small model to obtain a probability dictionary, and a negative value of emotion between 0 and 1 is obtained by combining the weighted sum of the emotion weight and the probability dictionary and the sigmoid conversion; the calculated features are brought into a risk level quantification model to generate a risk score, and the risk score is truncated and converted into a percentage form to obtain a percentage risk score; When the outbound NLP interface is called, the risk score and the features of the corresponding user are input, and a corresponding dialogue strategy is selected; a one-way transformer model is used to fine-tune the dialogue strategy to obtain a relatively optimal dialogue strategy.

[0104] The device for selecting a collection dialogue based on a multi-modal overdue payment risk portrait of the embodiment of the application realizes end-to-end intelligent decision-making from data collection, feature extraction to risk quantification and dialogue generation. The device effectively connects the whole process of heterogeneous data integration, user portrait construction, risk accurate assessment and personalized strategy output, significantly improves the accuracy and system adaptive ability of the collection operation, effectively balances the collection efficiency and user experience while optimizing resource allocation, and enhances the overall performance of the system in complex business scenarios.

[0105] The above only describes the preferred embodiments of the application and is not intended to limit the application. For those skilled in the art, the application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the application shall be included in the protection scope of the application.

[0106] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms is not necessarily directed to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, the skilled in the art can combine and combine the different embodiments or examples described in the present specification and the features of the different embodiments or examples without contradiction.

[0107] Furthermore, the terms "first", "second", "third", "fourth", "fifth" and "sixth" are used herein for descriptive purposes only and are not to be construed as indicating or implying relative importance or a significant nature of so described technical features. It is to be understood that a technical feature described with the "first", "second", "third", "fourth", "fifth" or "sixth" can implicitly or explicitly include at least one of the technical features described with the "first", "second", "third", "fourth", "fifth" or "sixth". In the description of the present application, the meaning of "a plurality" is at least two, for example, two, three, etc., unless otherwise specifically defined.

Claims

1. A method for selecting a collection script based on a multi-modal delinquent risk portrait, characterized in that, The application relates to a method for constructing a dynamic risk quantification model based on multi-source user data, and a system thereof. The method comprises the following steps: S1, collecting multi-source user data required by a communication system; wherein the multi-source user data comprises basic billing records, payment flow records and user package usage data in structured data, and historical call records and historical payment reminder message data in unstructured data; S2, pre-processing the multi-source user data, extracting the payment overdue times of the structured data, and performing speech signal framing and windowing processing, text cleaning and intent recognition on the unstructured data to generate user emotion features and repayment intention features; S3, constructing a dynamic risk quantification model based on the payment overdue times, emotion features and repayment intention features, mapping the repayment intention confidence by using a truncated normal distribution to generate a repayment intention value, combining the emotion weight to perform weighted summation and then performing sigmoid conversion to calculate the user risk score; 2. The method of claim 1, wherein, S4, dividing the risk score into high / medium / low risk levels according to a preset risk threshold, and calling corresponding dialogue strategy templates based on the risk levels, wherein strong dialogue strategies of penalty details and shutdown countdown are adopted for high-risk users, preferential inducement dialogue strategies combined with package usage habits are adopted for low-risk users, and mixed strategy dialogue strategies are adopted for medium-risk users. The S1 comprises the following steps: The structured data is synchronously transmitted to a data processing platform in a Hive daily partitioning mode every T+1 day; 3. The method of claim 2, wherein, The unstructured data is stored in a Hbase+MinI0 object storage mode, and the target user's unstructured data is collected to the data processing platform through ES fast query. The S2 comprises the following steps: Dirty data cleaning and repeated data elimination are performed on the basic billing records and payment flow records, and the payment overdue times of a user in the last 12 months are counted according to the processed basic billing records and payment flow records; The audio tracks of two channels of the historical voice call records are extracted, and the noise in the historical voice call records is preliminarily removed; The reply rate of the user of the historical payment reminder message is counted, and the reply text of the user is cleaned; 4. The method of claim 3, wherein, The two channels comprise a left channel and a right channel, the left channel is a customer service, and the right channel is a user; the text cleaning comprises removing irrelevant characters, unifying coding, standardizing, and removing stop words. The S3 comprises the following steps: The historical voice call records of the user are converted into text messages, the emotion of the user is recognized, and the emotional negative value of the user is extracted; 5. The method of claim 4, wherein, The text messages are manually labeled and input into an nlu module for fine-tuning processing, the nlu module is iteratively optimized multiple times, and the classification of the user's intention is realized. The S4 comprises the following steps: The payment overdue times are logarithmically transformed and compressed, the distribution parameters of the classification of the user's intention are defined, then a truncated normal distribution function is used to construct the probability distribution of the distribution parameters to obtain a value mapped to 0-1 between the classification of the intention, the historical voice call records are input into a SenseVoice-Small model to obtain a probability dictionary, the emotional negative value between 0-1 is obtained by combining the emotion weight and the weighted summation of the probability dictionary and the sigmoid conversion, and the related features are calculated. The calculated relevant features are brought into a risk level quantification model to generate a risk score, and the risk score is subjected to truncation processing and converted into a percentage form to obtain a risk score in percentage.

6. The method of claim 5, wherein, Further comprising S5: The risk score and features of the corresponding user are collected when the outbound NLP interface is called, and the corresponding dialogue strategy is selected; The dialogue strategy is fine-tuned by using a one-way transformer model to obtain a relatively optimal dialogue strategy.

7. A device for selecting a collection script based on a multi-modal delinquent risk portrait, characterized in that, Comprise: A multi-source data acquisition and synchronization module is used to acquire multi-source user data from a communication system, including structured data and unstructured data, and to synchronize the structured data to a data processing platform in time by using daily partition synchronization technology, and to transmit the unstructured data to the data processing platform by using fast query and collection technology, so as to establish a complete and real-time accessible data source basis; A structured data cleaning and feature extraction module is used to clean dirty data and remove duplicate data from basic bills and payment records, and to extract user behavior features from the cleaned structured data by using statistical modeling technology, including the number of payment overdue times in the past 12 months, to obtain a standardized and analyzable clean data set; A multi-modal unstructured data feature extraction module is used to extract features from voice call records and payment reminder message records, and to convert voice signals into text and recognize user emotions at the same time by using a multi-task voice processing model, and to extract user reply rate and intention value by using text cleaning technology and intention recognition model, to obtain a multi-modal interaction data set containing negative emotion value and repayment intention value; A risk level quantification model training module is used to train a risk level quantification model by using a clean data set and a multi-modal interaction data set, and to fuse overdue times, repayment intention value, negative emotion value and reply rate into a risk score by using a weighted linear model technique combined with logarithmic transformation and normalization processing, and to output a risk score in percentage by truncation processing and percentage conversion; A risk classification and collection strategy application module is used to classify risk scores into three categories of high, medium and low risk by threshold division technology, and to automatically select corresponding collection dialogue processes and operation strategies according to risk categories by using strategy mapping algorithm.

8. The apparatus of claim 7, wherein, The multi-modal unstructured data feature extraction module comprises: The user's historical voice call records are converted into text messages, and the user's emotions are recognized to extract the negative emotion value of the user; The text messages are manually labeled and input into the NLU module for fine-tuning processing to optimize the NLU module through multiple iterations to classify the user's intention.

9. The apparatus of claim 7, wherein, The risk level quantification model training module comprises: Logarithmic transformation is performed on the number of overdue payment times and the range is compressed; the distribution parameters of the classification of the user's intention are defined, and then the probability distribution of the distribution parameters is constructed by using the truncated normal distribution function to obtain the mapping of the classification of the intention to a value between 0 and 1; the historical voice call record is input into a SenseVoice-Small model to obtain a probability dictionary, and a negative value of emotion between 0 and 1 is obtained by combining the weighted sum of the emotion weight and the probability dictionary and the sigmoid conversion to calculate the relevant features; The relevant features calculated are brought into a risk level quantification model to generate a risk score, and the risk score is truncated and converted into a percentage form to obtain a percentage risk score.

10. The apparatus of claim 7, wherein, The risk classification and collection strategy application module comprises: When the outbound nlp interface is called, the risk score and features of the corresponding user are collected, and the corresponding dialogue strategy is selected; A one-way transformer model is used to fine-tune the dialogue strategy to obtain a relatively optimal dialogue strategy.

Images