Neural network-based household situation evaluation method and device, and storage medium

Abstract

The application provides a neural network-based family situation evaluation method and device and a storage medium. The method first converts structured family numerical data and unstructured text descriptions into computable feature vectors through artificial rule coding and large model semantic coding, achieving comprehensive digital characterization of the family situation. Furthermore, the core economic indicators are accurately predicted through a classic neural network, and comprehensive reasoning is performed under the guidance of structured prompts by a large language model, finally outputting evaluation results with both quantitative risk scores and interpretable basis. This scheme not only greatly improves the accuracy and automation of family economic situation evaluation and policy matching, but also ensures the continuous evolution of the system through modular design and periodic fine-tuning mechanism, providing efficient, reliable and traceable intelligent decision support for precise assistance.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus and storage medium for assessing family circumstances based on neural networks. Background Technology

[0002] With the deepening application of big data and artificial intelligence technologies in public services, using intelligent methods for family situation assessment and precise assistance has become an important direction for improving social governance efficiency. Currently, existing technical solutions mostly rely on traditional machine learning models (such as decision trees and logistic regression) or early neural networks. These methods primarily process structured numerical data (such as income and population), using historical statistical patterns for risk classification or policy recommendations. However, these existing technologies have significant limitations: First, they typically cannot effectively process and integrate the large amount of unstructured textual descriptive information in family situations (such as descriptions of difficulties and health conditions), leading to incomplete information utilization; second, existing models lack the ability to deeply understand complex policy texts, and policy matching is often based on keywords or simple rules, resulting in insufficient accuracy; third, traditional systems are often open-loop static models, making it difficult to continuously self-optimize and iterate using the actual assistance results generated after assessment, and the accuracy and adaptability of the assessment model decline over time. Therefore, there is an urgent need for an intelligent assessment method that can deeply integrate multimodal data, understand policy semantics, and continuously evolve. Summary of the Invention

[0003] This invention provides a method, apparatus, and storage medium for assessing family circumstances based on neural networks, in order to overcome the deficiencies in the prior art and improve the accuracy, efficiency, and reliability of family circumstances assessment.

[0004] This invention provides a method for assessing family circumstances based on neural networks, comprising: Acquire policy text data and family multimodal data of the households to be assessed, wherein the family multimodal data includes structured numerical data and textual description data; The structured numerical data is mapped according to rules by an artificial feature encoding unit to obtain a first feature vector; The text description data is semantically encoded using a semantic encoding unit to obtain a second feature vector. The first feature vector and the second feature vector are input into the economic indicator prediction model, and the predicted income and expenditure values ​​of the household to be evaluated are output. Based on a structured prompt engineering template, the predicted income values, predicted expenditure values, and policy text data are encapsulated to form prompt instructions; The prompts are input into the family situation assessment model to generate a comprehensive assessment result that includes a family risk score and the basis for the score.

[0005] According to a neural network-based family situation assessment method provided by the present invention, the structured numerical data includes per capita annual family income, total family debt, and number of family members in the labor force. The step of obtaining a first feature vector by mapping the structured numerical data based on rules through a manual feature encoding unit includes: The average annual income per capita of the households is divided into groups and converted into income level labels; the total debt of the households is logarithmically normalized. Calculate the debt-to-asset ratio based on the total household liabilities and total household assets; calculate the labor force ratio based on the number of working-age members and the total number of members in the household; The income level label, normalized total household debt, debt-to-asset ratio, and labor force ratio are concatenated to form the first feature vector.

[0006] According to the present invention, a method for assessing family circumstances based on a neural network, wherein the text description data includes textual descriptions of family members' health status, employment difficulties, or special circumstances, and the step of semantically encoding the text description data through a semantic encoding unit to obtain a second feature vector specifically includes: The text description data is input into the semantic encoding unit, which is a pre-trained large language model based on the Transformer architecture; Obtain the [CLS] tag vector output by the last hidden layer of the semantic encoding unit, which corresponds to the overall semantics of the text description data, as the initial semantic representation; The initial semantic representation is subjected to layer normalization to output a dense vector of fixed dimensions, which serves as the second feature vector.

[0007] According to the present invention, a method for assessing family circumstances based on a neural network is provided, wherein the economic indicator prediction model is a multilayer perceptron model, which includes an input layer, at least one hidden layer and an output layer. The output layer contains two neurons, which are used to output the predicted income value and the predicted expenditure value, respectively.

[0008] According to the present invention, a family situation assessment method based on neural networks is provided, wherein the comprehensive assessment result is a structured data object, which includes at least a family situation risk score and corresponding key factor analysis.

[0009] According to the present invention, a method for assessing household circumstances based on neural networks, after the step of acquiring policy text data and multimodal household data of the household to be assessed, further includes... The structured numerical data is filled with missing values ​​or corrected for outliers. Sensitive information desensitization and key information enhancement processing are performed on the text description data.

[0010] According to a neural network-based family situation assessment method provided by the present invention, after the step of inputting the prompt instruction into the family situation assessment model to generate a comprehensive assessment result including a family risk score and the scoring basis, the method further includes: The training sample set is formed by periodically collecting the first feature vector, the second feature vector, the predicted income and expenditure values ​​of the households to be evaluated, and the comprehensive evaluation results. The parameters of the economic indicator prediction model and / or the household situation assessment model are fine-tuned using the training sample set.

[0011] The present invention also provides a family situation assessment device based on a neural network, comprising: The data acquisition module is used to acquire policy text data and family multimodal data of the households to be assessed. The family multimodal data includes structured numerical data and text description data. The manual coding module is used to perform rule-based mapping on the structured numerical data through the manual feature coding unit to obtain the first feature vector; A semantic encoding module is used to perform semantic encoding on the text description data through a semantic encoding unit to obtain a second feature vector; The indicator prediction module is used to input the first feature vector and the second feature vector into the economic indicator prediction model, and output the income prediction value and expenditure prediction value of the household to be evaluated. The data encapsulation module is used to encapsulate the income forecast, expenditure forecast, and policy text data based on a structured prompt engineering template to form prompt instructions; The comprehensive assessment module is used to input the prompts into the family situation assessment model to generate a comprehensive assessment result that includes a family risk score and the basis for the score.

[0012] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the neural network-based family situation assessment method described above.

[0013] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the neural network-based family situation assessment method as described above.

[0014] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the neural network-based family situation assessment method as described above.

[0015] The present invention provides a neural network-based family situation assessment method, device, and storage medium that, through innovative architectural design, effectively overcomes the shortcomings of existing technologies, bringing significant technological advancements and practical value. First, this method creatively proposes and implements the collaborative processing and fusion analysis of "family multimodal data" (structured numerical data and unstructured text). By using "artificial feature encoding units" to perform domain-knowledge-based rule-based mapping of numerical data (such as income and debt), and by using "semantic encoding units" (based on a large language model) to perform deep semantic encoding of text descriptions, the two types of heterogeneous information are unified into feature vectors that can be efficiently processed by machines. This achieves the construction of a comprehensive and three-dimensional portrait of the family situation, solving the problem of single information utilization in traditional methods. Second, this method designs a two-stage analysis and reasoning process: first, an "economic indicator prediction model" (such as MLP) accurately predicts key economic indicators (income and expenditure) based on the fused feature vectors, achieving a quantitative grasp of the family's economic situation; then, a "family situation assessment model" (based on a large language model) performs comprehensive reasoning and decision generation based on the economic prediction results and policy text encapsulated in a structured prompting template. This design combines the stability of classical neural networks in numerical prediction with the powerful capabilities of large language models in complex semantic understanding and logical reasoning. This ensures the accuracy of core economic indicator calculations while achieving high-level risk assessment and interpretable policy matching. Ultimately, the system outputs structured comprehensive assessment results, including quantified risk scores and clear scoring criteria. This provides intelligent support for assistance decisions that is objective, accurate, and interpretable, significantly improving the precision, efficiency, and reliability of family situation assessments. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the family situation assessment method based on neural networks provided by the present invention; Figure 2 This is a network structure diagram of the family situation assessment method based on neural networks provided by the present invention; Figure 3This is a schematic diagram of the structure of the family situation assessment device based on neural networks provided by the present invention; Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0019] To address the problems in existing technologies, this invention proposes a neural network-based family situation assessment method to improve the accuracy, efficiency, and reliability of family situation assessment. The neural network-based family situation assessment method is described below, as follows: Figure 1 As shown, including but not limited to the following steps: Step 110: Obtain policy text data and family multimodal data of the families to be evaluated. The family multimodal data includes structured numerical data and textual description data.

[0020] In step 110, the system obtains information from two main data sources.

[0021] Policy text data: Obtain the full text or key clauses of currently effective policy documents at all levels related to family assistance from government public information platforms and internal policy databases. For example, obtain the specific chapters on monitoring standards and assistance measures from a province's "Implementation Plan for Dynamic Monitoring and Assistance of Family Status".

[0022] Household multimodal data: Complete files of the households to be assessed are obtained through grassroots information collection systems, historical databases, or API interfaces. This household multimodal data includes: Structured numerical data: Quantitative information stored in database tables, such as: annual household income (unit: yuan), annual household expenditure (unit: yuan), number of household members, number of working-age members, age list of members, total liabilities (unit: yuan), total assets (unit: yuan), per capita living area (unit: square meters), etc.

[0023] Textual description data: Descriptive information recorded in natural language, such as: "The head of household suffers from a chronic disease and needs to take medication for a long time, making him unable to engage in heavy physical labor" in the "Explanation of Family Difficulties" field; "The household's house is a timber structure that has been in disrepair for many years and poses a safety hazard" recorded in the "Grassroots Visit Record"; or "There is a student in the family who is studying at university, and the education expenses are relatively large" mentioned in the "Remarks".

[0024] Step 120: The structured numerical data is mapped according to rules by an artificial feature encoding unit to obtain the first feature vector.

[0025] In step 120, the system invokes a feature engineering module (i.e., a manual feature encoding unit) with preset rules to automatically transform the structured numerical data obtained in step 110. This module presets a series of mapping rules based on domain knowledge, converting the original numerical data into features that are more model-friendly and have higher information density. A specific embodiment is as follows: The "average annual household income (yuan)" is binned: a threshold is set (e.g., 5,000 yuan, 10,000 yuan), and it is mapped to a discrete "income level label" (e.g., 0 represents "below 5,000", 1 represents "5,000-10,000", 2 represents "above 10,000").

[0026] Logarithmic normalization is performed on “Total Household Debt (Yuan)”: calculate log(debt amount + 1) to alleviate data skew and scale the result to the [0,1] interval.

[0027] Feature Derivation: Calculate the "debt-to-asset ratio" (liabilities / assets) based on "total household liabilities" and "total household assets"; calculate the "labor force ratio" based on "labor force population" and "total household population".

[0028] Finally, all the scalar or categorical features processed above (such as income level labels, normalized liabilities, debt-to-equity ratio, labor force ratio, etc.) are concatenated in a fixed order into a one-dimensional numerical array, which is the first feature vector, for example, a vector with a dimension of [1×32].

[0029] Step 130: Semantically encode the text description data using a semantic encoding unit to obtain a second feature vector.

[0030] In step 130, the system uses a pre-trained large language model (such as BERT, ERNIE, etc.) based on the Transformer architecture as the semantic encoding unit. The specific process is as follows: Input the text description data obtained in step 110 (which may have been processed by word segmentation, truncation or concatenation) into the large language model.

[0031] Extract the semantic representation of the entire input sequence from the output of the last hidden layer of the model. Typically, the output vector corresponding to the special marker [CLS] is taken as the aggregate semantic representation of the entire text.

[0032] The [CLS] vector is subjected to layer normalization to stabilize training and accelerate convergence.

[0033] Output a dense vector of fixed dimensions (e.g., 768 dimensions) as the second feature vector. This vector captures deep semantic information from the text description regarding family difficulties, health conditions, and special circumstances.

[0034] Step 140: Input the first feature vector and the second feature vector into the economic indicator prediction model, and output the income prediction value and expenditure prediction value of the household to be evaluated.

[0035] In step 140, the system achieves the fusion of multimodal features and prediction of key indicators.

[0036] First, the first feature vector (numerical feature) obtained in step 120 and the second feature vector (semantic feature) obtained in step 130 are concatenated to form a fused feature vector.

[0037] Subsequently, the fused feature vector is input into a pre-trained economic indicator prediction model. In a preferred embodiment, the model is a multilayer perceptron, the structure of which includes: an input layer (the dimension of which matches the fused feature vector), two fully connected hidden layers (each using the ReLU activation function), and an output layer.

[0038] This output layer contains two neurons, corresponding to the "predicted annual household income" and the "predicted annual household expenditure," respectively. These two neurons typically use a linear activation function to directly output the numerical results of the regression predictions, in units of yuan.

[0039] Step 150: Based on the structured prompt engineering template, encapsulate the income forecast, expenditure forecast, and policy text data to form prompt instructions.

[0040] In step 150, the system organizes the numerical prediction results and policy text into task instructions that the large language model can understand. The system maintains one or more predefined structured prompt engineering templates. Each template is a parameterized text string containing the following parts: System role instruction: "You are a family economic situation and policy matching assessment expert." Task Context and Data Input: "Please analyze the following family's economic forecast data and related support policies: Forecasted annual income is {income_pred} yuan, and forecasted annual expenditure is {expense_pred} yuan. The relevant policy text is as follows: {policy_text}" Specific task requirements: "Please perform the following operations: 1) Assess the family's income and expenditure balance and economic vulnerability risk; 2) Determine whether the family meets the assistance criteria in the above policies; 3) Provide a comprehensive risk score from 1 to 100; 4) List the main basis for the score in detail." Output format instructions: "Please output in JSON format, including the risk_score and assessment_reasons fields." During implementation, the system will fill the corresponding placeholders in the template with the specific values ​​of {income_pred} and {expense_pred} predicted in step 140, as well as the {policy_text} obtained in step 110, to generate a complete and structured prompt instruction text.

[0041] Step 160: Input the prompt instruction into the family situation assessment model to generate a comprehensive assessment result that includes a family risk score and the basis for the score.

[0042] In step 160, the system performs the final comprehensive reasoning and decision generation.

[0043] The prompts generated in step 150 are input into a dedicated family situation assessment model. This model is typically a finely tuned large language model (such as one based on LLaMA or ChatGLM) with strong reasoning and instruction-following capabilities.

[0044] The model analyzes prompts and instructions, comprehensively understands economic forecast data and policy provisions, and performs risk assessment, policy matching, and logical reasoning.

[0045] Finally, the model strictly follows the JSON format required by the instructions to generate and output structured comprehensive evaluation results. For example: "risk_score": 85, "assessment_reasons": "1. The projected income-expenditure gap is negative (-5000 yuan), indicating a serious risk of deficit. 2. The household's debt-to-asset ratio is as high as 60%, resulting in significant financial pressure. 3. According to Article X of the policy, 'Medical assistance will be provided to households with per capita income below Y yuan and a family with a serious illness,' this household meets the criteria. 4. The proportion of the workforce is low, limiting the potential for income growth." As a further optional embodiment, the structured numerical data includes per capita annual household income, total household debt, and number of household labor force members. The step of mapping the structured numerical data to a first feature vector using a manual feature encoding unit based on rules includes: The average annual income per capita of the households is divided into groups and converted into income level labels; the total debt of the households is logarithmically normalized. Calculate the debt-to-asset ratio based on the total household liabilities and total household assets; calculate the labor force ratio based on the number of working-age members and the total number of members in the household; The income level label, normalized total household debt, debt-to-asset ratio, and labor force ratio are concatenated to form the first feature vector.

[0046] In this embodiment, the specific rule mapping operation performed by the artificial feature encoding unit is further clarified and optimized. First, for the key continuous variable of "average annual household income per capita," multiple threshold intervals are preset (e.g., below 5000 yuan, 5000-10000 yuan, above 10000 yuan), and binning is used to convert it into discrete "income level labels" (e.g., 0, 1, 2). This operation not only reduces the noise sensitivity of the data but also transforms continuous values ​​into more business-interpretable categories, making it easier for the model to capture nonlinear relationships. Second, for the data of "total household debt," which is usually distributed in a long tail, "logarithmic normalization" is used, that is, its logarithm (log(debt amount + 1)) is calculated first, and then min-max scaling is performed to the [0,1] interval. This method can effectively compress the range of extreme values, making the data distribution closer to normal, and improving the stability and convergence speed of model training.

[0047] More importantly, this embodiment introduces feature derivation rules based on domain knowledge. By correlating "total household debt" with "total household assets," the "debt-to-asset ratio" (liabilities / assets) is calculated. This ratio directly reflects the household's financial leverage and risk level, making it a more discerning indicator than simply the amount of debt. Simultaneously, by combining "number of working-age members" with "total number of household members," the "labor force ratio" is calculated. This indicator quantifies the household's potential self-sufficiency and is one of the core dimensions for assessing its development potential and vulnerability. Finally, the four key features generated by the above rules—"income level label," "normalized total household debt," "debt-to-asset ratio," and "labor force ratio"—are concatenated in a predetermined order into a one-dimensional numerical array, thus forming a concise and business-meaningful first feature vector. This design ensures that the features input to the downstream prediction model not only contain original quantitative information but also deeply integrate expert experience and business logic, laying a solid foundation for subsequent accurate predictions.

[0048] As a further optional embodiment, the text description data includes textual descriptions of family members' health status, employment difficulties, or special circumstances. The step of semantically encoding the text description data through a semantic encoding unit to obtain a second feature vector specifically includes: The text description data is input into the semantic encoding unit, which is a pre-trained large language model based on the Transformer architecture; Obtain the [CLS] tag vector output by the last hidden layer of the semantic encoding unit, which corresponds to the overall semantics of the text description data, as the initial semantic representation; The initial semantic representation is subjected to layer normalization to output a dense vector of fixed dimensions, which serves as the second feature vector.

[0049] In this embodiment, the core task of the semantic encoding unit is to transform unstructured textual descriptions into machine-understandable and computable numerical features. Specifically, a Transformer-based large language model (such as BERT, RoBERTa, or similar variants) pre-trained on a large amount of general text is selected as the base encoder. These models, through their self-attention mechanism, can deeply capture the long-range dependencies and contextual semantics between words in the text, making them particularly suitable for processing narrative texts containing complex logic and implicit information, such as "family member health status" or "description of employment difficulties."

[0050] During encoding, the model automatically adds a special [CLS] (classification) tag at the beginning of the input text. This tag is optimized during model training to aggregate global semantic information from the entire input sequence. Therefore, during inference, the hidden state vector corresponding to the [CLS] tag, output from the last layer of the encoder, is taken as the "initial semantic representation" of the entire text description. This vector is a high-dimensional representation (e.g., 768-dimensional or 1024-dimensional) containing rich contextual information.

[0051] To further improve the quality of features and the stability of subsequent processing, layer normalization is performed on the acquired initial semantic representation. This step adjusts the values ​​of each dimension in the vector to a similar scale range, effectively alleviating the internal covariate shift problem, accelerating model convergence, and enhancing the consistency of features across different samples. After normalization, a dense vector with fixed dimensions is finally output, which is the second feature vector. This vector condenses key semantic information about health, employment, and special circumstances from the original text, providing high-quality, structured textual semantic feature input for downstream economic indicator prediction models. It is a crucial step in achieving multimodal information fusion and accurate evaluation.

[0052] As a further optional embodiment, the economic indicator prediction model is a multilayer perceptron model, which includes an input layer, at least one hidden layer and an output layer. The output layer contains two neurons, which are used to output the predicted income value and the predicted expenditure value, respectively.

[0053] In this embodiment, the economic indicator prediction model is specifically implemented using the classic and efficient multilayer perceptron model. The model's architecture comprises three key parts: First, an input layer whose number of neurons strictly matches the dimension of the fused feature vector generated in the upstream steps, ensuring that all feature information can be input into the model without loss. Second, at least one hidden layer; these fully connected layers use nonlinear activation functions (such as ReLU or GELU) to learn and extract complex, high-level abstract patterns and nonlinear relationships from the input that integrates numerical regular features and textual semantic features. This is crucial for the model's powerful predictive capabilities. Finally, an output layer, specially designed to contain only two neurons, each typically employing a linear activation function. The output value of the first neuron is directly parsed as the predicted household income (in yuan), and the output value of the second neuron is parsed as the predicted expenditure (in yuan). This design clearly defines the model's core prediction task, enabling it to accurately fit these two core continuous indicators of household economic status through regression learning. This MLP model is trained in a supervised manner using a large amount of historical household data (including features and real income / expenditure labels) and optimized using loss functions such as mean squared error. It learns to map reliable economic indicator predictions from complex multimodal household features, providing accurate and structured quantitative input for subsequent comprehensive evaluation based on large models.

[0054] As a further optional embodiment, the comprehensive assessment result is a structured data object, which includes at least a family situation risk score and corresponding key factor analysis.

[0055] In this embodiment, the system's final output is strictly formatted as a machine-readable and human-understandable structured data object to maximize its practical application value. The household situation risk score is a quantified integer value, typically defined between 1 and 100; a higher score indicates a greater economic vulnerability risk to the household. This score is not generated through simple weighting, but rather as a quantitative judgment derived by the household situation assessment model through comprehensive reasoning, weighing the matching degree between predicted economic data and policy provisions.

[0056] More importantly, the key factor analysis, presented in the form of structured text fields, explicitly explains the logic behind the scoring. This analysis is not a general overview, but a specific and traceable list of key points. For example, the analysis clearly states: "1. The projected deficit reaches XX yuan, exceeding the warning threshold"; "2. The debt-to-asset ratio (XX%) is in the high-risk range"; "3. According to Article Y of the 'XX Policy,' this household fully meets the conditions for medical assistance due to the presence of a member with a serious illness and an average income below Z yuan per capita"; "4. The text description mentions 'safety hazards in the housing,' constituting an additional risk dimension." This method of directly linking quantitative conclusions with specific numerical evidence, policy provisions, and detailed textual descriptions greatly enhances the credibility, interpretability, and auditability of the assessment results. These structured results can be directly integrated and utilized by subsequent decision support systems, early warning platforms, or reporting tools, achieving seamless integration from intelligent analysis to automated business processes and completing the assessment loop.

[0057] As a further optional embodiment, after the step of acquiring policy text data and household multimodal data of the households to be assessed, the method further includes... The structured numerical data is filled with missing values ​​or corrected for outliers. Sensitive information desensitization and key information enhancement processing are performed on the text description data.

[0058] In this embodiment, a crucial data preprocessing and enhancement stage is introduced before executing the core analysis process. This stage aims to improve the quality, security, and information value of the input data, thereby ensuring the accuracy and reliability of subsequent model evaluation. This stage mainly includes two parallel processing flows: 1. Cleaning and Correction of Structured Numerical Data For structured numerical data (such as income, liabilities, and population) obtained from databases or forms, the system automatically performs data quality verification and repair procedures. For detected missing values, the system intelligently fills in the missing data according to preset strategies: for example, for a missing "annual household income" field, it can be filled using the median income of similar households in the same region; for a missing "labor force population," it can be inferred based on family member age rules. For outliers, the system identifies and corrects them using statistical methods (such as outlier detection based on quantiles) or business rules: for example, when "per capita household medical expenditure" significantly exceeds a reasonable range, the system will correct it to the reasonable upper limit for that population in that region, or mark it for manual review. These processes effectively reduce the interference of noisy data on model training and improve the robustness of feature engineering.

[0059] 2. Security and Information Enhancement Processing of Text-Described Data For textual descriptions containing personal privacy and sensitive details, the system first performs sensitive information anonymization. Using a pre-trained named entity recognition model or rule engine, it automatically identifies and masks sensitive information such as personal ID numbers, phone numbers, and detailed addresses in the text, replacing them with uniform placeholders (e.g., [PHONE]), preserving the text's grammatical and semantic structure while protecting personal privacy. Simultaneously, the system performs key information enhancement processing: using natural language processing techniques, it standardizes and supplements abbreviated, vague, or colloquial descriptions. For example, "elderly person is ill" is automatically associated and labeled as "there is an elderly family member (age > 65 years) with poor health"; or "unable to find a job" is enhanced to describe "there is difficulty in finding employment, and the main workforce is unemployed." This process not only makes the text information more standardized and complete but also explicitly highlights the risk dimensions crucial for assessment, providing higher-quality, more discriminative text input for subsequent semantic encoding.

[0060] Through the preprocessing and enhancement steps in this embodiment, the system ensures that the data flowing into the core analysis pipeline is clean, safe, standardized, and information-enhanced, laying a solid data foundation for the high accuracy and reliability of the entire evaluation method.

[0061] As a further optional embodiment, after the step of inputting the prompt instruction into the family situation assessment model to generate a comprehensive assessment result including a family risk score and the basis for the score, the method further includes: The training sample set is formed by periodically collecting the first feature vector, the second feature vector, the predicted income and expenditure values ​​of the households to be evaluated, and the comprehensive evaluation results. The parameters of the economic indicator prediction model and / or the household situation assessment model are fine-tuned using the training sample set.

[0062] In this embodiment, the system introduces a dynamic, continuously self-optimizing learning loop, which is the core mechanism to ensure that the evaluation model can adapt to policy changes, data distribution shifts, and continuously improve performance. This mechanism, as an offline background task, is executed automatically at a preset period (e.g., weekly, monthly) and includes two key stages: 1. Dynamic construction of training sample sets During each execution cycle, the system automatically collects family case data that have been assessed and have received final assistance decisions in the previous cycle. For each case, historical family multimodal data and historical policy text data constitute the input features (X), while the actual assistance outcome serves as the training label (Y). The actual assistance outcome refers to the record of assistance measures that have been finally confirmed and implemented by humans, such as "being included in the minimum living allowance," "receiving medical assistance of XX yuan," or "being assigned to a public service position." This record is considered the "real-world situation" of the family's true risk and needs, and is the gold standard for model learning. The system automatically organizes and aligns these "input-output" pairs to form new training samples, which are then added to the continuously accumulating historical training sample set.

[0063] 2. Efficient fine-tuning of model parameters Using the newly constructed training sample set, the system performs targeted updates to the core models. For the economic indicator prediction model, fine-tuning aims to make its income and expenditure predictions closer to the actual economic situation and optimize the loss function used (such as mean squared error). For the household situation assessment model, fine-tuning aims to optimize its comprehensive reasoning ability, making its generated risk scores and basis more consistent with the risk levels determined based on actual assistance results.

[0064] To achieve efficient and stable model updates while avoiding catastrophic forgetting, this embodiment preferably employs efficient parameter fine-tuning techniques, such as the Low-Rank Adaptive (LoRA) method. This method does not update all parameters of a large language model, but rather adjusts model behavior by injecting and training a small number of low-rank trainable matrices. This significantly reduces computational overhead and storage requirements while achieving excellent fine-tuning results. The entire fine-tuning process is completed offline, without interfering with the normal operation of the online evaluation service.

[0065] The newly refined model version undergoes a rigorous offline validation process, evaluating its performance on an independent test set. Only when the new model's evaluation metrics (such as prediction accuracy and F1 score) significantly outperform the current online version will an orderly model switchover process be triggered, deploying the new model online to replace the old one. Through this closed-loop design, this embodiment enables the entire evaluation system to continuously learn and iterate from actual business feedback, thereby maintaining its evaluation accuracy, timeliness, and policy adaptability in the long term.

[0066] Understandably, the distinction between the aforementioned three models should not be a limitation of the patent; the method of combining semantic encoding and evaluation models and unifying training through fine-tuning is the claim of this patent.

[0067] Figure 2A system architecture diagram of a preferred embodiment of the present invention is shown. As shown, the system is integrated into a unified intelligent evaluation model. Its core architecture adopts a dual-path encoding, phased prediction, and fusion inference design. The specific data flow and module functions are as follows: 1. Multimodal data input and split coding The system's input data is divided into two main categories and processed by different encoders: Left-hand path: Structured / Categorical data encoding The input is the structured and categorical data of families managed by the “Job Description (8)” module, specifically including: age, gender, education level, physical disability status and other category data, as well as income range, employment status and other data that may be covered by “Other Category Data (8)”.

[0068] This type of data is processed by the "Artificial Vector Encoding (2)" unit. This unit converts the raw data into a numerical first feature vector based on predefined business rules and mapping tables (e.g., mapping education level to an ordered value, binning age, etc.). This process requires no model training and is entirely driven by domain knowledge.

[0069] Right-hand path: Encoding of text and unstructured data The input is text description information such as "other family description information, including sudden serious illness, accident (7)", "already enjoying the policy (7)" and "related policies (7)".

[0070] This type of unstructured text data is uniformly input into the "Large Model (Vector Encoding)(1)" unit. This unit is based on a pre-trained Transformer large language model (such as BERT, ERNIE, etc.) to perform deep semantic understanding and encoding of the text, and outputs a second feature vector that can capture the core semantics of the text.

[0071] 2. Economic Indicator Forecasting Stage The first feature vector output by the "artificial vector encoding (2)" is fed into a lightweight "small neural network (3)". This network is preferably a multilayer perceptron model.

[0072] Based on the structured features of the input, the MLP network performs a regression prediction task and outputs two key numerical results: income prediction (6) and expenditure prediction (6). These two predictions form the quantitative basis for subsequent comprehensive evaluation.

[0073] 3. Comprehensive Assessment and Reasoning Stage The core of this stage is the "large model (4)" on the right, which is a large language model with powerful natural language understanding and reasoning capabilities.

[0074] The input to this model is a comprehensive context integrating information from multiple sources, including: Income and expenditure forecasts from the left (6).

[0075] Semantic feature vectors (i.e., second feature vectors) from “Large Model (Vector Encoding) (1)” regarding descriptions of family hardship and policy texts.

[0076] A well-designed "Comprehensive Tips (9)" template. This template defines the evaluation expert role of the model, specific tasks (such as analyzing economic vulnerability and matching policies), and strict output format requirements (such as requiring JSON output, including scores and justifications).

[0077] The “large model (4)” integrates all the above information, performs logical reasoning, numerical analysis and policy matching, and finally generates a structured “assessment result (5)”. This result includes at least a quantitative family risk score and a textual analysis of key factors, which clearly explains the reasoning behind the score.

[0078] 4. Model continuous optimization closed loop The “Large Model Fine-tuning Module (10)” at the top of the figure represents the system’s self-learning capability. This module periodically collects historical data generated during the system’s operation (including input data, model prediction results, and the actual assistance results verified by the final human verification) to form training samples.

[0079] Using these samples, module (10) employs parameter-efficient fine-tuning techniques (such as LoRA) to update the trainable parameters in the system. The main targets include "large model (vector encoding) (1)", "small neural network (3)" and "large model (4)". In this way, it can continuously evolve with the accumulation of business data, thereby improving the accuracy and adaptability of the evaluation.

[0080] In summary, this preferred embodiment clearly illustrates the complete technical implementation path of the method described in this invention through illustrations: from the diversion and feature extraction of multimodal data, to the accurate prediction of core economic indicators by classical neural networks, to the comprehensive reasoning and interpretation generation of a large language model guided by structured prompts, and finally to the continuous optimization of performance through a fine-tuning module. This integrated, end-to-end architecture is the key to ensuring the evaluation accuracy, interpretability, and practicality of this invention.

[0081] The following describes the family situation assessment device based on neural networks provided by the present invention, such as... Figure 3 As shown, the neural network-based family situation assessment device described below and the neural network-based family situation assessment method described above can be referred to and correspond to each other.

[0082] A family situation assessment device based on neural networks includes: The data acquisition module 310 is used to acquire policy text data and family multimodal data of the households to be assessed, wherein the family multimodal data includes structured numerical data and text description data; The manual coding module 320 is used to perform rule-based mapping on the structured numerical data through the manual feature coding unit to obtain a first feature vector; Semantic encoding module 330 is used to perform semantic encoding on the text description data through a semantic encoding unit to obtain a second feature vector; The indicator prediction module 340 is used to input the first feature vector and the second feature vector into the economic indicator prediction model, and output the income prediction value and expenditure prediction value of the household to be evaluated. The data encapsulation module 350 is used to encapsulate the income forecast, expenditure forecast and policy text data based on a structured prompt engineering template to form prompt instructions; The comprehensive assessment module 360 ​​is used to input the prompts into the family situation assessment model to generate a comprehensive assessment result that includes a family risk score and the basis for the score.

[0083] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4 As shown, the electronic device may include: a processor 410, a communications interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communications interface 420, and the memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute a neural network-based family situation assessment method, which includes: Acquire policy text data and family multimodal data of the households to be assessed, wherein the family multimodal data includes structured numerical data and textual description data; The structured numerical data is mapped according to rules by an artificial feature encoding unit to obtain a first feature vector; The text description data is semantically encoded using a semantic encoding unit to obtain a second feature vector. The first feature vector and the second feature vector are input into the economic indicator prediction model, and the predicted income and expenditure values ​​of the household to be evaluated are output. Based on a structured prompt engineering template, the predicted income values, predicted expenditure values, and policy text data are encapsulated to form prompt instructions; The prompts are input into the family situation assessment model to generate a comprehensive assessment result that includes a family risk score and the basis for the score.

[0084] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a 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 described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0085] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is able to execute the neural network-based family situation assessment method provided by the above methods, the method comprising: Acquire policy text data and family multimodal data of the households to be assessed, wherein the family multimodal data includes structured numerical data and textual description data; The structured numerical data is mapped according to rules by an artificial feature encoding unit to obtain a first feature vector; The text description data is semantically encoded using a semantic encoding unit to obtain a second feature vector. The first feature vector and the second feature vector are input into the economic indicator prediction model, and the predicted income and expenditure values ​​of the household to be evaluated are output. Based on a structured prompt engineering template, the predicted income values, predicted expenditure values, and policy text data are encapsulated to form prompt instructions; The prompts are input into the family situation assessment model to generate a comprehensive assessment result that includes a family risk score and the basis for the score.

[0086] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the neural network-based family situation assessment method provided by the methods described above, the method comprising: Acquire policy text data and family multimodal data of the households to be assessed, wherein the family multimodal data includes structured numerical data and textual description data; The structured numerical data is mapped according to rules by an artificial feature encoding unit to obtain a first feature vector; The text description data is semantically encoded using a semantic encoding unit to obtain a second feature vector. The first feature vector and the second feature vector are input into the economic indicator prediction model, and the predicted income and expenditure values ​​of the household to be evaluated are output. Based on a structured prompt engineering template, the predicted income values, predicted expenditure values, and policy text data are encapsulated to form prompt instructions; The prompts are input into the family situation assessment model to generate a comprehensive assessment result that includes a family risk score and the basis for the score.

[0087] The device embodiments described above are merely illustrative. 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 modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0088] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for assessing family circumstances based on neural networks, characterized in that, include: Acquire policy text data and family multimodal data of the households to be assessed, wherein the family multimodal data includes structured numerical data and textual description data; The structured numerical data is mapped according to rules by an artificial feature encoding unit to obtain a first feature vector; The text description data is semantically encoded using a semantic encoding unit to obtain a second feature vector. The first feature vector and the second feature vector are input into the economic indicator prediction model, and the predicted income and expenditure values ​​of the household to be evaluated are output. Based on a structured prompt engineering template, the predicted income values, predicted expenditure values, and policy text data are encapsulated to form prompt instructions; The prompts are input into the family situation assessment model to generate a comprehensive assessment result that includes a family risk score and the basis for the score.

2. The family situation assessment based on neural networks according to claim 1, characterized in that, The structured numerical data includes per capita annual household income, total household debt, and number of working-age members in the household. The step of mapping the structured numerical data to the first feature vector using a manual feature encoding unit based on rules includes: The average annual income per capita of the households is divided into groups and converted into income level labels; the total debt of the households is logarithmically normalized. Calculate the debt-to-asset ratio based on the total household liabilities and total household assets; calculate the labor force ratio based on the number of working-age members and the total number of members in the household; The income level label, normalized total household debt, debt-to-asset ratio, and labor force ratio are concatenated to form the first feature vector.

3. The family situation assessment based on neural networks according to claim 1, characterized in that, The text description data includes textual descriptions of family members' health status, employment difficulties, or special circumstances. The step of semantically encoding the text description data using a semantic encoding unit to obtain a second feature vector specifically includes: The text description data is input into the semantic encoding unit, which is a pre-trained large language model based on the Transformer architecture; Obtain the [CLS] tag vector output by the last hidden layer of the semantic encoding unit, which corresponds to the overall semantics of the text description data, as the initial semantic representation; The initial semantic representation is subjected to layer normalization to output a dense vector of fixed dimensions, which serves as the second feature vector.

4. The family situation assessment based on neural networks according to claim 1, characterized in that, The economic indicator prediction model is a multilayer perceptron model, which includes an input layer, at least one hidden layer and an output layer. The output layer contains two neurons, which are used to output the predicted income value and the predicted expenditure value, respectively.

5. The family situation assessment based on neural networks according to claim 1, characterized in that, The comprehensive assessment results are structured data objects, including at least a family situation risk score and corresponding key factor analysis.

6. The family situation assessment based on neural networks according to claim 1, characterized in that, Following the step of acquiring policy text data and household multimodal data of the households to be assessed, the following is also included: The structured numerical data is filled with missing values ​​or corrected for outliers. Sensitive information desensitization and key information enhancement processing are performed on the text description data.

7. The family situation assessment based on neural networks according to claim 1, characterized in that, After the step of inputting the prompt instruction into the family situation assessment model to generate a comprehensive assessment result including a family risk score and the basis for the score, the method further includes: The training sample set is formed by periodically collecting the first feature vector, the second feature vector, the predicted income and expenditure values ​​of the households to be evaluated, and the comprehensive evaluation results. The parameters of the semantic coding unit, the economic indicator prediction model, and the family situation assessment model are fine-tuned using the training sample set.

8. A method for assessing family circumstances based on neural networks, characterized in that, include: The data acquisition module is used to acquire policy text data and family multimodal data of the households to be assessed. The family multimodal data includes structured numerical data and text description data. The manual coding module is used to perform rule-based mapping on the structured numerical data through the manual feature coding unit to obtain the first feature vector; A semantic encoding module is used to perform semantic encoding on the text description data through a semantic encoding unit to obtain a second feature vector; The indicator prediction module is used to input the first feature vector and the second feature vector into the economic indicator prediction model, and output the income prediction value and expenditure prediction value of the household to be evaluated. The data encapsulation module is used to encapsulate the income forecast, expenditure forecast, and policy text data based on a structured prompt engineering template to form prompt instructions; The comprehensive assessment module is used to input the prompts into the family situation assessment model to generate a comprehensive assessment result that includes a family risk score and the basis for the score.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the neural network-based family situation assessment method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the neural network-based family situation assessment method as described in any one of claims 1 to 7.

Images