Method and device for intelligently grading complaint-related risks based on multi-model fusion

Abstract

The invention relates to the technical field of data processing, in particular to a complaint-related risk intelligent grading method and device based on multi-model fusion, and the method comprises the steps: carrying out the preprocessing of target text data; performing sentence segmentation on the preprocessed target text data, and matching the preprocessed target text data with preset emotion keywords to obtain emotion conflict strength features; laws and regulations associated features are extracted from the preprocessed target text data; establishing a historical event library, extracting historical reference features from the preprocessed target text data according to the historical event library, and determining a complaint-related risk level in combination with the emotional conflict strength features, the law and regulation association features and the historical reference features; and matching the pre-processed target text data with a preset public issue keyword and a preset third-party subject keyword to determine a public issue matching degree and a subject interleaving degree, and correcting a complaint-related risk level in combination with the active correlation coefficient. The invention can effectively reduce the incidence of litigation.

Description

A method and device for intelligent classification of litigation risks based on multi-model fusion Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and apparatus for intelligent classification of litigation risks based on multi-model fusion. Background Technology

[0002] In the property management sector, disputes between property management companies and homeowners over issues such as fee payment, service quality, and maintenance of public facilities are on the rise. Currently, property management companies primarily rely on customer service personnel to manually handle homeowner complaints and record issues through work order systems. Faced with a large volume of text-based complaint information, customer service departments struggle to quickly and accurately identify high-risk cases that could escalate into legal proceedings.

[0003] Human judgment is prone to problems such as strong subjectivity, inconsistent standards, and low efficiency. This can easily lead to high-risk complaints not being prioritized, missing the best opportunity for mediation, and ultimately escalating conflicts, triggering litigation, and increasing the company's operating costs and reputational risks.

[0004] Therefore, there is an urgent need for a technological solution that can automatically and intelligently analyze text complaints and provide early warnings of litigation risks. Summary of the Invention

[0005] The purpose of this invention is to provide a method and apparatus for intelligent classification of litigation risks based on multi-model fusion, so as to solve at least one of the problems existing in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for intelligent classification of litigation risks based on multi-model fusion, comprising: establishing a historical event database, extracting historical citation features from the preprocessed target text data based on the historical event database, and determining the litigation risk level by combining emotional conflict intensity features, legal association features, and historical citation features; matching the preprocessed target text data with preset public issue keywords and preset third-party subject keywords to determine the public issue matching degree and subject intertwining degree, and correcting the litigation risk level by combining the active association coefficient.

[0007] Furthermore, the method also includes: preprocessing the target text data; segmenting the preprocessed target text data into sentences and matching them with preset sentiment keywords to obtain sentiment conflict intensity features; extracting legal association features from the preprocessed target text data; the preprocessing process of the target text data is as follows: cleaning and normalizing the target text data to obtain clean text, and segmenting the clean text into words to obtain a word list; inputting the clean text and the word list into the BERT-BiLSTM-CRF model to identify and extract key entities in the clean text and output an entity list.

[0008] Further, the clean text is segmented into sentences to generate a list of sentences, and the first sentence in the list is marked as the first sentence. The list of sentences is then matched with a first preset sentiment keyword. If a match is successful and also matches the first sentence, the first match count is incremented by 2. If a match is successful but does not match the first sentence, the first match count is incremented by 1. If a match fails, the first match count is incremented by 0. After traversing the list of sentences, the first match count m1 is obtained. The list of sentences is then matched with a second preset sentiment keyword, and the second match count m2 is obtained using the same method as the first match count. The first match count m1 and the second match count m2 are then merged to obtain the sentiment conflict intensity feature Eq.

[0009] Further, each word in the word segmentation list is traversed and matched with the preset terminology dictionary. The total number of successful matches is recorded as cm, and the base density J is calculated, set as J=cm / (N+5). The average base density of the corresponding user in the word segmentation list over the historical period is calculated and recorded as Jp. If Jp is less than or equal to the preset density j0, the historical factor is set to 0; otherwise, the historical factor is set to ln[2×(Jp-j0)+1] / ln3. The regulatory association feature Lw is determined based on the base density J and the historical factor, Lw=min{1,J×(1+η×historical factor)}; where η is the preset adjustment coefficient.

[0010] Furthermore, for each event template k in the historical event database, if the keyword clusters of the entity list and event template k intersect and have an intersection or inclusion relationship, then the current entity list is determined to be a successful match with event template k, and the matching coefficient kp between the current entity list and event template k is set to 1; otherwise, the current entity list is determined to be a failed match with event template k, and the matching coefficient kp between the current entity list and event template k is set to 0. For each successfully matched event template k, its time-decrease weight wk is calculated, wk=exp(-ts / τ). Where ts is the number of days between the start time of the current analysis period and the most recent activation time of event template k, and τ is the time decay constant; for each event template k that fails to match, its time decay weight wk is set to 0; the original historical feature value Hr is calculated based on the matching coefficient and the time decay weight; the maximum value Hrmax and the minimum value Hrmin of the original historical feature value of each target text data in the historical window are extracted, and the historical reference feature Ht of the current target text data is calculated, Ht=(Hr-Hrmin) / (Hrmax-Hrmin).

[0011] Furthermore, if r1×Eq+r2×Lw+r3×Ht is less than or equal to the first preset risk threshold, the litigation risk level of the current target text data is determined to be low risk; if r1×Eq+r2×Lw+r3×Ht is greater than the first preset risk threshold and less than or equal to the second preset risk threshold, the litigation risk level of the current target text data is determined to be medium risk; if r1×Eq+r2×Lw+r3×Ht is greater than the second preset risk threshold, the litigation risk level of the current target text data is determined to be high risk; where r1, r2, and r3 are weighting coefficients.

[0012] Furthermore, the preprocessed target text data is matched with preset public issue keywords and preset third-party subject keywords to determine the public issue matching degree, and the group contagion potential is determined by combining the active correlation coefficient.

[0013] Furthermore, the process for determining the potential for group transmission is as follows: the clean text is matched with preset public topic keywords. If the clean text contains any preset public topic keyword, the public topic matching degree is determined to be 1; otherwise, the public topic matching degree is determined to be 0. The potential for group transmission Q is determined based on the public topic matching degree and the active correlation coefficient, where Q = [0.5 × public topic matching degree + exp(3 × active correlation coefficient - 3)] / 2.

[0014] Furthermore, the preprocessed target text data is matched with preset third-party subject keywords, and the number of successfully matched preset third-party subject keywords is denoted as H. The ratio of H to the preset quantity threshold Hmax is taken as the subject interweaving degree Z. When α1×Q+α2×Z is less than or equal to the preset correction threshold, the litigation risk level is not updated; otherwise, the litigation risk level is forcibly updated to high risk. Here, α1 and α2 are weighting coefficients, and α1+α2=1.

[0015] According to another aspect of this application, a multi-model fusion-based intelligent classification device for litigation risk is provided, comprising: a preprocessing unit for preprocessing target text data; an emotional conflict analysis unit for segmenting the preprocessed target text data into sentences and matching them with preset emotional keywords to obtain emotional conflict intensity features; a legal association analysis unit for extracting legal association features from the preprocessed target text data; a classification unit for establishing a historical event database and extracting historical citation features from the preprocessed target text data based on the historical event database, and determining the litigation risk level by combining the emotional conflict intensity features, legal association features, and historical citation features; and a classification correction unit for matching the preprocessed target text data with preset public issue keywords and preset third-party subject keywords to determine the public issue matching degree and subject intertwining degree, and correcting the litigation risk level by combining the active association coefficient.

[0016] The beneficial effects of this invention are as follows: By integrating multi-model analysis technology, this invention can automatically and in real time extract multi-dimensional risk characteristics such as emotional conflict, legal awareness, historical recurrence, and group diffusion from massive customer service texts, and perform refined risk classification accordingly. This method transforms the past passive and lagging dispute handling mode into a proactive and accurate risk warning mode, helping property management companies optimize resource allocation, prioritize the handling of high-risk complaints, effectively reduce the probability of conflicts escalating to the litigation stage, and improve the intelligence level of customer service and risk management. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 is a flowchart illustrating the intelligent classification method for litigation risks based on multi-model fusion in this embodiment.

[0019] Figure 2 is a flowchart illustrating the method for determining the litigation risk level in this embodiment.

[0020] Figure 3 is a flowchart illustrating the method for correcting the litigation risk level in this embodiment.

[0021] Figure 4 is a schematic diagram of the intelligent classification device for litigation risks based on multi-model fusion in this embodiment. Detailed Implementation

[0022] To more clearly illustrate the present invention, the following description, in conjunction with preferred embodiments and accompanying drawings, further explains the invention. Similar components in the drawings are indicated by the same reference numerals. Those skilled in the art should understand that the specific description below is illustrative rather than restrictive and should not be construed as limiting the scope of protection of the present invention.

[0023] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0024] Specifically, this embodiment is applied to intelligent early warning and classification of litigation risks in the field of property service contract disputes. It specifically handles interactive text data between property companies and owners of specific communities due to issues such as property services, fee payment, and maintenance of public facilities. The aim is to automatically identify and quantify the litigation risks contained in multi-source heterogeneous texts through real-time analysis, and to classify them in a refined manner according to the level of risk, so as to provide the property company's customer complaint department with accurate priority suggestions for handling.

[0025] Please refer to Figure 1, which is a flowchart illustrating the intelligent classification method for litigation risks based on multi-model fusion in this embodiment. Before executing the method, target text data is acquired. The target text data consists of two types of core texts that can be obtained in a standardized manner within the property management company to ensure the legality and usability of the data. These include: structured customer service work order system text and semi-structured online communication records. The structured customer service work order system text is collected from the property management CRM or work order system via API, including text content such as "problem description," "processing," and "owner feedback." The semi-structured online communication records are collected from the property management company's WeChat, DingTalk workbench, or official owner service mini-program via interface, including "one-on-one chat records" between owners and property managers / customer service personnel. During collection, only the text messages sent by the owners are extracted and aggregated according to the conversation and time sequence to form "conversation text blocks."

[0026] The method includes: step S1, preprocessing the target text data.

[0027] Specifically, the target text data is cleaned and normalized to obtain clean text, and the clean text is segmented to obtain a segmentation list. The clean text and the segmentation list are then input into the BERT-BiLSTM-CRF model to identify and extract key entities from the clean text and output an entity list.

[0028] For example, the text cleaning and normalization process is as follows: Automatically generated fixed phrases such as "Hello, customer service XXX is at your service," URL links, emoticon encodings, and special garbled characters are removed from the target text data. Full-width characters are converted to half-width characters, traditional Chinese characters are converted to simplified Chinese characters, and numbers and English units are unified into Chinese expressions, such as "3 square meters" being converted to "3 square meters," to obtain clean text. The text segmentation and part-of-speech tagging process is as follows: The clean text is input into the Jieba segmenter, and a custom dictionary for the property management field is loaded. This dictionary contains professional terms such as "property management fee," "shared area," "parking space management fee," and "maintenance unit" to accurately segment the text. The key entity recognition process is as follows: The clean text and the segmented list are input into the BERT-BiLSTM-CRF model. The BERT-BiLSTM-CRF model is a named entity recognition (NER) model finely tuned on a large corpus of property contracts, maintenance reports, and complaint documents to identify and extract key entities from the text. Model fine-tuning uses pre-trained BERT-base-Chinese as initial weights, and the annotation system adopts the BIOES (Begin, Inside, Outside, End, Single) labeling scheme to perform sequence labeling for each character. The training data consists of approximately 120,000 text fragments extracted from historical customer service work orders, repair reports, and litigation documents after anonymization. These fragments were precisely labeled by domain experts according to predefined entity types and divided into training, validation, and test sets in an 8:1:1 ratio. The main hyperparameters are set as follows: learning rate 2e-5, batch size 16, and number of training epochs 10, employing a linear learning rate warm-up and decay strategy. The entity types defined in this embodiment include: building numbers, such as "Building 3" and "Building B"; unit numbers, such as "Unit 2" and "Room 1001"; public facilities, such as "elevator", "fire escape", and "underground parking garage"; ambiguous personnel designations, such as "Mr. Zhang" and "Owner of 701" (the system automatically maps these to the hash value of the owner's ID after identification, which is used to associate historical records. This mapping relationship is stored in an internal mapping table to ensure a balance between data availability and privacy protection); issue time, such as "last Friday" and "since handover"; specific issues, such as "water leakage", "abnormal noise", and "lighting not working"; and finally, an entity list is obtained.

[0029] Specifically, by performing deep cleaning and domain-specific entity recognition on customer service text, irrelevant noise was effectively eliminated, and unstructured user statements were transformed into structured information containing key questions, facilities, locations, and other elements. This provides a high-quality, standardized data foundation for subsequent accurate risk analysis and solves the problem of the original text being messy and difficult to use directly.

[0030] Please refer to Figure 1. The intelligent classification method for litigation risks based on multi-model fusion further includes: step S2, segmenting the preprocessed target text data into sentences and matching them with preset emotional keywords to obtain emotional conflict intensity features. The preset emotional keywords include a first preset emotional keyword and a second preset emotional keyword. The first preset emotional keyword is an anger keyword, and the second preset emotional keyword is a disappointment and helplessness keyword.

[0031] Specifically, the clean text is segmented into sentences to generate a sentence list, and the first sentence in the sentence list is marked as the first sentence. The sentence list is matched with the first preset sentiment keyword. If the match is successful and also matches the first sentence, the first match count is incremented by 2. If the match is successful but does not match the first sentence, the first match count is incremented by 1. If the match fails, the first match count is incremented by 0. After traversing the sentence list, the first match count m1 is obtained. The sentence list is matched with the second preset sentiment keyword, and the second match count m2 is obtained in the same way as the first match count. The first match count m1 and the second match count m2 are merged to obtain the sentiment conflict intensity feature Eq, Eq=min{1,(γ1×m1+γ2×m2) / (β×S)}; where γ1 and γ2 are weight coefficients, S is the total number of sentences in the target text data after sentence segmentation, and β is a preset correction coefficient.

[0032] Specifically, in this embodiment, the preset correction coefficient is 2.

[0033] For example, in this embodiment, the first preset keyword can be "lawsuit", "court", "unscrupulous property management", "expose to the media", "compensation", "lawyer", "sue you", "garbage property management", "inaction", "illegal", etc., and the second preset keyword can be set to include "utter disappointment", "no one to solve the problem", "heartbroken", "helpless", "perfunctory", "shirking responsibility", "despair", "helpless", "no way to complain", "no one cares", etc.

[0034] Specifically, the counting logic for the "first matching quantity m1" and the "second matching quantity m2" is as follows: traverse the list of sentences obtained after sentence segmentation. If a sentence contains any preset emotional keyword, the sentence is counted as a valid match. In order to emphasize the significance of the owner expressing strong emotions at the initial stage of communication, the first sentence is doubled in the count to obtain the emotional conflict intensity feature Eq.

[0035] Specifically, by identifying strong emotional signals contained in the text, especially paying attention to intense emotional expressions at the beginning of the conversation, it is possible to capture the key turning point where homeowners' dissatisfaction escalates into confrontation at an early stage. This helps the customer complaints department to prioritize intervention before the conflict intensifies, allocate resources to the most emotionally charged cases, and improve the initiative and timeliness of crisis resolution.

[0036] Please refer to Figure 1. The intelligent classification method for litigation risks based on multi-model fusion also includes: step S3, extracting regulatory association features from the preprocessed target text data.

[0037] Specifically, each word in the word segmentation list is traversed, matched against a preset terminology dictionary, and the total number of successful matches is recorded as cm. The base density J is calculated, and J = cm / (N+5). The average base density of the corresponding user in the word segmentation list over the historical period is calculated and recorded as Jp. If Jp is less than or equal to the preset density j0, the historical factor is set to 0; otherwise, the historical factor is set to ln[2×(Jp-j0)+1] / ln3. The regulatory association feature Lw is determined based on the base density J and the historical factor, Lw = min{1, J×(1+η×historical factor)}, where η is the preset adjustment coefficient and N is the total number of words in the target text data after sentence segmentation.

[0038] Specifically, in this embodiment, the preset adjustment coefficient is 0.2 and the preset density is 0.05.

[0039] For example, in this embodiment, the preset terminology dictionary consists of legal terms frequently used in property disputes, which can be set as litigation, lawsuit, court, tribunal, defendant, plaintiff, evidence, hearing, judgment, ruling, appeal, arbitration, rights protection, breach of contract, liquidated damages, compensation, claim, tort, obligation, right, legitimate rights and interests, illegal, violation, etc.

[0040] Specifically, by analyzing the density and changing trends of legal terms in the text, it is possible to quantify the explicitness and historical changes of homeowners' awareness of their rights. This feature does not depend on the specific content of their demands, but rather identifies homeowners who are more inclined to take formal legal action based on their expression, providing a stable and forward-looking observation dimension for risk assessment.

[0041] Please refer to Figure 1. The intelligent classification method for litigation risk based on multi-model fusion further includes: Step S4, establishing a historical event database, and extracting historical citation features from the preprocessed target text data based on the historical event database. Combining the emotional conflict intensity features, legal association features, and historical citation features, the litigation risk level is determined. The historical event database is a pre-generated structured database used to store typical common problems that repeatedly occur in the community. The database consists of several event templates. Each event template is a structured data object containing the following fields: Template identifier: a unique ID of the event; Core issue summary: a summary text description of the core content of this type of historical issue; Related facility set: a list containing the names of public facilities strongly related to the issue; Keyword cluster: a set of keywords and phrases that frequently appear in the text describing this type of issue; First recording time: the timestamp when the issue pattern is first identified and entered into the system.

[0042] Last Activation Time: The timestamp of the last time the target text data was successfully matched with this template.

[0043] For example, a lightweight and efficient Sentence-BERT model is used to generate sentence vector representations of cleaning texts from the past 1-2 years. Then, the DBSCAN density clustering algorithm is applied to the sentence vectors of all cleaning texts for cluster analysis. DBSCAN can automatically identify high-density regions, cluster complaints describing similar issues into one class, and filter out independent, occasional noise points. For each cluster, an event template is automatically generated, which includes: a core issue summary (key sentences extracted from all texts in the cluster using algorithms such as TextRank); and related facilities (statistics on the cluster). The frequency of occurrence of all FACILITY entities in the cluster is used to identify the 1-3 facilities with the highest frequency as associated facilities; Keyword cluster: All ISSUE entities in the cluster are merged, and high-frequency nouns and verbs extracted from the clustering text are added to form the core vocabulary set of the problem; Time record: The latest work order time in the cluster is used as the initial value of the first recording time and the most recent activation time, and is dynamically updated every month; For example, in this embodiment, after experimental analysis of the sentence vector distribution of a large number of property customer complaint texts, the core parameter settings of the DBSCAN algorithm are determined as follows: Neighborhood radius parameter eps The parameter eps is set to 0.4, and the minimum number of samples (min_samples) is set to 5. The determination of this parameter combination is based on the following: in the preprocessed sentence vector space, the contour coefficient is combined with manual verification for optimization. eps=0.4 can capture semantically similar complaint clusters in the current vector space, while min_samples=5 can effectively filter out occasional and scattered noise points, ensuring that the event templates generated in the end represent common problems with a certain degree of universality. Those skilled in the art can fine-tune this parameter based on the vector distribution characteristics of the actual corpus.

[0044] For example, the BERT-BiLSTM-CRF model used for key entity recognition in this embodiment is fine-tuned based on a specially constructed property domain-annotated corpus. This corpus contains approximately 120,000 text fragments extracted from historical customer service work orders, maintenance reports, and litigation documents after being anonymized, and accurately annotated by domain experts according to predefined entity types. The model fine-tuning uses a pre-trained BERT-base-Chinese model as the initial weights, with the main hyperparameters set as follows: learning rate 2e-5, batch size 16, and number of training epochs 10. A linear learning rate warm-up and decay strategy is used. During training, the annotated data is divided into training set, validation set, and test set in a ratio of 8:1:1.

[0045] Specifically, by establishing a dynamically updated historical event database, the system can intelligently identify whether the current complaint is a recurring "old problem" in the community. For complaints that match historical patterns, the system will give a higher risk rating based on their recurrence frequency and timeliness, thereby reminding the property management to pay close attention to those long-standing problems that have not been eradicated and may cause collective dissatisfaction.

[0046] Please refer to Figure 2. The method for determining the litigation risk level includes: step S41, establishing a historical event database, and extracting historical citation features from the preprocessed target text data based on the historical event database.

[0047] Specifically, for each event template k in the historical event database, the following matching criteria are applied: First, check if there is any intersection between the keyword clusters of entities of type "specific problem" in the entity list and event template k. If there is no intersection, the match is deemed to have failed. Second, check if there is any intersection between the entities of type "public facility" in the entity list and the associated facility set of event template k. If there is no intersection, the match is deemed to have failed. If both of the above criteria pass, the current entity list is deemed to have matched event template k successfully, and the matching coefficient kp is set to 1. Otherwise, the match is deemed to have failed, and the matching coefficient kp is set to 0. For each successfully matched event template k, its time decay weight wk is calculated, wk = exp(-ts / τ), where ts is the number of days between the start time of the current analysis period and the most recent activation time of event template k, and τ is the time decay constant. For each unmatched event template k, its time decay weight wk is set to 0. The original historical feature value Hr is calculated based on the matching coefficient and the time decay weight. K represents the number of event templates; extract the maximum value Hrmax and the minimum value Hrmin of the original historical feature values ​​of each target text data in the history window, and calculate the historical reference feature Ht of the current target text data, Ht=(Hr-Hrmin) / (Hrmax-Hrmin).

[0048] Specifically, in this embodiment, the time decay constant is 365 days.

[0049] Please refer to Figure 2. The method for determining the litigation risk level includes: step S42, combining the characteristics of emotional conflict intensity, legal association characteristics and historical citation characteristics to determine the litigation risk level.

[0050] Specifically, if r1×Eq+r2×Lw+r3×Ht is less than or equal to the first preset risk threshold, the litigation risk level of the current target text data is determined to be low risk; if r1×Eq+r2×Lw+r3×Ht is greater than the first preset risk threshold and less than or equal to the second preset risk threshold, the litigation risk level of the current target text data is determined to be medium risk; if r1×Eq+r2×Lw+r3×Ht is greater than the second preset risk threshold, the litigation risk level of the current target text data is determined to be high risk; where r1, r2, and r3 are weighting coefficients.

[0051] Specifically, in this embodiment, the first risk threshold is 0.4, the second risk threshold is 0.7, r1 is 0.3, r2 is 0.4, and r3 is 0.3.

[0052] Please refer to Figure 1. The intelligent classification method for litigation risks based on multi-model fusion further includes: Step S5, matching the preprocessed target text data with preset public issue keywords and preset third-party subject keywords to determine the public issue matching degree and subject intertwining degree, and correcting the litigation risk level by combining the active correlation coefficient. The active correlation coefficient is calculated by querying the database using the current target text data building number as the filtering condition, counting the total number of target text data in the building whose current status is "pending" or "overdue", dividing this value by the total number of households in the building, and setting the active correlation coefficient to 1 when it is greater than 1. For example, the preset third-party subject keywords refer to a preset set of words used to identify whether the text mentions external entities other than the property service company that may bear joint liability for the dispute or constitute obstacles to its resolution. These can be set as "developer", "construction party", "power supply bureau", "water company", "heating company", "telecommunications company", and "elevator maintenance". "Greening Company", etc.; The preset public issue keywords refer to a set of preset words used to identify whether the text involves the common rights and interests of all owners in the community and public management decisions. They can be set as "public revenue", "green space renovation", "property management selection", "owners' committee", "maintenance fund", "common parts", etc.

[0053] Specifically, by detecting whether a text involves public interest issues or third-party liability entities, the system can identify cases that go beyond individual disputes, are likely to attract widespread attention, or have complex liability definitions. This ensures that cases with potential for diffusion or complexity are not underestimated by calculations based on conventional characteristics, thereby achieving mandatory attention to systemic risks.

[0054] Please refer to Figure 3. The method for correcting the litigation risk level includes: step S51, matching the preprocessed target text data with preset public issue keywords and preset third-party subject keywords to determine the public issue matching degree, and combining the active correlation coefficient to determine the group contagion potential.

[0055] Specifically, the clean text is matched with preset public issue keywords. If the clean text contains any preset public issue keyword, the public issue matching degree is determined to be 1; otherwise, the public issue matching degree is determined to be 0. The group contagion potential Q is determined based on the public issue matching degree and the active correlation coefficient, Q=[0.5×public issue matching degree+exp(3×active correlation coefficient-3)] / 2.

[0056] Please refer to Figure 3. The method for correcting the litigation risk level also includes: step S52, matching the pre-processed target text data with preset third-party subject keywords to determine the subject interweaving degree, and correcting the litigation risk level in combination with the potential for group contagion.

[0057] Specifically, the preprocessed target text data is matched with preset third-party subject keywords, and the number of successfully matched preset third-party subject keywords is denoted as H. The ratio of H to the preset quantity threshold Hmax is used as the subject interweaving degree Z. When α1×Q+α2×Z is less than or equal to the preset correction threshold, the litigation risk level is not updated; otherwise, the litigation risk level is forcibly updated to high risk. Here, α1 and α2 are weighting coefficients, and α1+α2=1.

[0058] Specifically, in this embodiment, α1 is 0.7, α2 is 0.3, the preset correction threshold is 0.6, and the preset quantity threshold is 3.

[0059] For example, in this embodiment, the analysis period is 1 hour, the historical period is 30 days before the current analysis period, and the historical window is 7 days before the current analysis period.

[0060] Specifically, the final values ​​of the various weight coefficients, thresholds, and formula constants involved in this method are determined based on a combination of historical data statistical analysis, grid search or cross-validation optimization, and evaluation by domain experts. The specific process is as follows: First, historical customer complaint texts within a time span and their tag data indicating whether they ultimately led to litigation are collected as a development set. Then, for the adjustable parameters in the formula, such as the weight coefficients w1, w2, w3 in step S42, the adjustment coefficient η in step S3, and the sentiment weights γ1, γ2 in step S2, within a reasonable range, for example, w1, w2, w3... Between 0 and 1, and with a sum of 1, a grid search is performed. The optimization objective is to find the optimal parameter combination by comparing the macro F1 score or AUC value of the risk level prediction results with the actual litigation results. For threshold parameters such as the first and second preset risk thresholds of 0.4 and 0.7, and the preset correction threshold of 0.6, the distribution of the comprehensive risk value on the development set and the correspondence with the litigation events are analyzed, and the determination is made in combination with the business party's requirements for the balance between the accuracy and recall rate of risk warning. All specific values ​​given in this embodiment are example values ​​after one optimization experiment. In actual deployment, the above optimization process can be repeated to adapt to the data characteristics of different property companies.

[0061] Please refer to Figure 4, which is a schematic diagram of the intelligent classification device for litigation risks based on multi-model fusion in this embodiment. It includes: a preprocessing unit for preprocessing target text data; an emotional conflict analysis unit for segmenting the preprocessed target text data into sentences and matching them with preset emotional keywords to obtain emotional conflict intensity features; a legal association analysis unit for extracting legal association features from the preprocessed target text data; a classification unit for establishing a historical event database and extracting historical citation features from the preprocessed target text data based on the historical event database, and determining the litigation risk level by combining the emotional conflict intensity features, legal association features, and historical citation features; and a classification correction unit for matching the preprocessed target text data with preset public issue keywords and preset third-party subject keywords to determine the public issue matching degree and subject interweaving degree, and correcting the litigation risk level by combining the active association coefficient.

[0062] The intelligent classification device for litigation risks based on multi-model fusion provided in this application can execute the intelligent classification method for litigation risks based on multi-model fusion provided in any embodiment of this application, and has the corresponding functional modules and beneficial effects of the execution method.

[0063] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. For those skilled in the art, other variations or modifications can be made based on the above description. It is impossible to exhaustively list all the implementation methods here. All obvious variations or modifications derived from the technical solutions of the present invention are still within the protection scope of the present invention.

Claims

1. A method for intelligent classification of litigation risks based on multi-model fusion, characterized in that, include: Establish a historical event database, and extract historical citation features from the preprocessed target text data based on the historical event database. Combine the emotional conflict intensity features, legal association features and historical citation features to determine the litigation risk level. The preprocessed target text data is matched with preset public issue keywords and preset third-party subject keywords to determine the public issue matching degree and subject intertwining degree, and the litigation risk level is corrected by combining the active correlation coefficient.

2. The intelligent classification method for litigation risks based on multi-model fusion according to claim 1, characterized in that, The method further includes: preprocessing the target text data; segmenting the preprocessed target text data into sentences and matching them with preset sentiment keywords to obtain sentiment conflict intensity features; extracting legal association features from the preprocessed target text data; the preprocessing process of the target text data is as follows: cleaning and normalizing the target text data to obtain clean text, and segmenting the clean text into words to obtain a word list; inputting the clean text and the word list into the BERT-BiLSTM-CRF model to identify and extract key entities in the clean text and output an entity list.

3. The intelligent classification method for litigation risks based on multi-model fusion according to claim 2, characterized in that, The clean text is segmented into sentences to generate a list of sentences, and the first sentence in the list is marked as the first sentence. The list of sentences is matched with a first preset sentiment keyword. If a match is successful and also matches the first sentence, the first match count is incremented by 2. If a match is successful but does not match the first sentence, the first match count is incremented by 1. If a match fails, the first match count is incremented by 0. After traversing the list of sentences, the first match count m1 is obtained. The list of sentences is then matched with a second preset sentiment keyword, and the second match count m2 is obtained using the same method as the first match count. The first match count m1 and the second match count m2 are then merged to obtain the sentiment conflict intensity feature Eq.

4. The intelligent classification method for litigation risks based on multi-model fusion according to claim 3, characterized in that, Iterate through each word in the word segmentation list, match it with the preset terminology dictionary, and record the total number of successful matches as cm. Calculate the base density J, setting J = cm / (N+5). Calculate the average base density of the corresponding user in the word segmentation list over the historical period, denoted as Jp. If Jp is less than or equal to the preset density j0, set the historical factor to 0; otherwise, set the historical factor to ln[2×(Jp-j0)+1] / ln3. Determine the regulatory association feature Lw based on the base density J and the historical factor, Lw = min{1, J×(1+η×historical factor)}; where η is the preset adjustment coefficient.

5. The intelligent classification method for litigation risks based on multi-model fusion according to claim 4, characterized in that, Iterate through each event template k in the historical event database. If the keyword clusters of the entity list and event template k intersect, and the entity list and event template k have an intersection or containment relationship, then the current entity list is considered to be a successful match with event template k, and the matching coefficient kp between the current entity list and event template k is set to 1. Otherwise, the current entity list is considered to be a failed match with event template k, and the matching coefficient kp between the current entity list and event template k is set to 0. For each successfully matched event template k, calculate its time-decrease weight wk, wk=exp(-ts / τ), where wk=exp(-ts / τ). In this context, ts represents the number of days between the start time of the current analysis period and the most recent activation time of event template k, and τ is the time decay constant. For each event template k that fails to match, its time decay weight wk is set to 0. The original historical feature value Hr is calculated based on the matching coefficient and the time decay weight. The maximum value Hrmax and the minimum value Hrmin of the original historical feature value of each target text data in the historical window are extracted, and the historical reference feature Ht of the current target text data is calculated, Ht=(Hr-Hrmin) / (Hrmax-Hrmin).

6. The intelligent classification method for litigation risks based on multi-model fusion according to claim 5, characterized in that, If r1×Eq+r2×Lw+r3×Ht is less than or equal to the first preset risk threshold, then the litigation risk level of the current target text data is determined to be low risk. If r1×Eq+r2×Lw+r3×Ht is greater than the first preset risk threshold and less than or equal to the second preset risk threshold, then the litigation risk level of the current target text data is determined to be medium risk. If r1×Eq+r2×Lw+r3×Ht is greater than the second preset risk threshold, then the current target text data is determined to be of high risk in terms of litigation risk level; where r1, r2 and r3 are weighting coefficients.

7. The intelligent classification method for litigation risks based on multi-model fusion according to claim 6, characterized in that, The preprocessed target text data is matched with preset public topic keywords and preset third-party subject keywords to determine the public topic matching degree, and the group contagion potential is determined by combining the active correlation coefficient.

8. The intelligent classification method for litigation risks based on multi-model fusion according to claim 7, characterized in that, The process for determining the potential for group contagion is as follows: the clean text is matched with preset public topic keywords. If the clean text contains any preset public topic keyword, the public topic matching degree is determined to be 1; otherwise, the public topic matching degree is determined to be 0. The potential for group contagion Q is determined based on the public topic matching degree and the active correlation coefficient, where Q = [0.5 × public topic matching degree + exp(3 × active correlation coefficient - 3)] / 2.

9. The intelligent classification method for litigation risks based on multi-model fusion according to claim 8, characterized in that, The preprocessed target text data is matched with preset third-party subject keywords, and the number of successfully matched preset third-party subject keywords is recorded as H. The ratio of H to the preset number threshold Hmax is used as the subject interweaving degree Z. When α1×Q+α2×Z is less than or equal to the preset correction threshold, the litigation risk level is not updated; otherwise, the litigation risk level is forcibly updated to high risk. Here, α1 and α2 are weight coefficients, and α1+α2=1.

10. A litigation risk intelligent classification device based on multi-model fusion, applied to the litigation risk intelligent classification method based on multi-model fusion as described in any one of claims 1-9, characterized in that, include: The preprocessing unit is used to preprocess the target text data; The sentiment conflict analysis unit is used to segment the preprocessed target text data into sentences and match them with preset sentiment keywords to obtain the sentiment conflict intensity features. The regulatory association analysis unit is used to extract regulatory association features from the preprocessed target text data. The classification unit is used to establish a historical event database and extract historical citation features from the preprocessed target text data based on the historical event database. The risk level of litigation is determined by combining the emotional conflict intensity features, legal association features and historical citation features. The rating correction unit is used to match the preprocessed target text data with preset public issue keywords and preset third-party subject keywords to determine the public issue matching degree and subject intertwining degree, and to correct the litigation risk level in combination with the active correlation coefficient.

Images