Case clue identification and analysis system

Abstract

The invention relates to the technical field of computer data processing and natural language processing, in particular to a case clue identification and analysis system, which comprises a data access and positioning module for calling a preset multi-source map service interface to concurrently analyze a network address, and determining a target position through voting arbitration; the cascade semantic analysis module is used for identifying attributes by utilizing coarse granularity of a first-level classification model, and routing to a vertical expert agent to generate a fine-granularity label; the element feature extraction module is used for eliminating interference digital features and then generating standard amount features, and analyzing intention intensity of the user; the value quantification and pricing module is used for mapping the heterogeneous features into multi-dimensional scores based on a tree-shaped weight system, and generating clue commercial value scores and pricing results; according to the invention, the problem of value evaluation caused by information non-standardization in a compliance consultation scene is solved.

Description

Technical Field

[0001] This invention relates to the fields of computer data processing and natural language processing technology, specifically a case clue identification and analysis system. Background Technology

[0002] In the application scenario of Internet compliance consulting services, the platform relies on an intelligent distribution system to connect with massive amounts of case clue data. The system usually needs to combine unstructured text descriptions and user network address data to analyze the location and commercial value of the case in real time.

[0003] Existing solutions for identifying and pricing case leads generally employ a single-channel processing architecture. This relies on a single map service provider's interface to parse network addresses and uses general natural language processing models for keyword matching or shallow classification of consultation texts, directly using the extracted results as the basis for lead value assessment. While this approach is feasible for standardized, simple consultations, its over-reliance on a single data source and lack of in-depth logical constraints within specific vertical domains leads to location drift due to data update lag when encountering cross-regional or non-standardized descriptions, resulting in incorrect jurisdiction determinations. Furthermore, general models struggle to accurately distinguish fine-grained case types within compliant domains, and simple numerical extraction algorithms easily misidentify everyday noise such as phone numbers and license plate numbers as features of high-value targets, leading to severely inflated value assessments. Simultaneously, the lack of a multi-dimensional, dynamic, quantitative pricing system makes it difficult for platforms to accurately price leads based on region, intent, and target value, resulting in ineffective distribution and a significant waste of lawyer resources.

[0004] Therefore, how to establish an analysis mechanism with multi-source calibration and anti-interference capabilities, effectively eliminating digital noise while improving the accuracy of case location and the reliability of commercial value assessment, has become an urgent technical problem to be solved. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a case clue identification and analysis system. Specifically, the technical solution of this invention includes:

[0006] The data access and positioning module is used to acquire unstructured text data and network address data, call the preset multi-source map service interface to perform multi-channel parsing of network address data, and determine the target geographical location based on the preset voting arbitration logic;

[0007] The cascaded semantic analysis module is used to perform coarse-grained recognition of unstructured text data using a primary classification model, determine primary classification attributes, and dynamically route to the corresponding vertical expert agent based on the primary classification attributes. The vertical expert agent then generates secondary and tertiary fine-grained classification labels.

[0008] The feature extraction module is used to perform numerical entity recognition on unstructured text data. After removing preset interference numerical features, it generates target value features and simultaneously analyzes the intensity of user intent. The value quantification and pricing module is used to map the target geographical location, fine-grained classification labels, target value features and intensity of user intent into multi-dimensional scores based on a preset tree-structured weight configuration system. It also generates clue commercial value scores and pricing results according to preset multi-dimensional numerical calculation logic.

[0009] Preferably, the data access and positioning module determines the target geographical location based on a preset voting arbitration logic, configured as follows: obtaining multiple candidate geographical location information returned by the multi-source map service interface; and detecting the consistency of the multiple candidate geographical location information.

[0010] If multiple candidate geographic location information are completely identical, then the identical information is determined as the target geographic location; if there are partially identical candidate geographic location information and the number of identical information exceeds a preset majority threshold, then the majority information is determined as the target geographic location.

[0011] If none of the above conditions are met, the candidate geographic location information corresponding to the preset authoritative source will be selected as the target geographic location, or a front-end interaction command will be triggered to obtain user confirmation information.

[0012] Preferably, the cascaded semantic analysis module dynamically routes to the corresponding vertical expert agent based on the primary classification attribute, including: constructing an agent group containing multiple independent operating environments, wherein each agent corresponds to a preset compliance business domain;

[0013] In response to the generation of the primary classification attribute, activate the target vertical category expert agent in the agent group that matches the primary classification attribute;

[0014] The target vertical category expert agent is used to perform deep semantic analysis on unstructured text data, and output the secondary and tertiary classification results belonging to the primary classification attribute.

[0015] Preferably, the feature extraction module generates target amount features after removing preset interfering numerical features, including: scanning unstructured text data using a preset Chinese amount regularization algorithm to extract the original numerical set;

[0016] Identify strings in unstructured text data that conform to preset mobile phone number rules and preset license plate number rules, and define them as interference number features; remove the parts that overlap with the interference number features from the original numerical set, and perform unit unification processing on the remaining numerical values ​​to generate target value features.

[0017] Preferably, the value quantification and pricing module is based on a preset tree-structured weight configuration system, which maps the target geographic location and fine-grained classification labels into multi-dimensional scores, including: calling the preset regional weight tree and classification weight tree, wherein the regional weight tree contains three-level nodes of province, city and county, and the classification weight tree contains first-level, second-level and third-level classification nodes;

[0018] Retrieve the corresponding regional weight score from the regional weight tree based on the target geographical location; retrieve the corresponding category weight score from the category weight tree based on the fine-grained category label.

[0019] Preferably, the value quantification and pricing module generates a lead commercial value score and pricing result based on a preset multi-dimensional numerical calculation logic, including: obtaining the amount score corresponding to the target value feature and the intent score corresponding to the user intent intensity based on a preset numerical mapping relationship;

[0020] The category weight score, region weight score, amount score, and intent score are summed to generate a comprehensive value score;

[0021] Obtain the preset global price adjustment coefficient, calculate the product of the comprehensive value integral and the global price adjustment coefficient, and define the calculated product as the pricing result.

[0022] Preferably, the feature extraction module analyzes the intensity of user intent, including: using a preset intent recognition model to analyze action verbs and time adverbs in unstructured text data;

[0023] If semantic features representing delegation or instant communication are identified, it is determined to be of strong intent strength; otherwise, it is determined to be of weak intent strength.

[0024] Preferably, the system also includes a security filtering module, which is used to clean unstructured text data before the data access and positioning module is executed;

[0025] The system matches unstructured text data based on a pre-defined blacklist and sensitive word database. If a match is found, the data is intercepted and the subsequent process is terminated. If no match is found, the cleaned data is transmitted to the cascaded semantic analysis module.

[0026] Preferably, the system also includes an auction ranking and display module for receiving pricing results and target geographical locations;

[0027] The pricing results are written back to the price attribute field in the database; in response to front-end query requests, a dual index is built based on the price attribute field and the target geographical location, and a list of leads is generated in descending order of pricing results.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] 1. This invention effectively solves the location drift problem caused by data lag from a single data source by introducing a multi-source map service concurrent parsing and voting arbitration mechanism. Unlike traditional solutions that rely on a single service provider, this system simultaneously calls multiple mainstream map interfaces to parse network address data in parallel and uses a decision algorithm based on consistency verification and majority principle to determine the target location. In extreme conflict situations, it can also trigger user interaction confirmation, ensuring the absolute accuracy of case jurisdiction determination in cross-regional scenarios and avoiding the generation of invalid clues.

[0030] 2. This invention adopts a cascaded semantic analysis architecture and vertical expert agent group technology, which breaks through the bottleneck of insufficient recognition accuracy of general models in compliance sub-fields; the system performs coarse-grained routing through a lightweight model, and then activates expert agents loaded with dedicated compliance criteria and case law libraries as needed. It uses a retrieval enhancement generation mechanism and a multi-task learning structure to perform deep semantic analysis of the case, thereby accurately identifying the secondary and tertiary fine-grained classification labels that are difficult for general models to capture, which greatly improves the professionalism and fine granularity of case identification;

[0031] 3. This invention constructs an element feature extraction mechanism based on anti-interference logic, which significantly reduces the misjudgment rate of target amount caused by digital noise in daily life scenarios; for interference features such as mobile phone numbers and license plate numbers mixed in unstructured text, the system pre-executes negative feature filtering and elimination operations according to specific rules, and only performs normalization and unit conversion on the remaining valid values, which effectively avoids misidentifying communication methods as features of high-value targets, and ensures the authenticity and reliability of the economic value assessment of case clues;

[0032] 4. This invention establishes a multi-dimensional value quantification system and dynamic pricing strategy, achieving precise anchoring of the commercial value of case leads. By integrating a tree-structured weight configuration system, user intent intensity analysis, and global price adjustment coefficient, the system can comprehensively consider heterogeneous factors such as regional economic level, case category weight, amount in dispute, and consultation period, transforming non-standardized compliance consultations into pricing results with clear gradations. It also supports value-based bidding and sorting display, thereby guiding the priority distribution of high-value leads and effectively avoiding ineffective investment of lawyer resources in low-conversion consultations. Attached Figure Description

[0033] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0034] Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0036] Example 1:

[0037] Please see Figure 1 A case clue identification and analysis system includes: a data access and positioning module, used to acquire unstructured text data and network address data, call a preset multi-source map service interface to perform multi-channel parsing of network address data, and determine the target geographical location based on a preset voting arbitration logic;

[0038] The cascaded semantic analysis module is used to perform coarse-grained recognition of unstructured text data using a primary classification model, determine primary classification attributes, and dynamically route to the corresponding vertical expert agent based on the primary classification attributes. The vertical expert agent then generates secondary and tertiary fine-grained classification labels.

[0039] The feature extraction module is used to perform numerical entity recognition on unstructured text data. After removing preset interference numerical features, it generates target value features and simultaneously analyzes the intensity of user intent. The value quantification and pricing module is used to map the target geographical location, fine-grained classification labels, target value features and intensity of user intent into multi-dimensional scores based on a preset tree-structured weight configuration system. It also generates clue commercial value scores and pricing results according to preset multi-dimensional numerical calculation logic.

[0040] This embodiment provides a case clue identification and analysis system. The system is built on a hardware architecture consisting of a high-performance computing server, a distributed database, and a multi-source API interface gateway. It aims to solve the value assessment problem caused by the non-standardization of information in compliance consulting scenarios.

[0041] Specifically, the mapping process uses a normalization operator. Transform heterogeneous features into a unified dimension; define the target geographic location score as... The fine-grained classification label score is The target amount is User intent strength score The multidimensional numerical operation logic adopts a weighted summation model, and its calculation formula is as follows:

[0042]

[0043] The data access and positioning module establishes a data channel with the Internet front end to receive unstructured text data input by users in natural language mode in real time, namely compliance consultation descriptions, as well as network address data of user terminals. This module does not rely solely on a single map service provider, but concurrently calls preset multi-source map service interfaces, such as Tencent, Baidu, and Gaode Map APIs, through the API gateway to perform parallel parsing of network address data.

[0044] The system incorporates voting arbitration logic, a decision-making algorithm based on multi-source data consistency verification. This algorithm determines the final target geographic location based on the majority or authoritative source principle when results from different map service providers are inconsistent. Building upon this, the cascaded semantic analysis module adopts a hierarchical routing architecture. It first uses a lightweight, first-level classification model built on the DistilBERT neural network architecture to perform coarse-grained text scanning. To conform to the tensor operation rules of deep learning models, the model's input layer receives a token sequence tensor generated by the token segmenter. and attention mask This outputs the hidden state tensor. And extract ,in To adapt to the dimensionality requirements of text classification tasks, the system performs pooling operations, specifically extracting the [CLS] vector from the start position of the sequence, defined as... This transforms the three-dimensional tensor from Stable collapse into two-dimensional eigenvectors This eliminates the risk that average pooling might dilute the keyword weights in long texts;

[0045] This two-dimensional vector must pass through an explicitly defined linear projection layer. ,in Explicitly defined as the total number of preset primary categories, for example The high-dimensional features are mapped to category Logits, and the output is discretized through the Softmax layer and Argmax operator to output first-level classification attributes such as marriage and family affairs and labor disputes. Based on these attributes, the data packets are distributed to the corresponding vertical expert agents through a dynamic routing gateway.

[0046] The feature extraction module performs numerical entity recognition. To prevent interference from numerical strings such as mobile phone numbers and ID card numbers in the amount extraction, the system extracts the remaining values ​​and normalizes them to generate target amount features after removing preset interference numerical features. At the same time, it combines NLP sentiment analysis to analyze the intensity of user intent. The value quantification and pricing module maps the above heterogeneous features into multi-dimensional scores based on a tree-structured weight configuration system, and generates the final clue commercial value score and pricing result based on multi-dimensional numerical calculation logic.

[0047] Example 2:

[0048] The data access and positioning module determines the target geographic location based on a preset voting arbitration logic, configured as follows: obtain multiple candidate geographic location information returned by the multi-source map service interface; detect the consistency of multiple candidate geographic location information; if multiple candidate geographic location information are completely consistent, then the consistent information is determined as the target geographic location; if there are partially consistent candidate geographic location information and the number of consistent information exceeds a preset majority threshold, then the majority information is determined as the target geographic location; if none of the above conditions are met, then the candidate geographic location information corresponding to the preset authoritative source is selected as the target geographic location, or a front-end interaction command is triggered to obtain user confirmation information.

[0049] This embodiment specifies the configuration of the voting arbitration logic in the data access and positioning module, aiming to solve the positioning drift problem often caused by a single data source; the system simultaneously sends... Send parsing requests to 12 different map service interfaces, among which Defined as a positive integer greater than or equal to 3, i.e. And preferably an odd number to avoid a tie, for example, the range of values ​​is [value range missing]. This allows us to obtain multiple candidate geographic location information. The system compares the administrative division codes of these candidate information and performs a consistency check.

[0050] In response to to If the data is completely consistent, the system determines the data to be reliable and directly identifies the consistent information as the target geographical location; if there is partial consistency, the system counts the frequency of each candidate location. ; responding to a certain location frequency Greater than the preset majority threshold For example, set to The system executes the majority decision and will The target geographic location is determined; if there are two or more sets of location information, the frequency is... Equal and none exceeding The system then introduces a preset authority weight coefficient. The system provides each map service interface Preset an authority weight For a given candidate geographical location Identify the set of all interfaces that return this location. Calculate the weighted score for that position. The calculation formula is as follows:

[0051]

[0052] The system iterates through all candidate positions and selects the calculated ones. The location corresponding to the maximum value is taken as the final target geographical location;

[0053] If none of the above conditions are met, i.e., the results returned by all interfaces have extremely high dispersion, the system will execute a fallback strategy: either select a preset authoritative source, i.e., the candidate geographic location information corresponding to the interface with the highest historical accuracy weighting, or trigger a front-end interaction command to pop up a location confirmation box in the user interface and force the acquisition of geographic information manually confirmed by the user.

[0054] In sensitive scenarios involving cross-regional compliance jurisdiction, this embodiment effectively eliminates positioning errors caused by data lag from a single service provider through multi-source voting and majority decision-making mechanisms. In extreme conflict situations, user interaction is introduced to ensure the absolute accuracy of the case source location attribute, avoiding situations where lawyers purchase leads only to find that they are from other regions, resulting in ineffective investment.

[0055] Example 3:

[0056] The cascaded semantic analysis module dynamically routes data to the corresponding vertical expert agents based on the primary classification attribute. This includes: constructing an agent group containing multiple independent operating environments, where each agent corresponds to a preset compliance business domain; activating the target vertical expert agent that matches the primary classification attribute in the agent group in response to the generation of the primary classification attribute; and using the target vertical expert agent to perform deep semantic analysis on the unstructured text data, outputting the secondary and tertiary classification results belonging to the primary classification attribute.

[0057] This embodiment details the deep parsing process based on agent groups in the cascaded semantic analysis module, aiming to address the poor performance of general models in the long-tail compliance domain; the system pre-builds... Each operating environment, among which Defined as a positive integer greater than 1 This represents the total number of subdivided compliant business areas that the system is pre-defined to cover, and is usually configured according to the actual business scale. ;

[0058] These are groups of agents operating in environments such as Docker containers or independent memory contexts; each agent... Each corresponds to a pre-defined compliant business area, such as Corresponding to marriage and family matters, It corresponds to traffic accidents and internally loads a compliance rule library and case library specific to this field; in response to the first-level classification attribute generated by the first-level classification model, the system activates only the target vertical category expert intelligent agents that match the attribute in the intelligent agent group through a message queue, instead of loading all of them; it uses the target vertical category expert intelligent agents to perform deep semantic parsing of the text;

[0059] Specifically, the agent is built on a pre-trained language model based on the architecture and integrates a retrieval enhancement generation mechanism. The system maps the exclusive compliance criterion library and case library into dense vector indexes and stores them in the vector database through a pre-trained sentence vector model. During the inference phase, the agent calculates the embedding vector of the input text, retrieves the Top-K most similar legal provision contexts in the vector library, concatenates the text and legal provisions, and inputs them into the model for attention interaction calculation, thereby identifying details that are difficult for general models to capture.

[0060] To clarify the computation graph topology of the multi-level output, the agent adopts a multi-task learning head structure, that is, two independent fully connected layers are connected in parallel after the Transformer backbone to map the secondary and tertiary label spaces respectively; through this parallel computing structure, the agent synchronously outputs the secondary and tertiary classification results belonging to the primary classification, if there is evidence and it is within the statute of limitations;

[0061] This embodiment significantly improves the accuracy of fine-grained classification when handling complex case descriptions by constructing a dedicated intelligent agent group architecture. At the same time, the mechanism of activating intelligent agents on demand avoids the waste of computing power in running the full model, enabling the system to maintain millisecond-level response speed and low operating costs when facing high-concurrency consultation requests.

[0062] Example 4:

[0063] The feature extraction module generates target amount features after removing preset interference numerical features, including: scanning unstructured text data using a preset Chinese amount regularization algorithm to extract the original numerical set; identifying strings in the unstructured text data that conform to preset mobile phone number rules and preset license plate number rules, defining them as interference numerical features; removing the parts that overlap with interference numerical features from the original numerical set, and unifying the units of the remaining values ​​to generate target amount features.

[0064] This embodiment describes the specific steps for generating the target amount feature in the feature extraction module. It employs anti-interference extraction logic to address the numerical noise mixed in unstructured text. The system uses a preset Chinese amount regularization algorithm to perform a full scan of the unstructured text data, extracting all possible values ​​to form the original numerical set. ;

[0065] The system executes the interference source definition step, identifying text elements that conform to preset mobile phone number rules (i.e., 11 consecutive digits or a specific pattern) and strings that conform to preset license plate number rules, defining them as interference number features. The system records the digital characteristics of interference. Starting offset in unstructured text and end offset ;

[0066] Traverse the original set of values Each element in ,like Original span in the text With any interference span There is an intersection, that is, it satisfies Then From the set Remove from the middle; unify the units of the remaining values, for example, convert 500,000 to 500,000, and take the maximum value or cumulative value as the target amount feature;

[0067] This embodiment addresses the common practice of users mixing phone numbers, license plate numbers, and monetary amounts in their consultation descriptions. By introducing a negative feature filtering mechanism based on specific life scenarios, it significantly reduces the false positive rate of target amount extraction, ensuring the reliability of subsequent assessment of the economic value of leads and avoiding the serious error of misjudging a mobile phone number as a target amount in the tens of millions.

[0068] Example 5:

[0069] The value quantification and pricing module, based on a pre-defined tree-structured weight configuration system, maps target geographic location and fine-grained classification labels into multi-dimensional scores. This includes: calling a pre-defined regional weight tree and a category weight tree, where the regional weight tree contains three levels of nodes: province, city, and county, and the category weight tree contains first-level, second-level, and third-level category nodes; retrieving the corresponding regional weight score from the regional weight tree based on the target geographic location; and retrieving the corresponding category weight score from the category weight tree based on the fine-grained classification labels.

[0070] This embodiment details the mapping mechanism in the value quantification and pricing module, introducing a tree-like weight configuration system to adapt to the differences in commercial value across different regions and case types; the system constructs a hierarchical regional weight tree. It includes provincial, municipal, and county-level nodes, with each node bound to a regional weight score. Simultaneously construct a classification weight tree. It contains first-level, second-level, and third-level classification nodes, with each leaf node bound to a classification weight score. The system, based on the determined target geographical location, in Traverse to the finest-grained node and retrieve the corresponding... Based on the determined fine-grained classification labels, in Searching for the corresponding ;

[0071] This embodiment utilizes the tree structure to support the inheritance and overriding of weights, enabling operators to quickly adjust the value strategy for the entire province or the entire business line. For example, when the compliance service market in a certain region is saturated, simply reducing the weight of the node in that province will have a global effect, greatly improving the system's operational flexibility and configurability.

[0072] Example 6:

[0073] The value quantification and pricing module generates lead commercial value scores and pricing results based on preset multi-dimensional numerical calculation logic, including: obtaining the amount score corresponding to the target value feature and the intent score corresponding to the user intent intensity based on preset numerical mapping relationships; accumulating the category weight score, region weight score, amount score and intent score to generate a comprehensive value score; obtaining a preset global price adjustment coefficient, calculating the product of the comprehensive value score and the global price adjustment coefficient, and defining the calculated product as the pricing result.

[0074] This embodiment details the calculation logic of the value quantification and pricing module in generating the final pricing result, aiming to integrate multi-dimensional features into a single commercial price. Based on a preset numerical mapping relationship, such as a piecewise function or a lookup table, the system converts physical values ​​into normalized scores to obtain the monetary score corresponding to the target value feature. and the intent score corresponding to the strength of user intent. ;

[0075] The specific configuration is as follows: Establish the target amount characteristics based on a piecewise function. With amount of points The mapping relationship is calculated using the following formula:

[0076]

[0077] For the intention score If determined to be a strong intent strength, then assign a value. If the intent is determined to be weak, then assign a value. The system uses the following formula to calculate the comprehensive value integral. The calculation formula is as follows:

[0078]

[0079] Among them, the regional weight score involved in the calculation The results are directly taken from the regional weight tree. To address the issue of magnitude balance, the values ​​in the regional weight tree have been normalized in the preset stage with reference to a benchmark constant, for example, preset to the range of [10, 100]. Therefore, no additional coefficient is needed during runtime. The calculation formula is as follows:

[0080]

[0081] in, This refers to the classification weight score, derived from the classification weight tree, representing the preset commercial value weight of the region; to ensure the validity of the calculation results, the region's weight score... The preset value range is The integer value is determined based on historical transaction data of the local compliant service market and a weighted average of GDP per capita. To eliminate the difference in magnitude between the average transaction amount and GDP per capita, this system adopts a weighted calculation strategy after component normalization. A normalization mapping function is defined. ,in This is the current indicator value. The historical statistical data set for this indicator is used, and its calculation formula is as follows:

[0082]

[0083] like Then let ;

[0084] In the pre-set phase, the system constructs a regional weight tree based on the region's economic indicators, and assigns weight scores to each node in the regional weight tree. The following formula is used for calculation and storage:

[0085] in, This represents the historical transaction amount of compliant services corresponding to the target geographical location. It represents the set of transaction amounts across all regions in the historical statistical data set; This represents the GDP per capita corresponding to the target geographical location. This represents the set of per capita GDP values ​​for all regions within the historical statistical data set; the weights are set to... This formula ensures that economic indicators of different dimensions are independently mapped to... The data is then merged after the intervals are defined, which avoids the large amount of data overwhelming the weights.

[0086] The system utilizes rounding operators to ensure that the result reaches the boundary value of 100. Based on this dynamic calculation formula, the typical value calculated by the system during actual operation is as follows: For first-tier cities, such as Beijing and Shanghai, the calculation result is... The calculation typically converges to 100; for second-tier provincial capital cities, the result... The values ​​typically fall within the [60, 90] range; for general prefecture-level cities or county-level areas, the calculation results are... It usually falls within the range of [10, 50]; The monetary score is obtained from the above interval mapping rules; The intention score is obtained from the aforementioned intention determination rules;

[0087] The system obtains the preset global price adjustment coefficient. In order for the program to automatically recognize the current time The system has a pre-defined time boundary determination logic for peak or off-peak periods: the system obtains the current server time. Extract the weekdays. Value and hours Value ;like and Determined as a weekday peak period, a time fluctuation factor is set. ;like or It was determined to be a nighttime off-peak period, and a time fluctuation factor was set. The remaining time periods are set to the default baseline value. To solve the integration in the preceding steps Pricing results due to dimensionless values To address the lack of a currency unit, the system introduces a preset currency conversion constant. ,For example At this time, the global price adjustment coefficient Calculated as The system calculates the pricing result based on this coefficient. The calculation formula is as follows:

[0088]

[0089] The final pricing result The physical unit is the Renminbi Yuan; due to ,in The dimension is yuan / minute. It is a dimensionless comprehensive integral, and the multiplication of the components realizes the physical dimension transformation from abstract credit assessment to specific monetary value.

[0090] Example 7:

[0091] The feature extraction module analyzes the intensity of user intent, including: using a preset intent recognition model to analyze action verbs and time adverbs in unstructured text data; if semantic features representing delegation or instant contact are identified, it is determined to be strong intent intensity; otherwise, it is determined to be weak intent intensity.

[0092] This embodiment describes the specific method for parsing user intent in the feature extraction module. It adopts a method based on a combination of rules and semantics to predict conversion rates. The system uses a preset intent recognition model, usually a sequence labeling model based on the Transformer architecture, to perform semantic analysis on action verbs in the text, such as suing and entrusting, and time adverbs, such as immediately and recently.

[0093] The model outputs a sequence of BIO tags corresponding to the text sequence, such as O, B-Action, I-Action, O, B-Time, ...; the system decodes the tag sequence and aggregates and extracts a set of action entities. and time entity set The system executes feature matching and judgment logic; the system has a pre-set strong intent keyword library. ;

[0094] Traverse the extracted action entities If it exists Make Or satisfy the combination conditions If the system identifies a user who simultaneously demonstrates a clear need for compliance and a sense of urgency, then the system determines that the clue is of strong intent and assigns it a value. Otherwise, it is judged as a weak intent strength and assigned a value. In other words, if the system identifies semantic features representing a desire for agency, such as wanting to hire a lawyer for a lawsuit, or immediate contact, such as asking to contact me by phone, the system determines that the clue has strong intent. Conversely, if the system only identifies features representing a simple inquiry into compliance knowledge or a distant, unrelated desire for contact, the clue is considered a miss. Furthermore, since there is no adverb of time to accompany it, the system judges it as a weak intent strength;

[0095] This embodiment distinguishes between strong and weak intent, enabling the selection and assignment of high-potential leads with urgent compliance service needs to high scores. In scenarios with limited legal resources, this mechanism effectively prevents lawyers from wasting valuable time on ineffective free consultations, significantly improving the conversion rate after lead distribution.

[0096] Example 8:

[0097] The security filtering module is used to clean unstructured text data before the data access and positioning module executes. It matches the unstructured text data based on a preset blacklist and sensitive word database. If a match is found, the data is intercepted and the subsequent process is terminated. If no match is found, the cleaned data is transmitted to the cascaded semantic analysis module.

[0098] This embodiment adds a security filtering module, placed at the forefront of the data processing flow, to ensure system security and compliance. Although the data access and positioning module is logically responsible for data acquisition, in the physical data flow, the raw traffic passes through the security filtering module at the gateway layer. The system loads a preset blacklist and sensitive word database to match and scan unstructured text data. In response to a match in any database, the system directly intercepts the data, logs it, and terminates the subsequent process, no longer consuming the computing resources of the backend modules. In response to a no-match, the system allows the cleaned compliant data to pass through, allowing it to enter the data access and positioning module for parsing and positioning, and then transmits it to the cascaded semantic analysis module for processing. This mechanism achieves cleaning before the core business modules execute from a physical architecture perspective, resolving logical timing contradictions.

[0099] This embodiment constitutes the first line of defense for the system, effectively blocking spam and risky content in scenarios with open internet data access. This not only protects the expensive backend computing resources from being abused, but also avoids compliance risks that may arise from the platform distributing illegal content.

[0100] Example 9:

[0101] It also includes: an auction ranking and display module, which receives pricing results and target geographical locations; writes the pricing results back to the price attribute field of the database; and responds to front-end query requests by creating a dual index based on the price attribute field and the target geographical location, and generating a list of leads to display in descending order of pricing results.

[0102] This embodiment adds an auction ranking display module for the final distribution and monetization of leads; after the pricing result is generated, the system will display the calculated pricing result. The system writes back to the price attribute field of the lead record in the database; in response to the query request from the front-end lawyer user, which usually includes regional filtering conditions, the system builds a dual index based on the price attribute field and the target geographical location; using this index, the system quickly retrieves leads that meet the regional conditions and generates a lead display list strictly in descending order of pricing results, and pushes it to the front-end interface;

[0103] This embodiment achieves the business logic of prioritizing the exposure of high-value leads by directly applying the value quantification results to the sorting display; especially in high-concurrency query scenarios, the dual indexing mechanism ensures that the system can still return accurate results sorted by value in milliseconds, which greatly improves the screening experience for lawyer users and the platform's lead conversion rate.

[0104] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A case clue identification and analysis system, characterized in that, include: The data access and positioning module is used to acquire unstructured text data and network address data, call a preset multi-source map service interface to perform multi-channel parsing on the network address data, and determine the target geographical location based on a preset voting arbitration logic; The cascaded semantic analysis module is used to perform coarse-grained recognition on the unstructured text data using a primary classification model, determine the primary classification attributes, and dynamically route to the corresponding vertical expert agent based on the primary classification attributes. The vertical expert agent then generates secondary and tertiary fine-grained classification labels. The feature extraction module is used to perform numerical entity recognition on the unstructured text data, generate target features after removing preset interference numerical features, and simultaneously analyze the intensity of user intent. The value quantification and pricing module is used to map the target geographical location, the fine-grained classification label, the target value feature and the user intent intensity into multi-dimensional scores based on a preset tree-structured weight configuration system, and generate lead commercial value scores and pricing results according to preset multi-dimensional numerical calculation logic.

2. The case clue identification and analysis system according to claim 1, characterized in that, The data access and positioning module determines the target geographical location based on a preset voting arbitration logic, and is configured as follows: Obtain multiple candidate geographic location information returned by the multi-source map service interface; Detect the consistency of the multiple candidate geographic location information; If the multiple candidate geographic location information are completely consistent, then the consistent information is determined as the target geographic location; If there are partially consistent candidate geographic locations and the number of consistent locations exceeds a preset majority threshold, then the majority information is determined as the target geographic location. If none of the above conditions are met, then the candidate geographic location information corresponding to the preset authoritative source is selected as the target geographic location, or a front-end interaction command is triggered to obtain user confirmation information.

3. The case clue identification and analysis system according to claim 1, characterized in that, The cascaded semantic analysis module dynamically routes data to the corresponding vertical expert agent based on the primary classification attribute, including: Construct a group of intelligent agents containing multiple independent operating environments, where each intelligent agent corresponds to a pre-defined compliance business domain; In response to the generation of the primary classification attribute, a target vertical category expert agent that matches the primary classification attribute is activated in the agent group; The target vertical category expert agent is used to perform deep semantic analysis on the unstructured text data, and output the secondary and tertiary classification results belonging to the primary classification attribute.

4. The case clue identification and analysis system according to claim 1, characterized in that, The feature extraction module generates target value features after removing preset interfering numerical features, including: The unstructured text data is scanned using a preset Chinese amount regularization algorithm to extract the original numerical set; Identify strings in the unstructured text data that conform to preset mobile phone number rules and preset license plate number rules, and define them as the interference number features; Remove the portions of the original numerical data that overlap with the interfering numerical features, and then standardize the units of the remaining numerical data to generate the target value features.

5. The case clue identification and analysis system according to claim 1, characterized in that, The value quantification and pricing module, based on a preset tree-structured weight configuration system, maps the target geographical location and the fine-grained classification labels into multi-dimensional scores, including: Call the preset regional weight tree and classification weight tree, wherein the regional weight tree contains three-level nodes of province, city and county, and the classification weight tree contains first-level, second-level and third-level classification nodes; Based on the target geographical location, retrieve the corresponding regional weight score in the regional weight tree; Based on the fine-grained classification label, the corresponding classification weight score is retrieved in the classification weight tree.

6. The case clue identification and analysis system according to claim 5, characterized in that, The value quantification and pricing module generates lead commercial value scores and pricing results based on preset multi-dimensional numerical calculation logic, including: Based on a preset numerical mapping relationship, the amount score corresponding to the target amount feature and the intent score corresponding to the user intent strength are obtained; The classification weight score, the regional weight score, the monetary score, and the intent score are summed to generate a comprehensive value score. Obtain a preset global price adjustment coefficient, calculate the product of the comprehensive value integral and the global price adjustment coefficient, and define the calculated product as the pricing result.

7. A case clue identification and analysis system according to claim 4, characterized in that, The feature extraction module analyzes the intensity of user intent, including: The action verbs and time adverbs in the unstructured text data are analyzed using a preset intent recognition model; If semantic features representing delegation or instant communication are identified, it is determined to be of strong intent strength; Otherwise, it is judged as a weak intent strength.

8. The case clue identification and analysis system according to claim 1, characterized in that, Also includes: A security filtering module is used to clean the unstructured text data before the data access and positioning module is executed; The unstructured text data is matched based on a preset blacklist and sensitive word database. If a match is found, the data is intercepted and the subsequent process is terminated; if no match is found, the cleaned data is transmitted to the cascaded semantic analysis module.

9. A case clue identification and analysis system according to claim 1, characterized in that, Also includes: The auction ranking and display module is used to receive the pricing results and the target geographical location; Write the pricing result back to the price attribute field in the database; In response to a front-end query request, a dual index is established based on the price attribute field and the target geographical location, and a list of leads is generated in descending order of the pricing results.

Images