Fraud call identification method and system based on causal graph and agent cooperation

Abstract

A fraud call recognition method and system based on causal graph and agent collaboration comprises the steps of obtaining a fraud ASR text, mapping extracted multi-category to-be-mapped objects into semantic concept variable vectors, then learning a causal structure based on a data matrix formed by all semantic concept variable vectors, constructing a fraud causal graph, and recognizing a fraud call in the fraud ASR text. Calculating a causal effect of the reason variable on the result variable, and storing the causal effect as a relation weight in a causal graph; the method comprises the steps of obtaining a to-be-recognized ASR text, mapping multiple categories of to-be-mapped objects extracted from the to-be-recognized ASR text into semantic concept variable vectors, then retrieving matched causal chains in a causal graph to obtain a plurality of candidate causal chains, and then generating fraud judgment and risk degree evaluation values based on a large language model inference engine. And determining whether the to-be-identified call is a fraud call. The invention relates to the field of telecommunication anti-fraud, can deeply fuse a causal knowledge base and a large language model, and effectively improves the effectiveness, adaptability and reliability of communication anti-fraud identification.

Description

Technical Field

[0001] This invention relates to a method and system for identifying fraudulent calls based on causal graphs and agent collaboration, and relates to the field of telecommunications anti-fraud. Background Technology

[0002] As large language models demonstrate their powerful capabilities in natural language processing tasks, their inherent problems, such as factual illusions, lack of industry knowledge, and lack of traceable reasoning, are becoming increasingly prominent. To alleviate these issues, the industry generally adopts a technical approach that combines large language models with external knowledge bases. Currently popular technical paradigms for combining large language models with knowledge bases include RAG (Vector Retrieval Paradigm) and GraphRAG (Graph Structure Paradigm).

[0003] RAG's core principle is to use text embedding models to convert unstructured knowledge into vectors, and then retrieve it based on semantic similarity. The main steps are as follows: 1) Document vectorization: After segmenting the document, convert it into vectors using an embedding model. 2) Knowledge base construction: Store the document vectors in a vector database to form a searchable knowledge base. 3) Query vectorization: Convert user queries into vectors using the same embedding model. 4) Retrieval and generation: Retrieve the most relevant document vectors from the knowledge base and combine them with a large language model to generate accurate answers. The limitation of the RAG scheme is that its retrieval process mainly relies on vector similarity search algorithms, which are difficult to effectively identify complex and varied fraudulent behaviors in anti-fraud applications, and lack understanding of deep intent, thus increasing the risk of false positives and false negatives.

[0004] GraphRAG's core principle is to describe various entities, concepts, and their relationships in the real world in a structured way. It can be used to represent relationships between knowledge and stores knowledge using graph databases (such as Neo4j), where nodes represent entities and edges represent relationships. The main steps are as follows: 1) Automatic Information Extraction: Entities are extracted from unstructured text using named entity recognition. Relationships between entities are identified using dependency parsing. Precise matching and constraints are performed by combining rule templates and dictionary matching. 2) Knowledge Base Construction: The extracted <entity-relationship-entity> triples are used as the basis for constructing a knowledge graph. 3) Subgraph Retrieval: Based on user queries, relevant subgraphs are retrieved from the constructed knowledge graph. 4) Enhanced Reasoning: The structured information of this subgraph is used as enhanced context and input into a large language model. The large language model integrates its own general knowledge and language capabilities with the structured knowledge input from external sources to generate more accurate, reliable, and traceable answers. The limitation of the GraphRAG scheme lies in its primary reliance on co-occurrence statistics in text to construct relationships, i.e., identifying frequently co-occurring entities and patterns in the text. For example, although A and B are frequently mentioned together, this does not mean that A caused B; their relationship may simply be correlation. This reliance can lead to the following problems: 1) False associations: The system may mistakenly associate two frequently co-occurring entities closely in the knowledge graph, when such association may be merely coincidental or caused by third-party factors, rather than a direct causal relationship. 2) Misjudgment of causal direction: The system identifies a strong association between two entities, but because it relies solely on text co-occurrence, it cannot determine which is the cause and which is the effect.

[0005] To address the limitations of the two techniques mentioned above, existing technologies have developed causal knowledge bases, aiming to explicitly model and manage causal relationships. The main steps are as follows: 1) Causal Extraction: Identify causal trigger words from the text and extract cause and effect events, forming triples of the form <cause entity, causal relationship, effect entity>. 2) Knowledge Storage: Store these causal triples in a graph database, where nodes represent causes or effects, and directed edges represent causal relationships, with the direction of the edges pointing from cause to effect. 3) Query and Reasoning: When a user query is received, the system matches the key entities in the query with nodes in the causal graph, and finds relevant causal chains by traversing the connected causal edges in the graph. This information is then provided as enhanced context to the large language model to assist in generating the answer.

[0006] Telecommunications fraud cases can lead to the leakage of personal data and financial losses. For telecom operators, there is an urgent need to implement anti-fraud solutions that can accurately identify and promptly alert users to telecommunications fraud in real time to protect people's property. Existing causal knowledge bases provide more logical causal evidence for large language models than traditional RAG and GraphRAG. However, in the field of telecommunications anti-fraud, cases of deep integration of causal knowledge bases and large language models are still scarce, and there are currently almost no related applications. The main technical challenges in deeply integrating causal knowledge bases and large language models are as follows:

[0007] 1. The problem of anti-fraud identification systems based on large language models lacking causal reasoning ability:

[0008] While current anti-fraud frameworks based on large language models and knowledge bases perform well in pattern recognition and association analysis, they fundamentally rely on statistical correlation principles and are more suitable as pattern matching tools. Large language models excel at discovering correlations between variables rather than causal relationships, while the core of anti-fraud decision-making lies in clarifying the causal chain of "why fraud occurs." This technological limitation leads to fundamental flaws in existing systems when dealing with complex and dynamic fraud scenarios. Specifically, the systems cannot proactively identify and establish causal chains in fraudulent behavior; they lack a deep understanding of the inherent causal logic of fraudulent methods, making it difficult to cope with variations and evolutions of fraudulent tactics; and the decision-making process lacks causal basis, resulting in insufficient interpretability.

[0009] 2. The dilemma of causal discovery caused by high-dimensional feature spaces and dynamic evolution fraud methods:

[0010] The variability of communication fraud scenarios and the frequent updates to methods require systems to extract a large number of features from multiple dimensions to achieve effective coverage. However, this high-dimensional feature strategy poses significant challenges to causal discovery algorithms: the computational complexity of causal discovery algorithms increases superlinearly with the number of feature variables, facing computational efficiency bottlenecks in high-dimensional feature spaces; statistical reliability decreases in high-dimensional spaces, affecting the accuracy and stability of causal relationship identification; fraudsters are highly adversarial and evolving, able to quickly create new fraud scripts, rhetoric, and strategies, making it difficult for any system based on a fixed feature set to maintain timeliness; existing causal knowledge bases cannot effectively respond to the identification and inference of new types of fraud. For example, when fraudsters quickly switch to other scripts, the original causal chain immediately becomes invalid, and the system's detection capability drops significantly.

[0011] 3. Statistical reliability issues in causal effect assessment under small sample conditions:

[0012] The field of telecommunications fraud prevention faces the inherent challenge of insufficient sample size, especially given the limited number of confirmed fraud cases and the fact that only a small number of cases can be observed in the early stages of new fraudulent tactics. This poses a severe challenge to the quantitative assessment of average causal effects: insufficient sample size leads to reduced statistical power, and even if a real causal effect exists, it may be wrongly concluded as "no effect" due to the inability to achieve statistical significance; the variance of causal effect estimation is huge in small sample environments, the confidence interval is too wide, and the estimation results are extremely unstable, with different sampling methods potentially yielding completely opposite conclusions; the influence of each sample point on the results is extremely significant, making the system exceptionally sensitive to outliers (such as abnormal data caused by ASR transcription errors, dialects, environmental noise, etc.), and a few outliers can severely distort the overall causal effect estimation.

[0013] The three technical issues mentioned above are interconnected and mutually reinforcing, collectively constraining the effectiveness, adaptability, and reliability of existing communication anti-fraud systems. Therefore, how to deeply integrate causal knowledge bases with large language models to effectively improve the effectiveness, adaptability, and reliability of communication anti-fraud identification has become a key technical concern for engineers. Summary of the Invention

[0014] In view of this, the purpose of this invention is to provide a fraud call identification method and system based on causal graph and intelligent agent collaboration, which can deeply integrate causal knowledge base with large language model and effectively improve the effectiveness, adaptability and reliability of communication anti-fraud identification.

[0015] To achieve the above objectives, this invention provides a method for identifying fraudulent calls based on causal graphs and agent collaboration, comprising:

[0016] Step 1: Obtain multiple fraud ASR text data from the fraud case corpus. Extract a multi-category object to be mapped from each fraud ASR text data and map each extracted multi-category object to a corresponding semantic concept variable vector. Then, use a causal discovery algorithm to learn the causal structure based on the data matrix composed of all semantic concept variable vectors, thereby constructing and generating a fraud causal graph. The causal variables in the fraud causal graph are all semantic concept variables, and the result variables are the fraud results. Then, use a causal effect evaluation algorithm to calculate the causal effect of each causal variable on the result variable and store it as a relation weight in the fraud causal graph. Finally, store the fraud causal graph in the knowledge base.

[0017] Step 2: Obtain the ASR text data of the call to be identified, extract a multi-category object from the ASR text data, and map the extracted multi-category object into the corresponding semantic concept variable vector. Then, search for matching causal chains in the fraud causal graph to obtain several candidate causal chains. Based on the large language model inference engine, generate a fraud judgment and corresponding risk assessment value for each candidate causal chain. Finally, determine whether the call to be identified is a fraudulent call.

[0018] To achieve the above objectives, the present invention also provides a fraud call identification system based on causal graphs and agent collaboration, comprising a data perception device built on an agent architecture, an NLP semantic parsing device, a semantic concept mapping device, a causal discovery device, a causal effect evaluation device, and a fraud identification agent, wherein:

[0019] The data sensing device monitors a corpus of fraud cases. When the monitoring detects an update to the corpus of fraud cases, it sends multiple updated fraud ASR text data to the fraud identification intelligent agent.

[0020] The NLP semantic parsing device extracts multiple categories of objects to be mapped from ASR text data;

[0021] A semantic concept mapping device maps multiple categories of objects to be mapped into corresponding semantic concept variable vectors;

[0022] The causal discovery device uses a causal discovery algorithm to learn causal structures and construct a fraud causal graph.

[0023] The causal effect assessment device calculates the causal effect of each cause variable on the outcome variable in the fraud causal graph and stores it as a relation weight in the fraud causal graph.

[0024] Fraud detection agents further include:

[0025] The causal graph training unit receives multiple fraud ASR text data sent by the data perception device, calls the NLP semantic parsing device to extract a multi-category object to be mapped from each fraud ASR text data, and calls the semantic concept mapping device to map each multi-category object to be mapped into a corresponding semantic concept variable vector. Then, all semantic concept variable vectors are combined into a data matrix, and the causal discovery device is called to learn the causal structure, thereby constructing and generating a fraud causal graph. The causal variables in the fraud causal graph are all semantic concept variables, and the result variables are the fraud results. Then, the causal effect evaluation device is called to calculate the causal effect of each causal variable on the result variable. Finally, the fraud causal graph is stored in the knowledge base.

[0026] The fraud identification unit acquires the ASR text data of the call to be identified, calls the NLP semantic parsing device to extract a multi-category object from the ASR text data, and calls the semantic concept mapping device to map the extracted multi-category object into a corresponding semantic concept variable vector. Then, it searches for matching causal chains in the fraud causal graph to obtain several candidate causal chains. Based on the large language model inference engine, it generates a fraud judgment and a corresponding risk assessment value for each candidate causal chain. Finally, it determines whether the call to be identified is a fraudulent call.

[0027] To achieve the above objectives, the present invention also provides a computing device, comprising:

[0028] Memory and processor;

[0029] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the fraud call identification method based on causal graph and intelligent agent collaboration.

[0030] To achieve the above objectives, the present invention also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the fraud call identification method based on causal graphs and agent collaboration.

[0031] Compared with the prior art, the beneficial effects of the present invention are: The present invention realizes a paradigm shift from relevance identification to causal reasoning, significantly improving the interpretability and decision reliability of the system: In response to the problem that existing large language model anti-fraud systems lack causal reasoning capabilities, by constructing a causal knowledge base and deeply integrating it with the large language model, the following significant technical effects are achieved: (1) Enhanced causal reasoning capability: The system is no longer limited to surface pattern matching, but can deeply understand the internal causal mechanism of fraudulent behavior, actively identify and construct a complete causal chain, and accurately answer the core question of "why it constitutes fraud"; (2) Dynamic causal relationship maintenance: Through the continuous learning and updating of the causal knowledge base, the system can capture To capture the evolution of fraud methods, identify variant fraud patterns in a timely manner, and effectively deal with complex and ever-changing fraud scenarios; (3) Improved decision interpretability: The decision-making process based on causal paths provides clear reasoning logic, making every anti-fraud judgment based on evidence, greatly enhancing the transparency and credibility of the system, and meeting strict regulatory compliance requirements; This invention achieves feature dimension compression and dynamic adaptation through a semantic abstraction layer, effectively solving the problem of high-dimensional dilemma and frequent changes in fraud methods: Facing the dilemma of causal discovery brought about by high-dimensional feature space and dynamic evolution of fraud methods, the innovative design of the semantic abstraction layer has achieved the following key technical effects: (1) Significantly improved computational efficiency, by mapping specific features (1) By reducing the number of abstract concepts, the feature dimension is effectively compressed, and the computational complexity of the causal discovery algorithm changes from superlinear growth to controllable growth, greatly improving the real-time performance of the system; (2) The dynamic adaptability is improved, and the system has the ability to respond to new fraud methods in real time. Even if there is an unprecedented fraud script, as long as its underlying semantics can be mapped to existing abstract concepts (such as "creating an emergency"), the system can accurately identify and process it; (3) The knowledge transfer capability is enhanced. The stability of the abstract semantic concepts enables the system to transfer the causal knowledge learned in one fraud scenario to other similar scenarios, accelerating the detection and response to new fraud patterns; This invention integrates the iterative learning of expert priors and data evidence. The method effectively overcomes the challenge of statistical reliability in small sample environments: In response to the problem of causal effect assessment in small sample environments in the field of communication anti-fraud, an expert knowledge and data-driven iterative fusion method is proposed, which produces the following outstanding technical effects: (1) Breakthrough in stability of small sample estimation: By introducing prior knowledge of domain experts, the system can provide stable causal effect estimation in small sample situations, effectively overcoming the problem of excessive variance in traditional methods; (2) Enhanced robustness to outliers: The introduction of expert priors reduces the system's sensitivity to single outlier samples. Even if there are data quality problems such as ASR transcription errors, the causal effect estimation results can still remain stable, improving the practicality and reliability of the system;(3) Adaptive learning mechanism: The system is designed with a smooth transition mechanism. When the sample size is small, it reasonably relies on expert prior knowledge. As the sample size accumulates, it gradually shifts to data-driven learning. This avoids the statistical unreliability under small sample size and prevents the bias that may be caused by prior knowledge. (4) Progressive knowledge evolution: Through an iterative update mechanism, the system can continuously integrate new data evidence, continuously optimize and correct causal knowledge, and achieve a natural transition from "expert experience-driven" to "data evidence-driven", ensuring the continuous improvement of system performance. Attached Figure Description

[0032] Figure 1 This is a flowchart illustrating an exemplary embodiment of the present invention of a fraud call identification method based on causal graphs and agent collaboration.

[0033] Figure 2 This is a flowchart illustrating a specific embodiment of the present invention, showing how to map each multi-category object to be mapped into a corresponding semantic concept variable vector.

[0034] Figure 3 This is a schematic diagram illustrating the structure of a fraud call identification system based on causal graphs and agent collaboration, as shown in an exemplary embodiment of the present invention.

[0035] Figure 4 This is a schematic diagram of the structure of a computer device shown in an exemplary embodiment of the present invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings.

[0037] (I) Terminology Explanation

[0038] To make the technical solution of the present invention clearer, the relevant core terms are explained below.

[0039] Traditional causal algorithms mainly revolve around two core tasks: causal discovery algorithms aim to answer whether a causal relationship exists (qualitative), while causal effect estimation focuses on the strength or degree of the causal relationship (quantitative).

[0040] (1) Causal discovery algorithms: These are tools that automatically infer causal relationships and their directions from observed data. Their core objective is to construct causal graphs. Their output is a series of causal atoms, thus providing a verifiable causal structure for the causal knowledge base. Commonly used causal discovery algorithms in this field include constraint-based PC and FCI algorithms, LiNGAM based on functional causal models, and NOTEARS based on continuous optimization.

[0041] (2) Causal effect estimation: By using values ​​such as average treatment effect (ATE) and risk ratio (RR), the causal relationship in the causal knowledge base is given a strength label, which makes the qualitative relationship "A may cause B" into a quantitative decision basis: "A will increase the probability of B by X%".

[0042] (3) Causal atom: It is the smallest, indivisible unit of causal relationship and the foundation of all complex causal structures. Its formal definition is as follows: ,

[0043] The globally unique identifier of this causal atom.

[0044] Causal variables are the active entities (which can also be events or states) that trigger causal effects.

[0045] Outcome variables are passive entities (or events or states) that bear the causal effects.

[0046] Relational descriptors are used for both quantitative and qualitative descriptions. right The causal effect. It is a complex structure, as shown below: ,

[0047] in,

[0048] Type: Relationship type, which can include, but is not limited to: direct relationship, indirect relationship, positive relationship, negative relationship.

[0049] Strength: A quantitative value of the strength of the causal effect. For example, ATE (Average Treatment Effect) = 0.8 means that, on average, make It increased by 80 percentage points.

[0050] Confidence: The confidence level of the system in the validity of the causal effect, usually a continuous value in the interval [0,1].

[0051] (4) Cause-effect graph: Usually represented by a directed acyclic graph. In this graph, nodes represent causes or effects, and directed edges are used to represent causal relationships. The direction of the edges is from the cause node to the effect node.

[0052] Intelligent agent: Generally refers to a system capable of autonomous perception, decision-making, and execution. In this invention, the intelligent agent is located at the top of the technology stack, forming a caller-initiated relationship with the underlying large language model (providing semantic understanding and generation), rule engine (providing deterministic logic), and knowledge base (providing domain facts). In anti-fraud scenarios, the intelligent agent acts as a task orchestration and decision-making hub, scheduling multiple modules in real time to achieve end-to-end automated response to fraudulent activities.

[0053] (II) Within the existing causal discovery framework, the operational steps for mining the causal graph of fraud corpora are as follows:

[0054] Multiple records from a fraud corpus are used as the analysis objects for causal discovery. Each record corresponds to a fraudulent call and is stored in ASR (Automatic Speech Recognition) text format. ASR technology is an intelligent processing technology that can automatically convert speech into text, thereby converting call speech into call text data.

[0055] Before applying the causal discovery algorithm, the ASR texts of the fraudulent corpus need to be feature-engineered to construct a data matrix of dimension N x Q. The matrix is ​​used as the input to the algorithm. Here, N represents the number of samples, for example, 1000 ASR texts; Q represents the number of feature variables extracted from each text. Each row of the matrix is ​​a feature vector of a fraudulent call sample, and each column corresponds to a feature variable. Observations across all samples.

[0056] Causal discovery algorithms use data matrices As input, analyze the set of variables. The joint distribution (especially the conditional independence relationship between variables) is used to identify and infer the implied causal relationships, thereby deriving a causal graph. This process can be represented as: ,

[0057] in,

[0058] This indicates a specific causal discovery algorithm.

[0059] This is the aforementioned N x Q data matrix.

[0060] It is the set of hyperparameters of the algorithm, such as the significance level of the conditional independence test, the regularization strength, etc.

[0061] In existing technical solutions, each This represents an observable or measurable feature. For example, the "emergency" scenario in a scam script can be encoded through multiple specific variables, such as:

[0062] : A binary variable indicating whether to mention "a relative's sudden illness".

[0063] : A binary variable indicating whether "cash flow" is mentioned.

[0064] : A binary variable indicating whether to mention "traffic accident".

[0065] Due to the ever-changing scenarios and frequent updates in methods of telecommunications fraud, in order to effectively capture various fraud patterns and improve coverage, it is necessary to extract a large number of features from multiple dimensions (Q is usually large). This strategy aims to enhance the coverage and robustness of the anti-fraud system through the breadth of features. However, the high-dimensional feature space may lead to significant computational challenges for causal discovery algorithms, manifested in: computational complexity increasing superlinearly with the number of variables Q, and decreased statistical reliability, thus affecting the accurate identification of causal relationships.

[0066] Furthermore, due to the adversarial and evolving nature of telecommunications fraud, once a script is identified and blocked, fraudsters often quickly create new scripts, scripts, and strategies. Therefore, it is virtually impossible to completely cover all fraud cases with specific characteristic variables. Any system based on a fixed set of features is destined to lag behind the latest fraud methods, and it is also impossible to predefine corresponding characteristic variables for fraudulent behaviors that have not yet been summarized and categorized.

[0067] This invention introduces a semantic abstraction layer on top of the existing causal discovery framework. Improvements will be made to the semantic abstraction layer. The formal definition is as follows: This semantic abstraction layer aims to cover the most critical semantic dimensions in the field of communication fraud prevention. Specifically, the semantic dimension of "emergency event" can be abstracted into semantic concept variables. .

[0068] Within this framework, It is the total number of semantic dimensions. Representing the Each semantic concept variable reflects a specific semantic dimension, and its value is obtained by mapping a set of enumerated values.

[0069] This invention first sets the variables Transform into a semantic abstraction layer Subsequently Causal discovery is performed. By introducing a semantic abstraction layer, the anti-fraud system no longer relies on infinitely expanding specific features, but instead maps specific features to a finite number of abstract fraud concepts. The advantage of this method is that even if new fraud scripts emerge, as long as the extracted features can be effectively mapped to existing abstract concepts, the system can identify and handle these newly emerging fraudulent behaviors. The specific implementation scheme of this invention is described in detail below.

[0070] like Figure 1 As shown, the present invention provides a method for identifying fraudulent calls based on causal graphs and agent collaboration, comprising:

[0071] Step 1: Obtain multiple fraud ASR text data from the fraud case corpus. Extract a multi-category object to be mapped from each fraud ASR text data and map each extracted multi-category object to a corresponding semantic concept variable vector. Then, use a causal discovery algorithm to learn the causal structure based on the data matrix composed of all semantic concept variable vectors, thereby constructing and generating a fraud causal graph. The causal variables in the fraud causal graph are all semantic concept variables, and the result variables are the fraud results. Then, use a causal effect evaluation algorithm to calculate the causal effect of each causal variable on the result variable and store it as a relation weight in the fraud causal graph. Finally, store the fraud causal graph in the knowledge base.

[0072] Step 2: Obtain the ASR text data of the call to be identified, extract a multi-category object from the ASR text data, and map the extracted multi-category object into the corresponding semantic concept variable vector. Then, search for matching causal chains in the fraud causal graph to obtain several candidate causal chains. Based on the large language model inference engine, generate a fraud judgment and corresponding risk assessment value for each candidate causal chain. Finally, determine whether the call to be identified is a fraudulent call.

[0073] In practical implementation, step one can run in offline batch processing mode, periodically learning new causal knowledge from an updated corpus of fraud cases. Step two can be deployed as an online service, processing real-time call ASR text and alerting users to identified fraudulent calls. Steps one and two achieve knowledge sharing and collaborative evolution through a unified causal knowledge base. The causal knowledge base is based on a graph database, supporting efficient graph structure queries and semantic similarity retrieval.

[0074] To identify entities, events, states, and relationships in causal graphs, this invention can also construct a dedicated NLP semantic parsing device based on an intelligent agent architecture. This NLP semantic parsing device utilizes existing NLP (Natural Language Processing) semantic parsing techniques to preprocess ASR text data. Figure 1In step one or step two, extracting a multi-category object from the ASR text data may further include:

[0075] Step A1: The NLP semantic parsing device preprocesses the ASR text, performing basic NLP processing such as text cleaning, word segmentation, part-of-speech tagging, dependency parsing, and named entity recognition (NER) to extract all entities, actions, states, and attributes from the ASR text. Specifically, this includes:

[0076] 1. Text Cleaning: Employing a combination of regular expressions and a domain-specific rule engine, this system systematically removes irrelevant characters, special punctuation marks, filler words, and stop words from dialogues. Cleaning rules are specifically designed to address typical noise patterns in communication anti-fraud scenarios, ensuring the purity and consistency of text semantics.

[0077] 2. Word segmentation and part-of-speech tagging: The standardized dialogue text is segmented into grammatically independent word units, and each word is labeled with its grammatical role in the sentence, including basic parts of speech such as nouns, verbs, and adjectives, as well as more granular subcategories;

[0078] 3. Dependency Syntax Analysis: This involves in-depth analysis of the internal grammatical structure of sentences, identifying dependency relationships between words, including core grammatical components such as subject-verb, verb-object, and attributive-head relationships. By constructing a syntactic dependency tree, the grammatical functions and interrelationships of each component in the sentence are clarified, providing a structured foundation for action recognition and event extraction.

[0079] 4. Named Entity Recognition (NER): NER tools are used to extract entity references with specific meanings from dialogue text, including but not limited to key information such as people's names, organization names, geographical locations, time expressions, and currency amounts;

[0080] 5. Sentiment Analysis: Apply sentiment analysis tools to determine the sentiment polarity and intensity of dialogue segments, identify changes in the speaker's emotional state, including basic sentiment categories such as positive, negative, and neutral, as well as more granular sentiment dimensions such as panic, anger, and trust.

[0081] 6. Grammatical Feature Analysis: Systematically detect specific sentence structures and grammatical patterns appearing in dialogues, including sentence patterns with specific pragmatic functions such as interrogative sentences, imperative sentences, and conditional sentences, as well as grammatical features with semantic importance such as negative expressions, passive voice, and comparative structures;

[0082] 7. Reference resolution: In dynamic, multi-turn dialogues, people use a lot of pronouns (such as "he" and "it") and demonstrative words (such as "this" and "that") to avoid repetition. This step associates these pronouns with the entities or events they refer to in the context.

[0083] Step A2: The NLP semantic parsing device will extract and deduplicat all entities, actions, states, and attributes to form multi-category objects to be mapped. : ,in, , , and They represent from the text Extract and deduplicate the set of entities, actions, states, and attributes.

[0084] This invention also constructs a dedicated semantic concept mapping device based on an intelligent agent architecture. This device achieves automated and intelligent mapping from concrete instances to semantic concept variables through a weight-adaptive multi-strategy collaborative mapping mechanism and a quantified confidence evaluation system. For example... Figure 2 As shown, Figure 1 In step one or two, mapping each multi-category object to be mapped into a corresponding semantic concept variable vector may further include:

[0085] Step B1: The semantic concept mapping device pre-sets multiple semantic dimensions, sets a semantic concept variable for each semantic dimension, and sets multiple enumeration values ​​for each semantic dimension.

[0086] Semantic concept variables must be precisely anchored to the core dimensions of the research domain, which directly affects the effectiveness of the causal graph. This invention can adopt a hybrid approach combining expert pre-definition and automatic agent identification. In the expert pre-definition stage, domain experts jointly formulate semantic dimensions and enumeration values ​​to ensure that semantic concept variables are aligned with business logic, avoid conceptual ambiguity, reduce the cost of misjudgment, and prevent the risks associated with relying solely on data-driven approaches. Experts will analyze fraud case corpora and use their professional knowledge to clarify the syntax and semantics of causal expressions, thereby providing a structured hypothesis space and verification benchmark for subsequent causal discovery algorithms. Since this invention focuses on the field of communication anti-fraud, semantic concept variables can cover the most critical semantic dimensions in the field of communication anti-fraud according to the following preset methods. In practical applications, these variables may include, but are not limited to, the following:

[0087] Semantic dimension Semantic concept variables Enumeration value Emergency <![CDATA[C urgency ]]> Family member falls seriously ill, cash flow problems, traffic accident, fabricated kidnapping, accidental activation of a million-dollar insurance policy with impending charges, package containing contraband, investment losses, flight delays, flight cancellations, frozen game accounts, stolen accounts Management Department <![CDATA[C authority_pret ense ]]> Securities companies, bank customer service, social security centers, medical insurance centers, tax bureaus, market supervision bureaus, telecommunications operators, disease control centers, Alipay customer service, WeChat Pay customer service Fund operations <![CDATA[C fund_operation ]]> Requests include: transferring funds, paying transaction fees, paying unfreezing fees, paying a security deposit, paying personal income tax, paying notary fees, purchasing gold, and purchasing cryptocurrency. Threat types <![CDATA[C threat_type ]]> Account freezing / suspension, suspected money laundering, negative credit history, asset seizure, ID card cancellation, termination of social security benefits, confiscation of virtual assets. Time constraints <![CDATA[C temporal_const raint ]]> Immediately, today, before midnight tonight, within half an hour, the last day, in a few minutes. Information Request <![CDATA[C information_re quest ]]> ID card number, bank card number, bank card password, payment password, SMS verification code, mobile phone number, credit card security code Relationship Roles <![CDATA[C relationship_r ole ]]> Your father, your mother, your son, your daughter, your wife, your husband, your friend, your classmate, your boss, your colleague Behavioral manipulation <![CDATA[C behavioral_coe rcion ]]> Enter your bank card number, provide your password, download the remote control software, share your screen, request an SMS verification code, verify your identity, keep your information confidential, prohibit communication with others, do not hang up, and enable Do Not Disturb mode. Psychological state <![CDATA[C psychological_ state ]]> Fear, greed, curiosity, anxiety, confusion Interactive control <![CDATA[C interaction_co ntrol ]]> Complete dominance, strong guidance, repeated questioning, demanding real-time progress reports, repeatedly emphasizing consequences, switching communication channels, questioning, and belittling. False promises <![CDATA[C false_commitme nt ]]> High returns, cashback on order boosting, bonuses, subsidies, high commissions, unsecured loans, shopping refunds, tax refunds, fast withdrawals, insider information.

[0088] Step B2: The semantic concept mapping device calculates the strength of evidence that each element in each multi-category object to be mapped maps to each enumerated value of each semantic concept variable. ,in, It is a multi-category object to be mapped The j-th element in, that is A word in the set of entities, actions, states, and attributes. It is the i-th enumerated value of the m-th semantic concept variable. yes Mapped to The strength of evidence, with a range of [0,1], It is an exact mapping function, when and Returns 1 if they match, otherwise returns 0. It is a semantic similarity mapping function, which calculates... and The semantic similarity is returned, and the similarity score in the interval [0,1] is returned. , These are weighting coefficients used to balance the contributions of the two mapping strategies, satisfying... and ;

[0089] Given the continuous evolution of fraud scenarios, the expert-predefined enumerated instances cannot cover all possible situations. To address this issue, in step B2, the semantic concept mapping device employs a weighted fusion approach, coordinating two mapping strategies: a precise mapping strategy and a semantic similarity mapping strategy. ① Precise mapping strategy: This strategy precisely matches elements in the multi-category objects to be mapped with the enumerated values ​​of semantic concept variables, achieving fast and accurate direct mapping. ② Semantic similarity mapping strategy: This strategy calculates the similarity between elements in the multi-category objects to be mapped and each semantic dimension in the vector space, thereby achieving generalized fuzzy mapping. This is achieved by dynamically adjusting the weight coefficients. , The semantic concept mapping device can balance the contribution of the two strategies in the final decision. This mechanism supports flexible configuration and optimization according to the business needs of different scenarios (such as prioritizing accuracy or coverage). It can not only handle known enumeration patterns, but also intelligently identify new and undefined fraudulent rhetoric and map it to the corresponding semantic concept variables.

[0090] Step B3: The semantic concept mapping device calculates the confidence level of the mapping result from each element to each semantic concept variable by aggregating the evidence strength of all enumerated values ​​of each element mapped to each semantic concept variable in each multi-category object: ,in, yes The confidence score of the mapping result to the m-th semantic concept variable is a continuous value between 0 and 1. Represents a mapping function. Let represent the 1st, 2nd, ..., Ith enumerated values ​​of the m-th semantic concept variable, where I is the total number of enumerated values ​​for the m-th semantic concept variable. , It is the i-th or k-th enumerated value of the m-th semantic concept variable. It is the weight coefficient of the i-th enumerated value. It is a normalization function, for example, using the sigmoid function to map the comprehensive score to the interval [0,1], and thereby determine the semantic concept variable mapped to each element: select the maximum value from the confidence scores of the mapping results from each element to all semantic concept variables, and the semantic concept variable corresponding to the maximum value is the semantic concept variable mapped to that element;

[0091] In step B3, Attention weights can also be introduced. In fraud detection scenarios, the attention mechanism can dynamically assign different levels of importance to each enumeration value. For attention scoring function, The calculation formula can be: ;

[0092] Step B4: The semantic concept mapping device determines the mapping confidence of each multi-category object to each semantic concept variable based on the semantic concept variable to which each element in the multi-category objects to be mapped is mapped and the corresponding mapping result confidence. ,in, It is a multi-category object to be mapped The mapping confidence to the m-th semantic concept variable, where A is a multi-class object to be mapped. The total number of elements mapped to the m-th semantic concept variable. It is a multi-category object to be mapped The a-th element of the m-th semantic concept variable is mapped to the m-th semantic concept variable. yes The confidence score of the mapping result to the m-th semantic concept variable. yes The weights can be set according to actual business needs, thereby generating a semantic concept variable vector corresponding to each multi-category object to be mapped: , It is a multi-category object to be mapped The corresponding semantic concept variable vector, , … yes The mapping confidence to the 1st, 2nd, ..., Mth semantic concept variables.

[0093] The semantic concept mapping device generates a comprehensive confidence score for each mapping result. This score is determined by the base confidence of each strategy and its weight coefficient, thereby ensuring that the mapping result truly reflects its credibility.

[0094] The semantic concept mapping device can also implement threshold-driven semantic similarity mapping. That is, a preset adjustable confidence threshold is used. If the maximum confidence value of the mapping result from an element in a multi-category object to each semantic concept variable is greater than or equal to the threshold, then the element is mapped to the semantic concept variable corresponding to the maximum value. If the confidence values ​​of all mapping results are less than the threshold, it is considered a potential new concept and sent to the semantic concept expansion device. The semantic concept expansion device expands the semantic dimension and adds new semantic concept variables, thereby continuously updating and improving the semantic concept system, realizing the system's learning loop and the self-evolution of the causal knowledge base. Specifically, during real-time analysis, there will be specific instances that cannot be mapped. For such cases, in-depth analysis can be conducted to ensure that the system can adapt to newly emerging fraud methods, exceed preset rules, and extract knowledge from unlabeled real-world data. This invention also constructs a dedicated semantic concept expansion device based on an intelligent agent architecture. In step B3, the semantic concept mapping device determines the semantic concept variable mapped to each element, which may further include:

[0095] Step B31: The semantic concept mapping device determines whether the maximum confidence value of the mapping result from each element to all semantic concept variables is greater than the confidence threshold. If yes, the semantic concept variable corresponding to the maximum value is the semantic concept variable mapped to the element, and this process ends. If no, it means that the element has no mapped semantic concept variable, and the element is sent to the semantic concept extension device.

[0096] Step B32: The semantic concept expansion device saves the elements sent by the semantic concept mapping device and determines at regular intervals (e.g., weekly) whether the number of saved unmapped elements has reached a preset threshold (e.g., 1000). If so, it performs cluster analysis and popularity assessment on all unmapped elements to identify high-frequency unmapped elements. Then, using a large language model, based on the existing semantic concept variables, it proposes concept expansion suggestions for high-frequency unmapped elements, thereby expanding the semantic dimension and adding new semantic concept variables.

[0097] Furthermore, this invention can also integrate other valuable feature variables to form a more comprehensive variable set. To conduct causal discovery. Specifically, the set of variables. It can also include communication behavior characteristics. User attribute characteristics and user's historical risk score , can be represented as: .

[0098] In step one, after obtaining the fraud causal graph using the causal discovery algorithm, this invention employs a causal effect assessment algorithm to accurately estimate each causal variable in order to quantify the causal relationships. Regarding the results Average causal effect These data will be stored as relation weights in the causal graph. Due to insufficient sample size in the field of telecommunications anti-fraud, the number of fraud cases (especially confirmed cases) is relatively small, and only a few cases are observed in the early stages of new fraudulent tactics. This poses a serious challenge to the quantitative assessment of average causal effects: ① Even if a real causal effect exists, insufficient samples may prevent statistical significance, leading to the erroneous conclusion of "no effect"; ② In small sample cases, the impact of each sample on the result is extremely significant, so one or two outliers (e.g., outliers due to ASR transcription errors, which are common in practical applications and are usually caused by dialects, noise, etc.) may severely distort the estimation of the entire causal effect. Therefore, this invention also constructs a dedicated causal effect assessment device based on an intelligent agent architecture: Considering the characteristics of the sample data volume of telecommunications anti-fraud corpus, a set of iterative fusion causal effect calculation methods combining domain expert knowledge and data-driven approaches is proposed. This aims to effectively integrate the qualitative knowledge of domain experts into the causal model and adaptively adjust it as sample data accumulates. In this way, the present invention can improve the stability of causal effect estimation in small sample cases, while relying heavily on data-driven results in large sample cases. Figure 1 In step one, a causal effect assessment algorithm is used to calculate the causal effect of each causal variable on the outcome variable, and these effects are stored as relation weights in the fraud causal graph. This may also include:

[0099] Step C1: The causal effect assessment device sets initial prior weight vectors for all predefined semantic concept variables. ;

[0100] Domain experts can be organized to evaluate predefined semantic concept variables, ranking or scoring each variable based on its relative importance in triggering fraudulent outcomes. These qualitative assessments are then transformed into a quantified initial weight vector. Expert evaluation can utilize a structured evaluation form with multiple evaluation dimensions. Each variable's importance is scored across these dimensions, assigning a score (between 0 and 10) to each variable. The scores from all experts are then normalized and their authority considered before being aggregated to obtain a final initial prior weight vector. ;

[0101] Step C2: The causal effect assessment device uses a causal effect assessment algorithm to calculate the causal effect assessment vector. And evaluate the causal effect vector Standardized to ;

[0102] Step C3: The causal effect evaluation device, based on the number of new samples added during each training iteration, employs an exponential smoothing-based weight update strategy to... Perform iterative updates and save: , It's the updated version. When executed for the first time, yes When the second round of iteration updates begins, It was saved from the previous iteration. , It is a smoothing factor used to control the speed at which weights are updated. An adaptive mechanism can be used to adjust the values, thereby adapting to changes in data quality. The calculation formula is as follows: , This is the number of new samples added during this training session. The sample size for reference (e.g., set to 500). Based on the smoothing factor, As the minimum smoothing factor, when the amount of new sample data is small in each iteration, the system will rely more on historical weights, while when the amount of new data is sufficient, the system will trust the newly estimated weights more.

[0103] Step C4: Use of the causal effect assessment device , causal effect evaluation vector Weighted adjustments are made to complete the cause-effect graph optimization. The adjustment formula can be as follows: tanh is the adjustment function. To prevent over-adjustment, this invention also sets control parameters for the adjustment. It is the causal variable in the causal diagram of fraud. Regarding the results The causal effect assessment value, It is the adjusted version , yes The b-th weight value in the equation, It is a scaling factor used to control the degree to which the weight size affects the adjustment magnitude. This is the global scaling factor, used to control the overall adjustment range to prevent over-adjustment. The adjustment coefficient ranges from [1-]. , 1+ ], yes The mean of all weights in the middle, when equal When the adjustment coefficient is 1, it means no adjustment is made. Higher than When the adjustment coefficient is greater than 1, the causal effect is enhanced. Below When the adjustment coefficient is less than 1, the effect is weakened.

[0104] The causal effect assessment device uses a closed loop of "causal discovery → effect assessment → weight update" to gradually converge the initial rough estimate to an accurate value that can truly reflect the causal relationship between variables.

[0105] Step two, determining whether the call to be identified is a scam call, may further include:

[0106] The risk assessment values ​​of all candidate causal chains pointing to fraudulent results are weighted to obtain a comprehensive score. The comprehensive score is then determined to be greater than a preset decision threshold. If it is, the call to be identified is a fraudulent call.

[0107] In summary, the inventiveness of the present invention compared with the prior art is summarized as follows:

[0108] 1. In existing technical solutions, each input feature of the causal discovery algorithm represents an observable or measurable feature. These features are information about entities, actions, states, and attributes extracted from text using NLP technology. Because existing solutions rely on specific instance features for causal mining, a large number of features need to be extracted to improve coverage. However, the high-dimensional feature space leads to a significant increase in the computational complexity of the causal discovery algorithm and a decrease in statistical reliability. Furthermore, the fixed feature set approach cannot cope with the rapidly evolving nature of fraud methods after they have been blocked. Therefore, existing solutions cannot effectively address the challenges of the ever-changing and rapidly evolving nature of communication fraud scenarios. To construct a causal knowledge base more suitable for the field of communication anti-fraud, this invention introduces a new semantic abstraction layer based on the existing causal discovery framework. This layer covers the most critical semantic dimensions in the field of communication anti-fraud, achieving automated and intelligent mapping from specific variables (i.e., features of existing technical solutions) to semantic concept variables through semantic mapping. Simultaneously, it can continuously expand and improve the semantic concept dimensions, thereby realizing a closed-loop learning system and self-evolution of the causal knowledge base. By introducing a semantic abstraction layer, this invention allows the anti-fraud system to no longer rely on infinitely expanding specific features, but instead maps specific features to a limited set of abstract fraud-related semantic concepts. Even if new fraud scenarios emerge, as long as the extracted features can be effectively mapped to existing abstract concepts, these new fraudulent behaviors can be identified and processed, enhancing the system's flexibility and adaptability.

[0109] 2. In existing technical solutions, causal effect assessment often relies on the large sample assumption when calculating ACE, neglecting the "small sample" or "zero sample" scenarios caused by insufficient overall sample size, especially in new types of fraud cases. First, even if a real causal effect exists, insufficient samples may prevent statistical significance from being achieved, leading to the erroneous conclusion of "no effect." Second, in the case of small samples, the impact of each sample on the result is extremely significant; one or two outliers may even severely distort the entire causal effect estimation result. Therefore, existing solutions have shortcomings when dealing with small sample sizes. This invention addresses the characteristics of the communication anti-fraud corpus sample data by proposing an iterative fusion causal effect calculation method that combines domain expert knowledge with data-driven approaches: initial prior weights are obtained through domain expert evaluation; based on this, a new round of weights is obtained using an exponential smoothing-based weight update strategy and an adaptive mechanism, thereby refining the causal effect assessment value. Weighted adjustments are made. This invention improves the stability of causal effect estimation in small sample situations by enhancing the estimation of causal effects of highly important variables while weakening the influence of variables with low weights. Furthermore, it can fully rely on data-driven results when the sample size is sufficient.

[0110] 3. Existing technologies primarily rely on traditional RAG and GraphRAG combined with large language models to construct anti-fraud systems. The former depends on vector similarity search algorithms, while the latter mainly relies on co-occurrence statistics in text to construct relationships. The limitation of the RAG approach lies in its reliance on vector similarity search algorithms, which struggles to effectively identify complex and varied fraudulent behaviors in anti-fraud applications and lacks understanding of deeper intent, thus increasing the risk of false positives and false negatives. The limitation of the GraphRAG approach is its reliance on co-occurrence statistics in text to construct relationships, i.e., identifying frequently co-occurring entities and patterns in text. This method may lead to false associations and misjudgments of causal direction. This invention addresses the lack of causal reasoning capabilities in existing large language model anti-fraud systems by proposing an innovative solution: deeply integrating a causal knowledge base with a large language model and applying it to the field of communication anti-fraud. This enables the system not only to learn from the textual information of fraud case corpora but also to deeply understand the causal relationships of fraudulent behaviors, thereby improving the accuracy and effectiveness of the anti-fraud system. By constructing a causal knowledge base and deeply integrating it with a large language model, the limitations of traditional methods are broken through. This allows the system to move beyond superficial pattern matching and gain a deeper understanding of the intrinsic causal mechanisms of fraudulent behavior. As a result, it can proactively identify and construct complete causal chains, accurately answering the core question of "why it constitutes fraud".

[0111] like Figure 3As shown, the present invention discloses a fraud call identification system based on causal graphs and agent collaboration, which may include a data perception device, an NLP semantic parsing device, a semantic concept mapping device, a causal discovery device, a causal effect evaluation device, and a fraud identification agent built on an agent architecture, wherein:

[0112] The data sensing device monitors a corpus of fraud cases. When the monitoring detects an update to the corpus of fraud cases, it sends multiple updated fraud ASR text data to the fraud identification intelligent agent.

[0113] The NLP semantic parsing device extracts multiple categories of objects to be mapped from ASR text data;

[0114] A semantic concept mapping device maps multiple categories of objects to be mapped into corresponding semantic concept variable vectors;

[0115] The causal discovery device uses a causal discovery algorithm to learn causal structures and construct a fraud causal graph.

[0116] The causal effect assessment device calculates the causal effect of each causal variable on the outcome variable in the fraud causal graph and stores it as a relation weight in the fraud causal graph.

[0117] Fraud detection agents further include:

[0118] The causal graph training unit receives multiple fraud ASR text data sent by the data perception device, calls the NLP semantic parsing device to extract a multi-category object to be mapped from each fraud ASR text data, and calls the semantic concept mapping device to map each multi-category object to be mapped into a corresponding semantic concept variable vector. Then, all semantic concept variable vectors are combined into a data matrix, and the causal discovery device is called to learn the causal structure, thereby constructing and generating a fraud causal graph. The causal variables in the fraud causal graph are all semantic concept variables, and the result variables are the fraud results. Then, the causal effect evaluation device is called to calculate the causal effect of each causal variable on the result variable. Finally, the fraud causal graph is stored in the knowledge base.

[0119] The fraud identification unit acquires the ASR text data of the call to be identified, calls the NLP semantic parsing device to extract a multi-category object from the ASR text data, and calls the semantic concept mapping device to map the extracted multi-category object into a corresponding semantic concept variable vector. Then, it searches for matching causal chains in the fraud causal graph to obtain several candidate causal chains. Based on the large language model inference engine, it generates a fraud judgment and a corresponding risk assessment value for each candidate causal chain. Finally, it determines whether the call to be identified is a fraudulent call.

[0120] See Figure 4 , Figure 4 This is a structural block diagram of a computing device 400 illustrated in an exemplary embodiment of this specification. The components of the computing device 400 include, but are not limited to, a memory 410 and a processor 420. The processor 420 is connected to the memory 410 via a bus 430, and a database 450 is used to store data.

[0121] The computing device 400 also includes an access device 440, which enables the computing device 400 to communicate via one or more networks 460. Examples of these networks include a Public Switched Telephone Network (PSTN), a Local Area Network (LAN), a Wide Area Network (WAN), a Personal Area Network (PAN), or a combination of communication networks such as the Internet. The access device 440 may include one or more of any type of wired or wireless network interface (e.g., a network interface card (NIC)), such as an IEEE 402.11 Wireless Local Area Network (WLAN) interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, a Near Field Communication (NFC) interface, and so on.

[0122] In one embodiment of this specification, the aforementioned components of the computing device 400 and Figure 4 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 4 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0123] The computing device 400 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or PCs. The computing device 400 can also be a mobile or stationary server or cloud server, etc.

[0124] The processor 420 is configured to execute the following computer-executable instructions, which, when executed by the processor, implement the steps of the above-described fraud call identification method based on causal graphs and agent collaboration.

[0125] The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solution of the fraud call identification method based on causal graph and intelligent agent collaboration described above belong to the same concept. For details not described in detail in the technical solution of the computing device, please refer to the description of the technical solution of the fraud call identification method based on causal graph and intelligent agent collaboration described above.

[0126] An embodiment of this specification also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the above-described fraud call identification method based on causal graphs and agent collaboration.

[0127] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium belongs to the same concept as the above-described fraud call identification method based on causal graph and agent collaboration. For details not described in detail in the technical solution of the storage medium, please refer to the description of the technical solution of the above-described fraud call identification method or system based on causal graph and agent collaboration.

[0128] An embodiment of this specification also provides a computer program, wherein when the computer program is executed in a computer, it causes the computer to perform the steps of the above-described fraud call identification method based on causal graphs and agent collaboration.

[0129] The above is an illustrative scheme of a computer program according to this embodiment. It should be noted that the technical solution of this computer program belongs to the same concept as the technical solution of the fraud call identification method based on causal graph and intelligent agent collaboration described above. For details not described in detail in the technical solution of the computer program, please refer to the description of the technical solution of the fraud call identification method or system based on causal graph and intelligent agent collaboration described above.

[0130] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0131] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0132] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0133] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying fraudulent calls based on causal graphs and agent collaboration, characterized in that, Including: Step 1: Obtain multiple fraud ASR text data from the fraud case corpus. Extract a multi-category object to be mapped from each fraud ASR text data and map each extracted multi-category object to a corresponding semantic concept variable vector. Then, use a causal discovery algorithm to learn the causal structure based on the data matrix composed of all semantic concept variable vectors, thereby constructing and generating a fraud causal graph. The causal variables in the fraud causal graph are all semantic concept variables, and the result variables are the fraud results. Then, use a causal effect evaluation algorithm to calculate the causal effect of each causal variable on the result variable and store it as a relation weight in the fraud causal graph. Finally, store the fraud causal graph in the knowledge base. Step 2: Obtain the ASR text data of the call to be identified, extract a multi-category object from the ASR text data, and map the extracted multi-category object into the corresponding semantic concept variable vector. Then, search for matching causal chains in the fraud causal graph to obtain several candidate causal chains. Based on the large language model inference engine, generate a fraud judgment and corresponding risk assessment value for each candidate causal chain. Finally, determine whether the call to be identified is a fraudulent call.

2. The method according to claim 1, characterized in that, Step one runs in offline batch processing mode, learning new causal knowledge from the updated fraud case corpus on a regular basis. Step two is deployed as an online service to process real-time call ASR text and alert users to identified fraudulent calls. Step one and step two achieve knowledge sharing and co-evolution through a causal knowledge base.

3. The method according to claim 1, characterized in that, Based on an intelligent agent architecture, an NLP semantic parsing device is constructed. In step one or step two, a multi-category object to be mapped is extracted from the ASR text data, which further includes: Step A1: The NLP semantic parsing device preprocesses the ASR text, completing NLP processing such as text cleaning, word segmentation, part-of-speech tagging, dependency parsing, and named entity recognition, thereby extracting all entities, actions, states, and attributes from the ASR text; Step A2: The NLP semantic parsing device will extract and deduplicat all entities, actions, states, and attributes to form multi-category objects to be mapped. : ,in, , , and They represent from the text Extract and deduplicate the set of entities, actions, states, and attributes.

4. The method according to claim 1, characterized in that, Based on the intelligent agent architecture, a semantic concept mapping device is constructed. In step one or two, each multi-category object to be mapped is mapped into a corresponding semantic concept variable vector, including: Step B1: The semantic concept mapping device pre-sets multiple semantic dimensions, sets a semantic concept variable for each semantic dimension, and sets multiple enumeration values ​​for each semantic dimension. Step B2: The semantic concept mapping device calculates the strength of evidence that each element in each multi-category object to be mapped maps to each enumerated value of each semantic concept variable. ,in, It is a multi-category object to be mapped The j-th element in It is the i-th enumerated value of the m-th semantic concept variable. yes Mapped to The strength of evidence, with a range of [0,1], It is an exact mapping function, when and Returns 1 if they match, otherwise returns 0. It is a semantic similarity mapping function, which calculates... and The semantic similarity is returned, and the similarity score in the interval [0,1] is returned. , These are weighting coefficients; Step B3: The semantic concept mapping device calculates the confidence level of the mapping result from each element to each semantic concept variable by aggregating the evidence strength of all enumerated values ​​of each element mapped to each semantic concept variable in each multi-category object: ,in, yes The confidence score of the mapping result to the m-th semantic concept variable is a continuous value between 0 and 1. Represents a mapping function. Let represent the 1st, 2nd, ..., Ith enumerated values ​​of the m-th semantic concept variable, where I is the total number of enumerated values ​​for the m-th semantic concept variable. It is the i-th enumerated value of the m-th semantic concept variable. It is the weight coefficient of the i-th enumerated value. It is a normalization function, and based on this, the semantic concept variable mapped to each element is determined: the maximum value is selected from the confidence scores of the mapping results from each element to all semantic concept variables, and the semantic concept variable corresponding to the maximum value is the semantic concept variable mapped to that element. Step B4: The semantic concept mapping device determines the mapping confidence of each multi-category object to each semantic concept variable based on the semantic concept variable to which each element in the multi-category objects to be mapped is mapped and the corresponding mapping result confidence. ,in, It is a multi-category object to be mapped The mapping confidence to the m-th semantic concept variable, where A is a multi-class object to be mapped. The total number of elements mapped to the m-th semantic concept variable. It is a multi-category object to be mapped The a-th element of the m-th semantic concept variable is mapped to the m-th semantic concept variable. yes The confidence score of the mapping result to the m-th semantic concept variable. yes The weights are used to generate a semantic concept variable vector for each multi-category object to be mapped: , It is a multi-category object to be mapped The corresponding semantic concept variable vector, , … yes The mapping confidence to the 1st, 2nd, ..., Mth semantic concept variables.

5. The method according to claim 4, characterized in that, In step B3, Introducing attention weights, For attention scoring function, The calculation formula is: , It is the kth enumeration value of the mth semantic concept variable.

6. The method according to claim 4, characterized in that, In step B3, which involves constructing a semantic concept extension device based on an intelligent agent architecture, determining the semantic concept variable mapped to each element further includes: Step B31: The semantic concept mapping device determines whether the maximum confidence value of the mapping result from each element to all semantic concept variables is greater than the confidence threshold. If yes, the semantic concept variable corresponding to the maximum value is the semantic concept variable mapped to the element, and this process ends. If no, it means that the element has no mapped semantic concept variable, and the element is sent to the semantic concept extension device. Step B32: The semantic concept expansion device saves the elements sent by the semantic concept mapping device and determines at regular intervals whether the number of saved unmapped elements has reached a preset threshold. If so, it performs cluster analysis and popularity assessment on all unmapped elements to identify high-frequency unmapped elements. Then, using a large language model, based on the existing semantic concept variables, it proposes concept expansion suggestions for high-frequency unmapped elements, thereby expanding the semantic dimension and adding new semantic concept variables.

7. The method according to claim 1, characterized in that, A causal effect assessment device is constructed based on an intelligent agent architecture. In step one, a causal effect assessment algorithm is used to calculate the causal effect of each causal variable on the outcome variable, and these effects are stored as relation weights in a fraud causal graph, including: Step C1: The causal effect assessment device sets initial prior weight vectors for all predefined semantic concept variables. ; Step C2: The causal effect assessment device uses a causal effect assessment algorithm to calculate the causal effect assessment vector. And evaluate the causal effect vector Standardized to ; Step C3: The causal effect evaluation device, based on the number of new samples added during each training iteration, employs an exponential smoothing-based weight update strategy to... Perform iterative updates and save: , It's the updated version. When executed for the first time, yes When the second round of iteration updates begins, It was saved during the previous iteration. , It is a smoothing factor used to control the speed at which weights are updated. The value is adjusted using an adaptive mechanism. The calculation formula is as follows: , This is the number of new samples added during this training session. For reference sample size, Based on the smoothing factor, It is the minimum smoothing factor; Step C4: Use of the causal effect assessment device , causal effect evaluation vector Weighted adjustments are made to complete the cause-effect graph optimization. The adjustment formula is as follows: tanh is the adjustment function. It is the causal variable in the causal diagram of fraud. Regarding the results The causal effect assessment value, It is the adjusted version , yes The b-th weight value in the equation, It is a scaling factor used to control the degree to which the weight size affects the adjustment magnitude. It is the global scaling factor, used to control the overall adjustment range. yes The mean of all weights in the middle.

8. The method according to claim 1, characterized in that, Step two, determining whether the call to be identified is a scam call, further includes: The risk assessment values ​​of all candidate causal chains pointing to fraudulent results are weighted to obtain a comprehensive score. The comprehensive score is then determined to be greater than a preset decision threshold. If it is, the call to be identified is a fraudulent call.

9. A fraud call identification system based on causal graphs and agent collaboration, characterized in that, This includes a data perception device based on an intelligent agent architecture, an NLP semantic parsing device, a semantic concept mapping device, a causal discovery device, a causal effect evaluation device, and a fraud detection intelligent agent, among which: The data sensing device monitors a corpus of fraud cases. When the monitoring detects an update to the corpus of fraud cases, it sends multiple updated fraud ASR text data to the fraud identification intelligent agent. The NLP semantic parsing device extracts multiple categories of objects to be mapped from ASR text data; A semantic concept mapping device maps multiple categories of objects to be mapped into corresponding semantic concept variable vectors; The causal discovery device uses a causal discovery algorithm to learn causal structures and construct a fraud causal graph. The causal effect assessment device calculates the causal effect of each causal variable on the outcome variable in the fraud causal graph and stores it as a relation weight in the fraud causal graph. Fraud detection agents further include: The causal graph training unit receives multiple fraud ASR text data sent by the data perception device, calls the NLP semantic parsing device to extract a multi-category object to be mapped from each fraud ASR text data, and calls the semantic concept mapping device to map each multi-category object to be mapped into a corresponding semantic concept variable vector. Then, all semantic concept variable vectors are combined into a data matrix, and the causal discovery device is called to learn the causal structure, thereby constructing and generating a fraud causal graph. The causal variables in the fraud causal graph are all semantic concept variables, and the result variables are the fraud results. Then, the causal effect evaluation device is called to calculate the causal effect of each causal variable on the result variable. Finally, the fraud causal graph is stored in the knowledge base. The fraud identification unit acquires the ASR text data of the call to be identified, calls the NLP semantic parsing device to extract a multi-category object from the ASR text data, and calls the semantic concept mapping device to map the extracted multi-category object into a corresponding semantic concept variable vector. Then, it searches for matching causal chains in the fraud causal graph to obtain several candidate causal chains. Based on the large language model inference engine, it generates a fraud judgment and a corresponding risk assessment value for each candidate causal chain. Finally, it determines whether the call to be identified is a fraudulent call.

10. A computing device, characterized in that, include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the fraud call identification method based on causal graph and agent collaboration as described in any one of claims 1-8.

11. A computer-readable storage medium, characterized in that, It stores computer-executable instructions, which, when executed by a processor, implement the steps of the fraud call identification method based on causal graph and agent collaboration as described in any one of claims 1-8.

Images