Intelligent decision support system based on multiple AI modules

Abstract

The invention belongs to the technical field of data processing, and particularly relates to an intelligent decision support system based on multiple AI modules. The intelligent quality inspection module outputs data after quality inspection through operations such as text conversion and natural language processing; the data analysis module calls a multi-factor dynamic attribution algorithm to position abnormal data and key influence factors, generates decision suggestions in combination with historical data, and constructs a visual human-computer interaction interface by relying on a generative AI; the knowledge graph module integrates the data after quality inspection and internal and external data of an enterprise, and builds a knowledge graph service database; the intelligent pair training module constructs a real business scene and an intelligent pair training robot based on a large model and a knowledge graph business database, and generates a training task according to decision suggestions. According to the invention, the problems of low decision-making efficiency and low decision-making precision of a traditional decision-making mode in the prior art can be solved.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and in particular relates to an intelligent decision support system based on multiple AI modules. Background Technology

[0002] In enterprise decision support scenarios, traditional decision-making models generally have the following problems: First, there is a "silo effect" at the data level. Business data (such as customer service dialogue recordings and customer consultation records), knowledge data (such as policy terms and business specifications), and operational data (such as customer satisfaction and business processing efficiency) are scattered and stored in different systems. There is a lack of effective integration mechanisms, which makes it impossible to fully explore the value of data and make it difficult to obtain comprehensive data support when making decisions. Secondly, the decision-making process is highly "experience-dependent," especially in scenarios such as customer service training, business risk assessment, and strategic direction formulation. It relies heavily on the personal experience of senior employees or managers, which is highly subjective and easily affected by the limitations of individual cognition, making it difficult to guarantee the consistency and accuracy of decisions. Third, the decision-making tools are "single-function". Although there are independent AI tools in the existing technology (such as single voice quality inspection software and basic data analysis tools), there is a lack of synergy between the tools, which cannot form a complete technical closed loop of "data input-processing-analysis-decision-feedback". For example, the quality inspection tool can only identify the violation of the script, but cannot link the quality inspection results with customer service training and customer satisfaction analysis, resulting in a broken decision-making process and low efficiency. Fourth, knowledge utilization is "inefficient." The internal and external knowledge accumulated by enterprises (such as policy documents, historical decision-making cases, and market research data) mostly exists in unstructured forms (such as documents and audio), lacking systematic sorting and association. When faced with vague business needs or complex decision-making scenarios, staff members find it difficult to quickly locate key knowledge, delaying decision-making opportunities. Summary of the Invention

[0003] The technical problem solved by this invention is to provide an intelligent decision support system based on multiple AI modules, so as to solve the problems of low decision efficiency and low decision accuracy in the traditional decision-making mode in the prior art.

[0004] The basic solution provided by this invention is an intelligent decision support system based on multiple AI modules, including a data acquisition module, an intelligent quality inspection module, a data analysis module, a knowledge graph module, and an intelligent training module, wherein: The data acquisition module is used to collect voice data from the business system; The intelligent quality inspection module is used to perform text data conversion, natural language processing, quality inspection, and derivative data generation on voice data to obtain the quality-inspected data. The data analysis module is used to locate abnormal data and key influencing factors by calling a preset multi-factor dynamic attribution algorithm based on the post-quality inspection data, integrate the post-quality inspection data and historical data to generate decision suggestions, and combine generative AI technology to build a visual human-computer interaction interface. The knowledge graph module is used to integrate post-quality inspection data with internal and external enterprise data to build a knowledge graph business database; The intelligent training module builds real business scenarios and intelligent training robots based on large-scale model AI technology and knowledge graph business database, and generates training tasks based on the decision suggestions of the data analysis module.

[0005] Furthermore, the intelligent quality inspection module includes an ASR intelligent voice unit, an NLP natural language processing unit, a quality inspection unit, and a derived data generation unit, wherein: The ASR intelligent voice unit is used to receive voice data and transcribe it into text data. The NLP (Natural Language Processing) unit is used for word segmentation, part-of-speech tagging, semantic parsing, keyword extraction, sentiment generation, and semantic logic analysis of text data. The quality inspection unit is used to perform semantic similarity analysis on natural language processed text data according to a preset quality inspection rule base and a preset matching algorithm; The derivative data generation unit is used to generate derivative data on hot word distribution and tag association based on big data statistical analysis of text data.

[0006] Furthermore, the NLP (Natural Language Processing) unit includes a BERT semantic analysis subunit. This subunit constructs a BERT model based on a business domain corpus. The BERT semantic analysis subunit is used to generate context-dependent word vectors from the input text data based on the BERT model, and to disambiguate the semantics of the current pronoun and preceding entities based on the word vectors. The expression is as follows:

[0007] in, For pronouns, BERT vectors This is the BERT vector of the entity mentioned above.

[0008] Furthermore, the quality inspection unit performs semantic similarity analysis on the natural language processed text data according to a preset quality inspection rule base and a preset matching algorithm, specifically as follows: Calculate the semantic similarity between the text data to be inspected and the standard phrases in the preset quality inspection rule base; For each standard phrase in the pre-defined quality inspection rule base, the core intent is labeled, an intent library is constructed, the BERT model is fine-tuned based on the intent library, the text data to be inspected is input into the fine-tuned BERT model, and the probability distribution of intent categories is output. ; If the highest probability intent of the text data to be inspected matches the core intent in the inspection rule base... Figure 1 If the result is correct, the score is the probability value; otherwise, it is 0. Based on the semantic similarity results between the text data to be inspected and the preset quality inspection rule base, and the intent matching results between the text data to be inspected and the intent base, a comprehensive quality inspection result is generated.

[0009] Furthermore, the data analysis module uses a preset multi-factor dynamic attribution algorithm to locate abnormal data and key influencing factors, specifically as follows: Receive post-inspection data from the intelligent quality inspection module, acquire historical data from the business system, and perform data cleaning processing; A multi-factor dynamic attribution algorithm is constructed. Based on the decision-making objective indicators, the contribution of influencing factors is analyzed, and outliers and key influencing factors are identified. The expression is as follows:

[0010] Among them, the set of decision-influencing factors is: , Let i be the i-th factor in the set of factors influencing decision-making; Indicates decision-making objectives and indicators. Let represent the partial derivative of the target indicator with respect to the i-th factor. The factor proportion coefficient, Let be the weight of the i-th factor.

[0011] Furthermore, the data analysis module is also used to predict future indicator trends based on historical data from the business system using time series forecasting algorithms.

[0012] Furthermore, the knowledge graph module includes a knowledge base construction unit, a knowledge association unit, a knowledge matching unit, and a report generation unit, wherein: The knowledge base construction unit is used to collect internal and external data from enterprises, perform structured processing and entity extraction, and generate a business database. The knowledge association unit associates entities in the business database based on the knowledge association algorithm to generate a knowledge graph business database; The knowledge matching unit is used to vectorize and store knowledge graphs based on vector embedding technology, and combined with semantic retrieval algorithms to perform knowledge matching of knowledge graph business databases under fuzzy requirements; The report generation unit automatically generates a benchmarking analysis report based on the knowledge matched by the knowledge matching unit.

[0013] Furthermore, the intelligent training module includes an AI customer model generation unit, a training task generation unit, a human-computer interaction unit, and an evaluation unit, wherein: The AI ​​customer model generation unit constructs AI customer models with real customer characteristics based on a knowledge graph business database using a large model scenario generation algorithm. The training task generation unit receives decision suggestions from the data analysis module and creates training scripts based on business scenario requirements to generate training tasks. The human-computer interaction unit is used for trainees to practice training tasks with AI customer models; The evaluation unit is used to evaluate the practical results based on the evaluation indicators.

[0014] The principles and advantages of this invention are as follows: In the technical solution of this application, the core architecture is based on the collaboration of multiple AI modules. First, the data acquisition module acquires business voice data, which is then processed by the intelligent quality inspection module through ASR transcription, NLP processing, rule matching quality inspection, and derivation data generation, and outputs the quality-inspected data. Next, the knowledge graph module integrates internal and external data to build a business database. The data analysis module calls a multi-factor dynamic attribution algorithm to locate anomalies and key influencing factors, and combines generative AI and time series prediction to generate decision suggestions and build a visualization interface. Finally, the intelligent training module builds real business scenarios and AI robots based on large models and knowledge graphs, and generates training tasks based on decision suggestions, forming a closed-loop technical link of "data acquisition - quality inspection and analysis - knowledge accumulation - decision support - training implementation".

[0015] The advantages are as follows: Through multi-module collaboration and the application of specialized AI technologies, it achieves intelligent processing of business voice data throughout the entire process. It not only improves the accuracy of quality inspection by leveraging the BERT model and semantic similarity matching, and accurately locates the root cause of business anomalies through multi-factor dynamic attribution algorithms, but also integrates full-domain data based on knowledge graphs to form a business knowledge base, and provides scientific decision-making suggestions by combining generative AI and time series prediction. At the same time, the intelligent training module transforms decision-making suggestions into practical training tasks, which not only improves the efficiency of business quality inspection, the scientific nature of decision-making, and the value of data utilization, but also strengthens the capabilities of business personnel in a targeted manner, helping enterprises achieve multiple goals such as business process optimization, early risk prediction, and team capability enhancement. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of an embodiment of the present invention. Detailed Implementation

[0017] The following detailed description illustrates the specific implementation method: The basic implementation examples are as follows: Figure 1 As shown: An intelligent decision support system based on multiple AI modules includes a data acquisition module, an intelligent quality inspection module, a data analysis module, a knowledge graph module, and an intelligent training module, wherein: The data acquisition module is used to collect voice data from the business system. In this embodiment, it connects to the business system through API interface, direct database connection or streaming data transmission protocol to collect voice data in real time. At the same time, it performs format standardization processing on the real-time collected voice data, such as uniformly converting it into MP3 format, and adding metadata such as timestamps and business identifiers (customer service ID, customer ID). Finally, it outputs standardized voice data with metadata and transmits it to the intelligent quality inspection module.

[0018] The intelligent quality inspection module is used to perform text data conversion, natural language processing, quality inspection, and derivative data generation on voice data to obtain quality-inspected data. The intelligent quality inspection module includes an ASR intelligent voice unit, an NLP natural language processing unit, a quality inspection unit, and a derivative data generation unit, wherein: The ASR intelligent speech unit is used to receive speech data and transcribe it into text data. In this embodiment, the ASR intelligent speech unit adopts the Conformer-Transducer, an end-to-end speech writing model based on deep learning, to perform noise reduction and track splitting on the speech data and convert it into text data. Noise reduction mainly removes background noise, and track splitting distinguishes the speech of the two parties in the dialogue. The generated text data is then automatically completed with punctuation marks and optimized with sentence segmentation to ensure text readability.

[0019] The Natural Language Processing (NLP) unit is used for word segmentation, part-of-speech tagging, semantic parsing, keyword extraction, sentiment analysis, and semantic logic analysis of text data. In this embodiment, word segmentation uses the jieba word segmentation tool and optimizes segmentation accuracy by combining it with a business domain dictionary (e.g., a financial or e-commerce terminology database). Part-of-speech tagging mainly involves tagging nouns, verbs, and adjectives. Semantic parsing analyzes the subject-verb-object structure of sentences in the text data and extracts core semantics. Sentiment analysis uses a BERT fine-tuning model to classify the text into sentiment categories, such as positive, negative, and neutral, assigning different scores to different sentiments and ultimately outputting a sentiment score. Semantic logic analysis uses the BERT semantic analysis subunit to disambiguate context-related words. Specifically: The BERT semantic analysis subunit constructs a BERT model based on a business domain corpus. This subunit generates context-dependent word vectors from the input text data using the BERT model, and disambiguates the semantics of the current pronoun with the preceding entity based on these word vectors. The expression is as follows:

[0020] in, For pronouns, BERT vectors This is the BERT vector of the preceding entity. For example, pronouns such as "he" or "the product" are used. By calculating the cosine similarity between their word vectors and the word vectors of the preceding entities, disambiguation is achieved, clarifying that "he" refers to "customer A".

[0021] The quality inspection unit is used to perform semantic similarity analysis on natural language processed text data according to a preset quality inspection rule base and a preset matching algorithm; specifically: The semantic similarity between the text data to be inspected and the standard phrases in the preset quality inspection rule base is calculated. In practice, firstly, compliant or non-compliant phrases (e.g., "prohibition of promising processing time") in the preset quality inspection rule base are encoded using BERT to generate standard phrase vectors. Then, the text data to be inspected is encoded using BERT to generate the text vector to be inspected, followed by cosine similarity calculation. The output value ranges from [value missing in original text]. Between these values, a higher value indicates a closer semantic similarity.

[0022] For each standard phrase in the preset quality inspection rule base, a core intent is labeled to construct an intent base. In this embodiment, the core intents include compliance intents and violation intents. The compliance intents and their meanings include: Accurate Response: Provide accurate and complete answers to customer inquiries regarding business processes, policies, and rules; Proactive reminders: Proactively inform customers of the implicit requirements for handling business, such as the validity period of materials and the processing time limit; Emotional reassurance: Effectively soothe customers' dissatisfaction, anxiety, and other emotions, and express understanding and willingness to resolve the issues; Process guidance: The cleaning process guides customers through the next steps in completing their business. Standardized transfer: When a task cannot be processed due to limitations in permissions or capabilities, clearly state the reason for the transfer and provide the transfer path; End the call politely: After the call ends, proactively confirm whether the customer's needs have been resolved and express your willingness to provide further service.

[0023] The intent and meaning of the violation include: Irregular promises: Promises that exceed the scope of policy or cannot be fulfilled, such as guarantees of processing time or results; Refusing to handle customer requests without a reasonable explanation, or directly shirking responsibility; Poor attitude: Using aggressive or impatient language in conversation, or deliberately interrupting the customer; Irrelevant answer: Failing to address the client's core issues, deliberately avoiding or changing the subject; Leaking sensitive information: Unauthorized disclosure of customer privacy or non-public internal company information; Misleading: Guiding customers to make choices that do not meet their actual needs.

[0024] Therefore, an intent library is constructed based on the intent tags mentioned above.

[0025] Subsequently, the BERT model was fine-tuned based on the intent database. During the fine-tuning process, the sample data consisted of historical customer service dialogue texts, divided into training and validation sets. After labeling these texts with intent tags, a BERT + fully connected layer + softmax architecture was adopted. The training set was input, and the probability distribution of intent categories was output. Training and optimization were performed based on the loss function and optimizer until the accuracy of the validation set reached the target, at which point training stopped. After inputting the text data to be quality inspected into the fine-tuned BERT model, the probability distribution of 12 intent categories was output. Take the intent corresponding to the highest probability as the main intent, and output the intent image label and its corresponding probability.

[0026] If the highest probability intent of the text data to be inspected matches the core intent in the inspection rule base... Figure 1 If the match is found, the score is the probability value; otherwise, it is 0. Each rule in the quality inspection rule base has a preset target intent, which is either a compliance requirement that the rule mandates for customer service representatives or a prohibited violation. Matching is achieved through the following logic: Each rule contains an "intent constraint" field, for example:

[0027] Therefore, during the matching process, for mandatory rules (such as R-001), if the main intent of the text to be inspected matches the target intent of the rule... Figure 1 If the intent matches, it is considered a match and not a violation; otherwise, it is considered a lack of intent and is a violation. For prohibition rules (such as R-002), if the main intent of the text to be inspected violates the target intent of the rule... Figure 1 If the intent matches, it is considered a violation; otherwise, it is considered compliant and there is no violation intent.

[0028] Finally, based on the semantic similarity results between the text data to be inspected and the preset quality inspection rule base, and the intent matching results between the text data to be inspected and the intent base, a comprehensive quality inspection result is generated.

[0029] The derivative data generation unit is used to generate derivative data such as hot word distribution and tag association based on big data statistical analysis of text data. In this embodiment, the derivative data generation unit performs statistical analysis on batch text data based on the Spark big data framework, generates hot word distribution through word cloud algorithm, such as the frequency of occurrence of "complaint", and generates tag association data through association rule mining, such as the co-occurrence frequency of "refund" and "logistics delay".

[0030] The data analysis module is used to locate abnormal data and key influencing factors based on the post-quality inspection data by calling a preset multi-factor dynamic attribution algorithm, integrating the post-quality inspection data and historical data to generate decision suggestions, and combining generative AI technology to build a visual human-computer interaction interface; specifically, the data analysis module calls the preset multi-factor dynamic attribution algorithm to locate abnormal data and key influencing factors as follows: Receive post-inspection data from the intelligent quality inspection module, acquire historical data from the business system, and perform data cleaning processing; A multi-factor dynamic attribution algorithm is constructed. Based on the decision-making objective indicators, the contribution of influencing factors is analyzed, and outliers and key influencing factors are identified. The expression is as follows:

[0031] Among them, the set of decision-influencing factors is: , Let i be the i-th factor in the set of factors influencing decision-making; Indicates decision-making objectives and indicators. Let represent the partial derivative of the target indicator with respect to the i-th factor. The factor proportion coefficient, Let be the weight of the i-th factor.

[0032] In this embodiment, decision-making target indicators are first defined, such as customer satisfaction, complaint rate, and a set of influencing factors, such as compliance of communication scripts, response speed, and product knowledge proficiency. Then, a multi-factor dynamic attribution algorithm is used to calculate the contribution of each factor, locate key influencing factors according to the contribution ranking, and mark the abnormal data corresponding to factors with abnormal contribution.

[0033] Meanwhile, the ARIMA model is used to predict the trend of indicators for the next three months based on historical data; and key influencing factors, abnormal data and predicted trends are integrated to generate structured decision suggestions through generative AI, such as the GPT-4 fine-tuning model.

[0034] Finally, generative AI is used to automatically generate interactive logic, which is combined with data dashboards to support users in obtaining visual results through natural language queries.

[0035] The knowledge graph module is used to integrate post-quality inspection data and internal and external enterprise data to build a knowledge graph business database. This module includes a knowledge base construction unit, a knowledge association unit, a knowledge matching unit, and a report generation unit. The knowledge base construction unit is used to collect internal and external data of the enterprise, perform structured processing and entity extraction, and generate a business database. In this embodiment, external data is obtained by web crawling, paper documents are recognized by OCR, structured data in the internal database is integrated, and unstructured data is structured. The BERT-CRF model is used to extract entities from the text, such as product names, customer levels, and violation types, to generate the business database.

[0036] The knowledge association unit associates entities in the business database based on the knowledge association algorithm to generate a knowledge graph business database. In this embodiment, the knowledge association unit embeds entity vectors based on the TransE knowledge graph and establishes associations between entities through vector similarity matching (e.g., cosine similarity > 0.8), such as associating "Customer A" with "Complaint Type: Logistics", and constructs a knowledge graph business database, which stores entities, relationships and attributes.

[0037] The knowledge matching unit is used to vectorize and store the knowledge graph based on vector embedding technology, and combined with semantic retrieval algorithms to perform knowledge matching of the knowledge graph business database under fuzzy requirements. In this embodiment, through the knowledge matching unit, users can query and retrieve target data through semantic retrieval, and fuzzy requirement queries are supported.

[0038] The report generation unit automatically generates a benchmarking analysis report based on the knowledge matched by the knowledge matching unit.

[0039] The intelligent training module constructs real-world business scenarios and intelligent training robots based on large-scale AI model technology and a knowledge graph business database, and generates training tasks based on decision suggestions from the data analysis module. The intelligent training module includes an AI customer model generation unit, a training task generation unit, a human-computer interaction unit, and an evaluation unit. The AI ​​customer model generation unit constructs an AI customer model with real customer characteristics based on a knowledge graph business database using a large model scene generation algorithm. In this embodiment, the AI ​​customer model generation unit constructs an AI customer model based on customer characteristic data in the knowledge graph, such as age, demand type, and complaint preference, using a large model scene generation algorithm (such as GPT-4 prompt word engineering) to simulate the language style, demand expression, and emotional reaction of real customers. For example, younger customers tend to use popular internet slang.

[0040] The training task generation unit receives decision suggestions from the data analysis module and creates training scripts based on business scenario requirements to generate training tasks. In this embodiment, the training task generation unit automatically generates training scripts based on key influencing factors in the decision suggestions from the data analysis module, such as insufficient product knowledge proficiency, combined with business scenarios, such as new customers inquiring about product functions. The training scripts include customer questions, standard response procedures, and marked training focus, such as accurately specifying the product warranty period.

[0041] The human-computer interaction unit is used for trainees to practice training tasks with the AI ​​client model. In this embodiment, the human-computer interaction unit provides a text or voice interaction interface. Trainees can talk to the AI ​​client model by inputting text or voice. The system uses a built-in ASR unit to transcribe the trainee's voice into text in real time and convert the AI ​​client model's text response into voice.

[0042] The evaluation unit is used to evaluate the practical results based on evaluation indicators. In this embodiment, based on preset evaluation indicators, including script compliance, semantic matching degree, response time, etc., the NLP unit analyzes the trainee's response text and combines it with the similarity algorithm scoring of the quality inspection unit to generate an evaluation report. In the report, weak links are matched with the knowledge graph business database to recommend targeted learning materials.

[0043] Therefore, in this application, the modules communicate in real time through data interfaces: the data acquisition module triggers the intelligent quality inspection module, and the quality inspection results are synchronized to the data analysis module and the knowledge graph module; the decision suggestions of the data analysis module drive the intelligent training module to generate training tasks, and the knowledge graph module provides knowledge support for quality inspection, analysis and training, forming a closed loop of "data acquisition-intelligent processing-decision output-training optimization", thereby improving the level of intelligence of the entire business process.

[0044] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. An intelligent decision support system based on multiple AI modules, characterized in that: It includes a data acquisition module, an intelligent quality inspection module, a data analysis module, a knowledge graph module, and an intelligent practice module, among which: The data acquisition module is used to collect voice data from the business system; The intelligent quality inspection module is used to perform text data conversion, natural language processing, quality inspection, and derivative data generation on voice data to obtain the quality-inspected data. The data analysis module is used to locate abnormal data and key influencing factors by calling a preset multi-factor dynamic attribution algorithm based on the post-quality inspection data, integrate the post-quality inspection data and historical data to generate decision suggestions, and combine generative AI technology to build a visual human-computer interaction interface. The knowledge graph module is used to integrate post-quality inspection data with internal and external enterprise data to build a knowledge graph business database; The intelligent training module builds real business scenarios and intelligent training robots based on large-scale model AI technology and knowledge graph business database, and generates training tasks based on the decision suggestions of the data analysis module.

2. The intelligent decision support system based on multiple AI modules according to claim 1, characterized in that: The intelligent quality inspection module includes an ASR intelligent voice unit, an NLP natural language processing unit, a quality inspection unit, and a derived data generation unit, wherein: The ASR intelligent voice unit is used to receive voice data and transcribe it into text data. The NLP (Natural Language Processing) unit is used for word segmentation, part-of-speech tagging, semantic parsing, keyword extraction, sentiment generation, and semantic logic analysis of text data. The quality inspection unit is used to perform semantic similarity analysis on natural language processed text data according to a preset quality inspection rule base and a preset matching algorithm; The derivative data generation unit is used to generate derivative data on hot word distribution and tag association based on big data statistical analysis of text data.

3. The intelligent decision support system based on multiple AI modules according to claim 2, characterized in that: The NLP (Natural Language Processing) unit includes a BERT semantic analysis subunit. This subunit constructs a BERT model based on a business domain corpus. The BERT semantic analysis subunit generates context-dependent word vectors from the input text data using the BERT model, and disambiguates the semantics of the current pronoun with the preceding entity based on the word vectors. The expression is as follows: in, For pronouns, BERT vectors This is the BERT vector of the entity mentioned above.

4. The intelligent decision support system based on multiple AI modules according to claim 3, characterized in that: The quality inspection unit performs semantic similarity analysis on the natural language processed text data according to a preset quality inspection rule base and a preset matching algorithm, specifically as follows: Calculate the semantic similarity between the text data to be inspected and the standard phrases in the preset quality inspection rule base; For each standard phrase in the pre-defined quality inspection rule base, the core intent is labeled, an intent library is constructed, the BERT model is fine-tuned based on the intent library, the text data to be inspected is input into the fine-tuned BERT model, and the probability distribution of intent categories is output. ; If the highest probability intent of the text data to be inspected is consistent with the core intent in the quality inspection rule base, the score is the probability value; otherwise, it is 0. Based on the semantic similarity results between the text data to be inspected and the preset quality inspection rule base, and the intent matching results between the text data to be inspected and the intent base, a comprehensive quality inspection result is generated.

5. The intelligent decision support system based on multiple AI modules according to claim 4, characterized in that: The data analysis module uses a preset multi-factor dynamic attribution algorithm to locate abnormal data and key influencing factors, specifically as follows: Receive post-inspection data from the intelligent quality inspection module, acquire historical data from the business system, and perform data cleaning processing; A multi-factor dynamic attribution algorithm is constructed. Based on the decision-making objective indicators, the contribution of influencing factors is analyzed, and outliers and key influencing factors are identified. The expression is as follows: Among them, the set of decision-influencing factors is: , Let i be the i-th factor in the set of factors influencing decision-making; Indicates decision-making objectives and indicators. Let represent the partial derivative of the target indicator with respect to the i-th factor. The factor proportion coefficient, Let be the weight of the i-th factor.

6. The intelligent decision support system based on multiple AI modules according to claim 5, characterized in that: The data analysis module is also used to predict future indicator trends based on historical data from the business system using time series forecasting algorithms.

7. The intelligent decision support system based on multiple AI modules according to claim 6, characterized in that: The knowledge graph module includes a knowledge base construction unit, a knowledge association unit, a knowledge matching unit, and a report generation unit, wherein: The knowledge base construction unit is used to collect internal and external data from enterprises, perform structured processing and entity extraction, and generate a business database. The knowledge association unit associates entities in the business database based on the knowledge association algorithm to generate a knowledge graph business database; The knowledge matching unit is used to vectorize and store knowledge graphs based on vector embedding technology, and combined with semantic retrieval algorithms to perform knowledge matching of knowledge graph business databases under fuzzy requirements; The report generation unit automatically generates a benchmarking analysis report based on the knowledge matched by the knowledge matching unit.

8. The intelligent decision support system based on multiple AI modules according to claim 7, characterized in that: The intelligent training module includes an AI customer model generation unit, a training task generation unit, a human-computer interaction unit, and an evaluation unit, wherein: The AI ​​customer model generation unit constructs AI customer models with real customer characteristics based on a knowledge graph business database using a large model scenario generation algorithm. The training task generation unit receives decision suggestions from the data analysis module and creates training scripts based on business scenario requirements to generate training tasks. The human-computer interaction unit is used for trainees to practice training tasks with AI customer models; The evaluation unit is used to evaluate the practical results based on the evaluation indicators.

Images