Customer survey report generation method based on artificial intelligence and RPA

Abstract

The invention provides a customer survey report generation method based on artificial intelligence and RPA, and the method comprises the steps: synchronously obtaining customer survey data from a structured data source and an unstructured data source through an application programming interface, a database query engine, an optical character recognition technology and a robot process automation technology; performing text denoising, key field extraction and structured conversion on the unstructured data based on a preset rule, and performing format standardization, data verification and field mapping processing on the structured data; a dynamic semantic analysis model is utilized to adjust a data analysis strategy according to user requirements and context information, and a semantic analysis result with weights is generated; and performing anomaly detection, duplicate data de-duplication and missing value filling on the cleaned structured data and the semantic analysis result through a multi-modal data fusion algorithm, and constructing a unified data pool with a quality score not lower than 95%.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically relating to a method for generating customer survey reports based on artificial intelligence and RPA. Background Technology

[0002] Traditional customer survey reports typically rely on manual data collection, organization, and analysis, resulting in inefficiency, significant subjective bias, difficulties in data integration, and insufficient timeliness. While some companies have attempted to improve report generation efficiency through automation tools (such as RPA) or AI models (such as natural language processing and machine learning), these solutions still have the following limitations: Data collection is limited to a single source: it mainly relies on structured data (such as databases and tables) and has difficulty integrating multi-source heterogeneous data (such as unstructured text, voice, and images).

[0003] Static analysis mode: Existing AI models lack the ability to dynamically adjust the semantic parsing of data and cannot optimize the analysis results in real time according to the context or user needs.

[0004] The generated reports are mostly in a fixed format and lack the ability to adapt to user needs, resulting in insufficient matching between the content and the actual business scenario.

[0005] Existing method steps / device structure and their relationships: Data acquisition phase: RPA is used to crawl information from structured data sources such as databases and APIs; for unstructured data, manual screening or simple keyword matching is relied upon.

[0006] Data processing and analysis phase: Applying NLP algorithms to analyze the sentiment of text data, and using machine learning models to predict customer behavior trends, etc.

[0007] Report generation stage: Fill in the analysis results according to the preset template to generate a standardized report.

[0008] While this process improves efficiency to some extent, it is limited by static analysis capabilities and data source limitations, failing to fully explore potential value and achieve true personalized report customization. Summary of the Invention

[0009] The present invention aims to at least partially solve one of the technical problems in the related art.

[0010] Therefore, the first objective of this invention is to propose a method for generating customer survey reports based on artificial intelligence and RPA.

[0011] The second objective of this invention is to provide a customer survey report generation device based on artificial intelligence and RPA.

[0012] The third objective of this invention is to provide a computer device.

[0013] A fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0014] To achieve the above objectives, a first aspect of the present invention proposes a method for generating customer survey reports based on artificial intelligence and RPA, comprising: S1 synchronously acquires customer survey data from structured and unstructured data sources through application programming interfaces, database query engines, optical character recognition technology, and robotic process automation technology. S2 performs text denoising, key field extraction, and structure transformation on unstructured data based on preset rules, and performs format standardization, data validation, and field mapping processing on structured data. S3 utilizes a dynamic semantic parsing model to adjust the data parsing strategy based on user needs and contextual information, generating weighted semantic analysis results. S4 uses a multimodal data fusion algorithm to perform anomaly detection, deduplication, and missing value filling on the cleaned structured data and semantic analysis results, constructing a unified data pool with a quality score of no less than 95%.

[0015] In one embodiment of the present invention, S1 includes: S11 enables secure transmission of scanned PDF / JPG / PNG files manually uploaded by users via HTTPS and SFTP protocols; S12 uses the BERT model to perform entity recognition and relation extraction on the text after OCR recognition, generating a structured event summary.

[0016] In one embodiment of the present invention, S2 includes: S21 executes field mapping through the rules engine, where the field mapping rules include a three-tiered tagging system of customer type, data category, and time period.

[0017] In one embodiment of the present invention, S3 further includes: S31, calculate the similarity weight between historical dialogue records and the current user query based on the attention mechanism; S32 updates model parameters in real time based on user feedback through an online fine-tuning mechanism. The fine-tuning process includes correcting the analysis results and recalculating the weights.

[0018] In one embodiment of the present invention, S4 includes: S41 uses the Z-score anomaly detection method, the formula is: The data is standardized, with an anomaly detection threshold of [value missing]. ; S42, through the quality scoring formula Quantitatively evaluate the effectiveness of data cleaning.

[0019] To achieve the above objectives, a second aspect of the present invention provides a customer survey report generation apparatus based on artificial intelligence and RPA, comprising: The data synchronization acquisition module is used to synchronously acquire customer survey data from structured and unstructured data sources through application programming interfaces, database query engines, optical character recognition technology, and robotic process automation technology. The unstructured data processing module is used to perform text denoising, key field extraction, and structured transformation on unstructured data based on preset rules, as well as to perform format standardization, data validation, and field mapping processing on structured data. The dynamic semantic parsing module is used to adjust the data parsing strategy according to user needs and context information using a dynamic semantic parsing model, and generate weighted semantic analysis results. The multimodal data fusion module is used to perform anomaly detection, deduplication, and missing value filling on cleaned structured data and semantic analysis results through multimodal data fusion algorithms, and to build a unified data pool with a quality score of no less than 95%.

[0020] This invention discloses a customer survey report generation method and apparatus based on artificial intelligence and RPA, which can realize intelligent collection and fusion of multi-source heterogeneous data. Through dynamic semantic parsing and user feedback-driven online fine-tuning mechanism, it can significantly improve the accuracy and scenario adaptability of customer survey reports, and support adaptive generation of personalized report content, effectively improving the overall automation level and processing efficiency.

[0021] To achieve the above objectives, a third aspect of this application provides a computer device comprising a processor and a memory; wherein the processor runs a program corresponding to the executable program code stored in the memory, for implementing a customer survey report generation method based on artificial intelligence and RPA as described in the first aspect embodiment.

[0022] To achieve the above objectives, a fourth aspect of this application provides a non-transitory computer-readable storage medium storing a computer program that, when executed by a processor, implements a customer survey report generation method based on artificial intelligence and RPA as described in the first aspect embodiment.

[0023] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0024] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a customer survey report generation method based on artificial intelligence and RPA according to an embodiment of the present invention; Figure 2 This is a flowchart of the data acquisition module according to an embodiment of the present invention; Figure 3 This is a flowchart of the adaptive report generation module according to an embodiment of the present invention; Figure 4 This is an automation flowchart according to an embodiment of the present invention; Figure 5 This is a structural diagram of a customer survey report generation device based on artificial intelligence and RPA according to an embodiment of the present invention; Figure 6 It is a computer device according to an embodiment of the present invention. Detailed Implementation

[0025] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0026] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0027] The following description, with reference to the accompanying drawings, illustrates a customer survey report generation method and apparatus based on artificial intelligence and RPA according to an embodiment of the present invention.

[0028] Example 1 Figure 1 This is a flowchart of a customer survey report generation method based on artificial intelligence and RPA according to an embodiment of the present invention, such as... Figure 1 As shown, it includes: S1 synchronously acquires customer survey data from both structured and unstructured data sources through application programming interfaces, database query engines, optical character recognition technology, and robotic process automation technology.

[0029] In some implementations, the system first communicates with structured data sources (such as CRM systems and ERP databases) both internally and externally through API interfaces, using RESTful or GraphQL protocols to obtain standardized data. Simultaneously, the database query engine, based on SQL or NoSQL languages, performs structured queries on local or cloud databases to extract customer survey-related fields. For unstructured data sources (such as scanned PDFs, images, and emails), the system utilizes OCR technology for image text recognition, employing high-precision recognition engines such as Tesseract or Baidu OCR, achieving an accuracy rate of over 98%. The recognized text undergoes denoising processing, including punctuation removal, stop word filtering, and key field extraction, and is finally converted into structured JSON format using Natural Language Processing (NLP) algorithms for subsequent analysis.

[0030] Specifically, the OCR recognition module needs to be configured with the recognition language (e.g., Chinese, English), image preprocessing parameters (e.g., binarization threshold, noise reduction intensity), key field extraction rules (e.g., regular expressions for recognizing ID numbers), and a field mapping table for structured transformation. The RPA module needs to set up UI automation scripts, element recognition strategies (e.g., XPath, image matching), and data anonymization rules (e.g., masking, field encryption) to ensure data security and compliance. The data fusion stage uses the Z-score standardization formula for anomaly detection, sets thresholds to identify abnormal data, and combines missing value imputation strategies (e.g., mean imputation, interpolation) to improve data integrity.

[0031] Furthermore, S1 includes: S11 enables secure transmission of scanned PDF / JPG / PNG files manually uploaded by users via HTTPS and SFTP protocols.

[0032] In some implementations, the HTTPS protocol uses TLS 1.3 or later, employing asymmetric encryption (such as RSA or ECC) for client-server authentication and key exchange, followed by symmetric encryption (such as AES-256-GCM) to encrypt file content during transmission, ensuring data is not eavesdropped on or tampered with. The SFTP protocol, based on the SSH 2.0 stack, establishes a secure channel through key authentication (such as SSH-RSA or Ed25519), supporting encrypted file uploads, downloads, and access control, making it suitable for scenarios with strict file operation control requirements. Both protocols support file chunked uploads and resume-from-break mechanisms, suitable for stable transmission of large scanned documents.

[0033] Specifically, HTTPS transmission requires the server to be configured with a valid X.509 digital certificate and support the HSTS (HTTPStrict Transport Security) policy to ensure mandatory encrypted connections. SFTP transmission requires configuring the SSH service port (default 22) and enabling strong encryption algorithms (such as AES-256-CBC or ChaCha20-Poly1305), while also setting a maximum file upload speed limit (such as 10MB / s) and a maximum number of concurrent connections (such as 50) to ensure system stability. Regarding file format compatibility, the system supports PDF 1.7 and above, JPG (JPEG) compression quality ≥85%, and PNG (lossless compression) formats to ensure image quality meets OCR recognition requirements.

[0034] S12 uses the BERT model to perform entity recognition and relation extraction on the text after OCR recognition, generating a structured event summary.

[0035] At the technical implementation level, text recognized by OCR often contains a lot of noise and inconsistent formatting, so text cleaning and standardization are necessary first. Then, the preprocessed text is input into a BERT-based entity recognition and relation extraction model. The BERT model, through its bidirectional Transformer architecture, can capture the contextual dependencies of words in the text, thereby achieving accurate recognition of entities (such as customer names, product numbers, time, and location). Relation extraction is achieved by introducing a relation classification layer on top of BERT, utilizing entity location information (such as the start and end positions of entities) and contextual semantics to identify semantic relationships between entities (such as "customer complains about the product" and "approval approval time").

[0036] At the parameter level, the input to the BERT model is typically a fixed-length sequence of tokens (e.g., 512 tokens), using a pre-trained BERT-base or BERT-large model, with different hidden layer dimensions (e.g., 768 or 1024) selected depending on the task requirements. In entity recognition tasks, the model outputs the entity label for each token (e.g., B-PER, I-ORG, etc.), and relation extraction outputs the relationship category between entity pairs. During model training, the loss function is typically cross-entropy loss, the optimizer is AdamW, and the learning rate is set to... to Furthermore, a tiered learning rate strategy is employed to perform differentiated updates for different layers.

[0037] S2 performs text denoising, key field extraction, and structure transformation on unstructured data based on preset rules, and performs format standardization, data validation, and field mapping processing on structured data.

[0038] In some implementations, unstructured data (such as scanned PDFs, images, and free text) is first processed using OCR technology for text recognition, with an accuracy rate of over 98%. The recognized text then undergoes denoising, removing punctuation, stop words, and garbled characters, typically using a combination of regular expression matching and part-of-speech tagging (POS tagging). Key field extraction relies on non-entity recognition (NER) models, such as BERT-based sequence labeling models, to identify core information like customer names, timestamps, and amounts, mapping them to predefined field structures for structured transformation.

[0039] For structured data (such as CSV, JSON, and database records), this step first performs format standardization, unifying field naming rules (e.g., using camelCase or underscore notation), and ensuring that fields such as time and numerical values ​​conform to ISO standards (e.g., ISO 8601 time format). The data validation phase is executed through a rules engine, including field integrity checks and numerical range validation (e.g., monetary fields must meet certain requirements). This includes format consistency verification, etc. Field mapping is achieved through a pre-defined mapping table, which unifies fields from different sources into a standard data model, ensuring data semantic consistency.

[0040] Specifically, the text denoising process requires setting character filtering thresholds (e.g., directly filtering words shorter than 3 characters), key field extraction requires defining entity type label sets (e.g., PER, ORG, DATE, AMOUNT), and structured data validation requires configuring field constraint rules (e.g., required fields, numerical precision, and enumeration value validation). The data quality scoring formula is... The final cleaned dataset must have a quality score of at least 95%.

[0041] Furthermore, S2 includes: S21 executes field mapping through the rules engine, where the field mapping rules include a three-tiered tagging system of customer type, data category, and time period.

[0042] In its specific operation, the rule engine first loads a pre-defined field mapping rule library, which consists of three layers of tags: the first layer is Customer Type, such as "Individual Customer," "Corporate Customer," and "VIP Customer"; the second layer is Data Category, covering "Transaction Records," "Credit Scores," "Complaint Content," and "Product Feedback"; and the third layer is Time Period, such as "Monthly," "Quarterly," and "Annual." Each rule contains metadata such as field name, regular expression matching pattern, tag mapping relationship, and data anonymization strategy. For example, the field "customer_type" can be matched using a regular expression and mapped to the corresponding customer type tag.

[0043] At the parameter level, the rule engine supports dynamic configuration, allowing users to adjust tag weights, matching thresholds, and field priorities according to business needs. For example, the minimum confidence level for field matching can be set to 0.85 to ensure the accuracy of the mapping results. Simultaneously, the rule engine supports field recognition in multilingual environments, is compatible with the UTF-8 encoding standard, and has a field conflict detection mechanism. When multiple rules match the same field, a priority ranking strategy is used for the final mapping decision.

[0044] S3 utilizes a dynamic semantic parsing model to adjust the data parsing strategy based on user needs and contextual information, generating weighted semantic analysis results.

[0045] In some implementations, the dynamic semantic parsing module introduces a context-sensitive semantic understanding mechanism to achieve real-time responses to user needs and historical interaction information, thereby dynamically adjusting the data parsing strategy and generating weighted semantic analysis results. This module is built on a Large Language Model (LLM) and combines an attention mechanism with online fine-tuning strategies to ensure that the analysis results are highly relevant and accurate across different business scenarios.

[0046] This module first receives structured data from a vector knowledge base, along with the user's current query context (e.g., "analyze customer complaint trends") and historical dialogue records. It then uses a pre-trained model like BERT to vectorize the context, obtaining the context embedding vector `context_embedding = BERT.encode(context)`. Subsequently, based on the interaction between the user's query and historical records, it employs an attention mechanism to calculate dynamic weights for each data source: `weights = softmax(attention_score(history, user_query))`, enabling focused processing of key information.

[0047] Specifically, during model inference, a temperature parameter of 0.7 is set to balance the diversity and stability of the generated results. Simultaneously, the model supports an online fine-tuning mechanism, which adjusts settings upon receiving user feedback. In this case, the parameters can be updated in real time by using model.fine_tune(analysis_result, user_feedback), thereby improving the accuracy of subsequent analysis and user satisfaction.

[0048] Furthermore, S3 includes: S31, calculate the similarity weight between historical dialogue records and the current user query based on the attention mechanism.

[0049] This step calculates the attention score between historical dialogues and the current query, and then uses the Softmax function to normalize it into a weight vector in the form of a probability distribution, thereby assigning differentiated attention to different historical information during semantic analysis.

[0050] The specific operation method is as follows: First, input the historical dialogue records and the current user query into the pre-trained BERT model to generate the corresponding context embedding vectors. and Subsequently, an attention mechanism is used to calculate the similarity score between the two, typically employing the dot product or cosine similarity as the attention scoring function. In this invention, the method of calculating the attention score is not limited, but its output should be a score vector with the same length as the historical dialogue record, representing the relevance of each historical segment to the current query.

[0051] Furthermore, the score vector is input into the Softmax function for normalization to obtain the weight vector. The Softmax function ensures that the weight values ​​are within the range of [0, 1] and that the sum is 1. This allows for weighted fusion of historical information during subsequent semantic parsing, improving the model's sensitivity to context and its responsiveness.

[0052] Specifically, the embedding dimension of the attention mechanism is typically 768 (BERT-base) or 1024 (BERT-large), and the output dimension of the Softmax function is consistent with the number of fragments in the historical dialogue records. In practical applications, this step is widely used in the semantic understanding stage of customer survey reports, especially when users submit complex queries with contextual dependencies, such as "How has customer satisfaction changed compared to the previous quarter?" The system can more accurately identify user intent and generate targeted analysis content through weighted information from historical dialogues.

[0053] S32 updates model parameters in real time based on user feedback through an online fine-tuning mechanism. The fine-tuning process includes correcting the analysis results and recalculating the weights.

[0054] In some implementations, the online fine-tuning mechanism of this invention updates model parameters in real time based on user feedback. This is a core optimization technique in the dynamic semantic parsing module, aiming to improve the model's accuracy in understanding user intent and its response speed. By introducing incremental learning and an online parameter update strategy, this mechanism enables the model to adjust local parameters based on real-time user feedback without retraining the entire model, thereby achieving continuous optimization of semantic parsing.

[0055] This fine-tuning process comprises two key sub-steps: revising the analysis results and recalculating the weights. When users provide feedback on the generated analysis results (such as labeling errors, adding information, or adjusting priorities), the system first semantically aligns the feedback with the original analysis results, identifying entities, relationships, or sentiment tendencies that require correction. Subsequently, the system dynamically adjusts the weight allocation of each data source in the model through an attention mechanism to reflect the user's level of attention to different information. For example, if a user emphasizes that "customer complaint trends" should be based on the latest data rather than historical data, the system will increase the weight coefficient of the current period's data and decrease the weight of historical data, thereby optimizing the model output.

[0056] Specifically, the attention score is normalized using the Softmax function during weight adjustment to ensure a reasonable weight distribution. During fine-tuning, the model's learning rate is typically set to... To ensure the stability of parameter updates, an L2 regularization term is introduced to avoid overfitting, with its coefficient... Furthermore, the model uses temperature parameters during fine-tuning. To balance the diversity and accuracy of generated content.

[0057] S4 uses a multimodal data fusion algorithm to perform anomaly detection, deduplication, and missing value filling on the cleaned structured data and semantic analysis results, constructing a unified data pool with a quality score of no less than 95%.

[0058] In some implementations, this fusion algorithm employs a hybrid approach combining statistics and rules. For anomaly detection, the system uses the Z-score method to identify outliers in numerical fields; the calculation formula is as follows: ,in For the current data point, For the average value of the field, is the standard deviation. When an outlier is detected, it is identified and marked or removed. For categorical fields, a frequency distribution-based outlier detection strategy is used, combined with TF-IDF vectorization and cosine similarity calculation, to identify semantically anomalous text.

[0059] For deduplication of duplicate data, the system employs a dual judgment mechanism based on hashing and semantic similarity. Structured data is quickly compared using field combination hashing (such as MD5 or SHA-256), while unstructured text uses pre-trained models such as BERT to generate semantic vectors and calculate vector similarity. When the similarity exceeds a preset threshold When the data is duplicated, it is determined to be a unique instance.

[0060] Missing value imputation employs a context-based strategy. For numerical fields, linear regression or KNN imputation methods are used; for text fields, semantic completion is performed based on the context generation capabilities of a large model. The imputed dataset must satisfy the quality scoring formula. ,in The number of valid data entries. The total number of data entries is used to ensure that the final data pool quality score is not lower than 95%.

[0061] Furthermore, S4 includes: S41 uses the Z-score anomaly detection method, the formula is: The data is standardized, with an anomaly detection threshold of [value missing]. .

[0062] Specifically, in the multi-source heterogeneous data intelligent acquisition module of this invention, the Z-score anomaly detection method is used to standardize the data. This technique is based on statistical principles and is used to identify and remove outliers in the dataset, thereby improving data quality and the reliability of subsequent analysis. The Z-score calculation formula is:

[0063] in, Indicates the current data point. The mean of the dataset. The standard deviation is used as the unit. This method transforms the original data into standardized values ​​centered at the mean and with the standard deviation as the unit, enabling data with different dimensions and distributions to be compared and integrated on a unified scale.

[0064] In its implementation, the system first performs statistical calculations on the numeric fields from structured and unstructured data sources to obtain their mean. with standard deviation Then, for each data point Calculate its Z-score and set the anomaly detection threshold as follows. This threshold is based on the 3σ principle of the normal distribution, meaning that under a normal distribution, approximately 99.7% of the data points lie within the normal distribution. Data points outside the specified range can be considered abnormal.

[0065] Furthermore, this step works in conjunction with the large model data fusion module during the data cleaning phase to ensure that the data in the input vector knowledge base possesses high-quality features. The system marks, removes, or corrects outlier data, thereby improving the overall quality score of the dataset. The goal is .

[0066] S42, through the quality scoring formula Quantitatively evaluate the effectiveness of data cleaning.

[0067] In some implementations, this step is embedded within the large model data cleaning module as a quality assessment step after data fusion and standardization. The cleaning process includes data format standardization and outlier detection (such as Z-score-based detection methods). ,in Operations include setting an outlier threshold, deduplicating duplicate data, and filling in missing values. After cleaning, the system automatically performs statistical analysis. and And substitute it into the quality scoring formula for calculation. If the score If the score is high enough, the dataset passes the quality review and proceeds to the subsequent data classification and vector knowledge base construction stage; if the score is lower than the threshold, a manual review mechanism is triggered, and reviewers intervene to handle high-risk items.

[0068] Specifically, The statistics must comply with bank compliance rules and data integrity standards, such as field non-emptiness, valid numerical range, and consistent text format. This represents the total number of records in the original dataset, typically imported into the system after being crawled by RPA or recognized by OCR. Quality Score The settings must meet the minimum data quality requirements of the business system, generally not less than 95%, to ensure the reliability of subsequent analysis results.

[0069] This invention discloses a customer survey report generation method based on artificial intelligence and RPA, which can realize intelligent collection and fusion of multi-source heterogeneous data. Through dynamic semantic parsing and user feedback-driven online fine-tuning mechanism, it can significantly improve the accuracy and scenario adaptability of customer survey reports, and support adaptive generation of personalized report content, effectively improving the overall automation level and processing efficiency.

[0070] Example 2 The following describes in detail, with reference to the accompanying drawings, a method for generating customer survey reports based on artificial intelligence and RPA according to an embodiment of the present invention.

[0071] The customer survey report generation method proposed in this invention mainly includes four core steps: intelligent collection of multi-source heterogeneous data, dynamic semantic parsing, adaptive report generation, and end-to-end automated process. The following is a detailed explanation of each step: 1. Intelligent acquisition module for multi-source heterogeneous data.

[0072] Technical Implementation: Integrating Application Programming Interface (API), database query engine, and natural language processing algorithms, it automatically captures structured data while utilizing Optical Character Recognition (OCR) technology and large-scale models to process unstructured text and image information, forming a unified data pool.

[0073] like Figure 2 As shown in Table 1, the data acquisition module receives data requests from different sources. After preprocessing (such as format conversion and noise filtering) and structuring (extracting key information), the data is stored in the data pool for subsequent analysis.

[0074] Table 1

[0075] 2. Dynamic semantic parsing module.

[0076] Technical Implementation: By leveraging a large language model, context-sensitive semantic understanding and dynamic adjustment are achieved, ensuring that the analysis results are optimized in real time as the scenario changes.

[0077] Process Description: Input data enters the dynamic context analysis unit, and the model parameters are adjusted through a real-time feedback mechanism to ensure that the output results closely match the current analysis objective.

[0078] Input data: Vector knowledge base retrieval results; User query context (e.g., "Analyze customer complaint trends"); Historical dialogue records.

[0079] The processing flow (pseudocode) is as follows: def dynamic_semantic_analysis(context, user_query, history): # 1. Context-Aware Embedding context_embedding = BERT.encode(context) # Context vectorization # 2. Dynamic weight adjustment weights = softmax(attention_score(history, user_query)) # Based on attention mechanism # 3. Semantic Analysis analysis_result = LLM.predict( prompt=f"Context-based analysis:{user_query}", weights=weights, temperature=0.7 ) # 4. Real-time feedback optimization if user_feedback:# User feedback model.fine_tune(analysis_result, user_feedback) # Online fine-tuning return analysis_result Output data: Optimized text.

[0080] Dynamic attention mechanism: dynamically adjusts the weights of different data sources based on user queries.

[0081] 3. Adaptive report generation module.

[0082] Technical Implementation: Design a rule-based reporting framework that combines machine learning recommendation algorithms to automatically adjust report styles, content depth, and visualization formats based on user preferences and business needs.

[0083] Recommendation Algorithm Framework: This unit employs a hybrid recommendation strategy that combines collaborative filtering with content feature fusion, achieving personalized content recommendation through the following three stages: User profile construction: Generate user preference vectors based on historical behavior data; Content feature extraction: Extract feature vectors of recommendable content from the report template library; Matching degree calculation: Predict user-content matching score through matrix factorization model.

[0084] Core computing process: 1. User Preference Vectorization Input: User's historical operation log Processing procedure: Feature Engineering: Building a User-Content Interaction Matrix Output: User latent factor vector 2. Content Feature Coding Input: Report template metadata (section type, data granularity) Processing procedure: TF-IDF text vectorization (for template description text) Structural feature standardization Output: Content feature vector 3. Matching degree prediction model Biased Matrix Decomposition (BiasSVD) is employed: Loss function optimization: Update parameters iteratively using stochastic gradient descent (SGD): Update bias terms b_u += lr * (err - λ * b_u) b_i += lr * (err - λ * b_i) Update the hidden factor vector u_vec += lr * (err * v_vec - λ * u_vec) v_vec += lr * (err * u_vec - λ * v_vec) 3. Output: Calculate the predicted score for the candidate content set.

[0085] Output a Top-N recommendation list sorted by rating (e.g., N=5). Work process as follows Figure 3 As shown in Table 2, the user interface collects customization requests and feeds them into the report generation engine. The engine then generates a personalized report draft based on built-in rules and a learning model. After user confirmation and fine-tuning, the draft is officially released.

[0086] Table 2

[0087] 4. End-to-end automated processes.

[0088] Technology Integration: The three core modules above are seamlessly connected through message queues and service orchestration technologies to form a fully automated process from front-end data collection to back-end report output.

[0089] System architecture such as Figure 4 As shown, the front-end data acquisition component is responsible for data capture, the middleware handles logical operations and data fusion, and the back-end generation module is responsible for report synthesis and presentation. The entire process follows the principles of microservice architecture to ensure high availability and scalability. Figure 4The specific relationships between the various levels of the system are shown in Table 3.

[0090] Table 3

[0091] Example 3 To achieve the above embodiments, such as Figure 5 As shown, this embodiment also provides a customer survey report generation device 10 based on artificial intelligence and RPA. The device 10 includes a data synchronization acquisition module 100, an unstructured data processing module 200, a dynamic semantic parsing module 300, and a multimodal data fusion module 400.

[0092] The data synchronization acquisition module 100 is used to synchronously acquire customer survey data from structured and unstructured data sources through application programming interfaces, database query engines, optical character recognition technology and robotic process automation technology. The unstructured data processing module 200 is used to perform text denoising, key field extraction and structure transformation on unstructured data based on preset rules, as well as to perform format standardization, data validation and field mapping processing on structured data. The dynamic semantic parsing module 300 is used to adjust the data parsing strategy according to user needs and context information using the dynamic semantic parsing model, and generate weighted semantic analysis results. The multimodal data fusion module 400 is used to perform anomaly detection, deduplication, and missing value filling on the cleaned structured data and semantic analysis results through multimodal data fusion algorithms, and to build a unified data pool with a quality score of no less than 95%.

[0093] Furthermore, the aforementioned data synchronization acquisition module 100 is also used for: The secure transmission submodule is used to securely transmit scanned PDF / JPG / PNG files manually uploaded by users via HTTPS and SFTP protocols. The entity recognition submodule is used to perform entity recognition and relation extraction on the text after OCR recognition using the BERT model, and generate a structured event summary.

[0094] Furthermore, the aforementioned unstructured data processing module 200 is also used for: Field mapping is performed through a rules engine, where field mapping rules include a three-tiered tagging system of customer type, data category, and time period.

[0095] Furthermore, the aforementioned dynamic semantic parsing module 300 is also used for: The similarity weight between historical dialogue records and the current user query is calculated based on the attention mechanism. The model parameters are updated in real time based on user feedback through an online fine-tuning mechanism. The fine-tuning process includes correcting the analysis results and recalculating the weights.

[0096] Furthermore, the aforementioned multimodal data fusion module 400 is also used for: The Z-score anomaly detection method is used, and the formula is: The data is standardized, with an anomaly detection threshold of [value missing]. ; Through the quality scoring formula Quantitatively evaluate the effectiveness of data cleaning.

[0097] This invention discloses a customer survey report generation device based on artificial intelligence and RPA, which can realize intelligent collection and fusion of multi-source heterogeneous data. Through dynamic semantic parsing and user feedback-driven online fine-tuning mechanism, it can significantly improve the accuracy and scenario adaptability of customer survey reports, and support adaptive generation of personalized report content, effectively improving the overall automation level and processing efficiency.

[0098] To implement the methods of the above embodiments, the present invention also provides a computer device, such as... Figure 6 As shown, the computer device 600 includes a memory 601 and a processor 602; wherein, the processor 602 reads executable program code stored in the memory 601 to run a program corresponding to the executable program code, so as to implement the various steps of the method described above.

[0099] To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in the foregoing embodiments.

[0100] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0101] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for generating customer survey reports based on artificial intelligence and RPA, characterized in that, include: S1 synchronously acquires customer survey data from structured and unstructured data sources through application programming interfaces, database query engines, optical character recognition technology, and robotic process automation technology. S2 performs text denoising, key field extraction, and structure transformation on unstructured data based on preset rules, and performs format standardization, data validation, and field mapping processing on structured data. S3 utilizes a dynamic semantic parsing model to adjust the data parsing strategy based on user needs and contextual information, generating weighted semantic analysis results. S4 uses a multimodal data fusion algorithm to perform anomaly detection, deduplication, and missing value filling on the cleaned structured data and semantic analysis results, constructing a unified data pool with a quality score of no less than 95%.

2. The method as described in claim 1, characterized in that, S1 includes: S11 enables secure transmission of scanned PDF / JPG / PNG files manually uploaded by users via HTTPS and SFTP protocols; S12 uses the BERT model to perform entity recognition and relation extraction on the text after OCR recognition, generating a structured event summary.

3. The method as described in claim 1, characterized in that, The S2 includes: S21 executes field mapping through the rules engine, where the field mapping rules include a three-tiered tagging system of customer type, data category, and time period.

4. The method as described in claim 1, characterized in that, The S3 further includes: S31, calculate the similarity weight between historical dialogue records and the current user query based on the attention mechanism; S32 updates model parameters in real time based on user feedback through an online fine-tuning mechanism. The fine-tuning process includes correcting the analysis results and recalculating the weights.

5. The method as described in claim 1, characterized in that, The S4 includes: S41 uses the Z-score anomaly detection method, the formula is: The data is standardized, with an anomaly detection threshold of [value missing]. ; S42, through the quality scoring formula Quantitatively evaluate the effectiveness of data cleaning.

6. A customer survey report generation device based on artificial intelligence and RPA, characterized in that, include: The data synchronization acquisition module is used to synchronously acquire customer survey data from structured and unstructured data sources through application programming interfaces, database query engines, optical character recognition technology, and robotic process automation technology. The unstructured data processing module is used to perform text denoising, key field extraction, and structured transformation on unstructured data based on preset rules, as well as to perform format standardization, data validation, and field mapping processing on structured data. The dynamic semantic parsing module is used to adjust the data parsing strategy according to user needs and context information using a dynamic semantic parsing model, and generate weighted semantic analysis results. The multimodal data fusion module is used to perform anomaly detection, deduplication, and missing value filling on cleaned structured data and semantic analysis results through multimodal data fusion algorithms, and to build a unified data pool with a quality score of no less than 95%.

7. The apparatus as claimed in claim 6, characterized in that, The data synchronization acquisition module is also used for: The secure transmission submodule is used to securely transmit scanned PDF / JPG / PNG files manually uploaded by users via HTTPS and SFTP protocols. The entity recognition submodule is used to perform entity recognition and relation extraction on the text after OCR recognition using the BERT model, and generate a structured event summary.

8. The apparatus as claimed in claim 6, characterized in that, The unstructured data processing module is also used for: Field mapping is performed through a rules engine, where field mapping rules include a three-tiered tagging system of customer type, data category, and time period.

9. A computer device, characterized in that, Including processor and memory; The processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the customer survey report generation method based on artificial intelligence and RPA as described in any one of claims 1-5.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the customer survey report generation method based on artificial intelligence and RPA as described in any one of claims 1-5.

Images