Method for analyzing customer feedback using artificial intelligence and electronic device for performing the same

Abstract

The present invention relates to data processing technology using artificial intelligence technology, and specifically, to a method and electronic device for analyzing unstructured customer feedback data collected from various channels and automatically deriving customer sentiment and potential requirements by utilizing algorithms based on natural language processing (NLP) and deep learning.

Description

Technology Field

[0001] The present invention relates to data processing technology using artificial intelligence technology, and specifically, to a method and electronic device for analyzing unstructured customer feedback data collected from various channels and automatically deriving customer sentiment and potential requirements by utilizing algorithms based on natural language processing (NLP) and deep learning. Background Technology

[0002] Analyzing customer feedback is essential for establishing corporate marketing and product improvement strategies. Traditional methods of analyzing customer feedback have primarily relied on predefined keyword matching or statistical data from surveys.

[0003] However, as the volume of unstructured data from social network services (SNS), online communities, and customer center voice recordings has increased rapidly in recent years, there are limitations to manually reviewing or analyzing it solely through simple keyword searches.

[0004] In particular, the simple keyword method suffers from low analysis accuracy because it fails to properly grasp sentence context, irony, and neologisms, and there is a problem in that it is difficult to identify hidden complaints or requests for improvement that customers have not directly mentioned.

[0005] Furthermore, processing over thousands of data points in real-time to match the speed of corporate response requires massive resources in terms of manpower and cost. Prior art literature

[0006] Korean Patent Publication 10-2023-0125396 A (2023.08.29) Korean Patent Publication 10-2011-0062063 A (2011.06.10) The problem to be solved

[0007] The technical problem that the present invention aims to solve is to provide an analysis method that improves the precision of feedback classification and sentiment analysis by applying Natural Language Understanding (NLU) technology to a large amount of unstructured customer feedback data to identify the accurate meaning according to context.

[0008] In addition, we aim to provide an electronic device that can automatically extract key elements related to specific product functions or service quality within feedback through a deep learning model based on an attention mechanism, and integrate and analyze multimodal data including voice and images in addition to text to understand the customer's intent from various angles.

[0009] In addition, the purpose is to maximize a company's risk response capabilities by having an artificial intelligence model learn the time-series changes of feedback to predict the potential for the spread of specific issues in advance and generate warnings in real time. means of solving the problem

[0010] One embodiment of the present invention relates to an electronic device for analyzing customer feedback data collected from multiple channels, comprising: a memory for storing unstructured customer feedback data and an artificial intelligence analysis algorithm; and a processor connected to the memory for executing the artificial intelligence analysis algorithm, wherein the processor comprises: a data acquisition step for collecting raw feedback data in the form of text or voice from various input paths and, if the raw feedback data is voice, converting it into text through an artificial intelligence-based speech-to-text engine; a preprocessing step for removing noise within the text data through an artificial intelligence-based language refinement model and correcting typos and non-standard words by referring to a domain-specific dictionary; and an aspect-based sentiment analysis step for extracting a plurality of product aspects from the preprocessed text data and determining the sentiment polarity for each aspect using a pre-trained deep neural network of a transformer structure. The present invention proposes an electronic device for analyzing customer feedback using artificial intelligence, characterized by being configured to perform an output step of calculating the risk regarding product quality based on the above sentiment analysis results and generating and outputting a real-time warning report to an administrator terminal when a preset threshold is exceeded. Effects of the invention

[0011] According to the present invention, by utilizing an advanced language model based on artificial intelligence, a vast amount of customer feedback can be automatically classified and analyzed in real time, thereby significantly reducing the time and cost required for data processing.

[0012] Furthermore, by deriving segmented sentiment scores that consider context beyond simple positive-negative classification, it is possible to precisely identify trends in customer satisfaction, providing objective and highly reliable insights for product quality improvement and the formulation of customized marketing strategies.

[0013] In particular, artificial intelligence can learn the correlations between customer feedback data on its own to proactively suggest potential problems or new business opportunities that humans may not have discovered, and it has an excellent effect in effectively preventing a decline in a company's brand value by preemptively detecting rapid changes in public opinion or an increase in claims. Brief explanation of the drawing

[0014] FIG. 1 is a conceptual diagram showing the overall system configuration of the invention for analyzing customer feedback using artificial intelligence according to one embodiment of the present invention. FIG. 2 is a block diagram showing detailed components of an electronic device for analyzing customer feedback according to one embodiment of the present invention. FIG. 3 is a diagram showing the internal architecture of an artificial intelligence analysis model and the deep learning-based data processing operation flow applied to one embodiment of the present invention. FIG. 4 is a flowchart sequentially illustrating the process of an electronic device according to an embodiment of the present invention processing customer feedback data and deriving analysis results based on artificial intelligence. Specific details for implementing the invention

[0015] Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the attached drawings. Terms and words used in this specification and claims should not be interpreted as being limited to their ordinary or dictionary meanings. Instead, based on the principle that the inventor can appropriately define the concepts of terms to best describe their invention, they should be interpreted in a meaning and concept consistent with the technical spirit of the present invention. Furthermore, the embodiments described in this specification and the configurations illustrated in the drawings are merely one preferred embodiment of the present invention and do not represent all aspects of the technical spirit of the present invention; therefore, it should be understood that various equivalents and modifications capable of replacing them may exist at the time of filing this application. Additionally, in the description with reference to the drawings, identical or corresponding components are assigned the same reference numerals, and redundant descriptions thereof are omitted. Moreover, if it is determined that a detailed description of related known technology may unnecessarily obscure the essence of the present invention, such detailed description is omitted.

[0016] Referring to FIG. 1, a conceptual diagram of the overall system configuration and data pipeline of the customer feedback analysis invention using artificial intelligence according to an embodiment of the present invention is illustrated. The system according to the present invention goes beyond simply collecting text data and establishes a multi-channel data collection environment that comprehensively collects vast amounts of unstructured and structured data accumulated in the company's internal Customer Relationship Management (CRM) server, including real-time voice recordings generated from a Computer Telephony Integration (CTI) system linked to a company's customer center (CS center), customer support bulletin boards on the company's official website, after-sales service (A / S) request records from offline stores and service centers, official email inquiry records, and data from the company's internal Customer Relationship Management (CRM) server. The collected heterogeneous raw data is temporarily merged into a data lake based on a distributed processing environment via an Application Programming Interface (API) gateway and a load balancer. As illustrated in FIG. 1, the merged data is streamed in real-time via a wired or wireless communication network to an AI-based electronic device or a cloud computing-based AI server, which is the core of the present invention, or transmitted according to a preset batch cycle. The electronic device does not stop at simply storing incoming data; instead, it structures it into a single analysis package by mapping metadata—such as the time customer feedback occurred, the customer's past product purchase history, previous consultation history, and the number of complaints received—to the original feedback text. Based on this advanced data pipeline, the electronic device activates a self-built deep neural network to multidimensionally evaluate the customer's emotional state, the importance of requirements, and the potential risk of developing into future claims hidden within the text and voice data. Finally, it operates to visualize and provide the inference results to the administrator terminal in the form of a dynamic dashboard and real-time alerts.

[0017] Referring to FIG. 2, a block diagram illustrating the detailed hardware and software architecture of an electronic device for analyzing large-scale customer feedback in real time according to an embodiment of the present invention is shown. The electronic device is basically configured to include a communication module, memory, a user interface, and a processor that controls the overall operation of said components and the computation of an artificial intelligence model. The communication module supports various network protocols such as TCP / IP, HTTP, and MQTT, and forms an encrypted communication channel with an external CRM database, a recording server of a customer center, and other third-party data provision servers to transmit and receive customer data while maintaining security. The memory includes both a non-volatile storage medium that guarantees high-speed read and write performance and a volatile memory for computational assistance, and systematically stores received raw feedback data and refined data that has undergone a preprocessing process, as well as a training dataset containing millions to billions of parameters, weight matrices, bias values, and labels for model training that are essential for operating a deep learning-based artificial intelligence model. In addition, a high-dimensional vector database that stores values ​​converted from natural language into mathematical vectors is logically constructed within memory to maximize the speed of similarity search and clustering operations. To overcome the limitations of a general central processing unit (CPU), the processor, which plays the most important role in the present invention, configures artificial intelligence hardware accelerators such as a graphics processing unit (GPU), a neural network processing unit (NPU), or a tensor processing unit (TPU) specialized for large-scale matrix operations and tensor operations in parallel as a single-chip system (SoC) or as separate acceleration cards.The processor loads and executes artificial intelligence algorithms stored in memory to perform morphological analysis, syntactic analysis, and Named Entity Recognition (NER) on input feedback data, and processes embedding operations at ultra-high speed to map the context of words or sentences to coordinates in vector space. The user interface converts the processor's inference results into visual graphs, heatmaps, or word clouds through a rendering engine and outputs them to a display screen, and provides an interactive function that provides the original data and the basis for the AI ​​analysis in a pop-up form when an administrator clicks on a specific feedback cluster.

[0018] Referring to FIG. 3, the internal architecture and data processing operation flow of an artificial intelligence analysis model, which is driven by a processor of an electronic device to deeply understand and analyze customer natural language feedback, are illustrated in detail. The artificial intelligence analysis model applied in the present invention deviates from existing methods that have limitations in the sequential processing of sentences and adopts a bidirectional encoder structure of the Transformer series pre-trained with a large-scale corpus. When customer feedback in the form of raw text is input into the model, a subword-based tokenizer operates to segment the text into optimal token units to address the problem of out-of-vocabulary (OOV). The segmented tokens pass through an input embedding layer and are each converted into high-dimensional continuous real vectors; at this time, a positional encoding value that mathematically encodes the absolute and relative positional information of the words is added to the vector, thereby preserving the order and contextual information of the words within the sentence. Subsequently, the input tensor sequentially passes through deep neural network blocks composed of multiple Multi-Head Attention layers and Feed-Forward Neural Networks. The attention mechanism is a high-level computational process that calculates the association between all word pairs within a sentence into a probability score through the dot product of the Query, Key, and Value matrices. For example, when feedback such as "The screen is clear, but the speaker sound quality is too distorted and uncomfortable" is input, the Multi-Head Attention layer calculates a high weight between the object 'screen' and the expression 'clear (positive),' while independently calculating weights between the object 'speaker sound quality' and the expressions 'distorted (negative)' and 'uncomfortable (negative).'The feature vectors extracted in this way are passed to an Aspect-Based Sentiment Analysis (ABSA) classifier layer, and through a Softmax or Sigmoid activation function, sentiment probability values ​​for each detailed attribute of the product (screen, sound quality, battery, etc.) are finally derived as numerical values ​​between 0 and 1. In addition, the artificial intelligence model includes a time series analysis module based on a Long Short-Term Memory (LSTM) network in parallel to analyze the trend of changes in feedback regarding a specific product accumulated over the past few months, thereby performing the function of mathematically determining whether the current increase rate of negative feedback is an anomaly that exceeds the statistical tolerance.

[0019] Referring to FIG. 4, a flowchart illustrates a series of methods for an electronic device according to an embodiment of the present invention to process complex customer feedback data step-by-step based on artificial intelligence and derive valuable business insights. First, a multi-channel data collection step is performed, in which the electronic device periodically crawls or receives data via a push method from sources such as call recording files or email inquiry records from a customer center. If the input data is a voice file, a speech-to-text (STT) engine, which combines a deep learning-based acoustic model and a language model installed in the electronic device, is activated to extract the speaker's voice features and convert them into a text sequence. Next, a preprocessing step is performed, and the collected or converted text data undergoes a noise removal algorithm to remove unnecessary special characters, HTML tags, etc. Furthermore, by referring to an AI dictionary that is continuously updated to match the company's specialized domain, slang, abbreviations, and typos used by customers are corrected into standard language or defined technical terms, and stop words are filtered out to maximize the efficiency of analysis.

[0020] Subsequently, the full-scale AI deep analysis phase begins. The refined data, having completed preprocessing, is input in batches into the pre-trained AI Transformer model described in detail in Figure 3. The processor's hardware accelerator determines the overall polarity (positive, negative, or neutral) of the feedback sentences through large-scale tensor operations. Simultaneously, through an attribute extraction model, it recognizes target attributes included in the sentences, such as product parts or service types, as named entities and independently maps and calculates detailed sentiment scores for each target attribute. In parallel, to detect new types of customer complaints for which no labels exist, unsupervised learning-based topic modeling and density-based clustering algorithms (such as DBSCAN) are executed. Through this process, the AI ​​groups semantically similar feedback sentences into a single cluster in a high-dimensional vector space, even if the keywords are not predefined, and automatically generates new issue categories by extracting key keywords representing those clusters. Finally, in the prediction and visualization stage, based on the derived sentiment scores and topic cluster data, artificial intelligence runs a time-series regression model to predict the future frequency of specific complaint issues and converts the impact of this on the company's after-sales service costs or brand credibility into a quantitative figure. As soon as the converted risk figure is detected to exceed a preset threshold on the electronic device, the device generates a comprehensive warning report containing a summary of the core cause of the complaint, links to relevant source data, and the estimated scale of damage, and transmits this to the administrator's terminal in the form of a real-time pop-up notification or email. This fully supports the company in taking preemptive defensive measures, such as correcting product defects or distributing customer response guidelines, before reaching critical quality issues or large-scale claim situations.

[0021] The processor performs a data acquisition step in which it collects raw feedback data in the form of text or voice from various input channels and, if the raw feedback data is voice, converts it into text using an AI-based speech-to-text engine. It also performs a preprocessing step in which it removes noise from the text data using an AI-based language refinement model and corrects typos and non-standard words by referring to a domain-specific dictionary. Furthermore, it performs an aspect-based sentiment analysis step in which it extracts multiple product aspects from the preprocessed text data and determines the sentiment polarity for each aspect using a pre-trained deep neural network with a transformer structure. Additionally, it is configured to perform an output step in which it calculates the risk level regarding product quality based on the results of the sentiment analysis and generates and outputs a real-time warning report to an administrator terminal if a preset threshold is exceeded.

[0022] Looking at the specific control logic of each step performed by the aforementioned processor, the data acquisition step collects unstructured data via real-time streaming from internal CRM servers, external online customer review platforms, and voice recording databases of customer centers. When voice data is received, the processor activates a high-precision STT engine combining a deep neural network-based acoustic model and a language model to convert the voice waveform into a digital signal and generate a text sequence while preserving the speaker's intonation and pause information. In the preprocessing step, the processor utilizes a BERT (Bidirectional Encoder Representations from Transformers) family language model that understands context to detect semantic errors within sentences, going beyond simple typo correction. Additionally, it loads a domain dictionary containing company-specific technical terms and product proper nouns from memory; by replacing technical keywords that general-purpose language models might misinterpret with accurate standard terms, it ensures the integrity of the data that forms the basis of the analysis.

[0023] In the attribute-based sentiment analysis step described above, the processor utilizes an advanced self-attention mechanism within the Transformer architecture. The processor calculates the association probability between specific attribute words, such as 'design,' 'battery,' 'price,' and 'delivery,' and adjective expressions modifying those attributes within the input sentence. Through this process, independent sentiment scores are derived for each attribute, even when a single sentence contains multiple attributes. Furthermore, rather than simply summing the sentiment polarity values ​​derived for each attribute, the processor assigns weights by considering the influence of each attribute on overall product quality and customer satisfaction. Additionally, the processor performs time-series feature extraction by analyzing the temporal relationships of the collected data to determine whether negative feedback regarding the same attribute is concentrated at a specific point in time or increases gradually over the long term.

[0024] In the process where the above processor quantitatively calculates the risk level regarding product quality, an advanced risk index calculation model is applied. The quality risk level (q) calculated according to one embodiment of the present invention is determined by the following mathematical formula.

[0025]

[0026] The definition and determination method of each parameter in the above mathematical formula is as follows.

[0027] First, 'a' represents the frequency of negative feedback regarding the same attribute occurring within a specific time window. The value of 'a' is determined by counting the number of feedbacks classified as the same issue by the processor through an unsupervised learning-based clustering algorithm.

[0028] Second, b represents the severity of the negative sentiment identified by the artificial intelligence model. The value of b is determined in proportion to the value assigned to the negative class among the softmax probability values ​​calculated in the output layer of the attribute-based sentiment analysis model. That is, if strong anger or profanity is included rather than simple dissatisfaction, the value of b is set high.

[0029] Third, c represents the importance weight (Customer Influence) of the customer who raised the feedback. The value of c is determined by referencing customer metadata stored in memory and is variably assigned based on the customer's total past purchases, VIP level, or social influence index.

[0030] Fourth, d represents the Time Decay Factor based on the elapsed time from the point of feedback occurrence to the present. The value of d is designed to increase over time in the form of a logarithmic or exponential function to maximize the influence of the latest feedback. Therefore, older feedback results in a larger value of d, which has the effect of lowering the overall risk q.

[0031] A specific example applying the above mathematical formula is as follows. Assume that 100 negative feedbacks regarding the 'battery overheating' attribute of a specific smartphone model were collected over the past hour (a=100). The average negative intensity of each sentence analyzed by the AI ​​model was calculated to be 0.8 (b=0.8), and the importance weight was set to 1.5 as the customer base that wrote the feedback was identified as primarily corporate customers making bulk purchases (c=1.5). Furthermore, if the feedback occurred very recently within 10 minutes and maintains a low time decay factor of 1.2 (d=1.2), the risk level q calculated by the processor becomes (100 * 0.8 * 1.5) / 1.2 = 100. If the warning threshold set in the system is 80, the processor determines the current situation as a serious quality risk and generates an immediate warning report.

[0032] In the output stage described above, as soon as the calculated risk level q exceeds a threshold, the processor extracts key sentences that are the root cause of the risk through an artificial intelligence summary engine. It also transmits a report to the administrator terminal that includes the extracted summary information along with a graph showing the real-time trend of changes in related attributes. Furthermore, the processor utilizes feedback received from the administrator after the report is transmitted (e.g., 'true positive', 'false positive', 'importance correction', etc.) as training data. Additionally, through this active learning process, the system continuously enhances the analysis accuracy of the system by fine-tuning the parameter determination logic of the mathematical formula in real time. Moreover, the processor is controlled to perform a predictive maintenance function by simulating various quality issue scenarios that may occur in a virtual space, and predicting the expected number of claims and corporate reputation degradation index within 24 hours if the current risk trend persists. Through this series of processes, the present invention operates as a preemptive risk management solution for enterprises, going beyond a simple data analysis device.

[0033] delete

[0034] The processor calculates the contextual association between a specific word and other words within a sentence using probability weights through a Multi-Head Attention mechanism within the Transformer structure, and based on this, maps and extracts the specific functional elements of a product or service components referred to by the customer's sentiment expression. The processor analyzes from multiple angles using parallel attention heads to determine which noun elements within the sentence, such as 'response speed,' 'UI structure,' or 'network module,' have the highest numerical correlation with adjective expressions like 'slow,' 'inconvenient,' or 'interrupted' within the input customer feedback text. The processor extracts key tokens that exceed a preset valid threshold among the attention scores for each token calculated in the multi-head attention layer, and performs a dynamic ontology anchoring process to identify whether the cause of a customer's complaint is a physical hardware defect or a software logic error by calculating the vector similarity between the key tokens and the technical nodes within the Product Specification Ontology stored in the memory. In the dynamic ontology anchoring process, the processor converts the extracted key tokens into coordinates in a high-dimensional embedding space and performs identification of technical causes beyond simple keyword matching by comparing them with nodes on an ontology graph in which the hierarchical structure of the product is defined. Additionally, to quantitatively evaluate the technical validity of the identified cause of complaint, the processor calculates the ontology anchoring strength (oas) according to the following mathematical formula 1.

[0035]

[0036] The definition and determination method of each parameter in the above mathematical formula 1 are as follows. First, attn represents the average attention weight assigned to the core token. The value of attn is determined by normalizing the sum of attention scores calculated from all heads within the multi-head attention structure by dividing it by the number of heads. Second, sim represents the vector similarity between the core token and the ontology description node. The value of sim is determined as a real number between 0 and 1 by calculating the cosine similarity between the input data's embedding vector and the description node vector predefined in memory. Third, conf represents the classification confidence derived from the sentiment analysis stage. The value of conf is determined by the probability value of the corresponding polarity class calculated by the Softmax function, which is the final output layer of the artificial intelligence model. Fourth, dist represents the semantic distance deviation calculated during the rumination process. The value of dist is determined by calculating the Euclidean distance between the state tensors before and after the attention weights are inverted during the rumination process, which will be described later.

[0037] Let us look at a specific example applying the above mathematical formula 1. Assume a case where feedback such as "I can't use it because the screen keeps flickering" is input. The processor calculates the average attention weight (attn) of the core token "because it flickers" as 0.88 and calculates the vector similarity (sim) with the "display panel controller" node in the product ontology as 0.94. Additionally, if the confidence (conf) that the sentence is classified as negative polarity is calculated as 0.85 and the distance (dist) between state tensors calculated through the rumination process is derived as 0.22, the processor calculates oas = (0.88 * 0.94 * 0.85) / 0.22 ≈ 3.19 according to the above mathematical formula 1. If the derived oas value exceeds a preset threshold score of 2.5, the processor determines that the cause of the feedback complaint is highly likely to be a hardware defect related to the display and reflects this in the technical report.

[0038] In addition, if the confidence level of the sentiment polarity calculated in the sentiment analysis step is below a threshold, the processor performs a counter-factual inference process for the corresponding feedback sentence with the attention weights reversed. The counter-factual inference process is a high-level verification procedure that intentionally reverses the attention distribution to observe changes in the model's output in order to exclude the possibility that the artificial intelligence model has excessively focused on a specific word or, conversely, missed important context. The processor re-executes the neural network operation with the influence of tokens that were assigned the highest attention weights minimized and the influence of background tokens that were assigned low weights forcibly increased. Furthermore, the processor analyzes the tensor deviation between the re-operation result and the previous operation result to detect and correct irony or double negation expressions hidden within the context. For example, in the sentence ‘It works really well,’ if positive attention is focused on the word ‘well’ and the initial confidence level is low, the processor captures the negative nuance of the entire sentence as a tensor deviation through a rumination process and ultimately corrects it to a strong negative.

[0039] Furthermore, if the confidence level still fails to reach a threshold even after undergoing this rumination process, the processor classifies the feedback as ambiguous data and performs a comparative analysis against a database of similar cases in memory. Reflecting the characteristics of specific product domains where irony or metaphorical expressions occur frequently, the processor detects in real-time neologisms or modified forms of complaint expressions that the AI ​​model has not encountered during the learning process. Additionally, based on the corrected analysis results, the processor recalculates the technical urgency of customer feedback and dynamically adjusts the priority of reports to be sent to the administrator terminal according to the severity of identified hardware defects or software errors. Moreover, the processor searches for potential defect nodes likely to occur sequentially on the ontology to generate predictive data regarding whether the currently collected feedback could spread beyond a simple single-component issue to a structural defect affecting the entire system. Through this process, the processor extracts objective and specific technical issues from customers' subjective linguistic expressions and demonstrates an excellent effect in maximizing the reliability of the analysis by independently verifying and supplementing the AI's reasoning process.

[0040] delete

[0041] The processor runs multiple deep neural network models with different hyperparameters in parallel to calculate a consistency index for the analysis results of each model, and for unstructured data where the consistency is lower than a preset lower limit, it additionally performs an ensemble-based prediction correction step by searching for similar cases within the high-dimensional vector database stored in memory and assigning a weighted average value. By simultaneously running multiple neural network architectures with different learning rates, hidden layer depths, and dropout rates for the artificial intelligence models, the processor minimizes bias that a single model may possess and ensures the objectivity of the analysis. Furthermore, the processor calculates a consistency index by calculating the difference in variance or entropy between the probability distributions output by each model, and identifies areas of low model confidence when disagreements between models occur regarding specific data. Instead of forcibly drawing a conclusion for data lacking consistency, the processor extracts the top K cases from a vast vector database in which correct answers have already been verified in the past and accumulated in memory that have characteristics most mathematically similar to the current input value. In addition, the processor determines the final result value by calculating a prediction correction index (pci) according to the following mathematical formula 2 in order to statistically combine the label information of extracted similar cases and the prediction values ​​of current models.

[0042]

[0043] The definition and determination method of each parameter in the above mathematical formula 2 are as follows. First, sim represents the average cosine similarity between the feature vector of the currently input feedback and adjacent vectors retrieved from the high-dimensional vector database. The value of sim ranges from -1 to 1, and a value closer to 1 indicates a greater match between the semantic context of past cases and the current feedback. Second, cnt represents the count of similar cases retrieved within the effective similarity radius. The value of cnt is determined by the density of data points existing within a specific distance threshold set by the processor in the vector space; the denser the data, the more sufficient the grounds for correction. Third, rel represents the historical reliability of the retrieved similar case group. The value of rel is determined as a real number between 0 and 1 based on the proportion of data clusters previously confirmed as correct by the administrator or the model's past prediction accuracy. Fourth, age represents the temporal age of the retrieved cases. The value of age mentioned above is determined by converting the elapsed time from when the case was stored in the database to the present into a logarithmic scale; as the data becomes more recent, the value of the denominator 'age' decreases, which serves to increase the overall pci index.

[0044] Looking at a specific embodiment applying the above mathematical formula 2, we assume a situation where multiple ensemble models produce different sentiment analysis results regarding a specific customer's figurative expression of dissatisfaction, causing the consistency index to fall below the lower limit. The processor immediately searches the vector database to find past cases with a similarity (sim) of 0.92 and discovers 15 valid data points in that area (cnt=15). Furthermore, if the past prediction accuracy (rel) of the data cluster is recorded as very high at 0.95 and the cases were updated approximately 3 months ago, resulting in a time aging (age) value of 1.1, the processor derives pci = (0.92 * 15 * 0.95) / 1.1 ≈ 11.91 according to the above mathematical formula 2. By utilizing the correction index calculated in this way as a weight to readjust the predictions of the ensemble models, the processor makes a sophisticated final judgment based on verified past experience, even for ambiguous sentences.

[0045] Furthermore, the processor receives feedback data from the administrator terminal regarding acceptance or modification of the warning report, and performs an active learning-based model update procedure that inputs the received feedback as an argument to a loss function to fine-tune the output layer weights of the attribute-based sentiment analysis model in real time. The processor converts actions such as the administrator checking the report content and selecting 'accurate' or 'inaccurate,' or directly modifying the classification of a specific attribute, into real-time learning signals. Additionally, the processor sets the administrator's modifications as the ground truth label, calculates the difference from the model's current prediction value as the loss value, and performs gradient descent on the weights of the last layer of the neural network. Moreover, instead of retraining the entire model, the processor freezes the weights of lower layers and updates only the weights of specific layers adjacent to the output layer, thereby dynamically evolving the system to meet the latest corporate requirements with minimal computational resources.

[0046] Furthermore, to prevent catastrophic forgetting of the model that may occur during this active learning process, the processor employs a replay technique that randomly extracts a portion of the past core training dataset and trains it alongside new training data. As manager feedback accumulates, the processor continuously monitors changes in the model's analytical performance regarding specific product families or complaint types. Additionally, if the scale of performance improvement stagnates or the error rate rises in a specific range, the processor performs a self-optimization process by automatically redesigning the hyperparameter configuration and replacing parts of the ensemble model. Moreover, the processor enhances the transparency of the analysis by visualizing the rationale behind "why the AI ​​reached this conclusion" within reports transmitted to the manager's terminal, along with the aforementioned PCI index and list of similar cases. Furthermore, the processor maintains a flexible data processing architecture that immediately accommodates customer linguistic changes from a technical perspective by updating the ontology structure and realigning the corresponding vector space distribution whenever external variables occur, such as internal technical specification changes or new product launches. Through these advanced control procedures, the electronic device functions as a self-evolving quality analysis platform whose precision improves over time.

[0047] delete

[0048] The processor executes an unsupervised learning-based topic modeling algorithm to group new types of customer feedback, which are not predefined, into multiple clusters based on semantic similarity, and automatically generates new complaint issue categories by extracting key keywords representing each cluster. The processor mathematically models the latent thematic structure embedded in customers' linguistic expressions by applying Latent Dirichlet Allocation or modern embedding-based topic modeling techniques to the input large-scale unstructured dataset. Furthermore, the processor monitors in real-time the cosine similarity between the embedding vectors of newly introduced feedback data and the centroids of existing clusters within the high-dimensional vector database stored in memory. If data density is concentrated in the boundary of a specific cluster exceeding a preset value, the processor performs a Dynamic Cluster Branching process to split the specific cluster by detailed attributes or redefine it into new subcategories.

[0049] In determining whether the above-mentioned dynamic cluster differentiation occurs, the processor calculates a differentiation index (bci) according to the following mathematical formula 3 in order to comprehensively consider the structural expansion and cohesion of the cluster, rather than simply calculating the number of data points.

[0050]

[0051] The definition and determination method of each parameter in the above mathematical formula 3 are as follows. First, dens refers to the density of data points per unit volume existing in the outer boundary area of ​​the cluster. The value of dens is determined by the processor defining an area exceeding a specific radius from the cluster center as the outer region and calculating it by dividing the number of new embedding vectors included within that region by the spatial volume. Second, dist refers to the average Euclidean distance between the data points located in the outer region and the cluster center. The value of dist represents the deviation between the geometric center of the outer region and the existing cluster center, and is determined by quantifying how semantically differentiated a new issue is from an existing issue. Third, stab refers to the temporal stability of the cluster center coordinates observed over a certain period. The value of stab is determined by calculating the standard deviation of the displacement distance of the center vector; the more stably the center is fixed, the smaller the value of stab in the denominator becomes, thereby serving to increase the overall differentiation index bci.

[0052] Looking at a specific embodiment applying the above mathematical formula 3, it is assumed that new feedback related to 'high temperature phenomenon during charging' is concentrated in the outer region of an existing cluster called 'heating problem'. The processor calculates the data density (dens) of the outer region as 450 and the average distance (dist) between these data and the existing center point as 0.35. Additionally, if the center point of the cluster has remained very stable for the past 24 hours and the variability index (stab) is derived as 0.05, the processor calculates bci = (450 * 0.35) / 0.05 = 3150 according to the above mathematical formula 3. If the processor confirms that the derived bci value exceeds 2000, which is the differentiation threshold set in the system, it performs a control procedure to separate the outer region into an independent subcategory and assign a new center point.

[0053] In addition, the processor calculates the cross-correlation coefficients between the differentiated multiple clusters to generate an undirected graph representing the association between each issue, and performs a multidimensional correlation inference process to identify pairs of clusters that belong to different categories but show a simultaneous increase in occurrence frequency at the same time, thereby deriving the possibility of a complex failure rather than a single issue. The processor sets each node within the graph as an individual issue category and determines the edge weight between nodes using the Pearson correlation coefficient, which is the similarity of the time-series occurrence pattern. Furthermore, the processor defines pairs of clusters with edge weights above a certain level as 'combined issues' and compares past repair history and design data stored in memory to analyze the causal relationship between them.

[0054] Furthermore, the processor identifies nodes exhibiting high centrality on the undirected graph to derive key failure factors that are the root cause of various quality issues. When a complex defect is detected, the processor does not simply output it as an individual report, but rather merges key keywords from all related clusters to generate a single integrated incident report. Additionally, the processor applies Natural Language Generation (NLG) technology to dynamically generated categories to automatically assign human-readable issue names. Moreover, the processor automatically updates the technical classification system by determining, based on vector similarity, where a newly generated issue category belongs within the existing product specification ontology system.

[0055] Furthermore, the processor resets the final warning priority by reflecting the risk level of complex defects derived through the multidimensional correlation inference process into the aforementioned quality risk calculation logic. Additionally, to optimize the computational load generated during the cluster differentiation and graph generation processes, the processor maintains processing speed by employing Principal Component Analysis (PCA) or t-SNE techniques in parallel to reduce data dimensionality. Moreover, the processor performs a self-optimization function that adjusts differentiation thresholds by receiving feedback on cluster differentiation results from an administrator terminal. Furthermore, the processor performs a comparative analysis step by referencing defect trend data from similar industries flowing in from an external network to determine whether a new category currently generated by the internal system is a common issue across the entire industry or a defect specific to that product. Through these sophisticated unsupervised learning and graph analysis procedures, the electronic device achieves an excellent effect in reconstructing constantly changing customer complaints into a technical and systematic structure, without being confined to predefined categories.

[0056] delete

[0057] The processor performs a structuring step for the clusters from which the potential for complex defects is derived; by referencing a product design knowledge base pre-built in memory, it matches them with the hardware hierarchy and automatically constructs a Hierarchical Issue Tree that categorizes and represents issues from the lowest component level to the highest system level. The processor does not stop at merely identifying linguistic similarities but generates semantic links between the technical specification data within the knowledge base—where the company's design drawings, Bill of Materials (BOM), and functional definition documents are formally stored—and the defect clusters derived by artificial intelligence. The processor mathematically infers which physical module of the hardware the identified complex defects originated from, or whether a functional malfunction of a sub-assembly was transferred to a defect in the upper system.

[0058] In addition, when constructing the hierarchical issue tree, the processor calculates a structural matching index (smi) according to the following mathematical formula to precisely target the physical location that serves as the source of the defect.

[0059]

[0060] The definition and determination method of each parameter in the above mathematical formula is as follows.

[0061] First, a node represents the cosine similarity between a core defect keyword vector extracted by artificial intelligence and a specific hardware node description vector within the product design knowledge base. The value of the node is determined as a normalized numeric value between 0 and 1 by calculating the angle between the two vectors through a transformer-based technical literature embedding model.

[0062] Second, freq refers to the cumulative frequency of occurrence of individual feedback data included in the corresponding defect cluster. The value of the above freq is determined by counting the total sum of unique user feedback mapped to the same technical node during a specific time interval, and is used as an indicator to support the statistical significance of the issue.

[0063] Third, path represents the shortest path distance between a defect phenomenon node and a cause component node on the hardware dependency graph within the knowledge base. The value of the path indicates how directly the two elements are connected in terms of design structure, and the shorter the path (i.e., the smaller the path value), the higher the probability that the corresponding component is the direct cause of the defect.

[0064] Fourth, the tree represents the depth at which the corresponding node is located within the hardware hierarchy. The value of the tree starts at 1 for the top-level system and increases in integer units as it descends to lower-level components. This is determined to differentiate the weighting from higher-level defects that disrupt the entire system by adjusting the overall index, such that the denominator becomes larger for defects at the micro-component level.

[0065] A specific example applying the above mathematical formula is as follows. When a complex defect cluster related to a 'refrigerator control panel communication error' is detected, the processor maps it to a 'Main PCB Module' node within the knowledge base. It is assumed that the vector similarity (node) is calculated to be very high at 0.95 and the related feedback frequency (freq) is tallied at 120 cases. Additionally, if the design path between the defect phenomenon and the Main PCB is a one-level direct connection structure and the Main PCB is located at a two-level depth (tree=2) in the hierarchical structure, the structural matching index smi becomes (0.95 * 120) / (1 * 2) = 57. Based on this figure, the processor confirms that the issue is not a simple software bug but a problem at the core hardware layer and places it at the top of the issue tree.

[0066] Furthermore, when the name of the automatically generated category is modified or cluster merging is performed by the administrator terminal, the processor replaces the administrator's manipulation data with weights of a loss function and performs a latent space remapping process to rearrange the coordinate values ​​of the corresponding data within the vector space of the unsupervised learning model so that they are automatically classified into the corresponding category upon subsequent similar feedback input. The processor defines the administrator's act of modifying the name as a signal for 'labeling correction by a human expert.' Additionally, when the administrator performs an operation to merge two different clusters into one, the processor updates the neural network parameters in a direction that forcibly narrows the Euclidean distance between the centroids of the two clusters.

[0067] Furthermore, in the aforementioned latent space remapping process, the processor does not merely move the coordinates of the corresponding data, but calculates a transformation matrix to move entire adjacent regions within the latent space that share similar semantic features. Additionally, the processor analyzes the frequency and accuracy of the manager's operations to dynamically adjust the learning rate applied to the loss function. Moreover, after the remapping is completed, the processor performs a post-processing step to re-verify previously misclassified historical data in the new vector space to review the consistency of category assignments. Furthermore, through this self-correction process, the processor controls the system so that as manager intervention is repeated, the system's classification system becomes perfectly synchronized with the company's internal professional technical classification standards.

[0068] Furthermore, the processor integrates information from the hierarchical issue tree and the remapped latent space to generate a visual heatmap illustrating the impact of a specific hardware component defect on the overall system reliability, which is then provided to the administrator terminal. Additionally, the processor probabilistically calculates which category new feedback data, received in real-time, settles into within the remapped space, thereby performing precise classification from the initial stages of issue occurrence. Moreover, to maintain consistency between knowledge base updates and latent space remapping, the processor parallelly performs a coordinate system normalization process that resets the reference axis of the vector space when the hardware structure is changed. Through this series of advanced structuring and remapping procedures, the electronic device transcends the limitations of mechanical classification and implements an intelligent quality management architecture that evolves by interacting in real-time with the expertise of actual industrial sites.

[0069] delete

[0070] The processor includes a Recurrent Neural Network (RNN) or regression analysis model for time-series data analysis and is configured to probabilistically predict the likelihood of a specific issue spreading and customer churn rates by learning through comparative analysis of the frequency of customer feedback accumulated in the past and the trend of changes in the currently analyzed sentiment index. The processor inputs the frequency signal of the collected feedback data into a neural network-based time-series decomposition algorithm to separate residual components, which represent temporary fluctuations, from trend components, which represent long-term defect trends. It then detects the inflection point of diffusion by identifying, using Bayesian Change Point Detection, the point where the rate of change of the slope of the trend component exceeds a preset acceleration threshold. Furthermore, based on the change point detection results, it performs an Adaptive Time-Window Control process by dynamically reducing the size of the data sliding window for time-series analysis. This process assigns high computational weights to the latest feedback data during periods of rapid issue diffusion and expands the window size during stable periods to maintain the smoothing of the analysis.

[0071] To describe the time series analysis process performed by the above processor in more detail, the processor utilizes a Long Short-Term Memory (LSTM) or Gate Circular Unit (GRU) architecture to reflect the sequence characteristics of feedback data input sequentially over time. The processor compresses and stores complaint occurrence patterns from a specific past period into a Hidden State vector, and determines whether the current issue is an anomaly that deviates from the statistical normal range by comparing the newly incoming real-time sentiment index with the Hidden State. Furthermore, the processor does not merely measure increases or decreases in frequency, but calculates the rate at which changes in the intensity of negative sentiment accelerate to infer which stage the issue is currently located in the Life Cycle.

[0072] In the neural network-based time series decomposition step described above, the processor utilizes a multilayer perceptron (MLP)-based decomposition layer to remove complex noise contained in the raw signal. The processor separates the input signal into a deterministic trend and a stochastic residual, and corrects the reliability of the trend component by re-extracting meaningful variability from the residual component excluding random noise. Additionally, the processor calculates the first and second derivative values ​​of the separated trend component to monitor abrupt changes in the slope. Furthermore, the processor calculates the posterior probability of abrupt changes in the statistical distribution of the signal through a Bayesian change point detection algorithm, thereby deriving the optimal point in time when a specific product defect reaches a critical mass and poses a risk of spreading into a large-scale claim.

[0073] In the above adaptive time window control process, the processor variably adjusts the window size to simultaneously ensure the precision and real-time nature of the analysis. Under normal conditions where no changes are detected, the window is expanded to its maximum extent to reflect the long-term data flow, thereby preventing false alarms caused by temporary data outliers. Conversely, in crisis situations where changes are detected, the window is reduced to its minimum size to concentrate the processor's computational resources on the most recent feedback data. Additionally, the processor controls the time-series model by increasing its learning rate in real-time in response to the speed of issue diffusion, thereby enabling the model to adapt immediately to changing patterns.

[0074] In addition, the processor calculates a diffusion index (e) according to the following mathematical formula to quantify the risk of customer churn and issue diffusion based on the analysis results.

[0075]

[0076] The definition and determination method of each parameter in the above mathematical formula is as follows.

[0077] First, f represents the average frequency of negative feedback occurring per unit of time within the current window. The value of f is determined by dividing the number of data points classified into specific issue clusters among the feedback collected during the data acquisition stage by the length of the corresponding time window.

[0078] Second, s represents the average negative sentiment intensity (Severity) of the analyzed feedback. The value of s is determined by the integral of the probability density function of the negative class among the results of the attribute-based sentiment analysis model, and is calculated as a real value between 0 and 1.

[0079] Third, w represents the sliding window size determined by the current adaptive time window control. The value of w is determined based on the Bayesian change point detection results, becoming smaller as the issue changes rapidly and larger as it stabilizes. As w decreases, the denominator becomes smaller, causing the total exponent e to increase.

[0080] Fourth, r represents the reliability residual of the trend component calculated during the time series decomposition step. The value of r is determined in proportion to the inverse of the error between the predicted value of the neural network model and the actual observed value; as the purity of the signal increases and the trend becomes clearer, it takes on a lower value, thereby amplifying the exponent e.

[0081] A specific example applying the above mathematical formula is described as follows. In a situation where a 'noise defect' issue of a specific home appliance is detected, the average frequency over the past hour was recorded as 120 cases (f=120). The negative sentiment intensity analyzed by the artificial intelligence model was found to be very high at 0.85 (s=0.85). It was determined by Bayesian change point detection that the issue had started to spread, so the window size was reduced to the minimum value of 2 hours (w=2). In addition, the trend component was very distinct as a result of time series decomposition, resulting in a small residual error, and the reliability residual value was determined to be 0.4 (r=0.4). In this case, the diffusion index e calculated by the processor becomes (120 * 0.85) / (2 * 0.4) = 102 / 0.8 = 127.5. The processor recognizes that the calculated diffusion index exceeds the risk threshold and immediately sends a high-risk warning to the administrator terminal.

[0082] Furthermore, the processor continuously tracks changes in the calculated diffusion index to determine in real time whether the issue is entering a decay phase. Additionally, the processor matches the potential for diffusion of a specific issue with the behavioral patterns of past churned customers to calculate the percentage of customers expected to actually churn among the customer group currently exposed to the issue. Moreover, the processor analyzes the impact of identified risk factors on specific departments or production processes within the company and generates decision-making support data that proposes priorities for technical solutions. Furthermore, the processor optimizes the system's hardware load by detecting changes in data density within a sliding window and dynamically adjusting the computational load. Through this intelligent time-series analysis and adaptive control mechanism, the present invention goes beyond simply recording changes in customer reactions to achieve the outstanding effect of predicting future risk situations based on sophisticated numerical evidence.

[0083] delete

[0084] The processor calculates the prediction error between the diffusion value predicted at a previous time point and the data actually collected at the current time point, and performs an online learning-based self-calibration procedure to improve the prediction accuracy of the next time point by real-time recalibrating the hidden state weights of the recurrent neural network (RNN) if the prediction error exceeds an acceptable range. In processing customer response data input in a time series, the processor forms a sophisticated feedback loop that does not simply follow past patterns but immediately projects the actual market response flowing in at every moment onto the model's internal parameters. If the intensity and range of dissatisfaction diffusion predicted at a previous time point show a mathematical discrepancy with the actual collected data, the processor detects this as a signal of declining model reliability. Furthermore, the processor fine-tunes the weight matrix stored in the hidden layer of the recurrent neural network using back-propagation according to the magnitude of the prediction error, thereby inducing the model to more accurately track the dynamics of the specific issue currently occurring.

[0085] In order to quantitatively evaluate the precision of the self-correction procedure performed by the above-mentioned processor and determine the correction strength, in one embodiment of the present invention, a correction judgment index (cpj) is calculated according to the following mathematical formula.

[0086]

[0087] The definition and determination method of each parameter in the above mathematical formula is as follows.

[0088] First, prd represents the Predicted Value derived by the recurrent neural network model at the previous time step. The value of prd is determined as the result of the operation between the hidden state vector of the previous time step and the current input vector, and is derived as a normalized probability distribution value between 0 and 1.

[0089] Second, rel refers to the actual occurrence value (Real-time Value) of feedback data actually collected and refined at the current point in time. The value of the above rel is determined by scaling the frequency and sentiment intensity of a specific issue, aggregated in real time through the data acquisition module of the electronic device, to the same scale as the above prd.

[0090] Third, RTE refers to the rate of change of actual collected data per hour. The value of the aforementioned RTE is determined by calculating the slope between the actual value at the previous point in time and the actual value at the current point in time, and is used as an indicator to show whether the speed of issue diffusion is accelerating or decelerating.

[0091] Fourth, cnf represents the internal confidence score of the current artificial intelligence model. The value of the cnf is determined by the maximum value among the softmax probability values ​​calculated in the output layer of the recurrent neural network, and represents the level of mathematical confidence the model has in its prediction results.

[0092] A specific example applying the above mathematical formula is as follows. Assume a situation where the defect diffusion index (prd) of a specific product predicted by the model at a previous point in time was 0.85, but the actual aggregated market response value (rel) was 0.60. In this case, if the rate of change (rte), which is the actual data inflow rate, was calculated as 0.45 and the current model's self-confidence (cnf) was calculated as 0.70, then the processor, according to the above mathematical formula, cpj = ...is calculated. If this is performed, cpj = = = √{0.125} ≈ 0.353. When the processor confirms that the calculated cpj value exceeds the allowable threshold of 0.20 stored in memory, it immediately activates a self-correction procedure to perform an operation that strongly corrects the hidden state weights of the recurrent neural network toward the actual value of 0.60.

[0093] Furthermore, based on the self-corrected predictive model, the processor executes a Monte Carlo-style multipath simulation that adjusts external environmental variables or corporate response speed variables to generate best-case response scenarios and worst-case diffusion scenarios, respectively. Through thousands of iterative sampling operations, the processor inputs uncontrollable probabilistic variables into the simulation model, such as changes in government regulations, the occurrence of similar issues by competitors, or shifts in customer public opinion. Additionally, the processor generates numerous hypothetical future paths by adjusting variable factors, such as the scale of response resource deployment set by the company via the administrator terminal or the timing of the official apology announcement.

[0094] Furthermore, the processor selects statistically significant top and bottom scenarios from the generated multipath data. In the best response scenario, which is the most optimistic path, it quantitatively depicts the process in which the issue is resolved early and brand credibility is restored. Conversely, in the worst diffusion scenario, which is the most pessimistic path, it simulates the process in which the issue spreads explosively through social media and escalates into a large-scale boycott or legal dispute. Additionally, by reflecting the latest information on self-corrected hidden states during the scenario generation process, the processor maintains consistency to ensure that the initial conditions of the simulation do not deviate from actual market conditions.

[0095] Furthermore, the processor quantifies the estimated scale of damage for each scenario and transmits it to the administrator terminal along with guidelines. In the damage quantification step, the processor does not stop at merely counting the number of negative feedbacks but converts the expected direct revenue loss, intangible asset damage resulting from the decline in brand value, and the increase in customer support center operating costs into monetary units or indices. Additionally, the processor generates technical guidelines listing key response measures that must be implemented to prevent the worst-case scenario, in order of importance for each scenario.

[0096] Furthermore, the processor provides a visualized dashboard to the administrator terminal, displaying how current issues are modeled through self-correction and the probability distribution for each response scenario in the form of heatmaps or graphs. Additionally, when an administrator inputs arbitrary response variables, the processor immediately re-runs the Monte Carlo simulation to provide real-time predictive data on how much the corresponding measure can mitigate the scale of damage. Moreover, once the analyzed report is delivered to the administrator and the company takes actual action, the processor simultaneously performs an optimization procedure to receive feedback on the resulting data and reset the correction strength of the self-correction algorithm.

[0097] Furthermore, the processor refers to a high-dimensional vector database to search for past actual cases similar to the current simulation scenario, and supplements the reliability of the report by comparing and analyzing the corporate response results in those cases with the current prediction results. Additionally, when the processor load increases due to a rapid influx of data, the processor maintains a constant processing speed by flexibly adjusting the number of simulation samples or dynamically allocating parallel processing units. Through this organic self-correction and simulation mechanism, the electronic device of the present invention performs a highly intelligent control function that minimizes the uncertainty inherent in artificial intelligence and supports companies in making the most rational and sophisticated decisions.

[0098] delete

[0099] The processor calculates an integrated quality index with weights applied according to the importance of each individual customer by analyzing metadata stored in the memory, including the customer's product purchase history, past consultation count, and after-sales service (A / S) request history, in conjunction with the customer feedback data. The processor does not simply process the original text of the feedback, but rather matches the feedback sender with the customer profiling database retrieved from the memory to quantify the economic value contributed by the customer to the company and the proficiency of product usage. Furthermore, the processor quantifies the customer's past quality tolerance based on the customer-specific metadata stored in the memory, and performs a profile-aware attention modulation process by combining a weight tensor corresponding to the tolerance with the text embedding vector generated in the attribute-based sentiment analysis step, thereby dynamically adjusting the sensitivity of the integrated quality index according to the customer's purchase history and loyalty, even for the same negative expression. The processor ensures objectivity in analysis by amplifying attention weights for minor inconveniences raised by long-term customers and attenuating attention weights for extreme expressions by customers who habitually raise complaints through the profile-based attention modulation process.

[0100] In addition, the processor calculates a statistical correlation coefficient between the text-based sentiment score derived in the sentiment analysis step and the actual after-sales service (A / S) reception frequency included in the metadata, and identifies cases where the sentiment score is positive but the A / S reception frequency surges as a hidden dissatisfaction, thereby performing a consistency-based weight optimization process that lowers the reliability weight of the corresponding data when calculating the integrated quality index and assigns it to a deep reanalysis queue. In order to quantify the consistency of the data during the consistency-based weight optimization process, the processor calculates a data reliability score (drs) according to the following mathematical formula.

[0101]

[0102] The definition and determination method of each parameter in the above mathematical formula is as follows.

[0103] First, sct stands for Sentiment Confidence of Text. The value of the sct is a positive probability value derived by an artificial intelligence model analyzing a customer's feedback sentence, and is determined as a real value between 0 (completely negative) and 1 (completely positive).

[0104] Second, asf represents the actual after-sales service frequency. The value of asf is determined by normalizing the number of after-sales service requests made by the customer during a recent period, extracted from the metadata in the memory, and maps to converge to 1 as the number of requests increases.

[0105] Third, ltp stands for Loyalty Tenure Period. The value of the above ltp is determined by converting the elapsed period from the time of the customer's initial subscription or purchase to the present into days.

[0106] Fourth, cpv stands for Cumulative Purchase Value. The value of the above cpv is determined by extracting the total value of products or services purchased by the customer to date from the above memory.

[0107] A specific example applying the above mathematical formula is as follows. It is assumed that text feedback left by a specific customer obtained a very positive score of 0.9 (sct=0.9) by an artificial intelligence model, but has a normalized value of 0.8 (asf=0.8) in the actual metadata due to the customer's recent frequency of after-sales service requests being very high. Additionally, if the customer's retention period (ltp) is 500 days and the cumulative purchase value (cpv) is 2 million won (processed as 200 numerically), the data reliability score drs becomes |0.9 - 0.8| / (500 * 200) = 0.1 / 100,000 = 0.000001. The processor determines that the data consistency is high when the absolute difference is small and the customer value in the denominator is large, as in this case. Conversely, if the text is 0.9 but the after-sales service frequency is 0.1, the difference in absolute value increases, or if it is a new customer with low purchase value, the processor detects this as a potential dissatisfaction concealment state as the reliability score of the data changes and switches to a detailed analysis stage.

[0108] In addition, the processor generates quality risk reports by product and customer segment in real time based on the calculated integrated quality index. The processor isolates feedbacks where the data reliability score is below a preset threshold into a separate deep reanalysis queue and calls a high-performance artificial intelligence model with more precise hyperparameters to re-verify whether the sentence is ironic or has hidden intent. Furthermore, the processor directly injects the weight tensor generated during the profile-based attention modulation process into the intermediate layer of the neural network, thereby controlling the artificial intelligence model to infer language patterns of specific high-value customer groups that it did not experience during training by correcting them in real time.

[0109] In addition, the processor probabilistically calculates the expected demand for after-sales service resources anticipated to occur within the next six months, based on the results of the combined analysis of the metadata and feedback data. The processor provides an advance warning of the possibility of high-value customer groups churn, which could have a fatal impact on the company's profitability, by weighting the customer's cumulative purchase value and the current rate of change in sentiment. Furthermore, if data lacking consistency is found concentrated in a specific region or product lineup, the processor identifies this as a structural defect in the data collection system or post-management process of the relevant department and notifies the manager.

[0110] In addition, to improve the calculation accuracy of the integrated quality index, the processor sets the result of the quality report finally confirmed by the manager as the correct label and activates a feedback loop to readjust the sensitivity of each parameter of the mathematical formula. By allowing artificial intelligence to autonomously learn the influence of customer retention period or cumulative purchase value on the index calculation, the processor generates customized quality management indicators optimized for the company's customer base over time. Furthermore, in response to real-time fluctuating market environments, the processor maximizes computational efficiency by performing a feature selection step in parallel, which selects variables within the metadata that show the most significant correlation with the index calculation and removes unnecessary variables. Through such complex metadata integration and attention modulation technologies, the electronic device of the present invention operates as a three-dimensional and highly reliable enterprise quality management system that goes beyond static text analysis.

[0111] delete

[0112] For a specific customer group in which the integrated quality index has entered a preset risk range, the processor constructs a Customer-Product Interaction Graph using the product usage cycle and feedback occurrence interval in the metadata as nodes, and performs a risk path-based scheduling process that probabilistically calculates the expected churn point of the customer group through path analysis within the graph and reallocates analysis priorities. Going beyond merely measuring the negative sensitivity of individual feedback, the processor models the correlation between time-series usage patterns from the time of product purchase to the present and the intervals at which feedback occurs, based on Graph Theory. Each node of the Customer-Product Interaction Graph consists of metadata such as the customer's product usage frequency, log records by function, and customer center inquiry cycle, and the edges between each node represent the probability of a customer state transition over time. The processor analyzes the shortest path and transition speed on the graph, showing how a customer's behavior moves from the 'Normal Usage' node through the 'Complaint Accumulation' node to the final 'Churn' node. In addition, if the processor determines that a specific customer group has reached a threshold of the general churn path, it raises the processing priority of feedback data flowing in from that customer group in real time and places it at the top of the analysis waiting queue.

[0113] In addition, the processor calculates the risk convergence probability (rcp) according to the following mathematical formula to quantitatively evaluate the potential risk of the customer based on the customer-product interaction graph.

[0114]

[0115] The definition and determination method of each parameter in the above mathematical formula is as follows.

[0116] First, fbi stands for Feedback Interval. The value of the fbi is determined by calculating the time difference in days between the time of the customer's previous feedback retrieved from the memory and the time of the current feedback. A longer feedback interval reflects the possibility that latent complaints have matured.

[0117] Second, ltp refers to the Loyalty Tenure Period. The value of the above ltp is determined by converting the total elapsed time from the time of the customer's initial product registration or service subscription to the current analysis time into days.

[0118] Third, ucy stands for Usage Cycle. The value of the above ucy is determined by dividing the total number of product activations during a specific analysis period by the total number of days in that period, and is used as an indicator of customer dependency on the product.

[0119] Fourth, pvs stands for Product Value Score. The value of the above pvs is determined by calling from a predefined weight table in memory based on the price group, technical importance, or contribution to sales within the company of the product used by the customer.

[0120] A specific example applying the above mathematical formula is described as follows. Assume that a long-term customer with a cumulative subscription period (ltp) of 1,000 days raised feedback again after 10 days, and the feedback interval (fbi) is measured as 10 days. If the daily average usage cycle (ucy) of the product is 5 times and the strategic value score (pvs) of the product is set to 20, the processor, according to the above mathematical formula, rcp = { {10 * 1000}} / {5 * 20} = { The processor calculates {10000} / {100} = {100} / {100} = 1. Since the calculated rcp value exceeds the preset management threshold of 0.8, the processor classifies the customer as a high-risk group and strengthens risk path analysis.

[0121] In addition, after the output stage, if the results of the company's response measures performed according to the risk path-based scheduling are updated again as metadata, the processor performs a self-correcting index calibration procedure by comparing and analyzing the integrated quality index prior to the measure and the amount of change in sentiment after the measure, and fine-tuning the contribution score for each metadata item within the integrated quality index calculation algorithm using a back-propagation method. After the company completes response measures, such as sending an apology, replacing parts, or paying compensation to a specific customer based on a report provided through a manager terminal, the processor tracks changes in the polarity of the sentiment of the feedback left by that customer in real time. Furthermore, if the customer's sentiment index has not improved as expected after the response measures, the processor determines whether the integrated quality index calculated in the previous stage relied excessively or insufficiently on specific metadata items (e.g., purchase history or number of consultations).

[0122] Furthermore, in the self-correcting index calibration procedure described above, the processor calculates the loss between the actual sentiment improvement and the predicted index value, and updates the weights for each parameter within the integrated quality index calculation formula in a manner similar to Gradient Descent to minimize this loss. If statistical significance is found indicating that the number of consultations with the customer center has a greater impact on actual complaint resolution than the product price, the processor adjusts the contribution value of the consultation frequency item upward. Additionally, if the AI ​​model fails to detect signs of churn in advance through changes in the customer's product usage cycle, the processor automatically lowers the trigger threshold of the risk path-based scheduling to increase sensitivity.

[0123] Furthermore, the processor stores a new algorithm updated based on the index calibration results in memory and applies it to all subsequent incoming new feedback analyses, thereby promoting the system's self-generating performance improvement. Additionally, the processor compares and analyzes the efficiency of sentiment improvement by type of corporate response measures to generate additional data recommending optimal response scenarios regarding which reward system is most effective for which customer group. Moreover, the processor concurrently performs a feature selection optimization process in which, if the reliability of a specific metadata item is determined to be low during the calibration process, that item is temporarily excluded from the index calculation factors or replaced with a new item.

[0124] Furthermore, the processor provides the manager with a visualized list of contributions for each item, adjusted through the self-correcting index calibration procedure, thereby transparently disclosing how the basis for the AI's judgment evolves over time. Additionally, if the effectiveness of countermeasures is minimal due to chronic defects in a specific product family, the processor redefines this not merely as a customer management issue but as a risk requiring fundamental modification of the product design, and automatically transfers the matter to the development department. Moreover, to prevent data bias that may occur during index calibration, the processor performs a regularization procedure that re-validates the entire training dataset at regular intervals. Through these sophisticated scheduling and self-correction mechanisms, the electronic device of the present invention functions as a dynamic intelligent platform that goes beyond a static analysis tool to interact with the company's response strategies in real time and autonomously optimize quality control indicators.

[0125] As described above, the artificial intelligence-based customer feedback analysis method and electronic device according to the present invention infer technical causal relationships hidden in unstructured data, such as text and voice, by combining them with an ontology, predict time-series diffusion paths, and variably operate hardware and software resources, thereby dramatically improving the accuracy and efficiency of the analysis. The present invention described above is not limited by the aforementioned embodiments and attached drawings, and it will be obvious to those skilled in the art that various substitutions, modifications, and changes are possible within the scope of the technical concept of the present invention. Therefore, the true technical scope of protection of the present invention should be determined by the claims set forth below.

[0126] delete

[0127] delete

[0128] delete

[0129] delete

[0130] delete

Claims

Claim 1 An electronic device for analyzing customer feedback data collected from multiple channels comprises: a memory for storing unstructured customer feedback data and an artificial intelligence analysis algorithm; and a processor connected to the memory for executing the artificial intelligence analysis algorithm, wherein the processor comprises: a data acquisition step for collecting raw feedback data in the form of text or voice from various input paths and, if the raw feedback data is voice, converting it into text through an artificial intelligence-based speech-to-text engine; a preprocessing step for removing noise within the text data through an artificial intelligence-based language refinement model and correcting typos and non-standard words by referring to a domain-specific dictionary; and an aspect-based sentiment analysis step for extracting multiple product aspects from the preprocessed text data and determining the sentiment polarity for each aspect using a pre-trained deep neural network of a transformer structure. The system is configured to perform an output step of calculating the risk regarding product quality based on the sentiment analysis results above, and generating and outputting a real-time warning report to an administrator terminal when a preset threshold is exceeded, wherein the processor calculates the contextual association between a specific word and other words within a sentence as a probability weight through a Multi-Head Attention mechanism within the Transformer structure, and maps and extracts specific functional elements of a product or service components referred to by the customer's sentiment expression based thereon, wherein the processor extracts key tokens that exceed a preset valid threshold among the attention scores for each token calculated in the Multi-Head Attention layer, and calculates the vector similarity between the technology nodes within the Product Specification Ontology stored in the memory and the key tokens,Performs a Dynamic Ontology Anchoring process to identify whether the cause of a customer's complaint is a physical hardware defect or a software logic error; if the Confidence Score of the sentiment polarity calculated in the sentiment analysis step is below a threshold, performs a Counter-factual Inference process to recalculate the corresponding feedback sentence with inverted attention weights; analyzes the tensor deviation between the recalculation result and the previous calculation result to detect and correct irony or double negation expressions hidden within the context; the processor further performs an ensemble-based prediction correction step by running multiple deep neural network models set with different hyperparameters in parallel to calculate a Consistency Index for the analysis results of each model, and for unstructured data where the consistency is lower than a preset lower limit, searches for similar cases within the high-dimensional vector database stored in memory and assigns a weighted average value; receives feedback data regarding acceptance or modification of the warning report from the administrator terminal, and processes the received feedback using a loss function (Loss An electronic device characterized by performing an active learning-based model update procedure that fine-tunes the output layer weights of the attribute-based sentiment analysis model in real time by inputting them as arguments of a Function. Claim 2 delete Claim 3 delete Claim 4 delete Claim 5 delete

Images