Strategy generation method and device based on data analysis, equipment and medium

Abstract

The invention relates to the technical field of data processing, can be applied to business scenes of financial science and technology, medical treatment and health and the like, and discloses a strategy generation method, device, equipment and medium based on data analysis. Extracting historical interaction features and generating an association relationship, constructing a knowledge base, receiving current interaction data and extracting current interaction features, performing feature matching through the knowledge base to generate a matching result, and generating a strategy result based on the matching result. According to the method, the features are extracted based on the historical interaction data, the knowledge base is constructed, and the current interaction features are matched with the historical interaction features, so that the accuracy and the dynamic adaptability of the recommendation strategy are effectively improved, and refined strategy recommendation can be realized according to the event result data; and a personalized strategy can be dynamically generated based on real-time interaction data, so that the recommendation effect and the user satisfaction are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a data analysis-based strategy generation method, device, equipment and storage medium. Background Art

[0002] In the fintech sector, recommendation technology often relies on manual experience or static rule bases. Script templates are typically designed by business experts or account managers based on their subjective experience and pre-entered into the system. While these static script templates can cover some common scenarios, they often fall short when faced with complex and diverse user needs and unexpected events. Due to the cognitive limitations and varying knowledge backgrounds of the template designers, the generated scripts can be subjective, resulting in recommendations that are poorly aligned with actual user needs. Furthermore, once script templates are set in static knowledge bases, they often lack dynamic update mechanisms, making it difficult to effectively incorporate high-success scripts and strategies from real-world interactions. This results in recommendations remaining unchanged over time and making it difficult to adapt to the rapid changes in financial services, such as responding to high-volume customer complaints, financial fraud risk alerts, and dynamic investment recommendations. In these scenarios, the lack of a static rule base for quantitative evaluation makes it difficult to evaluate the effectiveness of recommendation strategies and determine which scripts truly improve customer satisfaction and business success.

[0003] In the healthcare business, recommended scripts are often used in online health consultations, smart health Q&A, and diagnosis and treatment services. Existing technologies rely on question-and-answer templates or scripts pre-set by medical experts, such as common symptom inquiries, medication instructions, and health management recommendations. However, due to the diversity of symptom descriptions of patients in healthcare scenarios and the fact that health conditions vary from person to person, static script templates are difficult to cover all scenarios, and scripts may even be missing in complex or critical situations. Due to the lack of real-time effect feedback and dynamic adjustment mechanisms, invalid or inefficient scripts in static templates cannot be identified and eliminated in a timely manner. In addition, the effectiveness of scripts is difficult to quantify and evaluate, and the system cannot determine the impact of different scripts on patient satisfaction or health improvement effects, and cannot achieve script optimization based on effect feedback, especially in high-demand scenarios such as emergency medical consultation, chronic disease management, and rehabilitation guidance.

[0004] In intelligent speech recommendation technology based on data analysis, although existing technologies have attempted to introduce multimodal data (such as voice, text, and interactive behavior) for analysis, their core still relies on manual rules or static knowledge bases. For example, in customer service systems, some intelligent recommendations are only matched based on pre-defined voice intonation and text content, while ignoring actual customer feedback and event evolution data. Due to the lack of dynamic analysis of historical interaction data and real-time updating of the knowledge base, recommendation strategies often fail to reflect high-success rate strategies in real scenarios. In addition, existing technologies have failed to establish an effective correlation analysis mechanism between speech effects and event outcomes, and are unable to evaluate the actual impact of speech on user satisfaction or problem resolution rate, resulting in difficulty in quantitative optimization of recommendation strategies. In the fields of voice customer service, online customer service, and automatic question and answer, this defect is particularly evident, and static rule bases often cannot adapt to the dynamic changes in customer needs and business environment. Summary of the Invention

[0005] The main purpose of the present invention is to provide a strategy generation method, device, equipment and storage medium based on data analysis, aiming to solve the technical problem in the existing technology that static strategy templates lack dynamic update and quantitative evaluation mechanism and cannot automatically optimize recommended strategies based on historical interaction data and event results.

[0006] To achieve the above objectives, the present invention provides a strategy generation method based on data analysis, comprising:

[0007] Collect historical interaction data and corresponding event result data in historical interaction scenarios;

[0008] Marking the historical interaction data as successful or failed according to predefined event result indicators;

[0009] Extract historical interaction features from the annotated historical interaction data;

[0010] Performing correlation analysis on the historical interaction features to generate correlation relationships between the historical interaction features and event result data;

[0011] Building a knowledge base based on the correlation between the historical interaction features and the event result data;

[0012] receiving input data of a current interaction scenario, and extracting current interaction features from the input data;

[0013] Matching the current interaction feature with the historical interaction features in the knowledge base to generate a matching result;

[0014] Based on the matching result, a strategy result for the current interaction scenario is generated.

[0015] Furthermore, to achieve the above-mentioned purpose, the present invention provides a strategy generation device based on data analysis, comprising:

[0016] The data collection module is used to collect historical interaction data and corresponding event result data in historical interaction scenarios;

[0017] An event marking module, configured to mark the historical interaction data as successful or failed according to predefined event result indicators;

[0018] Feature extraction module, used to extract historical interaction features from the annotated historical interaction data;

[0019] A feature association analysis module, configured to perform association analysis on the historical interaction features and generate an association relationship between the historical interaction features and event result data;

[0020] A knowledge base construction module, configured to construct a knowledge base based on the association between the historical interaction features and the event result data;

[0021] A real-time data access module is used to receive input data of the current interaction scene and extract current interaction features from the input data;

[0022] A feature matching module, configured to match the current interaction feature with the historical interaction features in the knowledge base to generate a matching result;

[0023] A strategy generation module is used to generate a strategy result for the current interaction scenario based on the matching result.

[0024] Furthermore, to achieve the above-mentioned purpose, the present invention also provides a determination machine device, which includes a memory, a processor, and a data analysis-based policy generation program stored in the memory and runnable on the processor. When the data analysis-based policy generation program is executed by the processor, the steps of the data analysis-based policy generation method as described above are implemented.

[0025] Furthermore, to achieve the above-mentioned purpose, the present invention also provides a machine-readable storage medium, on which a policy generation program based on data analysis is stored. When the policy generation program based on data analysis is executed by a processor, the steps of the policy generation method based on data analysis as described above are implemented.

[0026] Beneficial effects: The present invention relates to the field of data processing technology and can be applied to business scenarios such as financial technology and medical health. A strategy generation method, apparatus, equipment and medium based on data analysis are disclosed, including: collecting historical interaction data and corresponding event result data in historical interaction scenarios, successfully marking or failing the historical interaction data based on the event result data according to predefined event result indicators, extracting historical interaction features from the marked historical interaction data, performing correlation analysis on the historical interaction features, generating a correlation relationship between the historical interaction features and the event result data, constructing a knowledge base based on the correlation relationship between the historical interaction features and the event result data, receiving input data of the current interaction scenario and extracting current interaction features from the input data, matching the current interaction features with the historical interaction features in the knowledge base, generating a matching result, and generating a strategy result for the current interaction scenario based on the matching result. The present invention effectively improves the accuracy and dynamic adaptability of the recommendation strategy by extracting features based on historical interaction data and constructing a knowledge base, matching the current interaction features with the historical interaction features, and can not only realize refined strategy recommendations based on event result data, but also dynamically generate personalized strategies based on real-time interaction data, significantly improving recommendation effects and user satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0028] Figure 1 A schematic diagram of an application environment of a data analysis-based strategy generation method according to an embodiment of the present invention;

[0029] Figure 2 This is a flow chart of an embodiment of a strategy generation method based on data analysis according to the present invention;

[0030] Figure 3 Schematic diagram of functional modules of a preferred embodiment of a strategy generation device based on data analysis of the present invention;

[0031] Figure 4 A schematic structural diagram of a determination device according to an embodiment of the present invention;

[0032] Figure 5 FIG. 2 is another structural diagram of a determination device in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0034] The data analysis-based strategy generation method provided by the embodiment of the present invention can be applied in Figure 1In an application environment, the user terminal communicates with the server terminal through a network. The server terminal can collect historical interaction data and corresponding event result data in historical interaction scenarios through the user terminal, mark the historical interaction data as successful or failed based on the event result data according to predefined event result indicators, extract historical interaction features from the marked historical interaction data, perform correlation analysis on the historical interaction features, generate a correlation relationship between the historical interaction features and the event result data, build a knowledge base based on the correlation relationship between the historical interaction features and the event result data, receive input data of the current interaction scenario and extract current interaction features from the input data, match the current interaction features with the historical interaction features in the knowledge base, generate a matching result, and generate a strategy result for the current interaction scenario based on the matching result. The present invention effectively improves the accuracy and dynamic adaptability of the recommendation strategy by extracting features based on historical interaction data and building a knowledge base, matching the current interaction features with the historical interaction features, and can not only realize refined strategy recommendations based on event result data, but also dynamically generate personalized strategies based on real-time interaction data, significantly improving recommendation effects and user satisfaction. The user terminal can be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, and portable wearable devices. The server side can be implemented by an independent server or a server cluster composed of multiple servers. The present invention will be described in detail below through specific embodiments.

[0035] See also Figure 2 , Figure 2 This is a flow chart of an embodiment of a data analysis-based strategy generation method provided by the present invention. It should be noted that although a logical order is shown in the flow chart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0036] like Figure 2 As shown, the data analysis-based strategy generation method proposed in the present invention includes the following steps:

[0037] S10, collecting historical interaction data and corresponding event result data in historical interaction scenarios;

[0038] In this embodiment, in interactive scenarios, historical interaction data and its corresponding event outcome data must first be collected to support subsequent analysis and recommendation. Historical interaction data includes voice data, text data, and interactive behavior data, while event outcome data serves as an indicator of interaction effectiveness. Event outcome data includes key performance indicators such as user satisfaction scores and complaint resolution success rates, which are used to evaluate the effectiveness of interactive strategies.

[0039] Specifically, the process of collecting historical interaction data usually includes the following steps. First, historical voice data is obtained by connecting to the historical interaction system interface. This data usually exists in the form of audio files or real-time streaming data, which contains the content of the conversation between the user and the service staff. Secondly, historical text data is obtained. This data usually comes from chat records or email content in the customer service system, recording the specific questions raised by the user and the answers of the service staff. Furthermore, interactive behavior data also needs to be collected. This data usually includes indicators such as the user's response time, speaking frequency, and conversation duration during the interaction process. These behavioral data can provide more dimensional information for subsequent analysis.

[0040] Closely related to historical interaction data is event outcome data, which records customer feedback after each interaction. For example, user satisfaction scores can be obtained through questionnaires or automated feedback systems to measure user satisfaction with the entire interaction process. Complaint resolution success rates measure the effectiveness of complaint handling and are typically calculated by tracking complaint records and handling outcomes. For each interaction data set, the event outcome data must be accurately linked to facilitate subsequent in-depth analysis.

[0041] To ensure data accuracy, all collected historical interaction data and event outcome data must be associated with a unique session identifier. This identifier ensures that voice, text, and interactive behavior data correspond to the corresponding event outcome data, thus forming a complete data chain. Furthermore, to ensure temporal consistency between data, all collected data must be timestamped to ensure that different types of data (such as voice, text, and behavior data) remain synchronized. This way, the collected historical interaction data and its corresponding event outcome data can provide a reliable data foundation for subsequent feature extraction and correlation analysis.

[0042] In the fintech sector, historical interaction data can be collected from various sources, including customer phone consultation records and online customer service chat logs. Customer service systems can automatically collect this data through API calls and correlate it with customer feedback (such as satisfaction scores). This historical data serves as the foundation for subsequent analysis and optimized recommendation strategies. During the data collection process, all data is stored in a central data warehouse and linked to corresponding event outcome data using unique identifiers, providing the necessary data foundation for subsequent feature extraction and model training.

[0043] In the healthcare sector, historical interaction data collected can come from patient consultations, doctor's consultation records, and other sources. Using speech recognition and natural language processing technologies, the system can automatically extract patient voice and text data and correlate it with subsequent patient feedback (such as treatment effectiveness and satisfaction). This data not only supports the optimization of medical services but also provides doctors with more personalized response strategies in subsequent strategic recommendations.

[0044] This embodiment accurately collects historical interaction data and event outcome data, providing high-quality data input for subsequent analysis. Accurately correlating this data ensures consistency and integrity, enabling more precise feature extraction and correlation analysis, thereby improving the matching and effectiveness of recommendation results. Compared to traditional static rule base methods, this approach can update data in real time, adapting to changing environments and needs, and improving the adaptability and accuracy of recommendation strategies.

[0045] S20, based on the event result data, marking the historical interaction data as successful or failed according to predefined event result indicators;

[0046] In this embodiment, when processing historical interaction data, an important first step is to label the historical interaction data as successful or failed based on predefined outcome metrics based on the event outcome data. Event outcome data is feedback directly related to the interaction process, typically including user satisfaction, problem resolution rate, and complaint resolution rate. This data directly reflects the results of the interaction and serves as the basis for evaluating interaction effectiveness and generating recommendation strategies.

[0047] The core of the annotation process is to categorize and classify interaction data based on event outcome data. Through these annotations, the system can classify historical interaction data into success and failure categories, providing clear foundational data for subsequent feature extraction and correlation analysis. Event outcome data, such as user satisfaction scores and complaint resolution success rates, serve as the basis for annotation, helping to determine whether interactions have achieved their intended effects.

[0048] Result indicators are used to evaluate the success of an interaction. Specifically, they include but are not limited to the following aspects:

[0049] User satisfaction rating: This is usually obtained through customer surveys, rating systems, etc. The user's rating of the interaction process directly reflects the quality and effectiveness of the service and serves as the main criterion for judging whether the interaction is successful.

[0050] Complaint Resolution Success Rate: This metric measures how effectively customer service personnel or the system handles customer complaints. A high success rate indicates that the issue was effectively resolved during the interaction, while a low success rate indicates a flaw in the service process.

[0051] First, based on the relevant content of historical interaction data, the system extracts the required evaluation metrics from the status outcome data. For example, the satisfaction survey completed by the user after the interaction will serve as the basis for the user satisfaction score; when handling complaints, the relevant handling records will serve as the basis for the complaint resolution success rate. Next, the system labels the interaction data by comparing these metrics with predefined thresholds. For example, if the user satisfaction score is above the set threshold, the interaction data is marked as "successful"; if it is below the threshold, it is marked as "failed." A similar comparison and labeling process is also performed for the complaint resolution success rate.

[0052] On this basis, the annotated historical interaction data will be further used for feature extraction and analysis. Especially in the subsequent association analysis, the annotated data can help the system identify which features are related to successful interaction data and which features are related to failed data, providing a strong basis for generating recommendation strategies.

[0053] In the financial sector, when processing customer service records, user feedback on satisfaction and service quality is annotated as outcome indicators. The customer service representative's handling results are marked as success or failure based on the customer satisfaction score. If the customer is satisfied with the resolution, the interaction is marked as "successful," otherwise, it is marked as "failed." This annotated data provides important information for subsequent recommendation strategy generation.

[0054] In healthcare, a patient's treatment journey can be annotated with the doctor's treatment effectiveness rating and patient feedback. For example, if the patient's treatment outcome meets expectations and they express satisfaction with the process, the interaction is marked as successful. If the treatment outcome is poor or the patient complains, the interaction is marked as failed. This annotated data can be used to improve healthcare services, enhance treatment quality, and improve patient satisfaction.

[0055] This embodiment uses success or failure labeling based on event outcome data to provide clear categorized data for subsequent analysis. This process converts historical interaction data into categorized success or failure data using quantitative metrics, providing a foundation for subsequent feature extraction, correlation analysis, and strategy generation. This not only improves data processing accuracy but also ensures that subsequent recommendation systems are adjusted based on real-world feedback, thereby optimizing the matching of recommendation results and their practical application effectiveness.

[0056] S30, extracting historical interaction features from the annotated historical interaction data;

[0057] In this embodiment, extracting historical interaction features from the annotated historical interaction data is a crucial part of the entire data processing process. The core of this process is to extract key features that can represent the quality of the interaction process, user intentions, and behavior patterns from the interaction data that has been marked as successful or failed. In actual implementation, voice data, text data, and interaction behavior data are first distinguished from the annotated historical interaction data. Voice data represents the sound information used by customers or service personnel during the interaction process, text data represents the written record of the actual interaction, and interaction behavior data records the behavior patterns and strategy changes during the interaction process. These three data forms together constitute complete interaction data, which can reflect the quality of interaction from different dimensions.

[0058] During speech data processing, each piece of speech data is first framed. Framing aims to segment the continuous speech signal into multiple short-term signals. Each short-term signal retains the speech characteristics of that period. This processing method can eliminate information redundancy in long-term speech. Within each frame segment, acoustic features are extracted, including the fundamental frequency variation curve and speech energy intensity. The fundamental frequency variation curve represents the fluctuation of the speech pitch, which can reflect the speaker's tone. For example, the fundamental frequency is usually higher when the speaker is emotionally excited, while it tends to be more stable when the tone is calm. The speech energy intensity represents the loudness of the speech signal. A higher energy intensity usually indicates that the speaker is emphasizing something, while a lower energy intensity may indicate a lower volume or a gentler tone. Through this processing method, the system can effectively capture key information from the speech characteristics.

[0059] The processing of text data focuses on the extraction of semantic features. Text data is parsed using natural language processing technology, first performing word segmentation and part-of-speech tagging, and then extracting emotional keywords and sentence structures. The distribution density of emotional keywords is used to identify emotional tendencies in the text, such as positive emotions (such as satisfaction and happiness) or negative emotions (such as dissatisfaction and complaints). This distribution density can reveal the intensity and changes of emotions in the interaction. The frequency of interrogative sentences measures the frequency of questions raised by customers during the interaction process. This frequency can reflect the customer's understanding of the information or the confusion they have. In addition, semantic structural features can be further extracted, such as the complexity of the sentence structure and the ratio of long to short sentences, to further enhance the accuracy of text analysis.

[0060] The extraction of interaction behavior data focuses on identifying interaction pattern features. Interaction pattern features describe behavioral changes during an interaction, such as the correspondence between question types and response strategies. In practical applications, by analyzing the question types (e.g., consultation, complaint, technical support) and their corresponding response strategies (e.g., explanation, soothing, suggestion) in interaction data, the system can generate a mapping table between question types and response strategies. This mapping table reveals which response strategies are most effective when handling different types of questions. Furthermore, by analyzing the turn-to-turn intervals, the system can identify the response speed during the interaction. Shorter turn-to-turn intervals typically indicate a quick response, while longer intervals may indicate waiting or delay. Through this analysis, the system can identify issues during the interaction and implement optimizations.

[0061] After extracting acoustic, semantic, and interaction pattern features, the system normalizes and encodes these features into feature vectors. The purpose of normalization is to eliminate the impact of numerical differences between features, allowing different features to be compared on the same scale. Feature vector encoding converts all extracted features into a multidimensional vector representation, which facilitates subsequent association analysis and knowledge base construction. In this feature vector representation, the numerical value of each feature represents the intensity or frequency of expression of that feature in a specific interaction. Through this structured representation, the system can achieve a precise quantitative description of interaction characteristics.

[0062] In the financial sector, the system can extract voice data, text data, and interactive behavior data from historical interaction data. For example, in bank customer service, voice data can reflect the customer's emotional state. For example, a high frequency of complaining tones represents customer dissatisfaction. Text data can be used to extract customer satisfaction with the service through sentiment analysis. For example, positive sentiment words (such as "rest assured" and "thank you") represent positive customer feedback, while negative sentiment words (such as "complaint" and "dissatisfaction") represent negative feedback. Interaction pattern features can identify different types of questions (such as account inquiries and transfer failures) and the customer service representative's response strategies (such as explanation, reiteration, and escalation). Through this feature extraction method, the system can generate a complete feature vector for each historical interaction, providing support for subsequent analysis and recommendation strategy generation.

[0063] In healthcare, voice data can represent verbal communication between doctors and patients, such as descriptions of symptoms expressed by patients and explanations of their questions by doctors. The system can extract fundamental frequency changes and speech energy intensity in patients' voices to identify changes in their emotions, such as fear, anxiety, or trust. Text data represents records of questions and answers during the diagnosis and treatment process. Through semantic analysis, it can identify patients' acceptance or concerns about treatment plans. Interaction pattern features can reflect the doctor's diagnosis and treatment process and the frequency of patients' questions. For example, frequent questions when explaining treatment plans may indicate that patients are uncertain about the treatment plan. Through this multi-dimensional feature extraction, the system can identify the characteristics of effective communication and provide doctors with suggestions for optimizing their speech.

[0064] This embodiment can achieve multi-dimensional interaction feature quantification and structuring by extracting historical interaction features from the annotated historical interaction data. In particular, acoustic features can reflect the changing characteristics of speech during interaction, semantic features can reveal the emotions and intentions in the interactive text, and interaction pattern features can identify the correspondence between question types and response strategies during the interaction process. This multimodal feature extraction method can ensure that the system fully covers the key information in the interaction and achieves unified representation through feature vector encoding, thereby providing an accurate data foundation for subsequent association analysis, knowledge base construction, and strategy generation.

[0065] S40, performing correlation analysis on the historical interaction features to generate correlation relationships between the historical interaction features and event result data;

[0066] In this embodiment, historical interaction features are correlated and analyzed, and a correlation between historical interaction features and event outcome data is generated, aiming to discover potential patterns and correlations between multi-dimensional interaction features and event outcomes. First, it is necessary to distinguish three major feature categories from the extracted historical interaction features: acoustic features, semantic features, and interaction pattern features. Acoustic features include fundamental frequency change curves and speech energy intensity, semantic features include the distribution density of emotional keywords and the frequency of interrogative sentences, and interaction pattern features include a mapping relationship table between question types and answer strategies and a conversation turn transition interval. Each feature category can reflect the changes in voice, text, and behavior patterns during the interaction process from different dimensions.

[0067] In correlation analysis, acoustic features are first analyzed. During this process, the fundamental frequency variation curve is considered as time series data representing fluctuations in voice pitch, reflecting changes in the speaker's intonation, while the speech energy intensity represents the loudness of the speech signal and reveals the intensity of the speaker's tone. In this case, by calculating the Pearson correlation coefficient between the fundamental frequency variation curve and the user satisfaction score, the system can quantify the linear relationship between the two. The value ranges from -1 to 1, with positive values ​​indicating positive correlation, negative values ​​indicating negative correlation, and values ​​close to 0 indicating no correlation. In addition, the Spearman correlation coefficient between speech energy intensity and complaint resolution success rate is used to identify nonlinear associations between the two, which is particularly suitable for non-normally distributed data. The Spearman correlation coefficient also ranges from -1 to 1, with larger values ​​indicating stronger correlation.

[0068] For semantic features, by calculating the conditional probability of high-frequency keywords and user satisfaction scores, it is possible to identify which keywords are more likely to appear in high-satisfaction scenarios. Conditional probability indicates the possibility of specific keywords appearing when user satisfaction reaches a certain level. For example, positive sentiment words (such as "rest assured" and "satisfied") are generally more likely to appear in high-satisfaction interactions, while negative sentiment words (such as "dissatisfied" and "complaint") may be associated with low satisfaction. In addition, the mutual information between the sentiment tendency in the semantic features and the success rate of complaint resolution is used to quantify the information intercommunication between the two. Mutual information is an indicator that measures the nonlinear dependency between two variables and can reveal the impact of emotional expression on the success rate of complaint resolution.

[0069] The correlation analysis of interaction pattern features primarily focuses on identifying the mapping between question types and response strategies. By calculating the success rate of each response strategy (such as explanation, appeasement, and escalation) for each question type (e.g., consultation, complaint, and technical support), the system can identify effective strategies and provide a basis for subsequent strategy recommendations. Furthermore, by calculating the linear regression coefficient between the interval between conversational turn transitions and user satisfaction scores, the relationship between response speed and customer satisfaction can be quantified. The linear regression coefficient represents the increase or decrease in user satisfaction scores for each one-second increase in response time.

[0070] After completing the aforementioned feature correlation analysis, the system integrates these analysis results into three correlation sub-models: an acoustic feature correlation sub-model, a semantic feature correlation sub-model, and an interaction pattern feature correlation sub-model. The acoustic feature correlation sub-model is composed of the Pearson correlation coefficient of the fundamental frequency variation curve and the Spearman correlation coefficient of the speech energy intensity. The semantic feature correlation sub-model is constructed based on the conditional probability of high-frequency keywords and the mutual information of sentiment tendency, and can identify efficient semantic expressions. The interaction pattern feature correlation sub-model is composed of the linear regression coefficient of the success rate of the response strategy and the interval between conversational turn transitions. The benefit of this structured representation of sub-models is that the system can independently evaluate the impact of different types of features on the outcome of the event and select the most valuable features for subsequent strategy generation.

[0071] In the financial sector, the system can extract acoustic and semantic features from historical conversations between customers and bank customer service representatives. For example, when a customer expresses dissatisfaction with service, their speech energy intensity may increase and their tone may become more rapid. This emotional shift can be reflected in the fundamental frequency variation curve and speech energy intensity. By calculating the Pearson correlation coefficient between these acoustic features and customer satisfaction scores, the system can quantify the impact of emotional changes on satisfaction. In text analysis, the system can identify frequently occurring emotional keywords, such as "safety" and "reassurance," which are often associated with high satisfaction, while "complaint" and "dissatisfaction" are often found in low satisfaction records. Regarding interaction pattern features, the system can calculate the success rate of different response strategies (such as explaining product features, recommending discounts, and transferring to a specialist) when customers inquire about different financial products (such as loans, credit cards, and insurance). In this way, the system can generate sub-models that reflect the correlation between acoustic, semantic, and interaction pattern features and event outcomes.

[0072] In the healthcare sector, the system can extract acoustic and semantic features from doctor-patient consultation records. Regarding acoustic features, large variations in the fundamental frequency and high speech energy in a doctor's voice may indicate that the doctor is emphasizing important health advice. These features can be positively correlated with patient treatment compliance (such as scheduling follow-up appointments and taking medication on time). Text analysis can identify emotional keywords used by patients during consultations. For example, "worried" and "doubtful" are common among anxious patients, while "reassured" and "thankful" indicate that the patient trusts the doctor. Furthermore, by analyzing the frequency of questions asked by patients during consultations and the doctor's response strategies (such as explaining the condition, recommending tests, and providing psychological reassurance), the system can identify effective consultation strategies. By correlating these features with outcome data (such as patient return visit rates and treatment effectiveness), the system can construct sub-models that correlate acoustic, semantic, and interaction patterns.

[0073] This embodiment can achieve structured modeling between multi-dimensional interaction features and event results by performing correlation analysis on historical interaction features and generating correlation relationships. In particular, the Pearson correlation coefficient and Spearman correlation coefficient of acoustic features can quantify the impact of voice features on user satisfaction and the success rate of complaint resolution, the conditional probability and mutual information of semantic features can reveal the impact of emotional expression on interaction effects, and the success rate and linear regression coefficient of interaction pattern features can quantify the effectiveness of question types and response strategies. This kind of correlation analysis can reveal the key factors that affect event results from multiple dimensions, and convert these factors into structured correlation sub-models, ensuring that the system can be optimized in a data-driven manner in subsequent strategy generation.

[0074] S50, constructing a knowledge base based on the association between the historical interaction features and the event result data;

[0075] In this embodiment, a knowledge base is constructed based on the relationship between historical interaction features and event outcome data. By structurally storing historical interaction features and their relationship with event outcome data, a continuously updated knowledge base is formed. First, three types of features need to be extracted from the historical interaction features: acoustic features, semantic features, and interaction pattern features. Acoustic features include fundamental frequency variation curves and speech energy intensity. Semantic features include the distribution density of emotional keywords and the frequency of interrogative sentences. Interaction pattern features include a mapping table between question types and response strategies and the interval between dialogue turn transitions.

[0076] When building the knowledge base, the system first creates nodes for these historical interaction features. In the knowledge base, each feature is represented as an independent node to ensure that each feature can be independently referenced and retrieved. For acoustic features, the system creates the fundamental frequency change curve and speech energy intensity as acoustic waveform feature nodes respectively. Each node records the numerical range of the feature and the interaction scene information from which it comes. For semantic features, the system creates a semantic structure feature node for each high-frequency keyword and semantic tendency. Each node contains the keyword text, the frequency of occurrence, and its associated emotional type. The interaction mode feature nodes are represented as question type nodes and answer strategy nodes, which record the mapping relationship and success rate of each answer strategy under different types of questions.

[0077] The system also creates a result node for each event outcome data point. This data includes user satisfaction scores and complaint resolution success rates, representing the customer's subjective evaluation after the interaction and the objective outcome of the problem resolution, respectively. In the knowledge base, each result node includes the score value and its source timestamp, ensuring that the event outcome maintains temporal correlation with the corresponding interaction feature node.

[0078] After the node is created, the system establishes an association edge between the feature node and the result node based on the correlation between the historical interaction features and the event result data. For the acoustic feature association sub-model, the system configures the association edge weights between the acoustic feature node and the result node based on the Pearson correlation coefficient and the Spearman correlation coefficient. The Pearson correlation coefficient represents the linear relationship between acoustic features (such as the fundamental frequency change curve) and user satisfaction scores, while the Spearman correlation coefficient represents the nonlinear relationship between acoustic features (such as speech energy intensity) and the success rate of complaint resolution.

[0079] In the semantic feature association sub-model, the system configures the confidence level of the edges between semantic feature nodes and result nodes using conditional probability and mutual information. Conditional probability indicates the probability of a specific keyword appearing in high-satisfaction interactions, while mutual information indicates the correlation between semantic orientation (such as positive or negative sentiment) and the success rate of complaint resolution. Semantic feature nodes with high conditional probability and high mutual information are assigned higher confidence levels.

[0080] For interaction pattern features, the system configures the path strength of the associated edges between the interaction pattern feature node and the outcome node based on the success rate and linear regression coefficient. The success rate indicates the effectiveness of each response strategy for a specific question type, while the linear regression coefficient represents the linear relationship between the interval between conversational turn transitions and user satisfaction scores. A higher path strength indicates a more significant impact of the interaction pattern feature on the outcome.

[0081] After configuring the nodes and associated edges, the system stores these feature nodes, result nodes, and their associated edges in a graph database. A graph database is a database structure specifically designed to represent nodes and their relationships, enabling efficient node retrieval, edge traversal, and associated path queries. Acoustic feature nodes, semantic structure feature nodes, interaction pattern feature nodes, and result nodes constitute the basic node types in the knowledge graph, while associated edges represent the causal relationships or correlations between these nodes.

[0082] To ensure the knowledge base can dynamically adapt to the latest interaction data, the system is configured with dynamic update rules. When new interaction records are collected, annotated, and feature extracted, the system automatically generates new feature nodes and result nodes, or updates the edge weights of existing nodes, based on the features extracted from the new data and their associated event outcomes. For infrequently used nodes or edges with weak association strength, the system can set weight decay rules to ensure the knowledge base can dynamically adapt to the latest interaction patterns and user feedback.

[0083] In the financial sector, knowledge bases can construct multi-dimensional knowledge graphs for customer service. Acoustic feature nodes represent the voice waveform characteristics of customers when consulting about loans, credit cards, or insurance products. For example, the fundamental frequency variation curve represents the customer's emotional fluctuations, and the voice energy intensity represents the strength of the customer's expression. Semantic feature nodes represent high-frequency keywords in customer questions, such as "interest rate," "installment," and "handling fee," and associate the probability of these keywords appearing in high-satisfaction interactions. Interaction pattern feature nodes represent response strategies in different product consultation scenarios, such as "explain product advantages," "recommend promotions," and "transfer to a specialist." The associated edges between these nodes and the result nodes represent the effectiveness of different strategies through success rates and linear regression coefficients. The result node records the customer's satisfaction score and the success rate of complaint resolution.

[0084] In the field of healthcare, the knowledge base can construct a knowledge graph for medical consultation records. Acoustic feature nodes represent the voice features of doctors when explaining the condition or health advice, such as the fundamental frequency change represents the emphasis of the doctor's tone, and the voice energy intensity represents the clarity of the voice. Semantic feature nodes represent the keywords that patients use to describe their condition or express their emotions, such as "pain", "worry", and "doubt", and the correlation between these keywords and patient satisfaction and return visit rate is determined based on conditional probability and mutual information. Interaction pattern feature nodes represent the doctor's response strategies in different diagnostic scenarios, such as "explain the cause of the disease", "recommend examination", and "provide psychological comfort". These strategies are associated with the result node through the success rate and linear regression coefficient. The result node records the patient satisfaction score and return visit rate.

[0085] This embodiment constructs a knowledge base based on the correlation between historical interaction features and event result data, which can structure the representation of multi-dimensional interaction features and their impact on event results, ensuring that the system can efficiently retrieve, update and optimize interaction strategies. The knowledge base not only records the characteristics of each feature node in the historical interaction and its correlation strength with the event results, but also can dynamically adapt to the addition of new data to ensure that its correlation relationship is always consistent with the latest interaction pattern. The acoustic feature node represents its correlation with the result node through the Pearson correlation coefficient and the Spearman correlation coefficient, the semantic feature node represents its impact on the result through conditional probability and mutual information, and the interaction pattern feature node represents its strategy effectiveness through success rate and linear regression coefficient. This knowledge base structure can achieve efficient organization and dynamic correlation update of multi-dimensional features, improving the accuracy and adaptability of strategy recommendations.

[0086] S60, receiving input data of the current interaction scene, and extracting current interaction features from the input data;

[0087] In this embodiment, input data from the current interaction scenario is received and current interaction features are extracted from it, aiming to structure the real-time interaction data into analyzable feature information to support subsequent strategy generation. First, the system receives input data from the real-time interaction scenario, which can be user voice, text, or interaction behavior records. In the financial field, the input data may be voice interaction data from users inquiring about loans, credit cards, or insurance products via phone or online customer service; in the healthcare field, the input data may be voice data from patients describing symptoms, seeking health advice, or conducting remote consultations.

[0088] The system first preprocesses the received speech input data, including noise removal, speech segmentation, and speech feature extraction. Noise removal is achieved through adaptive filters or speech enhancement algorithms to ensure that the speech data does not contain environmental noise or background noise. Speech segmentation is achieved through voice activity detection (VAD), which divides the continuous speech signal into speech segments, each representing an independent piece of speech information. The segmented speech data is further processed to extract the current acoustic features, including the fundamental frequency variation curve and speech energy intensity. The fundamental frequency variation curve represents the pitch changes in the speaker's voice, and the speech energy intensity indicates the loudness and clarity of the sound.

[0089] For text data, the system uses natural language processing (NLP) technology to extract current semantic features from the input text. NLP technology includes text segmentation, keyword extraction, sentiment analysis and semantic structure parsing. Word segmentation is to divide the text into independent words or phrases, and keyword extraction uses TF-IDF (term frequency-inverse document frequency) or BERT (Bidirectional Encoder Represented Transformer) models to identify important words in the text. Sentiment analysis determines the emotional tendency of text expression, such as positive, negative or neutral, through a dictionary or sentiment classification model. Semantic structure parsing identifies semantic relationships in the text, such as inquiries, explanations or suggestions, through dependency analysis or semantic role labeling. These semantic analysis results are integrated into current semantic features, including high-frequency keywords, emotional tendencies and sentence structures.

[0090] The system also extracts current interaction pattern features from the interaction behavior data. Interaction behavior data includes how users interact with the system, such as clicking buttons, selecting menus, entering text, or responding by voice. For voice interaction, the system analyzes changes in the user's question types and response strategies. For example, in the financial field, question types may include "inquiring about loan interest rates," "applying for credit cards," and "complaining about insurance services," and response strategies may include "direct answers," "recommended products," and "transferring to a specialist." For the medical and health field, question types may include "inquiring about symptoms," "getting health advice," and "arranging examinations," and response strategies may include "explaining the cause of the disease," "recommending examinations," and "calming emotions." The system records the question types and response strategies in each interaction and integrates them into interaction pattern features.

[0091] After extracting the current acoustic, semantic, and interaction mode features, the system structures these features into a current interaction feature set. The current acoustic features are represented by a fundamental frequency curve and a feature vector of speech energy intensity; the current semantic features are represented by a vector of high-frequency keywords and sentiment; and the current interaction mode features are represented by a sequence of question types and response strategies. This feature set serves as the basis for subsequent matching and strategy generation, ensuring that the system can recommend the optimal strategy in real time based on the current user input.

[0092] In the financial sector, when a user inquires about loan products over the phone, the system receives the user's voice data. It first performs noise suppression and speech segmentation to extract the user's acoustic features, including the fundamental frequency variation curve and speech energy intensity. Simultaneously, the system converts the speech into text through speech recognition and applies NLP techniques to extract semantic features from the text, such as high-frequency keywords like "loan interest rate," "installment repayment," and "early repayment." It also analyzes the user's semantic preferences, such as "concerned about fees" and "desiring flexible repayment." For interaction pattern features, the system records the user's inquiry sequence, such as "inquiring about interest rates → inquiring about fees → selecting installments," and identifies the system's response strategy, such as "explaining the interest rate → recommending discounts → providing calculation examples." Ultimately, the system integrates these acoustic, semantic, and interaction pattern features into the current interaction feature set.

[0093] In the healthcare sector, when patients consult about health issues via remote consultations, the system receives their voice data, performs speech segmentation and noise removal, and extracts acoustic features such as fundamental frequency variation and speech energy intensity. Natural language processing (NLP) technology analyzes the patient's speech text, identifying symptom keywords such as "headache," "fever," and "cough," and assessing the patient's emotional tendencies, such as "worry" and "anxiety." Interaction pattern features include the type of questions asked by the patient, such as "symptom description" and "request for treatment advice," as well as the doctor's response strategy, such as "explaining the cause," "recommending examinations," and "providing reassurance." The system integrates these features into the current interaction feature set to support subsequent diagnosis and health advice generation.

[0094] This embodiment can achieve structuring and characterization of real-time data by extracting current interaction features from the input data of the current interaction scenario, ensuring that the system can dynamically adapt to the interaction needs of different users. Acoustic feature analysis ensures that the system can perceive the emotional changes in the user's voice, semantic feature analysis ensures that the system can understand the core information expressed by the user, and interaction pattern feature analysis ensures that the system can identify the user's question type and their response strategy preferences. Whether it is product consultation in the financial field or consultation interaction in the medical and health field, the system can accurately recommend strategies based on the current interaction features, significantly improving user experience and service efficiency.

[0095] S70, matching the current interaction feature with the historical interaction features in the knowledge base to generate a matching result;

[0096] In this embodiment, the system receives input data from the current interaction scenario and extracts current interaction features, including acoustic features, semantic features, and interaction pattern features. Acoustic features indicate the acoustic properties of the current user's speech, such as pitch variation and speech energy intensity. Semantic features include keywords, sentence structure, and emotional tendencies in the text content, representing the user's semantic expression. Interaction pattern features describe the user's behavioral patterns during the interaction, such as the type of questions asked and the order of response strategies.

[0097] During the acoustic feature matching process, the system traverses all acoustic feature nodes in the knowledge base and compares the current acoustic feature with each historical acoustic feature. Similarity is determined based on the directional similarity between the current and historical acoustic features. That is, if their acoustic patterns have similar fluctuation trends and energy distributions, they are considered similar. The system then selects the historical acoustic features that are closest to the current acoustic feature based on a similarity threshold, forming a set of candidate acoustic feature nodes.

[0098] For semantic feature matching, the system traverses all semantic structure feature nodes in the knowledge base and compares the current semantic feature with each historical semantic feature. Semantic similarity is determined by the overlap of keywords and the consistency of sentiment expressed in the two features. The repetition of high-frequency keywords and similar sentiment significantly increase semantic similarity. The system then selects semantic feature nodes with a similarity greater than a preset threshold to form a set of candidate semantic feature nodes.

[0099] During interaction pattern feature matching, the system traverses all interaction pattern feature nodes in the knowledge base and compares the differences between the current interaction pattern features and those of historical interaction patterns. The similarity of interaction pattern features depends on the question type, the sequence of response strategies, and the order of user interactions. The system selects historical interaction pattern feature nodes whose differences are less than a preset threshold as candidate nodes.

[0100] After obtaining a set of candidate nodes for acoustic, semantic, and interaction pattern features, the system assigns a weighted score based on the edge parameters associated with these feature nodes in the knowledge base. Acoustic feature candidate nodes are scored based on the strength of their association with event outcome data, such as success rate or user satisfaction. Semantic feature candidate nodes are scored based on their probability of occurrence and semantic relevance in event outcomes. Interaction pattern candidate nodes are scored based on their strategic effectiveness in historical interaction scenarios.

[0101] The system calculates a weighted score for each candidate node and integrates the acoustic, semantic, and interaction feature scores in a preset proportion to generate a composite score. All candidate nodes are sorted from high to low by composite score, and the top N highest-scoring candidate nodes are selected to form the matching result set. The candidate nodes included in this set serve as the basis for subsequent strategy generation.

[0102] In the financial sector, when a user consults about loan or insurance products via voice, the system receives the user's voice input in real time and extracts acoustic features from the speech, including intonation, speech rate, and energy variations. Simultaneously, the system uses speech-to-text technology to capture the user's text content, extracting semantic features such as keywords like "loan interest rate," "repayment period," and "early repayment," and analyzes sentiment within the text, such as "concerned about repayment pressure" and "desiring flexible repayment." Interaction pattern features capture the user's inquiry sequence, such as "inquiring about interest rate → inquiring about repayment period → choosing early repayment." The system then matches these current interaction features with historical interaction features in the knowledge base. Acoustic features are selected by comparing them with acoustic patterns in historical speech data. Semantic features are matched based on keyword overlap and sentiment consistency. Interaction pattern features are matched based on the sequence of question types and response strategies. The system calculates a score for each candidate node based on its matching performance and associated edge parameters, selecting the highest-scoring nodes to form a matching result set.

[0103] In the healthcare sector, when patients consult about health issues via remote consultations, the system receives their voice input and extracts acoustic features, such as speech rate and volume fluctuations. It also extracts semantic features from the speech text, such as symptom keywords like "headache," "fever," and "cough," and analyzes emotional tendencies, such as "anxiety" and "worry." Interaction pattern features record the type of patient question, such as "symptom description" and "request for treatment advice." The system matches these current interaction features with historical interaction features in the knowledge base. Acoustic features are filtered based on speech pattern similarity, semantic features are matched based on symptom keyword similarity and emotional consistency, and interaction pattern features are matched based on the sequence of question descriptions and response patterns. The system scores the strategic effectiveness of each candidate node in the historical data and selects the highest-scoring nodes to form a set of matching results.

[0104] This embodiment achieves dynamic recommendation of the optimal strategy in real-time scenarios by performing multimodal matching of current interaction features with historical interaction features in the knowledge base. Acoustic feature matching ensures that the system can judge emotional changes and communication intentions based on user voice patterns, semantic feature matching ensures that the system can accurately identify the needs and concerns expressed by users, and interaction pattern feature matching ensures that the system can dynamically adapt to the user's interaction style. Through multimodal feature weighted scoring and comprehensive ranking, the system can prioritize the strategy that best meets user needs in a variety of scenarios, improving the accuracy of recommendation results and user satisfaction.

[0105] S80: Generate a strategy result for the current interaction scenario based on the matching result.

[0106] In this embodiment, the system generates a policy outcome for the current interaction scenario based on the matching results. First, the system extracts the event outcome data associated with each candidate node from the matching result set. This event outcome data typically includes a user satisfaction score and a complaint resolution success rate, representing user satisfaction with the service in historical interaction scenarios and the probability of successful resolution of user complaints or disputes, respectively.

[0107] The system comprehensively analyzes the event outcome data for each candidate node and calculates a comprehensive success rate indicator. This indicator is derived through a weighted calculation: the user satisfaction score is multiplied by a first weighting factor, and the complaint resolution success rate is multiplied by a second weighting factor, with the sum of the two weighting factors equaling one. The first weighting factor is greater than the second weighting factor, reflecting the relative priority of user satisfaction in policy evaluation.

[0108] The system compares the overall success rate of each candidate node with a preset success rate threshold, selecting those with a success rate greater than the threshold to form a set of valid strategy nodes. Each node in the set represents an interaction strategy that has achieved a high success rate in historical scenarios and has high-priority reference value.

[0109] Based on the set of valid policy nodes, the system generates a policy combination. This combination consists of three parts: a response template, voice and intonation adjustment suggestions, and conversation pacing control parameters. The response template defines the textual response to the user's question. Voice and intonation adjustment suggestions are based on acoustic features to ensure that the voice broadcast aligns with the user's emotions and context. Conversation pacing control parameters adjust the interaction rhythm based on the characteristics of the interaction pattern to ensure smooth communication.

[0110] The generated strategy combination is immediately applied to the current interaction scenario, and the system records the execution effect in real time during the interaction process, including the user feedback satisfaction score and the actual complaint resolution success rate.

[0111] To achieve dynamic optimization, the system evaluates strategy combinations in real time based on their effectiveness. When actual user satisfaction scores reach or exceed a preset threshold, and the actual complaint resolution success rate reaches or exceeds a preset threshold, the system adds that strategy combination, along with its corresponding current interaction characteristics and outcome data, to the knowledge base. This dynamic update mechanism ensures that the knowledge base continuously incorporates strategies with high success rates, adapting to changing user needs and interaction patterns.

[0112] In the financial sector, when a user uses voice to inquire about loan interest rates or insurance claims, the system receives the user's voice input and generates matching results based on the current interaction characteristics. From the matching result set, the system extracts historical event outcome data for each candidate node, such as user satisfaction scores for previous loan inquiries and the percentage of successfully resolved complaints. The system calculates a comprehensive success rate metric for each candidate node, screens nodes with high success rates, and generates a set of valid strategy nodes. Based on this set of valid strategy nodes, the system generates a strategy combination. For example, a "response template for loan interest rates" might include "The current interest rate is X%, and you can choose flexible repayment options based on your repayment ability." Suggestions for adjusting voice intonation might include "Use a gentle and patient tone," while conversation pacing parameters might specify "Maintain a moderate speaking speed to ensure user understanding." This strategy combination is immediately applied to the current user interaction and the user's feedback, including the satisfaction score and complaint resolution results, is recorded. If the user satisfaction score is above 90% and the complaint resolution success rate exceeds 80%, the system adds the strategy combination, along with its corresponding interaction characteristics and event outcome data, to the knowledge base.

[0113] In the healthcare sector, when patients consult remotely about health issues, the system extracts historical consultation strategies with high success rates from a set of matching results and generates response templates based on these strategies. For example, "The symptoms you mentioned may be related to an upper respiratory tract infection. Please describe the duration of the symptoms and any other discomfort you experience." Suggestions for adjusting voice intonation might include "speak more slowly and in a gentler tone," while parameters for controlling conversational pacing stipulate "allow ample time for the patient to describe their symptoms." The system applies this strategy combination in actual consultations and records patient feedback and consultation results in real time. If patient satisfaction ratings exceed 95%, the system adds the strategy combination, along with its corresponding interaction features and outcome data, to the knowledge base.

[0114] This embodiment generates strategy results based on matching results. The system automatically applies strategy combinations with high success rates to the current interaction scenario and dynamically updates the knowledge base in real time based on user feedback and performance data. This mechanism ensures that the system can quickly adapt to changing user needs and recommend optimal strategies for different scenarios. Strategy generation is not only based on historically high-success strategies, but also validates their effectiveness with real-time user feedback, achieving closed-loop optimization.

[0115] The present invention relates to the field of data processing technology and can be applied to business scenarios such as financial technology and medical health. A strategy generation method, device, equipment and medium based on data analysis are disclosed, including: collecting historical interaction data and corresponding event result data in historical interaction scenarios, successfully marking or failing the historical interaction data based on the event result data according to predefined event result indicators, extracting historical interaction features from the marked historical interaction data, performing correlation analysis on the historical interaction features, generating a correlation relationship between the historical interaction features and the event result data, building a knowledge base based on the correlation relationship between the historical interaction features and the event result data, receiving input data of the current interaction scenario and extracting current interaction features from the input data, matching the current interaction features with the historical interaction features in the knowledge base, generating a matching result, and generating a strategy result for the current interaction scenario based on the matching result. The present invention effectively improves the accuracy and dynamic adaptability of the recommendation strategy by extracting features based on historical interaction data and building a knowledge base, matching the current interaction features with the historical interaction features, and not only realizing refined strategy recommendation based on the event result data, but also dynamically generating personalized strategies based on real-time interaction data, significantly improving recommendation effects and user satisfaction.

[0116] In one embodiment, the above step S10 includes:

[0117] S101, capturing historical voice data and historical text data in historical interaction scenarios, and generating a unique session identifier for each interaction session;

[0118] S102, extracting historical interaction behavior indicator data from historical interaction logs;

[0119] S103, associating the historical voice data, historical text data, and historical interaction behavior indicator data using the unique session identifier;

[0120] S104, obtaining historical event result data corresponding to the unique session identifier from a historical database;

[0121] S105, performing timestamp alignment processing on the historical voice data, historical text data, historical interaction behavior indicator data, and historical event result data;

[0122] S106 , integrating the historical voice data, historical text data, and historical interaction behavior indicator data after the timestamp alignment processing into historical interaction data.

[0123] In this embodiment, during the data collection phase, the system first captures historical voice data and text data from historical interaction scenarios. Historical voice data typically comes from recorded voice interactions between users and the system or service personnel, including phone call recordings, online voice conversations, and other forms. Historical text data includes text content entered by users during the interaction process, such as online chat logs, text message conversations, or form submissions. By capturing and storing this data, the system ensures that subsequent data processing has a multimodal data source.

[0124] To accurately associate data, the system generates a unique session identifier for each interactive session. This identifier is a separate identification tag, typically composed of a timestamp, user identifier, and session number, such as "202305101230_USER001_SESSION001." This unique identifier ensures the uniqueness of each session data and prevents confusion between different sessions.

[0125] After generating a unique session identifier, the system extracts historical interaction behavior metrics from historical interaction logs. These metrics include customer response time, conversation rounds, issue resolution rate, and user wait time. These metrics reflect the effectiveness of user interactions with the system or service personnel. By analyzing these interaction behavior metrics, the system can further understand the behavioral characteristics of users during interactions.

[0126] The system associates historical voice data, historical text data, and historical interactive behavior indicator data using a unique session identifier. Specifically, the system performs data association operations in the historical data table based on the unique session identifier to ensure that the voice data, text data, and behavioral indicator data form a complete data record under the same session. For example, if the voice data is identified as "SESSION001_Audio.wav", the text data is identified as "SESSION001_Text.txt", and the behavioral indicator data record is "SESSION001_Log.json", then the three are associated as a complete data unit through the unique session identifier "SESSION001".

[0127] The system further retrieves historical event outcome data corresponding to the unique session identifier from the historical database. This historical event outcome data includes user feedback scores (such as satisfaction scores) and complaint resolution success rates, representing the user's feedback after each interaction. This event outcome data can be directly derived from explicit user feedback (such as ratings and comments) or automatically calculated by the system (such as problem resolution rates).

[0128] After ensuring the completeness of the data association, the system performs timestamp alignment on historical voice data, historical text data, historical interactive behavior indicator data, and historical event outcome data. Timestamp alignment is a data synchronization technology that ensures temporal consistency among different data types by aligning the timestamps of each data type. Specifically, the system first sorts each data type based on its collection timestamp, and then aligns the voice, text, and behavioral indicator data in the same session along the timeline. For example, the voice segment corresponding to "00:15 seconds" in the voice data is consistent with the text generated at "00:15 seconds" in the text data, and the response time recorded at "00:15 seconds" in the behavioral indicator data is also associated with it.

[0129] After timestamp alignment is complete, the system integrates the aligned historical voice data, historical text data, and historical interaction behavior indicator data into historical interaction data. During this integration process, the system uses structured data tables or document storage formats, such as JSON, CSV, or relational database tables, to store each session data in a unified format. The integrated historical interaction data includes voice data, text data, behavioral indicator data, and event outcome data, forming a complete and consistent set of interaction data.

[0130] This embodiment collects historical voice data, text data, interaction behavior indicator data, and event outcome data from historical interaction scenarios. The system achieves comprehensive multimodal data collection and integration, ensuring that each interactive session is supported by multi-dimensional data. The introduction of unique session identifiers ensures accurate data association, and timestamp alignment ensures temporal consistency of voice, text, and behavior indicator data, forming a complete and high-quality historical interaction dataset. This mechanism enables the system to accurately capture the details of user interactions with the system or service personnel, providing a reliable data foundation for subsequent data analysis and strategy generation.

[0131] In one embodiment, the above step S30 includes:

[0132] S301, performing frame processing on the voice data in the annotated historical interaction data, and extracting acoustic waveform features from the frame-processed voice data;

[0133] S302, performing semantic analysis on the text data in the annotated historical interaction data to generate semantic structure features;

[0134] S303, analyzing the interaction behavior data in the annotated historical interaction data to generate interaction pattern features;

[0135] S304 , normalizing and encoding the acoustic waveform features, semantic structure features, and interaction pattern features into feature vectors to generate a historical interaction feature set.

[0136] In this embodiment, extracting historical interaction features from the annotated historical interaction data is the core process of realizing data characterization. By processing voice, text and interaction behavior data, complex multimodal data is converted into a structured feature set. In this process, framing is first performed on the voice data. Framing is to divide the continuous voice signal into time windows of fixed length (such as 25 milliseconds) to ensure that subsequent feature extraction can capture the detailed features of the voice signal. For each frame of voice data, the system extracts acoustic waveform features, including the fundamental frequency change curve and voice energy intensity. The fundamental frequency change curve reflects the frequency fluctuations in the voice signal and can identify information such as speech rate and intonation changes, while the voice energy intensity is used to measure the loudness of the voice and reflect the changes in the user's emotional intensity during the interaction. The fundamental frequency change curve is obtained by analyzing the autocorrelation function or Fourier transform of each frame of voice, while the voice energy intensity is calculated based on the sum of the squares of the amplitude of each frame of voice signal.

[0137] In the feature extraction of text data, the system performs semantic parsing, which aims to generate semantic structural features from text data. Semantic parsing first performs word segmentation and part-of-speech tagging on the text data to identify the grammatical relationship between keywords and words in the text. The system calculates the distribution density of emotional keywords based on the predefined emotional dictionary. The distribution density of emotional keywords represents the frequency ratio of positive and negative emotional words in the text, which can reflect the user's emotional tendencies. The system also counts the frequency of interrogative sentences in the text. Interrogative sentences are usually identified by specific grammatical structures (such as question words or question marks), which can reflect the user's questions or intentions to seek help. Semantic parsing is based on natural language processing technologies, such as word vector models (Word2Vec), sentence vector models (BERT) or sentiment classification models.

[0138] Feature extraction from interaction behavior data aims to generate interaction pattern features. The system analyzes user behavior patterns during interactions, including a mapping table of question types and response strategies, as well as the interval between conversational turns. This mapping table is generated by analyzing the relationship between common questions and corresponding response strategies in historical interaction data, such as the "account balance inquiry" and "balance reminder" strategies. The system generates this mapping table by counting the frequency of each question type and its corresponding response strategy in historical interactions. The interval between conversational turns represents the difference in response time between users and the system or service personnel during multiple rounds of interaction. This feature is obtained by calculating the time difference between adjacent conversational turns. This feature can reveal user response habits during interactions. For example, a quick response indicates user familiarity with the system, while a slow response may indicate difficulty understanding the system or a need that is unclear.

[0139] After completing the feature extraction of speech, text, and interactive behavior data, the system normalizes the acoustic waveform features, semantic structure features, and interactive pattern features. Normalization ensures that different features have consistent numerical magnitudes, thereby preventing certain features from dominating in subsequent analysis due to excessively large or small values. Common normalization methods include min-max normalization (scaling the data to a specified interval, such as 0 to 1) or standardization (adjusting the data to a mean of 0 and a standard deviation of 1). The normalized feature data is then feature vector encoded, which integrates different types of features into a vector representation to form a set of historical interaction features. Feature vector encoding can integrate speech, text, and interactive pattern features into the same feature space, facilitating subsequent feature association analysis and strategy generation.

[0140] This embodiment extracts historical interaction features from the annotated historical interaction data, and the system realizes the structured conversion of multimodal data, which can comprehensively capture the user's interaction features from voice, text and interactive behavior. The acoustic waveform feature extraction of frame processing ensures that the voice data can reflect the user's emotions and tone changes, the emotional keyword distribution and question sentence frequency of semantic analysis can reveal the user's textual expression of emotions and intentions, and the interaction pattern features capture the user's behavior patterns during the interaction process. Normalization processing and feature vector encoding integrate multimodal features into a unified feature vector, ensuring that different features are consistent, which is convenient for subsequent data analysis and strategy generation. Therefore, the system can automatically extract and quantify interaction features based on the user's historical interaction behavior without relying on fixed rules, thereby improving the accuracy and adaptability of the recommendation strategy.

[0141] In one embodiment, the above step S40 includes:

[0142] S401, determining a Pearson correlation coefficient between a fundamental frequency change curve in an acoustic feature subset in the historical interaction features and a user satisfaction score in the event result data;

[0143] S402, determining the Spearman correlation coefficient between the speech energy intensity in the acoustic feature subset in the historical interaction features and the complaint resolution success rate in the event result data;

[0144] S403, determining the conditional probability of the high-frequency keywords in the semantic feature subset in the historical interaction features and the user satisfaction score in the event result data;

[0145] S404, determining the mutual information between the sentiment tendency in the semantic feature subset in the historical interaction features and the complaint resolution success rate in the event result data;

[0146] S405, determining the success rate of each strategy in the mapping relationship table between question types and answer strategies in the interaction pattern feature subset in the historical interaction features;

[0147] S406, determining a linear regression coefficient between the conversation turn transition interval in the interaction pattern feature subset in the historical interaction features and the user satisfaction score in the event result data;

[0148] S407, constructing an acoustic feature correlation sub-model between the acoustic feature subset and the event result data based on the Pearson correlation coefficient and the Spearman correlation coefficient;

[0149] S408, constructing a sub-model for associating the semantic feature subset with the event result data based on the conditional probability and mutual information;

[0150] S409: Based on the success rate and the linear regression coefficient, a sub-model of the association between the interaction pattern feature subset and the event result data is constructed.

[0151] In this embodiment, the purpose of performing correlation analysis on historical interaction features is to determine the correlation between these features and the event result data, thereby establishing a multi-dimensional feature-result correlation model. First, the system analyzes a subset of acoustic features in the historical interaction features. In this subset, the fundamental frequency change curve is a feature that reflects the pitch fluctuations in the voice signal, which can capture the emotional changes and expression intentions in the user's voice. The system traverses the fundamental frequency change curve in each historical interaction data and performs correlation analysis with the user satisfaction score in the corresponding event result data. The calculation method used is the Pearson correlation coefficient. The Pearson correlation coefficient is an indicator used to measure the linear correlation between two variables. The value range is between negative one and one. The closer the absolute value is to one, the stronger the correlation. By calculating the Pearson correlation coefficient of the fundamental frequency change curve and the user satisfaction score, the system can quantify the linear relationship between the user's voice emotional expression and service satisfaction.

[0152] Within the acoustic feature subset, the system also analyzes the correlation between speech energy intensity and the success rate of complaint resolution in the event outcome data. Speech energy intensity represents the loudness or sound pressure of a speech signal and is typically used to identify the intensity of emotion expressed by a user and the clarity of their language. The system evaluates the relationship between speech energy intensity and the success rate of complaint resolution by calculating the Spearman correlation coefficient between the two. The Spearman correlation coefficient is a non-parametric correlation measure based on ranking and is suitable for analyzing nonlinear correlations. The closer its absolute value is to one, the stronger the correlation. Using the Spearman correlation coefficient, the system can quantify the correlation between the intensity of speech expression and whether the user's problem was successfully resolved.

[0153] In the association analysis of semantic feature subsets, the system first identifies high-frequency keywords, which represent the core intent or emotion expressed by users in textual interactions. The system calculates the frequency of these high-frequency keywords in the text and combines them with the user satisfaction scores in the event outcome data to calculate conditional probabilities. Conditional probabilities represent the probability that a user's satisfaction score will reach a preset level if a particular keyword appears in the user's text. This analysis can reveal the impact of specific keywords (such as "satisfaction" and "complaint") on user satisfaction.

[0154] Within the semantic feature subset, the system also analyzes the relationship between sentiment and complaint resolution success rates. Sentiment represents the proportion of positive or negative sentiment in text data, such as positive sentiment ("satisfaction," "gratitude") or negative sentiment ("dissatisfaction," "complaint"). The system measures the degree of information correlation between sentiment and complaint resolution success rates by calculating the mutual information between the two. Mutual information is a metric used to measure the degree of information sharing between two random variables; a higher value indicates a stronger correlation. This analysis can reveal how emotional expression in text affects service outcomes.

[0155] In the correlation analysis of the interaction pattern feature subset, the system focuses on analyzing the mapping table between user question types and response strategies. This table records the association between each question type and its most commonly used response strategy. The system generates success rate data by calculating the success rate of each response strategy for different question types. The success rate indicates the probability that a response strategy will successfully resolve a specific question type. This data is used to measure the effectiveness of different strategies.

[0156] The system also analyzes the correlation between turn-to-turn intervals and user satisfaction scores. Turn-to-turn intervals represent the difference in response time between users and the system during multiple rounds of interaction. The system determines this correlation by calculating the linear regression coefficient between turn-to-turn intervals and user satisfaction scores. The linear regression coefficient indicates the impact of response time on satisfaction. A positive value indicates that an increase in response time increases satisfaction, while a negative value indicates that a decrease in response time increases satisfaction.

[0157] After completing the correlation analysis of each subset, the system constructs three types of correlation sub-models based on these correlation parameters. First, an acoustic feature correlation sub-model is constructed based on the Pearson correlation coefficient and the Spearman correlation coefficient. This model is used to predict the impact of acoustic features (such as voice fundamental frequency and energy) on user satisfaction and the success rate of complaint resolution. Secondly, a semantic feature correlation sub-model is constructed based on conditional probability and mutual information. This model is used to quantify the impact of specific keywords and emotional tendencies in the text on user feedback. Finally, an interaction mode feature correlation sub-model is constructed based on the success rate and linear regression coefficient to predict the impact of different response strategies and interaction modes on user satisfaction and service success rate.

[0158] In this embodiment, by performing correlation analysis on historical interaction features and generating correlation relationships between historical interaction features and event outcome data, the system can quantify and reveal the multi-dimensional correlations between acoustic, semantic and interaction pattern features and user feedback results (such as user satisfaction and complaint resolution success rate) from historical data. Pearson and Spearman correlation analysis of the acoustic feature subset enables the system to quantify the linear and nonlinear relationships between user voice expression features and satisfaction. Conditional probability and mutual information analysis of the semantic feature subset ensure that the system can identify the impact of keywords and emotional tendencies in the text on user feedback. The success rate and linear regression analysis of the interaction pattern feature subset provide a quantitative relationship between user behavior patterns (such as question type, response time) and user feedback. By constructing an association sub-model of acoustic, semantic and interaction pattern features, the system can provide personalized recommendations based on precise feature-result associations in subsequent strategy generation.

[0159] In one embodiment, the above step S50 includes:

[0160] S501, creating corresponding acoustic waveform feature nodes, semantic structure feature nodes, and interaction pattern feature nodes for each acoustic waveform feature, semantic structure feature, and interaction pattern feature in the historical interaction features;

[0161] S502, creating a result node for each event result data;

[0162] S503, configuring the association edge weight between the acoustic feature node and the result node according to the Pearson correlation coefficient and the Spearman correlation coefficient in the acoustic feature association sub-model in the association relationship;

[0163] S504, configuring the confidence of the association edge between the semantic feature node and the result node according to the conditional probability and mutual information in the semantic feature association sub-model in the association relationship;

[0164] S505, configuring the association edge path strength between the interaction pattern feature node and the result node according to the success rate and the linear regression coefficient in the interaction pattern feature association sub-model in the association relationship;

[0165] S506: Store the acoustic feature nodes, semantic feature nodes, interaction mode feature nodes, result nodes and associated edges in a graph database to complete the construction of the knowledge base.

[0166] In this embodiment, the process of building a knowledge base begins by converting the association between historical interaction features and event outcome data into a knowledge graph structure. First, corresponding feature nodes are created for the acoustic waveform features, semantic structure features, and interaction pattern features in the historical interaction features. The acoustic waveform feature node represents the audio features of the user's speech expression, including the fundamental frequency variation curve and speech energy intensity. The fundamental frequency variation curve captures the pitch variation of the speech and is a core feature that reflects emotion and intonation fluctuations. The speech energy intensity represents the loudness of the speech and the energy distribution in the speech signal, which can reflect the intensity of the user's expression. The semantic structure feature node represents the semantic features extracted from the text data, including the distribution density of emotional keywords and the frequency of interrogative sentences. Emotional keywords represent the core words used to express user emotions, such as "satisfaction" and "dissatisfaction", while interrogative sentences represent the types of questions asked by users during interactions, such as "how" and "why". The interaction pattern feature node represents the user's behavioral patterns during the interaction, including a mapping table between question types and response strategies and the interval between dialogue turns. The mapping relationship table reflects the correspondence between the types of questions raised by users and the response strategies selected by the system, while the dialogue turn transition interval represents the changes in the response time between users and the system in multiple rounds of interaction.

[0167] When creating feature nodes, the system also creates result nodes for each event outcome data. Result nodes represent the actual outcome data obtained by the user after the interaction, including user satisfaction scores and complaint resolution success rates. User satisfaction scores are a quantitative user feedback metric that indicates user satisfaction with the interaction outcome, while complaint resolution success rates indicate whether the user's concerns were resolved after the interaction.

[0168] After the feature nodes and result nodes are created, the system constructs the association edges between these nodes based on the association between historical interaction features and event result data. The association edges between acoustic feature nodes and result nodes are configured based on the Pearson correlation coefficient and Spearman correlation coefficient in the acoustic feature association sub-model. The Pearson correlation coefficient represents the linear correlation between acoustic features (such as the fundamental frequency change curve) and user satisfaction, and the Spearman correlation coefficient represents the nonlinear correlation between acoustic features (such as speech energy intensity) and the success rate of complaint resolution. These coefficients are assigned to the association edges between acoustic feature nodes and result nodes in the form of weights. The larger the weight value, the stronger the correlation between the feature and the result.

[0169] When configuring the edges between semantic feature nodes and result nodes, the system uses conditional probability and mutual information from the semantic feature association submodel. Conditional probability represents the probability that user satisfaction reaches a certain level when a specific keyword (such as "satisfied") appears in the text. Mutual information represents the degree of information sharing between sentiment (such as positive or negative sentiment) and the success rate of complaint resolution. Conditional probability and mutual information are assigned to the edges between semantic feature nodes and result nodes in the form of confidence values. Higher confidence values ​​indicate a stronger correlation between the feature and the result.

[0170] When configuring the association edges between interaction pattern feature nodes and outcome nodes, the system uses the success rate and linear regression coefficient from the interaction pattern feature association submodel. The success rate indicates the effectiveness of a specific response strategy for a specific question type, while the linear regression coefficient represents the linear relationship between the interval between conversational turn transitions and user satisfaction. The success rate is assigned to the association edge as a path strength, indicating the effectiveness of the strategy for a specific question type, while the linear regression coefficient indicates the impact of user response time on the outcome.

[0171] After configuring all nodes and associated edges, the system stores the acoustic feature nodes, semantic feature nodes, interaction pattern feature nodes, result nodes, and the associated edges between them in a graph database. A graph database is a type of database designed specifically for processing node and edge relationships, enabling efficient storage and retrieval of complex node-edge-node structures. Through the structured storage of a graph database, the system enables efficient querying and dynamic updating of the knowledge base, ensuring that subsequent policy generation is based on the most relevant historical interaction features and association rules.

[0172] This embodiment converts the association between historical interaction features and event result data into a knowledge base structure, so that the system can efficiently store and manage complex feature-result association information in the form of a graph database. The acoustic feature nodes, semantic feature nodes, and interaction mode feature nodes in the knowledge base record the user's features in voice expression, text expression, and interaction mode respectively, and the result nodes represent the user feedback results associated with these features. Through dynamic configuration based on association edge weights, confidence, and path strength, the system can automatically select the features and strategies that best suit the current scenario in subsequent strategy generation. The weight of the acoustic feature represents the impact of voice expression on satisfaction and complaint resolution, the confidence of the semantic feature represents the impact of keywords and emotional tendencies on user feedback, and the path strength of the interaction mode feature represents the impact of response strategy and response time on service effectiveness. Through efficient query and dynamic update of the graph database, the system can quickly match the optimal strategy in real-time interaction and achieve personalized recommendations.

[0173] In one embodiment, the above step S70 includes:

[0174] S701, traversing the acoustic waveform feature nodes in the knowledge base, determining the cosine similarity between the current acoustic waveform feature and each historical acoustic waveform feature in the current interaction feature, and screening a set of acoustic feature candidate nodes whose cosine similarity is greater than a first threshold;

[0175] S702, traversing the semantic structure feature nodes in the knowledge base, determining the Jaccard similarity between the current semantic structure feature in the current interaction feature and each historical semantic structure feature, and selecting a set of semantic feature candidate nodes whose Jaccard similarity is greater than a second threshold;

[0176] S703, traversing the interaction pattern feature nodes in the knowledge base, determining the Euclidean distance between the current interaction pattern feature and each historical interaction pattern feature in the current interaction feature, and screening a set of interaction pattern feature candidate nodes whose Euclidean distance is less than a third threshold;

[0177] S704: performing weighted scoring on the acoustic feature candidate node set, the semantic feature candidate node set, and the interactive pattern feature candidate node set based on the associated edge weights of the acoustic feature associated submodel, the associated edge confidences of the semantic feature associated submodel, and the associated edge path strengths of the interactive pattern feature associated submodel in the knowledge base;

[0178] S705, fusing the weighted scores of each candidate node set according to a preset weight ratio to generate a comprehensive matching score;

[0179] S706 , globally sorting all candidate nodes according to the comprehensive matching scores, and screening out candidate nodes with comprehensive matching scores higher than a preset output threshold, to generate a matching result set.

[0180] In this embodiment, when matching the current interaction features with the historical interaction features in the knowledge base, the system first extracts the current acoustic waveform features, the current semantic structure features, and the current interaction mode features from the current interaction scene. The current acoustic waveform features include features extracted from the current user's voice data, such as the fundamental frequency change curve and the voice energy intensity. These features capture the user's voice expression style and emotional fluctuations. The current semantic structure features are derived from the semantic analysis results of the text data, such as the frequency of emotional keywords and question sentences expressed by the user. These features describe the user's text expression intention and semantic information. The current interaction mode features represent the user's behavior pattern in multiple rounds of interaction, such as conversation turns, response time, and answer strategy.

[0181] The system first traverses the acoustic waveform feature nodes in the knowledge base. For each acoustic waveform feature node, the system calculates the similarity between the current acoustic waveform feature and each historical acoustic waveform feature. Cosine similarity is used as the metric here. Cosine similarity is a similarity measure that measures the angle between two vectors and is suitable for measuring the similarity of acoustic waveform features. The closer the similarity value is to 1, the more similar the two features are. The system filters out all acoustic feature nodes whose cosine similarity is greater than the first threshold to form a set of acoustic feature candidate nodes.

[0182] For semantic structure features, the system traverses the semantic structure feature nodes in the knowledge base and calculates the similarity between the current semantic structure feature and each historical semantic structure feature one by one. Jaccard similarity is used here, which measures similarity by comparing the degree of overlap between two sets (such as keyword sets). The higher the Jaccard similarity, the greater the overlap between the current semantic structure feature and the historical semantic structure feature. The system filters out all semantic feature nodes with a Jaccard similarity greater than the second threshold to form a set of semantic feature candidate nodes.

[0183] For interaction pattern features, the system traverses the interaction pattern feature nodes in the knowledge base and calculates the similarity between the current interaction pattern feature and each historical interaction pattern feature. Euclidean distance is used as the metric. Euclidean distance is a method for calculating the straight-line distance between two vectors. The smaller the Euclidean distance, the more similar the two interaction pattern features are. The system then selects all interaction pattern feature nodes whose Euclidean distance is less than a third threshold to form a set of candidate interaction pattern feature nodes.

[0184] After obtaining the candidate node sets of acoustic, semantic, and interaction pattern features, the system performs weighted scoring on each candidate node set based on the predefined association edge parameters in the knowledge base. Specifically, the weighted score of the acoustic feature candidate node set is calculated based on the association edge weight in the acoustic feature association sub-model. The association edge weight represents the strength of the association between the acoustic feature and the event outcome data. The weighted score of the semantic feature candidate node set is calculated based on the association edge confidence in the semantic feature association sub-model. The confidence represents the reliability of the association between the semantic feature and the event outcome data. The weighted score of the interaction pattern feature candidate node set is calculated based on the association edge path strength in the interaction pattern feature association sub-model. The path strength represents the strategic effect between the interaction pattern feature and the event outcome data.

[0185] The system combines the weighted scores of the three candidate node sets according to a preset weight ratio to generate a comprehensive match score for each candidate node. The preset weight ratio is determined through historical data verification and can typically be adjusted based on the business scenario. For example, in voice customer service scenarios, the weight of acoustic features can be increased, while in text customer service, the weight of semantic features can be increased. The system then globally sorts all candidate nodes based on the comprehensive match score and selects candidate nodes with a comprehensive match score above the preset output threshold to generate the final set of matching results. These high-scoring candidate nodes represent the historical interaction features most similar to the current interaction features, and their associated event outcome data will serve as the basis for subsequent policy generation.

[0186] In this embodiment, by performing multi-dimensional matching of current interaction features with historical interaction features in the knowledge base, the system can achieve comprehensive matching of acoustics, semantics, and interaction patterns. Acoustic features capture the emotions and intonation of user voice expressions through cosine similarity calculation, semantic features capture the keywords and emotions expressed in text through Jaccard similarity, and interaction pattern features capture the rhythm and response methods of user interaction behaviors through Euclidean distance calculation. The weighted scoring mechanism ensures that the matching results of the three features are reasonably integrated according to their importance in specific scenarios, and the generated comprehensive matching score can accurately reflect the matching effect of candidate nodes. Through global sorting and threshold screening, the system ensures that only candidate nodes with high relevance are included in the final matching result set. This multi-dimensional matching method can not only improve the accuracy of matching results, but also provide the best strategy reference for the current interaction scenario through successful cases in historical data, thereby realizing personalized recommendation strategies.

[0187] In one embodiment, the above step S80 includes:

[0188] S801, for each candidate node in the matching result set, extracting the user satisfaction score and complaint resolution success rate in the associated event result data;

[0189] S802, determining a comprehensive success rate indicator based on the user satisfaction score and the complaint resolution success rate;

[0190] S803, screening out candidate nodes whose comprehensive success rate index is greater than a preset success rate threshold, and generating a valid strategy node set;

[0191] S804: Generate a strategy combination including a response speech template, voice intonation adjustment suggestions, and dialogue rhythm control parameters based on the interaction pattern characteristics in the set of valid strategy nodes;

[0192] S805, inputting the strategy combination into the interactive system for execution, and recording the actual user satisfaction score and actual complaint resolution success rate obtained after execution;

[0193] S806, when the actual user satisfaction score is equal to or greater than the preset satisfaction threshold, and the actual complaint resolution success rate is equal to or greater than the preset resolution threshold, the current interaction characteristics and event result data corresponding to the strategy combination are added to the knowledge base.

[0194] In this embodiment, when generating a policy result for the current interaction scenario based on the matching results, the associated event result data is first extracted from each candidate node in the matching result set. This event result data typically includes a user satisfaction score and a complaint resolution success rate. The user satisfaction score is a rating that measures the user's satisfaction with the interaction result, usually expressed on a percentage or five-point scale. For example, a score of 80 indicates overall user satisfaction, and a score of 95 indicates high satisfaction. The complaint resolution success rate indicates the proportion of user complaints that have been effectively resolved in a specific scenario. For example, 100% indicates that all complaints have been resolved, while 70% indicates that some complaints have not been successfully resolved.

[0195] The system calculates a comprehensive success rate indicator for each candidate node, which is used to quantify the potential effect of each candidate node in the current interaction scenario. The comprehensive success rate indicator is calculated based on the weighted sum of the user satisfaction score and the complaint resolution success rate. Specifically, the system will set different weights for the user satisfaction score and the complaint resolution success rate based on business needs. For example, the weight of the user satisfaction score can be set to 0.7, and the weight of the complaint resolution success rate can be set to 0.3. This weighting method ensures that the focus of the strategy effect evaluation can be flexibly adjusted in different application scenarios. For example, in scenarios where user experience is prioritized, the weight of the user satisfaction score can be increased, while in scenarios where problem-solving efficiency is prioritized, the weight of the complaint resolution success rate can be increased.

[0196] The system compares the calculated overall success rate with a preset success rate threshold, selecting candidate nodes with a success rate exceeding the threshold and generating a set of valid policy nodes. The preset success rate threshold is determined through historical data validation; for example, it can be set at 85% for customer service in the financial sector or 90% for online consultations in healthcare. These thresholds ensure that the system only selects policy nodes with high success rates, thereby improving policy execution effectiveness.

[0197] After obtaining a set of valid strategy nodes, the system generates a strategy combination based on the interaction pattern characteristics within these nodes. This strategy combination is automatically generated based on historical interaction experience within the valid strategy nodes and includes response templates, voice intonation adjustment suggestions, and conversation pacing control parameters. Response templates are automatically generated based on keywords and sentence structures derived from semantic features. For example, in the financial sector, this might be, "Your credit limit has met the requirements for an increase. Would you like to learn more?" In the healthcare sector, this might be, "Your symptoms may be related to anxiety. I recommend deep breathing and relaxation. Do you need further psychological counseling?" Voice intonation adjustment suggestions are based on the speech energy intensity and fundamental frequency variation curves derived from acoustic features. For example, for a stressed user, the voice agent might be advised to reduce the speech speed and intensity, while for a positive user, the intonation and speed might be increased. Conversation pacing control parameters are set based on the number of conversation turns and response time derived from interaction pattern characteristics. For example, in high-frequency interaction scenarios, the response time might be shortened, while in low-frequency interaction scenarios, the response time might be appropriately extended.

[0198] The system inputs the generated strategy combinations into the interactive system for execution, and during this process, it records user feedback in real time, including actual user satisfaction scores and complaint resolution success rates. User satisfaction scores are typically derived through user ratings or voice recognition sentiment analysis, while complaint resolution success rates are determined based on whether users subsequently file complaints or appeals. The system compares these actual performance data against pre-set thresholds, such as whether actual user satisfaction scores reach 90 or above, and whether actual complaint resolution success rates reach above 95%.

[0199] When the actual user satisfaction score reaches or exceeds the preset satisfaction threshold, and the actual complaint resolution success rate reaches or exceeds the preset resolution threshold, the system adds the current interaction characteristics and event outcome data corresponding to this strategy combination to the knowledge base. This dynamic update mechanism ensures that the knowledge base continuously absorbs strategies with high success rates and empirical data, forming an adaptive recommendation strategy optimization system. Through this mechanism, the system can automatically learn from real interactions and gradually build a more efficient and accurate strategy library.

[0200] Example: In a customer service application in the financial sector, the system first collects historical voice and text data from historical interaction scenarios, including call recordings and chat logs between customers and customer service representatives. Simultaneously, the system extracts interaction behavior metrics from the interaction logs, such as customer response time and issue resolution cycle time. A unique session identifier is generated for each interaction session to ensure data independence and traceability. The unique session identifier is used to link historical voice and text data with interaction behavior metrics. The system further extracts event outcome data from the historical database, such as user satisfaction scores and complaint resolution success rates. All data is aligned based on timestamps to ensure synchronization across different data sources, ultimately integrating them into historical interaction data.

[0201] The system annotates the collected historical interaction data based on predefined event outcome indicators (such as user satisfaction scores greater than or equal to 80 points for success and less than 80 points for failure). Successful interaction data is marked as positive samples, and failed interaction data is marked as negative samples. For the annotated historical interaction data, the system performs frame processing on the voice data to extract acoustic waveform features, such as the fundamental frequency change curve and voice energy intensity; performs semantic parsing on the text data to generate semantic structure features, such as the distribution of emotional keywords and the frequency of interrogative sentences; and analyzes and generates interaction pattern features from the interaction behavior data, such as the mapping relationship table between question type and response strategy and the interval between dialogue turn transitions. These feature data are normalized and feature vector encoded to ultimately form a set of historical interaction features.

[0202] Next, the system performs correlation analysis on these historical interaction features, quantifying their association with event outcome data. The linear correlation between the fundamental frequency variation curve in acoustic features and user satisfaction scores is quantified using the Pearson correlation coefficient, while the nonlinear correlation between speech energy intensity and complaint resolution success rate is quantified using the Spearman correlation coefficient. The correlation between high-frequency keywords in semantic features and user satisfaction scores is calculated using conditional probability, and the correlation between sentiment and complaint resolution success rate is calculated using mutual information. The correlation between question type and response strategy success rate in interaction pattern features is determined through statistical analysis, while the correlation between conversation turn transition intervals and user satisfaction is determined through linear regression analysis. Based on these correlations, the system constructs correlation sub-models for acoustic features, semantic features, and interaction pattern features, forming a multi-dimensional correlation network between features and outcomes.

[0203] The system constructs a knowledge base based on these feature associations. Each historical interaction feature is converted into a node, including acoustic waveform feature nodes, semantic structure feature nodes, and interaction pattern feature nodes. Each event outcome data point is also converted into an outcome node. Based on the correlation coefficients and conditional probabilities in the association sub-model, the system connects the associations between feature nodes and outcome nodes using weighted edges. This graph-structured knowledge base is stored in a graph database and supports subsequent rapid retrieval and real-time matching.

[0204] When the system receives input data from a current customer, such as a voice inquiry asking "How can I increase my credit limit?", it extracts the current acoustic waveform features, semantic structure features, and interaction pattern features from the input data. The system traverses the historical acoustic waveform feature nodes in the knowledge base, calculates the cosine similarity between the current acoustic waveform feature and each historical acoustic feature, and selects a set of acoustic feature candidate nodes whose cosine similarity exceeds a set threshold. Similarly, the system calculates the Jaccard similarity between the current semantic structure feature and historical semantic features, and selects a set of semantic feature candidate nodes whose cosine similarity exceeds a threshold. For interaction pattern features, the system calculates the Euclidean distance between the current feature and historical features, and selects a set of interaction pattern feature candidate nodes whose Euclidean distance is less than a threshold. Each candidate node is weighted and scored based on the associated edge weights, confidence, and path strength in the knowledge base. The scoring results are combined according to a preset weight ratio to generate a comprehensive matching score. The system sorts all candidate nodes from high to low according to their comprehensive matching scores, selecting those with scores above the output threshold to generate a matching result set.

[0205] After generating matching results, the system extracts event outcome data from each candidate node, including user satisfaction scores and complaint resolution success rates. Based on this data, a comprehensive success rate metric is calculated. For example, the user satisfaction score and complaint resolution success rate are multiplied by weights of 0.7 and 0.3, respectively, and the sum is calculated. The system selects candidate nodes whose comprehensive success rate metrics exceed a preset threshold and generates a set of valid strategy nodes. Based on the interaction pattern characteristics within these strategy nodes, the system generates strategy combinations, including response templates (such as "Your credit limit has met the requirements for an increase. Would you like to increase it immediately?"), voice tone adjustment suggestions (such as a gentle and slow tone), and conversation pacing control parameters (such as a quick response). These strategy combinations are input into the real-time interaction system, which records the actual results during customer interactions, such as actual user satisfaction scores and complaint resolution success rates. When these actual results reach a preset threshold, the system dynamically updates the strategy combination, its corresponding interaction characteristics, and event outcome data to the knowledge base.

[0206] In the healthcare field, when a patient describes themselves by voice, "I've been having frequent headaches and poor sleep lately," the system captures the voice data and analyzes speech waveform features, such as voice energy and fundamental frequency variations. It also extracts semantic structural features, such as the high frequency of the keywords "headache" and "sleep," and the temporal expression reflected by the phrase "often lately." The system then identifies the question type as "symptom consultation" and the response strategy as "health advice" from the interaction pattern. These features are then matched against historical features in the knowledge base. Nodes with high phonetic similarity indicate a correlation between the symptom "headache" and "tension," while nodes with high semantic similarity indicate a correlation between "poor sleep" and "anxiety." The system then generates a strategy combination, such as the response line "Based on your description, you may be experiencing anxiety symptoms. I recommend staying relaxed. Do you require further psychological counseling?" The system also recommends that the voice customer service representative use a gentle tone and slow down their speech. During this implementation, patient satisfaction ratings reached 92 points, with no subsequent complaints. Because both the satisfaction rating and complaint resolution success rate reached pre-set thresholds, the system dynamically added this strategy combination and the corresponding interaction features to the knowledge base.

[0207] This embodiment generates the strategy results of the current interaction scenario based on the matching results, and the system can further optimize the strategy effect based on the matching results. The comprehensive success rate index is a weighted combination of the user satisfaction score and the complaint resolution success rate, so that the system can flexibly respond to the strategy effect evaluation in different business scenarios. In the financial field, the user satisfaction weight can be increased to ensure a high user experience; in the medical and health field, the complaint resolution success rate weight can be increased to ensure a high problem solving rate. The strategy combination realizes intelligent and personalized recommendations through multi-dimensional strategy generation of voice, text and interaction modes. The dynamic update mechanism ensures that the system can absorb high-success rate strategies in real time, incorporate effective strategies in real scenarios into the knowledge base, and form a self-optimizing strategy library. Over time, the effectiveness of the strategies recommended by the system will continue to improve.

[0208] In one embodiment, a data analysis-based policy generation device is provided, and the data analysis-based policy generation device corresponds one-to-one to the data analysis-based policy generation method in the above embodiment. Figure 3 , Figure 3 This is a functional module diagram of a preferred embodiment of a data analysis-based strategy generation device according to the present invention. It includes a data acquisition module 10, an event annotation module 20, a feature extraction module 30, a feature association analysis module 40, a knowledge base construction module 50, a real-time data access module 60, a feature matching module 70, and a strategy generation module 80. Each functional module is described in detail below:

[0209] The data collection module 10 is used to collect historical interaction data and corresponding event result data in historical interaction scenarios;

[0210] An event marking module 20 is configured to mark the historical interaction data as successful or failed according to predefined event result indicators;

[0211] A feature extraction module 30 is used to extract historical interaction features from the annotated historical interaction data;

[0212] A feature association analysis module 40 is configured to perform association analysis on the historical interaction features to generate an association relationship between the historical interaction features and event result data;

[0213] A knowledge base construction module 50 is used to construct a knowledge base based on the association between the historical interaction features and the event result data;

[0214] A real-time data access module 60 is configured to receive input data of a current interaction scenario and extract current interaction features from the input data;

[0215] A feature matching module 70 is configured to match the current interaction feature with the historical interaction features in the knowledge base to generate a matching result;

[0216] The strategy generating module 80 is configured to generate a strategy result for the current interaction scenario based on the matching result.

[0217] In one embodiment, the data acquisition module 10 is specifically configured to:

[0218] Capturing historical voice data and historical text data in historical interaction scenarios, and generating a unique session identifier for each interaction session;

[0219] Extract historical interaction behavior indicator data from historical interaction logs;

[0220] Associating the historical voice data, historical text data, and historical interaction behavior indicator data through the unique session identifier;

[0221] Acquire historical event result data corresponding to the unique session identifier from a historical database;

[0222] Performing timestamp alignment processing on the historical voice data, historical text data, historical interaction behavior indicator data, and historical event result data;

[0223] The historical voice data, historical text data and historical interaction behavior indicator data after timestamp alignment are integrated into historical interaction data.

[0224] In one embodiment, the feature extraction module 30 is specifically configured to:

[0225] Performing frame processing on the voice data in the annotated historical interaction data, and extracting acoustic waveform features from the frame-processed voice data;

[0226] Perform semantic parsing on the text data in the annotated historical interaction data to generate semantic structure features;

[0227] Analyze the interaction behavior data in the annotated historical interaction data to generate interaction pattern features;

[0228] The acoustic waveform features, semantic structure features and interaction pattern features are normalized and feature vector encoded to generate a historical interaction feature set.

[0229] In one embodiment, the feature association analysis module 40 is specifically configured to:

[0230] Determining the Pearson correlation coefficient between the fundamental frequency change curve in the acoustic feature subset in the historical interaction features and the user satisfaction score in the event result data;

[0231] Determining the Spearman correlation coefficient between the speech energy intensity in the acoustic feature subset in the historical interaction features and the complaint resolution success rate in the event outcome data;

[0232] Determining the conditional probability of high-frequency keywords in the semantic feature subset and user satisfaction scores in the event result data in the historical interaction features;

[0233] Determining the mutual information between the sentiment tendency in the semantic feature subset in the historical interaction features and the complaint resolution success rate in the event outcome data;

[0234] Determine the success rate of each strategy in the mapping relationship table between question types and answer strategies in the interaction pattern feature subset in the historical interaction features;

[0235] Determining a linear regression coefficient between a conversation turn transition interval within an interaction pattern feature subset in the historical interaction features and a user satisfaction score in the event outcome data;

[0236] Based on the Pearson correlation coefficient and the Spearman correlation coefficient, a sub-model of the acoustic feature association between the acoustic feature subset and the event result data is constructed;

[0237] Based on the conditional probability and mutual information, a sub-model for associating the semantic feature subset with the semantic feature of the event outcome data is constructed;

[0238] Based on the success rate and the linear regression coefficient, a sub-model of the association between the interaction pattern feature subset and the event outcome data is constructed.

[0239] In one embodiment, the knowledge base construction module 50 is specifically configured to:

[0240] Creating corresponding acoustic waveform feature nodes, semantic structure feature nodes, and interaction pattern feature nodes for each acoustic waveform feature, semantic structure feature, and interaction pattern feature in the historical interaction features;

[0241] Create a result node for each event result data;

[0242] configuring the association edge weight between the acoustic feature node and the result node according to the Pearson correlation coefficient and the Spearman correlation coefficient in the acoustic feature association sub-model in the association relationship;

[0243] configuring the confidence of the association edge between the semantic feature node and the result node according to the conditional probability and mutual information in the semantic feature association sub-model in the association relationship;

[0244] configuring the association edge path strength between the interaction pattern feature node and the result node according to the success rate and the linear regression coefficient in the interaction pattern feature association sub-model in the association relationship;

[0245] The acoustic feature nodes, semantic feature nodes, interaction mode feature nodes, result nodes and associated edges are stored in a graph database to complete the construction of the knowledge base.

[0246] In one embodiment, the feature matching module 70 is specifically configured to:

[0247] Traversing the acoustic waveform feature nodes in the knowledge base, determining the cosine similarity between the current acoustic waveform feature and each historical acoustic waveform feature in the current interaction feature, and screening a set of acoustic feature candidate nodes whose cosine similarity is greater than a first threshold;

[0248] Traversing the semantic structure feature nodes in the knowledge base, determining the Jaccard similarity between the current semantic structure feature and each historical semantic structure feature in the current interaction feature, and screening a set of semantic feature candidate nodes whose Jaccard similarity is greater than a second threshold;

[0249] Traversing the interaction pattern feature nodes in the knowledge base, determining the Euclidean distance between the current interaction pattern feature and each historical interaction pattern feature in the current interaction feature, and screening a set of interaction pattern feature candidate nodes whose Euclidean distance is less than a third threshold;

[0250] Performing weighted scoring on the acoustic feature candidate node set, the semantic feature candidate node set, and the interactive pattern feature candidate node set based on the associated edge weights of the acoustic feature associated submodel, the associated edge confidences of the semantic feature associated submodel, and the associated edge path strengths of the interactive pattern feature associated submodel in the knowledge base;

[0251] The weighted scores of each candidate node set are combined according to the preset weight ratio to generate a comprehensive matching score;

[0252] All candidate nodes are globally sorted according to the comprehensive matching scores, and candidate nodes with comprehensive matching scores higher than a preset output threshold are screened out to generate a matching result set.

[0253] In one embodiment, the policy generation module 80 is specifically configured to:

[0254] For each candidate node in the matching result set, extract the user satisfaction score and complaint resolution success rate in the associated event result data;

[0255] Determine a comprehensive success rate indicator based on the user satisfaction score and complaint resolution success rate;

[0256] Filter out candidate nodes whose comprehensive success rate index is greater than the preset success rate threshold and generate a set of valid strategy nodes;

[0257] Generate a strategy combination including a response script template, voice intonation adjustment suggestions, and dialogue rhythm control parameters based on the interaction pattern characteristics in the effective strategy node set;

[0258] Input the strategy combination into the interactive system for execution, and record the actual user satisfaction score and actual complaint resolution success rate obtained after execution;

[0259] When the actual user satisfaction score is equal to or greater than the preset satisfaction threshold, and the actual complaint resolution success rate is equal to or greater than the preset resolution threshold, the current interaction characteristics and event result data corresponding to the strategy combination are added to the knowledge base.

[0260] In one embodiment, a determination device is provided. The determination device may be a server, and its internal structure diagram may be as follows: Figure 4 As shown. The determination machine device includes a processor, a memory, a network interface and a database connected via a system bus. Among them, the processor of the determination machine device is used to provide determination and control capabilities. The memory of the determination machine device includes a non-volatile and / or volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a determination machine program and a database. The internal memory provides an environment for the operation of the operating system and the determination machine program in the non-volatile storage medium. The network interface of the determination machine device is used to communicate with an external user terminal through a network connection. When the determination machine program is executed by the processor, it realizes a function or step on the service side of a policy generation method based on data analysis.

[0261] In one embodiment, a determination device is provided. The determination device may be a user terminal, and its internal structure diagram may be as follows: Figure 5As shown. The determination machine device includes a processor, memory, network interface, display screen and input device connected via a system bus. Among them, the processor of the determination machine device is used to provide determination and control capabilities. The memory of the determination machine device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a determination machine program. The internal memory provides an environment for the operation of the operating system and the determination machine program in the non-volatile storage medium. The network interface of the determination machine device is used to communicate with an external server via a network connection. When the determination machine program is executed by the processor, it realizes the functions or steps on the user side of a policy generation method based on data analysis

[0262] In one embodiment, a determination machine device is provided, including a memory, a processor, and a determination machine program stored in the memory and executable on the processor. When the processor executes the determination machine program, the following steps are implemented:

[0263] Collect historical interaction data and corresponding event result data in historical interaction scenarios;

[0264] Based on the event result data, marking the historical interaction data as successful or failed according to predefined event result indicators;

[0265] Extract historical interaction features from the annotated historical interaction data;

[0266] Performing correlation analysis on the historical interaction features to generate correlation relationships between the historical interaction features and event result data;

[0267] Building a knowledge base based on the correlation between the historical interaction features and the event result data;

[0268] receiving input data of a current interaction scenario, and extracting current interaction features from the input data;

[0269] Matching the current interaction feature with the historical interaction features in the knowledge base to generate a matching result;

[0270] Based on the matching result, a strategy result for the current interaction scenario is generated.

[0271] In one embodiment, a determination machine readable storage medium is provided, on which a determination machine program is stored. When the determination machine program is executed by a processor, the following steps are implemented:

[0272] Collect historical interaction data and corresponding event result data in historical interaction scenarios;

[0273] Based on the event result data, marking the historical interaction data as successful or failed according to predefined event result indicators;

[0274] Extract historical interaction features from the annotated historical interaction data;

[0275] Performing correlation analysis on the historical interaction features to generate correlation relationships between the historical interaction features and event result data;

[0276] Building a knowledge base based on the correlation between the historical interaction features and the event result data;

[0277] receiving input data of a current interaction scenario, and extracting current interaction features from the input data;

[0278] Matching the current interaction feature with the historical interaction features in the knowledge base to generate a matching result;

[0279] Based on the matching result, a strategy result for the current interaction scenario is generated.

[0280] It should be noted that the above functions or steps that can be implemented by the machine-readable storage medium or the machine device can be referred to the relevant descriptions on the server side and the user side in the aforementioned method embodiment. To avoid repetition, they will not be described one by one here.

[0281] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a determination machine program, and the determination machine program can be stored in a non-volatile determination machine-readable storage medium. When the determination machine program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0282] Those skilled in the art will clearly understand that for the sake of convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0283] It should be noted that if any software tools or components other than those of the Company appear in the embodiments of this application, they are merely for illustration and do not represent actual use. The above embodiments are intended only to illustrate the technical solutions of the present invention, not to limit them. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some of the technical features therein with equivalents. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the scope of protection of the present invention.

Claims

1. A strategy generation method based on data analysis, characterized in that: The following steps are involved: Collect historical interaction data and corresponding event result data in historical interaction scenarios; Based on the event result data, marking the historical interaction data as successful or failed according to predefined event result indicators; Extract historical interaction features from the annotated historical interaction data; Performing correlation analysis on the historical interaction features to generate correlation relationships between the historical interaction features and event result data; Building a knowledge base based on the correlation between the historical interaction features and the event result data; receiving input data of a current interaction scenario, and extracting current interaction features from the input data; Matching the current interaction feature with the historical interaction features in the knowledge base to generate a matching result; Based on the matching result, a strategy result for the current interaction scenario is generated.

2. The data analysis-based strategy generation method according to claim 1, wherein: Collect historical interaction data and corresponding event result data in historical interaction scenarios, including: Capturing historical voice data and historical text data in historical interaction scenarios, and generating a unique session identifier for each interaction session; Extract historical interaction behavior indicator data from historical interaction logs; Associating the historical voice data, historical text data, and historical interaction behavior indicator data through the unique session identifier; Acquire historical event result data corresponding to the unique session identifier from a historical database; Performing timestamp alignment processing on the historical voice data, historical text data, historical interaction behavior indicator data, and historical event result data; The historical voice data, historical text data and historical interaction behavior indicator data after timestamp alignment are integrated into historical interaction data.

3. The data analysis-based strategy generation method according to claim 1, wherein: Extract historical interaction features from the annotated historical interaction data, including: Performing frame processing on the voice data in the annotated historical interaction data, and extracting acoustic waveform features from the frame-processed voice data; Perform semantic parsing on the text data in the annotated historical interaction data to generate semantic structure features; Analyze the interaction behavior data in the annotated historical interaction data to generate interaction pattern features; The acoustic waveform features, semantic structure features and interaction pattern features are normalized and feature vector encoded to generate a historical interaction feature set.

4. The data analysis-based strategy generation method according to claim 1, wherein: Performing correlation analysis on the historical interaction features to generate correlation relationships between the historical interaction features and event result data includes: Determining the Pearson correlation coefficient between the fundamental frequency change curve in the acoustic feature subset in the historical interaction features and the user satisfaction score in the event result data; Determining the Spearman correlation coefficient between the speech energy intensity in the acoustic feature subset in the historical interaction features and the complaint resolution success rate in the event outcome data; Determining the conditional probability of high-frequency keywords in the semantic feature subset and user satisfaction scores in the event result data in the historical interaction features; Determining the mutual information between the sentiment tendency in the semantic feature subset in the historical interaction features and the complaint resolution success rate in the event outcome data; Determine the success rate of each strategy in the mapping relationship table between question types and answer strategies in the interaction pattern feature subset in the historical interaction features; Determining a linear regression coefficient between a conversation turn transition interval within an interaction pattern feature subset in the historical interaction features and a user satisfaction score in the event outcome data; Based on the Pearson correlation coefficient and the Spearman correlation coefficient, a sub-model of the acoustic feature association between the acoustic feature subset and the event result data is constructed; Based on the conditional probability and mutual information, a sub-model for associating the semantic feature subset with the semantic feature of the event outcome data is constructed; Based on the success rate and the linear regression coefficient, a sub-model of the association between the interaction pattern feature subset and the event outcome data is constructed.

5. The data analysis-based strategy generation method according to claim 1, wherein: Based on the correlation between the historical interaction features and the event result data, a knowledge base is constructed, including: Creating corresponding acoustic waveform feature nodes, semantic structure feature nodes, and interaction pattern feature nodes for each acoustic waveform feature, semantic structure feature, and interaction pattern feature in the historical interaction features; Create a result node for each event result data; configuring the association edge weight between the acoustic feature node and the result node according to the Pearson correlation coefficient and the Spearman correlation coefficient in the acoustic feature association sub-model in the association relationship; configuring the confidence of the association edge between the semantic feature node and the result node according to the conditional probability and mutual information in the semantic feature association sub-model in the association relationship; configuring the association edge path strength between the interaction pattern feature node and the result node according to the success rate and the linear regression coefficient in the interaction pattern feature association sub-model in the association relationship; The acoustic feature nodes, semantic feature nodes, interaction mode feature nodes, result nodes and associated edges are stored in a graph database to complete the construction of the knowledge base.

6. The data analysis-based strategy generation method according to claim 1, wherein: Matching the current interaction feature with the historical interaction feature in the knowledge base to generate a matching result includes: Traversing the acoustic waveform feature nodes in the knowledge base, determining the cosine similarity between the current acoustic waveform feature and each historical acoustic waveform feature in the current interaction feature, and screening a set of acoustic feature candidate nodes whose cosine similarity is greater than a first threshold; Traversing the semantic structure feature nodes in the knowledge base, determining the Jaccard similarity between the current semantic structure feature and each historical semantic structure feature in the current interaction feature, and screening a set of semantic feature candidate nodes whose Jaccard similarity is greater than a second threshold; Traversing the interaction pattern feature nodes in the knowledge base, determining the Euclidean distance between the current interaction pattern feature and each historical interaction pattern feature in the current interaction feature, and screening a set of interaction pattern feature candidate nodes whose Euclidean distance is less than a third threshold; Performing weighted scoring on the acoustic feature candidate node set, the semantic feature candidate node set, and the interactive pattern feature candidate node set based on the associated edge weights of the acoustic feature associated submodel, the associated edge confidences of the semantic feature associated submodel, and the associated edge path strengths of the interactive pattern feature associated submodel in the knowledge base; The weighted scores of each candidate node set are combined according to the preset weight ratio to generate a comprehensive matching score; All candidate nodes are globally sorted according to the comprehensive matching scores, and candidate nodes with comprehensive matching scores higher than a preset output threshold are screened out to generate a matching result set.

7. The data analysis-based strategy generation method according to claim 1, wherein: Generating a strategy result for the current interaction scenario based on the matching result includes: For each candidate node in the matching result set, extract the user satisfaction score and complaint resolution success rate in the associated event result data; Determine a comprehensive success rate indicator based on the user satisfaction score and complaint resolution success rate; Filter out candidate nodes whose comprehensive success rate index is greater than the preset success rate threshold and generate a set of valid strategy nodes; Generate a strategy combination including a response script template, voice intonation adjustment suggestions, and dialogue rhythm control parameters based on the interaction pattern characteristics in the effective strategy node set; Input the strategy combination into the interactive system for execution, and record the actual user satisfaction score and actual complaint resolution success rate obtained after execution; When the actual user satisfaction score is equal to or greater than the preset satisfaction threshold, and the actual complaint resolution success rate is equal to or greater than the preset resolution threshold, the current interaction characteristics and event result data corresponding to the strategy combination are added to the knowledge base.

8. A strategy generation device based on data analysis, characterized in that: The strategy generation device based on data analysis includes: The data collection module is used to collect historical interaction data and corresponding event result data in historical interaction scenarios; An event marking module, configured to mark the historical interaction data as successful or failed according to predefined event result indicators; Feature extraction module, used to extract historical interaction features from the annotated historical interaction data; A feature association analysis module, configured to perform association analysis on the historical interaction features and generate an association relationship between the historical interaction features and event result data; A knowledge base construction module, configured to construct a knowledge base based on the association between the historical interaction features and the event result data; A real-time data access module is used to receive input data of the current interaction scene and extract current interaction features from the input data; A feature matching module, configured to match the current interaction feature with the historical interaction features in the knowledge base to generate a matching result; A strategy generation module is used to generate a strategy result for the current interaction scenario based on the matching result.

9. A determination device, characterized in that: The determination machine device includes a memory, a processor, and a data analysis-based policy generation program stored in the memory and capable of running on the processor. When the data analysis-based policy generation program is executed by the processor, the steps of the data analysis-based policy generation method as described in any one of claims 1-7 are implemented.

10. A machine-readable storage medium, characterized in that: The storage medium stores a data analysis-based policy generation program, which, when executed by a processor, implements the steps of the data analysis-based policy generation method according to any one of claims 1 to 7.

Images