Human-computer interaction processing method and device based on digital service and storage medium

Abstract

The embodiment of the invention provides a human-computer interaction processing method and device based on digital service and a storage medium, and the method comprises the steps: carrying out the multi-mode joint analysis of text semantics, voice intonation and interaction behavior sequence features through an intention analysis model after receiving a user interaction request, obtaining a structured user intention representation containing an intention type, a demand scene and an expected output format; based on the representation, retrieving a historical interaction database and a user preference feature library, and fusing retrieval results to generate personalized enhanced features; the features are input into an intention-driven response generation model, and digital service response content is constructed and pushed in combination with a service content template library and a dynamic information filling interface; and finally, the user preference feature library is updated according to the user service comments, so that the accuracy of user intention recognition and the individuation degree of response content are improved, and the fitness between the naturalness of man-machine interaction and the user experience is enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer interaction, more particularly, to a human-computer interaction processing method and device based on digital service and a storage medium. BACKGROUND

[0002] With the popularity of digital services, human-computer interaction technology has been widely used in the fields of finance, e-commerce, customer service, etc. Users' demand for the accuracy, personalization and real-time of interaction is increasing. In the prior art, the analysis of user interaction requests depends on a single text mode, and the user's intention is identified through keyword matching or simple semantic analysis. Some technologies will assist in judgment by combining the speed and tone of voice data, but they do not systematically integrate the sequence characteristics of interaction behaviors (such as the operation sequence and stay time of the user), resulting in an incomplete capture of the user's deep needs. In the response generation stage, the existing technology usually fills in fixed information based on a preset template, or generates response content only by referring to the user's historical interaction records, without effectively associating real-time preference data from the user preference feature library, making it difficult to meet the user's personalized needs. At the same time, the existing technology lacks effective use of feedback data after the service response, and cannot dynamically update the preference feature library according to the user's evaluation, resulting in insufficient optimization and iteration capabilities of the interaction system.

[0003] Therefore, the prior art has the problem of low accuracy of user intention recognition and low degree of personalization of response content. SUMMARY

[0004] Therefore, the present application provides a human-computer interaction processing method and device based on digital service and a storage medium.

[0005] The present application provides a human-computer interaction processing method based on digital service, applied to a human-computer interaction processing device, the method comprising: receiving a user interaction request sent by a user terminal device, performing multi-modal joint analysis on the user interaction request through an intention analysis model, disassembling the text semantics, voice tone and sequence characteristics of the interaction behavior in the user interaction request, and integrating to obtain a structured user intention representation containing intention type, demand scenario and expected output format; based on the intention type in the structured user intention representation, matching the retrieval dimensions of the historical interaction database and the retrieval dimensions of the user preference feature library according to the demand scenario, respectively performing retrieval operations, obtaining a first retrieval result corresponding to the historical interaction database and a second retrieval result corresponding to the user preference feature library; By associating the historical interaction intent matching degree and historical response satisfaction degree in the first search result with the preference feature similarity degree and preference update timestamp in the second search result, the first search result and the second search result are fused through intent-preference association rules to generate personalized enhanced features that represent the user's current intent. The personalized enhancement features are input into the intent-driven response generation model. Through the service content template library and dynamic information filling interface in the response generation model, combined with the expected output format in the structured user intent representation, digital service response content containing service content structure, information presentation order and interactive guidance options is constructed in real time. The digital service response content is pushed to the user terminal device, and the service comments returned by the user terminal device for the digital service response content are obtained. The intent matching evaluation and content satisfaction evaluation in the service comments are extracted. Based on the service content structure and information presentation order in the digital service response content, the preference feature similarity weight and preference update timestamp in the user preference feature library are updated.

[0006] This invention also provides a human-computer interaction processing device, including: a memory and a processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions. When the computer programs / instructions are executed by the processor, they implement the above-described human-computer interaction processing method based on digital services.

[0007] This invention also provides a computer storage medium storing a computer program / instruction, which, when executed by a processor, implements the aforementioned human-computer interaction processing method based on digital services.

[0008] This invention employs an intent parsing model to perform multi-modal joint analysis of the text semantics, voice tone, and interactive behavior sequence features of user interaction requests. This analysis integrates these features to form a structured user intent representation that includes intent type, demand scenario, and expected output format. This overcomes the limitations of single-modal parsing, achieving a precise and comprehensive characterization of user intent. Furthermore, based on this structured user intent representation, historical interaction databases and user preference feature libraries are retrieved separately. Personalized enhanced features are generated by fusing historical interaction intent matching degree, historical response satisfaction, preference feature similarity, and preference update timestamps through intent-preference association rules. This effectively links historical user interaction patterns with real-time preference needs, solving the problem of insufficient response personalization caused by the separation of historical data and preference features. The personalized enhanced features are then input into an intent-driven response generation model, which, combined with a service content template library and dynamic information filling interface, constructs digital service response content in real time, ensuring the structural rationality, real-time information, and format adaptability of the response content. Finally, intent matching evaluation and content satisfaction evaluation are extracted through service reviews, which are used to update the preference feature similarity weights and preference update timestamps in the user preference feature library. This improves the accuracy, personalization level, and continuous optimization capabilities of human-computer interaction, significantly enhancing user experience and increasing service response efficiency. Attached Figure Description

[0009] The accompanying drawings, which are incorporated herein by reference and form part of the embodiments of the invention, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0010] Figure 1 This is a flowchart illustrating the steps of a human-computer interaction processing method based on digital services, provided in an embodiment of the present invention.

[0011] Figure 2 This is a schematic diagram of the functional modules of a human-computer interaction processing device provided in an embodiment of the present invention.

[0012] Figure 3 This is a structural block diagram of a human-computer interaction processing device provided in an embodiment of the present invention. Detailed Implementation

[0013] Numerous specific details are set forth in the following description to provide a thorough understanding of the embodiments of the invention. However, the embodiments of the invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the embodiments of the invention. Therefore, the embodiments of the invention are not limited to the specific implementations disclosed below.

[0014] The terminology used in one or more embodiments of the present invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of the present invention. The singular forms “a,” “the,” and “the” used in one or more embodiments of the present invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” used in one or more embodiments of the present invention refers to and includes any or all possible combinations of one or more associated listed items.

[0015] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of the present invention, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of the present invention, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0016] See Figure 1 , Figure 1 The flowchart of a human-computer interaction processing method based on digital services provided by an embodiment of the present invention is shown, specifically including the following steps 110-150.

[0017] Step 110: Receive user interaction requests sent by user terminal devices, perform multi-modal joint analysis on the user interaction requests through the intent parsing model, decompose the text semantics, voice tone and interaction behavior sequence features in the user interaction requests, and integrate them to obtain a structured user intent representation that includes intent type, demand scenario and expected output format.

[0018] In this embodiment of the invention, the user interaction request sent by the user terminal device is a multimodal data consisting of the text "I want to know about the recent returns and redemption rules of stable funds," a stable and emotionally neutral tone, and the interaction behavior sequence of the text entered after clicking the "Fund Investment" entry. The intent parsing model is a multimodal fusion neural network model, specifically a cascaded structure composed of a text semantic parsing module, a speech tone analysis module, and an interaction behavior sequence feature extraction module. The text semantic parsing module uses a bidirectional long short-term memory network with parameters configured as follows: LSTM_H1 hidden layers, LSTM_N1 neurons per layer, and a learning rate of LSTM_LR1. The speech tone analysis module uses a convolutional neural network with parameters configured as follows: CNN_K1 convolutional kernels, CNN_S1×CNN_S1 kernel size, and CNN_P1×CNN_P1 pooling window size. The interaction behavior sequence feature extraction module uses a gated recurrent unit (GRU) with parameters configured as follows: GRU_H1 hidden layers, GRU_N1 neurons per layer, and a forget gate bias of GRU_B1.

[0019] When the intent parsing model performs multimodal joint analysis of user interaction requests, it first inputs text data into the text semantic parsing module to obtain a text semantic feature vector. This vector is multi-dimensional and contains information such as intent keywords and demand themes. Then, it inputs speech data into the speech intonation analysis module to obtain a speech intonation feature vector, which contains information such as speech rate and intonation fluctuations. Finally, it inputs interaction behavior sequence data into the interaction behavior sequence feature extraction module to obtain an interaction behavior sequence feature vector, which contains information such as operation order and dwell time. These three feature vectors are then concatenated and fused to obtain a multimodal fused feature vector. Finally, a structured user intent representation is output through a fully connected layer. In this representation, the intent type is "querying stable fund returns and redemption rules," the demand scenario is "preservation and appreciation of personal medium- to long-term assets," and the expected output format is "text list."

[0020] Step 120: Based on the search dimensions of the intent type matching historical interaction database and the search dimensions of the demand scenario associated user preference feature library in the structured user intent representation, perform search operations respectively to obtain the first search result corresponding to the historical interaction database and the second search result corresponding to the user preference feature library.

[0021] In this embodiment of the invention, the intent type in the structured user intent representation is "querying the returns and redemption rules of a stable fund." The retrieval dimensions of the historical interaction database include intent type, interaction time, response content, etc. The human-computer interaction processing device searches the historical interaction database according to the intent type, matches historical interaction records related to the intent type, and obtains a first retrieval result. This result contains three historical interaction records, namely History_Rec1, History_Rec2, and History_Rec3. Each record contains information such as historical interaction intent matching degree and historical response satisfaction. The demand scenario in the structured user intent representation is "preservation and appreciation of personal medium- and long-term assets." The retrieval dimensions of the user preference feature database include demand scenario, preference feature items, preference update timestamp, etc. The human-computer interaction processing device searches the user preference feature database according to the demand scenario, matches preference feature items related to the demand scenario, and obtains a second retrieval result. This result contains two preference feature items, namely Pref_Item1 and Pref_Item2. Each preference item contains information such as preference feature similarity and preference update timestamp.

[0022] Step 130: Associate the historical interaction intent matching degree and historical response satisfaction in the first search result, and associate the preference feature similarity and preference update timestamp in the second search result. By fusing the first search result and the second search result through intent-preference association rules, a personalized enhanced feature representing the user's current intent is generated.

[0023] In this embodiment of the invention, in the first search result, the historical interaction intent matching score of History_Rec1 is Match_Score1, and the historical response satisfaction score is Satis_Score1; the historical interaction intent matching score of History_Rec2 is Match_Score2, and the historical response satisfaction score is Satis_Score2; the historical interaction intent matching score of History_Rec3 is Match_Score3, and the historical response satisfaction score is Satis_Score3. In the second search result, the preference feature similarity score of Pref_Item1 is Sim_Score1, and the preference update timestamp is Time_Stamp1; the preference feature similarity score of Pref_Item2 is Sim_Score2, and the preference update timestamp is Time_Stamp2.

[0024] The intent-preference association rule is a weighted summation-based fusion rule. This rule first weights and fuses the historical interaction intent matching degree and historical response satisfaction degree from the first search result, with weights of Weight_H1 and Weight_H2, respectively, to obtain the historical interaction contribution value. The historical interaction contribution value for History_Rec1 is: Match_Score1×Weight_H1+Satis_Score1×Weight_H2; The historical interaction contribution value of History_Rec2 is: Match_Score2×Weight_H1+Satis_Score2×Weight_H2; The historical interaction contribution value of History_Rec3 is: Match_Score3×Weight_H1+Satis_Score3×Weight_H2.

[0025] Then, the preference feature similarity and preference update timestamp in the second search results are weighted and fused, with weights of Weight_P1 and Weight_P2, respectively, to obtain the preference feature contribution value, where the preference feature contribution value of Pref_Item1 is: Sim_Score1×Weight_P1+Time_Stamp1×Weight_P2; The contribution value of the preference feature of Pref_Item2 is: Sim_Score2×Weight_P1+Time_Stamp2×Weight_P2.

[0026] Finally, the historical interaction contribution value and the preference feature contribution value are concatenated and fused to obtain the personalized enhancement feature. This feature is a multi-dimensional vector that contains the fusion result of historical interaction information and preference feature information.

[0027] Step 131: Extract the historical interaction intent matching degree and the corresponding historical response satisfaction of each historical interaction record in the first search result. The historical interaction intent matching degree reflects the degree of overlap between the historical interaction request and the intent type of the current structured user intent representation. The historical response satisfaction reflects the user's evaluation of the historical digital service response content.

[0028] In this embodiment of the invention, the historical interaction intent matching degree of History_Rec1 in the first search result is Match_Score1, which is obtained by calculating the proportion of overlapping words between the intent type of the historical interaction request and the intent type of the current structured user intent representation; the corresponding historical response satisfaction is Satis_Score1, which is obtained by converting the user's star rating of the historical digital service response content, with star ratings ranging from 1 to 5 stars, and the conversion method is to divide the star rating by 5. The historical interaction intent matching degree of History_Rec2 is Match_Score2, and the corresponding historical response satisfaction is Satis_Score2; the historical interaction intent matching degree of History_Rec3 is Match_Score3, and the corresponding historical response satisfaction is Satis_Score3. The human-computer interaction processing device extracts the above information by traversing each record in the first search result and stores it in a temporary data structure.

[0029] Step 132: Extract the preference feature similarity and corresponding preference update timestamp of each preference feature item in the second search result. The preference feature similarity reflects the correlation between the preference feature item and the current structured user intent representation of the demand scenario, and the preference update timestamp reflects the most recent update time of the preference feature item.

[0030] In this embodiment of the invention, the preference feature similarity of Pref_Item1 in the second search result is Sim_Score1, which is obtained by calculating the semantic similarity between the semantics of the preference feature item and the semantics of the current structured user intent representation of the demand scenario; the corresponding preference update timestamp is Time_Stamp1, which is the time when the user last updated the preference feature item. The preference feature similarity of Pref_Item2 is Sim_Score2, and the corresponding preference update timestamp is Time_Stamp2. The human-computer interaction processing device extracts the above information by traversing each preference feature item in the second search result and stores it in a temporary data structure.

[0031] Step 133: Set the weight allocation strategy in the intent-preference association rule; where the weight of historical interaction intent matching degree is positively correlated with historical response satisfaction; calculate the time interval based on the current time and preference update timestamp, determine the recency score according to the time interval, the recency score is negatively correlated with the time interval, and the weight of preference feature similarity is positively correlated with the recency score.

[0032] In this embodiment of the invention, the weight calculation method for the historical interaction intent matching degree is as follows: if the historical response satisfaction is Satis_Score1, then the weight is Satis_Score1 × Base_W1, where Base_W1 is the base weight coefficient; if the historical response satisfaction is Satis_Score2, then the weight is Satis_Score2 × Base_W1; if the historical response satisfaction is Satis_Score3, then the weight is Satis_Score3 × Base_W1. The time interval is calculated by subtracting the preference update timestamp from the current time. If the preference update timestamp is Time_Stamp1, then the time interval is Current_Time - Time_Stamp1; if the preference update timestamp is Time_Stamp2, then the time interval is Current_Time - Time_Stamp2. The recentity score is calculated as follows: If the time interval is Current_Time-Time_Stamp1, then the recentity score is 1 / (Current_Time-Time_Stamp1+1); If the time interval is Current_Time-Time_Stamp2, then the recentity score is 1 / (Current_Time-Time_Stamp2+1).

[0033] The weighting method for preference feature similarity is as follows: If the recentity score is 1 / (Current_Time-Time_Stamp1+1), then the weight is 1 / (Current_Time-Time_Stamp1+1)×Base_W2; If the recentity score is 1 / (Current_Time-Time_Stamp2+1), then the weight is 1 / (Current_Time-Time_Stamp2+1)×Base_W2, where Base_W2 is the base weight coefficient.

[0034] Step 134: For each historical interaction record, calculate the historical interaction intent matching degree multiplied by the corresponding weight to obtain the historical interaction contribution value; for each preference feature item, calculate the preference feature similarity multiplied by the corresponding weight to obtain the preference feature contribution value.

[0035] In this embodiment of the invention, the historical interaction contribution value of History_Rec1 is Match_Score1×(Satis_Score1×Base_W1); the historical interaction contribution value of History_Rec2 is Match_Score2×(Satis_Score2×Base_W1); and the historical interaction contribution value of History_Rec3 is Match_Score3×(Satis_Score3×Base_W1).

[0036] The contribution value of the preference feature of Pref_Item1 is: Sim_Score1×(1 / (Current_Time-Time_Stamp1+1)×Base_W2); The contribution value of the preference feature of Pref_Item2 is: Sim_Score2×(1 / (Current_Time-Time_Stamp2+1)×Base_W2).

[0037] Step 135: Calculate the average historical interaction contribution value of all historical interaction records as the average historical interaction contribution, and calculate the average preference feature contribution value of all preference feature items as the average preference feature contribution.

[0038] In this embodiment of the invention, the average historical interaction contribution is: (Match_Score1×(Satis_Score1×Base_W1)+Match_Score2×(Satis_Score2×Base_W1)+Match_Score3×(Satis_Score3×Base_W1)) / 3; The average contribution of the preference feature is: (Sim_Score1×(1 / (Current_Time-Time_Stamp1+1)×Base_W2)+Sim_Score2×(1 / (Current_Time-Time_Stamp2+1)×Base_W2)) / 2.

[0039] Step 136: Adjust the fusion coefficient according to the ratio of average historical interaction contribution to average preference feature contribution; wherein, the fusion coefficient tilts towards the historical interaction direction as the proportion of average historical interaction contribution increases, and tilts towards the preference feature direction as the proportion of average preference feature contribution increases.

[0040] In this embodiment of the invention, the fusion coefficient is calculated as follows: if the average historical interaction contribution is Avg_H_Contri and the average preference feature contribution is Avg_P_Contri, then the fusion coefficient is Avg_H_Contri / (Avg_H_Contri+Avg_P_Contri) used for fusion of historical interaction contribution values, and Avg_P_Contri / (Avg_H_Contri+Avg_P_Contri) used for fusion of preference feature contribution values.

[0041] Step 137: The average historical interaction contribution and the average preference feature contribution are weighted and summed using the fusion coefficient to generate personalized enhanced features.

[0042] In this embodiment of the invention, the personalized enhancement feature is: (Avg_H_Contri / (Avg_H_Contri+Avg_P_Contri))×Avg_H_Contri+(Avg_P_Contri / (Avg_H_Contri+Avg_P_Contri))×Avg_P_Contri, this feature is a multi-dimensional vector that contains the fusion result of historical interaction information and preference feature information.

[0043] Step 140: Input the personalized enhancement features into the intent-driven response generation model. Through the service content template library and dynamic information filling interface in the response generation model, combined with the expected output format in the structured user intent representation, construct digital service response content in real time, including service content structure, information presentation order and interactive guidance options.

[0044] In this embodiment of the invention, the personalized enhancement feature is a multi-dimensional vector containing historical interaction contribution values ​​and preference feature contribution values. The intent-driven response generation model is a generative adversarial network model. The specific type of this model is an adversarial structure composed of a generator and a discriminator. The generator adopts a transformer architecture with the following parameter configurations: the number of encoder layers is Trans_E1, the number of decoder layers is Trans_D1, and the number of multi-head attention heads is Trans_H1. The discriminator adopts a convolutional neural network with the following parameter configurations: the number of convolutional kernels is CNN_K2, the kernel size is CNN_S2×CNN_S2, and the pooling window size is CNN_P2×CNN_P2.

[0045] The service content template library includes a "Fund Returns and Redemption Rules Inquiry" template, which contains placeholders for fund name, latest net asset value, returns for the past month, and redemption rules. The dynamic information population interface is an API interface that connects to the fund data server, with parameters configured as a request timeout of API_TO (1 second) and retries of API_RT (1 time). The expected output format in the structured user intent representation is a "text list".

[0046] When constructing digital service response content, the intent-driven response generation model first inputs personalized enhancement features into the generator, which then generates an initial service content structure based on templates in the service content template library. Next, it retrieves the latest fund return data and subscription / redemption rule data from the fund data server via a dynamic information population interface and populates the placeholders in the initial service content structure. Finally, according to the desired output format, the populated service content structure is converted into a text list, with the information presentation order determined as fund name, latest net asset value, one-month return, and subscription / redemption rules. Interactive guidance options are generated as "View more stable funds" and "Consult a fund manager."

[0047] Step 141: Extract the intent type keywords and demand scenario keywords from the personalized enhancement features, match them with the corresponding intent type and demand scenario templates in the service content template library, and obtain the initial service content structure.

[0048] In this embodiment of the invention, the intent type keywords in the personalized enhancement features are "stable fund," "return status," and "subscription / redemption rules," while the demand scenario keywords are "medium- to long-term asset preservation and appreciation." The "Fund Returns and Subscription / Redemption Rules Inquiry" template in the service content template library includes placeholders such as fund name, latest net asset value, returns for the past month, and subscription / redemption rules. The human-computer interaction processing device selects this template from the service content template library as the initial service content structure through keyword matching.

[0049] Step 142: Call the dynamic information population interface to obtain corresponding real-time information from the external real-time information data source based on the information placeholders in the initial service content structure. The real-time information includes the latest service rules, available resource status and interaction cases related to the demand scenario.

[0050] In this embodiment of the invention, the placeholders in the initial service content structure are the fund name, latest net asset value, one-month return, and subscription / redemption rules. The external real-time information data source is a fund data server, which stores the latest data of all funds. The human-computer interaction processing device calls the dynamic information filling interface to send a query request containing the keyword "stable fund" to the fund data server, and obtains real-time information such as the fund name "XX Stable Growth Hybrid Fund", the latest net asset value is Fund_NV1 yuan, the one-month return is Fund_R1%, and the subscription / redemption rules are "T+1 redemption, redemption fee rate is 0.5% (no redemption fee if held for more than 30 days)".

[0051] Step 1421: For the service rule placeholder, connect to the service rule management system through the dynamic information filling interface, send a query request containing keywords of the requirement scenario, and obtain the latest service rule text related to the requirement scenario returned by the service rule management system. The latest service rule text includes the rule's effective time and applicable conditions.

[0052] In this embodiment of the invention, the service rule placeholder is the subscription and redemption rule. The service rule management system stores the latest data on fund subscription and redemption rules. The human-computer interaction processing device connects to the service rule management system through a dynamic information filling interface, sends a query request containing the keyword "medium- and long-term asset preservation and appreciation", and obtains the subscription and redemption rule text as "T+1 redemption, redemption fee rate of 0.5% (no redemption fee if held for 30 days or more), rule effective date is January 1, 2024, applicable conditions are conservative funds".

[0053] Step 1422: For resource status placeholders, connect to the resource management system through the dynamic information filling interface, send a query request containing keywords of the required scenario, and obtain the available resource types, quantities, and occupancy status information returned by the resource management system.

[0054] In this embodiment of the invention, the resource status placeholders are the latest net asset value and the return over the past month. The resource management system stores the latest net asset value and return data of the fund. The human-computer interaction processing device connects to the resource management system through a dynamic information filling interface, sends a query request containing the keyword "medium-to-long-term asset preservation and appreciation", and obtains the latest net asset value as Fund_NV1 yuan and the return over the past month as Fund_R1%.

[0055] Step 1423: For the interaction case placeholder, connect to the interaction case library through the dynamic information filling interface, send a query request containing the intent type and requirement scenario keywords, and obtain historical interaction cases in the interaction case library that match the current intent and scenario.

[0056] In this embodiment of the invention, the placeholder for the interaction case class is the fund name. The interaction case library stores fund recommendation cases from historical interactions. The human-computer interaction processing device connects to the interaction case library through a dynamic information filling interface, sends a query request containing the keywords "stable fund" and "medium-to-long-term asset preservation and appreciation", and obtains the historical interaction case as "User A queries the returns of stable funds and recommends XX Stable Growth Hybrid Fund".

[0057] Step 1424: Filter the obtained service rule text, resource status information and interaction cases, and fill the filtered information into the corresponding information placeholders; if the information placeholder of one of the real-time information types does not obtain the corresponding real-time information, return the default prompt message, which includes the reason why the real-time information cannot be obtained and the alternative acquisition method.

[0058] In this embodiment of the invention, the human-computer interaction processing device filters the obtained subscription / redemption rule text, latest net asset value, one-month return, and fund name. After removing invalid information, it fills the subscription / redemption rule text into the subscription / redemption rule placeholder, the latest net asset value into the latest net asset value placeholder, the one-month return into the one-month return placeholder, and the fund name into the fund name placeholder. If the one-month return is not obtained, the default prompt message "Unable to obtain one-month return data, it is recommended to view the three-month return data" is returned.

[0059] Step 1425: After the information is filled, if there are any unfilled information placeholders, resend the query request until all information placeholders are filled or the default prompt message is returned.

[0060] In this embodiment of the invention, if the fund name placeholder is not filled, the human-computer interaction processing device resends the query request to the interaction case library until the fund name is obtained or the default prompt information is returned.

[0061] Step 143: Fill the information placeholders in the initial service content structure with the acquired real-time information to generate the filled service content structure.

[0062] In this embodiment of the invention, the human-computer interaction processing device fills the fund name "XX Steady Growth Hybrid Fund", the latest net asset value Fund_NV1 yuan, the one-month return Fund_R1%, and the redemption rule "T+1 redemption, redemption fee rate of 0.5% (no redemption fee if held for 30 days)" into the corresponding placeholders of the initial service content structure to generate the filled service content structure.

[0063] Step 144: Extract the expected output format from the structured user intent representation. If the expected output format is a text list, the format adaptation module converts the filled service content structure into a hierarchical list. If the expected output format is a visual chart, it is converted into a visual structure containing data dimensions and chart types to determine the information presentation order.

[0064] In this embodiment of the invention, the expected output format in the structured user intent representation is a "text list". The format adaptation module is a text format conversion module, configured with the following parameters: the number of list levels is List_L1, and the indentation of each level is List_I1 characters. The format adaptation module converts the populated service content structure into a hierarchical list format, with the information presented in the following order: fund name, latest net asset value, returns for the past month, and subscription / redemption rules.

[0065] Step 145: Based on the filled service content structure and information presentation order, generate interactive guidance options, which include further information requests, operation confirmations and feedback entry points related to the service content structure.

[0066] In this embodiment of the invention, the filled-in service content structure includes the fund name, latest net asset value, one-month return, and subscription / redemption rules. The information is presented in the following order: fund name, latest net asset value, one-month return, and subscription / redemption rules. The interactive guidance options generated by the human-computer interaction processing device are "View more stable funds," "Consult a fund manager," and "Feedback on service content."

[0067] Step 1451: Identify key information points in the populated service content structure, including the core rules of the service content, resource constraints, and factors affecting the results; generate corresponding progressive information request options for each key information point.

[0068] In this embodiment of the invention, the key information points in the filled service content structure are the redemption rule "T+1 redemption, redemption fee rate of 0.5% (no redemption fee if held for 30 days)" and the recent month's return Fund_R1%. The human-computer interaction processing device generates a progressive information request option for the redemption rule as "view detailed redemption rule"; and generates a progressive information request option for the recent month's return as "view the recent three months' return".

[0069] Step 1452: Identify logical nodes in the information presentation sequence. Logical nodes include step demarcation points and decision points of service content; generate corresponding operation confirmation options for each logical node.

[0070] In this embodiment of the invention, the logical nodes in the information presentation order are the fund name, the latest net asset value, the return over the past month, and the dividing point of the subscription and redemption rules. The human-computer interaction processing device generates an operation confirmation option of "Confirm to view the fund" for the fund name; and generates an operation confirmation option of "Confirm to understand the subscription and redemption rules" for the subscription and redemption rules.

[0071] Step 1453: Extract the overall structure of the digital service response content, locate the feedback entry point, and the feedback entry point content is an invitation for users to evaluate the service response content.

[0072] In this embodiment of the invention, the overall structure of the digital service response content is a text list, including fund information and interactive guidance options. The human-computer interaction processing device positions the feedback entry at the end of the interactive guidance options, with the content being "Feedback Service Content".

[0073] Step 1454: Insert the progressive information request option, the operation confirmation option, and the feedback entry into the filled service content structure in the order of information presentation.

[0074] In this embodiment of the invention, the human-computer interaction processing device inserts "View detailed subscription and redemption rules" and "View returns for the past three months" after the subscription and redemption rules and the returns for the past month; inserts "Confirm to view the fund" and "Confirm to understand the subscription and redemption rules" after the fund name and the subscription and redemption rules; and inserts "Feedback service content" at the end of the interactive guidance options.

[0075] Step 1455: Incorporate the generated set of interactive guidance options, which includes further information requests, operation confirmations, and feedback entry points, into the digital service response content.

[0076] In this embodiment of the invention, the human-computer interaction processing device integrates "view detailed subscription and redemption rules", "view returns for the past three months", "confirm viewing the fund", "confirm understanding of subscription and redemption rules", and "feedback service content" into digital service response content.

[0077] Step 146: Generate the digital service response content based on the service content structure, the information presentation order, and the interactive guidance options.

[0078] In this embodiment of the invention, the service content structure includes the fund name, latest net asset value, one-month return, and subscription / redemption rules; the information presentation order is fund name, latest net asset value, one-month return, and subscription / redemption rules; the interactive guidance options are "View detailed subscription / redemption rules," "View three-month return," "Confirm viewing the fund," "Confirm understanding of subscription / redemption rules," and "Feedback on service content." The human-computer interaction processing device integrates the above content to generate digital service response content, which is in the form of a text list and includes fund information and interactive guidance options.

[0079] Step 150: Push the digital service response content to the user terminal device, obtain the service comments returned by the user terminal device for the digital service response content, extract the intent matching evaluation and content satisfaction evaluation from the service comments, and update the preference feature similarity weight and preference update timestamp in the user preference feature library in combination with the service content structure and information presentation order in the digital service response content.

[0080] In this embodiment of the invention, the digital service response content is a text list containing the fund name, latest net asset value, recent one-month return, subscription / redemption rules, and interactive guidance options. The human-computer interaction processing device pushes this content to the user's terminal device via a push interface, configured with a push timeout of Push_TO (1 second) and a retries count of Push_RT (1 time). The user's terminal device returns a service review stating "The content is comprehensive, but the description of the subscription / redemption rules is not clear enough." The human-computer interaction processing device extracts the intent matching evaluation from the service review as "Comprehensive content" and the content satisfaction evaluation as "The description of the subscription / redemption rules is not clear enough." Then, combining the service content structure of the digital service response content (fund information, information presentation order: fund name, latest net asset value, recent one-month return, subscription / redemption rules), it updates the preference feature similarity weight in the user preference feature library to 1 times the original Upd_W and sets the preference update timestamp to the current time.

[0081] Step 151: After encapsulating the digital service response content, push it to the user terminal device and set the comment collection period. During the comment collection period, listen for the service comments returned by the user terminal device.

[0082] In this embodiment of the invention, the digital service response content is encapsulated in JSON format, including content fields, format fields, and interaction option fields. The comment collection time is set to Collect_T1 minutes. The human-computer interaction processing device listens for service comments returned by the user's terminal device within Collect_T1 minutes.

[0083] Step 152: Perform text semantic parsing on the obtained service reviews to extract the intent matching evaluation and content satisfaction evaluation from the service reviews; wherein, the intent matching evaluation reflects the degree to which the user believes that the service response content matches their own intent, and the content satisfaction evaluation reflects the user's satisfaction with the service response content.

[0084] In this embodiment of the invention, the service review is "The content is comprehensive, but the description of the redemption rules is not clear enough." The human-computer interaction processing device extracts the intent matching evaluation as "comprehensive content" and the content satisfaction evaluation as "the description of the redemption rules is not clear enough" through text semantic parsing.

[0085] Step 1521: Perform semantic segmentation on the service review, dividing the service review into multiple text segments with independent semantic meanings.

[0086] In this embodiment of the invention, the service review "The content is comprehensive, but the description of the redemption rules is not clear enough" is divided into two text segments: "The content is comprehensive" and "The description of the redemption rules is not clear enough".

[0087] Step 1522: Perform core semantic node extraction for each text segment to identify key semantic units in the text segment that represent user evaluation tendencies.

[0088] In this embodiment of the invention, the core semantic node of the text fragment "the content is comprehensive" is "comprehensive content"; the core semantic node of the text fragment "the description of the redemption rules is not clear enough" is "the description of the redemption rules is not clear enough".

[0089] Step 1523: Establish the association chain between each core semantic node based on the semantic similarity algorithm, and connect the core semantic nodes with semantic association in sequence to form a continuous reasoning path.

[0090] In this embodiment of the invention, the core semantic nodes are "comprehensive content" and "unclear description of redemption rules". The semantic similarity algorithm is the cosine similarity algorithm. The human-computer interaction processing device calculates the semantic similarity between two core semantic nodes as Sim_Cos1, and establishes an association chain of "comprehensive content" -- "unclear description of redemption rules".

[0091] Step 15231: Perform context window extraction processing on each core semantic node, extract the text fragments before and after the core semantic node in the service review, and generate the context semantic representation of the core semantic node. The context semantic representation includes the adjacent text semantics and positional relationship of the core semantic node.

[0092] In this embodiment of the invention, the contextual text fragment for the core semantic node "comprehensive content" is "the content is comprehensive, but..."; the contextual text fragment for the core semantic node "the description of redemption rules is unclear" is "but the description of redemption rules is not clear enough". The contextual semantic representation includes the semantic information of adjacent text and the positional information of the node in the comment.

[0093] Step 15232: Perform semantic encoding operation on each core semantic node to generate an independent semantic representation of the core semantic node, and combine it with the context semantic representation to obtain a comprehensive semantic representation of the core semantic node.

[0094] In this embodiment of the invention, the semantic encoding operation adopts the bag-of-words model, encoding the core semantic node "comprehensive content" into vector form to obtain an independent semantic representation; combined with the contextual semantic representation, a comprehensive semantic representation is obtained. Similarly, the core semantic node "unclear description of redemption rules" is handled.

[0095] Step 15233: Calculate the comprehensive semantic similarity between any two core semantic nodes. The comprehensive semantic similarity is a weighted fusion of the similarity of the independent semantic representations of the nodes and the similarity of the contextual semantic representations.

[0096] In this embodiment of the invention, the independent semantic representation similarity between the core semantic nodes "comprehensive content" and "unclear description of redemption rules" is Sim_Ind1, and the contextual semantic representation similarity is Sim_Con1. The overall semantic similarity is Sim_Ind1×0.6+Sim_Con1×0.4.

[0097] Step 15234: Select node pairs with similarity higher than a preset threshold based on comprehensive semantic similarity, and establish initial connection relationships between core semantic nodes.

[0098] In this embodiment of the invention, the preset threshold is Thres_S1. If the comprehensive semantic similarity Sim_Ind1×0.6+Sim_Con1×0.4 is higher than Thres_S1, then an initial connection relationship of "comprehensive content" -- "unclear description of redemption rules" is established.

[0099] Step 15235: Based on the text order of the service reviews, construct an association chain by combining the initial connection relationship. The starting node of the association chain is the core semantic node at the beginning of the service review, and subsequent nodes are arranged in descending order of comprehensive semantic similarity.

[0100] In this embodiment of the invention, the text order of the service review is "The content is comprehensive, but the description of the redemption rules is not clear enough." The order of the core semantic nodes is "Comprehensive content" and "Unclear description of redemption rules." The initial connection relationship is "Comprehensive content" -- "Unclear description of redemption rules." The associated chain is "Comprehensive content" -- "Unclear description of redemption rules."

[0101] Step 15236: Divide the association chain into hierarchical levels to distinguish between main nodes and child nodes. The main node is the core semantic unit that plays a leading role in user evaluation tendency, and the child node is the core semantic unit that assists in explaining the main node.

[0102] In this embodiment of the invention, the main node is "comprehensive content" and the child node is "unclear description of redemption rules".

[0103] Step 15237: Perform reverse reasoning verification from the last node of the association chain back to the starting node to confirm the semantic and logical consistency between nodes. If there is a logical contradiction, adjust the connection order of the nodes.

[0104] In this embodiment of the invention, the last node of the association chain is "unclear description of redemption rules," and the starting node is "comprehensive content." The reverse derivation is "unclear description of redemption rules" -- "comprehensive content," verifying semantic and logical consistency: "comprehensive content" but "unclear description of redemption rules" is logically consistent.

[0105] Step 15238: Identify the key node hierarchy in the association chain. The first-level key node is the node with the highest comprehensive semantic similarity among the main nodes, and the second-level key node is the node with the highest comprehensive semantic similarity among the child nodes.

[0106] In this embodiment of the invention, the primary key node is "comprehensive content" and the secondary key node is "unclear description of redemption rules".

[0107] Step 15239: Adjust the path structure of the association chain based on the key node hierarchy, and retain the connection relationship between the first-level key nodes to form an optimized inference path.

[0108] In this embodiment of the invention, the optimized reasoning path is "comprehensive content" -- "unclear description of redemption rules".

[0109] Step 1524: Combining the service content structure and information presentation order of the digital service response content, map the core semantic nodes in the association chain to the corresponding service links, and determine the correspondence between each core semantic node and each link of the service response content.

[0110] In this embodiment of the invention, the service content structure of the digital service response content is fund information, and the information presentation order is fund name, latest net asset value, recent one-month return, and subscription / redemption rules. The core semantic nodes in the association chain are "comprehensive content" and "unclear description of subscription / redemption rules". The human-computer interaction processing device maps "comprehensive content" to the fund information stage and "unclear description of subscription / redemption rules" to the subscription / redemption rules stage.

[0111] Step 15241: Extract information from each stage of the service content structure of the digital service response content, divide the service stages into semantic hierarchical levels, and divide them into first-level stages and second-level sub-stages to generate a set of hierarchical stage semantic representations.

[0112] In this embodiment of the invention, the process information in the service content structure includes fund name, latest net asset value, one-month return, and subscription / redemption rules. The first-level process is fund information, and the second-level sub-processes are fund name, latest net asset value, one-month return, and subscription / redemption rules. The hierarchical process semantic representation set includes the semantic representation of the first-level process "fund information" and the semantic representations of the second-level sub-processes "fund name," "latest net asset value," "one-month return," and "subscription / redemption rules."

[0113] Step 15242: Calculate the comprehensive semantic similarity between each core semantic node in the association chain and each first-level and second-level sub-link in the semantic representation set of the hierarchical links, and obtain the matching degree between the node and each level of links.

[0114] In this embodiment of the invention, the comprehensive semantic similarity between the core semantic node "comprehensive content" and the first-level segment "fund information" is Sim_L1; the comprehensive semantic similarity with the second-level sub-segment "fund name" is Sim_L2; the comprehensive semantic similarity with "latest net asset value" is Sim_L3; the comprehensive semantic similarity with "recent one-month return" is Sim_L4; and the comprehensive semantic similarity with "subscription and redemption rules" is Sim_L5. The comprehensive semantic similarity between the core semantic node "unclear description of subscription and redemption rules" and the first-level segment "fund information" is Sim_L6; and the comprehensive semantic similarity with the second-level sub-segment "subscription and redemption rules" is Sim_L7.

[0115] Step 15243: Perform multi-level mapping on the nodes based on the matching degree, mapping the nodes to the first-level link with the highest matching degree, and mapping them to the second-level sub-links with the highest matching degree under that first-level link.

[0116] In this embodiment of the invention, the core semantic node "comprehensive content" has the highest matching degree with the first-level stage "fund information" and the second-level sub-stage "recent one-month returns," and therefore is mapped to the first-level stage "fund information" and the second-level sub-stage "recent one-month returns." Similarly, the core semantic node "unclear description of subscription and redemption rules" has the highest matching degree with the first-level stage "fund information" and the second-level sub-stage "subscription and redemption rules," and therefore is mapped to the first-level stage "fund information" and the second-level sub-stage "subscription and redemption rules."

[0117] Step 15244: Combine the feedback data of the second-level sub-step in the user's historical interaction to cross-validate the multi-level mapping results to confirm the semantic matching rationality between the node and the sub-step. If the verification fails, the matching degree is recalculated.

[0118] In this embodiment of the invention, the feedback data for the second-level sub-step "Recent Monthly Earnings" in the user's historical interaction is "The user previously reported that the recent month's earnings data was useful." Cross-validation results show a reasonable match. The feedback data for the second-level sub-step "Redemption Rules" in the user's historical interaction is "The user previously reported that the description of the redemption rules needs to be clearer." Cross-validation results show a reasonable match.

[0119] Step 15245: Classify the mapped nodes and merge the core semantic nodes belonging to the same second-level sub-link into node groups. Each node group contains a main node and child nodes.

[0120] In this embodiment of the invention, the core semantic node belonging to the second-level sub-section "recent month's income" is "comprehensive content", and the core semantic node belonging to the second-level sub-section "redemption rules" is "unclear description of redemption rules", so they are divided into two node groups.

[0121] Step 15246: Analyze the comprehensive semantic similarity of each core semantic node in the node group to determine the core node of the node group. The core node is the core semantic unit with the highest comprehensive semantic similarity in the node group.

[0122] In this embodiment of the invention, the core semantic node in the node group "Recent Monthly Returns" is "Comprehensive Content," which has the highest overall semantic similarity and is therefore considered a core node. Similarly, the core semantic node in the node group "Redemption Rules" is "Unclear Description of Redemption Rules," which also has the highest overall semantic similarity and is therefore considered a core node.

[0123] Step 15247: Dynamically adjust the node groups. If nodes in a node group are mapped to different secondary sub-links, split the node group; if multiple node groups are mapped to the same secondary sub-link, merge the node groups.

[0124] In this embodiment of the invention, the node group's "recent month's returns" and "redemption rules" are mapped to different secondary sub-stages, so no adjustments are made.

[0125] Step 15248: Determine the correspondence between node groups and secondary sub-links based on the semantic meaning of core nodes, generate a node-sub-link mapping table, calculate the mapping quality score of the node-sub-link mapping table by combining the comprehensive semantic similarity between node groups and sub-links and the cross-validation pass rate, and feed the mapping quality score back to the service content structure optimization module to adjust the semantic description of secondary sub-links.

[0126] In this embodiment of the invention, the core node of the node group "Recent Monthly Revenue" is "Comprehensive Content," corresponding to the second-level sub-stage "Recent Monthly Revenue"; the core node of the node group "Redemption Rules" is "Unclear Description of Redemption Rules," corresponding to the second-level sub-stage "Redemption Rules." A node-sub-stage mapping table is generated. The mapping quality score is calculated as comprehensive semantic similarity × 0.7 + cross-validation pass rate × 0.3. The mapping quality score is fed back to the service content structure optimization module, adjusting the semantic description of the second-level sub-stage "Redemption Rules" to "Redemption Rules (Clear Version)."

[0127] Step 1525: Extract the set of core semantic nodes related to intent matching evaluation and content satisfaction evaluation based on the mapping relationship, and generate a comprehensive evaluation semantic structure that includes surface evaluation and potential evaluation.

[0128] In this embodiment of the invention, the mapping relationship is as follows: "Comprehensive content" maps to the "Recent Month's Returns" section, and "Unclear Description of Redemption Rules" maps to the "Redemption Rules" section. The core semantic node set is "Comprehensive content" and "Unclear Description of Redemption Rules". The comprehensive evaluation semantic structure is "Comprehensive content (Recent Month's Returns)" and "Unclear Description of Redemption Rules (Redemption Rules Section)".

[0129] Step 153: Convert the intent matching evaluation into the corresponding weight adjustment direction. If the intent matching evaluation is positive, increase the similarity weight of the corresponding preference feature item. If the intent matching evaluation is negative, decrease the similarity weight of the corresponding preference feature item.

[0130] In this embodiment of the invention, the intent matching evaluation is "comprehensive content", which is a positive evaluation. Therefore, the similarity weight of the "stable fund return query" preference feature item in the user preference feature library is increased by 1 times the original Upd_W.

[0131] Step 154: Convert the content satisfaction rating into the corresponding timestamp update status. If the content satisfaction rating is positive, set the preference update timestamp of the corresponding preference feature item to the current time; if the content satisfaction rating is negative, keep the preference update timestamp of the corresponding preference feature item unchanged.

[0132] In this embodiment of the invention, the content satisfaction rating is "the description of the redemption rules is not clear enough", which is a negative evaluation. Therefore, the preference update timestamp of the "redemption rule query" preference feature item in the user preference feature library is kept unchanged.

[0133] Step 155: Based on the service content structure and information presentation order in the digital service response content, determine the core preference features that affect the intent matching evaluation and content satisfaction evaluation, and adjust the similarity weights of the core preference features.

[0134] In this embodiment of the invention, the service content structure of the digital service response content is fund information, and the information presentation order is fund name, latest net asset value, returns for the past month, and subscription / redemption rules. The core preference feature affecting intent matching evaluation is "stable fund return query," and the core preference feature affecting content satisfaction evaluation is "subscription / redemption rule query." The human-computer interaction processing device increases the similarity weight of the preference feature "stable fund return query" to Upd_W1 times the original value, and decreases the similarity weight of the preference feature "subscription / redemption rule query" to Upd_W2 times the original value.

[0135] Step 156: Based on the weight adjustment direction and timestamp update status, update the preference feature similarity weight and preference update timestamp of the corresponding preference feature item in the user preference feature library.

[0136] In this embodiment of the invention, the weight adjustment direction is to increase the similarity weight of the "Conservative Fund Return Inquiry" preference feature item and decrease the similarity weight of the "Subscription and Redemption Rule Inquiry" preference feature item. The timestamp update status is to keep the preference update timestamp of the "Subscription and Redemption Rule Inquiry" preference feature item unchanged. The human-computer interaction processing device updates the similarity weight of the "Conservative Fund Return Inquiry" preference feature item in the user preference feature database to Upd_W1 times the original value and the preference update timestamp to the current time; it updates the similarity weight of the "Subscription and Redemption Rule Inquiry" preference feature item to Upd_W2 times the original value and keeps the preference update timestamp unchanged.

[0137] Optionally, the method further includes: Step 210: Extract the core semantic node set from the structured user intent representation. The core semantic node set includes intent type nodes, demand scenario nodes, and expected output format nodes.

[0138] In this embodiment of the invention, the structured user intent is represented as the intent type "querying the returns and redemption rules of a stable fund", the demand scenario "preserving and increasing the value of personal medium- and long-term assets", and the expected output format "text list". The core semantic node set extracted by the human-computer interaction processing device is "stable fund", "returns and redemption rules", "preserving and increasing the value of medium- and long-term assets", and "text list".

[0139] Step 220: Label each historical interaction record in the historical interaction database with core semantic nodes to generate a set of historical nodes.

[0140] In this embodiment of the invention, the historical interaction request for History_Rec1 in the historical interaction database is "querying the returns of a conservative fund," with the core semantic nodes labeled as "conservative fund" and "returns." The historical interaction request for History_Rec2 is "understanding the subscription and redemption rules of a conservative fund," with the core semantic nodes labeled as "conservative fund" and "subscription and redemption rules." The historical interaction request for History_Rec3 is "querying the returns and subscription and redemption rules of a mixed fund," with the core semantic nodes labeled as "mixed fund" and "returns and subscription and redemption rules." The generated set of historical nodes is "conservative fund," "returns," "subscription and redemption rules," and "mixed fund."

[0141] Step 230: Calculate the semantic similarity between the core semantic nodes of the core semantic nodes in the structured user intent representation and the core semantic nodes of each historical interaction record in the historical node set, determine the matching historical interaction records, and generate historical node association weights based on the historical interaction intent matching degree and historical response satisfaction of the matching historical interaction records.

[0142] In this embodiment of the invention, the core semantic node set of the structured user intent representation is "stable fund", "return and redemption rules", "medium-to-long-term asset preservation and appreciation", and "text list". The core semantic nodes of the historical interaction record History_Rec1 are "stable fund" and "return status", with a semantic similarity of Sim_H1; the core semantic nodes of History_Rec2 are "stable fund" and "redemption rules", with a semantic similarity of Sim_H2; and the core semantic nodes of History_Rec3 are "mixed fund" and "return and redemption rules", with a semantic similarity of Sim_H3. The matched historical interaction records are History_Rec1 and History_Rec2. The historical interaction intent matching degree of History_Rec1 is Match_Score1, and the historical response satisfaction degree is Satis_Score1; the historical interaction intent matching degree of History_Rec2 is Match_Score2, and the historical response satisfaction degree is Satis_Score2. The calculation method for the historical node association weight is as follows: (Match_Score1×Satis_Score1+Match_Score2×Satis_Score2) / 2.

[0143] Step 240: Calculate the semantic similarity between the core semantic node set of the structured user intent representation and the preference feature items in the user preference feature library, determine the matching preference feature items, and generate preference node association weights based on the preference feature similarity of the matching preference feature items and the preference update timestamp.

[0144] In this embodiment of the invention, the core semantic node set of the structured user intent representation is "stable fund", "return and redemption rules", "medium-to-long-term asset preservation and appreciation", and "text list". The core semantic nodes of Pref_Item1 in the user preference feature library are "stable fund" and "medium-to-long-term asset preservation and appreciation", with a semantic similarity of Sim_P1; the core semantic nodes of Pref_Item2 are "redemption rules" and "text list", with a semantic similarity of Sim_P2. The matched preference features are Pref_Item1 and Pref_Item2. The preference feature similarity of Pref_Item1 is Sim_Score1, and the preference update timestamp is Time_Stamp1; the preference feature similarity of Pref_Item2 is Sim_Score2, and the preference update timestamp is Time_Stamp2. The calculation method for the preference node association weight is as follows: (Sim_Score1×(Current_Time-Time_Stamp1+1)+Sim_Score2×(Current_Time-Time_Stamp2+1)) / 2.

[0145] Step 250: Combine the historical node association weights and the preference node association weights to enhance the core semantic node set, generate an enhanced core semantic node set, and input the enhanced core semantic node set into the intent-preference association rule to participate in the generation process of personalized enhanced features.

[0146] In this embodiment of the invention, the historical node association weight is: (Match_Score1×Satis_Score1+Match_Score2×Satis_Score2) / 2; The association weights of the preference nodes are: (Sim_Score1×(Current_Time-Time_Stamp1+1)+Sim_Score2×(Current_Time-Time_Stamp2+1)) / 2.

[0147] The enhancement process involves weighted summation of the association weights of historical nodes and preference nodes to obtain the enhancement weights. Then, each node in the core semantic node set is multiplied by its enhancement weight to generate an enhanced core semantic node set. This enhanced core semantic node set is then input into the intent-preference association rule to participate in the generation of personalized enhancement features.

[0148] Optionally, the method further includes: Step 310: Extract the core semantic nodes of each link from the service content structure of the digital service response content, and generate a service node set.

[0149] In this embodiment of the invention, the service content structure of the digital service response content includes the fund name, latest net asset value, recent one-month return, and subscription / redemption rules. The core semantic nodes extracted by the human-computer interaction processing device are "XX Steady Growth Hybrid Fund", "Fund_NV1 yuan", "Fund_R1%", and "T+1 redemption, redemption fee rate is 0.5% (no redemption fee if held for 30 days)". The generated service node set is "XX Steady Growth Hybrid Fund", "Fund_NV1 yuan", "Fund_R1%", and "Subscription / Redemption Rules".

[0150] Step 320: Extract core semantic nodes from the service comments returned by the user terminal device to generate a set of comment nodes.

[0151] In this embodiment of the invention, the service review is "The content is comprehensive, but the description of the redemption rules is not clear enough." The core semantic nodes extracted by the human-computer interaction processing device are "comprehensive content" and "unclear description of redemption rules." The generated review node set is "comprehensive content" and "unclear description of redemption rules."

[0152] Step 330: Calculate the semantic similarity of each node in the service node set and the comment node set, determine the node matching degree, filter out node pairs with matching degrees higher than a preset threshold based on the node matching degree, establish a service-comment node association chain, perform path analysis on the service-comment node association chain, identify the key path in the association chain, and the key path is the node connection sequence with the highest association strength.

[0153] In this embodiment of the invention, the service node set is "XX Steady Growth Hybrid Fund", "Fund_NV1 yuan", "Fund_R1%", and "Redemption Rules", and the comment node set is "Comprehensive Content" and "Unclear Description of Redemption Rules". Semantic similarity is calculated: the semantic similarity between "XX Steady Growth Hybrid Fund" and "Comprehensive Content" is Sim_M1, between "Fund_NV1 yuan" and "Comprehensive Content" is Sim_M2, between "Fund_R1%" and "Comprehensive Content" is Sim_M3, and between "Redemption Rules" and "Unclear Description of Redemption Rules" is Sim_M4. A preset threshold is Thres_M1. Node pairs with a matching degree higher than Thres_M1 are selected as "Fund_R1%" and "Comprehensive Content", and "Redemption Rules" and "Unclear Description of Redemption Rules". A service-comment node association chain is established as "Fund_R1%" -- "Comprehensive Content" -- "Unclear Description of Redemption Rules" -- "Redemption Rules". The path analysis method involves calculating the association strength of each path, which is the sum of the semantic similarities of all node pairs in the path. The critical path is "redemption rules" -- "redemption rules are not clearly described".

[0154] Step 340: Based on the semantic meaning of the critical path, determine the links in the service response content that need to be optimized; extract the core semantic nodes of the links that need to be optimized to generate an optimization node set; adjust the corresponding template structure in the service content template library according to the semantic meaning of the optimization node set, and update the data source priority of the dynamic information filling interface.

[0155] In this embodiment of the invention, the semantic meaning of the critical path is "unclear description of redemption rules". The human-computer interaction processing device determines that the part of the service response content that needs optimization is the redemption rules part. The extracted set of optimization nodes is "redemption rules" and "unclear description". Then, the template structure of "fund income and redemption rules query" in the service content template library is adjusted, and the description of the redemption rules placeholder is modified to "redemption rules (detailed version)"; the data source priority of the dynamic information filling interface is updated, and the priority of the fund data server is increased to Pri_Upd1 times the original.

[0156] This invention employs an intent parsing model to perform multi-modal joint analysis of the text semantics, voice tone, and interactive behavior sequence features of user interaction requests. This analysis integrates these features to form a structured user intent representation that includes intent type, demand scenario, and expected output format. This overcomes the limitations of single-modal parsing, achieving a precise and comprehensive characterization of user intent. Furthermore, based on this structured user intent representation, historical interaction databases and user preference feature libraries are retrieved separately. Personalized enhanced features are generated by fusing historical interaction intent matching degree, historical response satisfaction, preference feature similarity, and preference update timestamps through intent-preference association rules. This effectively links historical user interaction patterns with real-time preference needs, solving the problem of insufficient response personalization caused by the separation of historical data and preference features. The personalized enhanced features are then input into an intent-driven response generation model, which, combined with a service content template library and dynamic information filling interface, constructs digital service response content in real time, ensuring the structural rationality, real-time information, and format adaptability of the response content. Finally, intent matching evaluation and content satisfaction evaluation are extracted through service reviews, which are used to update the preference feature similarity weights and preference update timestamps in the user preference feature library. This improves the accuracy, personalization level, and continuous optimization capabilities of human-computer interaction, significantly enhancing user experience and increasing service response efficiency.

[0157] See Figure 2 , Figure 2 This diagram illustrates a block diagram of a human-computer interaction processing device according to an embodiment of the present invention. The human-computer interaction processing device includes: The multimodal joint analysis module is used to receive user interaction requests sent by user terminal devices, perform multimodal joint analysis on the user interaction requests through the intent parsing model, decompose the text semantics, voice tone and interaction behavior sequence features in the user interaction requests, and integrate them to obtain a structured user intent representation that includes intent type, demand scenario and expected output format. The multidimensional retrieval processing module is used to perform retrieval operations based on the retrieval dimensions of the historical interaction database and the retrieval dimensions of the user preference feature library associated with the demand scenario in the structured user intent representation, respectively, to obtain a first retrieval result corresponding to the historical interaction database and a second retrieval result corresponding to the user preference feature library. The intent feature enhancement module is used to associate the historical interaction intent matching degree and historical response satisfaction in the first search result, and associate the preference feature similarity and preference update timestamp in the second search result. By fusing the first search result and the second search result through intent-preference association rules, personalized enhanced features representing the user's current intent are generated. The service response construction module is used to input the personalized enhancement features into the intent-driven response generation model. Through the service content template library and dynamic information filling interface in the response generation model, combined with the expected output format in the structured user intent representation, it constructs digital service response content in real time, including service content structure, information presentation order and interactive guidance options. The interactive feedback update module pushes the digital service response content to the user terminal device, obtains the service comments returned by the user terminal device regarding the digital service response content, extracts the intent matching evaluation and content satisfaction evaluation from the service comments, and updates the preference feature similarity weight and preference update timestamp in the user preference feature library by combining the service content structure and information presentation order in the digital service response content.

[0158] See Figure 3 , Figure 3 A structural block diagram of a human-computer interaction processing device 1200 is shown, including a memory 1210 and a processor 1220; the memory 1210 is used to store computer programs / instructions, and the processor 1220 is used to execute the computer programs / instructions, which, when executed by the processor, implement the steps of the aforementioned human-computer interaction processing method based on digital services.

[0159] Furthermore, a computer storage medium is provided that stores a computer program / instruction, which, when executed by a processor, implements the aforementioned human-computer interaction processing method based on digital services.

[0160] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0161] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments of the present invention.

[0162] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0163] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments of the present invention. These embodiments are selected and specifically described to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention.

[0164] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A human-computer interaction processing method based on digital services, characterized in that, include: The system receives user interaction requests sent by user terminal devices, performs multi-modal joint analysis on the user interaction requests through an intent parsing model, decomposes the text semantics, voice tone and interaction behavior sequence features in the user interaction requests, and integrates them to obtain a structured user intent representation that includes intent type, demand scenario and expected output format. Based on the search dimensions of the intent type matching historical interaction database and the search dimensions of the demand scenario associated user preference feature library in the structured user intent representation, search operations are performed respectively to obtain a first search result corresponding to the historical interaction database and a second search result corresponding to the user preference feature library. By associating the historical interaction intent matching degree and historical response satisfaction degree in the first search result with the preference feature similarity degree and preference update timestamp in the second search result, the first search result and the second search result are fused through intent-preference association rules to generate personalized enhanced features that represent the user's current intent. The personalized enhancement features are input into the intent-driven response generation model. Through the service content template library and dynamic information filling interface in the response generation model, combined with the expected output format in the structured user intent representation, digital service response content containing service content structure, information presentation order and interactive guidance options is constructed in real time. The digital service response content is pushed to the user terminal device, and the service comments returned by the user terminal device for the digital service response content are obtained. The intent matching evaluation and content satisfaction evaluation in the service comments are extracted. Based on the service content structure and information presentation order in the digital service response content, the preference feature similarity weight and preference update timestamp in the user preference feature library are updated.

2. The method as described in claim 1, characterized in that, The first search result is associated with the historical interaction intent matching degree and historical response satisfaction degree, and the second search result is associated with the preference feature similarity and preference update timestamp. The first and second search results are then fused using intent-preference association rules to generate personalized enhanced features representing the user's current intent, including: Extract the historical interaction intent matching degree and the corresponding historical response satisfaction of each historical interaction record in the first search result. The historical interaction intent matching degree reflects the degree of overlap between the historical interaction request and the intent type of the current structured user intent representation. The historical response satisfaction reflects the user's evaluation of the historical digital service response content. Extract the preference feature similarity and corresponding preference update timestamp of each preference feature item in the second search result. The preference feature similarity reflects the correlation between the preference feature item and the current structured user intent representation of the demand scenario, and the preference update timestamp reflects the most recent update time of the preference feature item. Set a weighting strategy in the intent-preference association rule; among which, the weight of historical interaction intent matching degree is positively correlated with historical response satisfaction; calculate the time interval based on the current time and preference update timestamp, determine the recency score according to the time interval, the recency score is negatively correlated with the time interval, and the weight of preference feature similarity is positively correlated with the recency score; For each historical interaction record, the historical interaction intent matching degree is calculated and multiplied by the corresponding weight to obtain the historical interaction contribution value. For each preference feature item, the preference feature similarity is calculated and multiplied by the corresponding weight to obtain the preference feature contribution value. The average historical interaction contribution value of all historical interaction records is calculated as the average historical interaction contribution, and the average preference feature contribution value of all preference feature items is calculated as the average preference feature contribution. The fusion coefficient is adjusted according to the ratio of average historical interaction contribution to average preference feature contribution; wherein, the fusion coefficient tilts towards the historical interaction direction as the proportion of average historical interaction contribution increases, and tilts towards the preference feature direction as the proportion of average preference feature contribution increases. The average historical interaction contribution and the average preference feature contribution are weighted and summed using a fusion coefficient to generate personalized enhanced features.

3. The method as described in claim 1, characterized in that, The personalized enhancement features are input into the intent-driven response generation model. Through the service content template library and dynamic information filling interface within the response generation model, combined with the expected output format in the structured user intent representation, digital service response content containing service content structure, information presentation order, and interactive guidance options is constructed in real time, including: Extract the intent type keywords and demand scenario keywords from the personalized enhancement features, match them with templates of the corresponding intent type and demand scenario in the service content template library, and obtain the initial service content structure. The dynamic information population interface is called to obtain corresponding real-time information from an external real-time information data source based on the information placeholders in the initial service content structure. The real-time information includes the latest service rules, available resource status and interaction cases related to the demand scenario. The acquired real-time information is filled into the information placeholders of the initial service content structure to generate the filled service content structure. Extract the expected output format from the structured user intent representation. If the expected output format is a text list, the format adaptation module converts the filled service content structure into a hierarchical list. If the expected output format is a visual chart, it is converted into a visual structure containing data dimensions and chart types to determine the information presentation order. Based on the filled service content structure and information presentation order, interactive guidance options are generated, including further information requests, operation confirmations and feedback entry points related to the service content structure. Based on the service content structure, the information presentation order, and the interactive guidance options, the digital service response content is generated.

4. The method as described in claim 3, characterized in that, The information placeholders in the initial service content structure are marked as the types of real-time information to be filled. These real-time information types include service rule classes, resource status classes, and interaction case classes. The dynamic information filling interface is invoked to obtain corresponding real-time information from an external real-time information data source based on the information placeholders in the initial service content structure, including: For service rule placeholders, a dynamic information population interface is used to connect to the service rule management system, send a query request containing keywords of the requirement scenario, and obtain the latest service rule text related to the requirement scenario returned by the service rule management system. The latest service rule text includes the rule's effective time and applicable conditions. For resource status placeholders, a dynamic information population interface is used to connect to the resource management system, send a query request containing keywords of the required scenario, and obtain the available resource types, quantities, and occupancy status information returned by the resource management system. For interactive case placeholders, a dynamic information filling interface is used to connect to the interactive case library, send a query request containing keywords of intent type and requirement scenario, and obtain historical interactive cases in the interactive case library that match the current intent and scenario; The system filters the obtained service rule text, resource status information, and interaction cases, and fills the filtered information into the corresponding information placeholders. If the information placeholder for a certain real-time information type does not obtain the corresponding real-time information, a default prompt message is returned, which includes the reason why the real-time information could not be obtained and alternative methods for obtaining it. After the information is filled, if there are any unfilled placeholders, the query request will be resent until all placeholders are filled or the default message is returned.

5. The method as described in claim 3, characterized in that, The interactive guidance options are generated based on the filled service content structure and information presentation order, including: Identify key information points in the populated service content structure, including the core rules of the service content, resource constraints, and factors affecting the results; generate corresponding progressive information request options for each key information point; Identify logical nodes in the information presentation sequence, including step demarcation points and decision points of service content; generate corresponding operation confirmation options for each logical node; Extract the overall structure of the digital service response content, locate the feedback entry point, and the feedback entry point content is an invitation for users to evaluate the service response content; Insert the progressive information request option, the operation confirmation option, and the feedback entry into the filled service content structure in the order of information presentation; The generated set of interactive guidance options, including further information requests, operation confirmations, and feedback entry points, is incorporated into the digital service response content.

6. The method as described in claim 1, characterized in that, The step of pushing the digital service response content to the user terminal device, obtaining service reviews returned by the user terminal device for the digital service response content, extracting intent matching evaluation and content satisfaction evaluation from the service reviews, and updating the preference feature similarity weight and preference update timestamp in the user preference feature library based on the service content structure and information presentation order in the digital service response content includes: The digital service response content is packaged and pushed to the user terminal device, and a comment collection period is set. During the comment collection period, the service comments returned by the user terminal device are listened for. The obtained service reviews are subjected to text semantic parsing to extract the intent matching evaluation and content satisfaction evaluation from the service reviews; wherein, the intent matching evaluation reflects the degree to which the user believes that the service response content matches their own intent, and the content satisfaction evaluation reflects the user's satisfaction with the service response content; The intent matching evaluation is converted into the corresponding weight adjustment direction. If the intent matching evaluation is positive, the similarity weight of the corresponding preference feature item is increased; if the intent matching evaluation is negative, the similarity weight of the corresponding preference feature item is decreased. The content satisfaction rating is converted into a corresponding timestamp update status. If the content satisfaction rating is positive, the preference update timestamp of the corresponding preference feature item is set to the current time; if the content satisfaction rating is negative, the preference update timestamp of the corresponding preference feature item is left unchanged. Based on the service content structure and information presentation order in the digital service response content, the core preference features that influence intent matching evaluation and content satisfaction evaluation are determined, and the similarity weights of the core preference features are adjusted. Based on the weight adjustment direction and timestamp update status, update the preference feature similarity weight and preference update timestamp of the corresponding preference feature item in the user preference feature library.

7. The method as described in claim 6, characterized in that, The step of performing text semantic parsing on the acquired service reviews to extract intent matching evaluations and content satisfaction evaluations from the service reviews includes: Semantic segmentation is performed on service reviews, dividing them into multiple text segments with independent semantic meanings; For each text segment, a core semantic node extraction operation is performed to identify the key semantic units in the text segment that represent user evaluation tendencies; Based on the semantic similarity algorithm, the association chain between each core semantic node is established, and the core semantic nodes with semantic association are connected in sequence to form a continuous reasoning path; Semantic inference is performed on the association chain to uncover potential evaluation tendencies that users have not directly expressed. These potential evaluation tendencies include implicit feedback on needs not covered by service response content and potential optimization suggestions for service interaction processes. By combining the service content structure and information presentation order of digital service response content, the core semantic nodes in the association chain are mapped to the corresponding service links, and the correspondence between each core semantic node and each link of the service response content is determined. Based on the mapping relationship, the core semantic node set related to intent matching evaluation and content satisfaction evaluation is extracted, and a comprehensive evaluation semantic structure containing surface evaluation and potential evaluation is generated.

8. The method as described in claim 7, characterized in that, The process of establishing association chains between core semantic nodes based on semantic similarity algorithms, and sequentially connecting semantically related core semantic nodes to form a continuous reasoning path, includes: For each core semantic node, a context window extraction process is performed to extract the text fragments before and after the core semantic node in the service review, and a context semantic representation of the core semantic node is generated. The context semantic representation includes the semantics of adjacent texts and positional relationships of the core semantic node. Perform semantic encoding operations on each core semantic node to generate an independent semantic representation of the core semantic node, and combine it with the context semantic representation to obtain a comprehensive semantic representation of the core semantic node. Calculate the comprehensive semantic similarity between any two core semantic nodes. The comprehensive semantic similarity is a weighted fusion of the similarity of the independent semantic representations of the nodes and the similarity of the semantic representations of the context. Based on the comprehensive semantic similarity, node pairs with similarity higher than a preset threshold are selected, and the initial connection relationship between core semantic nodes is established. Based on the text order of service reviews, an association chain is constructed by combining the initial connection relationship. The starting node of the association chain is the core semantic node at the beginning of the service review, and subsequent nodes are arranged in descending order of comprehensive semantic similarity. The association chain is hierarchically divided to distinguish between main nodes and child nodes. The main node is the core semantic unit that plays a leading role in user evaluation tendencies, while the child node is the core semantic unit that assists in explaining the main node. The process involves reverse reasoning from the last node of the association chain back to the starting node to verify the semantic and logical consistency between nodes. If there is a logical contradiction, the connection order of the nodes is adjusted. Identify the key node hierarchy in the association chain. The first-level key node is the node with the highest comprehensive semantic similarity among the main nodes, and the second-level key node is the node with the highest comprehensive semantic similarity among the child nodes. The path structure of the association chain is adjusted based on the key node hierarchy, and the connection relationship between the first-level key nodes is retained to form an optimized inference path. The service content structure and information presentation order, combined with the digital service response content, map the core semantic nodes in the association chain to the corresponding service links, determining the correspondence between each core semantic node and each link of the service response content, including: Extract information from each stage of the service content structure of digital service response content, divide the service stages into semantic hierarchies, and divide them into first-level stages and second-level sub-stages to generate a set of hierarchical stage semantic representations. Calculate the comprehensive semantic similarity between each core semantic node in the association chain and each first-level and second-level sub-link in the semantic representation set of the hierarchical links to obtain the matching degree between the node and each level of links; Based on the matching degree, the nodes are mapped in multiple layers, mapping the nodes to the first-level link with the highest matching degree, and then to the second-level sub-link with the highest matching degree under that first-level link; The multi-level mapping results are cross-validated by combining feedback data from the user's historical interactions with the second-level sub-step to confirm the semantic matching rationality between the node and the sub-step. If the verification fails, the matching degree is recalculated. The mapped nodes are categorized, and core semantic nodes belonging to the same second-level sub-link are merged into node groups, which contain main nodes and child nodes. Analyze the comprehensive semantic similarity of each core semantic node in the node group to determine the core node of the node group. The core node is the core semantic unit with the highest comprehensive semantic similarity in the node group. The node groups are dynamically adjusted. If nodes in a node group are mapped to different second-level sub-links, the node group is split; if multiple node groups are mapped to the same second-level sub-link, the node groups are merged. Based on the semantic meaning of the core nodes, the correspondence between node groups and secondary sub-links is determined, and a node-sub-link mapping table is generated. The mapping quality score of the node-sub-link mapping table is calculated by combining the comprehensive semantic similarity between node groups and sub-links and the cross-validation pass rate. The mapping quality score is then fed back to the service content structure optimization module to adjust the semantic description of the secondary sub-links.

9. A human-computer interaction processing device, characterized in that, include: Memory and processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions, which, when executed by the processor, implement the steps of the method according to any one of claims 1 to 8.

10. A computer storage medium, characterized in that, The device stores a computer program / instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 8.

Images