Intelligent Interaction State Tracking and Reasoning Methods and Systems Applied to Multi-Turn Dialogue

Abstract

This application provides a method and system for intelligent interaction state tracking and reasoning applied to multi-turn dialogues, relating to the field of natural language processing technology. First, it retrieves the initial interaction state of the multi-turn dialogue, covering the user's initial needs, the system's initial response location, and dialogue scenario-related content. Then, it captures the state evolution triggering conditions for each round of interaction. Next, it constructs a bidirectional correlation evolution model based on the initial interaction state and triggering conditions to characterize the correlation and influence between preceding and subsequent interaction states. The current interaction state is derived through this model, integrating the state evolution results of all rounds and the bidirectional correlation influence. Finally, a dynamic guidance strategy is generated based on the current interaction state, adapting to subsequent interaction trends and adjusting the preceding correlation logic in reverse. This invention improves the intelligence and coherence of multi-turn dialogues.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, and more specifically, to an intelligent interactive state tracking and reasoning method and system applied to multi-turn dialogue. Background Technology

[0002] In the fields of artificial intelligence and natural language processing, multi-turn dialogue systems have become a key technology for achieving efficient human-computer interaction, and are widely used in various scenarios such as intelligent customer service, intelligent assistants, and online education. However, existing methods for tracking and reasoning the interaction state in multi-turn dialogues have significant shortcomings.

[0003] On the one hand, traditional methods often focus only on user input and system responses in the current round, lacking a comprehensive review and integration of the dialogue's historical state. This makes it difficult for the system to accurately grasp the evolution of user needs when dealing with complex dialogue scenarios, and it cannot adjust its interaction strategies in a timely manner according to the dynamic development of the dialogue. This can easily lead to problems such as irrelevant answers or repetitive responses, seriously affecting the user experience.

[0004] On the other hand, existing methods mostly employ a one-way reasoning model for state tracking, that is, deriving the subsequent state only from the preceding interaction state, while ignoring the backtracking effect of the subsequent state on the preceding state. This one-way reasoning approach struggles to accurately capture implicit information and logical connections in multi-turn dialogues with high uncertainty and complexity, resulting in inaccurate state tracking results and consequently affecting the guiding effect of subsequent dialogues. Summary of the Invention

[0005] In view of this, the purpose of this application is to provide a method and system for intelligent interactive state tracking and reasoning applied to multi-turn dialogue.

[0006] According to a first aspect of this application, a method for intelligent interaction state tracking and reasoning applied to multi-turn dialogue is provided, the method comprising:

[0007] Retrieve the initial interaction state of a multi-turn dialogue, which includes the user's initial needs expressed when the dialogue starts, the system's initial response positioning content, and dialogue scenario-related content.

[0008] Capture the state evolution triggering conditions for each round of interaction in a multi-round dialogue. The state evolution triggering conditions are related factors such as user expression adjustments, system response optimizations, and dialogue scene switching that cause changes in the interaction state.

[0009] Based on the initial interaction state and the state evolution triggering conditions, a bidirectional correlation evolution model for multi-turn dialogue is constructed. The bidirectional correlation evolution model is used to characterize the driving correlation effect of the preceding interaction state on the subsequent interaction state and the retrospective correlation effect of the subsequent interaction state on the preceding interaction state.

[0010] The current interaction state of a multi-turn dialogue is derived through the bidirectional correlation evolution model, and the current interaction state integrates the state evolution results of all rounds of interaction and the bidirectional correlation influence content.

[0011] Based on the current interaction state, a dynamic guidance strategy for multi-turn dialogue is generated. The dynamic guidance strategy adapts to the state evolution trend of subsequent interactions and adjusts the related logic of previous interactions in reverse.

[0012] According to a second aspect of this application, an intelligent interaction state tracking and reasoning system for multi-turn dialogue is provided. The intelligent interaction state tracking and reasoning system for multi-turn dialogue includes a machine-readable storage medium and a processor. The machine-readable storage medium stores machine-executable instructions. When the processor executes the machine-executable instructions, the intelligent interaction state tracking and reasoning system for multi-turn dialogue implements the aforementioned intelligent interaction state tracking and reasoning method for multi-turn dialogue.

[0013] According to a third aspect of this application, a computer-readable storage medium is provided, wherein computer-executable instructions are stored therein, and when the computer-executable instructions are executed, the aforementioned intelligent interactive state tracking and reasoning method applied to multi-turn dialogue is implemented.

[0014] Based on any of the above aspects, the technical effect of this application is as follows:

[0015] By retrieving the initial interaction states of multiple rounds of dialogue, encompassing key information such as the user's initial needs at the start of the dialogue, the system's initial response positioning, and the dialogue scenario associations, and then capturing the state evolution triggering conditions of each round of interaction, it can accurately identify various factors that cause changes in the interaction state, including adjustments in user expression, optimization of system responses, and switching of dialogue scenarios, ensuring real-time perception of state evolution. The bidirectional correlation evolution model built based on the initial interaction state and state evolution triggering conditions breaks through the limitations of traditional unidirectional reasoning. It not only represents the driving effect of the preceding interaction state on the subsequent interaction state but also considers the retrospective correlation influence of the subsequent interaction state on the preceding interaction state, achieving comprehensive and accurate tracking of multi-round dialogue states. The current interaction state derived through this model integrates the state evolution results of all rounds of interaction and the bidirectional correlation influences, accurately reflecting the current state of the dialogue. The dynamic guidance strategy generated based on the current interaction state can adapt to the state evolution trend of subsequent interactions and adjust the related logic of previous interactions in reverse. This allows the dialogue system to flexibly adjust the interaction strategy according to the dynamic development of the dialogue, significantly improving the coherence, accuracy and intelligence of multi-turn dialogues, and providing users with a better and more efficient interactive experience. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the intelligent interaction state tracking and reasoning method for multi-turn dialogue provided in an embodiment of this application is shown.

[0017] Figure 2 This illustration shows a schematic diagram of the component structure of an intelligent interactive state tracking and reasoning system for multi-turn dialogue provided in an embodiment of this application. Detailed Implementation

[0018] Figure 1 This paper illustrates a flowchart of an intelligent interaction state tracking and reasoning method and system for multi-turn dialogue provided in an embodiment of this application. The detailed steps include:

[0019] Step S110: Retrieve the initial interaction state of the multi-turn dialogue, which includes the user's initial demand expression when the dialogue starts, the system's initial response positioning content, and the dialogue scenario-related content.

[0020] In the context of IoT-enabled smart home systems, initiating a dialogue requires retrieving the initial interaction state. This initial interaction state encompasses the user's initial needs, the system's initial response, and related dialogue scenario information. The user's initial needs may be complex, for example, the user expressing a desire for "automatic opening of the bedroom curtains at 7 AM to let in sunlight; maintaining a comfortable living room temperature from 9 AM to 11 AM while simultaneously activating the air purifier; and preheating of the smart oven in the kitchen at 3 PM to prepare for baking; with these settings automatically adjusting based on weekends and weekdays." The system's initial response, based on the user's initial needs, identifies the relevant smart home devices, such as the bedroom curtain motor, living room thermostat, air purifier, and kitchen smart oven, and clarifies the direction of interaction with these devices. Related dialogue scenario information includes whether the current date is a weekday or weekend, outdoor weather conditions, and whether the user is at home. For example, if it is a weekday and the user is not at home, the startup time and operating mode of some devices may differ.

[0021] Step S120: Capture the state evolution triggering conditions for each round of interaction in a multi-turn dialogue. The state evolution triggering conditions are related factors such as user expression adjustments, system response optimizations, and dialogue scene switching that cause changes in the interaction state.

[0022] During multi-round dialogues, each round of interaction may trigger state evolution. User-expressed adjustments may involve modifications or additions to the initial requirements during the dialogue. For example, a user might mention in one round, "Preheating the oven at 3 PM is changed to 4 PM because I'll be home later today." System-optimized responses are improvements to previous responses based on user feedback or new requirements. For instance, the original response, "Living room temperature is adjusted via air conditioning," might be optimized to "Living room temperature is adjusted via the underfloor heating system, which is more energy-efficient and environmentally friendly" based on the user's mention of underfloor heating. Scene transitions may occur due to changes in the external environment. For example, if a user was initially at home but later went out, the scene changes from "indoor" to "remote outdoor control," requiring corresponding adjustments to the interaction state, such as adding the ability to remotely monitor the operating status of home devices.

[0023] Step S130: Construct a bidirectional correlation evolution model for multi-turn dialogue based on the initial interaction state and the state evolution triggering conditions. The bidirectional correlation evolution model is used to characterize the driving correlation effect of the preceding interaction state on the subsequent interaction state and the retrospective correlation effect of the subsequent interaction state on the preceding interaction state.

[0024] Step S131: Extract the core requirement dimension content corresponding to the user's initial requirement expression content in the initial interaction state, and decompose the core requirement dimension content into multiple requirement sub-items. Each requirement sub-item content corresponds to a specific implementation direction of the initial requirement.

[0025] Based on the user's initial requirements—"They want the bedroom curtains to open automatically at 7 AM to let in sunlight; the living room temperature to remain within a comfortable range from 9 AM to 11 AM while the air purifier starts; and the smart oven in the kitchen to preheat at 3 PM to prepare for baking; and these settings to automatically adjust according to weekends and weekdays"—the core requirements can be extracted and categorized into curtain control, temperature control, air purification, kitchen appliance control, and time mode adjustment. These core requirements are then further broken down into sub-items. The requirements under the curtain control dimension include "automatically opening bedroom curtains to full open at 7 AM" and "automatically adjusting the curtain opening degree according to outdoor light intensity," etc.; the requirements under the temperature control dimension include "maintaining the living room temperature within a comfortable range from 9 AM to 11 AM" and "automatically switching between comfortable temperature ranges for different seasons," etc.; the requirements under the air purification dimension include "starting the air purifier and running in automatic mode from 9 AM to 11 AM" and "adjusting the purification intensity according to the indoor air quality index," etc.; the requirements under the kitchen appliance control dimension include "starting the smart oven to preheat to the specified temperature at 3 PM" and "sending a reminder message after preheating is complete," etc.; the requirements under the time mode adjustment dimension include "executing according to the above time settings on weekdays" and "delaying the curtain opening time by 1 hour on weekends," etc.

[0026] Step S132: Analyze the user expression adjustment content in the state evolution trigger condition content, distinguish the three types of content corresponding to the user expression adjustment content: supplementary content of requirement sub-items, modified content of requirement sub-items, and newly added content of requirement sub-items, and mark the interaction round node content corresponding to each type of content.

[0027] During the dialogue, users may adjust their needs. For example, in the second round of interaction, the user says, "I want the smart lights in the bedroom to turn on at 7 a.m. when I open the curtains." This is a supplement to the sub-item "The bedroom curtains automatically open to the fully open state at 7 a.m." under the curtain control dimension, and is considered a supplementary sub-item. The corresponding interaction round is marked as round 2. In the fourth round of interaction, the user mentions, "Change the oven preheating time from 3 p.m. to 4 p.m.." This is a correction to the sub-item "The smart oven starts preheating to the specified temperature at 3 p.m." under the kitchen appliance control dimension, and is considered a correction to the sub-item. The interaction round is marked as round 4. In the sixth round of interaction, the user adds, "I want the smart speaker in the living room to automatically play soft music at 8 p.m.." This is a newly added sub-item, and is marked as round 6.

[0028] Step S133: Identify the system response optimization content in the state evolution trigger condition content, associate the system response optimization content with the corresponding requirement sub-item content, record whether the system response optimization content has completely solved, partially solved or not solved the corresponding requirement sub-item content, and record the interaction round node content corresponding to the identification result.

[0029] During the interaction, the system response is optimized. For example, in the third round of interaction, in response to the user's supplementary request to "open the curtains and turn on the smart bedroom light at 7 am," the optimized system response is "the bedroom curtains will automatically open at 7 am, and the smart bedroom light will turn on to 30% brightness." The related sub-item of this response is the supplementary content of "the bedroom curtains will automatically open to full open at 7 am." Analysis shows that this optimized response meets the user's needs and is considered a complete solution; the corresponding interaction round node is recorded as round 3. As another example, in the fifth round of interaction, the user mentioned wanting more energy-efficient living room temperature control. The optimized system response is "maintaining the living room temperature at the lower limit of the comfortable range from 9 am to 11 am." The related sub-item of this response is "maintaining the living room temperature within a comfortable range from 9 am to 11 am." Since this only adjusts the upper and lower limits of the temperature range without changing the essence of temperature control, it is considered a partial solution; the interaction round node is recorded as round 5.

[0030] Step S134: Extract the dialogue scene switching content from the state evolution triggering condition content, analyze the correlation and impact of the dialogue scene switching content on all requirement sub-item content, and mark the requirement sub-item content affected by the scene switching content and the corresponding interaction round node content.

[0031] The switching of dialogue scenarios can affect the content of demand sub-items. For example, in the 7th round of interaction, the user was originally at home, but later received a notification that they needed to go out, and the scenario switched from "indoor interaction" to "remote control". Analyzing the impact of this scenario switch on the relevance of all demand sub-items, the demand sub-items "maintaining the living room temperature within a comfortable range from 9:00 AM to 11:00 AM" and "simultaneously starting the air purifier" may need to be adjusted to energy-saving mode after the user leaves, and are therefore affected demand sub-items. The demand sub-item "starting the smart oven to preheat to the specified temperature at 4:00 PM" is not affected by the scenario switch and still needs to be executed on time. The affected demand sub-items are marked as "maintaining the living room temperature within a comfortable range from 9:00 AM to 11:00 AM" and "starting the air purifier and running in automatic mode from 9:00 AM to 11:00 AM", corresponding to the 7th round of interaction.

[0032] Step S135: Using the interaction round as the timeline content, integrate the changes in the demand sub-items, the impact of system response optimization, and the impact of scene switching in each interaction round into round state evolution unit content. Each round state evolution unit content includes all state change factors in that round.

[0033] Using the interaction rounds as a timeline, the content of the round state evolution unit is integrated starting from round 1. In round 1, the changes in the demand sub-items are the initial demand sub-items, the system response optimization impact is the initial response to the initial demand, and there is no scene switching related impact. These are integrated into the round state evolution unit content for round 1. In round 2, the changes in the demand sub-items are supplementary content to "opening the curtains and turning on the smart bedroom light at 7 AM," the system response optimization impact is the response to this supplementary content, and there is no scene switching related impact. These are integrated into the round state evolution unit content for round 2. The round state evolution unit content for each round includes all factors that cause state changes in that round.

[0034] Step S136: Establish the driving relationship between the content of the state evolution unit in the previous round and the content of the state evolution unit in the subsequent round, and deduce the initial state basic content of the state evolution unit in the subsequent round based on the change content of the demand sub-items in the content of the state evolution unit in the previous round.

[0035] Establish a driving relationship between preceding and subsequent rounds. For example, in the state evolution unit content of round 2, the change in the demand sub-item is the addition of the demand to turn on the smart light in the bedroom. This will drive the state evolution unit content of round 3, so that the initial state basic content of round 3 includes the control information of the smart light in the bedroom. The system response optimization in round 3 will be based on this initial state basic content, thus establishing a driving relationship between preceding and subsequent rounds.

[0036] Step S1361: Traverse the content of all round state evolution units, arrange them in the time order of the interaction rounds to form a round sequence content, and record the position of each round state evolution unit content in the sequence and the relationship between adjacent rounds.

[0037] Iterate through the state evolution unit content of all rounds. Assuming there are 8 rounds of interaction, arrange them in chronological order as round 1, round 2... round 8 to form a round sequence. Record the position of each round's state evolution unit content in the sequence, such as position 1 for round 1, position 2 for round 2, and so on. Also record the relationships between adjacent rounds, such as round 2 being the round following round 1, round 1 being the round preceding round 2, and round 3 being the round following round 2, etc.

[0038] Step S1362: Extract the change content of the requirement sub-items in the state evolution unit content of the previous round, distinguish the three forms of change of the requirement sub-items: new content, supplementary content and modified content, and record the specific content of the requirement sub-items corresponding to each form of change.

[0039] Taking round 2 as an example, the changes in the demand sub-items in its round state evolution unit are supplements to the demand sub-items under the curtain control dimension. These are supplementary contents, and the specific content of the corresponding demand sub-item is recorded as "The bedroom curtains automatically open at 7:00 AM, and the smart bedroom light turns on to 30% brightness." The changes in the demand sub-items in round 4 are corrections to the demand sub-items under the kitchen appliance control dimension. These are correction contents, and the specific content of the corresponding demand sub-item is recorded as "The smart oven starts preheating to the specified temperature at 4:00 PM."

[0040] Step S1363: Analyze the system response optimization content in the state evolution unit content of the previous round, record the adaptation of the system response optimization content to the change content of the requirement sub-items, and record the content of the satisfied state of the requirement sub-items after adaptation.

[0041] The third round of system response optimizations was conducted in response to the supplementary requirements in the second round. The adaptation was to enable the smart lights to turn on simultaneously with the opening of the curtains, according to user needs. The adapted requirement met the requirement of "being able to meet the requirement of opening the bedroom curtains and turning on the smart lights to 30% brightness at 7 am".

[0042] Step S1364: Extract the scene switching content from the state evolution unit content of the previous round, and mark the requirement sub-items that remain related after the scene switching content and the content whose degree of association changes.

[0043] In the 7th round of scene switching, the scene changed from "indoor interaction" to "remote control". After extracting the scene switching content, the sub-items that remained related were "the smart oven starts preheating to the specified temperature at 4 pm" and "the smart speaker in the living room automatically plays light music at 8 pm". These sub-items were not affected by the scene switching and their relevance remained unchanged. However, the relevance of the sub-items "the living room temperature is maintained within a comfortable range from 9 am to 11 am" and "the air purifier starts and runs in automatic mode from 9 am to 11 am" changed and needed to be adjusted to energy-saving mode.

[0044] Step S1365: Integrate the changes in the form of demand sub-items, the content of satisfied states, and the changes in the degree of scenario correlation in the content of the previous round state evolution unit into a set of previous driving factors. Each driving factor corresponds to a specific factor that affects the subsequent state.

[0045] Taking round 7 as an example, the content of changes in the form of demand sub-items (no new, supplementary, or modified content), the content of the fulfillment status (some demand sub-items need adjustment), and the content of changes in the degree of scenario relevance (the degree of relevance of some demand sub-items changes) are integrated into the set of preceding driving factors. One of the driving factors is "When the scenario is switched to remote control, the degree of relevance of the living room temperature control demand sub-item decreases, and it needs to be adjusted to energy-saving mode."

[0046] Step S1366: Record the initial state components of the state evolution unit in the subsequent rounds. The initial state components include the basic content of the initial demand sub-items, the initial direction of the system response, and the initial setting of the scenario adaptation.

[0047] Record the initial state components of subsequent rounds (e.g., round 8). The basic content of the initial requirement sub-items is the various requirement sub-items after the adjustment in round 7. The initial direction of the system response is the response direction for the remote control scenario. The initial setting of the scenario adaptation is the device operating parameter setting under the remote control scenario.

[0048] Step S1367: Match each driving factor content in the preceding driving factor set with the initial state constituent element content of the subsequent round state evolution unit content, and record the initial state element content corresponding to each driving factor content.

[0049] Match each driving factor in the set of preceding driving factors in round 7 with the initial state constituent elements in round 8. For example, the driving factor "The scene is switched to remote control, the correlation of the living room temperature control demand sub-item is reduced, and it needs to be adjusted to energy-saving mode" corresponds to the living room temperature control demand sub-item in the "basic content of initial demand sub-items" of the initial state constituent elements in round 8.

[0050] Step S1368: Based on the matching results, integrate the influence of the preceding driving factors into the initial state components of the subsequent round state evolution unit, and adjust the priority of the initial demand sub-item basic content, the adaptation direction of the system response initial direction content, and the targeted content of the scenario adaptation initial setting content.

[0051] Based on the matching results, the influence of the preceding driving factors is integrated into the initial state components of round 8. The priority of the basic content of the initial requirement sub-items is adjusted, with the priority of the living room temperature control requirement sub-item being reduced; the adaptation direction of the system response initial direction content is adjusted to focus more on the convenience and energy efficiency of remote control; the targeted content of the scenario adaptation initial setting content is adjusted to the energy-saving parameters of the equipment in the remote control scenario.

[0052] Step S1369: Integrate the adjusted initial state constituent elements into the initial state basic content of the subsequent round state evolution unit content. This initial state basic content fully inherits the driving and related influences of the previous round state evolution unit content.

[0053] The adjusted initial state components of the 8th round are integrated into the initial state base content of the 8th round state evolution unit, so that it fully inherits the driving and related influences of the 7th round, and ensures that the interactive state of the 8th round evolves reasonably based on the changes of the previous rounds.

[0054] Step S13610: Record the driving relationship path between the content of the state evolution unit in the preceding round and the content of the state evolution unit in the following round. The record of each driving relationship path includes: the preceding round number, the following round number, the driving factor type, the name of the affected subsequent state constituent element, and the specific parameter value of the corresponding element after being affected.

[0055] Record the driving relationship path between round 7 and round 8. The preceding round is numbered 7 and the following round is numbered 8. The driving factor type is scene switching impact. The name of the affected subsequent state component is "living room temperature control requirement sub-item in the basic content of the initial requirement sub-item". The specific parameter value of the corresponding element after being affected is "priority reduced and temperature adjustment range adjusted to energy saving range".

[0056] Step S137: Construct a retrospective relationship between the content of the state evolution unit in the subsequent round and the content of the state evolution unit in the previous round. If the content of a certain requirement sub-item is modified in the content of the state evolution unit in the subsequent round, a modification retrospective mark is added to the content of the requirement sub-item in the content of the state evolution unit in the previous round, and the weight value of the requirement sub-item in the subsequent state derivation is adjusted up or down according to the modification type.

[0057] In multi-turn dialogues, subsequent turns may modify the requirements of previous turns, thus establishing a backtracking relationship. For example, in the fifth turn of interaction, the user modifies the requirement for living room temperature control. In this case, a correction backtracking marker needs to be added to the state evolution unit of the previous turn (e.g., turn 3) for this requirement. If the correction reflects the user's belief that the previous temperature setting was too high (an optimization), then the weight of this requirement in subsequent state derivations is increased to prioritize the user's latest needs. If the correction is due to user error (an error), then the weight is decreased.

[0058] Step S138: Perform a network-like interweaving process on the driving and backtracking relationships of all round state evolution unit contents to form a relational evolution structure containing multi-dimensional relational path contents. Each relational path contents corresponds to a set of mutually influential round state evolution unit contents.

[0059] The driving and retrogressive relationships between all rounds are integrated to form a network-like evolutionary structure. For example, round 2 has a driving relationship with round 3, round 5 has a retrogressive relationship with round 3, round 3 has a driving relationship with round 4, round 6 has a retrogressive relationship with round 4, and so on. These relationships are intertwined to form multi-dimensional relationship paths. Each relationship path corresponds to a set of mutually influential round state evolution units, such as round 2-round 3-round 4-round 6 forming a relationship path.

[0060] Step S139: Integrate the core requirement dimension content of the initial interaction state as the basic anchor content of the related evolution structure content, so that all related path content can start the evolution process from the initial requirement content.

[0061] The core requirements from the initial interaction state (such as curtain control and temperature control) are integrated into the relational evolution structure as basic anchor points. All relational paths originate from these core requirements and evolve as the rounds progress. For example, the relational path for the curtain control dimension starts with the initial requirement "the bedroom curtains open automatically at 7 a.m.," and evolves through subsequent additions and corrections.

[0062] Step S1310: Integrate the content of driving association strength, retrospective association range, and priority of demand sub-items in the association evolution structure into the parameter content of the bidirectional association evolution model to form a bidirectional association evolution model. This bidirectional association evolution model is used to present the driving association effect of the preceding interaction state on the subsequent interaction state and the retrospective association effect of the subsequent interaction state on the preceding interaction state.

[0063] In the relational evolution structure, the driving correlation strength indicates the magnitude of the influence of preceding rounds on subsequent rounds, the retrospective correlation range indicates the scope of the influence of subsequent rounds on preceding rounds, and the demand sub-item priority indicates the importance of each demand sub-item in the evolution process. Integrating these elements into parameters for a bidirectional relational evolution model, such as driving correlation strength parameters, retrospective correlation range parameters, and demand sub-item priority parameters, forms a bidirectional relational evolution model capable of demonstrating the mutual influence between preceding and subsequent interactive states.

[0064] Step S140: Derive the current interaction state of the multi-turn dialogue through the bidirectional correlation evolution model. The current interaction state integrates the state evolution results of all rounds of interaction and the bidirectional correlation influence content.

[0065] In the context of smart home systems within the Internet of Things (IoT), when deriving the current interaction state, the system first retrieves the state evolution unit content from all rounds of the bidirectional correlation evolution model. These unit contents are arranged chronologically according to the interaction's occurrence, and each unit includes changes in user needs, optimizations in system responses, and scene transitions within that round. Next, based on the model's pre-defined push and backtracking correlation rules, the correlation analysis of these unit contents is performed. Push correlation rules clarify how preceding rounds influence subsequent rounds; for example, a user's adjustment of bedroom temperature in a preceding round directly affects the temperature control strategy in subsequent rounds. Backtracking correlation rules determine how subsequent rounds correct the content of preceding rounds; for instance, a user activating energy-saving mode in a subsequent round will adjust the temperature settings in the preceding round accordingly. During the analysis, the state evolution results of each round of interaction need to be superimposed, and adjustments made to the mutual influences based on the correlation rules. For example, if a user sets the bedroom temperature to 20 degrees Celsius in a previous round, and then adjusts it to 18 degrees Celsius in a subsequent round after activating energy-saving mode, these adjustments will be taken into account when deriving the current interaction state. The final current interaction state includes the cumulative changes from the initial interaction to the current round, as well as the bidirectional correlation between previous and subsequent rounds, and can comprehensively reflect the actual situation of the current dialogue.

[0066] Step S141: Determine the content of the current round node in the multi-round dialogue, and use the content of the round state evolution unit corresponding to the content of the current round node as the starting point for derivation, and extract all state change factors in the content of the round state evolution unit.

[0067] In the multi-round dialogue process of a smart home system, the current round node is determined based on the actual progress of the dialogue. Assuming the current interaction is in round 8, then the content of the current round node is round 8. Taking the round state evolution unit content corresponding to round 8 as the starting point, this unit content contains all factors that caused state changes in round 8. Specifically, it is necessary to extract the content of changes in demand sub-items, such as the user's new demand in round 8: "I want the smart refrigerator to automatically remind me to clean up expired food at night"; system response optimization content, such as the system's response optimization of "playing cleaning reminders through the smart speaker" in response to the user's new demand; and content related to scene switching effects. Since the current scene is still "remote control," the characteristics of the remote control scene need to be considered when processing demands, such as displaying reminder information synchronously through a mobile app. When extracting these state change factors, it is necessary to perform structured analysis of the round state evolution unit content to distinguish different types of state change factors.

[0068] Step S142: Based on the driving path content in the bidirectional correlation evolution model, trace all previous round state evolution unit contents of the current round state evolution unit content, and extract the demand sub-item change content, system response optimization impact content, and scene switching correlation impact content in each previous round state evolution unit content.

[0069] Based on the driving path content in the bidirectional correlation evolution model, starting from the current round (round 8), the state evolution unit content of all previous rounds (rounds 1 to 7) is traced back along the timeline. The driving path content records the sequential influence relationship between rounds. By querying this path content, the correlation between each previous round and the current round can be clarified. For each previous round, the content of demand sub-item changes, system response optimization impact, and scene switching correlation impact must be extracted. For example, the demand sub-item change content in round 2 is "open the curtains and turn on the smart bedroom light at 7 am," and the system response optimization impact content is "turn on the smart bedroom light to a suitable brightness"; the scene switching correlation impact content in round 7 is "the scene changes from indoor interaction to remote control, and some demand sub-items are adjusted to energy-saving mode." During the extraction process, the state evolution unit content of each previous round needs to be analyzed in the same structured manner as the current round to ensure that the extracted content accurately reflects the state changes of each round. Through the above tracing and extraction, we can understand the formation process and influencing factors of the current round's state evolution unit content, and clarify the specific impact of each preceding round on the current round.

[0070] Step S143: Based on the backlinking rules in the bidirectional correlation evolution model, filter the subsequent round state evolution unit content that has a backlinking impact on the current round state evolution unit content, and extract the corresponding requirement sub-item correction content and priority adjustment result content.

[0071] In the bidirectional association evolution model, the backtracking association rules define the conditions and methods by which subsequent rounds influence previous rounds. For example, when a subsequent round modifies a requirement item from a previous round, a backtracking association effect is triggered. Based on these rules, subsequent rounds that have a backtracking association effect on the current round's (round 8) state evolution unit are selected. Suppose that in the subsequent round 9 interaction, the user modifies the requirement item "smart refrigerator reminds you to clean expired food" from round 8, changing the reminder method from playing only through the smart speaker to simultaneously pushing a message on the mobile app, and adjusting the priority of this requirement item from medium to high. Then, it is necessary to extract the modified content and priority adjustment results of these requirement items. During extraction, it is necessary to determine, based on the backtracking association rules, which content in subsequent rounds is relevant to the current round, and accurately extract the relevant modified content and priority adjustment results, incorporating them into the derivation process of the current interaction state to reflect the backtracking influence of subsequent rounds on the current round.

[0072] Step S144: Associate and integrate the state change factors of the current round state evolution unit with the relevant influencing factors of all previous round state evolution units to form a forward evolution influence set, which includes the positive state cumulative change from the initial interaction state to the current round.

[0073] First, establish the correspondence between the current round (round 8) and all previous rounds (rounds 1 to 7), clarifying the specific impact of state change factors from each previous round on the state changes in the current round. This requires determining, based on the driving path content, which state change factors from each previous round affect which aspects of the current round. For example, the change in the "turn on the smart bedroom light" requirement sub-item in round 2 affects the lighting control-related states in the current round; the scene switching content in round 7 affects the device operating mode settings in the current round. Then, integrate the above relationships, classifying and summarizing the influencing factors from each previous round according to different dimensions such as requirement sub-item association, system response association, and scene association. During the integration process, the weight of the influencing factors from each previous round needs to be considered; factors closer to the current round in time and with greater influence have higher weights. For example, round 7 is closest to the current round in time, so its scene switching factor has a higher weight in influencing the current round; round 1 is the furthest away in time, so its influence weight is lower. Finally, a forward evolution impact set is formed, which systematically presents all positive state cumulative changes from the initial interaction state to the current round, including the addition, supplementation and correction of requirement sub-items, optimization of system response and the impact of scene switching, etc.

[0074] Step S1441: Taking the current round state evolution unit content as the endpoint content, extract the state change factor content of each previous round state evolution unit content in reverse time order. The state change factor content of each previous round includes the demand sub-item change content, the system response optimization impact content, and the scene switching related impact content.

[0075] Starting with the current round (round 8) of state evolution unit content as the endpoint, extract the state change factors from the rounds 7, 6, ..., 1 in reverse chronological order. The purpose is to trace back from the most recent preceding round to gradually understand the formation process of the current round's state. For each preceding round, extract the complete content of its requirement sub-item changes, system response optimization impacts, and scene switching related impacts. For example, when extracting round 7, the requirement sub-item change content might be "no new or modified requirement sub-items," the system response optimization impact content might be "adjust the living room temperature control to energy-saving mode," and the scene switching related impact content might be "switch from indoor interaction to remote control." When extracting round 6, the requirement sub-item change content might be "add the requirement for the living room smart speaker to automatically play light music at 8 PM," the system response optimization impact content might be "set the smart speaker playlist to light music," and the scene switching related impact content might be "no scene switching." During the extraction process, it is necessary to ensure that the state change factors of each preceding round are accurately extracted and recorded in a uniform format for subsequent correlation and integration processing. By extracting in reverse chronological order, the formation process of the current round's state can be traced, revealing the specific impact of each preceding round on the current state.

[0076] Step S1442: For each previous round of changes in the preceding requirement sub-item, analyze its relationship with the current round's requirement sub-item content, and record the continuation or expansion effect of the previous requirement sub-item changes on the current requirement sub-item content.

[0077] For each change in the sub-item of the requirement in the previous round, carefully analyze its relationship with the sub-item of the requirement in the current round. This requires comparing the changes in the sub-items of the requirement in the previous round with the sub-items of the requirement in the current round one by one to find the connections. For example, if the change in the sub-item of the requirement in round 2 was "turn on the smart bedroom light at 7 am when the curtains are opened," and the sub-items of the requirement in the current round (round 8) still contain content related to "controlling the smart bedroom light," it means that the change in the sub-item of the requirement in the previous round has a continuing effect on the sub-item of the requirement in the current round, that is, the current sub-item of the requirement is developed based on the sub-item of the requirement in round 2. As another example, if the sub-item of the requirement "automatically play light music on the smart speaker in the living room at 8 pm" added in round 6 is expanded in the current round to "automatically adjust the music type according to the user's mood," then record the expanding effect of the change in the sub-item of the requirement in the previous round on the sub-item of the requirement, that is, the current sub-item of the requirement has added new functions or conditions on the original basis. Through the above analysis and recording, we can clearly understand the continuation and expansion of changes in previous requirement sub-items in the current round, ensuring the integrity and coherence of the current requirement sub-items.

[0078] Step S1443: Analyze the system response optimization impact content of each previous round, determine its preparatory role for the system response direction content of the current round, and record the continuation effect of the optimization impact content in the current round.

[0079] Analyze the impact of system response optimizations in previous rounds to determine their foundational role in shaping the current round's response direction. System response optimizations refer to improvements and adjustments made to the system's response methods in previous rounds to meet user needs; these optimizations may lay the groundwork for the current round's response direction. For example, the optimization of the bedroom smart light's brightness (turning it to a suitable brightness) in round 3 laid the foundation for the system's response direction in adjusting the brightness of other lighting devices in the current round—setting appropriate brightness based on user habits and scenario requirements. Record the continuation effect of this optimization in the current round; for instance, the adjustment of the living room lights in the current round also references similar brightness setting logic, automatically adjusting brightness based on different time periods and user activities. During the analysis, it's necessary to analyze the specific measures and effects of the system response optimizations in previous rounds, and how these measures and effects influence the current round's system response direction. Through this analysis and recording, we can understand the historical context of system response optimizations and their impact on the current response direction, making the system response more aligned with user expectations.

[0080] Step S1444: Evaluate the impact of scene switching in each previous round, analyze its foundational role in the current round's scene-related content, and record the impact traces retained in the current round after the scene switching content.

[0081] Assess the impact of scene transitions in previous rounds and analyze their foundational role in the current round's scene-related content. Scene transition impact refers to the influence of scene transitions in previous rounds on the interaction state; these impacts may continue into the current round, laying the foundation for the current scene-related content. For example, the scene transition in round 7 from "indoor interaction" to "remote control" establishes that the scene-related content in the current round (round 8) will still be "remote control." Record the traces of impact retained in the current round after the scene transition, such as the device operating mode remaining in energy-saving mode, control commands being sent via remote network transmission, and the user interface displaying a dedicated remote control interface. During the assessment, it is necessary to analyze the impact of scene transitions on various interaction parameters and the specific manifestation of these impacts in the current round. Through the above assessment and recording, the continuous impact of scene transitions on the current interaction state can be clearly identified, ensuring that the system's various functions and responses in the current scenario accurately adapt to the scenario requirements.

[0082] Step S1445: Assign associated weight content to the state change factor content of each previous round, and perform corresponding association processing between the state change factor content of each previous round and its associated weight content and the state change factor content of the current round to form multiple sets of inter-round associated combination content. Each set of combination content contains the association relationship between a previous factor content and a current factor content.

[0083] Assign association weights to the state change factors from each preceding round, taking into account both the time distance and the degree of influence between the preceding and current rounds. Preceding rounds with a closer time distance and a greater impact on the current round are assigned higher weights, while those with a greater time distance and a smaller impact are assigned lower weights. For example, round 7 is the closest in time to the current round, and its scene transition has a significant impact, thus it receives a higher weight; round 1 is the furthest in time, and its impact is relatively smaller, so it receives a lower weight. A clear set of evaluation criteria needs to be established when assigning association weights. For example, the time distance and the degree of influence can be divided into several levels, with different levels corresponding to different weight values. Then, associate the state change factors from each preceding round and their associated weights with the state change factors from the current round, forming inter-round association combinations. Examples include (round 7 scene transition factor, high weight, round 8 scene-related factor), (round 2 requirement sub-item supplement factor, medium weight, round 8 requirement sub-item content), etc. Through the above correlation processing, the degree of influence of previous rounds on the current round can be quantified.

[0084] Step S14451: Extract the time series content of the multi-turn dialogue, record the difference between the content of the current turn number and the content of each previous turn number, and determine the time distance parameter content according to the size of the difference content. The smaller the difference content, the larger the proportional coefficient content of the time distance parameter content.

[0085] Extract the round time series of multi-turn dialogues, identifying the current round number and the numbers of each preceding round. The round time series is a set of round numbers arranged in the chronological order of the dialogue, such as [1,2,3,4,5,6,7,8]. Calculate the difference between the current round number and each preceding round number; the smaller the difference, the closer the preceding round is to the current round in time. For example, if the current round is round 8, the difference between round 7 and the current round is 1, the difference between round 6 and the current round is 2, and so on. Determine the time distance parameter based on the magnitude of the difference; the proportional coefficient corresponding to the time distance parameter increases as the difference decreases. This is because preceding rounds that are closer in time usually have a greater impact on the current round, so a higher proportional coefficient is needed. For example, a larger proportional coefficient is set when the difference is 1, a second larger value is set when the difference is 2, and the proportional coefficient gradually decreases as the difference increases. Through this method, preceding rounds that are closer in time have a higher proportional coefficient in the association weight allocation, reflecting their greater influence on the current round.

[0086] Step S14452: Analyze the number of sub-factors contained in the state change factors of each previous round, and count the coverage of each sub-factor to the state change factors of the current round.

[0087] Analyze the number of sub-factors included in the state change factors of each preceding round. Different preceding rounds may contain different numbers of sub-factors. For example, the state change factors in round 7 include two sub-factors: scene switching and adjustment of the correlation degree of requirement sub-items; the state change factors in round 2 include one sub-factor: supplementation of requirement sub-items. Simultaneously, calculate the coverage of each sub-factor with the state change factors of the current round, i.e., which aspects of state change the current round are affected by the sub-factor. For example, the scene switching sub-factor in round 7 covers the scene-related content of the current round, and the requirement sub-item correlation degree adjustment sub-factor covers the requirement sub-item priority content of the current round; the requirement sub-item supplementation sub-factor in round 2 covers the lighting control part of the requirement sub-item content of the current round. During the analysis and statistics, each sub-factor and its corresponding coverage need to be listed in detail to determine the degree of influence later. Through the above analysis and statistics, the specific coverage of each sub-factor with the state change of the current round can be determined.

[0088] Step S14453: Evaluate the depth of the impact of the state change factors in each previous round on the core requirement dimension of the current round, and record the number of key related points of influence.

[0089] Assess the depth of impact of each preceding round's state change factors on the core needs dimensions of the current round. Core needs dimensions include curtain control, temperature control, air purification, kitchen appliance control, and time mode adjustment. The depth of impact is measured by the number of key correlations, which are specific aspects that significantly influence the core needs dimension. For example, the bedroom smart light requirement added in round 2 has a significant impact on the lighting control core needs dimension, with key correlations including light on-time and brightness adjustment. Scene switching in round 7 has a certain impact on the depth of impact of all core needs dimensions, with key correlations including device operating modes and control methods. During the assessment, it is necessary to analyze the specific impact of preceding rounds' state change factors on each core needs dimension and count the number of key correlations. A higher number of key correlations indicates a greater depth of impact.

[0090] Step S14454: Integrate the coverage content and the number of key related points into the influence parameter content. The wider the coverage content and the more key related points the content contains, the larger the proportional coefficient content corresponding to the influence parameter content will be.

[0091] The coverage scope and number of key related points of each preceding round are integrated to form an influence degree parameter. A wider coverage indicates a broader impact of the state change factors in the preceding round on the state change of the current round; a greater number of key related points indicates a deeper impact. Combining these two aspects, preceding rounds with wider coverage and more key related points have a larger proportional coefficient corresponding to the influence degree parameter. For example, round 7 has a wider coverage (involving scenario associations and requirement sub-item priorities) and a larger number of key related points, resulting in a larger proportional coefficient for the influence degree parameter; round 2 has a narrower coverage (only involving requirement sub-item content) and fewer key related points, resulting in a smaller proportional coefficient for the influence degree parameter. During the integration process, the coverage scope and the number of key related points can be divided into different levels, and then the proportional coefficient can be determined according to the level. This integration comprehensively reflects the degree of influence of preceding rounds on the current round.

[0092] Step S14455: Construct the association weight calculation process, wherein the association weight is determined based on the time proximity level and the significance level of influence.

[0093] First, the temporal proximity level is determined based on the time distance parameter; the closer the time distance, the higher the temporal proximity level. Then, the influence significance level is determined based on the influence degree parameter; the greater the influence degree, the higher the influence significance level. The calculation of the association weight combines these two levels. Specifically, the temporal proximity level and influence significance level can be combined according to a certain weight ratio. For example, the temporal proximity level accounts for 40%, and the influence significance level accounts for 60%. A weighted average is obtained, and then the association weight is determined based on the overall level. Previous rounds with both high temporal proximity and high influence significance levels have a larger association weight; previous rounds with low temporal proximity or low influence significance levels have a smaller association weight. By constructing the above calculation process, the determination of association weights can be made more scientific and reasonable, accurately reflecting the comprehensive influence of previous rounds on the current round.

[0094] Step S14456: Determine the time proximity level for each preceding round's state change factor based on the time distance parameter content. The closer the time distance content, the higher the time proximity level.

[0095] The temporal proximity level is determined based on the temporal distance parameter, which measures the difference between the current round number and the previous round number. A smaller difference indicates a closer temporal distance and a higher temporal proximity level. For example, the temporal proximity level can be divided into 5 levels: a difference of 1 corresponds to level 5 (highest), a difference of 2 to level 4, a difference of 3 to level 3, a difference of 4 to level 2, and a difference of 5 or higher to level 1 (lowest). For the current round being round 8, round 7 has a difference of 1 and a temporal proximity level of 5; round 6 has a difference of 2 and a level of 4; round 5 has a difference of 3 and a level of 3; round 4 has a difference of 4 and a level of 2; and rounds 1-3 have a difference of 5 or higher and a level of 1.

[0096] Step S14457: Based on the impact degree parameter, determine the impact significance level of each previous round's state change factor content. The wider the coverage and the more key related points, the higher the impact significance level.

[0097] The significance level of influence is determined for each preceding round's state change factors based on the influence degree parameter. The influence degree parameter is composed of the coverage area and the number of key related points; the wider the coverage area and the greater the number of key related points, the higher the influence degree parameter and the higher the significance level. For example, the significance level can be divided into 5 levels, with the preceding round having the highest influence degree parameter at level 5 (highest), the second highest at level 4, and so on. Assume that round 7 has the highest influence degree parameter, with a significance level of 5; round 2 has the second highest at level 4; round 5 has a moderate influence degree parameter at level 3; and other preceding rounds have their corresponding significance levels determined based on their influence degree parameters. When determining the level, it is necessary to make reasonable divisions based on the actual coverage area and the distribution of key related points to ensure that the level accurately reflects the differences in influence degree. Through the above level division, the significance level of influence for different preceding rounds is clearly defined.

[0098] Step S14458: Based on the time proximity level and the significance level of the impact, determine a comprehensive level for the state change factors of each previous round.

[0099] Taking into account both the time proximity level and the significance level of influence, a comprehensive level is determined for each preceding round's state change factors. The comprehensive level can be determined by weighting the time proximity level and the significance level according to certain weights. For example, if the time proximity level accounts for 40% and the significance level accounts for 60%, then the comprehensive level = time proximity level × 40% + significance level × 60%. Assuming the time proximity level and significance level are both 5 in round 7, the comprehensive level = 5 × 40% + 5 × 60% = 5; the time proximity level is 1 in round 2 (difference of 6) and the significance level is 4, so the comprehensive level = 1 × 40% + 4 × 60% = 0.4 + 2.4 = 2.8; the time proximity level is 3 in round 5 and the significance level is 3, so the comprehensive level = 3 × 40% + 3 × 60% = 3. Through the above calculations, the comprehensive level of each preceding round is obtained. The higher the comprehensive level, the greater the overall impact of the preceding round on the current round. By determining the overall level, the temporal proximity and significance of the preceding rounds can be comprehensively reflected.

[0100] Step S14459: Normalize the comprehensive level of all previous round state change factors to generate the associated weight value of each previous round state change factor.

[0101] The overall rating of all preceding rounds' status change factors is normalized, converting the overall rating into a correlation weight value between 0 and 1. The purpose of normalization is to ensure that the sum of the correlation weight values ​​for all preceding rounds is 1, facilitating subsequent correlation integration analysis. Specifically, the normalization method is to divide the overall rating of each preceding round by the sum of the overall ratings of all preceding rounds. For example, if the sum of the overall ratings of all preceding rounds is 20, and the overall rating of round 7 is 5, then its correlation weight value is 5 / 20 = 0.25; the overall rating of round 2 is 2.8, and the correlation weight value is 2.8 / 20 = 0.14; the overall rating of round 5 is 3, and the correlation weight value is 3 / 20 = 0.15. This normalization process ensures that the correlation weight values ​​accurately reflect the relative influence of each preceding round; the higher the overall rating of a preceding round, the greater its correlation weight value.

[0102] Step S144510: Normalize the associated weight values ​​of all previous round state change factors, and associate and store the normalized associated weight values ​​with the corresponding previous round state change factors and round number to form a weight allocation record set.

[0103] The associated weight values ​​are normalized again to ensure that rounding errors during calculation do not affect the total weight sum of 1. Specifically, all associated weight values ​​after the first normalization are added together to obtain a sum. If the sum is not equal to 1, each associated weight value is divided by this sum to re-normalize it. For example, if the total weight sum after the first normalization is 1.02, each weight value is divided by 1.02 to obtain a new weight value, ensuring the sum is 1. Then, the normalized associated weight values ​​are associated and stored with the corresponding previous round's state change factors and round number, forming a weight allocation record set. For example, (Round 7, Scene Switching Factor, 0.25), (Round 2, Requirement Sub-item Supplement Factor, 0.14), (Round 5, System Response Optimization Factor, 0.15), etc. This associated storage facilitates quick access and use of the associated weight values ​​from each previous round during subsequent association integration, ensuring the accuracy and efficiency of the association integration process.

[0104] Step S1446: Classify the related content of all rounds into three dimensions: related content of requirement sub-items, related content of system response, and related content of scenario. Each dimension contains the corresponding related content.

[0105] The inter-round related content is categorized into three dimensions: demand sub-item association, system response association, and scenario association. The demand sub-item association dimension includes the association between changes in demand sub-items in previous rounds and the content of demand sub-items in the current round, such as (Round 2 demand sub-item supplement, 0.14, Round 8 lighting control demand sub-item). The system response association dimension includes the association between the impact of system response optimization in previous rounds and the direction of system response in the current round, such as (Round 3 system response optimization, 0.12, Round 8 lighting brightness adjustment response). The scenario association dimension includes the association between the impact of scene switching in previous rounds and the scenario-related content in the current round, such as (Round 7 scene switching, 0.25, Round 8 remote control scenario). During the classification process, it is necessary to accurately categorize the related content into the corresponding dimensions based on its nature and the fields involved. This classification process makes the related content more organized and clear, facilitating subsequent overlay and integration analysis, and improving analytical efficiency and accuracy.

[0106] Step S1447: Within each dimension, the preceding factor influence content in the associated combination content is overlaid with the current factor content to form the cumulative influence result content under that dimension. The cumulative influence result content includes the sum of the state changes of that dimension from the initial state to the current state.

[0107] Within the demand sub-item association dimension, the influence of preceding factors in each association combination is superimposed with the content of the current factor. For example, the influence of the supplementary factor of the second round of demand sub-item on the current lighting control demand sub-item (weight 0.14) is superimposed with the influence of other preceding rounds on this demand sub-item (such as the correction factor in the fifth round, weight 0.15) to obtain the cumulative influence result of the demand sub-item association dimension. The superposition method can be to add the weights of each association combination and combine them with the specific influence content for comprehensive analysis. Similarly, in the system response association dimension and the scene association dimension, the influence of each association combination is superimposed to form the cumulative influence result of the corresponding dimension. Through the above superposition process, the total state change from the initial to the current state under each dimension can be comprehensively reflected.

[0108] Step S1448: Integrate the cumulative impact results from the three dimensions, supplement the round source information corresponding to each cumulative impact result, and record the previous round from which the cumulative impact originated.

[0109] The cumulative impact results from three dimensions—requirement sub-item association, system response association, and scenario association—are integrated to form a comprehensive cumulative impact result. During integration, the results from the three dimensions need to be summarized to ensure that all state changes are reflected without duplication or omission. Simultaneously, round-based source information is added to each cumulative impact result, recording which preceding rounds the cumulative impact originated from. For example, the cumulative impact of a certain requirement sub-item might originate from the influence of preceding factors in rounds 2 and 5; the round-based source information would then be recorded as rounds 2 and 5. Through this integration and supplementation, the source of each cumulative impact result can be identified.

[0110] Step S1449: Incorporate the integrated cumulative impact result content and the round source information content into the forward evolution impact set content, which includes the positive state cumulative change content from the initial interaction state to the current round.

[0111] The integrated cumulative impact results and round-based source information are combined into the forward evolution impact set, ensuring that this set fully encompasses all positive state cumulative changes from the initial interaction state to the current round. The forward evolution impact set is stored in a structured data format, such as a table or tree structure, containing the cumulative impact results for each dimension, the corresponding round-based source information, and related weight information. Through this set, a comprehensive understanding of the formation process and historical influencing factors of the current interaction state can be obtained.

[0112] Step S145: Merge the state change factors of the current round state evolution unit with the retrospective correlation factors of the subsequent round state evolution unit to form a reverse retrospective influence set, which includes the corrective effect of subsequent interactions on the current round state.

[0113] Assuming the current round is round 8, with subsequent rounds 9 and 10, and these later rounds have a retrospective impact on round 8, the state change factors of round 8 are integrated with the retrospective impact factors of rounds 9 and 10. For example, round 9 modifies the "smart refrigerator reminds you to clean expired food" requirement from round 8, changing the reminder method from playing only through the smart speaker to simultaneously pushing a message to the mobile app, and adjusting the priority of this requirement from medium to high. The integration process involves incorporating these modifications into the state change factors of round 8, specifically updating the attributes of the requirement (reminder method and priority). Through this integration process, a set of reverse retrospective impacts is formed, reflecting the corrective effects of subsequent interactions on the current round's state, ensuring that the current interaction state reflects the latest requirement changes and priority adjustments.

[0114] Step S146: Using the core requirement dimension content of the initial interaction state as the baseline content, perform correlation matching processing on each factor content in the forward evolution influence set content, and record the direction of influence of each factor content on the core requirement dimension content.

[0115] Based on the core requirement dimensions of the initial interaction state (such as curtain control, temperature control, air purification, kitchen appliance control, time mode adjustment, etc.), each factor in the forward evolution influence set is matched with the core requirement dimension. For example, if a factor complements a requirement sub-item under the temperature control dimension, matching determines that it belongs to the temperature control core requirement dimension. Then, the direction of this factor's effect on the core requirement dimension is recorded. The direction can be positive enhancement (e.g., adding a new requirement sub-item), negative weakening (e.g., deleting a requirement sub-item), or adjustment (e.g., modifying the parameters of a requirement sub-item). Through this correlation matching and recording of the direction of effect, the specific impact of each factor on the core requirement dimension can be clearly defined.

[0116] Step S147: Based on the corrected backtracking marker contained in the backtracking influence set, locate the corresponding factor content in the forward evolution influence set, and adjust the weight value associated with the factor content by increasing or decreasing according to the type of the corrected backtracking marker.

[0117] The correction backtracking marker is found within the backtracking impact set. This marker indicates the changes made by subsequent rounds to the requirements of previous rounds, such as changes to reminder methods or priority adjustments. Based on this marker, the corresponding requirement sub-item factor is located within the forward evolution impact set. For example, the correction backtracking marker for the requirement sub-item "Smart refrigerator reminds you to clean expired food" in round 9 corresponds to the same requirement sub-item factor in round 8 of the forward evolution impact set. The weight associated with this factor is adjusted according to the type of correction backtracking marker. If the correction is an optimization made because the user felt the previous settings were unreasonable, the weight associated with this factor is increased to reflect its increased importance; if the correction is an error caused by user misoperation, the weight is decreased. Through this quantitative adjustment, the forward evolution impact set reflects the backtracking correction effect of subsequent rounds.

[0118] Step S148: Integrate the adjusted forward evolution influence set and the backward backtracking influence set, and extract core influencing factors from them according to preset filtering rules. The filtering rules include: the correlation weight value of the factor is higher than the weight threshold, or the number of core feature fields of the current interaction state explained by the factor exceeds the number threshold.

[0119] The adjusted forward evolution influence set and backward regression influence set are integrated to form a comprehensive influence set. Then, core influencing factors are extracted from this set according to preset screening rules. These rules include two aspects: first, the correlation weight value of a factor is higher than a set weight threshold (e.g., a weight threshold of 0.2), and factors with a correlation weight value greater than 0.2 are selected; second, the number of core feature fields of the current interaction state explained by a factor exceeds a quantity threshold (e.g., a quantity threshold of 3), and factors that can explain 3 or more core feature fields are selected. Through this screening process, core factors that have a significant impact on the current interaction state can be extracted.

[0120] For example, step S1481: Extract all factor content from the adjusted forward evolution influence set and classify and organize it according to the content related to the demand sub-items, the content related to the system response, and the content related to the scenario.

[0121] All factors in the adjusted forward evolution impact set were extracted, covering multiple aspects such as requirement sub-items, system responses, and scenario associations. These factors were then categorized and organized according to the dimensions of requirement sub-item association, system response association, and scenario association. The requirement sub-item association dimension includes factors affecting the changes in each requirement sub-item, such as additions, corrections, and deletions; the system response association dimension includes factors affecting system response optimization, such as algorithm adjustments and interface expansions; and the scenario association dimension includes factors affecting scenario switching, such as changes in scenario type and parameter adjustments. This categorization makes the factor content more organized and facilitates subsequent comparative analysis with the backward backtracking impact set.

[0122] Step S1482: Extract all factor content from the reverse backtracking influence set and classify and organize it according to the same three dimensions.

[0123] The factors in the backtracking impact set are also categorized and organized according to three dimensions: demand sub-item association, system response association, and scenario association. The classification criteria are consistent with those of the forward evolution impact set. For example, demand sub-item modification factors in the backtracking impact set are categorized under the demand sub-item association dimension, system response modification factors under the system response association dimension, and scenario association modification factors under the scenario association dimension. This classification ensures that the factors in both sets can be compared under the same criteria, facilitating subsequent analysis of their direction and intensity of influence.

[0124] Step S1483: Within each dimension, compare the direction of influence of forward evolution factors with that of backward regression factors, and mark the combination of factors with the same direction and the combination of factors with opposite directions.

[0125] Within each dimension, the content of forward evolutionary influencing factors and backward retrospective influencing factors are compared one by one to analyze whether their directions of action are consistent. For example, in the dimension of demand sub-item association, if the forward factor is to increase the preheating time of the smart oven, and the backward factor is also to increase the preheating time, it indicates that their directions of action are consistent, and this combination is marked as a factor combination with consistent directions. If the forward factor is to increase the temperature, and the backward factor is to decrease the temperature, then their directions of action are opposite, and this combination is marked as a factor combination with opposite directions. Through the above comparison and marking, the relationship between forward and backward influencing factors can be understood.

[0126] Step S1484: For the combination of factors with the same direction, based on a unified evaluation scale, the evaluation values ​​of the influence of the two factors are superimposed to generate the comprehensive influencing factor content under this dimension. The intensity evaluation value of the comprehensive influencing factor content is the result of the superposition.

[0127] For combinations of factors with consistent directions, the strength evaluation values ​​of the forward evolutionary influencing factors and the backward retrospective influencing factors are superimposed based on a unified evaluation scale. The evaluation scale can be a value between 0 and 1, where 0 represents no influence and 1 represents the greatest influence. For example, if the strength evaluation value of the forward factor is 0.5 and the strength evaluation value of the backward factor is 0.3, the strength evaluation value of the combined influencing factor after superposition is 0.8. Through the above superposition process, the comprehensive influencing factor content under this dimension is generated, comprehensively reflecting the combined effect of forward and backward factors.

[0128] Step S1485: For combinations of factors with opposite directions, calculate the difference in their intensity evaluation values ​​based on a unified evaluation scale, retain the factor content with a positive difference as the comprehensive influencing factor content under that dimension, and update the intensity evaluation value of that factor content to the difference.

[0129] For combinations of factors with opposite directions, the difference in the evaluation values ​​of the forward evolutionary influencing factors and the backward retrospective influencing factors is calculated based on a unified evaluation scale. For example, if the evaluation value of the forward factor's influence is 0.6 and the evaluation value of the backward factor's influence is 0.4, the difference is 0.2. Since the forward factor has a greater influence, it is retained as the comprehensive influencing factor, and its influence evaluation value is updated to 0.2. If the backward factor has a greater influence, it is retained, and its influence evaluation value is the difference between the two (backward intensity minus forward intensity). Through the above processing, the comprehensive influencing factor content for combinations of factors with opposite directions is generated.

[0130] Step S1486: Extract the key indicator content from the core feature set of the current interaction state. Each key indicator content corresponds to a core attribute content of the current interaction state.

[0131] Extract key metrics from the core feature set of the current interaction state. These key metrics are important parameters reflecting the core attributes of the current interaction state. Examples include demand satisfaction, system response accuracy, scenario adaptability, and completion rate of sub-demands. Each key metric corresponds to a core attribute; for example, demand satisfaction corresponds to the extent to which user needs are met, and system response accuracy corresponds to the degree to which the system response matches the user's needs.

[0132] Step S1487: Compare each comprehensive influencing factor in the comprehensive influencing factor set with the predefined state fields in the core feature set of the current interaction state. If the data or logical reasoning of the comprehensive influencing factor can directly support or deduce the value of a certain state field, it is counted as a match; count the number of matched state fields for each comprehensive influencing factor.

[0133] Each comprehensive influencing factor in the comprehensive influencing factor set is compared with a predefined state field in the core feature set of the current interaction state. State fields are the specific manifestations of core features, such as "temperature control state" or "lighting control state." If the data or logical reasoning of a certain comprehensive influencing factor can directly support or deduce the value of a certain state field—for example, the factor "smart oven preheating time adjustment" can deduce the value of the "kitchen appliance control state" field—it is counted as a match. The number of matched state fields for each comprehensive influencing factor is counted; the more matched fields, the stronger the explanatory power of that factor for the current interaction state.

[0134] Step S1488: Select comprehensive influencing factors whose number of status fields exceeds the set value as candidate core influencing factors.

[0135] Set a threshold, such as 3, and filter out comprehensive influencing factors whose number of matching state fields exceeds this threshold as candidate core influencing factors. For example, if a comprehensive influencing factor matches 4 state fields, exceeding the threshold of 3, it will be selected as a candidate core influencing factor. Through this filtering process, factors that have a significant impact on the current interaction state can be preliminarily identified.

[0136] Step S1489: Analyze the relationships between factors in the candidate core influencing factor set. If two or more factors point to the same core feature field of the current interaction state, and the credibility score of the core feature field is higher than the score when any factor appears alone, then the factors are marked as a group of related factors. If a factor explains a significantly larger number of core feature fields than other factors, then it is marked as a key independent factor.

[0137] Analyze the relationships between factors in the candidate core influencing factor set to identify interrelated factors. If two or more factors point to the same core feature field, and the confidence score of that core feature field is higher when these factors appear simultaneously than when any one factor appears alone, it indicates a synergistic effect between these factors, and they are marked as a group of related factors. For example, if factors A and B both point to the "temperature control status" field, and the confidence score of this field is 0.9 when both appear simultaneously, higher than the 0.7 of factor A alone and the 0.6 of factor B alone, then A and B are marked as a group of related factors. If one factor explains significantly more core feature fields than other factors, for example, factor C explains 5 fields while other factors explain a maximum of 3, then it is marked as a key independent factor. Through the above analysis, the relationships and importance between the candidate core influencing factors can be further clarified.

[0138] Step S14810: Combine the related factor group with the key independent factors to form the final core influencing factor set.

[0139] By integrating related factor groups and key independent factors, a final set of core influencing factors is formed. Related factor groups, as a whole, affect the current interaction state, while key independent factors play important roles individually. Through this combination, the final set of core influencing factors comprehensively and accurately reflects the factors that have a key impact on the current interaction state.

[0140] Step S149: Integrate the core influencing factors with the core demand dimensions of the initial interaction state to form the core feature set of the current interaction state. This core feature set includes demand fulfillment status content, unfulfilled demand direction content, and state evolution trend content.

[0141] The core influencing factors are integrated with the core requirement dimensions of the initial interaction state. This integration process requires consolidating the effects of the core influencing factors on the core requirement dimensions. For example, the content related to temperature control among the core influencing factors is integrated with the initial temperature control dimension content to form the current temperature control state. The integrated core feature set includes requirement fulfillment status (e.g., which requirement sub-items have been met and to what extent), unfulfilled requirement direction (e.g., which requirement sub-items have not yet been met and how they should be developed subsequently), and state evolution trend (e.g., whether the change trend of requirement sub-items is increasing or decreasing, and the direction of adjustment of the system response strategy). Through this integration process, a core feature set that comprehensively reflects the current interaction state is formed.

[0142] Step S1410: The core feature set content is presented in a structured manner to form the current interaction state. The current interaction state integrates the state evolution results of all rounds of interaction and the bidirectional correlation influence content.

[0143] The core feature set is presented in a structured manner, organized using a clear and standardized data format, such as JSON or XML. The structured presentation includes a requirement fulfillment state vector, an unfulfilled requirement matrix, and a state evolution trend curve. The requirement fulfillment state vector uses numerical values ​​to represent the degree of fulfillment of each requirement sub-item; the unfulfilled requirement matrix lists unfulfilled requirement sub-items and their priorities; and the state evolution trend curve shows the changes of each core feature over multiple rounds. Through this structured processing, the current interaction state is formed, which integrates the state evolution results from all rounds of interaction and the bidirectional correlation effects.

[0144] Step S150: Generate a dynamic guidance strategy for multi-turn dialogue based on the current interaction state. The dynamic guidance strategy adapts to the state evolution trend of subsequent interactions and adjusts the related logic of previous interactions in reverse.

[0145] When generating dynamic guidance strategies based on the current interaction state, the core feature set of the current interaction state is first analyzed, especially the state of fulfilled needs, unfulfilled needs, and state evolution trends. Based on the unfulfilled needs and state evolution trends, the possible development direction of subsequent interactions is predicted, thereby formulating corresponding guidance strategies. For example, if the unfulfilled need is "smart speaker music type setting," and the state evolution trend indicates that the user may want to adjust the music type according to their mood, then the dynamic guidance strategy may include asking the user about their favorite music types and recommending related music lists. Simultaneously, based on the bidirectional correlations in the current interaction state, the related logic of previous interactions is adjusted in reverse. For example, the correlation logic regarding the user's music preferences in previous rounds is strengthened to make the guidance strategy more targeted. Dynamic guidance strategies need to be flexible and able to be adjusted according to the actual situation of subsequent interactions to ensure that the guidance content adapts to the state evolution trend of subsequent interactions.

[0146] Step S151: Extract the demand satisfaction status content in the current interaction state, distinguish between demand sub-items that have been satisfied and demand sub-items that have not yet been satisfied, and determine the processing order of demand sub-items that have not yet been satisfied according to preset rules. The preset rules include the order specified by the user and the logical dependencies between demand sub-items.

[0147] Extract the demand fulfillment status from the current interaction state, which displays the fulfillment status of each demand sub-item. Based on the demand fulfillment status, distinguish between fulfilled and unfulfilled demand sub-items. For example, demand sub-items such as "open the curtains at 7 am" and "bedroom temperature control" are fulfilled, while demand sub-items such as "smart refrigerator cleaning reminder" and "smart speaker music playback" are not. Then, determine the processing order of the unfulfilled demand sub-items according to preset rules. Preset rules include the user-specified order, such as the user explicitly stating that "smart speaker music playback" should be processed before "smart refrigerator cleaning reminder"; and logical dependencies between demand sub-items, such as "smart oven preheating" needing to be done after "ingredient preparation". By applying these preset rules, the processing order of unfulfilled demand sub-items is determined, ensuring that the guidance strategy can orderly drive the dialogue forward.

[0148] Step S152: Analyze the unfinished demand direction content in the current interaction state, and combine it with the state evolution trend content in the bidirectional correlation evolution model to deduce the subsequent evolution path content of the unfinished demand sub-item content.

[0149] By analyzing the unfulfilled needs in the current interaction state, we can understand what the user's unmet needs are and how they hope to evolve. Combined with the state evolution trend content in the bidirectional correlation evolution model, this trend reflects the changing patterns and development directions of the sub-items of the need across rounds. For example, if the unfulfilled need is "smart speaker music playback," and the state evolution trend shows that the user previously expanded from "playing light music" to "adjusting the music genre according to their mood," then we can deduce that subsequent evolutionary paths might include "selecting a specific music genre," "setting music playback duration," and "linking music playback with scene switching," etc.

[0150] Step S153: Based on the subsequent evolution path of the unfinished requirement sub-item content, generate targeted user expression guidance content. Each guidance content corresponds to a deepening or explicit path of the unfinished requirement sub-item content.

[0151] Based on the subsequent evolution path of the unfinished requirement sub-items, generate targeted user expression guidance content for each sub-item. This guidance content should be specific and clear, guiding users to further express their needs or make choices. For example, for the requirement sub-item "smart speaker music playback," if the subsequent evolution path is "select a specific music genre," the guidance content could be "What type of music would you like to play? Pop, classical, or something else?"; if the subsequent evolution path is "set music playback duration," the guidance content could be "How long would you like the music to play?" Each guidance content corresponds to a deepening or clarification path for an unfinished requirement sub-item, driving the completion of the requirement sub-item by guiding user responses.

[0152] Step S154: Based on the interaction history of the sub-items that have met the requirements, select the appropriate system response mode content and integrate the system response mode content into the response design phase of the subsequent guidance strategy.

[0153] Review the interaction history of met sub-items to analyze user feedback and preferences regarding system responses during the fulfillment of these sub-items. For example, users respond well to concise and clear responses for the "bedroom temperature control" sub-item; and are more satisfied with responses including confirmation information for the "curtain control" sub-item. Based on this interaction history, filter out suitable system response patterns, such as concise responses, confirmation responses, and recommendation responses. Integrate these selected system response patterns into the response design phase of subsequent onboarding strategies to ensure that subsequent system responses align with user preferences and improve user experience. For example, after guiding users to set their music genre, use a confirmation response: "We have recorded your preferred music genre as classical music; we will create a playlist for you."

[0154] Step S155: Combine the scene-related content in the current interaction state to adjust the expression style and information density of the guidance strategy so that the guidance content is adapted to the characteristics of the current dialogue scene.

[0155] The current interactive scenario is associated with "remote control," a scenario where the user may be away from home and needs to operate via a mobile device. Therefore, the guidance content needs to be concise and directional. Based on these characteristics, the guidance strategy should be adjusted to be concise and direct, avoiding redundant information; the information density should be moderate to ensure the user can quickly understand and operate. For example, the guidance could be "Please select music type: 1. Pop 2. Classical 3. Other," which is concise, clear, and suitable for quick remote operation. If the scenario is associated with "indoor interaction," the style can be more friendly and natural, and the information density can be appropriately increased, providing more details and options. Through these adjustments, the guidance content is adapted to the characteristics of the current dialogue scenario, improving the guidance effect.

[0156] Step S156: Based on the backtracking association logic content of the bidirectional association evolution model, associate the key guidance point content in the subsequent guidance strategy with the state evolution unit content of the previous round, so that the guidance strategy can reinforce the effective association content of the previous interaction in reverse.

[0157] Based on the retrospective association logic of the bidirectional association evolution model, this logic defines how to strengthen the effective associations in previous interactions through subsequent interactions. Key guidance points in subsequent guidance strategies (such as asking about music genres) are associated with state evolution unit content from previous rounds (such as the user mentioning they like listening to music at night). For example, when guiding a user to choose a music genre, the user's music preferences from previous rounds are mentioned: "You mentioned you like listening to light music at night; would you like to continue with light music this time?" Through this association processing, the effective associations regarding the user's music preferences in previous interactions are strengthened, making the guidance strategy more targeted and enhancing the user's consistent experience.

[0158] Step S1561: Extract the effective association identifier content of the preceding round state evolution unit content in the bidirectional association evolution model. Each effective association identifier content corresponds to a set of preceding interactive association content with a positive driving effect.

[0159] This study extracts effective association markers from the content of previous round state evolution units in a bidirectional association evolution model. Effective association markers refer to association marks that have a positive driving effect on the current or subsequent interactions in previous interactions. For example, in round 6, the user mentions "I like listening to light music at night," which corresponds to the positively driving previous interaction association content (the user's preference for music genres). By extracting these effective association markers, it becomes clear which association content in previous interactions is valuable for subsequent guidance strategies.

[0160] Step S1562: Analyze the core guiding content of the key guiding points in the subsequent guiding strategy, and record the related direction of the requirement sub-items and the system response adaptation direction of each key guiding point.

[0161] Analyze the key guidance points in the subsequent guidance strategy to clarify the core guiding content of each guidance point. For example, the core guidance of "selecting music genre" is to clarify the user's music preferences, the corresponding sub-item related to the entertainment device control dimension, and the system response adaptation direction is to provide music genre options and record the user's selection. Through the above analysis and recording, the purpose and direction of the key guidance points can be clarified.

[0162] Step S1563: Match the core guiding content of the key guiding point with the effective association identifiers of the previous round state evolution unit content, and filter out the effective association content of the previous round with consistent guidance.

[0163] The core guiding content of key guidance points is matched with the effective related content of previous rounds to identify consistent guiding principles. For example, the core guidance of "choosing a music genre" clarifies the user's music preferences, which aligns with the effective related guidance of "preferring light music in the evening" in round 6. Therefore, this effective related content is selected. Through this matching process, subsequent guidance strategies can be seamlessly integrated with previous interactive content, enhancing the targeting and coherence of the guidance.

[0164] Step S1564: Extract the feature data of the round state evolution unit content corresponding to the successfully matched preceding valid associated content, including: the name of the associated demand sub-item, the fragment of the system response content, and the dialogue scene tag.

[0165] Extract the feature data of the successfully matched preceding valid related content (such as round 6). The associated request sub-item name is "The smart speaker in the living room will automatically play light music at 8 pm," the system response content fragment is "Your request to play light music at 8 pm has been recorded," and the dialogue scenario label is "indoor interaction." By extracting this feature data, we can understand the specific details of the preceding valid related content.

[0166] Step S1565: The implementation path content of the key guiding point content is fused with the core feature content of the preceding effective related content to generate enhanced related adaptation content, which includes the enhanced path content of the guiding point to the preceding association.

[0167] The implementation path of the key guidance point "select music genre" (asking about genre - recording selection - setting playback) is integrated with the core feature of the preceding effective related content (user likes light music) to generate enhanced related content. For example, the implementation path of the guidance point is adjusted to "Based on your previous favorite light music, we recommend the following list of light music for you: 1. Classical light music 2. Natural sound effect light music 3. Piano light music. Please select or add other genres." This enhanced related content includes a path that reinforces the preceding association, that is, it guides the current choice by mentioning previous preferences, enhancing user identification and engagement.

[0168] Step S1566: Based on the enhanced correlation adaptation content, adjust the correlation weight allocation of the content of the previous round state evolution unit, and increase the influence ratio of the previous correlation content that is consistent with the content orientation of the key guidance point.

[0169] Based on the enhanced relevance of the content, the relevance weight allocation of the content in the previous rounds of state evolution units has been adjusted. For example, the weight of the relevance content "I like listening to light music at night" in round 6 has been adjusted from 0.1 to 0.2, increasing its influence in the overall relevance. Through these adjustments, the relevance content of the previous rounds that is consistent with the content orientation of the key guidance points plays a greater role in subsequent interactions, strengthening the effective relevance of the previous interactions.

[0170] Step S1567: Extract the weak links in the preceding interactive content, combine them with the core guiding content of the key guiding points, generate supplementary and adaptive content, and integrate the enhanced and adaptive content with the supplementary and adaptive content into the preceding association enhancement plan content. The preceding association enhancement plan content includes specific association adjustment path content and implementation sequence content.

[0171] Weak links in the preceding interactive content are identified, such as the lack of clear instructions regarding music playback duration. Combined with the core guidance of the key prompt "Select Music Type," supplementary content is generated, such as "After selecting a music type, is it necessary to set the playback duration?". This strengthened content is integrated with the supplementary content into a preceding association enhancement plan, clarifying the adjustment path (first strengthening music type preference, then supplementing playback duration settings) and the implementation order (guiding playback duration setting immediately after guiding music type selection). This integration forms a complete preceding association enhancement plan, comprehensively strengthening the effective association of preceding interactions.

[0172] Step S1568: Integrate the content of the preceding association enhancement scheme into the implementation process of the dynamic guidance strategy, and record the interaction round node content and execution conditions corresponding to each enhancement step.

[0173] The content of the preceding association enhancement plan is integrated into the implementation process of the dynamic guidance strategy, clarifying the position and execution conditions of each enhancement step within the guidance strategy. For example, when guiding the user to select a music type in the 9th round, enhanced association adaptation content is executed, and the execution condition is that the user has entered the music settings stage; when guiding the user to set the playback duration in the 10th round, supplementary association adaptation content is executed, and the execution condition is that the user has selected a music type. Through the above integration and recording, it is ensured that the preceding association enhancement plan can be implemented in an orderly and accurate manner.

[0174] Step S1569: By implementing the dynamic guidance strategy, the effective related content of the preceding interaction is strengthened. At the same time, the related supplementary and adapted content is used to improve the logical content of the preceding interaction, forming a reverse-strengthened preceding interaction related system content.

[0175] By implementing a dynamic guidance strategy, the effective related content of previous interactions (such as user music preferences) is strengthened, allowing users to perceive the continuity and relevance of the guidance content with previous interactions. Simultaneously, supplementary and adaptive content (such as playback duration settings) refines the logic of the previous interactions, making the system of related interactions more complete and logical. Ultimately, this results in a reverse-strengthened system of related interactions that better supports the guidance and fulfillment of needs in subsequent interactions.

[0176] Step S157: Plan the phased implementation of the guidance strategy, assign one or more planned interaction rounds to each unmet requirement sub-item, and set a clear guidance goal for each planned interaction round.

[0177] For unmet needs, such as "smart refrigerator cleaning reminders" and "smart speaker music playback," a phased implementation plan for the guidance strategy should be developed. One or more planned interaction rounds should be assigned to each need; for example, rounds 9 and 10 should be assigned to "smart speaker music playback," and round 11 to "smart refrigerator cleaning reminders." Specific guidance objectives should be set for each planned interaction round; for example, the guidance objective for round 9 is "determine the music type," the guidance objective for round 10 is "set the playback duration," and the guidance objective for round 11 is "set the refrigerator cleaning reminder time." Through this planning, the implementation of the guidance strategy will be systematic and goal-oriented, ensuring that each need is gradually met.

[0178] Step S158: Incorporate the predictive process content of state evolution trigger conditions, set alternative guidance schemes in the guidance strategy, and respond to possible user expression adjustments or scene switching content.

[0179] The guidance strategy incorporates a predictive process for state evolution trigger conditions, anticipating potential adjustments in user behavior or scene changes. For example, it anticipates that a user might adjust the playback time after selecting a music genre, or that the scene might switch from "remote control" to "indoor interaction." Alternative guidance scenarios are set up for these potential trigger conditions. For instance, if the user adjusts the playback time, the alternative scenario could be, "What time would you like to change the playback time from 8 PM to?"; if the scene switches to indoor interaction, the alternative scenario could be, "We have switched to indoor mode. Would you like to adjust the music volume?" Through these predictions and alternative scenario settings, the guidance strategy becomes flexible and adaptable, capable of handling various possible changes.

[0180] Step S159: Integrate the targeted guidance direction content, system response mode content, scenario adaptation adjustment content, preceding association reinforcement content, step-by-step implementation rhythm content, and alternative guidance plan content into the preliminary guidance strategy structure content.

[0181] The generated targeted guidance content, the selected system response patterns, the adjusted scenario adaptation content, the established pre-implementation reinforcement content, the planned phased implementation schedule, and the set alternative guidance schemes are integrated to form the initial guidance strategy structure. This initial guidance strategy structure is organized in a modular manner, with each module corresponding to an aspect of the guidance strategy, such as the guidance direction module, response pattern module, and scenario adaptation module.

[0182] Step S1510: Logically connect the components of the initial guidance strategy structure to ensure that the implementation of each component adapts to the state evolution trend of subsequent interactions. At the same time, adjust the related logical content of the preceding interactions in reverse through the reinforcement of the preceding associations to form a dynamic guidance strategy.

[0183] The initial guidance strategy structure is logically linked to ensure smooth transitions and a logical sequence. For example, it begins with scenario adaptation adjustments, followed by targeted guidance, then the application of a system response model, incorporating reinforcing content from preceding interactions, and proceeding in a phased manner, activating alternative guidance options when necessary. Throughout this process, each component is designed to adapt to the evolving state of subsequent interactions, such as adjusting playback duration guidance based on the user's music genre selection. Simultaneously, reinforcing content from preceding interactions is used to adjust the logic of those interactions, such as emphasizing the weight of user music preferences. This results in a dynamic guidance strategy that flexibly adapts to various changes during the interaction process, effectively guiding the dialogue and meeting user needs.

[0184] Figure 2 This application illustrates an intelligent interactive state tracking and reasoning system 100 for multi-turn dialogue, including a processor 1001, a memory 1003, and program code stored in the memory 1003. The processor 1001 executes the program code to implement the steps of the intelligent interactive state tracking and reasoning method for multi-turn dialogue.

[0185] The processor 1001 and memory 1003 are connected, for example, via a bus 1002. Optionally, the intelligent interactive state tracking and inference system 100 for multi-turn dialogue may further include a transceiver 1004. The transceiver 1004 can be used for data interaction between the intelligent interactive state tracking and inference system for multi-turn dialogue and other intelligent interactive state tracking and inference systems for multi-turn dialogue, such as sending and / or receiving data. It should be noted that in actual scheduling, the transceiver 1004 is not limited to one, and the structure of the intelligent interactive state tracking and inference system 100 for multi-turn dialogue does not constitute a limitation on the embodiments of this application.

[0186] The memory 1003 is used to store program code for executing the embodiments of this application, and its execution is controlled by the processor 1001. The processor 1001 is used to execute the program code stored in the memory 1003 to implement the steps shown in the foregoing method embodiments.

[0187] This application provides a computer-readable storage medium storing program code, which, when executed by a processor, can implement the steps and corresponding content of the aforementioned method embodiments.

[0188] The above description is only an optional implementation method for some implementation scenarios of this application. It should be noted that for those skilled in the art, other similar implementation methods based on the technical concept of this application, without departing from the technical concept of this application, also fall within the protection scope of the embodiments of this application.

Claims

1. An intelligent interaction state tracking and reasoning method applied to multi-turn dialog, characterized in that, The method includes: Retrieve the initial interaction state of a multi-turn dialogue, which includes the user's initial needs expressed when the dialogue starts, the system's initial response positioning content, and dialogue scenario-related content. Capture the state evolution triggering conditions for each round of interaction in a multi-round dialogue. The state evolution triggering conditions are related factors such as user expression adjustments, system response optimizations, and dialogue scene switching that cause changes in the interaction state. Based on the initial interaction state and the state evolution triggering conditions, a bidirectional correlation evolution model for multi-turn dialogue is constructed. The bidirectional correlation evolution model is used to characterize the driving correlation effect of the preceding interaction state on the subsequent interaction state and the retrospective correlation effect of the subsequent interaction state on the preceding interaction state. The current interaction state of a multi-turn dialogue is derived through the bidirectional correlation evolution model, and the current interaction state integrates the state evolution results of all rounds of interaction and the bidirectional correlation influence content. A dynamic guidance strategy for multi-turn dialogue is generated based on the current interaction state. The dynamic guidance strategy adapts to the state evolution trend of subsequent interactions and adjusts the related logic of previous interactions in reverse. The bidirectional relational evolution model for multi-turn dialogue, constructed based on the initial interaction state and the state evolution triggering conditions, includes: Extract the core requirement dimension content corresponding to the user's initial requirement expression content in the initial interaction state, and decompose the core requirement dimension content into multiple requirement sub-items. Each requirement sub-item content corresponds to a specific implementation direction of the initial requirement. Analyze the user expression adjustment content in the state evolution trigger condition content, distinguish the three types of content corresponding to the user expression adjustment content: supplementary content of requirement sub-items, modified content of requirement sub-items, and newly added content of requirement sub-items, and mark the interaction round node content corresponding to each type of content. Identify the system response optimization content in the state evolution trigger condition content, associate the system response optimization content with the corresponding requirement sub-item content, record whether the system response optimization content has completely solved, partially solved or not solved the corresponding requirement sub-item content, and record the interaction round node content corresponding to the identification result; Extract the dialogue scene switching content from the state evolution triggering condition content, analyze the correlation and impact of the dialogue scene switching content on all requirement sub-items, and mark the requirement sub-items affected by the scene switching content and the corresponding interaction round node content. Using the interaction round as the timeline, the content of changes in the sub-items of requirements, the impact of system response optimization, and the impact of scene switching in each interaction round are integrated into the content of the round state evolution unit. The content of each round state evolution unit includes all state change factors in that round. Establish the driving relationship between the content of the state evolution unit in the previous round and the content of the state evolution unit in the subsequent round, and derive the initial state basic content of the state evolution unit in the subsequent round based on the change content of the demand sub-items in the content of the state evolution unit in the previous round; Construct a retrospective relationship between the content of the state evolution unit in the subsequent round and the content of the state evolution unit in the previous round. If the content of a certain requirement sub-item is modified in the content of the state evolution unit in the subsequent round, a modification retrospective mark is added to the content of the requirement sub-item in the content of the state evolution unit in the previous round, and the weight value of the requirement sub-item in the subsequent state derivation is adjusted up or down according to the modification type. The driving and retrospective relationships of all round state evolution units are interwoven in a network to form a relational evolution structure containing multi-dimensional relational path content. Each relational path content corresponds to a set of mutually influential round state evolution units. The core requirement dimension content integrated into the initial interaction state serves as the basic anchor content for the related evolution structure content, so that all related path content unfolds its evolution process starting from the initial requirement content. The driving correlation strength, retrospective correlation range, and priority of demand sub-items in the correlation evolution structure are integrated into the parameter content of this bidirectional correlation evolution model to form a bidirectional correlation evolution model. This bidirectional correlation evolution model is used to present the driving correlation effect of the preceding interaction state on the subsequent interaction state and the retrospective correlation effect of the subsequent interaction state on the preceding interaction state.

2. The intelligent interaction state tracking and reasoning method for multi-turn dialogue as described in claim 1, characterized in that, The process of deriving the current interaction state of a multi-turn dialogue through the bidirectional correlation evolution model includes: Determine the content of the current round node in the multi-round dialogue, and use the content of the round state evolution unit corresponding to the content of the current round node as the starting point for derivation, and extract all state change factors in the content of the round state evolution unit. Based on the driving path content in the bidirectional correlation evolution model, trace the content of all previous rounds of state evolution units in the current round of state evolution unit content, and extract the demand sub-item change content, system response optimization impact content, and scene switching correlation impact content in each previous round of state evolution unit content; Based on the backlinking rules in the bidirectional correlation evolution model, the content of subsequent round state evolution units that have a backlinking impact on the content of the current round state evolution unit is selected, and the corresponding requirement sub-item correction content and priority adjustment result content are extracted. The state change factors of the current round state evolution unit are associated and integrated with the relevant influencing factors of all previous round state evolution units to form a forward evolution influence set, which includes the positive state cumulative change from the initial interaction state to the current round. The state change factors of the current round state evolution unit are merged with the retrospective correlation factors of the subsequent round state evolution units to form a reverse retrospective influence set, which includes the corrective effect of subsequent interactions on the current round state. Using the core requirement dimension content of the initial interaction state as the baseline content, perform correlation matching processing on each factor content in the forward evolution influence set content, and record the direction of each factor content's effect on the core requirement dimension content; Based on the corrected backtracking markers contained in the backtracking influence set, the corresponding factor content is located in the forward evolution influence set, and the weight value associated with the factor content is quantitatively adjusted by increasing or decreasing according to the type of the corrected backtracking marker. The integrated and adjusted forward evolution influence set and backward backtracking influence set are used to extract core influencing factors according to preset filtering rules. The filtering rules include: the correlation weight value of the factor is higher than the weight threshold, or the number of core feature fields of the current interaction state explained by the factor exceeds the number threshold. The core influencing factors are integrated with the core demand dimensions of the initial interaction state to form the core feature set of the current interaction state. This core feature set includes the demand fulfillment status, the direction of unfulfilled demands, and the state evolution trend. The core feature set content is presented in a structured manner to form the current interaction state. This current interaction state integrates the state evolution results of all rounds of interaction and the bidirectional correlation influence content.

3. The intelligent interaction state tracking and reasoning method for multi-turn dialogue as described in claim 1, characterized in that, The dynamic guidance strategy for generating multi-turn dialogues based on the current interaction state includes: Extract the demand fulfillment status content from the current interaction state, distinguish between demand sub-items that have been satisfied and demand sub-items that have not yet been satisfied, and determine the processing order of demand sub-items that have not yet been satisfied according to preset rules. The preset rules include the order specified by the user and the logical dependencies between demand sub-items. Analyze the unfinished demand directions in the current interaction state, and combine them with the state evolution trend in the bidirectional correlation evolution model to deduce the subsequent evolution path of the unfinished demand sub-items. Based on the subsequent evolution path of the unfinished requirement sub-item, generate targeted user expression guidance content, with each guidance content corresponding to a deepening or clarification path of the unfinished requirement sub-item; Based on the interaction history of the sub-items that have met the requirements, select the appropriate system response mode content and integrate the system response mode content into the response design stage of the subsequent guidance strategy. Based on the contextual content of the current interaction state, adjust the expression style and information density of the guidance strategy to make the guidance content adapt to the characteristics of the current dialogue scenario. Based on the back-tracking association logic of the bidirectional association evolution model, the key guidance points in the subsequent guidance strategy are associated with the state evolution unit content of the previous round, so that the guidance strategy can reinforce the effective association content of the previous interaction in reverse. Plan the phased implementation of the guidance strategy, assign one or more planned interaction rounds to each unmet need sub-item, and set a clear guidance goal for each planned interaction round; The process incorporates the prediction of state evolution trigger conditions and sets up alternative guidance schemes in the guidance strategy to deal with possible adjustments to user expressions or scene switching. The content of targeted guidance, system response mode, scenario adaptation and adjustment, pre-related reinforcement, phased implementation rhythm and alternative guidance solutions are integrated into the initial guidance strategy structure. The components of the initial guidance strategy structure are logically linked to ensure that the implementation of each component adapts to the state evolution trend of subsequent interactions. At the same time, the related logical content of the preceding interactions is adjusted in reverse through the reinforcement of the preceding related content, thus forming a dynamic guidance strategy.

4. The intelligent interaction state tracking and reasoning method for multi-turn dialogue as described in claim 1, characterized in that, The process of establishing the driving relationship between the content of the preceding round state evolution unit and the content of the following round state evolution unit, and deriving the initial state base content of the following round state evolution unit based on the changes in the demand sub-items in the content of the preceding round state evolution unit, includes: Traverse all round state evolution unit contents, arrange them in the time order of interaction rounds to form round sequence contents, and record the position of each round state evolution unit contents in the sequence and the relationship between adjacent rounds; Extract the changes in the requirement sub-items in the state evolution unit of the previous round, distinguish the three types of changes in the requirement sub-items: new content, supplementary content, and modified content, and record the specific content of the requirement sub-items corresponding to each type of change. Analyze the system response optimization content in the state evolution unit of the previous round, record the adaptation of the system response optimization content to the changes in the requirement sub-items, and record the content of the requirement sub-item satisfaction after adaptation. Extract the scene switching content from the state evolution unit content of the previous round, and mark the requirement sub-items that remain related after the scene switching content and the content whose degree of relatedness changes. The changes in the form of demand sub-items, the content of satisfied states, and the changes in the degree of scenario relevance in the state evolution unit of the previous round are integrated into the set of previous driving factors. Each driving factor corresponds to a specific factor that affects the subsequent state. Record the initial state components of the state evolution unit in subsequent rounds. The initial state components include the basic content of the initial requirement sub-items, the initial direction of the system response, and the initial setting of the scenario adaptation. The content of each driving factor in the preceding driving factor set is matched with the initial state constituent elements of the subsequent round state evolution unit, and the initial state constituent elements corresponding to each driving factor are recorded. Based on the matching results, the influence of the preceding driving factors is integrated into the initial state components of the subsequent round state evolution unit, and the priority of the initial demand sub-item basic content, the adaptation direction of the system response initial direction content, and the targeted content of the scenario adaptation initial setting content are adjusted. The adjusted initial state components are integrated into the initial state base content of the subsequent round state evolution unit content. This initial state base content fully inherits the driving and related influences of the previous round state evolution unit content. Record the driving relationship paths between the content of the state evolution unit in the preceding round and the content of the state evolution unit in the following round. Each driving relationship path record includes: the preceding round number, the following round number, the driving factor type, the name of the affected subsequent state constituent element, and the specific parameter value of the corresponding element after being affected.

5. The intelligent interaction state tracking and reasoning method for multi-turn dialogue as described in claim 2, characterized in that, The process of associating and integrating the state change factors of the current round's state evolution unit with the relevant influencing factors of all previous rounds' state evolution units to form a forward evolution influence set includes: Taking the current round's state evolution unit content as the endpoint content, extract the state change factors of each previous round's state evolution unit content in reverse chronological order. The state change factors of each previous round include the content of changes in demand sub-items, the content of system response optimization impact, and the content of scene switching related impact. For each previous round of changes to the sub-items of the preceding requirements, analyze their relationship with the sub-items of the current round of requirements, and record the continuation or expansion effect of the changes to the sub-items of the preceding requirements on the current sub-items. Analyze the impact of each previous round's system response optimization, determine its role in laying the groundwork for the current round's system response direction, and record the continuation effect of this optimization impact in the current round; Evaluate the impact of scene switching in each previous round, analyze its foundational role in the current round's scene-related content, and record the impact traces retained in the current round after scene switching. Assign associated weights to the state change factors of each previous round, and associate the state change factors and their associated weights with the state change factors of the current round to form multiple sets of inter-round associated content. Each set of content contains the association relationship between a previous factor and a current factor. All related combinations of content across rounds are categorized and divided into three dimensions: related content of requirement sub-items, related content of system responses, and related content of scenarios. Each dimension contains the corresponding related combinations of content. Within each dimension, the preceding factors in the related combined content are overlaid with the current factors to form the cumulative impact result content under that dimension. This cumulative impact result content includes the sum of the state changes of that dimension from the initial state to the current state. The cumulative impact results from the three dimensions are integrated and processed, and the corresponding round source information for each cumulative impact result is supplemented. The round source information records the previous round from which the cumulative impact originated. The integrated cumulative impact results and round-based source information are combined into the forward evolution impact set, which includes the cumulative changes in the positive state from the initial interaction state to the current round.

6. The intelligent interaction state tracking and reasoning method for multi-turn dialogue as described in claim 3, characterized in that, The backtracking association logic based on the bidirectional association evolution model associates key guidance points in subsequent guidance strategies with state evolution units from previous rounds, enabling the guidance strategy to reinforce effective associations from previous interactions, including: Extract the effective association identifiers of the preceding round state evolution units in the bidirectional association evolution model. Each effective association identifier corresponds to a set of preceding interactive association contents with a positive driving effect. Analyze the core guiding content of key guiding points in subsequent guiding strategies, and record the related direction of the requirement sub-items and the system response adaptation direction of each key guiding point; The core guiding content of the key guiding point is matched with the effective association identifiers of the previous round state evolution unit content, and the effective association content of the previous round with consistent guidance is selected. Extract the feature data of the round state evolution unit content corresponding to the successfully matched preceding valid associated content, including: the name of the associated demand sub-item, the fragment of the system response content, and the dialogue scene tag; The implementation path content of the key guiding point content is integrated with the core feature content of the preceding effective related content to generate enhanced related adaptation content. This enhanced related adaptation content includes the enhanced path content of the guiding point to the preceding association. Based on strengthening the relevance and adaptation of content, the relevance weight allocation of the content in the previous round state evolution unit is adjusted to increase the influence ratio of the previous relevance content that is consistent with the content orientation of the key guidance point. Weak links in the preceding interactive content are extracted and combined with the core guiding content of the key guiding points to generate supplementary and adaptive content. The enhanced and adaptive content is then integrated with the supplementary and adaptive content into a preceding association enhancement plan. This preceding association enhancement plan includes specific association adjustment paths and implementation order. The content of the preceding association enhancement scheme is integrated into the implementation process of the dynamic guidance strategy, and the interaction round node content and execution conditions corresponding to each enhancement step are recorded. By implementing a dynamic guidance strategy, the effective related content of the preceding interactions is strengthened. At the same time, the related supplementary and adapted content is used to improve the logical content of the preceding interactions, forming a reverse-strengthened system of related content for the preceding interactions.

7. The intelligent interaction state tracking and reasoning method for multi-turn dialogue as described in claim 4, characterized in that, The driving relationship path between the content of the preceding round state evolution unit and the content of the following round state evolution unit is recorded, and the specific effect of each driving factor on the basic content of the following initial state is recorded, forming the driving relationship record content, including: Assign a unique identifier to each content in the related promotion path. This identifier includes the previous round number, the next round number, and the association type identifier. Extract the set of preceding driving factors from the content of the preceding round state evolution unit, and record the type of each driving factor, the associated demand sub-item identifier, and the list of specific constituent elements in the subsequent round state evolution unit to which it points. Analyze the structural distribution of the initial state components of the state evolution unit in subsequent rounds, and record the priority of each component and its associated interfaces. Match the associated interface content of each driving factor content with the content of the constituent elements of the subsequent initial state, and record the position and mode of action of the driving factor content; Simulate the process of the driving factor acting on the subsequent initial state of the basic content, and record the parameter changes, priority adjustments and direction corrections after the action. The unique identifier of the content driving the associated path will be associated and bound with the corresponding driving factor content, the content of the location of action, the content of the method of action, and the content of the result of action to form the associated record item content of a single path; All associated record items are grouped according to the subsequent round number, and each group corresponds to all the driving associated path contents of a subsequent round state evolution unit. Within each group, the related record items are arranged in chronological order according to the previous round number, presenting the temporal logic that drives the association. Extract the interactive scene feature content corresponding to the content of each associated record item, and perform associated storage processing on the scene feature content and the associated record item content to enrich the scene background information content of the records; All grouped related record items are integrated into a driving related record set, and a multi-dimensional retrieval index is established for this driving related record set. The multi-dimensional retrieval index is classified according to round number, driving factor type, and mode of action.

8. The intelligent interaction state tracking and reasoning method for multi-turn dialogue as described in claim 5, characterized in that, The process of assigning associated weights to the state change factors in each previous round, based on the ratio of the time distance and influence between the previous and current rounds, includes: Extract the time sequence content of multiple rounds of dialogue, record the difference between the content of the current round number and the content of each previous round number, and determine the time distance parameter content according to the size of the difference content. The smaller the difference content, the larger the proportional coefficient content of the time distance parameter content. Analyze the number of sub-factors contained in the state change factors of each previous round, and calculate the coverage of each sub-factor to the state change factors of the current round; Evaluate the depth of the impact of each previous round's status change factors on the core requirements of the current round, and record the number of key related points that influence the impact. The scope of coverage and the number of key related points are integrated into the influence parameter. The wider the scope of coverage and the more key related points, the larger the proportional coefficient of the influence parameter. The process for calculating association weights is constructed, and the association weights are determined based on both the time proximity level and the significance level of the impact. The temporal proximity level is determined based on the temporal distance parameter for each preceding round's state change factors. The closer the temporal distance, the higher the temporal proximity level. The significance level of influence is determined based on the influence degree parameter for each preceding round of state change factors. The wider the coverage and the more key related points, the higher the significance level of influence. Based on the time proximity level and the significance level of impact, a comprehensive level is determined for the content of the state change factors in each previous round; Based on the comprehensive level of all previous rounds' state change factors, normalization is performed to generate the associated weight value of each previous round's state change factor content. The associated weight values ​​of all previous round state change factors are normalized. The normalized associated weight values ​​are then associated and stored with the corresponding previous round state change factors and round numbers to form a weight allocation record set.

9. An intelligent interactive state tracking and reasoning system applied to multi-turn dialogue, characterized in that, The method includes a processor and a computer-readable storage medium storing machine-executable instructions that, when executed by the processor, implement the intelligent interactive state tracking and reasoning method for multi-turn dialogue as described in any one of claims 1-8.

Images