An optimization method and system of a dialogue interaction strategy based on artificial intelligence

Abstract

The application relates to the technical field of data processing, and discloses an optimization method and system for a dialogue interaction strategy based on artificial intelligence, which comprises the following steps: determining a user intention sequence based on multi-modal data, determining a candidate interaction strategy based on the user intention sequence and constraint conditions, determining a corresponding potential trigger event based on the user intention sequence, screening the candidate interaction strategy based on the potential trigger event to obtain a first interaction strategy, obtaining feedback data of a user executing the first interaction strategy, updating the first interaction strategy based on the feedback data to obtain a second interaction strategy. According to the technical scheme of the application, the timing evolution of the user intention can be accurately captured, the potential trigger event can be predicted, and the dynamic optimization of the interaction strategy can be realized based on the user feedback under the premise of guaranteeing the interaction safety and real-time performance, so that the accuracy, individualization and robustness of the dialogue interaction are improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically to an optimization method and system for dialogue interaction strategies based on artificial intelligence. Background Technology

[0002] With the rapid development of artificial intelligence and multimodal interaction technologies, dialogue systems have been widely applied in various data processing scenarios such as intelligent customer service, virtual assistants, and information retrieval. Users' demands for the accuracy, real-time performance, and personalization of interactions are increasing. Current optimization of dialogue interaction strategies mainly relies on single-modal data for intent recognition, which struggles to fully capture the semantic context, environmental features, and historical behavioral correlations in user interactions. This results in an intent recognition accuracy rate of less than 65%, especially in multi-turn dialogues or semantically ambiguous scenarios, where strategy response biases are prone to occur.

[0003] Existing technologies suffer from the following shortcomings: Firstly, they lack multimodal data fusion capabilities. Semantic interaction data, user profile information, and dynamic environmental data are often isolated, failing to form a unified feature representation and hindering the comprehensiveness of strategy generation. Secondly, strategy optimization lacks a dynamic closed-loop mechanism. Most systems employ static preset strategy libraries, passively responding only to known scenarios and lacking the ability to predict and adapt to potential triggering events, resulting in over 30% of interactions requiring manual intervention for correction. Thirdly, user feedback utilization is inefficient. Traditional methods often rely on single-round feedback for parameter adjustment, neglecting the cumulative correlation and scenario generalization of feedback data. Furthermore, balancing privacy protection and data utilization is difficult; directly using raw feedback data risks user privacy leaks, while excessive anonymization reduces the accuracy of strategy optimization. Additionally, some systems prioritize strategy complexity at the expense of computational resource constraints, leading to response delays and severely impacting user experience. These issues make existing dialogue interaction systems insufficient for the data processing domain. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of this application is to provide an optimization method and system for dialogue interaction strategy based on artificial intelligence, which can accurately capture the temporal evolution of user intent, predict potential triggering events, and dynamically optimize the interaction strategy based on user feedback, while ensuring the security and real-time performance of the interaction, thereby improving the accuracy, personalization and robustness of dialogue interaction.

[0005] Firstly, embodiments of this application provide an optimization method for dialogue interaction strategies based on artificial intelligence, which can be specifically manifested as follows:

[0006] The user intent sequence is determined based on multimodal data. The user intent sequence refers to the temporal evolution sequence of a user from the initial intent to the first moment during the dialogue interaction process. Multimodal data includes: semantic interaction data, user information, and environmental data.

[0007] Based on the user intent sequence and constraints, candidate interaction strategies are determined;

[0008] Identify potential triggering events based on the user intent sequence;

[0009] Based on the potential triggering events, the candidate interaction strategies are filtered to obtain the first interaction strategy. The first interaction strategy refers to the strategy with the best response adaptability to the potential triggering events selected from the candidate interaction strategies.

[0010] Obtain feedback data from users who execute the first interaction strategy, update the first interaction strategy based on the feedback data, and obtain the second interaction strategy.

[0011] In one possible implementation, candidate interaction strategies are determined based on the user intent sequence and constraints, which can be specifically manifested as follows:

[0012] Temporal features are extracted from the user intent sequence to obtain intent conversion frequency, duration of key intents and intent jump path features. Key intents are the intents with the highest probability distribution in the user intent sequence and whose duration exceeds the first threshold.

[0013] The input feature vector is obtained by fusing time-series features with multimodal data.

[0014] An initial strategy set is generated based on the input feature vector. The initial strategy set includes dialogue templates, information inquiry order, response rhythm parameters, and branch response logic.

[0015] The initial policy set is filtered based on constraints to obtain candidate interaction policies.

[0016] In another possible embodiment, determining the corresponding potential triggering event based on the user intent sequence can be specifically manifested as follows:

[0017] Construct an intent-event association graph. In the intent-event association graph, nodes represent the intent types in historical interactions and the associated event types corresponding to the intent types, while edges represent the co-occurrence probability and temporal association strength of intents and events.

[0018] Based on the user intent sequence, an intent transition anomaly point in the user intent sequence is identified by the temporal pattern mining algorithm. The intent transition anomaly point is the intent transition point whose transition probability is lower than the second threshold of the average transition probability of the user intent sequence.

[0019] In the intent-event association graph, potential triggering events are retrieved. Potential triggering events refer to the associated event types in the intent-event association graph that have an association strength greater than the third threshold with the intent transition anomaly point and the user intent at the first moment.

[0020] Determine a candidate set of potential triggering events based on the types of associated events;

[0021] The potential triggering events are obtained by filtering the candidate set of potential triggering events.

[0022] In another embodiment, the potential triggering event candidate set is filtered to obtain potential triggering events, which can be specifically manifested as follows:

[0023] Obtain the probability of occurrence of events in the potential triggering event candidate set;

[0024] Events whose probability of occurrence is greater than the fourth threshold are selected from the candidate set of potential triggering events and designated as potential triggering events.

[0025] In another embodiment, a first interaction strategy is obtained by filtering candidate interaction strategies based on potential triggering events, including:

[0026] For each strategy in the candidate interaction strategy set, the strategy execution process is simulated when a potential triggering event occurs, and the total triggering event adaptability score of each strategy is calculated.

[0027] The strategy with a total adaptability score greater than the fifth threshold is selected as the first interaction strategy.

[0028] In another embodiment, feedback data from the user executing the first interaction strategy is obtained, and the first interaction strategy is updated based on the feedback data to obtain the second interaction strategy, which can be specifically manifested as follows:

[0029] Obtain user feedback data;

[0030] Calculate the deviation between the feedback data and the expected effect of the first interaction strategy;

[0031] The strategy to be optimized in the first interaction strategy is determined based on the type of deviation value;

[0032] The strategy to be optimized is updated based on the preset parameter iteration rules to obtain the second interaction strategy.

[0033] In another embodiment, the strategy to be optimized is updated based on a preset parameter iteration rule to obtain a second interaction strategy, which can be specifically manifested as follows:

[0034] An optimized sample set is constructed based on feedback data, and quantitative adjustment coefficients are generated.

[0035] The parameters of the optimization strategy are iteratively adjusted based on the quantization adjustment coefficient;

[0036] Perform constraint verification on the adjusted strategy;

[0037] The adjusted strategy that passes the verification will be used as the second interaction strategy.

[0038] In another embodiment, the method provided in this application can be further implemented as follows: obtaining feedback data from the user executing the second interaction strategy, updating the second interaction strategy based on the feedback data, and obtaining a third interaction strategy.

[0039] Based on the technical means of this solution, by integrating multimodal data to accurately capture the temporal evolution of user intent in in-vehicle interaction scenarios, it breaks through the limitations of single-round semantic understanding and makes the strategy more in line with the user's real needs. By predicting potential triggering events and selecting the optimal response strategy, it improves the adaptability to sudden situations such as sudden changes in intent and command execution failures during driving. Based on user feedback, it dynamically updates the strategy to achieve personalized iterative optimization of the interaction logic, taking into account both driving safety and real-time response, and effectively improving the accuracy, robustness and user experience of in-vehicle dialogue interaction.

[0040] Secondly, embodiments of this application provide an optimization system for dialogue interaction strategies based on artificial intelligence, which can be specifically manifested as follows:

[0041] Data processing module: used to determine the user intent sequence based on multimodal data. The user intent sequence refers to the temporal evolution sequence of the user from the initial intent to the first moment during the dialogue interaction. Multimodal data includes: semantic interaction data, user information and environmental data.

[0042] Strategy determination module: used to determine candidate interaction strategies based on user intent sequences and constraints;

[0043] Intent parsing module: used to determine the corresponding potential triggering events based on the user's intent sequence;

[0044] Strategy filtering module: used to filter candidate interaction strategies based on potential triggering events to obtain the first interaction strategy. The first interaction strategy refers to the strategy with the best response adaptability to potential triggering events selected from the candidate interaction strategies.

[0045] Strategy update module: Used to obtain feedback data from users who execute the first interaction strategy, update the first interaction strategy based on the feedback data, and obtain the second interaction strategy.

[0046] This solution addresses the shortcomings of traditional interactive systems, such as incomplete intent recognition, delayed strategy response, and difficulty in dynamic optimization, through multi-module collaboration. The data processing module integrates multimodal data, breaking down information silos and improving the accuracy of intent sequence recognition; the intent parsing module predicts potential triggering events, avoiding passive responses; the strategy selection module accurately matches the optimal strategy, improving event response efficiency; and the strategy update module achieves closed-loop iteration based on feedback. Ultimately, this upgrades interaction strategies from static to dynamic and from single to multi-dimensional, significantly improving system response accuracy and user experience satisfaction.

[0047] The solution provided in the second aspect above is used to implement the method provided in the first aspect above, and its specific implementation will not be described in detail here. The technical effects corresponding to any implementation method of the solution provided in the second aspect above can be found in the technical effects corresponding to any implementation method of the first aspect above, and will not be described in detail here.

[0048] It should be noted that any of the possible implementations of any of the above aspects can be combined, provided that the solutions do not contradict each other. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of this application or the background art, the accompanying drawings used in the embodiments of this application will be described below.

[0050] Figure 1 This is a flowchart illustrating an optimization method for a dialogue interaction strategy based on artificial intelligence, as disclosed in an embodiment of this application.

[0051] Figure 2 This is a flowchart illustrating another optimization method for a dialogue interaction strategy based on artificial intelligence disclosed in an embodiment of this application.

[0052] Figure 3 This is a schematic diagram of the structure of an AI-based dialogue interaction strategy optimization system disclosed in an embodiment of this application. Detailed Implementation

[0053] The terms "first," "second," etc., are used for descriptive purposes only and have no sequential or technical meaning, nor should they be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. In the description of the embodiments of this application, unless otherwise expressly specified and limited, the term "connection," etc., should be interpreted broadly. For example, "connection" can be a detachable connection or a non-detachable connection; it can be a direct connection or an indirect connection through an intermediate medium. "Fixed connection" refers to a connection where the relative positional relationship remains unchanged after the connection.

[0054] In the embodiments of this application, "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0055] The embodiments of this application are described below with reference to the accompanying drawings.

[0056] Please see Figure 1 , Figure 1This is a flowchart illustrating an AI-based dialogue interaction strategy optimization method disclosed in an embodiment of this application. The AI-based dialogue interaction strategy optimization method provided in this application can be applied to intelligent customer service scenarios, such as e-commerce customer service handling order inquiries and after-sales complaints. It needs to understand the user's demand sequence through multimodal data, predict potential events such as "missing order information" and "refund process questions," and dynamically optimize response strategies to improve efficiency. It can also be applied to intelligent virtual assistant scenarios, such as mobile phone voice assistants and smart home control systems. It needs to integrate data such as voice commands, user behavior habits, and device status to identify intent evolution, filter and adapt strategies for trigger events such as "ambiguous commands" and "device offline," and iteratively optimize interaction logic based on user feedback. Furthermore, it can be applied to online education and tutoring scenarios, such as AI teaching assistants. It determines the student's learning intent sequence through text-based Q&A and video interaction data, predicts events such as "knowledge point misunderstanding deviations" and "delayed practice feedback," and dynamically adjusts tutoring strategies to adapt to personalized learning needs.

[0057] The flowchart of the AI-based dialogue interaction strategy optimization method provided in this embodiment includes the following steps:

[0058] S101: Determine the user intent sequence based on multimodal data.

[0059] Specifically, the multimodal data are first preprocessed to eliminate noise interference and standardize the data format. For semantic interaction data, if it is text data, it is converted into a low-dimensional semantic vector through word segmentation, error correction, and encoding; if it is speech data, acoustic features such as Mel-frequency cepstral coefficients are extracted. For user information, structured data such as historical interaction records and user profile tags are feature-mapped to form user feature vectors. For environmental data, information such as time, scene type, and device status is standardized to extract scene context features.

[0060] For example, a mid-term fusion strategy is adopted, which uses an attention mechanism to achieve cross-modal feature interaction, dynamically adjusts the weight allocation of each modality feature, enhances semantic complementarity, and forms a unified input feature representation. This feature representation is input into a pre-trained sequence labeling model to identify the user's immediate intent time step by time, and then integrates the intents at each time step according to the time sequence of the dialogue interaction to form a temporal evolution sequence from the initial intent to the first time step, where each intent in the sequence is accompanied by a corresponding confidence score.

[0061] Among them, the user intent sequence refers to the temporal evolution sequence of a user from the initial intent to the first moment during the dialogue interaction process.

[0062] Pre-trained sequence labeling models refer to models that are pre-trained based on large-scale in-vehicle dialogue interaction corpora and multimodal feature data, specifically designed for time-by-time intent recognition of temporal multimodal fusion features, and have the ability to capture temporal dependencies of intent and adapt to the interactive characteristics of in-vehicle scenarios.

[0063] The specific structure includes: an input layer that receives the feature sequence after multimodal fusion and performs dimensionality normalization on the features; an encoder layer that uses the BERT-base model as the basic encoder, introduces a multi-head attention mechanism to enhance the ability to capture long-distance contextual intent dependencies, and embeds a multimodal feature adaptation sublayer to optimize the semantic understanding of features such as speech and environment; a decoder layer that uses a conditional random field layer to model the transition probability between intent labels and improve the accuracy of sequence labeling; and an output layer that includes an intent classifier and a confidence calculation unit, outputting the intent label and corresponding confidence score at each time step. The input data refers to the unified input feature representation sequence formed after processing by the intermediate fusion strategy. The feature vector at each time step fuses semantic interaction features, user feature vectors, and scene context features, and the sequence length corresponds one-to-one with the rounds or time steps of the dialogue interaction. The output data refers to the temporal intent labeling results with the same length as the input feature sequence, including the user's immediate intent label at each time step and the confidence score corresponding to each intent label, used for subsequent integration to form a user intent sequence with confidence.

[0064] Optionally, the first moment can be the current moment.

[0065] Multimodal data includes semantic interaction data, user information, and environmental data.

[0066] In some embodiments, the weight allocation of multimodal data adopts an adaptive mechanism, which is dynamically adjusted based on the current dialogue scenario.

[0067] For example, in consultation scenarios, the weight of semantic interaction data can be increased, while in personalized recommendation scenarios, the weight of user information can be increased. For preprocessed multimodal data, if there are missing or outliers, interpolation methods can be used to supplement the missing data, and outlier samples that deviate from the data mean by three times the standard deviation can be removed to ensure feature quality.

[0068] Optionally, the sequence labeling model can introduce a multi-head attention mechanism to enhance the ability to capture long-distance contextual dependencies and improve the accuracy of intent recognition. It can also combine knowledge graphs to supplement semantic association information between intents, smooth the identified user intent sequences, eliminate unreasonable intent mutations, and make the sequences more consistent with the user's real interaction logic. The determination of key intents must meet the conditions that the confidence score is higher than the preset threshold and the duration exceeds the first threshold.

[0069] The first threshold refers to the duration threshold used to determine the effectiveness of key intentions. Its value is set based on the statistical analysis of historical interaction data and the characteristics of business scenarios, and can be expressed as the number of interaction rounds or the actual time.

[0070] Specifically, the determination of the first threshold must meet the following condition: at a 95% confidence level, the probability that an intent exceeding the threshold is verified as a core user need in historical interactions is ≥80%.

[0071] For example, in customer service scenarios, the first threshold can be set to ≥3 rounds of interaction, with each round of interaction consisting of a complete dialogue between the user and the system or lasting ≥120 seconds, to filter out short-term exploratory intentions; in quick query scenarios, such as weather queries or balance queries, the first threshold can be reduced to ≥1 round of interaction or lasting ≥30 seconds to suit the user's quick and clear needs.

[0072] In addition, the first threshold supports dynamic iteration. The system will periodically recalculate the optimal threshold based on new interaction data. By comparing the accuracy and false positive rate of key intent recognition under different thresholds, the threshold that gives the highest F1 score is retained. This ensures that the key intent judgment neither excessively misses core needs nor includes redundant short-term intents, providing accurate core intent basis for subsequent strategy generation.

[0073] S102: Determine candidate interaction strategies based on the user intent sequence and constraints.

[0074] Among them, constraints refer to the set of restrictive rules that must be met during the dialogue interaction process, including business rule constraints, such as prohibiting promises of returns in the financial field and avoiding absolute treatment recommendations in the medical field; system performance constraints, such as single-round response latency ≤1.5 seconds and concurrent interaction support ≥1000 paths; ethical and security constraints, such as filtering discriminatory expressions and protecting user privacy information; and user preference constraints, such as restrictions on the style of speech based on user profiles.

[0075] Candidate interaction strategies refer to a subset of interaction strategies that, after being filtered by constraints, possess basic intent adaptability and compliance, and can provide a basic strategy pool for subsequent screening of potential triggering events.

[0076] In some embodiments, temporal features are extracted from the user intent sequence to obtain intent conversion frequency, key intent duration and intent jump path features. The temporal features are fused with multimodal data to obtain an input feature vector. An initial strategy set is generated based on the input feature vector. The initial strategy set includes dialogue templates, information inquiry order, response rhythm parameters and branch response logic. The initial strategy set is filtered based on constraints to obtain candidate interaction strategies.

[0077] Specifically, the user intent sequence is first processed using temporal feature engineering: a sliding window method is used, with the window size set to 5-10 interaction rounds. The sequence is traversed, and the number of changes in intent type per unit time is counted to obtain the intent conversion frequency, such as "1 intent conversion occurs every 3 rounds of interaction"; through continuous intent type matching, the total number of rounds or duration of key intents in the sequence is calculated to obtain the duration of key intents, such as "'account query' intent lasts for 4 rounds of interaction"; the sequential relationship of intent conversion is recorded in the form of a directed graph, and topological features such as the length and degree of the conversion path are extracted to form intent jump path features, such as "the path sequence from initial intent to account query to transfer to end".

[0078] Among them, the key intent is the intent with the highest probability distribution in the user intent sequence and whose duration exceeds the first threshold.

[0079] The first threshold refers to the standard for judging the validity of intent based on the business scenario. It can be determined by statistical analysis of historical interaction data. For example, in the customer service scenario, it can be set to a continuous interaction of ≥3 rounds or a duration of ≥120 seconds. The threshold must meet the user intent stability requirement under a 95% confidence level.

[0080] Temporal feature extraction of user intent sequences refers to transforming discrete intent sequences into quantifiable feature indicators, capturing the dynamic patterns of intent evolution, and providing a temporal dimension for policy generation.

[0081] Intent switching frequency refers to the number of times a user's intent type changes per unit of time. The calculation formula is:

[0082]

[0083] The frequency of intent switching reflects the stability of user needs.

[0084] The duration of key intent refers to the time span or number of interaction rounds during which key intent appears consecutively in the user's intent sequence, and is an important indicator for judging the user's core needs.

[0085] Intent jump path features refer to the trajectory features of a user's intent from the initial state to the current state, including path length, jump node type, number of repeated jumps, etc., which can be used to identify user interaction patterns.

[0086] The parameter fusion of temporal features and multimodal data to obtain the input feature vector involves: first, aligning the temporal features (dimension D1) with the preprocessed multimodal data features (dimensions D2 for semantic interaction features, D3 for user information features, and D4 for environmental data features); then employing an attention-weighted fusion algorithm, using the formula:

[0087] The fusion feature = α × temporal feature + β × semantic feature + γ × user feature + δ × environmental feature, where α + β + γ + δ = 1. The weights are optimized by gradient descent to generate an input feature vector with dimensions max(D1, D2, D3, D4), thereby achieving complementary enhancement of cross-modal information.

[0088] An initial policy set is generated based on the input feature vector. Specifically, the input feature vector is fed into a pre-trained policy generation model, which captures the correlations between features through a multi-head attention mechanism and outputs the initial policy set. During training, the policy generation logic is optimized using intent achievement rate and user dwell time as reward functions.

[0089] The initial strategy set includes dialogue templates, information inquiry order, response rhythm parameters, and branching response logic.

[0090] For example, the script templates are categorized by intent type, such as "account inquiry" corresponding to "Do you want to inquire about the balance or transaction details?"; the order of information inquiries follows the principle of decreasing importance, such as asking "account type" first and then "inquiry time period"; the response rhythm parameter is set to a single round waiting time of ≤5 seconds and a continuous inquiry interval of ≥2 seconds; the branch response logic covers the handling rules for abnormal scenarios such as "user refuses to provide information" and "intent change".

[0091] For example, the initial set of strategies is filtered based on constraints to obtain candidate interaction strategies.

[0092] Specifically, a three-level verification is performed:

[0093] Business rule verification: Remove strategies containing illegal statements, such as templates in financial scenarios that "guarantee an annualized return of 10%";

[0094] Performance constraint verification: Strategies with response latency ≤1.5 seconds are selected through stress testing, and complex branch logic with high time consumption is eliminated;

[0095] Ethical and safety verification: A pre-trained text classification model is used to detect discriminatory and privacy-sensitive content in the script template, filtering out unqualified strategies. Strategies that pass verification are formed into candidate interaction strategies, while strategies that fail verification are recorded as violations and fed back to the strategy generation model for iterative optimization.

[0096] S103: Determine the corresponding potential triggering events based on the user intent sequence.

[0097] In some embodiments, an intent-event association graph is constructed. Based on the user intent sequence, anomalies in intent transitions in the user intent sequence are identified through a time-series pattern mining algorithm. Potential triggering events are queried in the intent-event association graph. A candidate set of potential triggering events is determined based on the associated event type. The candidate set of potential triggering events is then filtered to obtain potential triggering events.

[0098] Specifically, constructing an intent-event association graph refers to: based on historical dialogue interaction data, using entity recognition and relation extraction algorithms, extracting intent types and event types related to intents from historical interactions, and using these as nodes in the graph; calculating the co-occurrence probability of intents and events, using the formula: co-occurrence probability = number of times intents and events co-occur / total number of times intents occur; and calculating the temporal association strength, using the formula: temporal strength = number of times the event occurs within t seconds after the intent occurs / total number of times intents occur, where t is set to 30-60 seconds according to the business scenario, and using this as the weight of the edges connecting the nodes, forming a weighted directed graph structure for the intent-event association graph.

[0099] Optionally, the graph supports dynamic updates, recalculating edge weights after every 10,000 new interaction data points to improve association accuracy.

[0100] In the intent-event association graph, nodes represent the intent types and associated event types in historical interactions, while edges represent the co-occurrence probability and temporal association strength of intents and events.

[0101] An intent conversion anomaly is an intent conversion point where the conversion probability is lower than the second threshold of the average conversion probability of the user intent sequence.

[0102] The second threshold refers to a critical value set based on the distribution characteristics of the conversion probabilities of all intents in the user intent sequence. It is used to define the boundary of "abnormal" conversions and is usually set to 50%-70% of the average conversion probability of the sequence.

[0103] For example, if the average conversion probability of a sequence is 20%, the second threshold can be set to 10%. The threshold must meet the following requirements: in historical data, conversions below the threshold are verified as "non-natural intent jumps", such as the proportion of jumps caused by user misoperation or ambiguous intent expression ≥ 90%, and it supports dynamic adjustment based on the scenario. For example, in complex business scenarios, the threshold can be reduced to 40% to capture more potential anomalies.

[0104] Potential triggering events refer to the associated event types in the intent-event association graph that have an affinity for intent transition anomalies and user intent at the first moment that is greater than the third threshold.

[0105] Related event types refer to events that have a significant temporal or causal relationship with a specific intent type in historical interactions. These events may affect the progress of the interaction process, such as the "order cancelled" and "logistics information not updated" events associated with the "query order" intent, and the "payment failed" and "balance exceeded" events associated with the "recharge" intent. They usually include the event name, occurrence conditions, and degree of impact on the interaction, such as "interruption type" and "requires explanation type".

[0106] The third threshold refers to the critical value of association strength used to filter events that are highly related to the current intent sequence. It is set based on the distribution of historical association strength in the intent-event association graph and is usually 1.2-1.5 times the average weight of all edges in the graph.

[0107] For example, if the average association strength of the graph is 0.3, the third threshold can be set to 0.36. The threshold must ensure that the co-occurrence accuracy of the selected events with the corresponding intent in historical interactions is ≥85%, and can cover more than 90% of key influencing events, avoiding the omission of important potential trigger points.

[0108] Specifically, in the intent-event association graph, the process of retrieving potential triggering events includes:

[0109] The intent node corresponding to the point of abnormal intent conversion is denoted as A, and the user intent node at the first moment is denoted as B.

[0110] In the graph, retrieve the event nodes directly connected to nodes A and B, and obtain the association strength between these events and A and B, i.e., the edge weights;

[0111] Calculate the overall correlation strength between event A and B:

[0112] Overall strength = 0.4 × correlation strength between event A + 0.6 × correlation strength between event B;

[0113] Events with a comprehensive correlation strength greater than the third threshold are selected as potential triggering events that are strongly correlated with both the intent conversion anomaly and the current intent.

[0114] If the screening results are empty, the third threshold is lowered by 10% and the search is repeated, up to a maximum of 2 times, to ensure that at least one potential triggering event is returned. If it is still empty, it is marked as "no significant potential triggering events".

[0115] For example, if the intention to change anomaly point A is "querying bill to modify delivery address" (non-natural jump), and the initial intention B is "modify delivery address", then the event associated with A and B is retrieved in the graph. If the association strength of "incorrect delivery address format" with A is 0.4 and the association strength with B is 0.6, the overall strength = 0.4 × 0.4 + 0.6 × 0.6 = 0.52. If the third threshold is 0.36, then this event is included in the potential triggering event.

[0116] For example, the occurrence probability of events in the potential triggering event candidate set is obtained, and events with an occurrence probability greater than a fourth threshold are selected as potential triggering events.

[0117] The probability of occurrence refers to the likelihood of each event in the potential trigger event candidate set actually occurring under the current user intent sequence and multimodal data scenario. The value ranges from 0 to 1, where 0 indicates that it is impossible and 1 indicates that it will definitely occur.

[0118] Its calculation employs Bayes' theorem, fusing historical statistics with real-time scene features. The specific formula is as follows:

[0119] The probability of occurrence P = prior probability P0 × scenario fit likelihood L / normalization constant.

[0120] The prior probability P0 is taken from the historical co-occurrence probability of the event and the corresponding intent in the intent-event association graph. For example, the historical co-occurrence probability of the "payment failure" event and the "recharge" intent is 0.35.

[0121] The scenario adaptation likelihood L is calculated based on the current multimodal data features. For example, when the user's current device is a newly logged-in device, the L value for "payment failure" is 0.8, reflecting the degree of adaptation between the event and the current interaction scenario.

[0122] The normalization constant is used to ensure that the sum of the probabilities of all candidate events is 1, which facilitates horizontal comparison.

[0123] This probability is a dynamically updated real-time indicator that is recalculated and updated as user interaction progresses, such as when a user enters "cannot pay" or environmental data changes, such as when the network becomes unstable, to ensure that the judgment on the probability of the event is consistent with the current situation.

[0124] The fourth threshold is a critical value used to filter high-probability potential triggering events. Its core function is to balance the accuracy and coverage of event identification, avoiding the omission of key events or the inclusion of low-value redundant events. Its setting rules are as follows:

[0125] The base threshold is determined based on the probability distribution of "potentially triggered events that actually occurred" in historical interaction data. It is usually taken as the 90th quantile of the distribution. For example, if the probability of occurrence of historical real events is ≥0.4, then the base value of the fourth threshold is set to 0.4.

[0126] Adjust the tolerance for false positives / false negatives based on business scenarios: For high-risk scenarios, the threshold can be increased to the 95th percentile to reduce false positives; for scenarios where service experience is the priority, the threshold can be decreased to the 85th percentile to reduce false negatives.

[0127] Supports dynamic iterative updates: Every time the system accumulates 5,000 new interaction data, it recalculates the quantiles of the historical probability distribution and adjusts the fourth threshold accordingly to ensure it adapts to changes in business scenarios.

[0128] The specific screening process is as follows: calculate the probability of occurrence of each event in the candidate set one by one, and retain events with a probability value greater than the fourth threshold; if there are more than 3 results after screening, sort them in descending order of occurrence probability, and take the top 3 as the final potential triggering events (to avoid computational overload during strategy screening); if there are no results after screening, lower the fourth threshold by 10% and re-screen, lowering it a maximum of 2 times, and if there are still no results, mark it as "no high probability potential triggering events".

[0129] For example, the candidate set of potential triggering events includes "payment failure", "insufficient balance" and "incorrect password", with probabilities of occurrence of 0.52, 0.38 and 0.25, respectively. If the fourth threshold is set to 0.4, then only the "payment failure" event will be selected as the final potential triggering event.

[0130] S104: Filter from candidate interaction strategies based on potential triggering events to obtain the first interaction strategy.

[0131] The first interaction strategy refers to the strategy selected from the candidate interaction strategies that has the best adaptability to the potential triggering event.

[0132] Optimal response adaptability refers to the characteristics of a candidate interaction strategy that best performs in terms of process effectiveness, resource economy, and user experience friendliness in response to potential triggering events. It can accurately match the response needs of triggering events and solve problems efficiently with the fewest steps; it can also reasonably control the consumption of system resources during execution and not exceed the system's operating constraints; and it can also fit the user's interaction habits and preferences to obtain high user acceptance and satisfaction.

[0133] In some embodiments, for each strategy in the candidate interaction strategy set, the strategy execution process when a potential triggering event occurs is simulated, the total triggering event adaptability score of each strategy is calculated, and the strategy with a total adaptability score greater than a fifth threshold is selected as the first interaction strategy.

[0134] Specifically, a three-dimensional simulation evaluation framework is used to perform the screening:

[0135] Strategy Execution Simulation: An interactive scenario simulation environment is built, inputting potential triggering events and corresponding multimodal context data, such as the user's current intent, device status, and historical interaction records. Each candidate strategy is then virtually executed. The simulation process progresses step-by-step, recording the strategy's response steps and resource usage, such as the number of API calls, computation time, and simulated user feedback. A feedback prediction model is then generated based on historical interaction data.

[0136] Calculation of multidimensional adaptability index:

[0137] Process adaptability weight 40%: Evaluates the degree of matching between strategy steps and the requirements for responding to triggered events, using the following formula:

[0138] Process adaptability = (Number of effective response steps / Total number of response steps) × 100%. For example, if a "payment failure" event requires 3 response steps, and a certain strategy completes 2 effective steps, the score is 66.7%.

[0139] Resource consumption, weighted at 30%, measures the system cost of strategy execution. The formula is: Resource consumption = 1 - (Actual response latency / Maximum allowable latency × 0.5 + Actual CPU utilization / Maximum allowable utilization × 0.5). For example, if the response latency is 1 second and the CPU utilization is 40%, with maximum thresholds of 1.5 seconds and 80% respectively, then the score is 1 - (1 / 1.5 × 0.5 + 40% / 80% × 0.5) = 66.7%.

[0140] User acceptance: Based on the matching degree of strategic dialogue templates, interaction rhythm and historical user preferences, the satisfaction prediction value is output through a pre-trained sentiment classification model.

[0141] Total Adaptability Score Aggregation: The scores from the three dimensions are weighted and summed, using the following formula:

[0142] Total adaptability score = process adaptability × 0.4 + resource consumption × 0.3 + user acceptance × 0.3, with the result rounded to two decimal places.

[0143] The weights for each dimension in the formula are as follows: 40% for process adaptability, representing the important proportion of process adaptability in the overall adaptability assessment; 30% for resource consumption, representing the important proportion of resource consumption in the overall adaptability assessment; and 30% for user acceptance, representing the important proportion of user acceptance in the overall adaptability assessment.

[0144] The weights of each dimension are determined based on the priority of core needs for optimizing dialogue interaction strategies in the field of data processing technology, specifically as follows: Process Adaptability Weight 40%: In dialogue interaction, potential triggering events directly affect the progress of the interaction process. Whether the strategy's response process matches the event requirements determines whether the event can be resolved accurately and efficiently. This is the core of ensuring the effectiveness of the interaction and the fundamental value of the strategy, hence it is given the highest weight. Resource Consumption Weight 30%: Data processing systems have objective constraints on resources such as computing power and response latency. If strategy execution consumes too many resources, it will lead to a decrease in system processing efficiency and insufficient concurrency, directly affecting the real-time performance of the interaction and the stability of the system. This is a key prerequisite for the strategy to be implemented, and its priority is second only to process adaptability. User Acceptance Weight 30%: The long-term application of interaction strategies depends on user experience. Whether the dialogue style and interaction rhythm match user preferences affects the user's trust in the interaction system and their willingness to use it. However, the experience needs to be improved on the basis of effectively resolving events and ensuring stable system operation, therefore its weight is slightly lower than process adaptability.

[0145] The overall suitability score is a quantitative indicator that comprehensively reflects the matching degree of a candidate strategy to the response process of potential triggering events, the rationality of system resource consumption, and user feedback expectations. The score ranges from 0 to 100, with higher scores indicating stronger strategy suitability. This score needs to be interpreted in conjunction with the specific triggering event scenario. For example, in a "payment failure" event, a strategy with a total score of 82.5 means that it performs well in terms of process completeness, response efficiency, and user experience.

[0146] The fifth threshold is a critical value used to filter highly adaptable strategies. Its core function is to ensure that the response effect of the first interaction strategy meets business requirements. Its setting rules are as follows:

[0147] Determining the base threshold: Based on the total score distribution of "strategies for successfully responding to triggering events" in historical interactions, the 90th percentile of this distribution is taken as the base threshold. For example, if the total score of historical successful strategies is ≥75 points, then the base value of the fifth threshold is set to 75 points.

[0148] Scenario risk coefficient adjustment: A risk coefficient K is introduced, with K=1.1 for high-risk scenarios, K=1.0 for medium-risk scenarios, and K=0.9 for low-risk scenarios. The final threshold = base threshold × K. For example, in the financial transaction scenario (high risk), the threshold is adjusted to 75 × 1.1 = 82.5 points;

[0149] Dynamic iteration mechanism: Every 3,000 new interaction data entries accumulated by the system, the 90th percentile of the historical total score distribution is recalculated, and the basic threshold is updated synchronously; the risk coefficient K is adjusted monthly according to changes in the business risk level.

[0150] The specific screening process is as follows:

[0151] Calculate the total fitness score for all candidate strategies;

[0152] The strategy of retaining the total score > the fifth threshold is adopted. If the number exceeds 5, the top 5 are selected in descending order of total score (to balance strategy diversity and computational efficiency).

[0153] If the filtering results are empty, the fifth threshold will be lowered by 5% and the filtering will be repeated, up to a maximum of 2 times. If there are still no results, the 3 strategies with the highest total scores will be selected as the first alternative interaction strategy.

[0154] For example, if the potential triggering event is "payment failure", the candidate strategies are A, B and C, with total suitability scores of 88.2, 79.5 and 72.1 respectively. If the fifth threshold is set to 82.5 points, then only strategy A will be selected as the first interaction strategy.

[0155] S105: Obtain feedback data from the user executing the first interaction strategy, update the first interaction strategy based on the feedback data, and obtain the second interaction strategy.

[0156] The quantitative adjustment coefficient refers to a quantitative indicator generated based on the deviation between feedback data and expected results, used to guide the adjustment range of the parameters of the strategy to be optimized. Its core function is to transform abstract deviation differences into executable parameter adjustment ratios, ensuring that the optimization process is controllable and the effect is quantifiable.

[0157] Specifically, the characteristics and generation rules of the quantization adjustment coefficient are as follows:

[0158] Correlation: It is positively correlated with the absolute value of the deviation, that is, the larger the deviation, the larger the coefficient value, and the more obvious the parameter adjustment range. For example, the coefficient when the deviation of the intention achievement rate is -20% is greater than the coefficient when the deviation is -5%.

[0159] Calculation formula: After standardization to ensure the coefficients are within a reasonable range, the formula is as follows:

[0160]

[0161] Where D is the actual deviation value, taken as the absolute value. The maximum permissible deviation is preset, such as a maximum permissible deviation of 30% for the intended achievement rate. α is the weight of the deviation's impact, ranging from 0.1 to 0.3, set according to the business's requirements for optimization sensitivity. β is the base adjustment coefficient, with a value of 0.05, ensuring that even small deviations trigger fine-tuning. The coefficient K is strictly limited to a range of 0.05-0.2 to avoid excessively large coefficients leading to drastic parameter adjustments that could compromise the stability of the core strategy logic.

[0162] Differentiated adaptation: The coefficients are applied differently depending on the type of the strategy module to be optimized.

[0163] If the module to be optimized is the intent-policy matching parameter, K is used to adjust the association weight between intent features and policy variables. For example, if the original weight is 0.6 and K=0.1, the adjusted weight is 0.6×(1+0.1)=0.66.

[0164] If it is "speech style generation logic", K is used to correct the semantic affinity parameter. For example, if the original parameter is 0.7 and K=0.15, the adjusted parameter = 0.7×(1+0.15)=0.805).

[0165] If it is the "response rhythm parameter", K is used to compress or extend the duration threshold. For example, if the original threshold is 5 seconds and the response is too slow, the adjusted threshold = 5 × (1 - K) = 5 × 0.85 = 4.25 seconds.

[0166] Dynamic decay mechanism: When the adjusted strategy fails the constraint verification, K will decay by 20%. For example, if the original K=0.15, it will decay to 0.12 and be reused for parameter adjustment. It will decay a maximum of 3 times to ensure that optimization is achieved as much as possible while meeting the constraints.

[0167] For example, if the actual achievement rate of the first interaction strategy is 70%, and the expected value is 85% (deviation D=15%). =30%, α=0.2, β=0.05, then the quantification adjustment coefficient K=0.2×(15% / 30%)+0.05=0.15, which means that the intent-policy matching parameters to be optimized need to be adjusted by 15% to improve the intent achievement effect.

[0168] Please see Figure 2 , Figure 2 This is a flowchart illustrating another method for optimizing a dialogue interaction strategy based on artificial intelligence, as disclosed in an embodiment of this application. The flowchart of the method for optimizing a dialogue interaction strategy based on artificial intelligence provided in this embodiment includes the following steps:

[0169] S201: Determine the user intent sequence based on multimodal data.

[0170] For a description of this step, please refer to S101. It will not be elaborated further here.

[0171] S202: Determine candidate interaction strategies based on the user intent sequence and constraints.

[0172] For a description of this step, please refer to S102. It will not be elaborated further here.

[0173] S203: Determine the corresponding potential triggering events based on the user intent sequence.

[0174] For a description of this step, please refer to S103. It will not be elaborated further here.

[0175] S204: Based on potential triggering events, filter from candidate interaction strategies to obtain the first interaction strategy.

[0176] For a description of this step, please refer to S104. It will not be elaborated further here.

[0177] S205: Obtain feedback data from the user executing the first interaction strategy, update the first interaction strategy based on the feedback data, and obtain the second interaction strategy.

[0178] For a description of this step, please refer to S105. It will not be elaborated further here.

[0179] S206: Obtain feedback data from the user executing the second interaction strategy, update the second interaction strategy based on the feedback data, and obtain the third interaction strategy.

[0180] Specifically, explicit feedback during the user's execution of the second interaction strategy is obtained, such as satisfaction rating and confirmation of intent achievement, and implicit feedback, such as interaction completion time and number of repeated operations. At the same time, historical feedback data is integrated to construct a comprehensive feedback dataset that includes immediate feedback and historical feedback, and abnormal feedback samples that deviate from the mean by 3 times the standard deviation are removed.

[0181] Cumulative deviation calculation: Based on the expected effect of the second interaction strategy, and combined with the weighting of current feedback and historical feedback (current feedback weight 0.7, historical feedback weight 0.3), the cumulative deviation is calculated as follows: Cumulative deviation = 0.7 × current deviation + 0.3 × historical deviation. The current deviation is calculated by subtracting the expected effect from the actual effect, and the historical deviation is the deviation value from stage S205.

[0182] Optimization strategy identification: Based on the dominant type of cumulative deviation value, such as the deviation ratio of intent achievement rate ≥ 50%, identify the corresponding optimization module in the second interaction strategy, such as intent-strategy matching parameters and dialogue templates. At the same time, combine the correlation of multi-round feedback to identify potential optimization points across modules, such as the collaborative optimization of process adaptability and user acceptance.

[0183] Dynamic parameter iterative adjustment: An adaptive step-size iteration rule is adopted, adjusting the update step size according to the cumulative deviation value: when the deviation value > the preset threshold, the step size is 0.15; when the deviation value ≤ the preset threshold, the step size is 0.08, avoiding over-adjustment or under-optimization. After quantifying and updating the parameters of each module to be optimized, the parameters are checked against the constraints. If the check fails, the step size is readjusted by a 20% decay.

[0184] Strategy verification and output: The adjusted strategy is verified in multiple scenarios in the simulation environment to ensure its adaptability under different intent sequences and triggering events. Once the verification is passed, it becomes the third interaction strategy.

[0185] The third interaction strategy refers to the interaction strategy formed after optimization through the initial strategy, the first feedback iteration, and the second feedback iteration. It integrates multiple rounds of user feedback data and accurately adapts to the user's real needs. It not only inherits the core effective logic of the first two rounds of strategies, but also corrects the optimization deviations that may be caused by a single round of feedback through in-depth mining of accumulated feedback. It achieves a better balance in terms of process adaptability, reasonable resource consumption, and user acceptance, and has a stronger ability to generalize scenarios and a higher degree of alignment with user needs.

[0186] For example, the third interaction strategy is executed, and the third interaction strategy is updated based on the corresponding user feedback data. The above update process continues to iterate until the user ends the current interactive dialogue.

[0187] Optionally, Table 1 provides an example of the specific update process:

[0188] Table 1

[0189] Iteration rounds Real-time feedback data that triggers updates Deviation analysis Specific adjustment operations Updated strategy version Real-time performance metrics after update Round 1 Explicit: User inputs "Too troublesome, too many steps"; Implicit: Dialogue interrupted twice, 18-second hesitation between "Check Balance" and "Enter Card Number". Intent achievement rate deviation -20% (incomplete query), satisfaction deviation -1.5 stars (currently 2.5 stars), response adaptability deviation +8 seconds, total deviation 0.18 1. The "Confirm Account Type" step has been removed; 2. The script has been optimized to "Checking your balance only requires two steps: ① Enter the last 4 digits of your card number ② Click Confirm." V1.1 Intent achievement rate 60% (+20%), satisfaction rate 3.2 stars (+0.7 stars), and hesitation time for redirection reduced to 7 seconds. Round 2 Explicit: User feedback "The last four digits of the number were entered incorrectly, but no prompt was given"; Implicit: After an incorrect entry, the user is prompted to enter the wrong number twice, with a pause time of 22 seconds. Intent achievement rate deviation -10% (still incomplete), satisfaction deviation -0.8 stars (currently 3.2 stars), response adaptability deviation +5 seconds, total deviation 0.08 1. Added real-time validation: A pop-up window will appear when an error is entered, indicating "The last four digits should be 4 digits"; 2. Retains input history and supports modifying single digits. V1.2 Intent achievement rate 85% (+25%), satisfaction rate 4.0 stars (+0.8 stars), number of typos reduced to 0. Round 3 Explicit: The user says, "I have a bad memory, can you show the last digit of my last input?"; Implicit: The user frequently switches pages to check their input history. Intent achievement rate deviation -5% (completed but poor experience), satisfaction deviation -0.5 stars (currently 4.0 stars), response adaptability deviation +3 seconds, total deviation 0.03 1. For users registered for ≥1 year, display "Last digits: ****1234 (Click to reuse)"; 2. Add voice prompt "Do you need help filling in the last digits?" V1.3 Intent achievement rate 95% (+10%), satisfaction rate 4.6 stars (+0.6 stars), input time reduced to 5 seconds. Round 4 Explicit: User feedback "The balance number is too small to see clearly"; Implicit: One screenshot zoom-in operation and a 15-second viewing time. Intent achievement rate deviation 0% (completed), satisfaction deviation -0.3 stars (currently 4.6 stars), response adaptability deviation +2 seconds, total deviation 0.01 1. The balance numbers have been enlarged to 16-point font; 2. A "High Definition Mode" button has been added (increasing contrast by 30% when clicked). V1.4 Intent achievement rate 100% (+5%), satisfaction rate 4.9 stars (+0.3 stars), viewing time reduced to 6 seconds.

[0190] Please see Figure 3 , Figure 3 This is a schematic diagram of the structure of an AI-based dialogue interaction strategy optimization system disclosed in an embodiment of this application.

[0191] like Figure 3 The system shown is an optimization system for dialogue interaction strategies based on artificial intelligence, comprising:

[0192] Data processing module: used to determine the user intent sequence based on multimodal data. The user intent sequence refers to the temporal evolution sequence of the user from the initial intent to the first moment during the dialogue interaction. Multimodal data includes: semantic interaction data, user information and environmental data.

[0193] Strategy determination module: used to determine candidate interaction strategies based on user intent sequences and constraints;

[0194] Intent parsing module: used to determine the corresponding potential triggering events based on the user's intent sequence;

[0195] Strategy filtering module: used to filter candidate interaction strategies based on potential triggering events to obtain the first interaction strategy. The first interaction strategy refers to the strategy with the best response adaptability to potential triggering events selected from the candidate interaction strategies.

[0196] Strategy update module: Used to obtain feedback data from users who execute the first interaction strategy, update the first interaction strategy based on the feedback data, and obtain the second interaction strategy.

[0197] Through the above description of the implementation methods, those skilled in the art will clearly understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the module can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, modules, and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0198] The method steps in this embodiment can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary embodiment couples a storage medium to a processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and storage medium can reside in an ASIC. Additionally, the ASIC can reside in a network device. Alternatively, the processor and storage medium can exist as discrete components in the network device. In the above embodiments, implementation can be entirely or partially achieved through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented entirely or partially as a computer program product. A computer program product includes one or more computer programs or instructions. When a computer program or instruction is loaded and executed on a computer, all or part of the processes or functions of the embodiments of this application are performed. The computer may be a general-purpose computer, a special-purpose computer, a computer network, a network device, a user equipment, or other programmable module. The computer program or instructions may be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another. For example, a computer program or instructions may be transferred from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The computer-readable storage medium may be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium, such as a floppy disk, hard disk, or magnetic tape; or an optical medium, such as a digital video disc (DVD); or a semiconductor medium, such as a solid-state drive (SSD). The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0199] Since the data processing apparatus in the embodiments of the present invention can be applied to the above-described method, the technical effects it can achieve can also be referred to the above-described method embodiments, and the embodiments of the present invention will not be repeated here. The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0200] It should be understood that the application of this application is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims. Those skilled in the art can understand that implementing all or part of the processes of the above embodiments and making equivalent changes according to the claims of this application still fall within the scope of this application.

Claims

1. An optimization method for dialogue interaction strategies based on artificial intelligence, characterized in that, The method includes: The user intent sequence is determined based on multimodal data. The user intent sequence refers to the temporal evolution sequence of a user from the initial intent to the first moment during the dialogue interaction process. The multimodal data includes: semantic interaction data, user information, and environmental data. Based on the user intent sequence and constraints, candidate interaction strategies are determined; Construct an intent-event association graph, in which nodes represent intent types in historical interactions and associated event types corresponding to intent types, and edges represent the co-occurrence probability and temporal association strength of intents and events; Based on the user intent sequence, an intent conversion anomaly point in the user intent sequence is identified by a time-series pattern mining algorithm. The intent conversion anomaly point is an intent conversion point whose conversion probability is lower than the second threshold of the average conversion probability of the user intent sequence. In the intent-event association graph, potential triggering events are obtained by querying. The potential triggering events refer to the associated event types in the intent-event association graph that correspond to the intent conversion anomaly point and the user intent at the first moment with an association strength greater than a third threshold. Based on the associated event types, a potential set of triggering event candidates is determined; The potential triggering event candidate set is filtered to obtain potential triggering events; Based on the potential triggering event, a first interaction strategy is obtained by filtering from the candidate interaction strategies. The first interaction strategy refers to the strategy with the best response adaptability to the potential triggering event selected from the candidate interaction strategies. Obtain feedback data from the user executing the first interaction strategy, update the first interaction strategy based on the feedback data, and obtain the second interaction strategy.

2. The method according to claim 1, characterized in that, The step of determining candidate interaction strategies based on the user intent sequence and constraints includes: Temporal features are extracted from the user intent sequence to obtain intent conversion frequency, duration of key intent, and intent jump path features. The key intent is the intent with the highest probability distribution and duration exceeding a first threshold in the user intent sequence. The temporal features are fused with the multimodal data to obtain the input feature vector; An initial strategy set is generated based on the input feature vector. The initial strategy set includes a dialogue template, information inquiry order, response rhythm parameters, and branch response logic. The initial strategy set is filtered based on the constraints to obtain candidate interaction strategies.

3. The method according to claim 1, characterized in that, The process of filtering the candidate set of potential triggering events to obtain potential triggering events includes: Obtain the probability of occurrence of events in the potential triggering event candidate set; Events whose occurrence probability is greater than the fourth threshold are selected from the candidate set of potential triggering events and designated as potential triggering events.

4. The method according to claim 1, characterized in that, The step of filtering candidate interaction strategies based on the potential triggering events to obtain a first interaction strategy includes: For each strategy in the candidate interaction strategy set, the strategy execution process is simulated when the potential triggering event occurs, and the total triggering event adaptability score of each strategy is calculated. The strategy with the total adaptability score greater than the fifth threshold is selected as the first interaction strategy.

5. The method according to claim 1, characterized in that, The step of obtaining feedback data from the user executing the first interaction strategy, and updating the first interaction strategy based on the feedback data to obtain the second interaction strategy includes: Obtain user feedback data; Calculate the deviation between the feedback data and the expected effect of the first interaction strategy; The strategy to be optimized in the first interaction strategy is determined based on the type of the deviation value; The strategy to be optimized is updated based on the preset parameter iteration rules to obtain the second interaction strategy.

6. The method according to claim 5, characterized in that, The second interaction strategy is obtained by updating the strategy to be optimized based on a preset parameter iteration rule, including: An optimized sample set is constructed based on the feedback data, and a quantitative adjustment coefficient is generated. The parameters of the strategy to be optimized are iteratively adjusted based on the quantization adjustment coefficient. Perform constraint verification on the adjusted strategy; The adjusted strategy that passes the verification will be used as the second interaction strategy.

7. The method according to claim 1, characterized in that, The method further includes: Obtain feedback data from the user executing the second interaction strategy, update the second interaction strategy based on the feedback data, and obtain the third interaction strategy.

8. An optimization system for dialogue interaction strategies based on artificial intelligence, characterized in that, The system includes: Data processing module: used to determine the user intent sequence based on multimodal data. The user intent sequence refers to the temporal evolution sequence of the user from the initial intent to the first moment during the dialogue interaction process. The multimodal data includes: semantic interaction data, user information and environmental data. Strategy determination module: used to determine candidate interaction strategies based on the user intent sequence and constraints; The intent parsing module is used to construct an intent-event association graph. Nodes in the graph represent intent types from historical interactions and their corresponding associated event types. Edges represent the co-occurrence probability and temporal association strength between intents and events. Based on the user intent sequence, a temporal pattern mining algorithm is used to identify intent transition anomalies in the sequence. These anomalies are intent transition points whose transition probability is lower than a second threshold of the average transition probability of the user intent sequence. Potential triggering events are retrieved from the intent-event association graph. These are associated event types in the graph that correspond to intent types with an association strength greater than a third threshold with the intent transition anomaly and the user intent at the first moment. A candidate set of potential triggering events is determined based on these associated event types. The candidate set is then filtered to obtain the potential triggering events. Strategy filtering module: used to filter from the candidate interaction strategies based on the potential triggering event to obtain a first interaction strategy, wherein the first interaction strategy refers to the strategy with the best response adaptability to the potential triggering event selected from the candidate interaction strategies; Strategy update module: used to obtain feedback data from users who execute the first interaction strategy, update the first interaction strategy based on the feedback data, and obtain the second interaction strategy.

Images