A method and apparatus for updating user impression of an interactive robot

Abstract

The application relates to a user impression updating method and device for an interactive robot. The method comprises the following steps: obtaining historical interaction data and current interaction data, establishing a historical interaction baseline; judging whether the current interaction data meets a topic avoidance trigger condition; extracting a topic avoidance section and determining a set of discrimination parameters when the trigger condition is met; performing state discrimination based on the set of discrimination parameters; performing temporary recording, local scene impression updating or formal user impression updating according to the state discrimination result; and controlling a subsequent interaction strategy based on the updated user impression result. The device comprises an interaction data processing module, a state discrimination module, an impression updating module, an interaction control module and a counter-evidence verification module. The application can reduce the probability of user impression mis-updating and improve the pertinence and stability of the subsequent interaction strategy.

Description

Technical Field

[0001] This invention relates to the field of intelligent interactive robots, and more particularly to a method and apparatus for updating user impressions of interactive robots. Background Technology

[0002] With the development of natural language processing, human-computer dialogue control, and smart terminal technologies, interactive robots have been widely used in scenarios such as companionship, intelligent customer service, health management, education and tutoring, and home services. During multi-round interactions, interactive robots typically need to record and update information such as the user's interests, communication habits, sensitive topics, and avoidance tendencies by combining the user's historical dialogue content, current responses, and long-term interaction behavior. This forms a user impression, and the robot adjusts its subsequent topic progression, follow-up questions, and prompting strategies based on this impression, thereby improving the coherence, adaptability, and intelligence of the interaction.

[0003] In existing technologies, user impression updates are mostly based on current response content, keyword matching results, sentiment analysis results, or statistical results from several rounds of dialogue, directly writing, overwriting, or incrementally correcting user impressions. When users refuse to answer, give brief replies, actively change topics, or continuously avoid a particular topic category, existing solutions typically only record this as a single type of avoidance behavior, lacking differentiation regarding the persistence, scenario dependence, cross-scenario diffusion, and recoverability of avoidance behavior. This easily leads to misjudging short-term abnormal avoidance as long-term stable avoidance, and also easily expands local scenario-based avoidance errors into global user impression features, resulting in problems such as insufficient topic progression, repeated avoidance, mismatched prompting strategies, or abnormal contraction of follow-up questioning strategies in subsequent interactions. Therefore, there is an urgent need for a user impression update method for interactive robots to reduce the probability of erroneous user impression updates and improve the targeting and stability of subsequent interaction strategies. Summary of the Invention

[0004] The purpose of this invention is to solve the problem of poor impression updating caused by the coarse recognition of topic avoidance states in existing technologies.

[0005] According to one aspect of the present invention, a method for updating user impressions for an interactive robot is provided, comprising: Obtain historical and current user interaction data around the target topic category, and establish a historical interaction baseline corresponding to the target topic category based on the historical interaction data; Based on the current interaction data and the historical interaction baseline, determine whether the current interaction data meets the topic avoidance triggering condition; When the current interaction data satisfies the topic avoidance triggering condition, the corresponding topic avoidance segment is extracted. Based on the topic avoidance segment, a set of discrimination parameters for characterizing the topic avoidance state is determined. The set of discrimination parameters includes avoidance intensity parameters, consecutive avoidance rounds parameters, scene locking parameters, cross-scene diffusion parameters, and recovery reversal parameters. Based on the discrimination parameter set, the topic avoidance segment is identified in terms of state, and when the preset state determination condition is met, one of the following is output: transient disturbance type avoidance state label, scene stable type avoidance state label, or global stable type avoidance state label, and the topic avoidance segment is labeled as one of the following: transient disturbance type avoidance state, scene stable type avoidance state, or global stable type avoidance state. Based on the state identification results, perform temporary recording, partial scene impression update, or formal user impression update respectively; Based on the updated user impressions, control the subsequent interaction strategy of the interactive robot around the target topic category.

[0006] Preferably, the step of performing state identification on the topic avoidance segment based on the discriminant parameter set includes: If the avoidance intensity parameter is greater than or equal to the avoidance trigger threshold, first determine whether the consecutive avoidance rounds parameter is less than or equal to the short duration threshold, and whether the recovery reversal parameter is greater than or equal to the recovery boundary threshold. When the consecutive avoidance round parameter is less than or equal to the short duration threshold and the recovery reversal parameter is greater than or equal to the recovery boundary threshold, output the transient disturbance-type avoidance state label corresponding to the topic avoidance segment and mark the topic avoidance segment as a transient disturbance-type avoidance state. Without outputting the transient disturbance-type avoidance status label, determine whether the continuous avoidance round parameter is greater than or equal to the continuous avoidance threshold, whether the scene locking parameter is greater than or equal to the scene concentration threshold, and whether the cross-scene diffusion parameter is less than the scene diffusion boundary threshold. When the continuous avoidance round parameter is greater than or equal to the continuous avoidance threshold, the scene locking parameter is greater than or equal to the scene concentration threshold, and the cross-scene diffusion parameter is less than the scene diffusion boundary threshold, the scene stable avoidance state label corresponding to the topic avoidance segment is output, and the topic avoidance segment is marked as a scene stable avoidance state. Without outputting the scene-stable avoidance state label, determine whether the continuous avoidance round parameter is greater than or equal to the continuous avoidance threshold, whether the recovery reversal parameter is less than the recovery boundary threshold, and whether the cross-scene diffusion parameter is greater than or equal to the scene diffusion boundary threshold. When the continuous avoidance round parameter is greater than or equal to the continuous avoidance threshold, the recovery reversal parameter is less than the recovery boundary threshold, and the cross-scene diffusion parameter is greater than or equal to the scene diffusion boundary threshold, the global stable avoidance state label corresponding to the topic avoidance segment is output, and the topic avoidance segment is marked as a global stable avoidance state. Wherein, the short duration threshold is less than the duration avoidance threshold.

[0007] Preferably, the step of outputting the transient disturbance-type avoidance state label corresponding to the topic avoidance segment and marking the topic avoidance segment as a transient disturbance-type avoidance state includes: Starting from the beginning of the topic avoidance phase, extract continuous interaction data surrounding the target topic category within a preset short observation window; Based on the avoidance and maintenance status of each round of interaction within the preset short observation window, the number of interaction rounds that continuously meet the avoidance trigger threshold is counted, and the statistical result is determined as the continuous avoidance round parameter. Based on the recovery status of each round of interaction within the preset short observation window, the number of interaction rounds that meet the conditions for restoring normal communication is counted, and the recovery reversal parameter is determined according to the ratio between the number of interaction rounds that meet the conditions for restoring normal communication and the total number of rounds in the preset short observation window. When the avoidance intensity parameter is greater than or equal to the avoidance trigger threshold, the consecutive avoidance rounds parameter is less than or equal to the short duration threshold, and the recovery reversal parameter is greater than or equal to the recovery boundary threshold, the transient disturbance-type avoidance state label corresponding to the topic avoidance segment is output, and the topic avoidance segment is marked as a transient disturbance-type avoidance state.

[0008] Preferably, the step of outputting the scene-stable avoidance state label corresponding to the topic avoidance segment and marking the topic avoidance segment as a scene-stable avoidance state includes: Based on the scene tags corresponding to the topic avoidance segment, extract historical avoidance events that are the same as the target topic category and have the same avoidance direction within the preset historical observation window; The number of historical avoidance events with the same scene tag as the topic avoidance segment is counted, and the number of historical avoidance events corresponding to other single scene tags in the historical avoidance events is counted. The scene locking parameter is determined based on the ratio between the number of historical avoidance events with the same scene tag as the topic avoidance segment and the number of historical avoidance events corresponding to other single scene tags. The number of different scene tags covered by the historical avoidance events is counted, and the cross-scene diffusion parameter is determined based on the number of different scene tags; When the avoidance intensity parameter is greater than or equal to the avoidance trigger threshold, the consecutive avoidance rounds parameter is greater than or equal to the continuous avoidance threshold, the scene locking parameter is greater than or equal to the scene concentration threshold, and the cross-scene diffusion parameter is less than the scene diffusion boundary threshold, the scene stable avoidance state label corresponding to the topic avoidance segment is output, and the topic avoidance segment is marked as a scene stable avoidance state.

[0009] Preferably, the step of outputting the globally stable avoidance state label corresponding to the topic avoidance segment and marking the topic avoidance segment as a globally stable avoidance state includes: Extract historical avoidance events that are of the same category as the target topic and have the same avoidance direction from the preset historical observation window; The number of different scene tags covered by the historical avoidance events is counted, and the cross-scene diffusion parameter is determined based on the number of different scene tags; In the subsequent observation window after the topic avoidance segment, the number of interaction rounds that meet the conditions for resuming normal communication is counted, and the restoration reversal parameter is determined based on the ratio between the number of interaction rounds that meet the conditions for resuming normal communication and the total number of rounds in the subsequent observation window. When the avoidance intensity parameter is greater than or equal to the avoidance trigger threshold, the consecutive avoidance rounds parameter is greater than or equal to the continuous avoidance threshold, the cross-scene diffusion parameter is greater than or equal to the scene diffusion boundary threshold, and the recovery reversal parameter is less than the recovery boundary threshold, the global stable avoidance state label corresponding to the topic avoidance segment is output, and the topic avoidance segment is marked as a global stable avoidance state.

[0010] Preferably, the determination of the avoidance intensity parameter and the consecutive avoidance rounds parameter includes: Based on the historical interaction data, extract two or more of the following: historical topic continuation probability, historical average response length, historical refusal or avoidance expression frequency, and historical active topic shifting frequency corresponding to the target topic category, as the baseline statistical items of the historical interaction baseline. Based on the current interaction data, extract the current topic continuation probability, current average response length, current frequency of refusal to answer or avoidance of expression, and current frequency of actively changing topics; The avoidance intensity parameter is determined based on the difference between the current topic continuation probability and the historical topic continuation probability, the difference between the current average response length and the historical average response length, the difference between the current frequency of refusal to answer or avoidance and its corresponding baseline statistical upper limit, and the difference between the current frequency of actively changing topics and its corresponding baseline statistical upper limit. When the probability of the current topic continuing is lower than the probability of the historical topic continuing, the current average reply length is lower than the historical average reply length, and at least one of the current frequency of refusing to answer or avoiding expression and the current frequency of actively changing topics is higher than the corresponding baseline statistical upper limit, the avoidance intensity parameter is determined to be greater than or equal to the avoidance trigger threshold. When the current interaction data meets the topic avoidance triggering condition, the current interaction data and several subsequent rounds of interaction data are extracted to form the topic avoidance segment; Each round of interaction in the topic avoidance segment is judged to determine whether the avoidance intensity parameter is greater than or equal to the avoidance trigger threshold, and the number of consecutive interaction rounds that meet the avoidance trigger threshold is accumulated to determine the consecutive avoidance round parameter.

[0011] Preferably, the determination of the scene locking parameter, the cross-scene diffusion parameter, and the recovery reversal parameter includes: Extract historical avoidance events that are the same as the target topic category and have the same avoidance direction within the preset historical observation window, and classify and count them according to scene tags to determine the scene locking parameter and the cross-scene diffusion parameter; The scene tags include time period tags and task type tags; The scene locking parameter is determined by the ratio between the number of historical avoidance events with the same scene label as the topic avoidance segment and the number of historical avoidance events corresponding to other single scene labels; The cross-scenario diffusion parameter is determined by the number of different scenario labels covered by the historical avoidance event; In the subsequent observation window, it is determined round by round whether the conditions for restoring normal communication are met, and the restoration reversal parameter is determined based on the ratio between the number of interaction rounds that meet the conditions for restoring normal communication and the total number of rounds in the subsequent observation window; The conditions for restoring normal communication are as follows: the probability of topic continuation in the corresponding interaction round is not lower than the probability of historical topic continuation, and at least two of the following conditions are met: there is no refusal to answer or avoidance of expression, and there is no active topic shifting behavior.

[0012] Preferably, the step of performing temporary recording, partial scene impression update, or formal user impression update according to the marked state of the topic avoidance segment includes: When the topic avoidance segment is marked as the transient disturbance-type avoidance state, only temporary records are generated for the topic avoidance segment, and the formal user impression is not updated. When the topic avoidance segment is marked as the stable avoidance state of the scene, the user's avoidance information of the target topic category is bound and stored with the scene tag corresponding to the topic avoidance segment in order to update the local scene impression, but not the formal user impression. When the topic avoidance segment is marked as the globally stable avoidance state, information about the user's stable avoidance tendency towards the target topic category is written into the formal user impression.

[0013] Preferably, based on the updated user impression results, the interactive robot is controlled to control at least one of the following: topic advancement intensity, follow-up question frequency, prompting method, and topic switching strategy around the target topic category; After the formal user impression update is completed, the counter-evidence verification is performed based on the recovery reversal parameters in the subsequent observation window and the cross-scenario diffusion parameters re-determined based on the avoidance events in the subsequent observation window. When the recovery reversal parameter indicates that the user has met the conditions for resuming normal communication and / or the redefined cross-scenario diffusion parameter indicates that there are fewer than two different scenario labels covered by subsequent avoidance events, the formal user impression update is revoked and the user impression state before the update is restored.

[0014] The present invention also provides a user impression updating apparatus for an interactive robot, applied to any of the above-described user impression updating methods for interactive robots, comprising: The interaction data processing module is used to process the user's historical interaction data and current interaction data around the target topic category, so as to form the historical interaction baseline and topic avoidance segment related data corresponding to the target topic category; The state identification module, connected to the interactive data processing module, is used to receive data related to the topic avoidance segment and output the identification result of the topic avoidance state based on the discrimination parameter set. The identification result includes one of the following: transient disturbance type avoidance state, scene stable type avoidance state, and global stable type avoidance state. An impression update module, connected to the state identification module, is used to receive the identification result and output the user impression update result, wherein the user impression update result includes one of temporary record result, local scene impression update result and formal user impression update result; An interactive control module, connected to the impression update module, is used to receive the user impression update result and control the interactive robot's subsequent interaction strategy around the target topic category; The counter-evidence verification module is connected to the interactive data processing module and the impression update module respectively. It is used to receive subsequent observation window data and the official user impression update result, and output a rollback control signal to the impression update module when the rollback condition is met. The discrimination parameter set includes avoidance intensity parameter, consecutive avoidance rounds parameter, scene locking parameter, cross-scene diffusion parameter and recovery reversal parameter. The impression update module is also used to revoke the formal user impression update result in response to the rollback control signal.

[0015] This invention offers the following advantages: By acquiring historical and current user interaction data related to a target topic category, a corresponding historical interaction baseline is established. Upon detecting a topic avoidance trigger condition, a topic avoidance segment is extracted. Furthermore, based on avoidance intensity parameters, consecutive avoidance round parameters, scene locking parameters, cross-scene diffusion parameters, and recovery reversal parameters, the topic avoidance segment is state-identified, subdividing it into transient disturbance-type avoidance states, scene-stable avoidance states, and globally stable avoidance states. This effectively distinguishes between short-term abnormal avoidance, local scene avoidance, and globally stable avoidance, preventing the miswriting of temporary avoidance into the final user impression and avoiding the inappropriate extension of avoidance limited to specific scenarios into global features. This improves the accuracy of user impression updates, reduces the probability of erroneous updates, and enhances the targeting and stability of subsequent topic advancement, follow-up question control, and topic switching by the interactive robot. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a logic block diagram of a user impression update method for an interactive robot according to an embodiment of the present invention. Detailed Implementation

[0018] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0020] Please refer to Figure 1This invention addresses the problems of existing interactive robots mostly treating user topic avoidance as a single avoidance, making it difficult to distinguish between short-term abnormal avoidance, local scenario avoidance, and global stable avoidance, which easily leads to incorrect updates of formal user impressions, misgeneralization of local features, and mismatch of subsequent topic advancement and follow-up questioning strategies. The invention proposes a user impression update method for interactive robots.

[0021] This invention does not merely determine whether a user "avoids" the topic, but rather, around the target topic category, sequentially completes the following steps: establishing a historical interaction baseline, determining avoidance triggers, extracting topic avoidance segments, constructing a set of discriminant parameters, identifying states, updating hierarchical impressions, and controlling subsequent interactions. This allows the avoidance recognition results to directly serve the control of the interactive robot.

[0022] One embodiment of the present invention provides a method for updating user impressions for an interactive robot, comprising: S10. Obtain historical and current interaction data of users around the target topic category. Based on the historical interaction data, establish a historical interaction baseline corresponding to the target topic category. Historical interaction data is used to characterize the normal communication pattern of the same user around the same target topic category, while current interaction data is used to characterize the immediate communication status around the target topic category in the current conversation. By establishing a historical interaction baseline first, rather than directly judging whether avoidance exists based on uniform empirical rules, subsequent identification can be based on the user's own historical communication habits, thereby reducing the impact of differences in expression style, response length, topic changing habits, and topic sensitivity among different users on the identification results.

[0023] S20. Based on the current interaction data and the historical interaction baseline, determine whether the current interaction data meets the topic avoidance trigger conditions. It should be noted that this step does not automatically conclude that the user has an avoidance tendency simply because of short replies, refusal to answer, or topic changes. Instead, it compares the current interaction data with the historical interaction baseline to identify whether the current communication state has deviated from the user's normal communication pattern around the target topic category. Preferably, the topic avoidance trigger conditions can be comprehensively judged by combining factors such as decreased topic continuity, decreased reply length, and an abnormal increase in refusal to answer or active topic shifting behavior, to avoid false triggering caused by occasional, transient communication fluctuations.

[0024] S30. When the current interaction data meets the topic avoidance triggering condition, extract the corresponding topic avoidance segment. Based on the topic avoidance segment, determine the set of discriminant parameters to characterize the topic avoidance state. The discriminant parameter set includes avoidance intensity parameters, consecutive avoidance rounds parameters, scene locking parameters, cross-scene diffusion parameters, and recovery reversal parameters. It should be noted that this step does not treat "whether avoidance occurred" as the sole analysis object, but further decomposes the avoidance behavior into multiple quantifiable dimensions. Specifically, the avoidance intensity parameter characterizes the degree of deviation of the current interaction from the historical interaction baseline; the consecutive avoidance rounds parameter characterizes the maintenance of avoidance behavior in consecutive rounds; the scene locking parameter characterizes whether avoidance is mainly concentrated in a specific scene; the cross-scene diffusion parameter characterizes whether avoidance has expanded to multiple different scenes; and the recovery reversal parameter characterizes whether normal communication is restored in subsequent interactions after avoidance occurs. By first extracting the topic avoidance segment and then constructing the discriminant parameter set, multi-dimensional evidence such as intensity, persistence, scene-based nature, diffusion, and recovery can be provided for subsequent state identification, thereby avoiding the conflation of avoidance behaviors with different characteristics.

[0025] S40. Based on the discriminant parameter set, perform state identification on the topic avoidance segment, and when the preset state determination conditions are met, output one of the following labels: transient disturbance-type avoidance state label, scene-stable avoidance state label, or globally stable avoidance state label, and label the topic avoidance segment as one of these three states. It should be noted that the technical focus of this step is not on determining whether the user has engaged in avoidance, but rather on identifying whether the avoidance belongs to a transient disturbance-type avoidance state, a scene-stable avoidance state, or a globally stable avoidance state based on the discriminant parameter set, thus providing a basis for subsequent hierarchical impression updates and interactive control. Preferably, this step uses a sequential splitting approach to complete state identification, i.e., first identifying short-term abnormal samples, then identifying local scene-stable samples, and finally identifying globally stable samples. The reason is that: first, excluding short-term disturbances can avoid mistakenly writing short-term abnormalities into the formal user impression; second, identifying scene stability can avoid mistakenly generalizing local scene features to global features; and finally, identifying global stability can ensure that only continuous, diffuse, and unrecovered avoidance behaviors enter the formal user impression update process.

[0026] S50. Based on the state identification results, perform temporary recording, local scene impression update, or formal user impression update respectively. It should be noted that the user impression update in this step is not a uniform write, but rather a layered process based on the state of the topic avoidance segment. For transient disturbance-type avoidance states, only temporary records are generated to suppress the contamination of formal user impressions by short-term abnormal samples; for scene-stable avoidance states, only local scene impressions bound to scene tags are updated to avoid local scene features being incorrectly extended to global user features; only for globally stable avoidance states is a formal user impression written to ensure that the formal user impression truly corresponds to a long-term, cross-scene, stable avoidance tendency.

[0027] S60. Based on the updated user impression results, control the subsequent interaction strategy of the interactive robot around the target topic category. It should be noted that this step ensures that the aforementioned identification and update results truly affect the subsequent control process of the interactive robot, rather than remaining at the static recording level. Specifically, the interactive robot can adjust at least one of the following based on the updated user impression results: the intensity of topic advancement, the frequency of follow-up questions, the prompting method, and the topic switching strategy for the target topic category. Specifically, for cases where only temporary records are generated, cautious probing can be maintained without immediate and comprehensive avoidance; for cases where a partial scene impression is formed, the advancement intensity can be reduced only in the corresponding scene; for cases where a formal user impression is formed, the frequency of follow-up questions can be reduced more broadly or alternative topics can be actively switched.

[0028] The technical solution implemented in this embodiment can first establish an individualized historical interaction baseline based on historical interaction data, then determine whether the current interaction data triggers topic avoidance, and after triggering, use the topic avoidance segment and multi-dimensional discriminant parameter set to perform fine-grained state identification of avoidance behavior, thereby effectively distinguishing between short-term abnormal avoidance, local scenario avoidance, and globally stable avoidance. This reduces the probability of erroneous updates to the user's impressions and suppresses misgeneralization of local features, making the interactive robot's topic advancement, follow-up questions, prompts, and switching control around the target topic category more closely match the user's actual communication state.

[0029] In one specific embodiment, state identification of the topic avoidance segment is performed based on a discriminant parameter set, including: S41. When the avoidance intensity parameter is greater than or equal to the avoidance trigger threshold, first determine whether the consecutive avoidance rounds parameter is less than or equal to the short duration threshold, and whether the recovery reversal parameter is greater than or equal to the recovery boundary threshold. It should be noted that this step prioritizes the avoidance intensity parameter as a unified entry condition to exclude ordinary communication fluctuations that do not reach an effective avoidance level. Then, it prioritizes judging "short duration + high recovery" because: if an avoidance behavior has been triggered but only appears in a small number of consecutive rounds and quickly recovers to normal communication in subsequent interactions, it is more likely to be a short-term disturbance sample rather than a stable impression feature. By prioritizing the identification of this type of sample, it is possible to avoid subsequent judgments on local and global stability based on short-term abnormal samples.

[0030] S42. When the consecutive avoidance round parameter is less than or equal to the short duration threshold, and the recovery reversal parameter is greater than or equal to the recovery boundary threshold, output the transient disturbance-type avoidance state label corresponding to the topic avoidance segment, and label the topic avoidance segment as a transient disturbance-type avoidance state. It should be noted that this step does not identify mild avoidance in the general sense, but rather an avoidance type that "has reached the avoidance trigger level, but the duration is short and the subsequent recovery is fast." Outputting and labeling this type of state separately is beneficial for subsequently executing only temporary recordings without directly updating the formal user impression, thereby suppressing the contamination of the formal user impression by short-term emotional fluctuations, temporary environmental interference, or momentary communication interruptions.

[0031] S43. Without outputting a transient disturbance-type avoidance status label, determine whether the continuous avoidance round parameter is greater than or equal to the persistent avoidance threshold, whether the scene locking parameter is greater than or equal to the scene concentration threshold, and whether the cross-scene diffusion parameter is less than the scene diffusion boundary threshold. It should be noted that this step, after excluding short-term disturbance samples, further identifies whether the avoidance behavior is stable within a local scene. Specifically, the continuous avoidance round parameter confirms whether the avoidance is persistent, the scene locking parameter confirms whether the avoidance is mainly concentrated in the current scene, and the cross-scene diffusion parameter confirms whether the avoidance has not yet spread to multiple different scenes. By jointly judging these three factors, "persistent but localized" avoidance behavior can be separated from short-term abnormal samples and globally diffused samples.

[0032] S44. When the continuous avoidance round parameter is greater than or equal to the continuous avoidance threshold, the scene locking parameter is greater than or equal to the scene concentration threshold, and the cross-scene diffusion parameter is less than the scene diffusion boundary threshold, output the scene-stable avoidance state label corresponding to the topic avoidance segment, and mark the topic avoidance segment as a scene-stable avoidance state. It should be noted that the avoidance state corresponding to this step is: the avoidance behavior has been stably present in continuous rounds and is mainly concentrated in a specific scene, but has not yet formed cross-scene diffusion. Therefore, this type of state is more suitable as the basis for updating local scene impressions, rather than being directly written into the formal user impression. By processing it in this way, we can avoid generalizing stable avoidance errors that only exist in a specific time period, a specific task type, or a specific interaction context into global user characteristics.

[0033] S45. Without outputting a scenario-stable avoidance state label, determine whether the consecutive avoidance rounds parameter is greater than or equal to the continuous avoidance threshold, whether the recovery reversal parameter is less than the recovery boundary threshold, and whether the cross-scenario diffusion parameter is greater than or equal to the scenario diffusion boundary threshold. It should be noted that this step is used to further filter globally stable avoidance states. Compared to scenario-stable avoidance states, this step requires not only that the avoidance behavior is persistent, but also that it has crossed local scenario boundaries and has not resumed normal communication in subsequent interactions. In other words, only when the avoidance behavior simultaneously meets the criteria of "persistent existence + diffusion to multiple scenarios + no subsequent recovery" can it better reflect that the user has formed a stable and widely applicable avoidance tendency around the target topic category.

[0034] S46. When the continuous avoidance round parameter is greater than or equal to the continuous avoidance threshold, the recovery reversal parameter is less than the recovery boundary threshold, and the cross-scene diffusion parameter is greater than or equal to the scene diffusion boundary threshold, output the global stable avoidance state label corresponding to the topic avoidance segment, and label the topic avoidance segment as a global stable avoidance state. It should be noted that this step does not identify an enhanced version of local scene avoidance, but rather a global stable avoidance state that can characterize a long-term, cross-scene stable avoidance tendency. After outputting and labeling this type of state, it can be used as the basis for writing formal user impressions, enabling the interactive robot to adjust the topic advancement intensity, follow-up question frequency, prompting method, or topic switching strategy around the target topic category in a wider range, thereby reducing repeated triggering and strategy mismatch in subsequent interactions.

[0035] Among them, the short-duration threshold is lower than the continuous avoidance threshold. Setting the short-duration threshold to be lower than the continuous avoidance threshold can logically separate the short-term disturbance judgment interval from the stable avoidance judgment interval, avoiding the same set of continuous avoidance round parameters from falling into the short-term disturbance judgment and stable avoidance judgment range at the same time, thereby improving the clarity of the state boundary and the consistency of the entire state identification process.

[0036] The technical solution implemented in this embodiment can effectively distinguish between short-term anomaly avoidance, local scene avoidance, and global stable avoidance, avoiding the miswriting of short-term anomaly samples into the formal user impression, and also avoiding the incorrect generalization of local scene features to global features, thereby improving the clarity of state identification results and the accuracy of user impression updates.

[0037] In one specific embodiment, the transient disturbance-type avoidance state label corresponding to the topic avoidance segment is output, and the topic avoidance segment is marked as a transient disturbance-type avoidance state, including: Starting from the initial round of the topic avoidance phase, continuous interaction data surrounding the target topic category is extracted within a preset short observation window. It should be noted that this step uses the initial round of the topic avoidance phase as the starting point for analysis. The purpose is to ensure that subsequent judgments strictly revolve around the same avoidance event, avoiding the mixing of normal communication data prior to the avoidance into the analysis process. The preset short observation window preferably covers several consecutive rounds of interaction after the initial round, used to characterize the evolution of this avoidance within a short timeframe. The reason for using a short observation window instead of a long window is that this embodiment targets transient, perturbation-type avoidance states. Its identification focus is on determining whether the avoidance behavior is only short-lived and quickly reversed. Therefore, it is necessary to examine its persistence and recoverability within the timeframe immediately adjacent to the avoidance starting point, rather than performing generalization analysis on a longer timescale.

[0038] Based on the avoidance maintenance status corresponding to each round of interaction within a preset short observation window, the number of interaction rounds that continuously meet the avoidance trigger threshold is counted, and the statistical result is determined as the continuous avoidance round parameter. It should be noted that the continuous avoidance round parameter in this step is used to characterize the continuous maintenance length of the avoidance behavior within the short observation window. Specifically, it can be determined round by round whether the avoidance intensity parameter corresponding to each round of interaction still reaches the avoidance trigger threshold. If the threshold is reached for multiple consecutive rounds, it indicates that the avoidance behavior is still maintained; if it is quickly interrupted within the short observation window, it indicates that the avoidance lacks stable continuity. By introducing the continuous avoidance round parameter, this embodiment no longer makes judgments based solely on single-round anomalies, but further expands "whether avoidance occurred" to "how long the avoidance lasted," thereby providing a time dimension basis for subsequently distinguishing between short-term disturbances and stable avoidance.

[0039] Based on the recovery status of each round of interaction within a preset short observation window, the number of interaction rounds that meet the conditions for resuming normal communication is counted. The recovery reversal parameter is determined by the ratio of this number to the total number of rounds within the preset short observation window. It should be noted that the recovery reversal parameter in this step characterizes whether the avoidance behavior can recover to a normal communication state close to the historical interaction baseline within the short observation window. The recovery judgment here preferably does not rely on a single surface phenomenon, but rather combines multiple communication characteristics such as topic continuation, refusal to answer or avoidance expressions, and proactive topic shifting, thereby avoiding misjudgment of recovery simply because a few rounds of responses are slightly longer or there is no refusal to answer temporarily. By ratioing the number of interaction rounds that meet the conditions for resuming normal communication to the total number of rounds within the short observation window, "whether recovery has occurred" can be transformed into a continuous indicator of "the degree of recovery," making the identification of transient disturbance-type avoidance states based on quantifiable recoverability.

[0040] When the avoidance intensity parameter is greater than or equal to the avoidance trigger threshold, the consecutive avoidance rounds parameter is less than or equal to the short duration threshold, and the recovery reversal parameter is greater than or equal to the recovery boundary threshold, a transient disturbance-type avoidance state label corresponding to the topic avoidance segment is output, and the topic avoidance segment is labeled as a transient disturbance-type avoidance state. It should be noted that this step identifies the transient disturbance-type avoidance state through the combination of "effective avoidance already formed + short duration + rapid subsequent recovery." Specifically, if the avoidance intensity parameter reaches the avoidance trigger threshold, it indicates that the sample is not a normal communication fluctuation; if the consecutive avoidance rounds parameter does not exceed the short duration threshold, it indicates that the avoidance does not possess long-term stable maintenance characteristics; if the recovery reversal parameter reaches the recovery boundary threshold, it indicates that the user has a high probability of resuming normal communication within a short observation window. Therefore, this embodiment identifies not a mild avoidance in the general sense, but a special avoidance state that is "triggered but reversible, short-term, and unstable." Outputting its state label separately and completing the state labeling is beneficial for generating only temporary records instead of directly writing them into the formal user impression, thereby avoiding short-term abnormal samples from contaminating the formal user impression and preventing the interactive robot from prematurely ceasing its reasonable exploration of the target topic category.

[0041] For example, when an interactive robot communicates with a user about the target topic category "health status", if the user makes a brief reply and an evasive expression during the current interaction, the current interaction data meets the topic evasion trigger condition.

[0042] Subsequently, within a preset short observation window starting from the initial round of the topic avoidance segment, only when the avoidance intensity parameter corresponding to two consecutive rounds of interaction is greater than or equal to the avoidance trigger threshold, and two out of the subsequent three rounds of interaction meet the conditions for resuming normal communication, the consecutive avoidance round parameter is less than or equal to the short duration threshold, and the recovery reversal parameter is greater than or equal to the recovery boundary threshold.

[0043] Accordingly, the system outputs a transient disturbance-type avoidance status label and marks the topic avoidance segment as a transient disturbance-type avoidance status, generating only temporary records without updating the formal user impression.

[0044] This example demonstrates that avoidance behaviors that occur briefly and recover quickly afterward are not directly solidified into long-term user characteristics.

[0045] The technical solution implemented in this embodiment can take the starting round of the topic avoidance segment as a unified starting point for analysis, and simultaneously examine the persistence and recoverability of avoidance behavior within a preset short observation window. This allows for the separation of short-term abnormal samples that "have reached the avoidance trigger level but have a short duration and rapid subsequent recovery" from stable avoidance samples. Therefore, on the one hand, it can reduce the probability of incorrect updates to formal user impressions caused by short-term emotional fluctuations, environmental interference, or momentary communication interruptions, preventing short-term abnormal errors from becoming long-term user characteristics. On the other hand, it can also provide more refined state information for the subsequent advancement, follow-up questioning, and prompting control of the interactive robot around the target topic category, improving the targeting and stability of interactive control.

[0046] In one specific embodiment, the scenario-stable avoidance state label corresponding to the topic avoidance segment is output, and the topic avoidance segment is marked as a scenario-stable avoidance state, including: Based on the scene tags corresponding to the topic avoidance segment, historical avoidance events with the same category and direction of avoidance as the target topic are extracted within a preset historical observation window. It's important to note that this step does not directly determine whether a situation is a stable avoidance state based solely on the current avoidance behavior. Instead, it first backtracks and extracts historical avoidance events with the same category and direction of avoidance as the target topic, based on the scene tags corresponding to the current topic avoidance segment. This is because a stable avoidance state emphasizes local stability, characterized by "continuous and repetitive occurrences in a specific scenario." Therefore, it's necessary to examine the distribution patterns of similar avoidance events across the scene dimension using a historical observation window, rather than relying solely on localized behavior in the current conversation. This step ensures that subsequent identification is based on the scene consistency of historical samples, thereby improving the reliability of identifying stable avoidance in local scenarios.

[0047] The process involves counting the number of historical avoidance events that share the same scene label as the topic avoidance segment, and then counting the number of historical avoidance events corresponding to other single scene labels. The scene locking parameter is determined based on the ratio between the number of historical avoidance events with the same scene label as the topic avoidance segment and the number of historical avoidance events corresponding to other single scene labels. It's important to note that the scene locking parameter in this step characterizes whether the current avoidance behavior is primarily concentrated in the same scene as the current topic avoidance segment. If the number of historical avoidance events with the same scene label as the current topic avoidance segment significantly outweighs the number of historical avoidance events corresponding to other single scene labels, it indicates that this type of avoidance behavior has a strong scene dependence, rather than being randomly distributed across scenes. By using a ratio method to determine the scene locking parameter, the qualitative judgment of "whether it is concentrated in the current scene" can be transformed into a comparable quantitative result, thus avoiding the misjudgment of a stable avoidance state based solely on occasional avoidance in a particular scene.

[0048] The number of different scenario labels covered by historical avoidance events is counted, and cross-scenario diffusion parameters are determined based on the number of different scenario labels. It should be noted that the cross-scenario diffusion parameter in this step is used to characterize whether the current avoidance behavior has expanded from a local scenario to multiple different scenarios. Unlike the scenario locking parameter, which focuses on "whether the current scenario is dominant," the cross-scenario diffusion parameter focuses on "how many scenarios the same type of avoidance event has spread to overall." If the number of different scenario labels covered by historical avoidance events is small, it indicates that the avoidance behavior is still mainly limited to a few scenarios; if the number of covered scenario labels is large, it indicates that the avoidance behavior may have a cross-scenario generalization trend. By introducing this parameter, it is possible to prevent avoidance behaviors that occur frequently in the current scenario but have actually spread to multiple scenarios from being misidentified as a stable avoidance state in a local scenario.

[0049] When the avoidance intensity parameter is greater than or equal to the avoidance trigger threshold, the consecutive avoidance rounds parameter is greater than or equal to the continuous avoidance threshold, the scene locking parameter is greater than or equal to the scene concentration threshold, and the cross-scene diffusion parameter is less than the scene diffusion boundary threshold, the scene-stable avoidance state label corresponding to the topic avoidance segment is output, and the topic avoidance segment is marked as a scene-stable avoidance state. It should be noted that this step identifies the scene-stable avoidance state through the combination of "effective avoidance has been formed + it has persistence + high scene concentration + it has not yet spread across scenes". Among them, the avoidance intensity parameter reaching the avoidance trigger threshold indicates that the current sample is not a normal communication fluctuation; the consecutive avoidance rounds parameter reaching the continuous avoidance threshold indicates that the avoidance behavior is not a short-term anomaly, but has been maintained in consecutive rounds; the scene locking parameter reaching the scene concentration threshold indicates that the avoidance behavior is mainly concentrated in the current scene; and the cross-scene diffusion parameter being lower than the scene diffusion boundary threshold indicates that the avoidance behavior has not yet formed a consistent stable avoidance tendency across scenes. It can be seen that what this embodiment identifies is not general stable avoidance, but a specific avoidance state that "persistently exists in a local scene but has not expanded into a global feature". Outputting and labeling it as a scene-stable avoidance state is beneficial because it allows for subsequent updates to only the local scene impressions bound to the scene tags, rather than directly writing them into the formal user impressions, thereby avoiding misgeneralization of local features.

[0050] For example, in the scenario of "night + health inquiry task", the interactive robot communicated with the user multiple times around the target topic category "sleep status". The user gave short replies and took the initiative to change the topic in multiple rounds, and the continuous avoidance round parameter reached the continuous avoidance threshold.

[0051] Further analysis revealed that within the preset historical observation window, the number of historical avoidance events with the same scene tag as the current topic avoidance segment was significantly higher than the number of historical avoidance events corresponding to other single scene tags, and the scene locking parameter reached the scene concentration threshold; at the same time, the number of different scene tags covered by historical avoidance events was relatively small, and the cross-scene diffusion parameter was lower than the scene diffusion boundary threshold.

[0052] Based on this, the system outputs a scenario-stable avoidance status label and marks the topic avoidance segment as a scenario-stable avoidance status. It only updates the local scenario impressions bound to the "night + health inquiry task" scenario, without updating the formal user impressions.

[0053] This example demonstrates that persistent avoidance behavior within a local scenario will not be incorrectly extended into a global avoidance feature.

[0054] The technical solution implemented in this embodiment can identify avoidance behaviors that persist within a specific scenario but have not yet expanded to multiple different scenarios by analyzing the scenario distribution of similar historical avoidance events. This allows for the differentiation of such behaviors from short-term abnormal avoidance and globally stable avoidance. Therefore, on the one hand, it prevents avoidance tendencies existing only in local scenarios from being mistakenly written into the formal user impression, reducing the risk of misgeneralization of local features; on the other hand, it enables the interactive robot to adjust the intensity of topic progression, follow-up questioning methods, or prompting strategies around the target topic category only in the corresponding scenario, rather than uniformly avoiding the topic in all scenarios. This improves the hierarchy of user impression updates and the targeting and stability of subsequent interactive control.

[0055] In one specific embodiment, the globally stable avoidance state label corresponding to the topic avoidance segment is output, and the topic avoidance segment is marked as a globally stable avoidance state, including: Extract historical avoidance events within the preset historical observation window that are of the same category as the target topic and have the same avoidance direction. It's important to note that this step does not directly conclude that a user has formed a globally stable avoidance tendency based solely on a single avoidance behavior. Instead, it first traces back historical avoidance events within the historical observation window that are of the same category as the current target topic and have the same avoidance direction to confirm whether this type of avoidance has historical continuity and a basis in sample accumulation. This is because a globally stable avoidance state should reflect a long-term, stable, and repeatable avoidance characteristic formed by the user around the target topic category, rather than an occasional judgment caused by a single or small number of local samples. This step ensures that subsequent identification of global stability is based on the continuous existence of similar historical events, improving the reliability of the basis for writing formal user impressions.

[0056] The number of different scenario tags covered by historical avoidance events is counted, and cross-scenario diffusion parameters are determined based on the number of different scenario tags. It should be noted that the cross-scenario diffusion parameters in this step are used to characterize whether the current avoidance behavior has expanded from a local scenario to multiple different scenarios. Unlike scenario-stable avoidance states, which mainly emphasize "whether it is concentrated in a specific scenario," globally stable avoidance states emphasize "whether it has broken through the limitations of a single scenario and recurred in multiple different scenarios." Therefore, by counting the number of different scenario tags covered by historical avoidance events, the distribution range of avoidance behavior can be transformed from a qualitative description into a comparable quantitative result. If the number of covered scenario tags is large, it indicates that the avoidance behavior no longer depends on a single time period, a single task type, or a single interaction context, but has cross-scenario consistency; conversely, if the number of covered scenarios is still small, it is more likely to be local scenario stable avoidance rather than globally stable avoidance.

[0057] Within the subsequent observation window following the topic avoidance phase, the number of interaction rounds that meet the conditions for resuming normal communication is counted. The recovery reversal parameter is determined based on the ratio of this number to the total number of rounds in the subsequent observation window. It's important to note that this step is used to determine whether the current avoidance behavior has recovered to normal communication in subsequent interactions. This is because even if an avoidance behavior has appeared in multiple scenarios, its spread alone cannot definitively classify it as globally stable avoidance. Further confirmation is needed regarding whether the avoidance persists in the subsequent observation window and whether significant recovery has occurred. By counting the number of interaction rounds that meet the conditions for resuming normal communication and then comparing this number to the total number of rounds in the subsequent observation window, "whether recovery has occurred" can be further characterized as a continuous indicator of "the degree of recovery." This not only helps distinguish between complete recovery and no recovery but also helps identify intermediate states such as partial recovery and intermittent recovery, thus avoiding misidentifying diffused samples that have entered the recovery phase as long-term stable global avoidance features.

[0058] When the avoidance intensity parameter is greater than or equal to the avoidance trigger threshold, the consecutive avoidance rounds parameter is greater than or equal to the sustained avoidance threshold, the cross-scene diffusion parameter is greater than or equal to the scene diffusion boundary threshold, and the recovery reversal parameter is less than the recovery boundary threshold, a globally stable avoidance state label corresponding to the topic avoidance segment is output, and the topic avoidance segment is marked as a globally stable avoidance state. It should be noted that this step identifies the globally stable avoidance state through a combination of "effective avoidance already formed + sustained + cross-scene diffusion + no subsequent recovery." Specifically, if the avoidance intensity parameter reaches the avoidance trigger threshold, it indicates that the current sample is not a normal communication fluctuation; if the consecutive avoidance rounds parameter reaches the sustained avoidance threshold, it indicates that the avoidance behavior has a continuous maintenance characteristic; if the cross-scene diffusion parameter reaches the scene diffusion boundary threshold, it indicates that its scope is no longer limited to a single scene; and if the recovery reversal parameter is less than the recovery boundary threshold, it indicates that the avoidance has not significantly recovered within the subsequent observation window. Therefore, this embodiment does not identify an enhanced version of local scene avoidance, but rather a global characteristic that reflects the user's long-term, cross-scene stable avoidance tendency around the target topic category. Outputting and labeling it as a globally stable avoidance state is beneficial for subsequently writing relevant information into the formal user impression, and based on this, allowing for broader control and adjustment of the interactive robot's subsequent topic advancement intensity, follow-up question frequency, prompting methods, and topic switching strategies around the target topic category.

[0059] For example, when the interactive robot communicates with users around the target topic category of "family relationships", users consistently exhibit behaviors such as refusing to answer, avoiding expression, and actively changing the topic in multiple scenarios, including daytime question-and-answer tasks, nighttime companionship tasks, and reminder tasks.

[0060] In this case, the continuous avoidance round parameter reaches the continuous avoidance threshold, and statistics show that the number of different scene labels covered by historical avoidance events has reached the scene diffusion boundary threshold, indicating that the avoidance behavior has cross-scene diffusion characteristics; at the same time, in the subsequent observation window, the proportion of interaction rounds that meet the conditions for resuming normal communication is low, and the recovery reversal parameter is lower than the recovery boundary threshold.

[0061] Based on this, the system outputs a globally stable avoidance status label and marks the topic avoidance segment as a globally stable avoidance status, writing the information that the user has a stable avoidance tendency towards the target topic category into the formal user impression.

[0062] This example demonstrates that only persistent, cross-scenario-spreading avoidance behaviors that are not subsequently reversed will solidify into a formal user impression.

[0063] The technical solution implemented in this embodiment can further identify avoidance behaviors that persist around the target topic category, have expanded to multiple different scenarios, and have not subsequently resumed normal communication, based on the elimination of short-term abnormal avoidance and local scenario avoidance. This distinguishes these behaviors from phased diffusion samples and local scenario samples. Therefore, on the one hand, it avoids erroneously recording avoidance behaviors limited to local scenarios or those that have subsequently resumed in the official user impression, reducing the risk of long-term distortion of the official user impression; on the other hand, it ensures that the official user impression truly corresponds to the user's long-term, cross-scenario, stable avoidance tendency around the target topic category, thereby improving the authenticity and stability of the official user impression and the targeting and consistency of subsequent control strategies for the interactive robot.

[0064] In one specific embodiment, the determination of the avoidance intensity parameter and the consecutive avoidance rounds parameter includes: Based on historical interaction data, two or more of the following are extracted as baseline statistics for the historical interaction baseline: historical topic continuation probability, historical average response length, historical frequency of refusal or avoidance, and historical frequency of proactive topic shifting, corresponding to the target topic category. It is important to note that the focus of this step is to first establish an individualized historical interaction baseline, rather than directly using uniform empirical rules to determine whether avoidance has occurred. Since different users exhibit different expression habits, response lengths, and topic-changing tendencies around the same target topic category, the lack of a historical interaction baseline as a reference can easily lead to interference from individual differences in subsequent avoidance identification. By extracting two or more of these statistics as baseline statistics, the user's normal communication patterns can be characterized from multiple dimensions, including communication continuity, expression sufficiency, and explicit avoidance behavior, providing a stable comparative benchmark for subsequent avoidance intensity assessment. Furthermore, the baseline statistics can be formed based on historical interaction data within a preset statistical period, ensuring that they reflect the user's long-term communication characteristics without introducing distortion due to outdated data.

[0065] Based on current interaction data, the following parameters are extracted: current topic continuation probability, current average response length, current frequency of refusal or avoidance, and current frequency of proactive topic shifting. It's important to note that this step aims to extract current statistical items that correspond one-to-one with the historical interaction baseline, creating a structurally consistent correlation between the current interaction state and historical norms. Specifically, the current topic continuation probability characterizes the user's immediate willingness to continue communicating around the target topic category; the current average response length characterizes the current level of expression; and the current frequency of refusal or avoidance and the frequency of proactive topic shifting characterize current explicit avoidance behavior. By extracting these current statistical items, rather than solely relying on the brevity of a single response or the occurrence of a refusal, the impact of occasional noise on subsequent avoidance identification results can be reduced, allowing the determination of avoidance intensity parameters to be based on a more complete understanding of the current interaction state.

[0066] The avoidance intensity parameter is determined based on the differences between the current topic continuation probability and the historical topic continuation probability, the current average response length and the historical average response length, the current frequency of refusal or avoidance of responses and its corresponding baseline statistical upper limit, and the current frequency of proactive topic shifting and its corresponding baseline statistical upper limit. It should be noted that this step does not separately determine whether multiple statistical items are abnormal, but rather maps multiple deviation features into a unified avoidance intensity parameter. Specifically, the difference between the current topic continuation probability and the historical topic continuation probability represents the degree of decline in the willingness to continue communicating around the target topic category; the difference between the current average response length and the historical average response length represents the degree of decline in the degree of expression development; and the differences between the current frequency of refusal or avoidance of responses and its corresponding baseline statistical upper limit, and the current frequency of proactive topic shifting and its corresponding baseline statistical upper limit, represent whether explicit avoidance behavior exceeds the historical normal fluctuation range. The baseline statistical upper limit is preferably determined based on the statistical distribution, fluctuation range, or preset tolerance interval of historical interaction data to ensure its clear source and clear comparison criteria. By comprehensively forming avoidance strength parameters, subsequent trigger judgments no longer rely on a single abnormal signal, but rather reflect the overall degree of deviation of the current interaction from historical norms.

[0067] When the probability of the current topic continuing is lower than the probability of historical topic continuing, the current average reply length is lower than the historical average reply length, and at least one of the current frequency of refusing to answer or expressing avoidance, or the current frequency of actively changing topics, is higher than the corresponding baseline statistical upper limit, the avoidance intensity parameter is considered to be greater than or equal to the avoidance trigger threshold. It should be noted that this step uses a multi-condition joint triggering mechanism, rather than using a single abnormal phenomenon as the basis for topic avoidance triggering. The reason is that a short reply alone does not necessarily mean avoidance, and a single topic change may be a natural communication switch. Only when the user's continuity around the target topic category decreases, the expression length decreases, and explicit avoidance behavior exceeds the historical normal range in at least one dimension, can it be more accurately stated that the current interaction has formed effective avoidance. This joint triggering method can distinguish genuine avoidance-related interaction deviations from ordinary communication fluctuations, thereby reducing the probability of false triggers and avoiding subsequent state identification based on unstable samples.

[0068] When the current interaction data meets the topic avoidance triggering condition, the current interaction data and several subsequent rounds of interaction data are extracted to form a topic avoidance segment. It should be noted that this step does not immediately make a stable impression judgment after a single round of interaction; instead, it extracts the current interaction data and several subsequent rounds of interaction data to form a topic avoidance segment. This is because a single round of interaction can only reflect an instantaneous state and cannot fully characterize whether the avoidance behavior persists, gradually weakens, or quickly recovers afterward. By forming a topic avoidance segment, subsequent identification can be based on continuous temporal segments of the same avoidance event, rather than on discrete, fragmented samples. This provides a unified analytical object for determining the parameters of consecutive avoidance rounds and lays the foundation for distinguishing between transient, scenario-stable, and globally stable avoidance states.

[0069] The process involves round-by-round evaluation of the avoidance intensity parameter for each interaction within the topic avoidance segment to determine if it is greater than or equal to the avoidance trigger threshold. The number of consecutive interaction rounds that meet the avoidance trigger threshold is accumulated to determine the consecutive avoidance round parameter. It's important to note that the consecutive avoidance round parameter in this step does not count "the total number of avoidances," but rather "the number of consecutive rounds of avoidance behavior." These are not the same: the former is easily affected by scattered anomalous samples, while the latter better reflects whether the avoidance behavior possesses stable maintenance characteristics. By comparing the avoidance intensity parameter with the avoidance trigger threshold round by round and accumulating the number of consecutive rounds that meet the threshold, the persistence of avoidance behavior over time can be quantified.

[0070] It should be noted that the probability of continuation of historical topics can be determined by statistically analyzing the percentage of interaction rounds in which users continuously answered, continued to communicate, or did not actively leave the target topic category; the historical average reply length can be determined by statistically analyzing the average number of words, sentence length, or semantic unit length of users' historical replies around the target topic category; the historical frequency of refusing to answer or avoiding expression can be determined by statistically analyzing the ratio between the number of times users expressed refusing to answer or avoiding expression in historical interaction data and the total number of interaction rounds; and the historical frequency of actively switching topics can be determined by statistically analyzing the ratio between the number of times users actively switched to other topic categories during communication around the target topic category and the total number of interaction rounds.

[0071] Accordingly, the current topic continuation probability, current average response length, current frequency of refusal to answer or avoidance of expression, and current frequency of actively shifting topics can be extracted from the current interaction data using the same or corresponding statistical methods as the historical interaction baseline, so as to ensure that the current statistical items are comparable to the historical statistical items.

[0072] Furthermore, the baseline statistical upper limit is preferably determined based on the fluctuation range, statistical distribution upper bound, or preset tolerance range of the corresponding statistical item in historical interaction data. For example, the normal fluctuation upper bound of the historical refusal to answer or avoidance expression frequency and the historical active topic shifting frequency within a preset statistical period can be used as the corresponding baseline statistical upper limit to characterize the acceptable range of behavioral fluctuations for the user under normal communication conditions.

[0073] The avoidance trigger threshold is preferably determined based on the comprehensive deviation between the historical interaction baseline and the current interaction data, and is used to distinguish between ordinary communication fluctuations and effective avoidance. The short duration threshold is used to limit the maximum number of consecutive avoidance rounds allowed in the short-term disturbance avoidance state, and the continuous avoidance threshold is used to limit the minimum number of consecutive avoidance rounds required to meet the stable avoidance state. The short duration threshold is preferably smaller than the continuous avoidance threshold so that the short-term disturbance judgment interval and the stable avoidance judgment interval are separated from each other.

[0074] The technical solution implemented in this embodiment can first establish an individualized historical interaction baseline based on historical interaction data, and then extract corresponding statistical items from the current interaction data to form an avoidance intensity parameter, which uniformly represents the degree of deviation of the current interaction from the historical norm. After the current interaction data meets the topic avoidance triggering condition, the continuous avoidance round parameter is determined by forming a topic avoidance segment and accumulating the number of rounds that continuously meet the avoidance triggering threshold, so as to quantify the persistence of avoidance behavior in the time dimension. Thus, on the one hand, it can reduce the probability of falsely triggering avoidance recognition based on a single round of anomaly or a single abnormal signal, and improve the accuracy of avoidance recognition; on the other hand, it can provide a unified and reliable parameter basis for the fine-grained distinction of subsequent transient disturbance-type avoidance state, scene-stable avoidance state, and globally stable avoidance state, thereby improving the accuracy and stability of user impression updates and the targeting of subsequent interaction control.

[0075] In one specific embodiment, the determination of the scene locking parameters, cross-scene diffusion parameters, and recovery inversion parameters includes: Within a pre-defined historical observation window, historical avoidance events that are of the same category as the target topic and have the same avoidance direction are extracted. These events are then categorized and statistically analyzed according to scene tags to determine scene locking parameters and cross-scene diffusion parameters. It's important to note that this step preferentially extracts historical avoidance events within the pre-defined historical observation window that are of the same category as the current target topic and have the same avoidance direction. This is because scene locking parameters and cross-scene diffusion parameters do not reflect the immediate characteristics of a single avoidance behavior, but rather the historical scene distribution characteristics of similar avoidance behaviors. By limiting the criteria to "same target topic category" and "consistent avoidance direction," it ensures that the historical samples included in the statistics are consistent with the current topic avoidance segment in terms of the analysis object and behavioral nature, thus avoiding the mixing of communication anomalies of different topics and types into the scene distribution analysis process. Furthermore, categorizing and statistically analyzing according to scene tags helps extend the analysis of avoidance behavior from simple temporal continuity to scene distribution analysis, providing a parameter basis for subsequently distinguishing between stable avoidance in local scenes and stable avoidance across scenes.

[0076] The scenario tags include time period tags and task type tags. It should be noted that the scenario tags in this step preferably include at least a time period tag and a task type tag, so as to describe the interaction scenario in which the current avoidance behavior occurs from both a time dimension and a task context dimension. The time period tag can be used to distinguish different communication times such as morning, daytime, and nighttime, while the task type tag can be used to distinguish different interaction modes such as question-and-answer tasks, health inquiry tasks, emotional companionship tasks, and reminder tasks.

[0077] The scene-locking parameter is determined by the ratio of the number of historical avoidance events with the same scene label as the topic avoidance segment to the number of historical avoidance events corresponding to other single scene labels. It should be noted that the scene-locking parameter in this step is used to characterize whether the current avoidance behavior is mainly concentrated in the same scene as the current topic avoidance segment and occurs repeatedly. If the number of historical avoidance events with the same scene label as the current topic avoidance segment is significantly greater than the number of historical avoidance events corresponding to other single scene labels, it indicates that the avoidance behavior has a strong scene dependence, rather than being evenly distributed across different scenes. Using a ratio method to determine the scene-locking parameter transforms the "whether it is concentrated in the current scene" from an empirical judgment into a quantifiable comparison result, thereby improving the clarity and feasibility of local scene identification.

[0078] The cross-scenario diffusion parameter is determined by the number of different scenario labels covered by historical avoidance events. It should be noted that the cross-scenario diffusion parameter in this step is used to characterize whether the avoidance behavior has expanded from a local scenario to multiple different scenarios. Unlike the scenario locking parameter, which mainly answers "whether it is concentrated in the current scenario," the cross-scenario diffusion parameter mainly answers "how many types of scenarios the avoidance behavior has expanded to overall." If the number of different scenario labels covered by historical avoidance events is small, it indicates that the avoidance behavior is more likely to be limited to a local scenario; if the number of different scenario labels covered is large, it indicates that the avoidance behavior has a cross-scenario consistency trend. By introducing the cross-scenario diffusion parameter, this embodiment can supplement the persistence judgment from the perspective of scenario coverage, avoiding misjudging high-frequency samples in local scenarios as globally stable samples.

[0079] In subsequent observation windows, the system assesses each round to determine if the conditions for resuming normal communication are met. The recovery reversal parameter is determined based on the ratio of the number of interaction rounds meeting the conditions to the total number of rounds in the subsequent observation window. It's important to note that the recovery reversal parameter in this step characterizes whether, after the avoidance behavior occurs, the user recovers to a normal communication state close to the historical interaction baseline in subsequent interactions. This doesn't mean the user has resumed normal communication simply because a single interaction round appears to have recovered. Instead, it prioritizes assessing each round in the subsequent observation window and quantifying the degree of recovery by the ratio of the number of interaction rounds meeting the conditions to the total number of rounds. Using a ratio method instead of a single assessment helps identify different scenarios such as complete recovery, partial recovery, and continued non-recovery, thus avoiding the long-term misinterpretation of samples that have entered the recovery phase as stable avoidance and the mistaking occasional recovery for overall recovery. Through this parameter, this embodiment further introduces a "reversibility" dimension to improve the accuracy of subsequent state identification and formal user impression updates.

[0080] The conditions for resuming normal communication are as follows: the probability of topic continuation in the corresponding interaction round is not lower than the probability of topic continuation in the past, and at least two of the following conditions must be met: no refusal to answer or avoidance of expression, and no active shifting of topic. It should be noted that the conditions for resuming normal communication are not determined based on a single feature, but rather preferably using a multi-condition joint judgment mechanism. This is because simply having longer responses or not having a single instance of refusal to answer is insufficient to indicate that the user has resumed normal communication around the target topic category. Only when the probability of topic continuation in the corresponding interaction round recovers to at least the probability of topic continuation in the past, and at least two of the following conditions are met: no refusal to answer or avoidance of expression, and no active shifting of topic, can it be more accurately stated that this round of interaction has approached the normal communication pattern represented by the historical interaction baseline. This setting improves the stability of the recovery judgment and avoids misjudging superficial, partial, or temporary improvements as true recovery.

[0081] It should be noted that the scene locking parameters, cross-scene diffusion parameters, and recovery reversal parameters in this invention are preferably determined based on historical interaction data and subsequent observation window data, rather than being abstractly set detached from specific interaction scenarios.

[0082] The scene tags preferably include at least a time period tag and a task type tag, and may also include a device status tag, a user interaction mode tag, or an environment tag if necessary. The scene locking parameter can be determined based on the relative size between the number of historical avoidance events with the same scene tag as the current topic avoidance segment and the number of historical avoidance events corresponding to other single scene tags; when the historical avoidance events corresponding to the same scene tag are significantly dominant within the preset historical observation window, it indicates that the avoidance behavior has a strong scene dependence.

[0083] The cross-scene diffusion parameter can be determined based on the number of different scene labels covered by historical avoidance events; when the number of different scene labels covered reaches or exceeds the preset number, the avoidance behavior can be considered to have cross-scene diffusion characteristics. The scene concentration threshold and scene diffusion boundary threshold are preferably determined based on the scene distribution results of historical avoidance events, for example, they can be set by combining the degree of scene clustering in historical samples, the number of scenes covered, or a preset empirical range.

[0084] The optimal recovery reversal parameter is determined based on the ratio between the number of interaction rounds that meet the conditions for restoring normal communication within the subsequent observation window and the total number of rounds in the subsequent observation window. The recovery boundary threshold can be set based on the proportion of recovery rounds within the subsequent observation window to distinguish between "normal communication restored" and "normal communication not restored".

[0085] Among them, the preferred method for restoring normal communication is a multi-condition joint judgment mechanism, which means that at least two of the following should be considered: whether the probability of topic continuation has recovered to a level not lower than the historical probability of topic continuation, whether there is no refusal to answer or avoidance of expression, and whether there is no active behavior of changing the topic. This avoids misjudging the restoration based on a single superficial phenomenon.

[0086] The technical solution implemented in this embodiment can further refine the characterization of avoidance behavior from the perspectives of scene distribution and recovery, beyond continuous analysis. Specifically, the scene locking parameter identifies whether avoidance behavior is concentrated in a specific scene, the cross-scene diffusion parameter identifies whether avoidance behavior has expanded to multiple different scenes, and the recovery reversal parameter identifies whether avoidance behavior returns to normal communication in subsequent interactions. Therefore, on the one hand, it provides unified and reliable parameter support for distinguishing between transient disturbance-type avoidance states, scene-stable avoidance states, and globally stable avoidance states, avoiding misidentification of local scene avoidance as globally stable avoidance, or misidentification of phased diffusion samples as long-term stable avoidance; on the other hand, it also enables subsequent user impression layering updates and adjustments to the control strategy of the interactive robot around the target topic category to be based on a more complete technical basis, thereby improving the accuracy of user impression updates, the stability of formal user impressions, and the targeting of subsequent interactive control.

[0087] In one specific embodiment, depending on the state of the topic avoidance segment, temporary recording, partial scene impression update, or formal user impression update are performed respectively, including: When a topic avoidance segment is marked as a transient disturbance-type avoidance state, only a temporary record is generated for that segment, without updating the official user impression. It's important to note that this step separates triggered but short-lived avoidance behaviors that can be recovered from the official user impression update process. Since transient disturbance-type avoidance states typically reflect abnormal samples caused by short-term emotional fluctuations, temporary environmental disturbances, or momentary communication interruptions, directly writing them into the official user impression could easily contaminate it with short-term anomalies. By generating only a temporary record, the analytical value of the avoidance event can be preserved while preventing it from prematurely solidifying into long-term user characteristics, thus allowing the subsequent interactive robot to retain appropriate room for exploration and dynamic adjustment around the target topic category.

[0088] When a topic avoidance segment is marked as a scenario-stable avoidance state, the user's avoidance information regarding the target topic category is bound and stored with the scenario tag corresponding to the topic avoidance segment. This updates the local scenario impression without updating the overall user impression. It's important to note that the purpose of this step is to limit the update of avoidance behavior that "only persists in a specific scenario" to the corresponding scenario scope, rather than directly extending it to global user characteristics. Since the scenario-stable avoidance state emphasizes local scenario dependence—that is, while the avoidance behavior is persistent, it hasn't expanded to multiple different scenarios—it's more suitable for management as a local scenario impression. By binding and storing avoidance information with scenario tags, the interactive robot can adjust the intensity of topic advancement, the frequency of follow-up questions, or the prompting method around the target topic category in subsequent identical or similar scenarios, while retaining normal interaction capabilities in other scenarios, thus avoiding misgeneralization of local features.

[0089] When a topic avoidance segment is marked as a globally stable avoidance state, information about the user's stable avoidance tendency towards the target topic category is written into the official user impression. It's important to note that this step aims to only write such behavior into the official user impression when it exhibits characteristics such as persistence, cross-scenario diffusion, and no subsequent recovery. This is because a globally stable avoidance state more accurately reflects the user's long-term, consistent, and stable avoidance tendency around the target topic category, making it suitable as the official basis for subsequent broad-based interaction control. By updating the official user impression only for this type of state, it ensures that the official user impression records not short-term abnormal samples or localized scenario samples, but rather real user communication characteristics with high stability and broad applicability.

[0090] The technical solution implemented in this embodiment can perform hierarchical update processing on user impressions based on the different states of the topic avoidance segment. Specifically, for transient disturbance-type avoidance states, only temporary records are generated, which can suppress the contamination of formal user impressions by short-term abnormal samples; for scene-stable avoidance states, only local scene impressions are updated, which can prevent local scene features from being incorrectly extended to global user features; and for globally stable avoidance states, they are written into formal user impressions, which can ensure that formal user impressions truly correspond to long-term, cross-scene stable avoidance tendencies.

[0091] In one specific embodiment, based on the updated user impression results, the interactive robot controls at least one of the following: topic advancement intensity, follow-up question frequency, prompting method, and topic switching strategy around the target topic category. It should be noted that the updated results are not merely static records but further influence the subsequent control of the interactive robot. Based on the updated user impression results, the interactive robot can adjust at least one of the following: topic advancement intensity, follow-up question frequency, prompting method, and topic switching strategy, thereby matching the subsequent interaction strategy with the user's actual communication state. For example, for cases where only temporary records are formed, the pace of advancement can be appropriately slowed down without immediately and completely avoiding the topic; for cases where a partial scene impression is formed, follow-up questions can be reduced or prompting methods adjusted in the corresponding scene; for cases where a formal user impression is formed, the advancement intensity around the target topic category can be reduced more broadly or the robot can actively switch to an alternative topic.

[0092] After the formal user impression update is completed, a counter-verification is performed based on the recovery reversal parameters in subsequent observation windows and the cross-scene diffusion parameters redefined based on avoidance events in subsequent observation windows. It should be noted that the robot further introduces a counter-verification and rollback mechanism after the formal user impression update. This is because even if a certain avoidance behavior has been identified as a globally stable avoidance state, it may still return to normal communication as subsequent interactions continue, or its cross-scene diffusion characteristics may gradually weaken.

[0093] Therefore, when the recovery reversal parameter indicates that the user has resumed normal communication, and / or the redefined cross-scenario diffusion parameter indicates that the number of different scenario labels covered by subsequent avoidance events is less than the predetermined number, the formal user impression update can be revoked and the user impression state before the update can be restored, so as to avoid the phased diffusion sample or subsequent recovery sample being retained as the formal user impression for a long time.

[0094] For example, a certain target topic category was written into the official user impressions because it met the conditions for a globally stable avoidance state.

[0095] During subsequent interactions, the interactive robot continues to acquire relevant interaction data within the observation window and recalculates the recovery reversal parameter and cross-scene diffusion parameter. If it is found that the user has resumed normal communication in subsequent rounds of interaction, the recovery reversal parameter reaches the recovery threshold, or the recalculated cross-scene diffusion parameter indicates that the different scene labels covered by the subsequent avoidance event are less than two, it means that the global stability condition on which the previous formal user impression was based no longer holds.

[0096] Accordingly, the counter-verification module outputs a rollback control signal to the impression update module, canceling the formal user impression update and restoring the user impression state before the update.

[0097] This example demonstrates that the formal user impression in this invention is not permanently fixed after being written once, but rather has the ability to be dynamically corrected based on subsequent real-world interaction states.

[0098] The technical solution implemented in this embodiment can directly use the updated results for topic advancement, follow-up questioning, prompting, and switching control of the interactive robot around the target topic category after the user impression update is completed, so that the interaction strategy matches the user's actual communication state. At the same time, by continuing to perform counter-verification and rollback when necessary after the formal user impression update, it is possible to avoid the phased diffusion samples or subsequent recovery samples from being retained as formal user impressions for a long time, thereby reducing the risk of long-term distortion of formal user impressions.

[0099] The present invention also provides a user impression updating device for an interactive robot, applied to the above-described user impression updating method for an interactive robot, comprising: The interaction data processing module is used to process historical and current interaction data of users around the target topic category to form the historical interaction baseline and topic avoidance segment related data corresponding to the target topic category.

[0100] The state identification module, connected to the interactive data processing module, is used to receive data related to the topic avoidance segment and output the identification result of the topic avoidance state based on the discrimination parameter set. The identification result includes one of the following: transient disturbance-type avoidance state, scene-stable avoidance state, and globally stable avoidance state.

[0101] The impression update module, connected to the state identification module, is used to receive the identification result and output the user impression update result. The user impression update result includes one of the following: temporary record result, local scene impression update result, and formal user impression update result.

[0102] The interaction control module, connected to the impression update module, is used to receive the user's impression update results and control the interactive robot's subsequent interaction strategies around the target topic category.

[0103] The rebuttal verification module is connected to the interactive data processing module and the impression update module respectively. It is used to receive subsequent observation window data and the impression update results of the actual user, and output a rollback control signal to the impression update module when the rollback conditions are met.

[0104] The discrimination parameter set includes avoidance intensity parameters, consecutive avoidance rounds parameters, scene locking parameters, cross-scene diffusion parameters, and recovery reversal parameters. The impression update module is also used to respond to rollback control signals to revoke the official user impression update results.

[0105] It should be noted that the interactive data processing module, state identification module, impression update module, interactive control module, and rebuttal verification module in this embodiment are not set up in isolation, but work together around the processing link of "historical interactive data input - discrimination parameter extraction - state label output - hierarchical impression update - subsequent interactive control - rebuttal verification rollback".

[0106] The interactive data processing module is used to form historical interaction baselines and topic avoidance segment related data, providing individualized data support for subsequent state identification; the state identification module is used to execute the state identification logic of sequential diversion, and output transient disturbance-type avoidance state, scene-stable avoidance state, or globally stable avoidance state; the impression update module is used to perform temporary recording, local scene impression update, or formal user impression update according to different states, and respond to the rollback control signal to revoke the formal user impression update result when necessary; the interactive control module is used to control the subsequent interactive strategy of the interactive robot around the target topic category according to the update result; the counter-verification module is used to re-verify the formal user impression update result based on the subsequent observation window data, and output the rollback control signal when the rollback condition is met.

[0107] The technical solution implemented in this embodiment can transform historical interaction data and current interaction data into historical interaction baseline and topic avoidance segment-related data. It then distinguishes between three types of avoidance states and performs temporary recording, local scene impression updates, or formal user impression updates. Finally, the updated results are applied to the subsequent control of the interactive robot around the target topic category, while dynamically correcting the formal user impression update results. This avoids the erroneous writing of short-term abnormal samples into formal user impressions, suppresses the misgeneralization of local scene features into global user features, and improves the stability of formal user impressions and the targeting of subsequent control.

[0108] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for updating user impressions for interactive robots, characterized in that, include: Obtain historical and current user interaction data around the target topic category, and establish a historical interaction baseline corresponding to the target topic category based on the historical interaction data; Based on the current interaction data and the historical interaction baseline, determine whether the current interaction data meets the topic avoidance triggering condition; wherein, the topic avoidance triggering condition can be judged by combining factors such as decreased topic continuity, decreased response length, and abnormal increase in refusal to answer or active topic shifting behavior. When the current interaction data satisfies the topic avoidance triggering condition, the corresponding topic avoidance segment is extracted. Based on the topic avoidance segment, a set of discrimination parameters is determined to characterize the topic avoidance state. The set of discrimination parameters includes avoidance intensity parameters, consecutive avoidance rounds parameters, scene locking parameters, cross-scene diffusion parameters, and recovery reversal parameters. The determination of the scene locking parameters, the cross-scene diffusion parameters, and the recovery reversal parameters includes: Extract historical avoidance events that are the same as the target topic category and have the same avoidance direction within the preset historical observation window, and classify and count them according to scene tags to determine the scene locking parameter and the cross-scene diffusion parameter; The scene tags include time period tags and task type tags; The scene locking parameter is determined by the ratio between the number of historical avoidance events with the same scene label as the topic avoidance segment and the number of historical avoidance events corresponding to other single scene labels; The cross-scenario diffusion parameter is determined by the number of different scenario labels covered by the historical avoidance event; In subsequent observation windows, it is determined round by round whether the conditions for restoring normal communication are met, and the restoration reversal parameter is determined based on the ratio between the number of interaction rounds that meet the conditions for restoring normal communication and the total number of rounds in the subsequent observation windows; The conditions for restoring normal communication are as follows: the probability of topic continuation in the corresponding interaction round is not lower than the probability of topic continuation in the past, and at least two of the following conditions are met: there is no refusal to answer or avoidance of expression, and there is no active topic shifting behavior. Based on the discrimination parameter set, the topic avoidance segment is identified in terms of state, and when the preset state determination condition is met, one of the following is output: transient disturbance type avoidance state label, scene stable type avoidance state label, or global stable type avoidance state label, and the topic avoidance segment is labeled as one of the following: transient disturbance type avoidance state, scene stable type avoidance state, or global stable type avoidance state. include: If the avoidance intensity parameter is greater than or equal to the avoidance trigger threshold, first determine whether the consecutive avoidance rounds parameter is less than or equal to the short duration threshold, and whether the recovery reversal parameter is greater than or equal to the recovery boundary threshold. When the consecutive avoidance round parameter is less than or equal to the short duration threshold and the recovery reversal parameter is greater than or equal to the recovery boundary threshold, output the transient disturbance-type avoidance state label corresponding to the topic avoidance segment and mark the topic avoidance segment as a transient disturbance-type avoidance state. Without outputting the transient disturbance-type avoidance status label, determine whether the continuous avoidance round parameter is greater than or equal to the continuous avoidance threshold, whether the scene locking parameter is greater than or equal to the scene concentration threshold, and whether the cross-scene diffusion parameter is less than the scene diffusion boundary threshold. When the continuous avoidance round parameter is greater than or equal to the continuous avoidance threshold, the scene locking parameter is greater than or equal to the scene concentration threshold, and the cross-scene diffusion parameter is less than the scene diffusion boundary threshold, the scene stable avoidance state label corresponding to the topic avoidance segment is output, and the topic avoidance segment is marked as a scene stable avoidance state. Without outputting the scene-stable avoidance state label, determine whether the continuous avoidance round parameter is greater than or equal to the continuous avoidance threshold, whether the recovery reversal parameter is less than the recovery boundary threshold, and whether the cross-scene diffusion parameter is greater than or equal to the scene diffusion boundary threshold. When the continuous avoidance round parameter is greater than or equal to the continuous avoidance threshold, the recovery reversal parameter is less than the recovery boundary threshold, and the cross-scene diffusion parameter is greater than or equal to the scene diffusion boundary threshold, the global stable avoidance state label corresponding to the topic avoidance segment is output, and the topic avoidance segment is marked as a global stable avoidance state. Wherein, the short duration threshold is less than the duration avoidance threshold; Based on the results of the state identification, perform temporary recording, partial scene impression update, or formal user impression update respectively; Based on the updated user impressions, control the subsequent interaction strategy of the interactive robot around the target topic category.

2. The user impression update method for interactive robots according to claim 1, characterized in that, The step of outputting the transient disturbance-type avoidance state label corresponding to the topic avoidance segment and marking the topic avoidance segment as a transient disturbance-type avoidance state includes: Starting from the beginning of the topic avoidance phase, extract continuous interaction data surrounding the target topic category within a preset short observation window; Based on the avoidance and maintenance status of each round of interaction within the preset short observation window, the number of interaction rounds that continuously meet the avoidance trigger threshold is counted, and the statistical result is determined as the continuous avoidance round parameter. Based on the recovery status of each round of interaction within the preset short observation window, the number of interaction rounds that meet the conditions for restoring normal communication is counted, and the recovery reversal parameter is determined according to the ratio between the number of interaction rounds that meet the conditions for restoring normal communication and the total number of rounds in the preset short observation window. When the avoidance intensity parameter is greater than or equal to the avoidance trigger threshold, the consecutive avoidance rounds parameter is less than or equal to the short duration threshold, and the recovery reversal parameter is greater than or equal to the recovery boundary threshold, the transient disturbance-type avoidance state label corresponding to the topic avoidance segment is output, and the topic avoidance segment is marked as a transient disturbance-type avoidance state.

3. The user impression update method for interactive robots according to claim 1, characterized in that, The step of outputting the scene-stable avoidance state label corresponding to the topic avoidance segment and marking the topic avoidance segment as a scene-stable avoidance state includes: Based on the scene tags corresponding to the topic avoidance segment, extract historical avoidance events that are the same as the target topic category and have the same avoidance direction within the preset historical observation window; The number of historical avoidance events with the same scene tag as the topic avoidance segment is counted, and the number of historical avoidance events corresponding to other single scene tags in the historical avoidance events is counted. The scene locking parameter is determined based on the ratio between the number of historical avoidance events with the same scene tag as the topic avoidance segment and the number of historical avoidance events corresponding to other single scene tags. The number of different scene tags covered by the historical avoidance events is counted, and the cross-scene diffusion parameter is determined based on the number of different scene tags; When the avoidance intensity parameter is greater than or equal to the avoidance trigger threshold, the consecutive avoidance rounds parameter is greater than or equal to the continuous avoidance threshold, the scene locking parameter is greater than or equal to the scene concentration threshold, and the cross-scene diffusion parameter is less than the scene diffusion boundary threshold, the scene stable avoidance state label corresponding to the topic avoidance segment is output, and the topic avoidance segment is marked as a scene stable avoidance state.

4. The user impression update method for interactive robots according to claim 1, characterized in that, The step of outputting the globally stable avoidance state label corresponding to the topic avoidance segment and marking the topic avoidance segment as a globally stable avoidance state includes: Extract historical avoidance events that are of the same category as the target topic and have the same avoidance direction from the preset historical observation window; The number of different scene tags covered by the historical avoidance events is counted, and the cross-scene diffusion parameter is determined based on the number of different scene tags; In the subsequent observation window after the topic avoidance segment, the number of interaction rounds that meet the conditions for resuming normal communication is counted, and the restoration reversal parameter is determined based on the ratio between the number of interaction rounds that meet the conditions for resuming normal communication and the total number of rounds in the subsequent observation window. When the avoidance intensity parameter is greater than or equal to the avoidance trigger threshold, the consecutive avoidance rounds parameter is greater than or equal to the continuous avoidance threshold, the cross-scene diffusion parameter is greater than or equal to the scene diffusion boundary threshold, and the recovery reversal parameter is less than the recovery boundary threshold, the global stable avoidance state label corresponding to the topic avoidance segment is output, and the topic avoidance segment is marked as a global stable avoidance state.

5. The user impression update method for an interactive robot according to any one of claims 2-4, characterized in that, The determination of the avoidance intensity parameter and the consecutive avoidance rounds parameter includes: Based on the historical interaction data, extract two or more of the following: historical topic continuation probability, historical average response length, historical refusal or avoidance expression frequency, and historical active topic shifting frequency corresponding to the target topic category, as the baseline statistical items of the historical interaction baseline. Based on the current interaction data, extract the current topic continuation probability, current average response length, current frequency of refusal to answer or avoidance of expression, and current frequency of actively changing topics; The avoidance intensity parameter is determined based on the difference between the current topic continuation probability and the historical topic continuation probability, the difference between the current average response length and the historical average response length, the difference between the current frequency of refusal to answer or avoidance and its corresponding baseline statistical upper limit, and the difference between the current frequency of actively changing topics and its corresponding baseline statistical upper limit. When the probability of the current topic continuing is lower than the probability of the historical topic continuing, the current average reply length is lower than the historical average reply length, and at least one of the current frequency of refusing to answer or avoiding expression and the current frequency of actively changing topics is higher than the corresponding baseline statistical upper limit, the avoidance intensity parameter is determined to be greater than or equal to the avoidance trigger threshold. When the current interaction data meets the topic avoidance triggering condition, the current interaction data and its subsequent several rounds of interaction data are extracted to form the topic avoidance segment; Each round of interaction in the topic avoidance segment is judged to determine whether the avoidance intensity parameter is greater than or equal to the avoidance trigger threshold, and the number of consecutive interaction rounds that meet the avoidance trigger threshold is accumulated to determine the consecutive avoidance round parameter.

6. The user impression update method for an interactive robot according to any one of claims 2-4, characterized in that, Based on the marked state of the topic avoidance segment, perform temporary recording, partial scene impression update, or formal user impression update respectively, including: When the topic avoidance segment is marked as the transient disturbance-type avoidance state, only temporary records are generated for the topic avoidance segment, and the formal user impression is not updated. When the topic avoidance segment is marked as the stable avoidance state of the scene, the user's avoidance information of the target topic category is bound and stored with the scene tag corresponding to the topic avoidance segment in order to update the local scene impression, but not the formal user impression. When the topic avoidance segment is marked as the globally stable avoidance state, information about the user's stable avoidance tendency towards the target topic category is written into the formal user impression.

7. The user impression update method for an interactive robot according to claim 6, characterized in that, Based on the updated user impression results, the interactive robot is controlled to control at least one of the following: topic advancement intensity, follow-up question frequency, prompting method, and topic switching strategy around the target topic category. After the formal user impression update is completed, the counter-evidence verification is performed based on the recovery reversal parameters in the subsequent observation window and the cross-scenario diffusion parameters re-determined based on the avoidance events in the subsequent observation window. When the recovery reversal parameter indicates that the user has met the conditions for resuming normal communication and / or the redefined cross-scenario diffusion parameter indicates that there are fewer than two different scenario labels covered by subsequent avoidance events, the formal user impression update is revoked and the user impression state before the update is restored.

8. A user impression updating device for an interactive robot, applied to the user impression updating method for an interactive robot according to any one of claims 1-7, characterized in that, include: The interaction data processing module is used to process the user's historical interaction data and current interaction data around the target topic category, so as to form the historical interaction baseline and topic avoidance segment related data corresponding to the target topic category; The state identification module, connected to the interactive data processing module, is used to receive data related to the topic avoidance segment and output the identification result of the topic avoidance state based on the discrimination parameter set. The identification result includes one of the following: transient disturbance type avoidance state, scene stable type avoidance state, and global stable type avoidance state. An impression update module, connected to the state identification module, is used to receive the identification result and output the user impression update result, wherein the user impression update result includes one of temporary record result, local scene impression update result and formal user impression update result; An interactive control module, connected to the impression update module, is used to receive the user impression update result and control the interactive robot's subsequent interaction strategy around the target topic category; The counter-evidence verification module is connected to the interactive data processing module and the impression update module respectively. It is used to receive subsequent observation window data and the official user impression update result, and output a rollback control signal to the impression update module when the rollback condition is met. The discrimination parameter set includes avoidance intensity parameter, consecutive avoidance rounds parameter, scene locking parameter, cross-scene diffusion parameter and recovery reversal parameter. The impression update module is also used to revoke the formal user impression update result in response to the rollback control signal.

Images