Intelligent interaction guiding method and system based on user intention recognition

Abstract

The invention relates to the technical field of computers, and provides an intelligent interaction guiding method and system based on user intention recognition. According to the embodiment of the invention, deep and comprehensive user interaction analysis is realized; the user demand dynamic state can be accurately captured by acquiring the interaction data set in real time; and context association features and interaction behavior features generated by multi-modal feature extraction provide a rich and meticulous information basis for intention analysis. On the basis, a preset intention recognition rule base is used for carrying out joint intention analysis processing, an intention label set and user portrait features are generated, accurate insight of the user intention is achieved, and then an interaction path guiding strategy is generated and pushed to a target user terminal in real time to update an interaction interface. Therefore, the method and the system can closely meet user requirements, provide personalized operation guidance, remarkably improve man-machine interaction efficiency and user operation accuracy, and enhance the intelligent service level of an intelligent factory.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of computer, and particularly relates to an intelligent interaction guiding method and system based on user intention recognition. BACKGROUND

[0002] With the intelligent development of manufacturing industry, the human-computer interaction technology of smart factory becomes the key to improve production efficiency and quality. However, the traditional human-computer interaction is mainly based on simple instruction input and feedback mode, and the operator needs to operate according to the fixed process, lacking flexibility and intelligence.

[0003] On the one hand, the existing human-computer interaction technology is single in obtaining and analyzing user interaction data, only focusing on the current operation and ignoring the influence of historical behavior on user intention, which is difficult to fully understand user demand. On the other hand, the existing intention recognition technology and interaction guiding technology lack depth and individualization, and cannot provide accurate guiding strategies according to the characteristics of different users and real-time scenes, resulting in low efficiency of human-computer interaction, and the operator is prone to operation errors, affecting the smooth operation of production process. SUMMARY

[0004] The present application provides an intelligent interaction guiding method and system based on user intention recognition, to improve the efficiency of human-computer interaction and the accuracy of user operation, and enhance the intelligent service level of smart factory.

[0005] In a first aspect, the present application provides an intelligent interaction guiding method based on user intention recognition, applied to an intelligent interaction guiding system, the method comprising: obtaining a set of interaction data of a target user in real time, the set of interaction data comprising a current interaction text sequence and a historical interaction behavior sequence; performing multi-modal feature extraction processing on the set of interaction data to generate context association features of the current interaction text sequence and interaction behavior features of the historical interaction behavior sequence; performing joint intention analysis processing on the context association features and the interaction behavior features according to a preset intention recognition rule library to generate a set of intention labels of the current interaction text sequence and user portrait features; based on the set of intention labels and the user portrait features, generating a set of interaction path guiding strategies, and pushing the set of interaction path guiding strategies to a target user terminal in real time to realize interaction interface update processing of the target user terminal.

[0006] In a second aspect, the present application provides an intelligent interaction guiding system, comprising a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0007] In a third aspect, the embodiments of the present application provide a computer readable storage medium comprising a computer program, when the computer program is run on an intelligent interactive guidance system, the computer program is configured to cause the intelligent interactive guidance system to perform the steps of the method described above.

[0008] In the embodiments of the present application, the real-time acquisition of the interactive data set can accurately capture the dynamic user demand; the context-related features and the interactive behavior features generated by the multi-modal feature extraction provide a rich and detailed information basis for intent analysis. Based on this, the preset intent recognition rule library is used for joint intent analysis processing to generate an intent label set and a user portrait feature, accurately understand the user intent, and then generate an interactive path guidance strategy and push it to the target user terminal in real time to update the interactive interface. In this way, the user demand can be closely matched, personalized operation guidance can be provided, the human-computer interaction efficiency and the user operation accuracy can be significantly improved, and the intelligent service level of the intelligent factory can be enhanced.

[0009] To sum up, the embodiments of the present application can accurately identify the user intent and provide personalized interactive path guidance by real-time acquisition of multi-dimensional interactive data, multi-modal feature extraction and joint intent analysis, effectively solving the deficiencies of the prior art in comprehensively understanding user demand and providing accurate guidance, and greatly improving the intelligence and efficiency of human-computer interaction in the smart factory. BRIEF DESCRIPTION OF DRAWINGS

[0010] Figure 1 A flowchart of an intelligent interactive guidance method based on user intent recognition provided by the embodiments of the present application.

[0011] Figure 2 A structural diagram of an intelligent interactive guidance system provided by the embodiments of the present application. DETAILED DESCRIPTION

[0012] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the technical scheme of the present application, rather than all the embodiments. Based on the embodiments described in the present application document, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the technical scheme of the present application.

[0013] Referring to Figure 1 It is an intelligent interactive guidance method based on user intent recognition provided by the embodiments of the present application, which can be applied to an intelligent interactive guidance system, and the specific process is as shown in steps 110-140.

[0014] Step 110: Real-time acquisition of the interaction data set of the target user, the interaction data set including a current interaction text sequence and a historical interaction behavior sequence.

[0015] In the application scenario of the smart factory, the system uses a data acquisition interface based on the OPCUA (Open Platform Communications Unified Architecture) protocol to acquire the interaction data set of the target user (covering the operating personnel, various devices, etc. in the factory) in real time. The OPCUA protocol has high openness, reliability and security, and can effectively realize data interaction and sharing between devices and systems of different manufacturers, which is very suitable for the complex and diverse device environment of the smart factory.

[0016] For example, the current interaction text sequence can be the text content such as "The running status of device M seems not normal, the value of parameter P fluctuates greatly, how to handle it" sent by the operating personnel through the terminal. Among them, "device M" clearly indicates the object, "the running status is not normal" describes the overall situation, "the value of parameter P fluctuates greatly" gives the specific problem point, and "how to handle it" expresses the intention of seeking a solution.

[0017] The historical interaction behavior sequence is a collection of a series of interaction operation records of the operating personnel and device M in the past. For example, at time t1, the operating personnel performed a calibration operation on device M, and the operation record details the specific parameter settings of the calibration (such as parameter Q set to q1), the calibration duration t1-t0 (t0 is the calibration start time); at time t2, device M triggered an alarm, and the alarm record contains the alarm type (such as parameter exception alarm), the alarm occurrence time t2, and part of the running parameters of device M at that time (such as parameter R is r1) and other information. The system relies on the data acquisition interface of the OPCUA protocol to accurately collect and integrate these real-time generated interaction data to form a complete interaction data set.

[0018] Step 120: Multi-modal feature extraction processing is performed on the interaction data set to generate context-related features of the current interaction text sequence and interaction behavior features of the historical interaction behavior sequence.

[0019] In the embodiments of the present application, in the smart factory environment, corresponding algorithms and models are used to deeply mine key features for the current interaction text sequence and the historical interaction behavior sequence. At the same time, to ensure the accuracy and effectiveness of the multi-modal features in subsequent processing, mapping, normalization, standardization and dimension adjustment operations are performed on these features to ensure dimensional consistency.

[0020] In an implementation manner, the multi-modal feature extraction processing on the interaction data set generates the context association feature of the current interaction text sequence and the interaction behavior feature of the historical interaction behavior sequence, and includes:

[0021] Step 121: performing semantic word segmentation processing on the current interaction text sequence to obtain a plurality of text units, and extracting the syntactic structure feature of each text unit.

[0022] In the intelligent factory scenario, taking the current interaction text sequence "The running status of device M seems not normal, the value of parameter P fluctuates greatly, how to handle it" sent by the operator as an example. The system processes the text by using a semantic word segmentation model based on deep learning. The model is trained on a large amount of text data in the intelligent factory field, and can accurately identify the lexical boundary and semantic unit in the text. After processing, a plurality of text units such as "device M", "running status", "seems", "not normal", "parameter P", "value", "fluctuates greatly", "how", and "handle" are obtained.

[0023] Then, for each text unit, the system extracts its syntactic structure feature by using a syntactic structure analysis tool based on dependency syntax analysis. For example, "device M" is a noun, which acts as the subject in the sentence and is the object of action and description; "running status" is a noun phrase, which is used to describe the status of the device; "seems" is an adverb, which expresses an uncertain tone; "not normal" is a predicate, which describes "running status"; "parameter P" is a noun, which is a specific index related to the operation of the device; "value" is a noun, which is associated with "parameter P"; "fluctuates greatly" is a description of the state of "value"; "how" and "handle" are interrogative adverbs, which lead to the inquiry intention; "handle" is a verb, which is the core action expressed in the sentence.

[0024] Step 122: based on the syntactic structure feature, calling a pre-trained context encoding model to perform cross-sentence association analysis to generate a context association feature containing a semantic coherence index, the semantic coherence index representing the semantic association strength between the current interaction text and the historical interaction text.

[0025] After obtaining the syntactic structure feature of each text unit, the system calls a context encoding model pre-trained based on the BERT (Bidirectional Encoder Representations from Transformers) architecture. The model is pre-trained on a large amount of text data in the intelligent factory field, including device operation manuals, fault reports, maintenance records, etc., so that it can deeply understand the semantics of industrial field text.

[0026] The context encoding model performs cross-sentence correlation analysis on the current interaction text sequence and the historical interaction text. For example, the historical interaction text may have a record of parameter changes of device M after a certain maintenance, and the handling measures when the parameter fluctuation occurred before. The context encoding model calculates the semantic similarity between them by matching and analyzing the key information such as "device M", "parameter P", "value fluctuation" in the current text and the related content in the historical text.

[0027] In the calculation process, to ensure dimensional consistency, the features are normalized. For example, the text feature vector is mapped to a standard normal distribution space with a mean of 0 and a variance of 1 through the formula (x-μ) / σ, where x is the original feature value, μ is the mean of the feature dimension, and σ is the standard deviation. Then the semantic coherence index is calculated through the existing algorithm. The index represents the semantic correlation strength between the current interaction text and the historical interaction text through a numerical range (for example, a-b), and the higher the value, the stronger the correlation. For example, if the problem of parameter fluctuation of device M in the current text is similar to the description of a certain abnormal parameter situation of device M in the historical text, the semantic coherence index may reach c (a<c<b), indicating that there is a strong semantic correlation between them. Thus, the context correlation feature containing the semantic coherence index is generated, providing more rich information for subsequent intent recognition.

[0028] Step 123: performing time series segmentation processing on the historical interaction behavior sequence to generate a plurality of interaction behavior segments.

[0029] In the intelligent factory scenario, the system performs time series segmentation processing on the historical interaction behavior sequence, aiming to divide the continuous historical interaction behavior into a plurality of meaningful interaction behavior segments according to certain rules, so as to better analyze and understand the interaction patterns in different time periods.

[0030] In an implementation manner, the performing time series segmentation processing on the historical interaction behavior sequence to generate a plurality of interaction behavior segments comprises:

[0031] Step 1231: obtaining a time stamp data sequence of each interaction event in the historical interaction behavior sequence, and calculating a time interval set between adjacent two interaction events.

[0032] In the embodiment of the application, the system first extracts the time stamp data sequence of each interaction event from the database of the historical interaction behavior sequence. For example, the time stamps of each interaction event are t1, t2, t3, …, and then the time intervals between adjacent two interaction events are obtained by calculating t2-t1=d1, t3-t2=d2, …, to generate a set containing all adjacent interaction event time intervals.

[0033] Step 1232: performing a dynamic segmentation operation on the historical interaction behavior sequence according to a comparison result of each element in the time interval set and a preset interval threshold, and inserting a segmentation point between adjacent interaction events corresponding to a target time interval when it is detected that the target time interval exceeds the preset interval threshold, so as to segment the historical interaction behavior sequence into a plurality of time-continuous interaction behavior segments.

[0034] In the embodiment of the present application, the system sets a preset interval threshold D in advance. Then, each element in the time interval set is compared with the preset interval threshold. When it is detected that one of the time intervals exceeds D, for example, a time interval d4 (d4>D), the system inserts a segmentation point between the two adjacent interaction events corresponding to the time interval d4. In the operation record of the device M, the time interval d4 between the device inspection operation and the next device parameter adjustment operation exceeds the preset threshold D, and therefore the system inserts a segmentation point between the time points corresponding to the two operations. Through the dynamic segmentation operation, the historical interaction behavior sequence is segmented into a plurality of time-continuous interaction behavior segments. These segments can clearly divide the interaction behaviors in different time periods, for example, one segment can include all the related operations during a device maintenance, and another segment can include various monitoring and adjustment operations of the device during a normal operation period.

[0035] Step 1233: performing a behavior density analysis based on the segmented interaction behavior segment set to generate an interaction operation frequency feature, wherein the interaction operation frequency feature includes a statistical value of the number of interaction operations per unit time in each interaction behavior segment and a distribution proportion vector of different operation types.

[0036] For each segmented interaction behavior segment, the system performs a behavior density analysis. Taking one of the interaction behavior segments as an example, the segment records the interaction operation of the device M in a time period T. In the time period T, the device M performs n1 times of parameter reading operations, n2 times of device state inspection operations, n3 times of maintenance operations and other types of operations.

[0037] First, the number of interaction operations per unit time in each interaction behavior segment is calculated, and the value is obtained by the formula N=(n1+n2+n3+…) / T, where N represents the number of interaction operations per unit time. Then, the distribution proportion vector of different operation types is calculated. For example, the distribution proportion of parameter reading operation is p1=n1 / (n1+n2+n3+…), the distribution proportion of device state checking operation is p2=n2 / (n1+n2+n3+…), the distribution proportion of maintenance operation is p3=n3 / (n1+n2+n3+…), and so on, forming a distribution proportion vector [p1, p2, p3, …]. Thus, the interaction operation frequency feature including the number of interaction operations per unit time and the distribution proportion vector of different operation types is generated, providing detailed behavior pattern information for subsequent analysis.

[0038] Step 124: For each interaction behavior segment, perform behavior pattern recognition processing to extract interaction behavior features, including at least one of interaction operation frequency feature, interaction path switching feature, and interface element dwell time feature.

[0039] It can be understood that for each segmented interaction behavior segment, the system further performs behavior pattern recognition processing. Taking an interaction behavior segment as an example, in addition to the interaction operation frequency feature calculated above, the interaction path switching feature is also analyzed. For example, the operation of device M in this segment involves multiple functional modules, the number of times of switching from functional module A to functional module B is m1, the number of times of switching from functional module B to functional module C is m2, and so on. By counting these switching times and the order relationship of switching, the interaction path switching feature can be obtained. For example, a switching matrix is formed, where the element a ij in the matrix represents the number of times of switching from functional module i to functional module j.

[0040] For the interface element dwell time feature, the operator interacts with the device M through the operation terminal, and the operation terminal interface has multiple elements, such as parameter display area, operation button area, etc. The dwell time of the operator on each interface element is recorded, such as the dwell time in the parameter display area t a , the dwell time in the operation button area t b , etc. By analyzing the distribution and variation of these dwell times, the interface element dwell time feature is obtained. For example, the proportion of the dwell time of each interface element in the total interaction time can be calculated to form a proportion vector [q a , q b , …], which represents the interface element dwell time feature. By integrating these interaction operation frequency features, interaction path switching features, and interface element dwell time features, the interaction behavior features of each interaction behavior segment are comprehensively extracted, providing more abundant and accurate information for subsequent intent analysis.

[0041] Step 130: According to the preset intent recognition rule base, the context association features and the interaction behavior features are jointly processed for intent analysis to generate an intent label set and a user portrait feature of the current interaction text sequence.

[0042] In the intelligent factory, the preset intent recognition rule base is constructed based on a large amount of historical interaction data and expert knowledge. The system inputs the previously extracted context association features and interaction behavior features into the intent recognition process. For example, combined with the context association feature of "device M parameter P value fluctuation is large" mentioned in the current interaction text, and the interaction behavior feature of frequent parameter adjustment operation of device M in the recent historical interaction behavior fragment. The rule base may have the following rule: when a certain parameter of a device fluctuates and the device has frequent parameter adjustment operation in the recent period, the intent can be "device parameter abnormal processing". Through matching and reasoning of these features and rules in the rule base, an intent label set of the current interaction text sequence is generated, such as {"device parameter abnormal processing", "seek technical support"} and the like.

[0043] At the same time, according to the interaction behavior features and the historical interaction data, the user portrait feature is generated. For example, if the operator often refers to the corresponding technical document when handling the device problem, or often consults the technical team, these information will be included in the user portrait feature. The user portrait feature can include the skill level of the operator (such as the familiarity with device M, which is evaluated by the accuracy and efficiency of the operation), knowledge preference (such as the type of technical material often consulted), and other multi-dimensional information, which is represented in the form of a multi-dimensional vector, such as [skill level value, knowledge preference vector, …].

[0044] In an optional embodiment, the construction process of the preset intent recognition rule base includes:

[0045] Step 131: Collect a multi-field user interaction sample set, and extract a key intent expression feature vector and an associated behavior pattern feature matrix in each user interaction sample.

[0046] In the intelligent factory scenario, a multi-field user interaction sample set from different production links, different device types, and different operators is collected. For example, interaction data from device maintenance, device monitoring, production scheduling, and other fields is collected. For each user interaction sample, such as an interaction record of handling a device fault, a key intent expression feature vector is extracted through natural language processing technology and data analysis method. For example, keywords related to fault type and processing demand are extracted from the interaction text and converted into vector form. For example, the key intent expression can be "device fault", "emergency repair", and the like, which is converted into a feature vector through a word vector model.

[0047] Meanwhile, an associated behavior pattern feature matrix is extracted. For example, the sequence and time interval of a series of operations performed by an operator during the device fault handling process are recorded to form a behavior pattern matrix. The elements in the matrix can represent the association between different operations or the time sequence of operation occurrence, etc. For example, a ij represents the association strength or time sequence between operation i and operation j.

[0048] Step 132: Based on the associated behavior pattern feature matrix, the key intent expression feature vector is labeled with multi-level intent categories to generate a labeled intent training sample set.

[0049] According to the operation logic and flow reflected by the associated behavior pattern feature matrix, the key intent expression feature vector is labeled with multi-level intent categories. For example, for the interactive sample of device fault handling, according to the sequence and type of operation behavior, if the fault diagnosis operation is performed first and then the maintenance resource is requested, the key intent expression feature vector can be labeled as “fault diagnosis” and “maintenance resource request” under the high-level intent category of “device fault handling”. Thus, each key intent expression feature vector is labeled with the corresponding intent category label to generate a labeled intent training sample set, which will be used for subsequent model training.

[0050] Step 133: The intent training sample set is jointly trained using a multi-task learning framework to generate a hybrid model containing an intent recognition main model and an auxiliary verification model, which is used to calculate the confidence score of the output result of the intent recognition main model.

[0051] Optionally, the labeled intent training sample set is input into the training process using a multi-task learning framework. The multi-task learning framework can simultaneously learn multiple related tasks to improve the generalization ability and accuracy of the model. During the training process, the intent recognition main model learns to identify intent categories from the key intent expression feature vector and the associated behavior pattern feature matrix, while the auxiliary verification model learns to evaluate the reliability of the output result of the main model. For example, the main model outputs an intent category “device maintenance”, and the auxiliary verification model calculates the confidence score of the output result, such as 0.8, indicating that the credibility of the main model determining the “device maintenance” intent is 80%. Through joint training, the parameters of the two models are continuously optimized to finally generate a hybrid model containing an intent recognition main model and an auxiliary verification model.

[0052] Step 134: When the auxiliary verification model detects that the confidence score of the output of the intent recognition main model is lower than a preset verification threshold, a review mechanism is triggered and a rule supplement instruction containing a contradiction feature identifier is generated.

[0053] Optionally, a verification threshold is preset, for example, 0.6. When the auxiliary verification model detects that the confidence score of the intent recognition main model output is lower than 0.6, it indicates that the judgment result of the main model may be inaccurate, at which time the review mechanism is triggered. The review mechanism analyzes the difference between the features input by the main model and the output result, and finds out the contradictory features that may cause inaccurate judgment. For example, the main model judges the intent to be "device upgrade" according to certain features, but the auxiliary verification model finds that there is no operation record related to device upgrade in the historical interaction data, which produces a contradiction. The system generates rule supplement instructions containing contradictory feature identifiers, such as indicating that "when judging the device upgrade intent, there is no relevant upgrade operation record in the historical interaction data".

[0054] Step 135: According to the contradictory feature identifier carried in the rule supplement instruction, an intent matching condition expression and an associated behavior pattern feature matching rule in the intent recognition rule library are constructed, and there is a logical mapping relationship between the intent matching condition expression and the behavior pattern feature corresponding to the contradictory feature identifier.

[0055] Optionally, according to the contradictory feature identifier in the rule supplement instruction, a new intent matching condition expression and an associated behavior pattern feature matching rule are constructed in the intent recognition rule library. For example, for the above-mentioned "when judging the device upgrade intent, there is no relevant upgrade operation record in the historical interaction data" contradictory feature, an intent matching condition expression is added in the rule library: if there is no device upgrade related operation record in the historical interaction data, the credibility of the current intent judgment as "device upgrade" is reduced. At the same time, the associated behavior pattern feature matching rule can stipulate that when the device upgrade related keywords appear but there is no corresponding operation record, the relevant information should be further verified. In this way, the intent recognition rule library is constantly improved, and the accuracy of intent recognition is improved.

[0056] Step 140: Based on the intent label set and the user portrait feature, an interaction path guidance strategy set is generated, and the interaction path guidance strategy set is pushed to the target user terminal in real time to realize the interaction interface update processing of the target user terminal.

[0057] In the intelligent factory, based on the generated intent label set and user portrait feature, the system starts to formulate an interaction path guidance strategy. For example, when the intent label set is {"device parameter abnormality processing", "seeking technical support"} and the user portrait feature shows that the operator has certain experience in adjusting the parameters of device M but needs more guidance on some complex problems.

[0058] In one implementation manner, the generating of the interaction path guidance strategy set based on the intent label set and the user portrait feature comprises:

[0059] Step 141: According to the priority of the dominant intention category identified in the intention label set, an initial basic guidance path configuration is matched from the preset interaction path template library.

[0060] In the embodiment of the application, the priority of the dominant intention in the intention label set is determined first, and in this example, "device parameter abnormality processing" is the highest priority. The preset interaction path template library stores a plurality of basic guidance path configurations for different intentions. The system matches in the template library according to the dominant intention category "device parameter abnormality processing". For example, the template library may have a basic guidance path configuration for device parameter abnormality processing, which first guides the operator to check the parameter monitoring system of the device M, obtains the detailed fluctuation data of the parameter P, and records the running state information of the current device.

[0061] Step 142: The user portrait features are input into the preset path optimization model to perform the following personalized adjustment operations: analyzing the historical path adoption rate matrix and interface element preference vector in the user portrait features; dynamically adjusting the weight distribution parameters of the branch nodes in the basic guidance path configuration based on the adoption probability data of different path nodes in the historical path adoption rate matrix; and replacing the layout arrangement of the corresponding interface elements in the basic guidance path configuration according to the preference score of each type of interface element in the interface element preference vector.

[0062] Optionally, the user portrait features are input into the preset path optimization model. The historical path adoption rate matrix in the user portrait features records the selection probability of the operator for different guidance path nodes when handling similar problems in the past. For example, the historical path adoption rate matrix shows that when handling device parameter abnormality problems, the operator has a high probability of first selecting to view the real-time monitoring data of the device. Based on this data, in the basic guidance path configuration, the weight distribution parameters of the branch node guiding the operator to view the real-time monitoring data of the device are increased, so that the possibility of this operation path being preferentially recommended in the new guidance strategy is increased.

[0063] At the same time, the interface element preference vector reflects the preference degree of the operator for different interface elements on the operation terminal. For example, the interface element preference vector indicates that the operator prefers a simple and clear parameter display interface. Then according to the preference score, the originally complex parameter display interface layout in the basic guidance path configuration is replaced by a more simple layout to improve the efficiency and operation experience of the operator in obtaining information. Through these personalized adjustment operations, the guidance path is more in line with the behavior habits and needs of the user.

[0064] Step 143: Generate multiple personalized adjusted interaction path guidance strategy instances by combining different weight distribution parameter adjustment schemes and layout arrangement adjustment schemes, and sort the multiple interaction path guidance strategy instances according to preset strategy priority scores to form the interaction path guidance strategy set.

[0065] Optionally, the system combines different weight distribution parameter adjustment schemes and layout arrangement adjustment schemes. For example, there are three adjustment schemes for weight distribution parameters and two adjustment schemes for layout arrangement. By combining them, 3x2=6 personalized adjusted interaction path guidance strategy instances can be generated. Each instance is optimized for different user portrait features and intentions.

[0066] Then, according to the preset strategy priority score rule, the multiple interaction path guidance strategy instances are sorted. The preset strategy priority score rule may consider factors such as the efficiency, accuracy, and acceptance of the operation personnel of the guidance path. For example, a guidance strategy can quickly solve the device parameter abnormal problem and at the same time meet the operation habit of the operation personnel, so its priority score will be higher. Through sorting, the strategies with high scores are placed in front to form the interaction path guidance strategy set, ensuring that the system can preferentially recommend the most suitable guidance strategy for the user.

[0067] In an implementation manner, the real-time pushing of the interaction path guidance strategy set to the target user terminal to realize the interaction interface update processing of the target user terminal comprises:

[0068] Step 144: Obtain a device type parameter set of the target user terminal, which includes a screen resolution parameter and an input method support identifier.

[0069] The system obtains the device type parameter set through the communication connection established with the target user terminal. For the terminal device used by the operation personnel, the screen resolution parameter can be a×b pixels (such as the common 1920x1080 pixels), and the input method support identifier can indicate that the terminal supports touch input, keyboard input, and other input methods. These parameter information is crucial for subsequent interface adaptation processing.

[0070] Step 145: Perform interface adaptation processing based on the device type parameter set: parse the screen resolution parameter to determine the display area size, dynamically adjust the interface element size parameter and layout density coefficient in the interaction path guidance strategy set; according to the input method support identifier, match the corresponding interaction control generation rule set, and reconstruct the touch response logic of the interface element.

[0071] Wherein, according to the obtained screen resolution parameter a x b pixels, the actual size of the display area is calculated. Based on this, the interface element size parameter in the interactive path guidance strategy set is dynamically adjusted. For example, if the screen resolution is low, in order to ensure that the interface elements can be clearly displayed, the size of the interface elements is appropriately reduced, and the layout density coefficient is adjusted accordingly, so that the interface layout is more compact and reasonable.

[0072] At the same time, according to the input mode support identifier, the corresponding interactive control generation rule set is matched. If the terminal supports touch input, the touch response logic of the interface elements is reconstructed. For example, for the operation button, the touch response area is increased to facilitate the touch operation of the operator; the feedback mechanism of the touch operation is optimized, and when the operator touches the button, visual or tactile feedback is given in time to improve the accuracy and fluency of the operation.

[0073] Step 146: Generate a device-adapted interactive interface configuration file and push it to the target user terminal.

[0074] Wherein, after the interface adaptation processing, the system integrates the adjusted interface element size parameter, layout density coefficient and new touch response logic and other information to generate a device-adapted interactive interface configuration file. The configuration file contains all the setting information about the interactive interface and is stored in a preset file format (such as XML format). Then, the configuration file is pushed to the target user terminal in real time through network communication. After the target user terminal receives the configuration file, the interactive interface is updated according to the setting information in the file, so that the interactive interface can better adapt to the characteristics of the terminal device.

[0075] Step 147: When an interface configuration file loading exception event is monitored, start an alternative interface rendering scheme to re-render the interface, and the exception event includes element rendering misplacement or control response failure.

[0076] Wherein, the system continuously monitors the interface configuration file loading of the target user terminal. When an element rendering misplacement, such as parameter display box position offset on the interface, or control response failure, such as operation button non-response, and other abnormal events are detected, the alternative interface rendering scheme is immediately started.

[0077] Preferably, the starting of the alternative interface rendering scheme to re-render the interface comprises:

[0078] Step 1471: Identify the conflict interface element identifier list and the abnormal layout parameter set in the current interactive interface configuration file.

[0079] In the embodiment of the present application, the system first analyzes the current interactive interface configuration file, and identifies the list of conflicting interface element identifiers that cause loading abnormalities through commonly used detection algorithms and rules. For example, if it is found that the layout parameters of one of the parameter display boxes do not match the screen resolution, causing misalignment in rendering, the identifier of the parameter display box will be included in the list of conflicting interface element identifiers. At the same time, a set of abnormal layout parameters, such as the incorrect position coordinates, size, and other parameters of the parameter display box, are collected.

[0080] Step 1472: According to the list of conflicting interface element identifiers, match the simplified interface template instance compatible with the device type parameter set from the preset alternative solution library.

[0081] In the embodiment of the present application, the preset alternative solution library stores a plurality of simplified interface template instances, each of which is designed for different device types and common interface problems. According to the list of conflicting interface element identifiers and the device type parameter set, matching is performed in the alternative solution library. For example, for a terminal device with a screen resolution of a x b pixels and touch input support, a simplified interface template instance that can solve the current parameter display box rendering problem is found. The template instance usually has a more concise layout and a more stable rendering method, which can avoid similar loading abnormality problems.

[0082] Step 1473: Map the core guide elements in the interactive path guide strategy set to the placeholder nodes of the simplified interface template instance, generate interface update instructions containing degraded rendering parameters, and re-push to the target user terminal.

[0083] In the embodiment of the present application, the core guide elements in the interactive path guide strategy set, such as operation step prompts and key information display, are mapped to the placeholder nodes of the simplified interface template instance. For example, there is a placeholder node in the simplified interface template instance for displaying important prompt information, and the key prompt information about device parameter abnormality handling in the interactive path guide strategy is mapped to the node. At the same time, according to the requirements of the simplified interface template, interface update instructions containing degraded rendering parameters are generated, which ensure that the interface can still be normally rendered under lower configuration or abnormal conditions. Finally, the interface update instructions are re-pushed to the target user terminal, so that the terminal re-renders the interface according to the new instructions.

[0084] Step 1474: Record the feature log of the interface configuration file loading abnormality event, and add the device compatibility data in the feature log to the verification rule library to optimize the template matching algorithm of the subsequent adaptation process.

[0085] In the embodiments of the present application, the system records the detailed feature logs of the interface configuration file loading exception event, including the time when the exception occurs, the specific information of the conflicting interface elements, the exception layout parameters, etc. From these logs, device compatibility data is extracted, for example, it is found that terminal devices with a certain combination of screen resolution and input method are prone to certain types of interface loading exceptions. These device compatibility data are added to the verification rule library, which is used to optimize the template matching algorithm of the subsequent adaptation process. By continuously accumulating and analyzing these data, the template matching algorithm can more accurately select the appropriate interface template, improve the success rate of interface adaptation, and reduce the occurrence of loading exception events.

[0086] In an optional embodiment, the updating process of the user portrait feature includes:

[0087] Step 210: Real-time monitoring of the interactive behavior change indicators triggered by the target user within a preset period, including the newly added interactive operation type code and the operation frequency fluctuation amplitude value.

[0088] Optionally, the system continuously monitors the interactive behavior of the target user within a preset period (for example, every hour). For example, in a smart factory, an operator may trigger a new interactive operation, such as using a new device diagnostic tool, and the system will assign a newly added interactive operation type code to this new operation. At the same time, the system calculates the fluctuation of the operation frequency, for example, the frequency of the device parameter reading operation changes from every 10 minutes to every 5 minutes within the past hour, and calculates the operation frequency fluctuation amplitude value. These newly added interactive operation type codes and operation frequency fluctuation amplitude values constitute the interactive behavior change indicators.

[0089] Step 220: Based on the interactive behavior change indicators, incrementally update the historical path adoption rate matrix in the user portrait feature: when a newly added interactive operation type code is detected, add a temporary weight coefficient to the corresponding interface element preference vector in the new dimension, and the temporary weight coefficient is dynamically adjusted according to the interaction frequency; when the operation frequency fluctuation amplitude value exceeds the preset fluctuation threshold, trigger the integrity verification process of the user portrait feature, and according to the missing dimension identifier determined by the integrity verification process, supplement the initial value of the corresponding portrait dimension feature.

[0090] Optionally, when a new interaction operation type code is detected, for example, a new operation code of a device diagnostic tool is added, a new dimension is added to the interface element preference vector of the user portrait feature to represent the preference for the new operation related interface element. At the same time, a temporary weight coefficient is added to this new dimension, and the initial value of the coefficient can be set to a small value, such as 0.1. As the frequency of the operator using the new operation increases, the temporary weight coefficient is dynamically adjusted. For example, each time the operation is used, the weight coefficient increases by 0.05, to reflect the gradual familiarity and preference level of the operator for the new operation.

[0091] When the operation frequency fluctuation amplitude value exceeds the preset fluctuation threshold (for example, the fluctuation amplitude exceeds 30%), the integrity verification process of the user portrait feature is triggered. The system checks whether there is a missing dimension related to the current operation frequency change in the user portrait feature. For example, if it is found that the frequency of the operator's device parameter adjustment operation has increased significantly, but there is no dimension related to the parameter adjustment skill level in the user portrait, then the "parameter adjustment skill level" is determined as the missing dimension identifier. Then, according to the preset rules or reference historical data, the initial value of the portrait dimension feature is supplemented. For example, according to the success rate of the operator's past handling of parameter adjustment problems, the initial value of the parameter adjustment skill level is set to a corresponding level (such as primary, intermediate, and advanced).

[0092] Step 230: generating a corrected user portrait feature based on the incrementally updated user portrait feature.

[0093] Optionally, the system integrates the incrementally updated parts to generate a corrected user portrait feature. For example, the updated historical path adoption rate matrix, the adjusted interface element preference vector, and the supplemented portrait dimension feature are fused together to form a user portrait feature that more accurately and comprehensively reflects the current interaction behavior and characteristics of the target user. This corrected user portrait feature will be used in subsequent intent recognition, guidance strategy generation, and other processes to improve the system's understanding and service capabilities for the user.

[0094] In an alternative embodiment, after the intent label set and the user portrait feature of the current interaction text sequence are generated, the method further comprises:

[0095] Step 310: inputting the intent label set into a preset conflict detection model to identify a set of logical contradiction pairs between different intent labels in the intent label set.

[0096] Optionally, the pre-set conflict detection model is constructed based on logical reasoning and machine learning algorithms. The generated set of intent labels, e.g., {“device parameter abnormal handling”, “device normal operation maintenance”}, is input into the conflict detection model. The model identifies possible pairs of logical contradictions among the intent labels through semantic analysis and logical relationship judgment of the intent labels. For example, “device parameter abnormal handling” implies that there is a problem with the device operation that needs to be solved, while “device normal operation maintenance” indicates maintenance operations under normal conditions of the device, and these two intents may have logical contradictions. The model determines the set of logical contradiction pairs by analyzing the relationship between the operations and states represented by the intent labels, such as {(“device parameter abnormal handling”, “device normal operation maintenance”)}.

[0097] Step 320: If a logical contradiction pair is detected in the set of logical contradiction pairs, extract the ambiguity expression segment index associated with the logical contradiction pair from the current interactive text sequence.

[0098] Optionally, if the conflict detection model detects a contradiction pair in the set of logical contradiction pairs, the system returns to the current interactive text sequence for analysis. For example, for the above contradiction pair, search for related expressions in the current interactive text. If the text mentions “the device parameter seems not quite right, but the device is currently running normally”, the system will extract the ambiguity expression segments associated with the contradiction pair in this text, such as “the device parameter seems not quite right” and “the device is currently running normally”, and determine their index positions in the text. These index position information will be used to generate clarification guide questions later.

[0099] Step 330: Based on the historical clarification record data in the user portrait features, generate a set of clarification guide questions for the ambiguity expression segments, and insert the set of clarification guide questions at the head of the priority queue of the interactive path guide strategy set.

[0100] Optionally, referring to the historical clarification record data in the user portrait features, these data record the user's feedback and handling methods in similar situations in the past. For example, the historical clarification record shows that when similar device state contradictory expressions occur, asking the user the question “Do you confirm that the current device running parameters are consistent with what you observed?” can effectively clarify the contradiction. Based on this, a set of clarification guide questions is generated for the currently extracted ambiguity expression segments. For example, {“Do you observe that the device parameter is abnormal, which affects the normal operation of the device?”, “Are you sure that the device is currently in a normal operation state?”}. Then, this set of clarification guide questions is inserted at the head of the priority queue of the interactive path guide strategy set, ensuring that the system prioritizes asking the user these questions to clarify the intent contradiction.

[0101] Step 340: receiving a feedback data stream of the target user for the set of clarification guide questions, correcting the set of intent labels according to the user selection identifier in the feedback data stream, and generating an updated set of intent labels.

[0102] Optionally, when the target user answers the set of clarification guide questions, the system receives a feedback data stream. For example, the user answers "the device is not running very well, and the parameter exception affects the running", and according to the user selection identifier (such as positive or negative answer) in the feedback data stream, the system corrects the set of intent labels. In this example, the set of intent labels is corrected to { "device parameter exception handling"}, and the "device normal running maintenance" intent label that is inconsistent with the user feedback is removed, generating an updated set of intent labels.

[0103] Step 350: regenerating the set of interactive path guide strategies based on the updated set of intent labels, and triggering the interactive interface of the target user terminal to perform an incremental update operation.

[0104] Optionally, according to the updated set of intent labels, the system regenerates the set of interactive path guide strategies. For example, for the intent label "device parameter exception handling", a new set of guide strategies is generated, including guiding the user to further check the parameter exception reason, providing related troubleshooting steps, etc. Then, the newly generated set of interactive path guide strategies is pushed to the target user terminal, triggering the terminal to perform an incremental update operation. The interactive interface updates the display content according to the new guide strategies, such as displaying new operation prompts, parameter monitoring interfaces, etc., to guide the user to more effectively handle the device problem.

[0105] In one non-limiting embodiment, the method further comprises:

[0106] Collecting interface interaction behavior data of the target user terminal, extracting a user response feature set containing behavior timing features and operation distribution features; generating a set of dynamic adaptation parameters of the interactive path based on the deviation calculation of the user response feature set and the preset strategy expected indicators; triggering a multi-dimensional priority reconstruction operation of the interactive path guide strategy according to the path node weight offset in the set of dynamic adaptation parameters; and synchronously calibrating the reconstructed interactive path guide strategy with the real-time interaction state of the user terminal.

[0107] In this embodiment, the system continuously collects interface interaction behavior data of the target user terminal, such as the click operation time of the operator on the terminal, the sliding operation trajectory, and other information. From these data, behavior timing features such as the order of operations, operation interval time, etc. are extracted; and operation distribution features such as the distribution of different operation types on the interface are extracted. These features constitute the user response feature set. The preset strategy expected index is the ideal user operation mode and interface interaction index set according to past experience and best practices. The deviation degree of the user response feature set and the preset strategy expected index is calculated. For example, if the preset strategy expects the operator to perform a key parameter check operation every 10 minutes, but the actual operator's operation interval time is 20 minutes, the deviation degree is calculated by a commonly used algorithm. Based on the calculation result of the deviation degree, a dynamic adaptation parameter set of the interaction path is generated, which is used to adjust the interaction path guidance strategy. For example, the dynamic adaptation parameter set contains path node weight offset. If it is found that the user's execution frequency of one of the operation steps is lower than expected, the weight of the step in the interaction path guidance strategy is correspondingly reduced; conversely, if one of the operation steps is frequently executed, the weight is increased. According to these weight offsets, a multi-dimensional priority reconstruction operation of the interaction path guidance strategy is triggered to rearrange the priority order of each step in the guidance strategy. Finally, the reconstructed interaction path guidance strategy is synchronized and calibrated with the real-time interaction state of the user terminal. For example, when the user is performing one of the operations, it is ensured that the guidance strategy can provide appropriate next-step guidance according to the user's current operation state, so that the guidance strategy is closely combined with the user's actual operation, improving the guidance effect and user experience.

[0108] In one non-limiting embodiment, the method further comprises:

[0109] Real-time monitoring of the interface rendering performance data of the target user terminal, identifying performance bottleneck indicators containing response delay features and resource occupation features; based on the association mapping relationship between the performance bottleneck indicators and the interaction path guidance strategy, generating hierarchical adjustment instructions of the interface rendering optimization strategy; according to the hierarchical adjustment instructions, asynchronously processing the dynamic element loading logic in the interactive interface, establishing a collaborative control model of rendering resource allocation and path guidance priority; and injecting the collaborative control model into the user terminal interaction engine through incremental updating.

[0110] In detail, the system monitors the interface rendering performance data of the target user terminal in real time. Through the performance monitoring tool deployed on the terminal device, the response delay characteristics such as the delay time of interface element display, the delay time of operation response, etc. are obtained; and the resource occupation characteristics such as the occupation of CPU, memory and other resources in the interface rendering process are obtained. These data constitute the performance bottleneck indicators. The correlation mapping relationship between the performance bottleneck indicators and the interaction path guidance strategy is analyzed. For example, it is found that when the guidance strategy contains more dynamic element loading, the response delay increases significantly, and the resource occupation also increases. Based on this relationship, the hierarchical adjustment instructions of the interface rendering optimization strategy are generated. These instructions formulate adjustment measures according to the severity and influence range of the performance bottleneck. For example, for a mild performance bottleneck, the instruction may require asynchronous processing of the loading logic of part of the dynamic elements, that is, let these dynamic elements load in the background thread, without affecting the running of the main thread, thereby reducing the interface response delay. For a moderate performance bottleneck, in addition to asynchronous processing, the allocation of rendering resources is also adjusted to prioritize the rendering of key interface elements and elements related to important steps in the guidance strategy. For a severe performance bottleneck, part of the interaction path guidance strategy may be simplified to reduce unnecessary dynamic element display. According to the hierarchical adjustment instructions, the dynamic element loading logic in the interactive interface is asynchronously processed. For example, the loading tasks of dynamic elements such as device state real-time update charts and operation prompt animations are allocated to background threads. In this way, when the user performs an operation, the main thread can respond to user input more quickly, improving the smoothness of the interface.

[0111] At the same time, a collaborative control model of rendering resource allocation and path guidance priority is established. The model dynamically adjusts the allocation of rendering resources according to the importance of each step in the interaction path guidance strategy and the real-time performance bottleneck indicators. For example, if the current guidance strategy focuses on device parameter setting, more rendering resources are preferentially allocated to the elements of the parameter setting interface to ensure that it can be quickly and accurately rendered.

[0112] Finally, the collaborative control model is injected into the user terminal interaction engine through an incremental update method. The system packages the updated part of the collaborative control model into an incremental update package and sends it to the user terminal. After receiving the update package, the interaction engine of the user terminal fuses the new collaborative control model part with the original engine function to continuously optimize the interface rendering performance. In this way, as the user uses and the performance bottleneck changes, the interaction engine can continuously adjust the rendering strategy to maintain good user experience.

[0113] In one non-limiting embodiment, the method further comprises:

[0114] The content browsing behavior data returned by the target user terminal is analyzed to extract cognitive load evaluation indexes including visual focus distribution features and information digestion rate; dynamic compression and expansion instruction sets of the interface information architecture are generated based on matching degree analysis of the cognitive load evaluation indexes and the content density of the guidance strategy; semantic aggregation and visual enhancement processing of the core guidance nodes in the current interaction path are performed based on the dynamic compression and expansion instruction sets to generate multi-granularity layered guidance views; dynamic switching logic of the layered guidance strategy is constructed based on the multi-granularity layered guidance views, and the dynamic switching logic includes weight allocation rules and trigger condition thresholds between different granularity levels; user cognitive rhythm change data is monitored in real time through the interaction state perception module of the target user terminal, and the granularity adaptive adjustment of the multi-granularity layered guidance views is triggered based on the dynamic switching logic.

[0115] Optionally, the system deeply analyzes the content browsing behavior data returned by the target user terminal. For example, through eye tracking technology or user operation record analysis deployed on the terminal interface, the visual focus distribution features of the user are obtained, that is, where the user's gaze is mainly concentrated when browsing the interface content; at the same time, the information digestion rate is calculated according to the time, operation behavior and other information of the user browsing content, such as the understanding and processing speed of the user for different types of information within a certain time. These data constitute cognitive load evaluation indexes.

[0116] Secondly, the cognitive load evaluation indexes are analyzed for matching degree with the content density of the guidance strategy. The content density of the guidance strategy refers to the amount of information presented in a unit interface space or a unit time in the interaction path. If the cognitive load of the user is low, that is, the information digestion rate is fast and the visual focus distribution is uniform, and the content density of the guidance strategy is low, the content density may need to be appropriately increased to provide more detailed information. Conversely, if the cognitive load of the user is high and the information digestion is difficult, the content density needs to be reduced. Based on this matching degree analysis, dynamic compression and expansion instruction sets of the interface information architecture are generated.

[0117] Then, according to the dynamic compression and expansion instruction sets, semantic aggregation and visual enhancement processing of the core guidance nodes in the current interaction path are performed. For example, for the guidance nodes of device operation steps, if the instruction set requires to compress information, some related and secondary operation prompts are semantically aggregated and summarized in simple language; at the same time, the core operation prompts are visually enhanced, such as increasing the font size, changing the color, etc., to highlight the key points. If the instruction set requires to expand information, the related knowledge and precautions of each operation step are expanded in detail. Through these processes, multi-granularity layered guidance views are generated, and different granularity levels contain guidance information of different detail levels.

[0118] Subsequently, based on the multi-granularity hierarchical guidance view, a dynamic switching logic of the hierarchical guidance strategy is constructed. A weight distribution rule between different granularity levels is determined, for example, when the user first contacts the device operation, the weight of the low granularity level (i.e. concise guidance) is higher, and as the user is familiar with the operation, the weight of the high granularity level (detailed guidance) is gradually increased. At the same time, the trigger condition threshold is set, such as when the user stays in one of the operation steps for more than a certain threshold, or repeats one of the operations for a certain number of times, triggering the switching from the low granularity level to the high granularity level.

[0119] Further, through the interaction state perception module of the target user terminal, real-time monitoring of user cognitive rhythm change data is performed. For example, the perception module can monitor the pause time of user operation, the continuity of operation and other information, so as to judge the cognitive rhythm of the user. When the monitored data meets the trigger condition threshold in the dynamic switching logic, the system triggers the granularity self-adaptive adjustment of the multi-granularity hierarchical guidance view based on the dynamic switching logic. In this way, the system can automatically adjust the detail level of the guidance view according to the real-time cognitive state of the user, provide the most suitable guidance information for the user, improve the understanding and execution efficiency of the user for the guidance strategy, and better complete various operation tasks in the smart factory.

[0120] It can be seen that in the smart factory application scenario, the embodiment of the present application realizes deep and comprehensive user interaction analysis: the real-time acquisition of the interaction data set can accurately capture the dynamic user demand; the context association features and interaction behavior features generated by multi-modal feature extraction provide a rich and detailed information basis for intent analysis. Based on this, the preset intent recognition rule library is used for joint intent analysis processing to generate an intent label set and a user portrait feature, accurately understand the user intent, and then generate an interaction path guidance strategy and push it to the target user terminal in real time to update the interaction interface. In this way, the user demand can be closely matched, personalized operation guidance can be provided, the human-computer interaction efficiency and user operation accuracy can be significantly improved, and the intelligent service level of the smart factory can be enhanced.

[0121] In summary, the embodiment of the present application can accurately identify user intent and provide personalized interaction path guidance by real-time acquisition of multi-dimensional interaction data, multi-modal feature extraction and joint intent analysis, effectively solving the deficiencies of the prior art in comprehensively understanding user demand and providing accurate guidance, greatly improving the intelligence and efficiency of human-computer interaction in the smart factory.

[0122] Based on the same inventive concept, the embodiment of the present application also provides an intelligent interaction guidance system. Referring to Figure 2 As shown in the figure, it is a possible structure schematic diagram of an intelligent interaction guidance system provided in the embodiment of the present application, Figure 2In some embodiments, the intelligent interaction guiding system 200 comprises a processor 210 and a memory 220. The memory 220 stores a computer program executable by the processor 210. The processor 210 can execute the steps of the intelligent interaction guiding method based on user intent recognition by executing the instructions stored in the memory 220.

[0123] Based on the same inventive concept, the present application provides a computer readable storage medium comprising a computer program, which, when executed on an intelligent interaction guiding system, causes the intelligent interaction guiding system to perform the steps of the intelligent interaction guiding method based on user intent recognition described above. In some possible implementations, various aspects of the intelligent interaction guiding method based on user intent recognition provided by the present application can also be implemented in the form of a program product, which comprises a computer program, which, when executed on an intelligent interaction guiding system, causes the intelligent interaction guiding system to perform the steps of the intelligent interaction guiding method based on user intent recognition described above, for example, the intelligent interaction guiding system can perform the steps shown in FIG. 13. Figure 1

[0124] In the technical solutions related to the above-mentioned embodiments of the present application, whether it is a multi-dimensional feature comparison calculation or a composite parameter construction, if there are problems caused by significant differences in the number of dimensions, units of dimensions, and semantic meanings of different features, those skilled in the art can fully understand that these differences need to be properly handled based on their professional knowledge and past practical experience, so that the calculation results are accurate and comparable, and avoid logical confusion, unclear mathematical meaning, and other conditions.

[0125] In detail, when facing features with different numbers of dimensions, in order to accurately calculate the similarity, matching degree or feature distance between different features, those skilled in the art can use a variety of strategies.

[0126] Feature selection is a commonly used method. For a high-dimensional feature set, a feature subset that matches the number of low-dimensional features and is most representative can be selected from the high-dimensional features according to indicators such as importance and correlation. By using chi-square test, information gain and other methods for feature selection, the most valuable features for the technical solution are screened out, so as to reduce the high-dimensional features to a dimension comparable to the low-dimensional features, and then calculate the similarity or distance.

[0127] ​Feature extraction is also an effective means, by constructing a suitable feature extraction model, different dimensions of features are mapped to a common low-dimensional feature space. Principal component analysis (PCA) can not only be used to deal with dimensional differences, but also project high-dimensional features to a low-dimensional space composed of principal components, so that features of different dimensions have comparability in this low-dimensional space. In addition, deep learning models such as autoencoders can also be used for feature extraction, which can automatically learn the latent representation of input features and convert features of different dimensions into feature vectors of the same dimension for subsequent similarity, matching or feature distance calculation.

[0128] In addition, the kernel method can also be used, and the kernel function can calculate the similarity between features in high-dimensional space without explicitly mapping the features to high-dimensional space. For features with different dimensions, appropriate kernel functions such as Gaussian kernel function and polynomial kernel function can be selected to directly calculate the similarity between them. This method avoids the difficulty of direct calculation caused by different feature dimensions and can effectively measure the relationship between features in the original feature space or implicit high-dimensional space.

[0129] In the process of comparing multi-dimensional features, in order to realize the comparability alignment of the feature space, those skilled in the art can use a variety of existing general technical means.

[0130] Standardization preprocessing is a widely used and effective method, which converts data into a standard normal distribution with a mean of 0 and a standard deviation of 1 by performing a specific linear transformation on the original feature data. This processing method can essentially eliminate the influence of the dimensions of different features and allow all features to be compared on the same scale. For example, in a dataset containing features of different dimensions, after standardization preprocessing, these features can be calculated for similarity or distance on the same scale, avoiding calculation bias caused by different dimensions.

[0131] Mapping conversion is also an effective way to solve the problem of dimensional difference, which can map the original features to a completely new space according to the specific properties of the features and the actual business requirements. In this new space, features of different dimensions can have better comparability. For features with nonlinear relationships, those skilled in the art can use logarithmic transformation, power transformation and other methods to convert them to linear relationships, making it easier to calculate similarity or distance. For example, when dealing with some features with exponential growth trend, logarithmic transformation can convert them to linear relationship, making subsequent calculation more accurate and convenient.

[0132] Spatial projection is also an important technical means, which projects high-dimensional feature space into low-dimensional space while preserving important information between features as much as possible. By carefully selecting the appropriate projection direction and projection dimension, those skilled in the art can effectively reduce the influence of dimensional difference on the calculation result while reducing the data dimension. Common spatial projection methods include principal component analysis (PCA), linear discriminant analysis (LDA), etc. Taking principal component analysis as an example, it projects high-dimensional data into a low-dimensional space composed of principal components, while simplifying the data structure and reducing the interference of dimensional difference on feature comparison.

[0133] In the construction process of composite parameters (such as loss function value), different parameter items often have different dimensions. Those skilled in the art can use normalization processing or adaptive weight distribution mechanism based on distribution characteristics.

[0134] Normalization processing is to unify the value range of different parameter items to a fixed interval, such as [0, 1]. This processing method can eliminate the influence of dimensional difference and ensure that each parameter item has the same importance in weighted fusion. Common normalization methods include min-max normalization, Z-score normalization, etc. Taking min-max normalization as an example, it performs linear transformation on the parameter item to scale its value range to the [0, 1] interval, so that parameter items of different dimensions can be weighted and fused under the same standard.

[0135] Adaptive weight distribution mechanism based on distribution characteristics dynamically adjusts the weights of different parameter items according to their distribution characteristics. For parameter items with large variance, those skilled in the art can appropriately reduce their weights; for parameter items with small variance, those skilled in the art can appropriately increase their weights. In this way, the composite loss function can pay more attention to parameter items with small variance, thereby improving the stability and generalization ability of the model. For example, in a composite loss function containing multiple parameter items, if the variance of a certain parameter item is large, it means that its fluctuation is large, which may adversely affect the stability of the model. Reducing its weight can reduce this adverse effect; for parameter items with small variance, increasing their weight can make the model pay more attention to the information reflected by the parameter item, thereby improving the overall performance of the model.

[0136] The above-mentioned general technical means for solving the problems of feature matching and loss balancing belong to the common knowledge in the art. These technical means have been fully verified and widely used in a large number of practical applications. Those skilled in the art can skillfully and flexibly use these methods to solve similar dimensional difference problems.

[0137] The formula and calculation process involved in the embodiments of the present application, whether for multi-dimensional feature comparison or composite loss function construction, strictly follow the principle of dimension correspondence. The variables in each formula have a clear and explicit physical meaning, and the operation logic is completely consistent with the basic mathematical and physical logic, and the operation result is necessarily the reasonable result expected by the present application. Those skilled in the art have the ability to effectively solve various problems caused by the number of dimensions, dimensional differences and the like in the multi-dimensional feature comparison calculation and the composite loss function construction in the embodiments according to the specific data situation and business requirements, and to ensure the accuracy, reliability and implementability of the technical solutions of the present application by comprehensively using the above general technical means.

Claims

1. A method for intelligent interaction guidance based on user intent recognition, characterized in that, The method comprises the following steps: real-time acquisition of an interaction data set of a target user, the interaction data set comprising a current interaction text sequence and a historical interaction behavior sequence; multi-modal feature extraction processing of the interaction data set to generate context association features of the current interaction text sequence and interaction behavior features of the historical interaction behavior sequence; joint intent analysis processing of the context association features and the interaction behavior features according to a preset intent recognition rule base to generate an intent label set of the current interaction text sequence and a user portrait feature; based on the intent label set and the user portrait feature, an interaction path guide strategy set is generated, and the interaction path guide strategy set is pushed to a target user terminal in real time to realize interaction interface update processing of the target user terminal.

2. The method of claim 1, wherein, The multi-modal feature extraction processing of the interaction data set to generate the context association features of the current interaction text sequence and the interaction behavior features of the historical interaction behavior sequence comprises: performing semantic word segmentation processing on the current interaction text sequence to obtain a plurality of text units, and extracting syntax structure features of each text unit; based on the syntax structure features, a pre-trained context encoding model is called to perform cross-sentence association analysis to generate context association features containing a semantic coherence index, the semantic coherence index representing the semantic association strength between the current interaction text and the historical interaction text; performing time series segmentation processing on the historical interaction behavior sequence to generate a plurality of interaction behavior segments; for each interaction behavior segment, performing behavior pattern recognition processing to extract interaction behavior features, the interaction behavior features including at least one of interaction operation frequency features, interaction path switching features and interface element dwell time features.

3. The method of claim 2, wherein, The time series segmentation processing on the historical interaction behavior sequence to generate a plurality of interaction behavior segments comprises: obtaining a timestamp data sequence of each interaction event in the historical interaction behavior sequence, and calculating a time interval set between adjacent two interaction events; based on the comparison result of each element in the time interval set with a preset interval threshold, performing dynamic segmentation operation on the historical interaction behavior sequence, when a target time interval is detected to exceed the preset interval threshold, inserting a segmentation point between adjacent interaction events corresponding to the target time interval to segment the historical interaction behavior sequence into a plurality of time-continuous interaction behavior segments; based on the segmented interaction behavior segment set, performing behavior density analysis to generate interaction operation frequency features, the interaction operation frequency features including a statistical value of the number of interaction operations per unit time in each interaction behavior segment and a distribution proportion vector of different operation types.

4. The method of claim 3, wherein, The generation of the interaction path guide strategy set based on the intent label set and the user portrait feature comprises: according to a dominant intent category identifier with the highest priority in the intent label set, an initial basic guide path configuration is matched and obtained from a preset interaction path template library; inputting the user portrait feature into a preset path optimization model, and performing the following personalized adjustment operations: analyzing a historical path adoption rate matrix and an interface element preference vector in the user portrait feature; dynamically adjusting a weight distribution parameter of a branch node in a basic guide path configuration based on adoption probability data of different path nodes in the historical path adoption rate matrix; and replacing a layout arrangement mode of a corresponding interface element in the basic guide path configuration according to a preference score of each type of interface element in the interface element preference vector; a plurality of interactive path guide strategy instances that are adjusted in a personalized manner are generated by combining different adjustment schemes of the weight distribution parameter and the layout arrangement mode, and the plurality of interactive path guide strategy instances are sorted according to preset strategy priority scores to form the interactive path guide strategy set.

5. The method of claim 4, wherein, The updating process of the user portrait feature includes: real-time monitoring of an interactive behavior change indicator triggered by the target user within a preset period, the interactive behavior change indicator including a newly added interactive operation type code and an operation frequency fluctuation amplitude value; based on the interactive behavior change indicator, incrementally updating the historical path adoption rate matrix in the user portrait feature: when the newly added interactive operation type code is detected, temporarily adding a weight coefficient to a new dimension in the corresponding interface element preference vector, the temporary weight coefficient being dynamically adjusted according to the interaction frequency; when the operation frequency fluctuation amplitude value exceeds a preset fluctuation threshold, triggering a completeness verification process of the user portrait feature, and supplementing initial values of corresponding portrait dimension features according to a missing dimension identifier determined by the completeness verification process; based on the incrementally updated user portrait feature, generating a corrected user portrait feature.

6. The method of claim 1, wherein, The construction process of the preset intent recognition rule library includes: collecting a multi-field user interaction sample set, and extracting a key intent expression feature vector and an associated behavior pattern feature matrix from each user interaction sample; based on the associated behavior pattern feature matrix, performing multi-level intent category labeling on the key intent expression feature vector to generate a labeled intent training sample set; using a multi-task learning framework to jointly train the intent training sample set to generate a hybrid model including an intent recognition main model and an auxiliary verification model, the auxiliary verification model being used to calculate a confidence score of an output result of the intent recognition main model; when the auxiliary verification model detects that the confidence score output by the intent recognition main model is lower than a preset verification threshold, triggering a review mechanism and generating a rule supplement instruction including a contradictory feature identifier; based on the contradictory feature identifier carried in the rule supplement instruction, constructing an intent matching condition expression and an associated behavior pattern feature matching rule in the intent recognition rule library, the intent matching condition expression and the behavior pattern feature corresponding to the contradictory feature identifier having a logical mapping relationship.

7. The method of claim 1, wherein, The real-time pushing of the interactive path guide strategy set to the target user terminal to realize the interactive interface updating process of the target user terminal includes: obtaining a device type parameter set of the target user terminal, the device type parameter set including a screen resolution parameter and an input mode support identifier; Perform interface adaptation processing based on the device type parameter set: parse the screen resolution parameter to determine the display area size, dynamically adjust the interface element size parameter and layout density coefficient in the interaction path guidance strategy set; according to the input mode support identifier, match the corresponding interaction control generation rule set, and reconstruct the touch response logic of the interface element; Generate the device-adapted interaction interface configuration file and push it to the target user terminal; When an interface configuration file loading exception event is monitored, start the alternative interface rendering scheme to re-render the interface, and the exception event includes element rendering misplacement or control response failure.

8. The method of claim 7, wherein, The method further comprises: Identify the conflict interface element identifier list and the abnormal layout parameter set in the current interaction interface configuration file; According to the conflict interface element identifier list, match the simplified interface template instance compatible with the device type parameter set from the preset alternative scheme library; Map the core guidance elements in the interaction path guidance strategy set to the placeholder nodes of the simplified interface template instance, generate interface update instructions containing degraded rendering parameters, and re-push them to the target user terminal; Record the feature log of the interface configuration file loading exception event, and add the device compatibility data in the feature log to the verification rule library to optimize the template matching algorithm of the subsequent adaptation process.

9. The method of claim 1, wherein, After generating the intent label set and user portrait feature of the current interaction text sequence, the method further comprises: Input the intent label set into the preset conflict detection model, identify the logical contradiction pair set between different intent labels in the intent label set; If a logical contradiction pair is detected in the logical contradiction pair set, extract the ambiguity expression segment index associated with the logical contradiction pair from the current interaction text sequence; Based on the historical clarification record data in the user portrait feature, generate a clarification guidance question set for the ambiguity expression segment, and insert the clarification guidance question set at the head of the priority queue of the interaction path guidance strategy set; Receive the feedback data stream of the target user for the clarification guidance question set, correct the intent label set according to the user selection identifier in the feedback data stream, and generate an updated intent label set; Based on the updated intent label set, regenerate the interaction path guidance optimization strategy, and trigger the incremental update operation of the interaction interface of the target user terminal.

10. An intelligent interactive guidance system, characterized by, It comprises a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method of any one of claims 1-9. It comprises a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method of any one of claims 1-9.

Images