Human-machine collaborative intelligent interaction intention alignment and correction method and system

Abstract

This application provides a method and system for aligning and correcting intelligent interaction intentions in human-computer collaboration, relating to the field of intelligent interaction technology. When an operator performs a human-computer interaction task, it captures any unexpected interface backtracking operations performed by the operator in response to the current interface content output by the intelligent interaction system, generates a backtracking trigger signal, and simultaneously acquires historical intention derivation path information prior to the backtracking time point. It then activates an intention divergence tracing module to analyze the feature differences between the historical output content corresponding to the backtracking target interface identifier and the subsequent operation input content, generating an intention divergence feature distribution map. This allows for node splitting and reconstruction of specific intention nodes, generating a corrected intention derivation path topology structure, which is then loaded into an online decision engine to regenerate interactive response content matching the operator's true intention. This achieves accurate alignment of human and machine intentions, significantly improving the efficiency and effectiveness of human-computer collaboration and enhancing the adaptability and reliability of the intelligent interaction system.

Description

Technical Field

[0001] This application relates to the field of intelligent interaction technology, and more specifically, to a method and system for aligning and correcting intelligent interaction intentions in human-computer collaboration. Background Technology

[0002] In human-machine collaborative intelligent interaction scenarios, effective communication and intent understanding between operators and intelligent interaction systems are crucial. Currently, intelligent interaction systems typically deduce the intent of operators' input based on preset algorithms and models, and generate corresponding interactive responses. However, due to the complexity and dynamic nature of operators' intents, deviations may occur during the deduction process, leading to a mismatch between the generated response and the operator's true intent.

[0003] In existing technologies, when the aforementioned mismatch occurs, operators often can only attempt to correct it by repeatedly inputting information or performing simple interface operations. However, these methods are inefficient and ineffective. Intelligent interaction systems themselves lack effective mechanisms to proactively identify intent discrepancies and correct and optimize historical intent derivation paths. While some systems possess simple error detection functions, they can only identify obvious input errors and cannot deeply analyze the root causes of intent discrepancies. Therefore, they cannot fundamentally solve the problem of human-machine intent misalignment, thus affecting the efficiency and effectiveness of human-machine collaboration. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a method and system for aligning and correcting intelligent interactive intentions in human-machine collaboration.

[0005] According to a first aspect of this application, a method for aligning and correcting intelligent interaction intentions in human-computer collaboration is provided, the method comprising:

[0006] During the process of the operator performing human-computer interaction tasks, the unexpected interface backtracking operation performed by the operator on the current interface content output by the intelligent interaction system is captured, and a backtracking trigger signal containing the backtracking time point and the backtracking target interface identifier is generated.

[0007] The system synchronously acquires historical intent deduction path information generated by the intelligent interaction system before the retrospective time point. The historical intent deduction path information includes multiple sequentially connected intent nodes and a decision basis feature vector corresponding to each intent node.

[0008] The intent divergence tracing module is activated to perform feature difference analysis on the historical output content corresponding to the target interface identifier and the subsequent operation input content collected after the time point of the backtracking, and generate an intent divergence feature distribution map describing the root cause of intent divergence.

[0009] Based on the intention divergence feature distribution map, a node splitting and reconstruction process is performed on a specific intention node in the historical intention derivation path information to generate a corrected intention derivation path topology containing multiple candidate intention branches.

[0010] The modified intent deduction path topology is loaded into the online decision engine of the intelligent interaction system, and the online decision engine regenerates interactive response content that matches the operator's true intent based on the modified intent deduction path topology.

[0011] According to a second aspect of this application, a system for aligning and correcting intelligent interaction intentions for human-machine collaboration is provided. The system includes a machine-readable storage medium and a processor. The machine-readable storage medium stores machine-executable instructions. When the processor executes the machine-executable instructions, the system implements the aforementioned method for aligning and correcting intelligent interaction intentions for human-machine collaboration.

[0012] Based on any of the above aspects, the technical effect of this application is as follows:

[0013] By capturing unexpected interface retracing operations from operators and generating retracing trigger signals, the system can promptly detect operator dissatisfaction with system output. Simultaneously, it acquires historical intent derivation path information, activates the intent divergence tracing module for feature difference analysis, and generates an intent divergence feature distribution map. This accurately pinpoints the root cause of intent divergence, providing a clear direction for subsequent corrections. Based on this map, specific intent nodes are split and reconstructed to generate a corrected intent derivation path topology, effectively optimizing historical intent derivation paths. Finally, this is loaded into the online decision engine to regenerate interactive response content, achieving accurate alignment of human and machine intents. This significantly improves the efficiency and effectiveness of human-machine collaboration and enhances the adaptability and reliability of the intelligent interaction system. Attached Figure Description

[0014] Figure 1 A flowchart illustrating the intelligent interaction intent alignment and correction method for human-machine collaboration provided in an embodiment of this application is shown.

[0015] Figure 2 This illustration shows a schematic diagram of the component structure of the intelligent interactive intent alignment and correction system for human-machine collaboration provided in an embodiment of this application. Detailed Implementation

[0016] Figure 1 This paper illustrates a flowchart of a method and system for aligning and correcting intelligent interaction intentions in human-machine collaboration, as provided in an embodiment of this application. The detailed steps include:

[0017] Step S110: During the operation of the human-computer interaction task, capture the unexpected interface backtracking operation performed by the operator on the current interface content output by the intelligent interaction system, and generate a backtracking trigger signal containing the backtracking time point and the backtracking target interface identifier.

[0018] During the interaction between the operator and the intelligent interactive system, the system monitors various user actions performed on the user interface in real time. This monitoring is achieved through the event listening mechanism of the underlying operating system. When the operator performs a specific type of operation on the interface, the system captures relevant information about that operation. Unintended interface backtracking operations refer to actions performed by the operator after viewing the current interface content, due to misunderstanding, dissatisfaction, or perception that the content does not conform to their intentions, resulting in a deliberate reversal to a previous interface. Examples include clicking the back button on the interface, performing specific touch gestures such as swiping from left to right on the screen, or pressing the physical back button. When such a backtracking operation is captured, the system immediately generates a backtracking trigger signal. This signal is a structured data object containing a backtracking time point field, which is obtained by calling the system clock function and records the precise moment of the backtracking operation in millisecond-level timestamps. It also contains a backtracking target interface identifier field, which is a unique identifier extracted from the interface object of the target interface returned by the backtracking operation, such as the unique hash value of the interface in the system interface stack. The generated backtracking trigger signal is temporarily stored in the system's shared memory area.

[0019] Step S120: Synchronously acquire the historical intent deduction path information generated by the intelligent interaction system before the retrospective time point. The historical intent deduction path information includes multiple sequentially connected intent nodes and the decision basis feature vector corresponding to each intent node.

[0020] Based on the backtracking time point in the backtracking trigger signal, the historical intent deduction path information corresponding to that time point is synchronously read from the decision log database of the intelligent interaction system. This historical intent deduction path information is a complete record of a series of intent reasoning processes performed by the system before the backtracking time point to generate the final interface content output to the operator. It is stored in a directed graph data structure containing multiple sequentially connected intent nodes. Each intent node is a data structure containing a node identifier, a node generation timestamp, and a decision basis feature vector. The node identifier is a globally unique code assigned when the intent node is generated. The node generation timestamp records the time when the node was created. The decision basis feature vector is a D-dimensional floating-point vector, which is obtained by encoding all input information that leads the system to select the current intent branch, such as the operator's historical operation sequence, current context information, and the system's preset knowledge base matching results. The obtained historical intent deduction path information is formatted into a graph data structure containing a node list and an edge list. Each element in the node list contains the above three fields, and each element in the edge list contains a predecessor node identifier and a successor node identifier.

[0021] Step S130: Activate the intent divergence tracing module to perform feature difference analysis on the historical output content corresponding to the target interface identifier and the subsequent operation input content collected after the backtracking time point, and generate an intent divergence feature distribution map describing the root cause of intent divergence.

[0022] Step S131: Parse the target interface identifier in the backtracking trigger signal, and retrieve the complete historical interface output content data corresponding to the target interface identifier from the interface history database.

[0023] First, the backtracking trigger signal is analyzed, and the target interface identifier field is extracted. Using this field as the query key, a search operation is performed in the interface history database of the intelligent interaction system. The interface history database stores complete snapshots of all previously output interfaces. The search result returns complete, uncompressed historical interface output content data. This data is a multimodal data object containing all text content, graphic elements, control layout information, etc., on the interface. It can be a structured document object model tree, where each node corresponds to a visible element on the interface, and the node attributes record the element's type, content, position coordinates, etc. The retrieved complete historical interface output content data is loaded into memory, ready for subsequent parsing and processing.

[0024] Step S132: Extract the sequence of subsequent operation input content collected within a preset time window after the retrospective time point from the operation log database. The sequence of subsequent operation input content includes at least one continuous touch operation trajectory data input by the operator.

[0025] Based on the backtracking time point in the backtracking trigger signal, all operation input content collected within a preset time window after that time point is extracted from the operation log database. The length of this preset time window is a configurable system parameter, set to T milliseconds after the backtracking time point. The extracted subsequent operation input content is a sequence, where each element is an operation event record containing at least one continuous touch operation trajectory data input by the operator. The touch operation trajectory data records the continuous movement path of the finger or stylus on the screen, containing a series of sampling points ordered by time. Each sampling point includes the screen's horizontal coordinate value, screen's vertical coordinate value, pressure value, and timestamp offset. These consecutive sampling points constitute a multidimensional time series describing the geometric characteristics of the operation behavior. The extracted sequence of subsequent operation input content is temporarily stored in a memory buffer.

[0026] Step S133: Perform content structure parsing processing on the complete historical interface output content data, and extract the core semantic feature vector of each output content unit contained in the complete historical interface output content data.

[0027] The complete historical interface output data is parsed. First, the document object model tree of the interface content data is traversed to identify each independent minimum output content unit, such as a text block, an icon, a button, or an input box. For each identified output content unit, its core semantic information is extracted. For text blocks, the text string is extracted; for graphic elements, their shape descriptors or icon category labels are extracted. The extracted raw information is then input into a pre-trained multimodal semantic encoder. This encoder uses a Transformer architecture, whose input layer receives embedded representations of text and graphics. After processing through multiple encoder layers using self-attention mechanisms and feedforward neural networks, each output content unit is mapped to a fixed-dimensional core semantic feature vector. Assuming this vector has a dimension of E, it encodes the position of the output content unit in the semantic space. Finally, a set of N_unit core semantic feature vectors is obtained.

[0028] Step S134: Perform sequence feature encoding processing on the subsequent operation input content sequence, convert each continuous touch operation trajectory data in the subsequent operation input content sequence into a corresponding operation trajectory geometric feature vector, and stack all operation trajectory geometric feature vectors to form an operation behavior temporal feature matrix.

[0029] The subsequent input sequence is processed by first separating all continuous touch operation trajectory data from the sequence. For each trajectory, since the number of sampling points is variable, normalization encoding is required. An original trajectory containing M sampling points, each containing two-dimensional coordinates and time, is input into a trajectory encoder based on a Long Short-Term Memory network. This encoder processes the coordinate sequence of each sampling point step by step through its recurrent units, and finally aggregates the hidden states of the entire sequence to output a fixed-dimensional operation trajectory geometric feature vector. Assuming the dimension of this operation trajectory geometric feature vector is F, it comprehensively describes the geometric characteristics of the trajectory, such as shape, direction, and length. If the subsequent input sequence contains K consecutive touch operation trajectories, K F-dimensional operation trajectory geometric feature vectors are obtained through the above processing. These K vectors are stacked in chronological order to form an operation behavior temporal feature matrix of shape K rows and F columns. Each row of this matrix represents the geometric feature of a trajectory, and the order of the rows reflects the temporal relationship of the operation.

[0030] Step S135: Input the core semantic feature vector and the temporal feature matrix of the operation behavior into the pre-trained cross-modal feature alignment network for feature space mapping and alignment processing, generate the aligned feature representation of the historical output content in the unified feature space and the aligned feature representation of the subsequent operation input content in the unified feature space, calculate the difference between the aligned feature representations, and generate an initial feature difference representation, which is used to describe the degree of difference between the historical output content and the subsequent operation input content in the unified feature space.

[0031] N_unit core semantic feature vectors and the temporal feature matrix of operational behavior are input into a pre-trained cross-modal feature alignment network. The core of this network is a dual-tower structure, comprising a semantic encoding tower and a behavioral encoding tower, with the top layers of both towers connected to a shared feature projection layer. The semantic encoding tower first fuses the N_unit core semantic feature vectors into a global interface semantic feature vector of dimension G through an attention pooling layer. The behavioral encoding tower first flattens the temporal feature matrix of operational behavior and maps it to a global operational behavior feature vector of dimension G through a multilayer perceptron. The shared feature projection layer projects the global interface semantic feature vector and the global operational behavior feature vector into the same unified feature space of dimension H through two independent linear transformations, obtaining aligned historical output content and aligned subsequent operation input content feature vectors. The element-wise difference between these two H-dimensional vectors is calculated to obtain an initial feature difference vector of dimension H, where each element represents the degree of difference between the historical output and the subsequent operation in the corresponding dimension of the unified feature space.

[0032] Step S136: Perform saliency enhancement processing on the initial feature difference vector, identify feature dimensions in the initial feature difference vector whose feature value difference exceeds a preset saliency threshold, mark the feature dimensions as key feature dimensions of intent divergence, and back locate the output content unit region associated with each key feature dimension of intent divergence in the output content data of the complete history interface according to the key feature dimensions of intent divergence, and generate a feature-region mapping relationship between each key feature dimension of intent divergence and the output content unit region.

[0033] The initial feature difference vector is analyzed and normalized to unify the value range of all elements to between zero and one. Each dimension of the vector is traversed, and its value is compared with a preset significance threshold T_sig. Dimensions with difference values ​​greater than T_sig are marked as key feature dimensions for intent divergence, and their dimension indices are recorded. For each marked key feature dimension for intent divergence, backpropagation using an attention mechanism is employed to calculate the contribution weight of each dimension of the unified feature space to each output content unit in the original interface. For the current key feature dimension for intent divergence, one or more output content units with the highest contribution weight are identified. The coordinate regions of these units on the interface are used as the output content unit regions associated with that dimension, establishing a mapping relationship from each key feature dimension for intent divergence to a specific interface region, stored in key-value pairs.

[0034] Step S137: Based on the intent divergence key feature dimension, reverse locate the operation trajectory segment associated with each intent divergence key feature dimension in the subsequent operation input content sequence, and generate the feature-trajectory mapping relationship between each intent divergence key feature dimension and the operation trajectory segment.

[0035] For each marked key feature dimension of intent divergence, its source is located in the subsequent operation input content. Using the interpretability analysis of the behavior coding tower in step S135, the response weight of each dimension of the unified feature space to each trajectory in the original operation behavior temporal feature matrix is ​​calculated. For the current key feature dimension of intent divergence, the one or more consecutive touch operation trajectory data with the highest contribution weight are identified. The original sampling point sequence of the above trajectory is used as the operation trajectory segment associated with that dimension, establishing a mapping relationship from each key feature dimension of intent divergence to the specific operation trajectory, stored in key-value pair format.

[0036] Step S138: Perform association and fusion processing on the feature-region mapping relationship and the feature-trajectory mapping relationship to construct an intent divergence feature distribution map with the key feature dimension of intent divergence as nodes and the feature-region mapping relationship and the feature-trajectory mapping relationship as node attributes.

[0037] The feature-region mapping and feature-trajectory mapping are fused, with each key feature dimension of intent divergence serving as a node in the graph. Each node has two core attributes: the first is a list of interface regions corresponding to that dimension, and the second is a list of operation trajectory segments corresponding to that dimension. The association strength between different nodes is calculated. If the regions corresponding to two feature dimensions overlap in the original interface or operation sequence, or if their cosine similarity in the unified feature space is high, an undirected edge is added between the two nodes, with the edge weight being their similarity. The final constructed intent divergence feature distribution graph is a heterogeneous graph data structure. Nodes are feature dimensions, node attributes contain the specific manifestations of the divergence point in both interface regions and operation trajectories, and edges represent the potential associations between different divergence points.

[0038] Step S140: Perform node splitting and reconstruction processing on specific intent nodes in the historical intent derivation path information according to the intent divergence feature distribution map, and generate a corrected intent derivation path topology structure containing multiple candidate intent branches.

[0039] Step S141: Parse the historical intent derivation path information to obtain the node identifier, node generation timestamp, and decision basis feature vector of each intent node contained in the historical intent derivation path information.

[0040] The detailed information of all intent nodes contained in the historical intent derivation path information is parsed out. The node list in the path information is traversed. For each intent node in the list, three core data items are extracted: node identifier, node generation timestamp, and decision basis feature vector with dimension D.

[0041] Step S142: Extract the feature representation of each key feature dimension of intent divergence from the intent divergence feature distribution map, and calculate the feature matching degree with the decision basis feature vector of each intent node to generate a feature matching degree score matrix between the feature representation of each key feature dimension of intent divergence and the decision basis feature vector of each intent node.

[0042] Information on all key feature dimensions of intent divergence is extracted from the intent divergence feature distribution map. For each key feature dimension, the corresponding projection weight vector is taken from the shared feature projection layer in step S135, assuming its dimension is P, and used as the feature representation of that key feature dimension. The D-dimensional decision basis feature vector of each intent node and the P-dimensional feature representation of each key feature dimension of intent divergence are mapped to the same dimension Q through linear transformation, and then their inner product is calculated to obtain a scalar value. A feature matching degree score matrix of shape M rows and N columns is formed, where the element in the i-th row and j-th column of the matrix represents the matching degree between the i-th key feature dimension of intent divergence and the j-th intent node.

[0043] Step S143: Identify intent nodes whose matching degree with the key feature dimension of intent divergence exceeds a preset matching degree threshold based on the feature matching degree score matrix, and mark the intent nodes as intent nodes to be split.

[0044] Set a preset matching threshold T_match, traverse the feature matching score matrix, and for each column, i.e., each intent node, find all rows whose matching score exceeds T_match, i.e., the key feature dimensions of intent divergence. If an intent node has at least one key feature dimension of intent divergence with a matching score exceeding the threshold, then mark the intent node as an intent node to be split, and record a list of all key feature dimensions of intent divergence associated with it.

[0045] Step S144: For each intent node to be split, perform feature decomposition processing on the decision basis feature vector of the intent node to be split according to the set of key feature dimensions of intent divergence that it matches, and decompose the decision basis feature vector of the intent node to be split into sub-feature vector components corresponding to each key feature dimension of intent divergence.

[0046] For each labeled intent node to be split, let its decision-based feature vector be V_node, with dimension D. There are L key intent divergence feature dimensions matched with this node, each corresponding to a projection weight vector W_i, with dimension P, already mapped to D dimensions. The projection coefficients c_i of V_node in each direction W_i are calculated as c_i = (V_node·W_i) / (W_i·W_i). The sub-feature vector component V_i corresponding to each key intent divergence feature dimension is calculated as V_i = c_i × W_i. The original V_node is approximately decomposed into the sum of L sub-feature vector components plus a residual vector.

[0047] Step S145: Based on the sub-feature vector components corresponding to the key feature dimension of each intent divergence, generate at least two initial split candidate intent branch nodes for each intent node to be split. The decision basis feature vector of each initial split candidate intent branch node is composed of the corresponding sub-feature vector components.

[0048] For each sub-feature vector component V_i, it is used as a new, independent decision-based feature vector for a candidate intent branch node. For an intent node to be split with L matching dimensions, L initial splitting candidate intent branch nodes are generated. The decision-based feature vector for each new node is the corresponding sub-feature vector component V_i. The node identifier is a new UUID automatically generated by the system, and the node generation timestamp is set to the current processing time.

[0049] Step S146: Assign an initial branch credibility weight to each initial split candidate intention branch node. The initial branch credibility weight is calculated based on the difference weight coefficient of the intention divergence key feature dimension corresponding to the initial split candidate intention branch node in the intention divergence feature distribution map.

[0050] For each initial split candidate intention branch node, calculate the initial confidence weight by extracting the difference value diff_i of that feature dimension and the confidence value conf_i of the operation trajectory segment associated with that feature dimension from the graph. The initial branch confidence weight w_i_init is calculated as w_i_init=(diff_i×conf_i) / (∑_{j=1}^{L}(diff_j×conf_j)), where L is the total number of initial split candidate intention branch nodes.

[0051] Step S147: Perform semantic similarity clustering analysis on the initial split candidate intention branch nodes, calculate the similarity of decision basis feature vectors between different initial split candidate intention branch nodes, and merge multiple initial split candidate intention branch nodes whose decision basis feature vector similarity exceeds a preset similarity threshold into the same aggregated intention branch node.

[0052] Calculate the pairwise cosine similarity between the decision-based feature vectors V_i of all L initial split candidate intention branch nodes. The cosine similarity is calculated as sim(V_i,V_j)=(V_i·V_j) / (||V_i||×||V_j||). Set a preset similarity threshold T_sim. If the cosine similarity between any two branch nodes V_i is greater than T_sim, they are grouped into the same candidate cluster. For each formed cluster, generate a new aggregated intention branch node. The decision-based feature vector of this aggregated node is the average of the decision-based feature vectors of all branch nodes within the cluster, and the initial confidence weight of this aggregated node is the sum of the initial confidence weights of all branch nodes within the cluster.

[0053] Step S148: Based on the initial branch credibility weight, recalibrate the credibility of the merged aggregated intent branch nodes and the initial split candidate intent branch nodes that did not participate in the merge, and generate the calibrated branch credibility weight for each retained candidate intent branch node.

[0054] The current weights of all retained nodes after merging are combined into a vector W. W is then transformed using a Softmax function with a temperature parameter T_temp to obtain the new calibrated weights W_cal. The formula for calculating the calibrated weights W_cal_i is W_cal_i=e^(W_i / T_temp) / (∑_{j=1}^{N}e^(W_j / T_temp)), where W_i is the original weight of the i-th node, N is the total number of retained candidate intent branch nodes, and T_temp is the preset temperature parameter.

[0055] Step S149: Remove candidate intent branch nodes whose calibrated branch confidence weight is lower than the preset retention threshold, and retain candidate intent branch nodes whose calibrated branch confidence weight is higher than the preset retention threshold as the final retained intent branch nodes. Based on the original connection relationship between each intent node in the historical intent derivation path information and the derivation relationship between the final retained intent branch node and the original intent node, construct a corrected intent derivation path topology structure that includes the final retained intent branch node and its logical connection relationship with each other.

[0056] A preset retention threshold T_keep is set. All calibrated candidate intent branch nodes are traversed, and their calibrated branch confidence weights are compared with T_keep. Nodes with weights lower than T_keep are removed, and nodes with weights higher than T_keep are retained as the final retained intent branch nodes. The edge list in the original historical intent derivation path information is parsed to determine the position of the original intent node to be split in the path. The original intent node to be split is removed from the path, and all the final retained intent branch nodes are inserted into the path as successors of their predecessor nodes. These final retained intent branch nodes all point to all the original successor nodes of the original intent node to be split. The final constructed corrected intent derivation path topology is a graph structure containing multiple parallel branches. The nodes in the graph include unaffected original intent nodes and newly generated final retained intent branch nodes, and the edges reflect the temporal and logical connections between them.

[0057] Step S150: Load the modified intent derivation path topology into the online decision engine of the intelligent interaction system, and the online decision engine regenerates the interactive response content that matches the operator's true intent based on the modified intent derivation path topology.

[0058] Step S151: Convert the modified intent derivation path topology into a structured decision path description file that can be recognized by the online decision engine. The structured decision path description file contains the node identifier of each final retained intent branch node, the calibrated branch credibility weight corresponding to each final retained intent branch node, and the encoding of the logical connection relationship between each final retained intent branch node.

[0059] The corrected intent derivation path topology is serialized to generate a structured decision path description file in JSON format that can be parsed by an online decision engine. Within the file, an object entry is created for each node that ultimately retains the intent branch. This entry contains three core fields: a node identifier field storing the node's unique identifier string; a weight field storing the calibrated branch confidence weight; and a logical succession relationship encoding field storing a list where each element represents the node identifier of the next node pointed to from the current node. The outermost layer of the entire file specifies the starting node identifier.

[0060] Step S152: Inject the structured decision path description file into the runtime context of the online decision engine, replacing the historical intent derivation path information currently cached by the online decision engine and associated with the traceback target interface identifier.

[0061] The structured decision path description file is injected into the engine's runtime context as new context data through the application programming interface provided by the online decision engine. Based on the backtracking target interface identifier in the backtracking trigger signal, the cache entry corresponding to the identifier is found in the runtime context, and the old historical intent derivation path information data structure stored in the cache entry is completely replaced with the new structured decision path description file.

[0062] Step S153: The online decision engine parses the structured decision path description file, determines the starting intent branch node to be activated based on the logical succession relationship encoding in the structured decision path description file, and starts from the starting intent branch node. The online decision engine traverses each final retained intent branch node in the corrected intent derivation path topology, and generates output content generation parameters corresponding to each final retained intent branch node based on the calibrated branch credibility weight and the decision basis feature vector of each final retained intent branch node.

[0063] The online decision engine parses the structured decision path description file, reads the root node identifier to determine the starting intent branch node that needs to be activated, and traverses each of the final retained intent branch nodes in the corrected intent derivation path topology starting from this starting node. For each traversed node, the calibrated branch credibility weight is extracted from its corresponding object entry, and the decision basis feature vector is read from the node's attributes. These two pieces of information, along with the node's depth-level information in the topology, are combined to form a multi-dimensional output content generation parameter. This parameter is a structure containing the above three elements.

[0064] Step S154: Input the output content generation parameters corresponding to each final retained intent branch node into the content generation model built into the online decision engine. The content generation model generates a branch output content candidate set corresponding to each final retained intent branch node according to the output content generation parameters. The branch output content candidate set is fused. The content elements in the branch output content candidate set are weighted and merged according to the weight coefficient of each final retained intent branch node in the modified intent derivation path topology structure to generate an initial fused interactive response content draft.

[0065] The output content generation parameters generated by each node are input into the content generation model built into the online decision engine. This model is a pre-trained Transformer-based generation model. After receiving the parameters, the model uses them as the initial hidden state for the generation process, generating output content word by word through autoregression. Each node generates an independent set of candidate content. After generating candidate content from all nodes, the content is merged into a unified response draft. The fusion process uses a weighted merging based on the weight coefficient of each node in the topology; nodes with higher weight coefficients have a higher proportion of their generated candidate content in the final draft. The final output is an initial fused interactive response content draft containing information from all active branches.

[0066] The content generation model adopts a Transformer-based generative model architecture, consisting of an input encoding layer, multiple parallel decoding branches, and an output mapping layer. The input encoding layer concatenates the decision-based feature vector, deep hierarchical embeddings, and weight embeddings into a 3D input encoded vector. The decoding branches include shared parameters in the L_shared layer and independent parameters in the L_independent layer. Each decoder layer has a structure similar to a multimodal semantic encoder but includes masked self-attention. The decoder layer contains a masked multi-head self-attention sublayer, a cross-attention sublayer, and a feedforward neural network sublayer. Masked self-attention prevents the seeing of future information, and cross-attention uses the input encoded vector as the key and value. The output mapping layer is a fully connected layer with a softmax layer, mapping the decoded output sequence to the probability distribution of the output content vocabulary. It is pre-trained on a large-scale human-computer interaction content dataset containing C_million sample pairs from intent nodes to output content. The input is the parameters for generating the output content, and the output is the corresponding output content sequence. An autoregressive language modeling loss is used, with AdamW as the optimizer, an initial learning rate of L_gen, a batch size of B_gen, and a training epoch of E_gen. The evaluation metric is perplexity. In step S290, KL divergence loss is used to fine-tune the last few layers of the model. During inference, the input encoded vector is processed through shared and independent layers, and an autoregressive layer generates output content unit codes word by word. The output mapping layer then converts these codes into actual vocabulary or graphical element codes. Multiple parallel decoding branches independently generate different candidate contents.

[0067] Step S155: Perform content coherence optimization processing on the initial fused interactive response content draft. Adjust the arrangement order and connection method of different content segments in the initial fused interactive response content draft according to the logical connection relationship between each final retained intent branch node in the modified intent derivation path topology structure to generate coherent optimized interactive response content. Extract core intent expression summary information from the coherent optimized interactive response content, and compare and verify the core intent expression summary information with the historical interface output content before the backtracking time point carried in the backtracking trigger signal. Encapsulate the verified coherent optimized interactive response content into a data packet format that conforms to the output interface standard of the intelligent interaction system, and add an intent correction identifier field to the header of the data packet format.

[0068] The initial draft undergoes content coherence optimization. The logical connections between nodes in the revised intent derivation path topology are analyzed, and the order of content fragments from different nodes is adjusted based on these relationships, inserting appropriate transition words or sentences between paragraphs. After optimization, a core intent expression summary is extracted from the coherence-optimized content. This summary is compared and verified with historical interface outputs prior to the previous time point to check if the new response truly resolves the intent divergence. If verification passes, the final coherence-optimized interactive response content is encapsulated, a data packet is created according to the system output interface standard, the content data is written to the payload area of ​​the data packet, and a specific intent correction identifier field is added to the header area of ​​the data packet.

[0069] Step S156: The data packet with the intent correction identifier field added is sent to the user interface rendering module of the intelligent interaction system, and the user interface rendering module generates the final interactive response content displayed on the user interface based on the content data in the data packet.

[0070] The encapsulated data packet with the intent correction identifier is sent to the user interface rendering module via the system's internal inter-process communication mechanism. After receiving the data packet, the rendering module parses the packet header, identifies the intent correction identifier field, and calls the underlying graphics library and text rendering engine based on the content data in the data packet payload. The text content is drawn onto the specified screen area, and the graphic elements are drawn according to their descriptors. Finally, a complete user interface frame is generated and output to the physical screen through the display driver for display to the operator.

[0071] Step S210: After loading the modified intent derivation path topology into the online decision engine, an intent alignment status visualization indicator is generated synchronously in the preset monitoring area on the user interface of the intelligent interaction system. The intent alignment status visualization indicator is used to present the summary information of the modified intent derivation path topology currently loaded by the online decision engine in real time.

[0072] After step S152, which loads the revised path topology into the online decision engine, an intent alignment status visualization indicator is immediately generated in a fixed, preset monitoring area of ​​the user interface. This indicator is initially presented in a concise summary form, such as a circular icon whose color represents the overall alignment status, or displays the number of currently active branches. The indicator is drawn by the front-end rendering logic based on runtime context information shared in real-time by the online decision engine, and its status is updated synchronously as the engine loads a new path structure.

[0073] Step S220: Continuously monitor the selective focusing operation performed by the operator in the area where the intention alignment status visualization indicator is located. When the operator's zoom-in gesture is captured, dynamically control the switching or loading of the visualization interface according to the gesture parameters of the zoom-in gesture, switching from displaying the summary information of the corrected intention derivation path topology to displaying the detailed topology diagram of the corrected intention derivation path topology.

[0074] The system's underlying event listeners continuously monitor user touch operations on the screen area where the intent alignment status visualization indicator is located. When a specific zoom-in gesture is detected, such as a pinch gesture with two fingers expanding outwards in the indicator area, the system triggers interface switching logic based on the gesture parameters. This logic first hides or minimizes the current summary information display, then reads the complete and corrected intent derivation path topology data from the runtime context of the online decision engine, and calls the graphics layout engine to dynamically calculate the coordinates of each node in the graph and the path of the connecting lines based on the screen size and the number of nodes. This renders and generates an interactive, detailed topology graph, which is then loaded into the front-end display layer of the current interface.

[0075] Step S230: After the detailed topology diagram is fully expanded, a visual progress bar for the calibrated branch confidence weight corresponding to each final retained intention branch node is generated synchronously near the display area of ​​each final retained intention branch node in the detailed topology diagram.

[0076] In the callback function after the detailed topology diagram rendering is complete, the system iterates through the graphical elements corresponding to each node of the final retained intention branch in the diagram. For each node, the calibrated branch confidence weight calculated in step S148 is read from its associated data object; this weight is a floating-point number between zero and one. A horizontal or vertical progress bar graphical element is dynamically generated near the graphical element of this node. The length or height of the progress bar is proportional to the weight value; the larger the weight value, the longer the fill portion of the progress bar.

[0077] Step S240: Simultaneously capture the node selection operation performed by the operator on the detailed topology diagram. When the operator selects any final retention intention branch node in the detailed topology diagram through touch operation, display a simplified summary information of the decision basis feature vector of the selected final retention intention branch node in the floating information window on the user interface.

[0078] While displaying a detailed topology diagram, a touch event listener is enabled to capture operator clicks or touch selections on nodes within the diagram. When an operator selects a graphical element that ultimately intends to retain a branch node, the front-end logic immediately captures the node's identifier. Based on this identifier, detailed attribute information of the node, particularly its decision-making feature vector, is retrieved from the local cache or from the back-end via an asynchronous request. A feature summarization generation module is invoked to map the high-dimensional vector to a low-dimensional space using a pre-trained dimensionality reduction encoder and further convert it into a series of human-readable keywords or brief descriptions. This simplified summary information is then popped up in a predefined floating information window, which subsequently appears near the selected node.

[0079] Step S250: Simultaneously generate a branch confirmation control and a branch rejection control in the floating information window. When the operator triggers the branch confirmation control, increase the calibrated branch credibility weight of the selected branch node with the final intention to retain by a preset weight adjustment step value.

[0080] In the floating information window generated in step S240, in addition to displaying summary information, two virtual button controls are generated side by side: a branch confirmation control marked as "Confirm" and a branch rejection control marked as "Reject". The system continuously listens for the trigger events of these two controls. When the operator clicks the confirmation button, the front end captures the event and sends an instruction containing the selected node identifier and the adjustment direction as "increase" to the online decision engine. After receiving the instruction, the decision engine finds the weight value of the node from the runtime context and increases it by a preset weight adjustment step value Δw. At the same time, in order to ensure that the sum of the weights of all branch nodes is still 1, the engine will re-normalize the weights of all sibling branch nodes at the same level.

[0081] Step S260: When the operator triggers the branch rejection control, the calibrated branch credibility weight of the selected final retention intention branch node is reduced by a preset weight adjustment step value, and the display length of the corresponding visual progress bar in the detailed topology diagram is updated in real time.

[0082] When an operator clicks the "Vote" button in the floating information window, the front-end also sends a command to the online decision engine. The engine reduces the confidence weight of the selected node's calibrated branch by a preset weight adjustment step size Δw, and also performs a normalization operation on the weights of sibling nodes at the same level so that the sum is 1. After the weight update is complete, the engine pushes the new weight values ​​to the front-end rendering module via a real-time communication protocol. Upon receiving the update, the front-end immediately redraws the visual progress bar of the corresponding node in the detailed topology diagram, adjusting its fill length to reflect the latest weight status in real time.

[0083] Step S270: Generate a global intent confirmation control in the edge area of ​​the user interface. When the operator triggers the global intent confirmation control, freeze the real-time adjustment function of the calibrated branch confidence weight of all final retained intent branch nodes in the detailed topology diagram, and encapsulate the current detailed topology diagram and the final calibrated branch confidence weight of all final retained intent branch nodes contained therein into an intent alignment snapshot after operator confirmation.

[0084] A global intent confirmation control, such as a prominent confirmation button, is generated in a fixed area at the edge of the user interface. After the operator manually adjusts the branch weights through steps S250 and S260, they can click this global confirmation control if they are satisfied with the adjustment. Once the trigger event of this control is captured, the system first sends an instruction to the front end to freeze the interactive functions of all nodes on the detailed topology diagram, especially disabling the branch confirmation and rejection controls to prevent further modification of the weights. At the same time, a snapshot generation request is sent to the online decision engine. After receiving the request, the engine serializes the complete corrected intent derivation path topology structure at the current moment, including all nodes, edges, and the latest calibrated branch confidence weights adjusted manually by the operator, adds the current timestamp and operator identifier, and encapsulates it into an immutable intent-aligned snapshot object after operator confirmation.

[0085] Step S280: Store the intent alignment snapshot confirmed by the operator in the intent alignment history database of the intelligent interaction system, and assign a unique timestamp and operator identifier to the intent alignment snapshot.

[0086] The intent alignment snapshot object generated in step S270 is written to the intent alignment history database of the intelligent interaction system via the database interface. This database is a time-series database specifically designed to store such historical states. Before writing, a globally unique timestamp, accurate to the millisecond level, is generated for this snapshot record, and an operator identifier for performing this confirmation operation is appended. This identifier is a de-identified user-unique code obtained from the current login session.

[0087] Step S290: Based on the final calibrated branch credibility weights of each final retained intent branch node recorded in the intent alignment snapshot, fine-tune and update the model parameters of the pre-trained content generation model in the online decision engine.

[0088] The final calibrated branch confidence weights of each final retained intent branch node are parsed from the intent alignment snapshot stored in step S280. These weights are used as a supervision signal and compared with the original output content generation parameters used when generating these nodes in step S153. A fine-tuning loss function is constructed, calculated as L_finetune = KL_divergence(P_model||P_user), where P_model is the probability distribution of the current model output, and P_user is the expected distribution indicated by the operator through weight adjustments. Using this loss function, the last few layers or all parameters of the pre-trained content generation model in the online decision engine are fine-tuned in small steps using the backpropagation algorithm, making the updated model more inclined to generate output content consistent with the intent branch weight distribution confirmed by the operator.

[0089] Step S310: parse the output content generation parameters corresponding to each final retained intention branch node. The output content generation parameters include the decision basis feature vector corresponding to the final retained intention branch node, the depth level information of the final retained intention branch node in the corrected intention derivation path topology, and the calibrated branch credibility weight corresponding to the final retained intention branch node.

[0090] Before executing step S154, which inputs the parameters for generating the output content into the content generation model, the parameters are first analyzed in detail. This parameter is a composite structure encapsulating three core components: the first is the decision-based feature vector corresponding to the final retained intention branch node, with a dimension of D; the second is the depth-level information of the final retained intention branch node in the corrected intention derivation path topology, which is an integer value; and the third is the calibrated branch confidence weight corresponding to the final retained intention branch node, which is a floating-point value between zero and one. The analysis process separates these three parts from the structure and stores them separately in three different registers or memory variables.

[0091] Step S320: Input the output content generation parameters into the input encoding layer of the content generation model. The input encoding layer performs feature concatenation processing on the decision basis feature vector, depth level information and calibrated branch credibility weight in the output content generation parameters to generate the input encoding vector corresponding to each final retained intent branch node.

[0092] The three components parsed in step S310 are simultaneously input into the input encoding layer of the content generation model. This input encoding layer first embeds the scalar integer of the deep-level information into an embedding vector with the same dimension as the decision-based feature vector. Similarly, it maps the scalar floating-point number of the calibrated branch confidence weight into an embedding vector of the same dimension through a linear transformation layer. This results in three vectors with the same dimensions: the original decision-based feature vector, the deep-level embedding vector, and the weight embedding vector. The input encoding layer concatenates these three vectors along the feature dimension to form a 3D fusion vector, which is the input encoding vector corresponding to each final intention-preserving branch node.

[0093] Step S330: Input the input encoding vector corresponding to each final retained intent branch node into multiple parallel decoding branches of the content generation model. Each parallel decoding branch is composed of at least one stacked decoder layer. Different parallel decoding branches share the bottom-level decoder parameters, but the top-level decoder parameters are independent of each other.

[0094] The input encoded vector corresponding to each final retained intent branch node generated in step S320 is copied multiple times and input into multiple parallel decoding branches in the content generation model. The number of these parallel decoding branches is a preset hyperparameter, for example, set to B. Each parallel decoding branch has a basically identical structure, consisting of multiple stacked Transformer decoder layers. The lower layers of these decoding branches, such as the first L_shared layer, share parameters across all branches, while the upper layers of each branch, such as the last L_independent layer, have independent parameters. The input encoded vector first passes through the shared layer and then is split into the independent layers of each branch for subsequent processing.

[0095] Step S340: Each parallel decoding branch independently generates a corresponding initial decoding output sequence based on the input encoding vector it receives. The initial decoding output sequence is composed of a series of sequentially arranged output content unit encodings.

[0096] Within each independent parallel decoding branch, the feature vector processed by an independent layer serves as the initial hidden state of the decoder. The decoder of that branch then begins the autoregressive decoding process. At the first time step, the decoder receives the embedding vector of a special start marker as input, combines it with the probability distribution of the first output content unit from the current hidden state, and samples or selects the unit encoding with the highest probability as the first output. This output unit encoding is then used as the input for the next time step, and the above process is repeated until a special end marker is generated. Ultimately, each parallel decoding branch generates an initial decoded output sequence consisting of a series of output content unit encodings.

[0097] Step S350: Process the initial decoding output sequence generated by each parallel decoding branch to adapt to the input format requirements of the subsequent output content mapping layer, and generate the adapted decoding output sequence.

[0098] Obtain the maximum length L_max of all B branch-generated sequences. For branch sequences with a length less than L_max, pad the end with a specific number of padding markers to make the length also reach L_max. Simultaneously, generate an attention mask matrix with the same shape as the sequence to mark which positions are actual generated content (value 1) and which positions are invalid padding content (value 0). After length unification and mask generation, the initial decoded output sequence of each branch is transformed into the adapted decoded output sequence.

[0099] Step S360: Input each adapted decoded output sequence into the output content mapping layer of the content generation model. The output content mapping layer encodes each output content unit in each adapted decoded output sequence into the corresponding actual output content vocabulary or graphic element encoding according to the preset output content vocabulary.

[0100] The batch data consisting of all B adapted and decoded output sequences is input into the output content mapping layer of the content generation model. This mapping layer is a fully connected neural network followed by a Softmax activation function. For each integer code at each position in the sequence, the mapping layer first maps it to a dense vector through an embedding table, and then computes a probability distribution over the entire output content vocabulary through the fully connected layer and the Softmax layer. The output content vocabulary is a pre-built mapping table containing all possible text words, punctuation marks, and predefined graphic element codes that the system can output. The word or graphic element code with the highest probability from this probability distribution is taken as the actual output at that position. After performing this operation on all positions, the adapted and decoded output sequence of each branch is transformed into a sequence consisting of the actual output content vocabulary or graphic element codes.

[0101] Step S370: Generate an initial text fragment or initial set of graphic elements corresponding to each final retained intent branch node based on the actual output content vocabulary or graphic element encoding, as the initial candidate content in the branch output content candidate set of each final retained intent branch node.

[0102] The actual output vocabulary or graphic element codes in each branch sequence obtained in step S360 are concatenated or combined according to their order in the sequence. For text-based sequences, the vocabulary codes are concatenated to form an initial text fragment. For sequences containing graphic element codes, the graphic element codes and their order relationships are parsed to form an initial set of graphic elements. Each parallel decoding branch generates such an initial candidate content for its corresponding final intention-preserving branch node, so each final intention-preserving branch node now has B different initial candidate contents.

[0103] Step S380: Perform content diversity enhancement processing on the initial candidate content in the branch output content candidate set of each final retained intention branch node, and generate at least one enhanced candidate content with similar content semantics but different expression form based on the initial candidate content through synonym replacement or sentence transformation.

[0104] The initial candidate content generated in step S370 is subjected to data augmentation. For the initial text fragment, a synonym replacement tool based on a pre-trained language model is used to identify key nouns, verbs, and adjectives in the text and randomly selects synonyms from a pre-built thesaurus to replace them, generating one or more semantically similar but word-different augmented text fragments. At the same time, a sentence transformation model is used to convert declarative sentences into rhetorical questions or exclamatory sentences, or to adjust the order of clauses, generating augmented text fragments with different sentence structures. For the initial set of graphic elements, the color, size, position, and other attributes of the elements can be fine-tuned to generate a set of visually slightly different but semantically unchanged augmented graphic elements. The newly generated augmented candidate content is added to the branch output content candidate set of this node.

[0105] Step S390: Merge the initial candidate content and the enhanced candidate content in the branch output content candidate set of each final retained intention branch node to form the complete branch output content candidate set corresponding to each final retained intention branch node.

[0106] The initial candidate content generated in step S370 and the enhanced candidate content generated in step S380 are merged and deduplicated to form a larger candidate set. This set contains multiple possible expressions for the same intent branch node, including both the original generated results and data-enhanced variants.

[0107] Step S3110: Assign a candidate content identifier to each candidate content in the candidate set of each complete branch output content, and associate each candidate content identifier with the node identifier of the final retention intention branch node from which the candidate content originates.

[0108] For each candidate content in the candidate set of output content for each complete branch formed in step S390, a globally unique candidate content identifier, such as a UUID string, is generated. An association record is created in a content mapping database to associate the candidate content identifier with the node identifier of the final reserved intention branch node to which it belongs. This record also contains metadata about the candidate content itself, such as content type, length, creation time, etc.

[0109] Step S410: Extract the node weight coefficient of each final retained intention branch node from the modified intention derivation path topology. The node weight coefficient is calculated based on the calibrated branch confidence weight corresponding to each final retained intention branch node and the number of out-degrees of the final retained intention branch node in the topology.

[0110] The number of out-degrees of each final retained intention branch node, i.e., the number of edges pointing from that node to other nodes, is obtained from the corrected intention derivation path topology and denoted as d_i. The node weight coefficient W_node_i is calculated as follows: W_node_i=(W_cal_i×(1+log(1+d_i))) / (∑_{j=1}^{N}(W_cal_j×(1+log(1+d_j)))), where W_cal_i is the calibrated branch credibility weight calculated in step S148, and N is the total number of final retained intention branch nodes. The node weight coefficient calculated by the above formula considers both the credibility of the branch itself and its pivotal position in the entire path structure, and the sum of all weight coefficients is 1.

[0111] Step S420: Sort all candidate contents in the candidate set of the complete branch output content corresponding to each final retained intention branch node according to their candidate content identifiers, and generate an ordered candidate content list corresponding to each final retained intention branch node.

[0112] For each final retained intention branch node, obtain its corresponding complete set of branch output content candidates. Read the metadata of each candidate content from the content mapping database, especially its quality score. Sort all candidate content according to quality score from highest to lowest to generate an ordered candidate content list, where the first element is the highest quality candidate content, and so on down the list.

[0113] Step S430: Based on the node weight coefficient of each final retention intention branch node, determine the number of candidate contents to be selected from the ordered candidate content list of each final retention intention branch node. The larger the node weight coefficient of the final retention intention branch node, the more candidate contents to be selected.

[0114] Set a total expected candidate content quantity budget Total_N. Allocate the number of content fragments N_i that each node should contribute based on its node weight coefficient W_node_i. The formula for calculating N_i is N_i = round(Total_N × W_node_i), where round represents the floor function. Nodes with larger weight coefficients are allocated a larger number of N_i.

[0115] Step S440: According to the number of candidate contents that need to be selected for each final retention intention branch node, select the candidate contents ranked first from the ordered candidate contents list of each final retention intention branch node in turn, and use them as the set of candidate contents to be merged for each final retention intention branch node.

[0116] Based on the quantity N_i calculated for each node in step S430, the first N_i candidate contents are selected sequentially from the ordered candidate content list generated in step S420 for that node. If the length of the ordered list is less than N_i, all contents in the list are selected. For each final retained intention branch node, a set of candidate contents to be merged consisting of N_i high-quality candidate contents is obtained.

[0117] Step S450: Classify the candidate content in the candidate content set to be merged of each final retained intention branch node according to the content type corresponding to its candidate content identifier, and group the candidate content of the same content type into the same content type group to obtain multiple content type groups.

[0118] Group all selected candidate content together. For each candidate content, iterate through it and query the content mapping database for its content type based on its candidate content identifier, such as a plain text fragment, a structured text list, an icon element, a button element, or an image element. Group candidate content with the same content type into the same group, resulting in multiple content type groups.

[0119] Step S460: Within each content type group, the candidate content within the group is weighted and summed or weighted and spliced ​​according to the node weight coefficient of the final retention intention branch node from which each candidate content originates, to generate the fused content unit corresponding to each content type group.

[0120] Each content type group undergoes independent fusion processing. Taking text groups as an example, if the text fragments are parallel, a weighted concatenation method is used: they are sorted from largest to smallest according to the node weight coefficient W_node_i from which they originate, and then the text fragments are concatenated sequentially. If the text fragments describe different attributes of the same thing, a weighted sum method is used: first, each text fragment is converted into a semantic vector by an encoder; then, the above vectors are weighted and summed, with the weight being W_node_i, resulting in a new fused semantic vector. Finally, a decoder re-decodes this vector into a new text fragment as the fused content unit. For graphic element groups, the weights determine which elements occupy a more central position or a larger size in the layout. Each group outputs a fused content unit after processing.

[0121] Step S470: Arrange the fused content units corresponding to each content type group in sequence according to the logical connection relationship between each final retained intention branch node in the modified intention derivation path topology structure to form a preliminary arranged content unit sequence.

[0122] After obtaining the merged content units grouped by content type, it is necessary to determine their display order in the final output. The system parses the directed edges in the topology structure to generate a topological sort of the nodes. Based on the topological order of the main source nodes of each merged content unit, all merged content units are sorted to form a preliminary sequence of content units.

[0123] Step S480: Insert logical transition words or transition graphic elements between adjacent merged content units in the initially arranged content unit sequence to generate a content unit sequence with added transition elements.

[0124] To make the transitions between content units more natural and smooth, appropriate logical transition elements are inserted between every two adjacent merged content units in the initially arranged sequence. If both content units are text, appropriate transition words such as "therefore," "however," and "in addition" are inserted based on the logical relationship between the nodes they represent. If the first is text and the second is a graphic, a guiding text can be inserted: "As shown in the attached figure XXXX" or "The corresponding operation interface is XXXX." If both are graphics, an arrow or a next-step icon can be inserted between them.

[0125] Step S490: Format the text-type and graphic-type merged content units in the content unit sequence after adding transition elements, so that the font and font size of the text-type merged content units are consistent with the default output style of the intelligent interactive system, and the resolution and color mode of the graphic-type merged content units are matched with the display screen parameters of the intelligent interactive system.

[0126] Iterate through each element in the sequence and determine its type. For text-based content units, set their font type, font size, line spacing, color, and other attributes to uniform values ​​corresponding to the system's default output style. For graphics-based content units, obtain the physical parameters of the current display screen, such as resolution, pixel density, and color mode, and then perform necessary scaling, cropping, and color space conversion on the graphics data to ensure optimal display quality on the current screen.

[0127] Step S500: Combine and encapsulate the formatted text-type merged content units and the graphic-type merged content units into an initial merged interactive response content draft.

[0128] All content units formatted in step S490 are combined into a unified data container according to their order in the sequence. This container can be a rich text object containing both text paragraphs and embedded graphics. All content is then integrated to form a complete preliminary draft of interactive responsive content.

[0129] Step S510: Analyze the logical connection relationship between each final retained intention branch node in the modified intention derivation path topology structure, and obtain at least one logical branch path composed of multiple final retained intention branch nodes in chronological order.

[0130] Traversing the corrected intention deduction path topology, all directed edges are identified to determine the complete path from the starting node to the ending node. Each path consists of a series of nodes representing the final retained intention branches, arranged in logical order. These logical branch paths are extracted, and each path is stored as an ordered list of node identifiers.

[0131] Step S520: Based on the number of logical branch paths and the number of final retained intent branch nodes contained in each logical branch path, divide the content fragments in the initial fused interactive response content draft into at least one content fragment group corresponding to each logical branch path.

[0132] The content fragments in the initial draft are associated with logical branch paths. Each content fragment is assigned to the logical branch path containing its primary source node. If a content fragment's source node exists in multiple paths, its primary path is determined based on context or node weight. After this process, the content fragments in the initial draft are divided into several groups, each corresponding to an independent logical branch path.

[0133] Step S530: For each logical branch path, analyze the semantic coherence between the content fragments within the content fragment group, quantify the degree of semantic difference between each pair of adjacent content fragments, and identify content fragment pairs within the content fragment group whose semantic difference exceeds a preset difference threshold.

[0134] For each logical branch path, the content fragments within a group are sequentially input into a pre-trained semantic similarity calculation model. This model outputs a semantic similarity score between two fragments, with the semantic difference defined as 1 minus the similarity score. The semantic difference between each pair of adjacent content fragments within the group is calculated sequentially. This value is compared to a preset difference threshold T_diff. If the difference between a pair of adjacent fragments exceeds T_diff, it is identified as a content fragment pair requiring optimization.

[0135] Step S540: For content fragment pairs whose semantic jumps exceed a preset jump threshold, insert supplementary explanatory content fragments generated by intermediate intent branch nodes located between the final retained intent branch nodes corresponding to these two content fragments in the logical branch path between the content fragment pairs.

[0136] When a semantic jump is identified between fragment A and fragment B, the node N_A corresponding to fragment A and the node N_B corresponding to fragment B are found in the corresponding logical branch path. All intermediate intent branch nodes between N_A and N_B in the path are searched. From the complete branch output candidate set of these intermediate nodes, the candidate content fragment that best matches the current context is selected as a supplementary explanatory content fragment and inserted between fragment A and fragment B in the initial draft.

[0137] Step S550: After performing the supplementary explanation content fragment insertion operation on all content fragment groups corresponding to the logical branch paths, the content fragment groups are recombined according to the parallel or serial relationship between multiple logical branch paths in the modified intention derivation path topology structure to generate the recombined content fragment sequence.

[0138] After supplementary explanatory fragments have been inserted within all logical branch paths, the aforementioned groups are recombined into a unified sequence based on the relationships between different paths in the topology. If multiple paths are parallel, their group contents can be presented side-by-side, distinguishing different reasoning threads using subheadings or separators. If the paths are sequential, the groups are arranged in order. Based on these relationships, the content fragments of all groups are reintegrated into a single, globally ordered sequence of content fragments.

[0139] Step S560: Perform referential resolution processing on the recombined content fragment sequence, identify pronouns or indicative words appearing in the recombined content fragment sequence, and replace the pronouns or indicative words with the specific intent node name or specific content object name they refer to.

[0140] The referential resolution module is activated to analyze the text content of the entire sequence. This module first identifies all pronouns and indicative words. Then, based on context and semantic analysis, it determines which specific intent node or described content object each pronoun or indicative word refers to. Finally, the pronouns or indicative words are replaced with the specific names they refer to.

[0141] Step S570: Adjust the tone consistency of the recombined content fragment sequence after the referential resolution process, and set the declarative tone, interrogative tone or imperative tone for the entire content fragment sequence according to the overall style attribute of the path topology structure of the modified intention.

[0142] The overall style attributes of the modified intention-derived path topology are obtained from its metadata. For example, directives are primarily imperative, informatives are primarily declarative, and inquisitives are primarily interrogative. Each sentence in the sequence is traversed, and a tone classifier is used to determine its current tone. If the tone of a sentence does not match the overall style attributes, a tone conversion model is invoked to convert the sentence's tone to the target tone.

[0143] Step S580: Input the content segment sequence adjusted for tone consistency into the pre-trained text fluency evaluation model, calculate the fluency score between adjacent content segments in the content segment sequence, and generate a fluency score sequence.

[0144] The entire sequence of content fragments, after tone adjustment, is fed into a pre-trained text fluency evaluation model. For each pair of adjacent sentences or content fragments in the sequence, the model calculates a fluency score; a higher score indicates a more natural transition. This process is repeated for all adjacent pairs to obtain a fluency score sequence that is related to the sequence length.

[0145] Step S590: Identify the content segment positions where the fluency score is lower than the preset fluency threshold based on the fluency score sequence, and automatically add preset fluency transition phrases at the content segment positions.

[0146] Identify the positions of adjacent content fragment pairs whose scores are below a preset fluency threshold T_fluent. For these positions, automatically select a suitable transition phrase from a predefined fluency transition phrase library based on the content and logical relationship between the two fragments (e.g., further, on the other hand), and insert it between the two fragments.

[0147] Step S600: Align the text content and graphic content in the content segment sequence after the fluency optimization to ensure that the layout of the text content surrounding the graphic content conforms to the user interface design specifications of the intelligent interactive system, and output the complete content after the layout alignment as the interactive response content after the coherence optimization.

[0148] The final visual layout of the content fragment sequence, after all language and logic optimizations, is performed. Text and graphic fragments within the sequence are laid out according to user interface design guidelines. The system invokes a layout engine to dynamically calculate the optimal position of each graphic on the page based on its size and the number of lines of text, and adjusts the text's flow layout accordingly. The complete content obtained after layout alignment is the final, coherent, and optimized interactive responsive content.

[0149] Step S610: Perform key information extraction processing on the coherence-optimized interactive response content, extract all intent-related keywords and feature descriptors of all intent-related graphic elements appearing in the coherence-optimized interactive response content, as preliminary core intent expression summary information.

[0150] The optimized interactive response content is parsed into plain text and graphical elements. For the plain text, a keyword extraction tool based on a pre-trained language model is used to identify all keywords strongly related to the intent, forming a list of intent-related keywords. For the graphical elements, feature descriptors are extracted for each element; for example, for icons, their shape feature vector, color histogram features, or they can be mapped to a feature vector using a pre-trained graphical encoding model, forming a list of intent-related graphical element feature descriptors.

[0151] Step S620: Perform word frequency statistics on the intent-related keywords in the preliminary core intent expression summary information, and select the preset number of intent-related keywords with the highest frequency as high-frequency core keywords.

[0152] The keyword list related to intent is processed by frequency statistics. The total number of times each keyword appears in the coherence-optimized interactive response content is counted to obtain the keyword frequency value for each keyword. All keywords are sorted in descending order of frequency value, and the top K keywords in the sorted list are selected as high-frequency core keywords.

[0153] Step S630: Perform cluster analysis on the feature descriptors of intent-related graphic elements in the preliminary core intent expression summary information, classify multiple intent-related graphic elements with feature descriptor similarity exceeding a preset similarity threshold into the same graphic element category, and select the representative graphic element in each graphic element category as the core representative graphic element.

[0154] Cluster analysis is performed on the set of feature descriptors for intent-related graphical elements, calculating the cosine similarity between all pairwise feature descriptors. A preset similarity threshold T_graph_sim is set, and graphical elements corresponding to feature descriptors with similarity exceeding T_graph_sim are grouped into the same graphical element category. For each category, a representative graphical element is selected, with selection criteria including the element with the highest frequency in the category or the element whose feature descriptor is closest to the center point of the category.

[0155] Step S640: Combine and encapsulate the high-frequency core keywords and the core representative graphic elements into the final core intent expression summary information.

[0156] The high-frequency core keyword list obtained in step S620 and the core representative graphic element set obtained in step S630 are combined and encapsulated into a structured data object. This object contains a keyword field that stores a list of K strings, and a graphic element field that stores a list of identifiers representing graphic elements or their compressed feature descriptors.

[0157] Step S650: Retrieve historical interface output data from the interface history record library that corresponds to the historical interface output content before the retrospective time point carried in the retrospective trigger signal.

[0158] Based on the backtracking time point carried in the backtracking trigger signal, the system retrieves and retrieves the historical interface output content data corresponding to that time point from the interface history database. This data is the complete interface content that the system last output to the operator before the operator performed the backtracking operation.

[0159] Step S660: Extract key information from the historical interface output data, which is the same as the interactive response content after coherence optimization. Extract historical high-frequency core keywords and historical core representative graphic elements from the historical interface output data as historical core intent expression summary information.

[0160] The key information extraction and summary generation process is performed on the historical interface output data in a manner completely consistent with steps S610 to S640, resulting in a set of historical high-frequency core keywords and a set of historical core representative graphic elements. These are then combined and encapsulated to form a summary information expressing the historical core intent.

[0161] Step S670: Compare the high-frequency core keywords in the final core intent expression summary information with the historical high-frequency core keywords in the historical core intent expression summary information one by one, and calculate the keyword overlap between the two; calculate the graphic similarity between the core representative graphic elements in the final core intent expression summary information and the historical core representative graphic elements in the historical core intent expression summary information, and generate graphic element similarity scores.

[0162] The final high-frequency core keyword list is intersected with the historical high-frequency core keyword list to obtain the number of keywords C_overlap appearing in both lists. The keyword overlap R_key is calculated as R_key = C_overlap / K. Each element in the final core representative graphic element set is paired with each element in the historical core representative graphic element set to calculate pairwise graphic similarity. The graphic similarity score is obtained by inputting the graphic elements into a pre-trained Siamese neural network and receiving a similarity score between 0 and 1. For each element in the final set, the maximum similarity score between it and all elements in the historical set is taken as the matching score for that element. Then, the average of the matching scores of all final elements is used to obtain the overall graphic element similarity score S_graph.

[0163] Step S680: Calculate the overall intent difference between the final core intent expression summary information and the historical core intent expression summary information based on the keyword overlap and the graphic element similarity score.

[0164] meaning Figure 1 The consistency score C_consistency is calculated as C_consistency = α × R_key + (1 - α) × S_graph, where α is a weighting factor between 0 and 1. The overall intent difference D_diff is calculated as D_diff = 1 - C_consistency.

[0165] Step S690: Compare the overall intent difference with a preset intent correction success threshold. If the overall intent difference is greater than the intent correction success threshold, it is determined that the coherence-optimized interactive response content and the historical interface output content have a clear intent correction relationship. If the overall intent difference is less than or equal to the intent correction success threshold, the instruction to re-execute the intent divergence tracing and node splitting reconstruction operation is triggered.

[0166] A preset threshold T_success for successful intent correction is set, and the overall intent difference D_diff is compared with T_success. If D_diff is greater than T_success, the intent correction is considered successful and verified. If D_diff is less than or equal to T_success, the correction is considered to have failed to effectively resolve the original intent divergence, and the system automatically triggers a re-execution instruction to jump the control flow back to step S130 to restart the intent divergence tracing and node splitting reconstruction process.

[0167] For example, step S710: Obtain the complete data of the interactive response content after verification and optimization, wherein the complete data includes two parts: text content data stream and graphic content data stream.

[0168] After step S690 determines that the verification is successful, the final complete and coherent optimized interactive response content data is obtained. This complete data is clearly divided into two parts: text content data stream and graphic content data stream.

[0169] Step S720: Perform compression encoding on the text content data stream, convert the text content data stream into compressed text data blocks, and assign a first data block identifier to the compressed text data blocks.

[0170] The text content data stream is input into a lossless data compression algorithm module. The Deflate or LZ4 compression algorithm analyzes the original text character sequence, removes redundant information, and encodes it into a more compact binary format, outputting a compressed text data block. After compression, the system generates a globally unique first data block identifier for this data block, such as a UUID string.

[0171] Step S730: Perform image compression processing on the graphic content data stream, convert the graphic content data stream into compressed image data blocks, and assign a second data block identifier to the compressed image data blocks.

[0172] The graphic content data stream is input into an image compression encoding module. Based on the specific format of the graphic content, an appropriate compression algorithm is selected: lossy JPEG compression or lossless PNG compression is used for bitmaps, while SVG text compression or conversion to a more compact binary format is used for vector graphics. After compression, the original graphic data stream is converted into a smaller compressed image data block, and the system generates a globally unique second data block identifier for this block.

[0173] Step S740: Create a data packet container that conforms to the output interface standard of the intelligent interactive system. The data packet container has a predefined container header area and a container payload area.

[0174] Call the system output interface API to create a new empty data packet container instance. The structure of the container fully conforms to the predefined output interface standard of the intelligent interactive system. The standard stipulates that the data packet container must contain two main areas: a container header area and a container payload area.

[0175] Step S750: Write an intent correction identifier field at the beginning of the container head region, and set the field value of the intent correction identifier field to a preset intent correction status code.

[0176] Begin filling the header area of ​​the packet container. At the very beginning of the header area, write a specific intent correction identifier field. The value of this field is set to a preset intent correction status code, such as a fixed 16-bit integer 0x5A5A.

[0177] Step S760: Write the total data packet length field in the container header area immediately after the random image correction identifier field. The value of the total data packet length field is calculated based on the total data size of the compressed text data block and the compressed image data block.

[0178] Write the total packet length field after the intent to correct identifier field. This field is an integer whose value is obtained by calculating the fixed length of the container header area itself, the length of the compressed text data block, the length of the compressed image data block, and finally the sum of any other possible padding bytes.

[0179] Step S770: Write the first data block identifier and the second data block identifier in the container header area, and write the starting storage offset and data block length information of the corresponding data block in the container load area after each data block identifier.

[0180] Continue filling the header area by writing the first data block identifier and the second data block identifier in sequence. After each identifier, write two description fields: the first is the starting storage offset, which indicates the starting storage position of the corresponding data block in the container load area in bytes; the second is the data block length, which indicates the actual size of the data block in bytes.

[0181] Step S780: Write the compressed text data block and the compressed image data block sequentially into the storage location specified by the starting storage offset in the container load area. Perform integrity verification calculation on the entire data packet after writing, generate a data packet checksum field, and write the data packet checksum field to the end of the container header area.

[0182] According to the starting storage offset set in step S770, the binary data of the compressed text data block and the compressed image data block are precisely written to the corresponding positions in the payload area. After all data blocks are written, integrity verification calculations are performed on the entire data packet, including the header and payload. For example, a fixed-length data packet checksum field is generated using a cyclic redundancy check algorithm, and finally, this checksum field is written to the end of the container header area.

[0183] Step S790: Output the complete data packet with the intent correction identifier field, total data packet length field, data block identifier, starting storage offset, data block length information, and data packet checksum field added, as the final encapsulated data packet conforming to the output interface standard of the intelligent interactive system, and send it to the output interface driver of the intelligent interactive system. The output interface driver parses the compressed text data block and compressed image data block according to the information in the header area of ​​the data packet container, and transmits them to the user interface rendering module for final output display.

[0184] The complete data packet is sent to the output interface driver of the intelligent interactive system via the system kernel's inter-process communication mechanism. Upon receiving the original data packet byte stream, the driver confirms complete reception based on the total packet length field, recalculates the checksum, and compares it with the checksum field at the end of the data packet to verify that the data was not corrupted during transmission. After successful verification, the driver parses the header area to obtain the identifiers of the first and second data blocks, their offsets, and lengths. Based on this information, it precisely segments compressed text data blocks and compressed image data blocks from the payload area. The compressed data is decompressed to recover the original text and image content data streams. Finally, the decompressed text and image data are passed to the user interface rendering module, which uses this data to draw the final, clear interactive response content on the screen for the operator.

[0185] Figure 2 This illustration shows an intelligent interaction intent alignment and correction system 100 for human-machine collaboration provided in an embodiment of this application. It includes a processor 1001, a memory 1003, and program code stored in the memory 1003. The processor 1001 executes the program code to implement the steps of the intelligent interaction intent alignment and correction method for human-machine collaboration. The processor 1001 and the memory 1003 are connected, for example, via a bus 1002. Optionally, the intelligent interaction intent alignment and correction system 100 for human-machine collaboration may further include a transceiver 1004. The transceiver 1004 can be used for data interaction between this intelligent interaction intent alignment and correction system for human-machine collaboration and other intelligent interaction intent alignment and correction systems for human-machine collaboration, such as sending and / or receiving data. It should be noted that in actual scheduling, the transceiver 1004 is not limited to one, and the structure of this intelligent interaction intent alignment and correction system 100 for human-machine collaboration does not constitute a limitation on the embodiments of this application. The memory 1003 is used to store the program code for executing the embodiments of this application, and its execution is controlled by the processor 1001. The processor 1001 is used to execute program code stored in the memory 1003 to implement the steps shown in the foregoing method embodiments.

[0186] This application provides a computer-readable storage medium storing program code, which, when executed by a processor, can implement the steps and corresponding content of the aforementioned method embodiments.

[0187] The above description is only an optional implementation method for some implementation scenarios of this application. It should be noted that for those skilled in the art, other similar implementation methods based on the technical concept of this application, without departing from the technical concept of this application, also fall within the protection scope of the embodiments of this application.

Claims

1. A method for human-robot collaborative intelligent interaction intention alignment and correction, characterized in that, The method includes: During the process of the operator performing human-computer interaction tasks, the unexpected interface backtracking operation performed by the operator on the current interface content output by the intelligent interaction system is captured, and a backtracking trigger signal containing the backtracking time point and the backtracking target interface identifier is generated. The system synchronously acquires historical intent deduction path information generated by the intelligent interaction system before the retrospective time point. The historical intent deduction path information includes multiple sequentially connected intent nodes and a decision basis feature vector corresponding to each intent node. The intent divergence tracing module is activated to perform feature difference analysis on the historical output content corresponding to the target interface identifier and the subsequent operation input content collected after the time point of the backtracking, and generate an intent divergence feature distribution map describing the root cause of intent divergence. Based on the intention divergence feature distribution map, a node splitting and reconstruction process is performed on a specific intention node in the historical intention derivation path information to generate a corrected intention derivation path topology containing multiple candidate intention branches. The modified intent deduction path topology is loaded into the online decision engine of the intelligent interaction system, and the online decision engine regenerates interactive response content that matches the operator's true intent based on the modified intent deduction path topology.

2. The method of claim 1, wherein, The activation intent divergence tracing module performs feature difference analysis on the historical output content corresponding to the target interface identifier and the subsequent operation input content collected after the backtracking time point, generating an intent divergence feature distribution map describing the root cause of intent divergence, including: The target interface identifier in the backtracking trigger signal is parsed, and the complete historical interface output content data corresponding to the target interface identifier is retrieved from the interface history database according to the target interface identifier. Extract the sequence of subsequent operation input content collected within a preset time window after the retrospective time point from the operation log database. The sequence of subsequent operation input content includes at least one continuous touch operation trajectory data input by the operator. The complete historical interface output content data is subjected to content structure parsing processing to extract the core semantic feature vector of each output content unit contained in the complete historical interface output content data. The subsequent operation input content sequence is subjected to sequence feature encoding processing, and each continuous touch operation trajectory data in the subsequent operation input content sequence is converted into a corresponding operation trajectory geometric feature vector. All operation trajectory geometric feature vectors are stacked to form an operation behavior temporal feature matrix. The core semantic feature vector and the temporal feature matrix of the operation behavior are input into a pre-trained cross-modal feature alignment network for feature space mapping and alignment processing. This generates aligned feature representations of the historical output content and the subsequent operation input content in the unified feature space. The difference between the aligned feature representations is calculated to generate an initial feature difference representation, which is used to describe the degree of difference between the historical output content and the subsequent operation input content in the unified feature space. The initial feature difference vector is subjected to saliency enhancement processing to identify feature dimensions in the initial feature difference vector whose feature value difference exceeds a preset saliency threshold. The feature dimensions are marked as key feature dimensions of intent divergence. Based on the key feature dimensions of intent divergence, the output content unit region associated with each key feature dimension of intent divergence is located in the output content data of the complete history interface in reverse. The feature-region mapping relationship between each key feature dimension of intent divergence and the output content unit region is generated. Based on the key feature dimension of intent divergence, the operation trajectory segment associated with each key feature dimension of intent divergence is located in reverse in the subsequent operation input content sequence, and the feature-trajectory mapping relationship between each key feature dimension of intent divergence and the operation trajectory segment is generated. The feature-region mapping relationship and the feature-trajectory mapping relationship are associated and fused to construct an intent divergence feature distribution map with the key feature dimension of intent divergence as nodes and the feature-region mapping relationship and the feature-trajectory mapping relationship as node attributes.

3. The method of claim 1, wherein, The step of performing node splitting and reconstruction processing on specific intent nodes in the historical intent derivation path information based on the intent divergence feature distribution map to generate a corrected intent derivation path topology containing multiple candidate intent branches includes: The historical intent derivation path information is parsed to obtain the node identifier, node generation timestamp, and decision basis feature vector of each intent node contained in the historical intent derivation path information. The feature representation of each key feature dimension of intent divergence is extracted from the intent divergence feature distribution map, and the feature matching degree is calculated with the decision basis feature vector of each intent node to generate a feature matching degree score matrix between the feature representation of each key feature dimension of intent divergence and the decision basis feature vector of each intent node. Based on the feature matching score matrix, identify intent nodes whose matching degree with the key feature dimension of intent divergence exceeds a preset matching degree threshold, and mark the intent nodes as intent nodes to be split. For each intent node to be split, the decision basis feature vector of the intent node to be split is decomposed according to the set of key feature dimensions of intent divergence matched with it, and the decision basis feature vector of the intent node to be split is decomposed into sub-feature vector components corresponding to each key feature dimension of intent divergence. Based on the sub-feature vector components corresponding to the key feature dimension of each intention divergence, at least two initial split candidate intention branch nodes are generated for each intention node to be split. The decision basis feature vector of each initial split candidate intention branch node is composed of the corresponding sub-feature vector components. An initial branch credibility weight is assigned to each initial split candidate intention branch node. The initial branch credibility weight is calculated based on the difference weight coefficient of the intention divergence key feature dimension corresponding to the initial split candidate intention branch node in the intention divergence feature distribution map. Semantic similarity clustering analysis is performed on the initial split candidate intention branch nodes to calculate the similarity of decision basis feature vectors between different initial split candidate intention branch nodes. Multiple initial split candidate intention branch nodes whose decision basis feature vector similarity exceeds a preset similarity threshold are merged into the same aggregated intention branch node. Based on the initial branch credibility weight, the credibility of the merged aggregated intent branch nodes and the initial split candidate intent branch nodes that did not participate in the merge are recalibrated to generate a calibrated branch credibility weight for each remaining candidate intent branch node. Candidate intent branch nodes whose calibrated branch confidence weight is lower than the preset retention threshold are removed, and candidate intent branch nodes whose calibrated branch confidence weight is higher than the preset retention threshold are retained as the final retained intent branch nodes. Based on the original connection relationship between each intent node in the historical intent derivation path information and the derivation relationship between the final retained intent branch node and the original intent node, a corrected intent derivation path topology structure containing the final retained intent branch node and its logical inheritance relationship is constructed.

4. The method for aligning and correcting intelligent interactive intentions in human-computer collaboration according to claim 1, characterized in that, The step of loading the modified intent deduction path topology into the online decision engine of the intelligent interaction system, and having the online decision engine regenerate interactive response content matching the operator's true intent based on the modified intent deduction path topology, includes: The modified intent derivation path topology is converted into a structured decision path description file that can be recognized by the online decision engine. The structured decision path description file contains the node identifier of each final retained intent branch node, the calibrated branch credibility weight corresponding to each final retained intent branch node, and the encoding of the logical connection relationship between each final retained intent branch node. The structured decision path description file is injected into the runtime context of the online decision engine to replace the historical intent derivation path information currently cached by the online decision engine and associated with the traceback target interface identifier. The online decision engine parses the structured decision path description file, determines the starting intent branch node to be activated based on the logical succession relationship encoding in the structured decision path description file, and starts from the starting intent branch node. The online decision engine traverses each final retained intent branch node in the corrected intent derivation path topology, and generates output content generation parameters corresponding to each final retained intent branch node based on the calibrated branch credibility weight and the decision basis feature vector of each final retained intent branch node. The output content generation parameters corresponding to each final retained intention branch node are input into the content generation model built into the online decision engine. The content generation model generates a branch output content candidate set corresponding to each final retained intention branch node according to the output content generation parameters. The branch output content candidate set is fused. The content elements in the branch output content candidate set are weighted and merged according to the weight coefficient of each final retained intention branch node in the modified intention derivation path topology structure to generate an initial fused interactive response content draft. The initial fused interactive response content draft is subjected to content coherence optimization processing. According to the logical connection relationship between each final retained intent branch node in the modified intent derivation path topology, the arrangement order and connection method of different content segments in the initial fused interactive response content draft are adjusted to generate coherent optimized interactive response content. Core intent expression summary information is extracted from the coherent optimized interactive response content, and the core intent expression summary information is compared and verified with the historical interface output content before the backtracking time point carried in the backtracking trigger signal. The verified coherent optimized interactive response content is encapsulated into a data packet format that conforms to the output interface standard of the intelligent interaction system, and an intent correction identifier field is added to the header of the data packet format. The data packet with the intent correction identifier field added is sent to the user interface rendering module of the intelligent interaction system, and the user interface rendering module generates the final interactive response content displayed on the user interface based on the content data in the data packet.

5. The method for aligning and correcting intelligent interactive intentions in human-computer collaboration according to claim 1, characterized in that, The method further includes: After the modified intent derivation path topology is loaded into the online decision engine, an intent alignment status visualization indicator is generated synchronously in the preset monitoring area on the user interface of the intelligent interaction system. The intent alignment status visualization indicator is used to present the summary information of the modified intent derivation path topology currently loaded by the online decision engine in real time. The system continuously monitors the selective focusing operation performed by the operator in the area where the intention alignment status visualization indicator is located. When the operator performs a zoom-in gesture on the intention alignment status visualization indicator, the system dynamically controls the switching or loading of the visualization interface based on the gesture parameters of the zoom-in gesture, switching from displaying the summary information of the corrected intention derivation path topology to displaying the detailed topology diagram of the corrected intention derivation path topology. Once the detailed topology diagram is fully expanded, a visual progress bar for the calibrated branch credibility weight corresponding to each final retained intention branch node is simultaneously generated near the display area of ​​each final retained intention branch node in the detailed topology diagram. Simultaneously, the node selection operation performed by the operator on the detailed topology diagram is captured. When the operator selects any final retention intention branch node in the detailed topology diagram through touch operation, a simplified summary information of the decision basis feature vector of the selected final retention intention branch node is displayed in a floating information window on the user interface. A branch confirmation control and a branch rejection control are generated simultaneously in the floating information window. When the operator triggers the branch confirmation control, the credibility weight of the selected branch node with the final intention to retain is increased by a preset weight adjustment step value. When the operator triggers the branch rejection control, the calibrated branch credibility weight of the selected final retention intention branch node is reduced by a preset weight adjustment step value, and the display length of the corresponding visual progress bar in the detailed topology diagram is updated in real time. A global intent confirmation control is generated in the edge area of ​​the user interface. When the operator triggers the global intent confirmation control, the real-time adjustment function of the calibrated branch confidence weight of all final retained intent branch nodes in the detailed topology diagram is frozen. The detailed topology diagram at the current moment and the final calibrated branch confidence weight of all final retained intent branch nodes contained therein are encapsulated as an intent alignment snapshot after operator confirmation. The intent alignment snapshot confirmed by the operator is stored in the intent alignment history database of the intelligent interaction system, and a unique timestamp and operator identifier are assigned to the intent alignment snapshot. Based on the final calibrated branch credibility weights of each final retained intent branch node recorded in the intent alignment snapshot, the model parameters of the pre-trained content generation model in the online decision engine are fine-tuned and updated.

6. The method of claim 4, wherein, The step involves inputting the output content generation parameters corresponding to each final retained intent branch node into the content generation model built into the online decision engine. The content generation model then generates a candidate set of branch output content for each final retained intent branch node based on these output parameters, including: The output content generation parameters corresponding to each final retained intention branch node are analyzed. The output content generation parameters include the decision basis feature vector corresponding to the final retained intention branch node, the depth level information of the final retained intention branch node in the corrected intention derivation path topology, and the calibrated branch credibility weight corresponding to the final retained intention branch node. The output content generation parameters are input into the input encoding layer of the content generation model. The input encoding layer performs feature concatenation processing on the decision basis feature vector, depth level information and calibrated branch credibility weight in the output content generation parameters to generate the input encoding vector corresponding to each final retained intention branch node. The input encoding vector corresponding to each final retained intent branch node is input into multiple parallel decoding branches of the content generation model. Each parallel decoding branch is composed of at least one stacked decoder layer. Different parallel decoding branches share the bottom decoder parameters, but the top decoder parameters are independent of each other. Each parallel decoding branch independently generates a corresponding initial decoding output sequence based on the input encoding vector it receives. The initial decoding output sequence is composed of a series of sequentially arranged output content units encoded. The initial decoding output sequence generated by each parallel decoding branch is processed to adapt to the input format requirements of the subsequent output content mapping layer, and an adapted decoding output sequence is generated. Each adapted and decoded output sequence is input into the output content mapping layer of the content generation model. The output content mapping layer encodes each output content unit in each adapted and decoded output sequence into the corresponding actual output content vocabulary or graphic element encoding according to a preset output content vocabulary. Based on the actual output content vocabulary or graphic element encoding, generate the initial text fragment or initial graphic element set corresponding to each final retained intention branch node, as the initial candidate content in the branch output content candidate set of each final retained intention branch node; For each branch output content candidate set of the final retained intention branch node, perform content diversity enhancement processing on the initial candidate content, and generate at least one enhanced candidate content with similar content semantics but different expression form based on the initial candidate content through synonym replacement or sentence transformation; The initial candidate content and the enhanced candidate content in the branch output content candidate set of each final retained intention branch node are merged to form the complete branch output content candidate set corresponding to each final retained intention branch node; Assign a candidate content identifier to each candidate content in the candidate set of output content for each complete branch, and associate each candidate content identifier with the node identifier of the final retention intention branch node from which the candidate content originates.

7. The method of claim 4, wherein, The process of fusing the candidate set of branch output content involves weighting and merging the content elements in the candidate set of branch output content according to the weight coefficient of each final retained intent branch node in the modified intent derivation path topology, to generate an initial fused interactive response content draft, including: The node weight coefficient of each final retained intention branch node is extracted from the modified intention derivation path topology. The node weight coefficient is calculated based on the calibrated branch confidence weight corresponding to each final retained intention branch node and the number of out-degrees of the final retained intention branch node in the topology. Sort all candidate contents in the candidate set of the complete branch output content corresponding to each final retained intention branch node according to their candidate content identifiers to generate an ordered candidate content list corresponding to each final retained intention branch node. Based on the node weight coefficient of each final retention intention branch node, determine the number of candidate contents to be selected from the ordered candidate content list of each final retention intention branch node. The larger the node weight coefficient of the final retention intention branch node, the more candidate contents to be selected. Based on the number of candidate contents that need to be selected for each final retention intention branch node, the candidate contents ranked first are selected sequentially from the ordered candidate contents list of each final retention intention branch node, and this is used as the set of candidate contents to be merged for each final retention intention branch node. The candidate content in the set of candidate content to be merged of each final retention intention branch node is classified according to the content type corresponding to its candidate content identifier. Candidate content with the same content type is grouped into the same content type group, resulting in multiple content type groups. Within each content type group, the candidate content within the group is weighted and summed or weighted and concatenated according to the node weight coefficient of the final retention intention branch node from which each candidate content originates, to generate the fused content unit corresponding to each content type group. The fused content units corresponding to each content type group are arranged sequentially according to the logical connection relationship between each final retained intention branch node in the modified intention derivation path topology structure to form a preliminary arranged content unit sequence. Logical transition words or transition graphic elements are inserted between adjacent merged content units in the initially arranged content unit sequence to generate a content unit sequence with added transition elements. The text-type and graphic-type merged content units in the content unit sequence after adding transition elements are formatted separately to ensure that the font and size of the text-type merged content units are consistent with the default output style of the intelligent interactive system, and that the resolution and color mode of the graphic-type merged content units are matched with the display screen parameters of the intelligent interactive system. The formatted text-based content units and the graphic-based content units are combined and encapsulated into an initial draft of the merged interactive response content.

8. The method of claim 4, wherein, The initial fusion of the interactive response content draft is subjected to content coherence optimization processing. Based on the logical connection relationship between each ultimately retained intent branch node in the corrected intent derivation path topology, the arrangement order and connection method of different content segments in the initial fusion of the interactive response content draft are adjusted to generate coherent optimized interactive response content, including: Analyze the logical connection relationship between each final retained intention branch node in the topology of the modified intention derivation path to obtain at least one logical branch path composed of multiple final retained intention branch nodes in chronological order. Based on the number of logical branch paths and the number of final retained intent branch nodes contained in each logical branch path, the content fragments in the initial fused interactive response content draft are divided into at least one content fragment group corresponding to each logical branch path. For each logical branch path, the semantic coherence between the content fragments within the content fragment group is analyzed, the degree of semantic difference between each pair of adjacent content fragments is quantified, and the content fragment pairs whose semantic difference exceeds a preset difference threshold are identified. For content fragment pairs whose semantic jumps exceed a preset jump threshold, a supplementary explanatory content fragment generated by an intermediate intent branch node located between the final retained intent branch nodes corresponding to these two content fragments is inserted between the content fragment pairs. After performing the supplementary explanation content fragment insertion operation on all content fragment groups corresponding to the logical branch paths, each content fragment group is recombined according to the parallel or serial relationship between multiple logical branch paths in the modified intention derivation path topology structure to generate a recombined content fragment sequence. The recombined content fragment sequence is subjected to referential resolution processing to identify pronouns or indicative words appearing in the recombined content fragment sequence, and the pronouns or indicative words are replaced with the specific intent node name or specific content object name they refer to; The tone consistency of the recombined content fragment sequence after the dereference resolution process is adjusted, and the declarative tone, interrogative tone, or imperative tone is uniformly set for the entire content fragment sequence according to the overall style attribute of the path topology structure of the modified intention. The content segment sequence adjusted for tone consistency is input into a pre-trained text fluency evaluation model to calculate the fluency score between adjacent content segments in the content segment sequence and generate a fluency score sequence. Based on the fluency score sequence, identify the content segment locations where the fluency score is lower than a preset fluency threshold, and automatically add preset fluency transition phrases at the content segment locations; The text and graphic content in the content fragment sequence after smoothness optimization are aligned to ensure that the layout of the text content surrounding the graphic content conforms to the user interface design specifications of the intelligent interaction system. The complete content after alignment is output as the interactive response content after coherence optimization.

9. The method of claim 4, wherein, The step of extracting core intent expression summary information from the coherence-optimized interactive response content and comparing and verifying the core intent expression summary information with the historical interface output content before the backtracking time point carried in the backtracking trigger signal includes: The coherence-optimized interactive response content is subjected to key information extraction processing. All intent-related keywords and feature descriptors of all intent-related graphic elements appearing in the coherence-optimized interactive response content are extracted as preliminary core intent expression summary information. The frequency of intent-related keywords in the preliminary core intent expression summary information is statistically analyzed, and the preset number of intent-related keywords with the highest frequency are selected as high-frequency core keywords. Cluster analysis is performed on the feature descriptors of intent-related graphic elements in the preliminary core intent expression summary information. Multiple intent-related graphic elements with feature descriptor similarity exceeding a preset similarity threshold are grouped into the same graphic element category, and a representative graphic element in each graphic element category is selected as the core representative graphic element. The high-frequency core keywords and the core representative graphic elements are combined and encapsulated into the final core intent expression summary information; Retrieve historical interface output data from the interface history library that corresponds to the historical interface output content before the backtracking time point carried in the backtracking trigger signal; The historical interface output data is processed to extract key information in the same way as the interactive response content after the coherence optimization. The historical high-frequency core keywords and historical core representative graphic elements in the historical interface output data are extracted as historical core intent expression summary information. The high-frequency core keywords in the final core intent expression summary information are compared one by one with the historical high-frequency core keywords in the historical core intent expression summary information, and the keyword overlap between the two is calculated; the core representative graphic elements in the final core intent expression summary information are compared with the historical core representative graphic elements in the historical core intent expression summary information to generate graphic element similarity scores. The overall intent difference between the final core intent expression summary information and the historical core intent expression summary information is calculated based on the keyword overlap and the graphic element similarity score. The overall intent difference is compared with a preset intent correction success threshold. If the overall intent difference is greater than the intent correction success threshold, it is determined that the coherence-optimized interactive response content and the historical interface output content have a clear intent correction relationship. If the overall intent difference is less than or equal to the intent correction success threshold, the instruction to re-execute the intent divergence tracing and node splitting reconstruction operation is triggered.

10. A human-in-the-loop intelligent interactive intent alignment and correction system, comprising: The method includes a processor and a computer-readable storage medium storing machine-executable instructions that, when executed by the processor, implement the intelligent interaction intent alignment and correction method for human-machine collaboration as described in any one of claims 1-9.

Images