Information guide identification and form filling system and method

Abstract

The invention discloses an information guide identification and form filling system and method, belongs to the technical field of intelligent information interaction and data processing, and aims to solve the problems of fuzzy demand positioning, fragmentation of information acquisition, low form filling efficiency and insufficient data credibility in traditional information interaction. After initial basic information of a user is collected, a question and answer association graph is constructed through a graph neural network, the initial information is mapped to graph generation node association features, and a guide label is generated through clustering analysis; matching a core guiding main line based on guiding labels and node weights, dividing interaction stages, generating a verbal skill construction guiding plan, starting self-adaptive question and answer to construct an original demand pool, and verifying and integrating the original demand pool into an initial demand information set; and matching a multi-modal acquisition target to generate a standardized data set, correcting structured information to position a target demand, matching a scale template, and realizing intelligent form filling through information quantitative conversion and deletion completion. According to the method, the cross-scene information interaction efficiency and the data availability are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent information interaction and data processing technology, and more specifically to an information guidance and identification, form filling system and method. Background Technology

[0002] In scenarios where user needs need to be clarified and data recording completed through information interaction, existing technologies generally lack deep adaptation to scenario characteristics and end-to-end collaborative optimization, resulting in significant shortcomings in practical applications. On the one hand, information processing lacks specificity. Whether it's the selection of dimensions for initial information collection or the basis for subsequent needs analysis, the core requirements of the scenario are not fully considered, often resulting in redundant information or missing key information. Furthermore, the colloquial and fragmented content expressed by users is difficult to directly convert into usable data, increasing the difficulty of subsequent processing. On the other hand, the efficiency and accuracy of the interaction and form-filling stages are insufficient. The interaction process often follows a fixed procedure, ignoring differences in user needs, easily leading to indiscriminate and redundant questions. Moreover, it is difficult to flexibly adjust to ambiguous user responses, resulting in a poor user experience and low efficiency. The form-filling stage either relies on manual input, which is inefficient and prone to errors, or uses generic templates with field designs that do not match the scenario requirements. Missing entries are often filled in with simple statistics, which are out of touch with actual user needs. The resulting data lacks standardization and credibility, failing to provide reliable support for subsequent scenario services and hindering the overall improvement of scenario service quality. Therefore, in order to overcome these limitations, this invention proposes an information-guided identification and form-filling system and method. Summary of the Invention

[0003] To address the shortcomings of existing technologies, the present invention aims to provide an information-guided identification and form-filling system and method, solving the problem of accurately acquiring user needs in different target scenarios, reducing invalid information interaction, and achieving structured processing of user information and efficient and accurate filling of question-and-answer interaction questionnaires. To achieve the above objectives, the present invention provides the following technical solution:

[0004] An information-guided identification and form-filling method, comprising:

[0005] The system collects initial basic information about users, constructs a question-and-answer association graph through a graph neural network, and quantifies the association weights between nodes in the question-and-answer association graph through a graph attention mechanism. The initial basic information is then mapped to the corresponding nodes in the question-and-answer association graph to generate node association features for users. These features are used to classify users into cluster categories through cluster analysis and generate user guidance tags.

[0006] Based on the quantitative association weight of the association features of the guidance tags and nodes, the core guidance main line is matched from the question and answer association graph, the information interaction stage is divided and the information collection points and collection targets are set to generate guidance scripts, so as to build a user information interaction guidance plan, which is used to start adaptive question and answer interaction, build the original demand information pool, generate structured units, and build the user initial demand information set through verification and integration.

[0007] Based on the initial set of user needs information, a multimodal information collection target is matched to generate a standardized multimodal scene information dataset, which is used to match and correct structured units in order to locate the target user needs.

[0008] Based on the needs of target users, a question-and-answer interaction scale template is matched, and the template is intelligently filled in through information quantification and intelligent missing completion.

[0009] Specifically, the steps for constructing a question-and-answer relationship graph include:

[0010] Based on the target scenario matching information collection dimensions, initial basic user information is collected; the target scenario refers to a specific application scenario that requires obtaining user needs through targeted information interaction and providing personalized guidance to realize the scenario function.

[0011] In the target scenario, a question-answering association graph containing scenario association dimensions, user needs, and target information is constructed using a graph neural network. The question-answering association graph is used to transform basic user information, user needs, and solution association logic in the target scenario into computable structured knowledge.

[0012] By using graph attention mechanism, combining the question-and-answer logic of the target scenario with historical interaction data, the association weights between nodes in the question-and-answer association graph are quantified through linear transformation of node feature vectors, attention score calculation and normalization.

[0013] Specifically, the steps for generating a user's onboarding tag include:

[0014] The collected initial basic information of users is parsed and mapped to the corresponding nodes in the question-and-answer association graph. The weight attributes of each mapped node are extracted and integrated into structured node association features.

[0015] Using node association features as input, cluster analysis is used to divide users into cluster categories, extract key scene features and user demand tendencies corresponding to the cluster categories, and generate guidance tags that are adapted to user demand tendencies.

[0016] Guiding tags are scenario-adaptive demand identifiers generated based on the key features of the scenario and user demand tendencies corresponding to the cluster category. They are used to clarify the priority of question-and-answer interactions and the core guidance direction for each type of user in the target scenario.

[0017] Specifically, the steps for developing a user information interaction guidance plan include:

[0018] The node association feature intensity matrix output by the graph attention mechanism is quantized, and the generated guide label is semantically encoded to generate a guide label semantic vector.

[0019] Calculate the semantic correlation between the semantic vector of the guide label and the feature vectors of all nodes in the intensity matrix, set the filtering ratio, sort the semantic correlation, and obtain the set of key node relationships based on the filtering ratio;

[0020] For key node relationship sets, a weighted directed graph path planning algorithm is adopted, with the association weight between nodes as the path edge weight; at the same time, the priority of the guidance label is incorporated, and a greedy strategy is used to carry out path search and construct the core guidance main line.

[0021] The nodes of the question-and-answer association graph are labeled at the node level. A hierarchical clustering algorithm is used to divide the information interaction stages of the core guiding line and set the preset information collection points and collection targets for each information interaction stage.

[0022] Based on a general large language model, we access the target scenario domain corpus and carry out incremental pre-training. For the information collection points and collection targets of each information interaction stage, we construct input text containing guidance label content, collection target nodes and scenario constraints to generate guidance dialogue.

[0023] Integrate the guiding scripts for each stage of information interaction to build a user information interaction guidance plan.

[0024] Specifically, the steps for constructing the original demand information pool and generating structured units include:

[0025] Initiate adaptive question-and-answer interaction, receive users' spoken language responses in real time, perform semantic understanding, extract explicit and implicit demand information from the response content, and build a pool of original demand information.

[0026] The information interaction state transition matrix of the pre-trained Hidden Markov Model is invoked. The information interaction state transition matrix takes the information collection point and the collection target as the core correlation dimension and takes the information interaction stage and the user's implicit demand tendency as the dual state dimension. It quantifies the joint transition probability of stage adjustment and information collection point completion triggered by changes in the user's implicit demand tendency under different information interaction stages.

[0027] When an information collection point fails to achieve its collection target, the current information interaction stage is dynamically adjusted based on the joint transition probability output by the information interaction state transition matrix, and supplementary guidance scripts are generated to match the information collection point that has failed to achieve its target.

[0028] The original demand information pool is annotated in three dimensions, including scenario dimensions, demand attributes, and preference details, to construct structured units.

[0029] Specifically, the steps for constructing a set of initial user requirement information include:

[0030] It calls the association rule library of target scenarios and user needs. The association rule library has built-in association criteria for information collection points, collection targets and guidance tags in each target scenario.

[0031] The structured units are validated based on the association rule base, the non-compliant information collection points are identified, the guiding scripts for the non-compliant information collection points are generated, and the adaptive question and answer interaction is carried out again.

[0032] The structured units are associated and matched with the corresponding information interaction stages, information collection points, collection targets and guiding labels, and the initial user demand information set is constructed by integration.

[0033] Specifically, the steps for matching and modifying structured units include:

[0034] The system analyzes the initial set of user needs information, extracts the collection targets of structured units, and matches multimodal information collection targets based on the node association weights between the collection targets and target information in the question-and-answer association graph and the priority of scenario requirements.

[0035] Based on the information interaction state transition matrix of the Hidden Markov Model, a scenario-based multimodal data acquisition guidance scheme is generated to guide users in real time to complete multimodal information acquisition that conforms to the target scenario specifications.

[0036] Feature extraction is performed on the collected multimodal information, and the feature vectors of the modal information are integrated to construct a standardized multimodal scene information dataset;

[0037] The basic matching degree between the structured unit and each modal feature vector is calculated. Then, based on the cross-modal attention mechanism, dynamic attention weights are assigned according to the different modal features of the target scene to calculate the comprehensive matching degree of each structured unit.

[0038] The overall matching degree is compared with the preset matching degree threshold of the target scene. If the overall matching degree is greater than the matching degree threshold, the initial requirement information of the structured unit is determined to be accurate; otherwise, a second adaptive question-and-answer interaction is triggered to correct the structured unit.

[0039] Specifically, the steps to identify user target needs include:

[0040] Structured units are associated and bound with standardized multimodal scene information datasets to form a chain of evidence for user needs, and the source and corresponding question-and-answer association graph node association weights are labeled for each element of the evidence chain.

[0041] Based on the quantitative association weights between structured units and user demand nodes within the question-and-answer association graph, potential user demand directions are screened.

[0042] Based on the pre-defined priority rules for the target scenario, and combined with the association weights of the question-and-answer association graph nodes corresponding to the potential demand directions, the comprehensive priority score of each potential user demand direction is calculated by using the priority coefficients set for each user demand direction in the pre-defined priority rules for the target scenario and the association weights of the question-and-answer association graph nodes corresponding to the potential demand directions, in order to select potential demand directions as target user demands.

[0043] Specifically, the steps for intelligently filling in the question-and-answer interaction scale template include:

[0044] Based on the target user's needs, locate the question-and-answer interaction scale template that matches the target needs dimension in the target scenario needs rule base;

[0045] The collection targets of the structured units are transformed into standardized content or quantitative scores required by the question-and-answer interaction scale through scene-adaptive information mapping algorithms and associated weight calibration mechanisms, and then filled into the corresponding question-and-answer interaction scale items.

[0046] The feature vectors of the standardized multimodal scene information dataset are transformed into scene descriptive information. By comparing the core field set and data filling specifications of the question-and-answer interaction scale, the scene descriptive information is matched to the corresponding question-and-answer interaction scale items through a multimodal feature and scale field cross-modal mapping mechanism.

[0047] Identify missing items in the question-and-answer interaction scale, obtain historical scales of similar target user needs archived in the target scenario requirement rule base, filter reference cases based on the similarity of the question-and-answer interaction scale items, extract the valid fill content of the corresponding missing items in the reference cases as target fill information, complete the missing items, and generate the question-and-answer interaction scale.

[0048] An information-guided identification and form-filling system includes: an initial information processing module, an interactive guidance module, a demand positioning module, and an intelligent form-filling module;

[0049] The initial information processing module collects basic user information based on the target scenario and generates guidance tags; the interaction guidance module constructs a core guidance thread based on the guidance tags and generates guidance scripts to build a user information interaction guidance plan. Through adaptive question-and-answer interaction, it constructs a set of initial user needs information; the needs positioning module is used to match multimodal information collection targets, match and correct structured units, and locate the needs of target users; the intelligent filling module is used to match question-and-answer interaction scale templates and intelligently fill them based on the needs of target users.

[0050] The beneficial effects of this invention are:

[0051] By using scenario-based question-and-answer association graphs and guiding tags, the initial information is structured, and the collection dimensions are accurately selected to avoid redundant information. Personalized interaction plans and adaptive question-and-answer based on guiding tags reduce indiscriminate and redundant questions, flexibly respond to ambiguous user responses, and improve interaction efficiency and user experience. Standardized processing and cross-validation of multimodal information, combined with the construction of demand evidence chains, significantly improve the accuracy of target demand positioning. Scenario-adaptive scale templates and historical case completion replace manual input and general templates, ensuring the standardization and credibility of scale data and providing reliable data support for subsequent scenario services. Overall, it achieves accurate, efficient, and standardized information guidance and form filling across scenarios. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of the structure of an information-guided identification and form-filling method according to the present invention;

[0053] Figure 2 A flowchart illustrating the specific steps involved in generating user guidance tags according to this invention;

[0054] Figure 3 This is a flowchart illustrating the steps involved in constructing the information interaction guidance plan of this invention.

[0055] Figure 4 A flowchart for constructing a set of initial user requirement information for this invention;

[0056] Figure 5 This is a flowchart illustrating the specific steps involved in identifying the needs of the target user in this invention. Detailed Implementation

[0057] Please see Figure 1 This embodiment introduces an information-guided identification and form-filling method, including:

[0058] Step S1: Collect the user's initial basic information, construct a question-and-answer association graph through a graph neural network, and quantify the association weights between nodes in the question-and-answer association graph through a graph attention mechanism; map the initial basic information to the corresponding nodes in the question-and-answer association graph one by one to generate the user's node association features, and perform cluster analysis based on the user's node association features using an improved K-means algorithm to generate the user's guidance tags;

[0059] In this embodiment, a question-and-answer association graph is constructed using a graph neural network to establish scene association dimensions, user needs, and target information. Combined with a graph attention mechanism, the association weights between nodes are quantified. This transforms the abstract association logic in various scenarios—such as region, travel preferences, and attraction types in a scenic area tour scenario; taste preferences, dietary restrictions, and dish classifications in a food recommendation scenario; and region, past medical history, chief symptoms, and syndrome types in a medical consultation scenario—into computable structured knowledge. This makes the quantification of association weights between different nodes more accurate, providing reliable domain knowledge support for subsequent personalized guidance and avoiding deviations in guidance direction caused by ambiguous association logic. The initial basic information of the user is mapped one by one to the corresponding graph. The node-generating node association features effectively solve the problems of scattered initial information, inconsistent formats, and difficulty in directly using it for interactive guidance. This significantly improves the structuring rate of previously fragmented information, providing high-quality feature input for subsequent clustering analysis. Based on the improved K-means clustering algorithm, which generates guidance tags, such as in-depth tour preferences for ancient towns and short-distance check-in preferences for scenic spot guides, and mild flavor preferences and non-spicy food preferences for food recommendations, the classification accuracy is improved. This can clearly define the exclusive question-and-answer interaction priority for each type of user, fundamentally reducing the situation of undifferentiated and redundant questions in general information interaction and reducing the proportion of invalid interaction steps.

[0060] Please see Figure 2 Preferably, the specific steps for generating user guide tags include:

[0061] To match the target scenario with the information collection dimensions, initial basic information of the user is collected. In this invention, the target scenario refers to a specific application scenario that requires obtaining user needs through targeted information interaction and providing personalized guidance to realize the scenario function. This includes, but is not limited to, intelligent consultation, chronic disease management and follow-up, TCM constitution identification, scenic spot guidance, and food recommendation. When the target scenario is intelligent consultation, the initial basic information of the user needs to cover the core dimensions related to diagnosis and treatment, specifically including the user's region, age, gender, past medical history, current chief complaint, recent basic signs, and recent dietary and lifestyle habits, to ensure that the initial information can provide basic data support for subsequent diagnosis and treatment-related guidance.

[0062] In the target scenario, a question-and-answer association graph is constructed using a graph neural network, encompassing scenario-related dimensions, user needs, and target information. This graph transforms abstract user basic information, user needs, and solution association logic within the target scenario into computable structured knowledge. This provides precise domain knowledge anchors for subsequent personalized interaction guidance, preventing the guidance direction from deviating from the scenario's functional objectives. Scenario-related dimensions refer to fundamental attribute dimensions directly related to the target scenario's functional implementation and influencing user need judgment. These dimensions serve as basic anchors for graph nodes, establishing a bridge between initial user basic information and user needs. This ensures that need judgment is based on scenario-adaptive fundamental features. For example, in an intelligent consultation scenario, scenario-related dimensions specifically include region, age, past medical history, and basic physical signs. User needs refer to the user's needs within the target scenario through… Information interaction aims to solve core problems or obtain key services, serving as core nodes in the graph. These nodes connect scenario-related dimensions with target information, clarifying the core direction of interaction guidance and avoiding redundant, undifferentiated questions. For example, in an intelligent consultation scenario, user needs specifically include symptom and etiology lookup, preliminary diagnosis of syndromes and diseases, targeted treatment advice, medication safety guidance, and follow-up appointment reminders. Target information refers to key information or service content that meets user needs and supports the implementation of scenario functions. These serve as output nodes in the graph, providing targeted content support for interaction guidance, ensuring that the guidance results align with user needs and are applicable to the scenario. For example, in an intelligent consultation scenario, target information specifically includes syndrome type, preliminary disease diagnosis results, targeted consultation direction, basic treatment advice, and medication precautions.

[0063] By employing a graph attention mechanism, and combining the question-and-answer logic of the target scenario with historical interaction data, the association weights between nodes in the question-and-answer association graph are quantified through linear transformation of node feature vectors, attention score calculation, and normalization. Specifically: First, the original features of nodes such as scenario association dimensions, user needs, and target information are vector-encoded. The original feature vectors are then spatially mapped using a preset linear transformation matrix to enhance the feature dimensions related to the scenario logic. Next, for any two potentially related nodes, the linearly transformed feature vectors are concatenated. Combining the professional logic of the target scenario with historical interaction data, an initial attention score is output through linear calculation with weighted parameters and an activation function, quantifying the degree of association between node pairs. Finally, the initial attention scores of all adjacent nodes associated with a core node are normalized using the Softmax function, mapping the scores to the 0-1 range. This ensures that the normalized scores both sum to 1 and intuitively reflect the differences in association weights between different adjacent nodes and the core node, thus completing the quantification of association weights between nodes. Taking the intelligent consultation scenario as an example, the original feature vectors of nodes such as scenario association dimension, user needs, and target information are first linearly transformed to strengthen consultation-related features. Then, the initial attention scores between nodes are calculated by combining consultation logic and historical interaction data. Finally, the association weight values ​​in the 0-1 range are obtained after normalization, which intuitively reflects the differences in the closeness of association between different nodes.

[0064] The collected initial user information is parsed and mapped one by one to the corresponding nodes in the question-and-answer association graph. The weight attributes of each mapped node are extracted and integrated into structured node association features, transforming fragmented and colloquial initial user information into standardized node association features that can be directly used for cluster analysis.

[0065] Using node association features as input, an improved K-means algorithm incorporating scene weight factors is employed. Through cluster analysis, users are categorized into clusters, and key scene features and user demand tendencies corresponding to each cluster are extracted to generate guidance tags tailored to those tendencies. The scene weight factor strengthens the weight of key scene features while mitigating the influence of features irrelevant or weakly correlated with scene needs, avoiding category bias caused by indiscriminate K-means clustering. It references the quantified association weights of nodes in the question-and-answer association graph, combines the professional logic of the target scene, and calibrates the weights using historical interaction data. Guidance tags, generated based on the key scene features and user demand tendencies corresponding to each cluster category, clarify the priority of question-and-answer interactions and the core guidance direction for each user type in the target scene. This avoids redundant and indiscriminate questioning in general information interactions, providing precise user demand guidance for the subsequent generation of structured interaction guidance scripts.

[0066] Step S2: Based on the quantified association weights of the guiding tags and node association features, the core guiding lines matching the user's needs and preferences are selected from the question-and-answer association graph. The information interaction stages are divided, and appropriate guiding scripts are generated for the preset information collection points of each information interaction stage. These are integrated to form a personalized information interaction guidance plan for the user. An adaptive question-and-answer interaction is initiated using a large language model to capture user needs information. During the adaptive question-and-answer interaction, an information interaction state transition matrix is ​​constructed based on a hidden Markov model. By learning from historical interaction data of the scenario, the joint transition probability of explicit stage switching and implicit need updates is quantified, and the information interaction stages and guiding scripts are dynamically adjusted. Conditional random fields are used to perform three-dimensional annotation of the captured colloquial needs information in terms of scenario dimensions, needs attributes, and preference details, transforming it into structured units and matching them with user needs to construct an initial set of user needs information.

[0067] In this embodiment, core guidance lines are selected based on guidance tags and node association features to generate personalized guidance plans, ensuring that the guidance direction accurately matches user needs and preferences, significantly reducing the proportion of invalid interaction steps and improving the smoothness of user interaction. Adaptive question answering is initiated using a large language model, which can flexibly respond to users' colloquial expressions and capture implicit demand information in answers in real time, significantly improving the coverage and accuracy of demand capture. Simultaneously, a state transition matrix constructed based on a Hidden Markov Model quantifies state transition probabilities by learning from historical data, intelligently handling situations such as ambiguous user responses and refusal to cooperate, dynamically adjusting the interaction stage and dialogue to avoid interaction interruptions. Furthermore, the three-dimensional annotation of colloquial information using conditional random fields can transform fragmented expressions into structured units, forming a complete set of initial user demand information when combined with a demand association rule base. This not only solves the problem of reusing colloquial information but also clarifies the precise target for multimodal information collection in step S3, providing high-quality data support for subsequent service recommendations.

[0068] Please see Figure 3 Preferably, the steps for constructing an information interaction guidance plan include:

[0069] The node association feature quantization strength matrix output by the graph attention mechanism is used. This strength matrix stores the pairwise association weights of three types of nodes in the question-and-answer association graph: scene association dimension, user needs, and target information. The generated guide tags are semantically encoded using the BERT model to generate guide tag semantic vectors. Then, a cosine similarity algorithm is used to calculate the semantic association degree between the guide tag semantic vector and the feature vectors of all nodes in the strength matrix, constructing a mapping table between guide tags and node association degrees. A selection ratio is set based on predefined complexity tags for the target scene, and the semantic association degrees are sorted. A list of node relationships is output in descending order. Based on the selection ratio, a set of key node relationships strongly related to the guide tags is obtained, ensuring that the node relationships are completely focused on the user needs corresponding to the guide tags.

[0070] For the selected set of key node relationships, a weighted directed graph path planning algorithm is used, with the association weights between nodes serving as the edge weights of the paths. Simultaneously, the priority of guidance tags is incorporated. Guidance tag priorities are pre-defined based on the scenario's professional logic, assigning node access priorities to different guidance tags. For example, in the "damp-heat stomach tendency" guidance tag, the tongue coating feature node has higher priority than the dietary preference node; in the "deep tour of ancient towns" guidance tag, the travel companion node has higher priority than the transportation mode node. A greedy strategy is used for path search, starting with scenario-related nodes and selecting the adjacent node with the highest association weight that matches the guidance tag priority as the next hop node. This process iterates until all key nodes are covered, forming the core guidance thread.

[0071] The question-and-answer relationship graph nodes are pre-labeled according to their functions. A hierarchical clustering algorithm is used with node hierarchical labels as clustering features. The core guidance thread is broken down into three progressive interaction stages: basic information confirmation, user needs refinement, and potential preference supplementation. Pre-set information collection points and collection targets are also set for each information interaction stage.

[0072] Based on a general-purpose large language model, a lightweight fine-tuning technique is employed. The model incorporates target scenario domain corpora: a medical scenario corpus containing anonymized medical dialogues, consultation guidelines, and labeled guidance tag dialogues; and a cultural tourism scenario corpus containing tour guide interaction records, user preference descriptions, and similar labeled data. Incremental pre-training is conducted during this process. The model's core parameters are frozen during fine-tuning, and only the scenario adaptation adapter is trained, enabling the model to learn the professional expressions of the scenario and the corresponding dialogue styles of the guidance tags. For each information collection point and target stage of information interaction, input text is constructed, including guidance tag content, target nodes, and scenario constraints. The large language model then generates guidance dialogues, ensuring that the dialogues accurately target the collection objectives and align with the user needs and preferences corresponding to the guidance tags.

[0073] This plan integrates the guiding scripts for each information interaction stage to construct a user information interaction guidance contingency plan. Using guidance tags as the core logical anchor, the plan includes a unique guidance tag ID, which is associated with the user's demand preferences and scenario attributes. It also incorporates a list of key node relationships selected from a question-and-answer relationship graph. Each relationship includes node pairs identified by related nodes in the graph, a quantified association weight based on the graph attention mechanism, and the semantic relevance between the guidance tag semantic vector and the node pair. Furthermore, it covers core guidance thread information built upon key node relationships. This information includes a sequence of path nodes arranged logically according to the progression of basic, demand, and preference layers, as well as node priority matching records indicating the priority of each path. In addition, the plan includes details of the information interaction stages broken down according to the core guidance thread node hierarchy. Each stage clearly defines its name, the corresponding graph hierarchy label, the quantified collection target for node information coverage requirements, the specific node information to be collected (i.e., the corresponding information collection point), and a set of adapted guiding scripts integrating the content corresponding to each information collection point in that stage. Ultimately, this forms a logically complete and comprehensive contingency plan system.

[0074] Please see Figure 4 Preferably, the specific steps for constructing a set of initial user requirement information include:

[0075] The large language model initiates adaptive question-and-answer interaction based on the information interaction guidance plan. Guided by the information collection points and collection objectives at each stage of information interaction, it receives users' spoken responses to the information collection points in real time, performs semantic understanding, interprets the explicit demand information in the response content, and extracts implicit demand information related to the user's demand tendency, forming a raw demand information pool covering explicit and implicit demands, providing a complete initial data foundation for subsequent information processing. Explicit demand information refers to the demand content directly expressed by the user in their response, which corresponds directly to the information collection point and the collection target. For example, for the tongue coating characteristic information collection point, the collection target is to confirm the color of the tongue coating. If the user answers: "My tongue coating is yellowish," this directly fills in the content corresponding to the information collection point, quickly achieving the quantitative collection target at each stage and ensuring the direct acquisition of basic information related to the user's needs. Implicit demand information refers to the demand content that the user does not express directly, but can be inferred from the response content and is associated with the guiding tags. For example, for the dietary situation information collection point, the collection target is to confirm dietary preferences. If the user answers: "I've been eating too much spicy food lately and my stomach feels uncomfortable," this can extract the information that spicy food is a cause of stomach discomfort. This is used to supplement the deficiencies of explicit demand information, improve the basis for judging user demand tendencies, and provide more comprehensive support for subsequent information completion, structured transformation, and demand matching.

[0076] In adaptive question-answering interaction, a Hidden Markov Model (HMM) information interaction state transition matrix, pre-trained using historical interaction data, is invoked. This matrix uses the information collection points and targets in the information interaction guidance plan as its core correlation dimensions, and incorporates the information interaction stage and the user's implicit needs as dual-state dimensions. By learning the correlation data between information interaction stage switching, user implicit needs updates, and the achievement of collection targets in the historical context, the matrix quantifies the joint transition probability of stage adjustments and information collection point completion triggered by changes in the user's implicit needs at different information interaction stages. The user's implicit needs refer to the user's preferences for each information collection point. Explicit demand information and derived implicit demand information are used to summarize and refine trend judgments related to the needs of users with guiding tags, rather than individual, scattered demand information points. Based on historical interaction data of different scenarios, user responses under different information collection points and collection targets are labeled to summarize typical user implicit demand tendencies and form a basic tendency library. Then, in real-time adaptive question answering, the explicit and implicit demand information of users' responses to the current information collection point are combined to match the most relevant user implicit demand tendencies from the basic tendency library, thus completing the setting of real-time user implicit demand tendencies and ensuring that they are highly consistent with the information collection point, collection target, and user needs of users with guiding tags.

[0077] Based on the joint transition probability output by the information interaction state transition matrix, the current information interaction stage is dynamically adjusted to ensure that the information collection point is supplemented without interrupting the interaction, thereby promoting the gradual achievement of the collection objectives at each stage and ensuring the integrity of the original demand information pool.

[0078] Using the information collection points and collection targets at each stage of information interaction as the annotation benchmark, the content in the original demand information pool is annotated in three dimensions: scenario dimension, demand attribute, and preference details. The annotation content of the scenario dimension includes: the application scenario to which the information belongs, and the scenario affiliation of the information collection point corresponding to the information. The demand attribute annotation information corresponds to the user demand type of the information collection point. The preference details annotation information includes the specific features required by the collection target. Through annotation, the scattered, colloquial demand information is transformed into structured units with a unified format and clear semantics that are directly bound to the information collection points and collection targets.

[0079] Subsequently, a pre-defined rule library for associating scenario information with user needs is invoked. This rule library contains association criteria for information collection points, collection targets, and guidance tags in various scenarios. For example, in a medical scenario, the color determination target of the tongue coating feature information collection point needs to match the syndrome feature requirements of the damp-heat stomach tendency guidance tag. The structured unit is then checked for consistency based on the association criteria: if the content of the structured unit does not conform to the collection target of the corresponding information collection point, or conflicts with the user needs of the guidance tag, the structured unit is removed and a secondary adaptive question-and-answer interaction is triggered. Guiding scripts are regenerated for the information collection points that do not meet the criteria to correct the information. After the verification is passed, the structured unit is associated and matched with the corresponding information interaction stage, information collection point, collection target, and guidance tag to ensure that each unit can accurately support the achievement of the collection target and the determination of user needs.

[0080] The matched structured units are integrated, first categorized by scenario, then further subdivided by information interaction stage. Each subcategory contains associated information collection points and collection targets. Finally, the structured units are bound to the node association strength calculated based on graph attention mechanism. A standardized format is used to organize the correspondence and association weights between each unit and the information collection points and collection targets. The final output is a complete set of initial user demand information. This initial demand information set not only clarifies the target achievement status of each information collection point, but also clearly presents the association logic between demand information and user needs. This provides a clear and accurate target direction for subsequent multimodal information collection, while also providing high-quality and highly adaptable data support for service recommendation.

[0081] Step S3: Match multimodal information collection targets based on the user's initial demand information set. Guide users to complete standardized information collection and target information positioning through multimodal collection guidance, feature extraction and cross-modal association matching. Integrate features of the collected target information to generate a standardized multimodal scene information dataset, which is used to match and correct structured units and locate target user needs.

[0082] In this embodiment, the generated set of initial user demand information serves as the core basis. By matching multimodal information collection targets and conducting standardized collection, the initial demand information is effectively expanded and verified. This accurately identifies the user's final target demand, such as the demand for diagnosis and treatment of damp-heat stomach syndrome in intelligent consultation, or the demand for in-depth parent-child tours of ancient towns and intangible cultural heritage in scenic area guides. This not only improves the accuracy and credibility of demand positioning, but also provides user demand support with both completeness and accuracy for subsequent scenario services.

[0083] Please see Figure 5 Preferably, the specific steps for identifying the needs of target users include:

[0084] The initial user demand information set is analyzed to extract the collection targets of structured units. Based on the collection targets, a joint matching mechanism is used, combining the node association weights between the collection targets and target information in the question-and-answer association graph with the priority rules of scenario requirements, to match multimodal information collection targets and determine the appropriate multimodal collection types. Specifically, the intelligent medical consultation scenario focuses on image and voice collection related to syndrome differentiation, while the scenic area tour scenario focuses on voice and image annotation collection related to experience preferences. This ensures that the determined multimodal collection types are highly adapted to user needs, providing a clear direction for subsequent standardized collection.

[0085] Based on the hidden Markov model's information interaction state transition matrix, a scenario-based multimodal data acquisition guidance scheme is generated. This scheme guides users in real time to complete multimodal information acquisition that conforms to the target scenario specifications, and dynamically adjusts the guidance strategy for non-compliance during the acquisition process. For example, for image acquisition, the scheme outputs operation guidance scripts and standard examples, clarifying the requirements for the acquisition environment and content specifications. For instance, in intelligent consultation, tongue coating acquisition must be performed under natural light to avoid food residue on the tongue coating surface. At the same time, the scheme detects the standardization of the acquired images in real time. If the images do not meet the requirements, the guidance content is dynamically adjusted based on the state transition matrix to prompt users to optimize the acquisition operation. This dynamic guidance mechanism differs from the fixed acquisition guidance in existing technologies. It can be flexibly adjusted according to the actual situation during the user's acquisition process to ensure that the final acquired multimodal information meets the standardization requirements.

[0086] A target scene-adaptive technical solution is adopted to extract features from the collected multimodal information, unify the feature format, and integrate the feature vectors of all multimodal information to form a standardized multimodal scene information dataset with a unified format. For example, for image information, core features related to scene requirements are extracted through convolutional neural networks, such as the color, thickness, and moisture characteristics of tongue coating in intelligent consultation, and key element features of activity scenes in scenic spot guidance. The extracted features are normalized to generate structured image feature vectors; this ensures that the extracted features are highly relevant to scene requirements, and the unified format processing solves the problem of difficulty in fusing multimodal features.

[0087] The basic matching degree between the structured unit and the feature vector of each modality is calculated by the cosine similarity algorithm. Then, based on the cross-modal attention mechanism, dynamic attention weights are assigned according to the different modal features of the target scene. For example, in the intelligent consultation scene, the tongue coating image features are more correlated with the syndrome determination, so higher attention weights are assigned to calculate the comprehensive matching degree of each structured unit.

[0088] The overall matching degree is compared with the preset matching degree threshold of the target scenario. If the overall matching degree is greater than the matching degree threshold, the initial demand information of the corresponding structured unit is determined to be accurate, and its annotation information is verified. If the overall matching degree is less than or equal to the matching degree threshold, such as the structured unit with white tongue coating in intelligent consultation not matching the yellow coating feature vector of the tongue coating image, a secondary adaptive question-and-answer interaction is triggered based on the information interaction state transition matrix of the Hidden Markov Model to generate targeted confirmation scripts. The structured unit is corrected according to user feedback to eliminate information bias. For the collection targets not covered in the user's initial demand information set, the corresponding multimodal feature vectors are transformed into new structured units and added to the initial demand set to improve the dimension of demand information.

[0089] Structured units are linked and bound to standardized multimodal scenario information datasets to form a user demand evidence chain. Each element of the evidence chain is labeled with its source and the corresponding question-and-answer association graph node association weight. This ensures that the elements of the evidence chain are traceable and the association logic is clear, solving the problem of insufficient support from traditional single text evidence chains and providing complete and three-dimensional data support for subsequent demand positioning.

[0090] The system invokes the logic of linking structured units in the question-and-answer association graph with nodes representing user needs. Using structured units in the demand evidence chain as the core input, it filters potential user demand directions based on the quantified association weights between structured units and user demand nodes in the question-and-answer association graph. This avoids deviations in demand direction due to ambiguity in the association logic and ensures a high degree of alignment between potential demands and the evidence chain.

[0091] The system integrates the pre-defined priority rules for the target scenario's needs. Combining this with the association weights of question-and-answer graph nodes corresponding to potential needs, a comprehensive priority score is calculated by multiplying the priority coefficient assigned to each user need direction in the pre-defined priority rules with the association weights of the question-and-answer graph nodes corresponding to that potential need direction. The potential need direction with the highest score is then selected as the target user need. This ensures that the target user needs both align with actual user demands and meet the requirements for implementing the scenario's functionalities.

[0092] Step S4: Based on the target user's needs, match the corresponding question-and-answer interaction scale template. Combine the corrected structured units with the standardized multimodal scenario information dataset, and achieve intelligent filling of the question-and-answer interaction scale through information quantification transformation, intelligent completion, and anomaly correction. First, convert the textual requirement information and multimodal feature vectors into scale quantification scores respectively, and automatically fill the scenario-adapted information collection scale. Complete missing items based on historical scales to avoid deviating from actual interaction needs. This not only realizes the automation and standardization of scenario information collection scales, but also outputs complete and reliable information for scenario interaction, effectively improving the efficiency and usability of cross-scenario information filling.

[0093] In this embodiment, by matching question-and-answer interactive scale templates, the redundancy or missing fields caused by general templates are avoided. For example, in the intelligent consultation scenario, the diagnosis and treatment needs of damp-heat accumulation in the stomach syndrome are accurately matched with the TCM initial diagnosis scale, and in the scenic spot tour scenario, the needs of parent-child in-depth intangible cultural heritage tours are matched with the parent-child cultural tourism preference scale. Quantitative conversion and filling replace manual input. In the intelligent consultation scenario, tongue coating image features and dietary inducement text are converted into scale scores and automatically filled. In the scenic spot tour scenario, voice preference keywords and companion characteristics are filled into corresponding fields. The form filling efficiency is significantly improved compared to manual filling, and the standardization of data format is greatly improved. Missing items are filled in according to historical scales, which solves the problem of pure statistical filling being out of touch with the needs. The accuracy of the filling is high, invalid data is avoided, and the data credibility is significantly improved.

[0094] Preferably, the specific steps for intelligent completion of the question-and-answer interactive scale include:

[0095] Based on the target user needs, the question-and-answer interaction questionnaire template that matches the target needs dimensions is located in the target scenario needs rule base. The target scenario needs rule base is a structured knowledge base that stores the matching criteria between user needs and questionnaire templates in the target scenario, questionnaire field specifications, data validation logic, and scenario-specific association rules. The question-and-answer interaction questionnaire consists of scenario identifiers, core field sets, field association logic, data entry specifications, and validation rules.

[0096] Based on the structured units and in accordance with the pre-set adaptation scoring criteria of the question-and-answer interaction scale, the collection targets of the structured units are transformed into standardized content or quantitative scores required by the question-and-answer interaction scale through a scenario-adaptive information mapping algorithm and an association weight calibration mechanism, and then filled into the corresponding question-and-answer interaction scale entries. The adaptation scoring criteria are formulated based on the target scenario requirement rules and are divided into levels according to the correlation strength between the structured units and the scale fields. Specifically, the collection targets in the structured units are first mapped to descriptive content of the scale fields according to the scoring criteria. During the transformation process, each filled entry is marked with a structured unit source identifier to ensure data traceability and avoid information misfilling.

[0097] For multimodal scene information datasets, feature vectors are transformed into scene descriptive information. Then, by referring to the core field set and data filling specifications of the question-and-answer interaction scale, a cross-modal mapping mechanism of multimodal features and scale fields is used. This mechanism is built based on scene-specific association rules in the target scene requirement rule base to match scene descriptive information to corresponding scale entries. During the filling process, each entry is labeled with a multimodal feature source identifier, including modality type, feature extraction algorithm identifier, and association weight, to ensure that multimodal information is presented collaboratively in the scale and to avoid the one-sidedness of single-modal information.

[0098] The process involves identifying missing entries in the question-and-answer interaction scale, retrieving historical scales of similar target user needs archived in the target scenario requirement rule base, filtering reference cases based on the similarity of the question-and-answer interaction scale entries, extracting valid fill-in content for corresponding missing entries from the reference cases as target fill-in information, completing the missing entries, and finally generating a complete question-and-answer interaction scale. The target scenario requirement rule base refers to a dedicated knowledge base that stores full-dimensional structured knowledge supporting the construction, filling, and missing entry completion of the question-and-answer interaction scale under the target scenario. This includes user needs and scale template matching criteria, scale field specifications, data validation logic, scenario-specific association rules, and a collection of historical scales archived according to target user need types.

[0099] This embodiment introduces an information-guided identification and form-filling system, including: an initial information processing module, an interactive guidance module, a demand positioning module, and an intelligent filling module;

[0100] The initial information processing module is used to collect initial basic information of users based on the target scenario and generate guidance tags; the interaction guidance module is used to build the core guidance main line based on the guidance tags and generate guidance scripts to build a user information interaction guidance plan. Through adaptive question and answer interaction, it builds a set of initial user demand information; the demand positioning module is used to match the multimodal information collection target, match and correct the structured units, and locate the target user demand; the intelligent filling module is used to match the question and answer interaction scale template based on the target user demand and perform intelligent filling.

[0101] The core function of the initial information processing module is to structure user's initial basic information and generate guidance tags adapted to the target scenario, providing a demand-oriented basis for subsequent personalized interaction guidance. Specifically, firstly, it collects user's initial basic information covering the core needs of the scenario, targeting the information collection dimension of the target scenario; secondly, it constructs a question-and-answer association graph containing scenario association dimensions, user needs, and target information through a graph neural network, combining the question-and-answer logic of the target scenario with historical interaction data, and uses a graph attention mechanism to quantify the association weights between graph nodes, transforming abstract association logic into computable structured knowledge; thirdly, it parses the collected initial basic information, mapping it one by one to the corresponding nodes in the question-and-answer association graph, extracting the weight attributes of each node and integrating them into structured node association features, solving the problems of fragmented and inconsistent format of initial information; finally, using the node association features as input, it performs cluster analysis using an improved K-means algorithm that incorporates scenario weight factors, classifies users into cluster categories, extracts key scenario features and demand tendencies for each category, generates scenario-adaptive guidance tags, and clarifies the user's exclusive question-and-answer interaction priority.

[0102] The core function of the interactive guidance module is to construct personalized information interaction guidance plans based on guidance tags, and generate a structured set of initial user needs through adaptive question-and-answer interaction, thus clarifying precise targets for multimodal information collection. Specifically, it calls the node association feature quantification intensity matrix output by the graph attention mechanism, performs semantic encoding on the guidance tags using the BERT model, calculates the semantic association degree between the semantic vector of the guidance tag and the feature vector of the nodes in the matrix, and obtains a set of key node relationships strongly related to the guidance tags by combining the scenario-preset screening ratio; it uses a weighted directed graph path planning algorithm to construct the core guidance main line, and uses a hierarchical clustering algorithm to break down the main line into three progressive interaction stages: basic information confirmation, user need refinement, and potential preference supplementation, setting the information collection points and collection targets for each stage; and it uses a general large language model as a base and accesses the target scenario domain corpus to conduct lightweight... Quantitative fine-tuning involves generating adaptive guiding scripts for each stage of data collection and targeting, integrating them into a user information interaction guidance plan. Based on the plan, adaptive question-and-answer interaction is initiated, receiving users' spoken responses in real time, extracting explicit and implicit demand information to construct an original demand information pool, and calling a pre-trained Hidden Markov Model information interaction state transition matrix to dynamically adjust the interaction stage and supplement the guiding scripts. Conditional random fields are used to perform three-dimensional annotation of the original demand pool in terms of scenario dimensions, demand attributes, and preference details, generating structured units. After verification and association matching between scenario information and user demand association rule base, these units are integrated into a complete set of structured initial user demands.

[0103] The core function of the demand positioning module is to collect, process, and match multimodal information to correct structured units and accurately locate the needs of target users, providing core demand support for subsequent scale filling. Specifically, the process involves parsing the initial set of user needs, extracting the collection targets of structured units, matching multimodal information collection targets and determining suitable collection types by using the node association weights between collection targets and target information in the question-and-answer association graph and the scenario demand priority rules; generating scenario-based multimodal collection guidance schemes based on the information interaction state transition matrix of a Hidden Markov Model, outputting operation guidelines and standard examples, real-time detection of the standardization of collected information, and dynamically adjusting the guidance strategy to ensure standardized collection; using scenario adaptation technology to extract core features of multimodal information, unifying the feature format, and integrating them into a standardized multimodal scenario information dataset; calculating the basic matching degree between structured units and feature vectors of each modality, calculating the comprehensive matching degree by combining a cross-modal attention mechanism, comparing the accuracy of units with the scenario preset threshold, triggering secondary adaptive question-and-answer correction if the standard is not met, and converting multimodal features of uncovered collection targets into new structured units to supplement the demand set; associating and binding structured units with multimodal datasets to form a user demand evidence chain, filtering potential demand directions based on the node association weights between structured units and user needs in the graph, calculating a comprehensive priority score by combining the scenario demand priority rules, and selecting the direction with the highest score as the target user demand.

[0104] The core function of the intelligent fill module is to automate and standardize the intelligent filling of question-and-answer interaction questionnaires based on the needs of target users and multi-source structured information, thereby improving the efficiency and reliability of cross-scenario information filling. Specifically, based on the needs of the target users, a question-and-answer interaction scale template that perfectly matches the needs dimension is located in the target scenario requirement rule base. The collection targets of the modified structured units are transformed into standardized content or quantitative scores required by the scale through scenario-adapted information mapping algorithms and association weight calibration mechanisms. After annotating the source identifier of the structured units, they are filled into the corresponding scale entries. The feature vectors of the standardized multimodal scenario information dataset are transformed into scenario descriptive information. Through a cross-modal mapping mechanism based on scenario professional association rules, they are matched to the corresponding scale entries. The modality type, feature extraction algorithm identifier, and association weight are annotated to ensure the collaborative presentation of multimodal information. Missing entries in the scale are identified. Historically verified scales under the same type of target user needs are retrieved from the target scenario requirement rule base. Reference cases are selected according to the similarity of the entries. The valid filled content of the corresponding missing entries in the cases is extracted as the completion information to complete the missing entries. Finally, a complete and standardized question-and-answer interaction scale is generated, realizing the automated processing of the scenario information collection scale and avoiding the problems of general template redundancy and pure statistical completion being out of touch with the requirements.

[0105] Working principle and its effects:

[0106] This invention focuses on information guidance and form filling needs in target scenarios such as intelligent consultation and scenic spot navigation. It takes scenario-adaptive information structuring, precise demand positioning, and standardized form filling as its core logic. Through end-to-end technical collaboration, it achieves efficient interaction and data processing, which not only solves the problems of information fragmentation, interaction redundancy, demand misjudgment, and non-standard form filling in existing technologies, but also provides reliable data support for the implementation of scenario services.

[0107] This invention first collects initial basic information for the target scenario. It then constructs a question-and-answer association graph containing scenario-related dimensions, user needs, and target information using a graph neural network. Combining graph attention mechanisms, it quantifies node association weights, mapping the initial information into structured node association features. After cluster analysis, it generates scenario-adaptive guidance tags. This process not only transforms abstract scenario-related logic into computable structured knowledge, solving the problem of initially information being formatted in a chaotic way and difficult to use directly for needs analysis, but also accurately captures user needs and preferences, clarifying priorities for subsequent interactions and reducing indiscriminate and redundant questions from the source. Based on the guidance tags, this invention further matches the core guidance thread, divides the interaction stages, generates guidance scripts using a lightweight, fine-tuned large language model to construct interaction plans, initiates adaptive question-and-answer interaction to extract explicit and implicit needs, and dynamically adjusts the interaction using a hidden Markov model. The strategy involves creating a structured initial set of requirements through 3D annotation and rule validation, effectively improving the flexibility and efficiency of interaction, transforming colloquial information into standardized data, and avoiding interaction interruptions or information omissions. Subsequently, the invention matches multimodal collection targets to generate a standardized multimodal scene information dataset. A cross-modal attention mechanism is used to validate and correct structured units, forming a chain of evidence for user needs and locating target needs. This overcomes the limitations of single text information, significantly improving the accuracy of need judgment and avoiding misjudgments due to incomplete information. Finally, based on a target need matching scenario-adapted scale template, structured information and multimodal descriptions are transformed into scale content. Historical scales of the same type are retrieved to complete missing items, replacing manual input and general templates. This improves form-filling efficiency while ensuring the standardization and credibility of scale data, solving the problem of supplementary content deviating from actual needs.

[0108] In summary, this invention achieves end-to-end optimization from information collection to questionnaire filling through deep collaboration of various technical links. It can balance interaction efficiency, demand accuracy and data quality in different target scenarios, effectively make up for the shortcomings of existing technologies, provide key technical support for the efficient implementation of scenario-based personalized services, and significantly improve user experience and service support capabilities.

[0109] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An information-guided identification and form-filling method, characterized in that, include: Collect initial basic information of users, construct a question-answering association graph through a graph neural network, and quantify the association weights between nodes in the question-answering association graph through a graph attention mechanism; The initial basic information is mapped to the corresponding nodes in the question-and-answer association graph to generate the user's node association features, which are used to divide the user's cluster category through cluster analysis and generate the user's guidance tags. Based on the quantified association weights of the guidance tags and node association features, the core guidance main line is matched from the question-and-answer association graph, the information interaction stage is divided and the information collection points and collection targets are set, and guidance scripts are generated to construct a user information interaction guidance plan for initiating adaptive question-and-answer interaction, constructing an original demand information pool, generating structured units, and constructing a user initial demand information set through verification and integration. Based on the user's initial demand information set, a multimodal information collection target is matched to generate a standardized multimodal scene information dataset, which is used to match and correct structured units in order to locate the target user's needs. Based on the needs of the target users, a question-and-answer interaction scale template is matched, and the template is intelligently filled in through information quantification and intelligent missing completion.

2. The information-guided identification and form-filling method as described in claim 1, characterized in that, The specific steps for constructing the question-answering association graph include: For the target scenario matching information collection dimension, initial basic user information is collected; under the target scenario, a question-answer association graph containing scenario association dimension, user needs and target information is constructed through graph neural network; By using graph attention mechanism, combining the question-and-answer logic of the target scenario with historical interaction data, the association weights between nodes in the question-and-answer association graph are quantified through linear transformation of node feature vectors, attention score calculation and normalization.

3. The information-guided identification and form-filling method as described in claim 2, characterized in that, The specific steps for generating user guide tags include: The collected initial basic information of users is parsed and mapped to the corresponding nodes in the question-and-answer association graph. The weight attributes of each mapped node are extracted and integrated into structured node association features. Using node association features as input, cluster analysis is used to divide users into cluster categories, extract key scene features and user demand tendencies corresponding to the cluster categories, and generate guidance tags that are adapted to user demand tendencies. The guidance tag refers to a scenario-adaptive demand identifier generated based on the key features of the scenario and user demand tendencies corresponding to the cluster category. It is used to clarify the priority of question-and-answer interaction and the core guidance direction for each type of user in the target scenario.

4. The information-guided identification and form-filling method as described in claim 1, characterized in that, The steps for constructing the user information interaction guidance plan include: The node association feature intensity matrix output by the graph attention mechanism is quantized, and the generated guide label is semantically encoded to generate a guide label semantic vector. Calculate the semantic correlation between the semantic vector of the guide label and the feature vectors of all nodes in the intensity matrix, set the filtering ratio, sort the semantic correlation, and obtain the set of key node relationships based on the filtering ratio; For key node relationship sets, a weighted directed graph path planning algorithm is adopted, with the association weight between nodes as the path edge weight; at the same time, the priority of the guidance label is incorporated, and a greedy strategy is used to carry out path search and construct the core guidance main line. The nodes of the question-and-answer association graph are labeled at the node level. A hierarchical clustering algorithm is used to divide the information interaction stages of the core guiding line and set the preset information collection points and collection targets for each information interaction stage. Based on a general large language model, we access the target scenario domain corpus and carry out incremental pre-training. For the information collection points and collection targets of each information interaction stage, we construct input text containing guidance label content, collection target nodes and scenario constraints to generate guidance dialogue. Integrate the guiding scripts for each stage of information interaction to build a user information interaction guidance plan.

5. The information-guided identification and form-filling method as described in claim 1, characterized in that, The steps of constructing the original demand information pool and generating structured units include: Initiate adaptive question-and-answer interaction, receive users' spoken language responses in real time, perform semantic understanding, extract explicit and implicit demand information from the response content, and build a pool of original demand information. The information interaction state transition matrix of the pre-trained Hidden Markov Model is invoked. The information interaction state transition matrix takes the information collection point and the collection target as the core correlation dimension and takes the information interaction stage and the user's implicit demand tendency as the dual state dimension. It quantifies the joint transition probability of triggering stage adjustment and information collection point completion after the user's implicit demand tendency changes under different information interaction stages. The current information interaction stage is dynamically adjusted based on the joint transition probability output by the information interaction state transition matrix. The original demand information pool is annotated in three dimensions, including scene dimension, demand attribute, and preference details, to construct structured units.

6. The information-guided identification and form-filling method as described in claim 5, characterized in that, The steps for constructing the user's initial requirement information set include: The system calls upon a rule base that associates target scenarios with user needs. This rule base contains association criteria for information collection points, collection targets, and guidance tags in each target scenario. Based on the association rule base, the structured units are validated, non-compliant information collection points are identified, guiding messages for non-compliant information collection points are generated, and adaptive question-and-answer interaction is performed again. The structured units are associated and matched with the corresponding information interaction stages, information collection points, collection targets and guiding labels, and the initial user demand information set is constructed by integration.

7. The information-guided identification and form-filling method as described in claim 5, characterized in that, The steps for matching and modifying structured units include: The system analyzes the initial set of user needs information, extracts the collection targets of structured units, and matches multimodal information collection targets based on the node association weights between the collection targets and target information in the question-and-answer association graph and the priority of scenario requirements. Based on the information interaction state transition matrix of the Hidden Markov Model, a scenario-based multimodal data acquisition guidance scheme is generated to guide users in real time to complete multimodal information acquisition that conforms to the target scenario specifications. Feature extraction is performed on the collected multimodal information, and the feature vectors of the modal information are integrated to construct a standardized multimodal scene information dataset; The basic matching degree between the structured unit and each modal feature vector is calculated. Then, based on the cross-modal attention mechanism, dynamic attention weights are assigned according to the different modal features of the target scene to calculate the comprehensive matching degree of each structured unit. The overall matching degree is compared with the preset matching degree threshold of the target scene. If the overall matching degree is greater than the matching degree threshold, the initial requirement information of the structured unit is determined to be accurate; otherwise, a second adaptive question-and-answer interaction is triggered to correct the structured unit.

8. The information-guided identification and form-filling method as described in claim 7, characterized in that, The steps for identifying the needs of target users include: Structured units are associated and bound with standardized multimodal scene information datasets to form a chain of evidence for user needs, and the source and corresponding question-and-answer association graph node association weights are labeled for each element of the evidence chain. Based on the quantitative association weights between structured units and user demand nodes within the question-and-answer association graph, potential user demand directions are screened. Based on the pre-defined priority rules for the target scenario, and combined with the association weights of the question-and-answer association graph nodes corresponding to the potential demand directions, the comprehensive priority score of each potential user demand direction is calculated by using the priority coefficients set for each user demand direction in the pre-defined priority rules for the target scenario and the association weights of the question-and-answer association graph nodes corresponding to the potential demand directions, in order to select potential demand directions as target user demands.

9. The information-guided identification and form-filling method as described in claim 1, characterized in that, The step of intelligently filling the question-and-answer interaction scale template includes: Based on the target user's needs, locate the question-and-answer interaction scale template that matches the target needs dimension in the target scenario needs rule base; The collection targets of the structured units are transformed into standardized content or quantitative scores required by the question-and-answer interaction scale through scene-adaptive information mapping algorithms and associated weight calibration mechanisms, and then filled into the corresponding question-and-answer interaction scale items. The feature vectors of the standardized multimodal scene information dataset are transformed into scene descriptive information. By comparing the core field set and data filling specifications of the question-and-answer interaction scale, the scene descriptive information is matched to the corresponding question-and-answer interaction scale items through a multimodal feature and scale field cross-modal mapping mechanism. Identify missing items in the question-and-answer interaction scale, obtain historical scales of similar target user needs archived in the target scenario requirement rule base, filter reference cases based on the similarity of the question-and-answer interaction scale items, extract the valid fill content of the corresponding missing items in the reference cases as target fill information, complete the missing items, and generate the question-and-answer interaction scale.

10. An information-guided identification and form-filling system, implemented based on the information-guided identification and form-filling method according to any one of claims 1-9, characterized in that, include: Initial information processing module, interactive guidance module, requirement positioning module and intelligent filling module; The initial information processing module collects the user's initial basic information based on the target scenario and generates guidance tags; The interactive guidance module constructs a core guidance thread based on guidance tags, generates guidance scripts, and builds a user information interaction guidance plan. Through adaptive question-and-answer interaction, it constructs a set of initial user needs information. The demand localization module is used to match multimodal information collection targets, match and correct structured units, and locate target user needs; the intelligent filling module is used to match question-and-answer interaction scale templates and intelligently fill them based on target user needs.

Images