User demand intention identification method and device, equipment and storage medium

Abstract

The present application relates to the technical field of user intention recognition, and particularly relates to a user demand intention recognition method, device and equipment and a storage medium, the method first acquires a pre-constructed multi-dimensional intention space, which includes a plurality of key dimensions, performs semantic matching processing on the acquired user initial input text based on the multi-dimensional intention space to construct an initial intention certainty heat map, determines a target confidence corresponding to each key dimension based on the initial intention certainty heat map, performs structured entropy calculation processing on the initial intention certainty heat map, compares the obtained global confidence discrete entropy with a preset convergence threshold, and compares each target confidence with a preset peak value score threshold, respectively, when the preset convergence condition is met, outputs a structured user intention vector based on the initial intention certainty heat map, aiming to realize efficient convergence of accurate structured user intention according to uncertain dynamic interaction questioning.

Description

Technical Field

[0001] This invention relates to the field of user intent recognition technology, and in particular to a method, apparatus, device, and storage medium for recognizing user needs and intents. Background Technology

[0002] Materials testing is a crucial link in industrial production quality control and engineering safety assurance, widely used in metallurgy, aerospace, manufacturing, and many other fields. In pre-sales business scenarios for materials testing, users often express their testing needs in a vague and incomplete manner. The system needs to identify multi-dimensional parameters such as material type, test object, test target, and testing standards from these fragmented user descriptions to generate a suitable testing plan. User needs are inherently uncertain and dynamically change during the interaction process. How to gradually converge to obtain precise user intent through strategic questioning has become a pressing technical challenge in this scenario.

[0003] In existing technologies, keyword matching rules are ill-suited for colloquial expressions and do not support multi-turn interactions. Single-turn deep learning classification models can only make static judgments based on a single input and cannot iteratively correct recognition results using user feedback. Traditional slot-filling methods collect information sequentially according to a preset order, failing to quantify the degree of uncertainty in different dimensions and easily leading to redundant questions or delayed acquisition of key information. While reinforcement learning adaptive strategies can dynamically select interaction actions, training data in specialized domains is scarce, resulting in poor model interpretability. Existing general dialogue solutions, in professional detection scenarios, struggle to effectively quantify intent uncertainty and select the optimal questioning strategy accordingly. User negative feedback cannot be effectively transmitted to the associated dimensions, nor can it fully utilize the dependencies and mutual exclusion constraints between domain dimensions to collaboratively update confidence. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, the present invention aims to provide a user demand intent recognition method, device, equipment and storage medium, which aims to solve the problems of ambiguous expression of user demand and difficulty in quantifying uncertainty of intent in material testing scenarios, realize dynamic interactive questioning based on uncertainty, and efficiently converge to obtain accurate structured user intent.

[0005] The first aspect of this invention provides a method for recognizing user intent, comprising: acquiring initial user input text, a pre-constructed multi-dimensional intent space, a preset convergence threshold, and a peak score threshold, wherein the multi-dimensional intent space includes multiple key dimensions; performing semantic matching processing on the initial user input text based on the multi-dimensional intent space to construct an initial intent deterministic heatmap; determining the target confidence level corresponding to each of the key dimensions based on the initial intent deterministic heatmap; performing structured entropy calculation processing on the initial intent deterministic heatmap to obtain a global confidence discrete entropy; comparing the global confidence discrete entropy with the convergence threshold, and comparing each of the target confidence levels with the peak score threshold respectively, to determine whether a preset convergence condition is met; and when the convergence condition is met, outputting a structured user intent vector based on the initial intent deterministic heatmap.

[0006] Optionally, in a first implementation of the first aspect of the present invention, the multidimensional intent space further includes multiple non-critical dimensions, with each non-critical dimension and each critical dimension respectively serving as a preset dimension, and each preset dimension including multiple candidate values; the step of performing semantic matching processing on the user's initial input text based on the multidimensional intent space to construct an initial intent deterministic heatmap includes: inputting the user's initial input text and all the candidate values ​​into a pre-trained language model for vector encoding processing to obtain a semantic vector and multiple candidate vectors corresponding to each preset dimension; calculating the matching score between the semantic vector and each candidate vector under the same preset dimension to obtain multiple matching scores corresponding to each preset dimension; normalizing each matching score under the same preset dimension to obtain multiple initial deterministic scores corresponding to each preset dimension; and performing heatmap assembly processing on all the initial deterministic scores based on the dimensional organization structure of the multidimensional intent space to construct the initial intent deterministic heatmap.

[0007] Optionally, in a second implementation of the first aspect of the present invention, determining the target confidence level corresponding to each of the key dimensions based on the initial intent deterministic heatmap includes: obtaining multiple initial deterministic scores corresponding to each of the preset dimensions based on the initial intent deterministic heatmap; taking the initial deterministic score with the highest score under each preset dimension as the dimension confidence level corresponding to that preset dimension; extracting the dimension confidence level corresponding to each key dimension from the dimension confidence levels corresponding to all preset dimensions as the target confidence level, thereby obtaining the target confidence level corresponding to each of the key dimensions.

[0008] Optionally, in a third implementation of the first aspect of the present invention, the multidimensional intent space further includes multiple non-critical dimensions, with each non-critical dimension and each critical dimension respectively serving as preset dimensions. The step of performing structured entropy calculation on the initial intent deterministic heatmap to obtain global confidence discrete entropy includes: obtaining preset dimension association rules based on the multidimensional intent space, and determining dimension pairs with direct correlation from all preset dimensions according to the dimension association rules to obtain multiple associated dimension pairs; obtaining multiple initial deterministic scores corresponding to each preset dimension based on the initial intent deterministic heatmap; performing independent entropy calculation on the corresponding preset dimensions based on the initial deterministic scores to obtain independent entropy corresponding to each preset dimension; performing mutual information calculation on the corresponding associated dimension pairs based on the initial deterministic scores to obtain mutual information corresponding to each associated dimension pair; and performing aggregation operations on all independent entropies and all mutual information to obtain the global confidence discrete entropy.

[0009] Optionally, in a fourth implementation of the first aspect of the present invention, after comparing the global confidence discrete entropy with the convergence threshold and comparing each target confidence level with the peak score threshold to determine whether a preset convergence condition is met, the method further includes: when the convergence condition is not met, taking the key dimension whose target confidence level is less than or equal to the peak score threshold as a non-converged dimension; determining the target question based on each non-converged dimension and the global confidence discrete entropy; obtaining feedback information input by the user based on the target question, and based on the feedback... Feedback information is used to update the initial intent deterministic heatmap to obtain an updated intent deterministic heatmap; the updated intent deterministic heatmap is then processed by structured entropy calculation to obtain an updated global confidence discrete entropy; based on the updated intent deterministic heatmap, the updated target confidence level corresponding to each of the key dimensions is determined; the updated global confidence discrete entropy is compared with the convergence threshold, and each of the updated target confidence levels is compared with the peak score threshold; when the convergence condition is met, a structured user intent vector is output based on the updated intent deterministic heatmap.

[0010] Optionally, in a fifth implementation of the first aspect of the present invention, determining the target question based on each of the non-converged dimensions and the global confidence discrete entropy includes: determining candidate question directions corresponding to each of the non-converged dimensions and acquiring historical dialogue data; determining multiple candidate answers corresponding to each candidate question direction and the occurrence probability corresponding to each candidate answer based on the historical dialogue data; calculating the expected information gain corresponding to each candidate question direction based on the occurrence probability corresponding to each candidate answer for each candidate question direction; taking the candidate question direction with the highest expected information gain value as the target question direction; acquiring a preset question granularity adaptation rule, and determining the target question based on the question granularity adaptation rule, the target question direction, and the global confidence discrete entropy.

[0011] Optionally, in a sixth implementation of the first aspect of the present invention, the step of updating the initial intent deterministic heatmap based on the feedback information to obtain an updated intent deterministic heatmap includes: taking the unconverged dimension corresponding to the feedback information as the dimension to be updated, and determining the feedback type of the feedback information; when the feedback type is a positive information type, determining a positive candidate value corresponding to the feedback information based on the dimension to be updated, and adjusting the initial deterministic score corresponding to the positive candidate value based on a preset score adjustment rule to obtain an adjusted score; when the feedback type is a negative information type, determining a negative candidate value corresponding to the feedback information based on the dimension to be updated and... For mutually exclusive candidate values ​​corresponding to the negative candidate values, the initial deterministic scores corresponding to the negative candidate values ​​are adjusted downwards based on preset contradiction rules, and the initial deterministic scores corresponding to the mutually exclusive candidate values ​​are adjusted upwards to obtain an adjusted score set. Based on the multidimensional intent space, preset dimension association rules are obtained, and other preset dimensions that are associated with the dimension to be updated are determined according to the dimension association rules. Based on the adjusted scores or the adjusted score set, combined with a preset association conditional probability matrix and a propagation strength factor, confidence propagation processing is performed on the other preset dimensions to update the initial intent deterministic heatmap, resulting in the updated intent deterministic heatmap.

[0012] A second aspect of the present invention provides a user intent recognition device, comprising: a data acquisition module for acquiring initial user input text, a pre-constructed multi-dimensional intent space, a preset convergence threshold, and a peak score threshold, wherein the multi-dimensional intent space includes multiple key dimensions; a semantic matching module for performing semantic matching processing on the initial user input text based on the multi-dimensional intent space to construct an initial intent deterministic heatmap; a confidence determination module for determining a target confidence level corresponding to each key dimension based on the initial intent deterministic heatmap; a structured entropy calculation module for performing structured entropy calculation processing on the initial intent deterministic heatmap to obtain a global confidence discrete entropy; a comparison module for comparing the global confidence discrete entropy with the convergence threshold, and comparing each target confidence level with the peak score threshold respectively, to determine whether a preset convergence condition is met; and an output module for outputting a structured user intent vector based on the initial intent deterministic heatmap when the convergence condition is met.

[0013] A third aspect of the present invention provides a user demand intent recognition device, the user demand intent recognition device comprising: a memory and at least one processor, the memory storing instructions; the at least one processor invokes the instructions in the memory to cause the user demand intent recognition device to perform the various steps of the user demand intent recognition method described in any of the preceding claims.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the user demand intent recognition method described in any of the preceding claims.

[0015] In the technical solution of this invention, the user's initial input text, a pre-constructed multi-dimensional intent space, a preset convergence threshold, and a peak score threshold are first obtained. The multi-dimensional intent space includes multiple key dimensions. Then, semantic matching processing is performed on the user's initial input text based on the multi-dimensional intent space to construct an initial intent deterministic heatmap. Next, the target confidence level corresponding to each key dimension is determined based on the initial intent deterministic heatmap. The structured entropy calculation processing is performed on the initial intent deterministic heatmap to obtain the global confidence discrete entropy. Finally, the global confidence discrete entropy is compared with the convergence threshold, and each target confidence level is compared with the peak score threshold to determine whether the preset convergence condition is met. When the convergence condition is met, a structured user intent vector is output based on the initial intent deterministic heatmap. This aims to solve the problems of ambiguous user demand expression and difficulty in quantifying intent uncertainty in material testing scenarios, realize dynamic interactive questioning based on uncertainty, and efficiently converge to obtain accurate structured user intent. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the user intent recognition method provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the user demand intent recognition device provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the user intent recognition device provided in an embodiment of the present invention. Detailed Implementation

[0017] This invention provides a method, apparatus, device, and storage medium for identifying user intent. In this invention, the terms "first," "second," "third," "fourth," etc. (if present)," in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0018] The user intent recognition method, apparatus, device, and storage medium provided in this invention belong to the field of artificial intelligence and can be applied to scenarios such as pre-sales consultation for material testing, intelligent customer service for industrial testing, and human-computer dialogue consultation in professional fields. These include, but are not limited to, multi-dimensional parameter analysis of testing needs, multi-round strategic interactive questioning, user feedback confidence updating, and progressive convergence recognition of uncertain intent. The core of this invention lies in constructing a multi-dimensional intent space containing key and non-key dimensions, building a deterministic heatmap of intent through semantic matching, quantifying overall intent uncertainty using structured entropy, obtaining user feedback by selecting the optimal questioning direction based on information gain, and completing confidence propagation updates by combining domain-dimensional association rules. Relying on uncertainty quantification, information gain-driven questioning, and domain-constrained collaborative confidence update mechanisms, this invention solves the problems of ambiguous user need expression, difficulty in quantifying intent uncertainty, and inability to transmit feedback to associated dimensions in professional scenarios. It provides interpretable and progressive user intent recognition underlying interactive support for intelligent consultation for material testing and industrial professional human-computer interaction systems.

[0019] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the user demand intent recognition method in this invention includes: 101. Obtain the user's initial input text, a pre-constructed multidimensional intent space, a preset convergence threshold, and a peak score threshold, wherein the multidimensional intent space includes multiple key dimensions; 102. Based on the multidimensional intent space, perform semantic matching processing on the user's initial input text to construct an initial intent deterministic heatmap; In this embodiment, a multi-dimensional intent space is pre-constructed for material testing business scenarios. This space is divided into key dimensions and non-key dimensions, with all dimensions collectively constituting preset dimensions. Each preset dimension is configured with several candidate values, and association rules between preset dimensions are defined within the space. Key dimensions are selected based on their impact on the generation of testing solutions. Candidate values ​​for key dimensions directly change the combination of testing items and core pricing parameters. For example, key dimensions by default include material type, test object, test target, and testing standard. Taking the key dimension of material type as an example, its corresponding candidate values ​​can include metal, plastic, rubber, and ceramic. The candidate values ​​corresponding to the key dimension of test target can include composition analysis, mechanical properties, thermal properties, electrical properties, and aging tests. The remaining non-key dimensions are classified as non-key dimensions. Non-key dimensions can retain a certain degree of uncertainty for subsequent business modules to handle. For example, budget range, timeliness requirements, and qualification requirements are non-key dimensions. The candidate values ​​corresponding to the budget range can be divided into multiple price ranges, the candidate values ​​corresponding to the timeliness requirements can include regular cycle and expedited cycle, and the candidate values ​​corresponding to the qualification requirements can include requiring CNAS qualification and not requiring special qualification. Within the multidimensional intent space, dimensional association rules can be set. For example, when testing the mechanical properties of the target, the sample morphology cannot be powder.

[0020] In this embodiment, the initial user input text is the natural language text of the detection request submitted by the user. For example, the user requests to test the strength of a plastic part. The initial user input text and all candidate values ​​are fed into a pre-trained language model to complete vector encoding, generating semantic vectors and candidate vectors corresponding to each preset dimension. The matching score between the semantic vector and each candidate vector within the same preset dimension is calculated. Normalization is performed on all matching scores within a single preset dimension to obtain the initial deterministic score corresponding to each preset dimension. Heatmap assembly processing is performed on all initial deterministic scores with reference to the dimensional organization structure of the multidimensional intent space to obtain the initial intent deterministic heatmap. The initial intent deterministic heatmap carries the initial deterministic scores corresponding to all candidate values ​​under each preset dimension, which can quantify the degree of matching of the user input text on each candidate value. This processing method relies on a pre-trained language model to achieve semantic matching between text and candidate values, overcoming the limitations of traditional keyword matching. It can fully capture the semantic information implicit in natural language. Normalization processing ensures that scores under different dimensions have comparable and unified dimensions. Heatmap assembly processing integrates the scattered dimensional scores into a unified data carrier, providing a complete and standardized data foundation for subsequent uncertainty quantification, convergence judgment, and multi-round feedback updates. At the same time, it can retain the complete score distribution of each candidate value and will not directly discard low-confidence candidates, reserving computational space for the confidence propagation of subsequent dimensional association rules.

[0021] 103. Determine the target confidence level corresponding to each of the key dimensions based on the initial intent deterministic heatmap; In this embodiment, after constructing the initial intent determinism heatmap, it is necessary to obtain the target confidence level corresponding to each key dimension from the initial intent determinism heatmap. The initial intent determinism heatmap stores the initial determinism scores corresponding to each candidate value under all preset dimensions. First, all initial determinism scores contained in each preset dimension within the initial intent determinism heatmap are read. The initial determinism score with the highest value within a single preset dimension is selected, and this score is identified as the dimension confidence level corresponding to this preset dimension. The dimension confidence level represents the credibility of the most likely item among all candidate values ​​of this preset dimension under the current conditions. The part belonging to the key dimension is filtered out from the dimension confidence levels corresponding to all preset dimensions. The dimension confidence level obtained by this filtering is the target confidence level corresponding to the key dimension. The target confidence level represents the level of certainty of the most likely candidate under each key dimension. This processing method uses the peak score within each dimension as the basis for calculating the dimension confidence score. It can intuitively reflect the degree of confidence of the best candidate in a single dimension. By distinguishing preset dimensions and extracting only the dimension confidence score of key dimensions as the target confidence score, it can filter out numerical interference from non-key dimensions. It can focus the calculation of subsequent convergence judgment on the key dimensions that directly affect the output of the business solution, avoid unnecessary computational overhead caused by secondary parameters, and retain the complete original heatmap data. The scores of non-key dimensions will not be deleted and can be used for subsequent confidence propagation processes.

[0022] 104. Perform structured entropy calculation on the initial intention deterministic heatmap to obtain the global confidence discrete entropy; In this embodiment, the structured entropy calculation process uses the initial intent deterministic heatmap as the input data source. First, it reads the pre-configured dimension association rules within the multi-dimensional intent space. Based on the dimension association rules, it filters out dimension pairs with direct correlation among all preset dimensions, generating several sets of associated dimension pairs. Then, it reads the initial deterministic scores corresponding to each preset dimension output by the initial intent deterministic heatmap. For each preset dimension, it performs independent entropy calculation, which is based on the initial deterministic scores corresponding to each candidate value under the preset dimension. Subsequently, it performs mutual information calculation on the filtered associated dimension pairs. The mutual information calculation relies on the set of candidate values ​​within the associated dimension pair, joint probability estimation, and marginal probability to complete the operation, which is used to characterize the amount of information shared between a set of associated dimensions. The independent entropies corresponding to all preset dimensions are summed, and then the mutual information corresponding to all associated dimension pairs is subtracted to complete the aggregation operation. Finally, the global confidence discrete entropy is output. The global confidence discrete entropy is a global uncertainty metric obtained by comprehensively considering the uncertainty of the dimension itself and the correlation constraints between dimensions, which is used to characterize the uncertainty level of the current overall user intent state. This approach avoids the computational burden caused by the excessively large parameter space of the complete joint distribution. While statistically analyzing the inherent uncertainties of each dimension, it introduces mutual information to eliminate the overestimation problem caused by information overlap between related dimensions. This allows the global confidence discrete entropy to better fit the inherent dimensional dependencies and mutual exclusion constraints within the business scenario, more realistically and objectively reflecting the overall uncertainty of multi-dimensional intentions, and providing a reliable quantitative measurement basis for subsequent convergence condition judgments.

[0023] 105. Compare the global confidence discrete entropy with the convergence threshold, and compare each target confidence level with the peak score threshold to determine whether the preset convergence condition is met. 106. When the convergence condition is met, a structured user intent vector is output based on the initial intent deterministic heatmap.

[0024] In this embodiment, the global confidence discrete entropy can characterize the uncertainty of the current overall intent, the target confidence level represents the level of certainty of the optimal candidate value under each key dimension, the convergence threshold is the quantitative threshold for judging the uncertainty of the overall intent, and the peak score threshold is used to judge the level of certainty of the candidate value of a single key dimension. Both thresholds are configured in combination with the actual output requirements of the material testing business and the statistical results of historical business samples. The preset convergence condition requires that the global confidence discrete entropy is less than the convergence threshold, and that the target confidence level corresponding to all key dimensions is greater than the peak score threshold. The convergence condition is set according to the accuracy requirements of the intent recognition results in the business scenario.

[0025] When the comparison results meet the convergence condition, the score distribution of all preset dimensions in the initial intent deterministic heatmap is read, the candidate value with the highest score in each dimension and the corresponding confidence value are extracted, and the vector is arranged according to the dimensional arrangement order defined in the multidimensional intent space to complete the output. This output vector is the structured user intent vector. The structured user intent vector realizes the standardized encapsulation of detection requirement parameters, which can directly provide complete input for downstream solution recommendation engines and support subsequent business processing such as detection project matching and quotation generation.

[0026] When the comparison result determines that the convergence condition is not met, the key dimensions whose target confidence does not exceed the peak score threshold constraint are identified as non-converged dimensions. Multiple sets of candidate question directions are generated for the non-converged dimensions. Based on historical dialogue data, the candidate answers and their corresponding occurrence probabilities for each candidate question direction are obtained. The expected information gain of each set of candidate question directions is calculated, and the candidate question direction with the largest expected information gain is selected as the target question direction. Then, the target question is generated by combining the question granularity adaptation rule and the global confidence discrete entropy. An interactive inquiry is initiated to the user. After receiving the user's feedback, the non-converged dimensions associated with the feedback are determined as dimensions to be updated. The corresponding score adjustment is performed by distinguishing between positive and negative information types. Then, the dimension association rule is retrieved to determine the other preset dimensions with association relationships. Based on the association conditional probability matrix and the propagation intensity factor, cross-dimensional confidence propagation is completed, and the heatmap is updated to obtain an update intention deterministic heatmap. The structured entropy calculation is performed on the update intention deterministic heatmap to obtain the updated global confidence discrete entropy. The update target confidence is extracted simultaneously, and the convergence condition comparison is carried out again. The interactive update process is executed cyclically until the convergence condition is met.

[0027] This convergence judgment mechanism takes into account both the global uncertainty of the overall intent and the single-dimensional confidence level of key dimensions. The dual threshold constraint can avoid the defect of high confidence in local dimensions but still unclear overall intent. When the convergence condition is not met, the multi-round interaction process is initiated to filter questions with information gain as the guide, which can reduce the overall uncertainty with the fewest number of interactions. At the same time, it relies on the dimension association rule to complete the cross-dimensional propagation of confidence, and makes full use of domain business constraints to correct the scores of each dimension, ensuring that the recognition results are consistent with the business logic of the material testing field. The final output structured user intent vector has a standardized format that is readable by machines, which is convenient for direct connection to downstream business systems.

[0028] In this embodiment of the invention, the multidimensional intent space further includes multiple non-critical dimensions, with each non-critical dimension and each critical dimension respectively serving as a preset dimension, and each preset dimension including multiple candidate values; the step of performing semantic matching processing on the user's initial input text based on the multidimensional intent space to construct an initial intent deterministic heatmap includes: inputting the user's initial input text and all the candidate values ​​into a pre-trained language model for vector encoding processing to obtain semantic vectors and multiple candidate vectors corresponding to each preset dimension; calculating the matching score between the semantic vector and each candidate vector under the same preset dimension to obtain multiple matching scores corresponding to each preset dimension; normalizing each matching score under the same preset dimension to obtain multiple initial deterministic scores corresponding to each preset dimension; and performing heatmap assembly processing on all the initial deterministic scores based on the dimensional organization structure of the multidimensional intent space to construct the initial intent deterministic heatmap.

[0029] In this embodiment, the pre-trained language model can adopt the network architecture of BERT and its variants. This architecture relies on a large-scale general text corpus to complete the pre-training. During the pre-training stage, parameters are iteratively optimized through two types of tasks: masked language model and next sentence prediction. It learns the semantic representation ability of general language. For the material inspection business scenario, domain text can also be introduced for fine-tuning to further adapt to the language expression habits of professional fields.

[0030] In this embodiment, when the user's initial input text and all candidate values ​​are input into the pre-trained language model for vector encoding, the input text and each candidate value are first segmented into words, and after adding special markers specified by the model, they are converted into a number sequence that the model can recognize. The sequence is then fed into the model's network layer to complete multi-layer feature transformation, and finally outputs a dense vector of fixed dimensions. The vector obtained by encoding the user's initial input text is a semantic vector, and the vector generated after encoding each candidate value under each preset dimension is the multiple candidate vectors corresponding to the preset dimension.

[0031] In this embodiment, the matching score between the semantic vector and each candidate vector under the same preset dimension is calculated. Generally, cosine similarity or dot product operation is used to measure the distance in the vector space. The result is multiple matching scores corresponding to each preset dimension. The matching score can reflect the semantic similarity between the user's expression and the corresponding candidate value. The matching score under the same preset dimension is normalized using the softmax function. All matching scores within the dimension are mapped to a numerical range of 0 to 1. The sum of the scores is 1. The set of values ​​obtained after the normalization operation is the multiple initial deterministic scores corresponding to each preset dimension. The initial deterministic scores represent the credibility of each candidate value under the current user input conditions.

[0032] In this embodiment, the dimensional organization structure of the multidimensional intent space specifies the division method of all preset dimensions, distinguishes between key and non-key dimensions, records the candidate value set under each preset dimension and the pre-configured association constraints between dimensions, and performs heatmap assembly processing on all initial deterministic scores based on the dimensional organization structure of the multidimensional intent space. This means that according to the dimensional hierarchy defined by the organization structure, the initial deterministic scores scattered under each preset dimension are aggregated in an orderly manner, and the dimension identifier, candidate value and corresponding initial deterministic score are associated and bound to form a complete and unified initial intent deterministic heatmap. Candidate values ​​that are not directly mentioned by the user input will be assigned a lower initial deterministic score (e.g., can be set to 0.1) and will not be directly eliminated, but will be retained for subsequent confidence propagation calculations.

[0033] The entire processing flow relies on pre-trained language models to obtain deep semantic representations, breaking away from the limitations of traditional keyword matching which can only rely on literal character matching. It can capture the implicit semantic relationships within natural language. Vector-level similarity calculation can objectively measure the matching relationship between user needs and business candidate values. Softmax normalization provides a unified and comparable probabilistic dimension for different candidate values. Heatmap assembly processing organizes fragmented dimensional scores into structured data carriers, providing complete basic data for subsequent independent entropy and mutual information operations, convergence condition judgment, and confidence iteration updates under multiple rounds of interaction. It fully preserves the score distribution of all candidate values, so that the mutual exclusion and association rules between dimensions can take effect normally in subsequent processes, adapting to the complex business constraints in the field of materials testing.

[0034] In this embodiment of the invention, determining the target confidence level corresponding to each of the key dimensions based on the initial intent deterministic heatmap includes: obtaining multiple initial deterministic scores corresponding to each of the preset dimensions based on the initial intent deterministic heatmap; taking the initial deterministic score with the highest score under each preset dimension as the dimension confidence level corresponding to that preset dimension; extracting the dimension confidence level corresponding to each key dimension from the dimension confidence levels corresponding to all preset dimensions as the target confidence level, thereby obtaining the target confidence level corresponding to each of the key dimensions.

[0035] In this embodiment, the initial intent determinism heatmap is a data carrier formed by assembling a multi-dimensional intent space using a heatmap. Internally, it establishes a binding relationship between preset dimensions, candidate values, and initial determinism scores. Obtaining multiple initial determinism scores corresponding to each preset dimension based on the initial intent determinism heatmap involves traversing the heatmap's internal storage structure and, according to the preset dimension's division entries, reading the set of initial determinism scores corresponding to all candidate values ​​under each preset dimension. Using the initial determinism score with the highest value under each preset dimension as the dimension confidence level means performing a numerical comparison on each set of read preset dimension scores and selecting the initial determinism score with the largest value within the set. This peak score represents the level of confidence of the most likely candidate under that preset dimension, and is thus used as the overall dimension confidence level. Extracting the target confidence scores from the confidence scores of all preset dimensions for each key dimension relies on the key dimension labeling information predefined in the multidimensional intent space. From the calculated confidence scores of all preset dimensions, the results for non-key dimensions such as budget range and timeliness requirements are filtered out. Only the confidence scores corresponding to key dimensions such as material type, test object, test target, and testing standard are retained. This retained portion of the results is the target confidence score for each key dimension. The confidence scores for non-key dimensions are not extracted, but the original score data is not deleted and is still stored in the heatmap for subsequent uncertainty assessment and confidence propagation calculations. This processing method uses the peak score within a dimension to represent the credibility of a single dimension. It does not require complex distribution fitting calculations for all candidate values, resulting in low computational overhead and high execution efficiency. By distinguishing between key and non-key dimensions, it only uses the confidence of dimensions that play a decisive role in the generation of detection schemes and pricing parameters as the target confidence. This constrains the focus of subsequent convergence judgment calculations to core business parameters, avoiding unnecessary computational burdens caused by secondary parameters. At the same time, it completely preserves the original score data of the heatmap and does not discard the candidate scores of non-key dimensions, providing complete and usable basic data for subsequent structured entropy calculations and multi-round interactive updates.

[0036] In this embodiment of the invention, the multidimensional intent space further includes multiple non-critical dimensions. Each non-critical dimension and each critical dimension are respectively used as preset dimensions. The step of performing structured entropy calculation on the initial intent deterministic heatmap to obtain global confidence discrete entropy includes: obtaining preset dimension association rules based on the multidimensional intent space, and determining dimension pairs with direct correlation from all preset dimensions according to the dimension association rules to obtain multiple associated dimension pairs; obtaining multiple initial deterministic scores corresponding to each preset dimension based on the initial intent deterministic heatmap; performing independent entropy calculation on the corresponding preset dimensions based on the initial deterministic scores to obtain independent entropy corresponding to each preset dimension; performing mutual information calculation on the corresponding associated dimension pairs based on the initial deterministic scores to obtain mutual information corresponding to each associated dimension pair; and performing aggregation operation on all independent entropies and all mutual information to obtain the global confidence discrete entropy.

[0037] In this embodiment, obtaining preset dimension association rules based on a multi-dimensional intent space involves reading pre-stored business constraint configurations within the multi-dimensional intent space. Dimension association rules represent the constraint relationships between different preset dimension candidate values ​​in the materials testing field, including types such as mutual exclusion constraints and conditional dependency constraints. These rules are jointly set by domain experts based on testing business experience and statistical results of historical testing commission samples. For example, the mutual exclusion between the test target of mechanical properties and the sample morphology of powder is a type of dimension association rule. Determining dimension pairs with direct association relationships from all preset dimensions based on the dimension association rules involves traversing all dimension association rules and extracting the two sets of preset dimensions that have a constraint effect. Each set of two preset dimensions with a constraint binding relationship is a dimension pair with a direct association relationship. Multiple associated dimension pairs constitute the set of dimension pairs obtained from all screening. Obtaining multiple initial deterministic scores corresponding to each preset dimension based on the initial intent deterministic heatmap involves reading the structured stored data within the heatmap and extracting the set of initial deterministic scores corresponding to all candidate values ​​under each preset dimension.

[0038] The specific formula for calculating independent entropy of the corresponding preset dimensions based on the initial deterministic score is expressed as follows: , in, This represents the independent entropy corresponding to each preset dimension. This represents the candidate values ​​under the preset dimensions. Indicates candidate value The corresponding initial deterministic score, independent entropy is used to measure the degree of uncertainty brought about by the distribution of candidate values ​​within a single preset dimension. The more dispersed the distribution, the larger the value of the independent entropy. The calculation process no longer performs repeated normalization on the initial deterministic score. The base of the logarithm can be 2 or the natural constant e. Changing the base only causes the entropy value to scale proportionally and does not change the entropy ranking or the relative comparison relationship of the thresholds. Therefore, this invention does not limit the specific base of the logarithm. For ease of engineering implementation, this embodiment preferably uses a logarithm with base 2, in bits. The expression for the specific calculation formula for mutual information calculation based on the initial deterministic score pairs corresponding to the associated dimensions is as follows: , in, This represents the mutual information corresponding to each associated dimension pair. , Representing preset dimensions With preset dimensions The set of candidate values, Indicates the preset dimension One of the candidate values, Indicates the preset dimension One of the candidate values, Represents marginal probability, i.e., the preset dimension. candidate values The corresponding initial deterministic score, Represents marginal probability, i.e., the preset dimension. candidate values The corresponding initial deterministic score, Indicates the preset dimension Take candidate values Simultaneously preset dimensions Take candidate values The joint probability estimate is derived from the joint probability distribution estimate obtained from actual observations. It originates from the joint normalization of the initial deterministic scores of two predefined dimensions, thus reflecting the true dependencies between the dimensions. When the two predefined dimensions are independent, the mutual information value is 0; when there is a strong correlation between the dimensions, the mutual information value is significantly greater than 0. Mutual information is used to characterize the amount of information shared between the two sets of related dimensions. The higher the mutual information value, the stronger the constraint dependency between the two sets of dimensions. The base of the logarithm can be 2 or the natural constant e. Changing the base only causes the entropy value to scale proportionally and will not change the entropy value sorting or the relative comparison relationship of the threshold. Therefore, this invention does not limit the specific base of the logarithm. For ease of engineering implementation, this embodiment preferably uses a logarithm with base 2, and the unit is bits.

[0039] The aggregation operation is performed on all independent entropies and all mutual information, with the corresponding formula: , First, sum the independent entropies obtained from solving all preset dimensions. Then, sum the mutual information of all associated dimensions. Subtract the sum of mutual information from the sum of independent entropies. The final output is the global confidence discrete entropy. This represents the global confidence discrete entropy. This indicates traversing all preset dimensions. This represents the independent entropy corresponding to each preset dimension. This represents the set consisting of all related dimension pairs. This represents the mutual information corresponding to each associated dimension pair. Index markers representing related dimension pairs are used to refer to a set of predefined dimensions. With preset dimensions The dimensional pairings of the composition Indicates traversing a collection Each group consists of preset dimensions. With preset dimensions For each pair of related dimensions, the mutual information of the corresponding pair is summed.

[0040] The entire structured entropy calculation process adopts a pairwise factorization approximation approach, avoiding the problem of a surge in computational load caused by the excessively large scale of the complete joint distribution parameter space. Independent entropy can characterize the inherent uncertainty of each dimension, and mutual information can deduct the statistical bias caused by the overlap of information between related dimensions, avoiding overestimation of overall uncertainty. This allows the global confidence discrete entropy to fully fit the constraints in areas such as mutual exclusion and dependence in the material testing business, objectively restoring the true uncertainty level of the multi-dimensional intention state. The output global confidence discrete entropy can serve as a standardized quantitative indicator, providing reliable data support for the subsequent determination of convergence conditions.

[0041] In this embodiment of the invention, after comparing the global confidence discrete entropy with the convergence threshold and comparing each target confidence level with the peak score threshold to determine whether a preset convergence condition is met, the method further includes: when the convergence condition is not met, designating key dimensions whose target confidence level is less than or equal to the peak score threshold as non-converged dimensions; determining target questions based on each non-converged dimension and the global confidence discrete entropy; obtaining feedback information from user input based on the target questions; updating the initial intent deterministic heatmap based on the feedback information to obtain an updated intent deterministic heatmap; performing structured entropy calculation on the updated intent deterministic heatmap to obtain an updated global confidence discrete entropy; determining the updated target confidence level corresponding to each key dimension based on the updated intent deterministic heatmap; comparing the updated global confidence discrete entropy with the convergence threshold and comparing each updated target confidence level with the peak score threshold; and when the convergence condition is met, outputting a structured user intent vector based on the updated intent deterministic heatmap.

[0042] In this embodiment, when the convergence condition is not met, key dimensions with a target confidence level less than or equal to the peak score threshold are considered as non-converged dimensions. This means that after completing the convergence condition comparison, the target confidence levels corresponding to all key dimensions are traversed, and key dimensions whose target confidence levels do not exceed the peak score threshold are determined as non-converged dimensions. The confidence level of the optimal candidate value under this type of dimension has not yet reached the business usability standard, and further information needs to be collected through interactive questioning. Non-key dimensions are not subject to this determination logic, and uncertainty is allowed to be retained for subsequent business engine processing.

[0043] In this embodiment, the target question is determined based on each non-converged dimension and the global confidence discrete entropy. First, candidate question directions corresponding to each non-converged dimension are generated. Historical dialogue data stored in the system is read, and several candidate answers and their occurrence probabilities are obtained for each candidate question direction based on the historical dialogue data. For each candidate question direction, the expected information gain is calculated based on the occurrence probabilities of the candidate answers. The expected information gain represents the expected effect of the question in reducing the uncertainty of the overall intent space. The candidate question direction with the highest expected information gain is selected as the target question direction. Then, a preset question granularity adaptation rule is invoked, and the question form is dynamically matched based on the target question direction and the value of the global confidence discrete entropy. The final interactive question obtained after granularity adaptation is the target question.

[0044] In this embodiment, obtaining feedback information from the user based on the target question input means receiving the text content of the user's reply to the target question output by the system. This feedback information is used to supplement or correct the currently obtained detection requirement information. The process of updating the initial intent deterministic heatmap based on the feedback information to obtain an updated intent deterministic heatmap involves first identifying the non-converged dimension corresponding to the feedback information as the dimension to be updated, identifying the feedback type of the current feedback information (divided into positive and negative information types), and if it is a positive information type, locating the positive candidate value matching the feedback information within the dimension to be updated, and increasing the deterministic score corresponding to the positive candidate value according to a preset score increase rule to obtain an increased score. If it is a negative information type, locating the negative candidate value rejected by the user within the dimension to be updated, and simultaneously obtaining mutually exclusive candidate values ​​that are mutually exclusive with the negative candidate value, decreasing the deterministic score of the negative candidate value and increasing the deterministic score of the mutually exclusive candidate value based on a preset contradiction rule base, generating an adjusted score set.

[0045] In this embodiment, after completing the single-dimensional score correction, the dimension association rules stored within the multi-dimensional intent space are read, and other preset dimensions with association constraints with the dimension to be updated are found. The association conditional probability matrix pre-built by the system is called, and cross-dimensional confidence propagation calculation is performed in combination with the propagation intensity factor. The scores of each candidate value within the associated dimension are smoothly migrated and corrected according to the score update formula. All score change results are synchronously written into the initial intent deterministic heatmap. After all the above score updates, the updated intent deterministic heatmap is obtained. The updated intent deterministic heatmap is a complete structured data carrier after completing the absorption of user feedback information and cross-dimensional confidence propagation, and it saves the deterministic scores after updating all candidate values ​​under all preset dimensions. The process involves performing structured entropy calculation on the heatmap of intent determination to obtain the updated global confidence discrete entropy. Reusing the complete structured entropy calculation flow, the updated deterministic scores of each preset dimension in the heatmap are read. Dimension association rules are used to determine associated dimension pairs. Independent entropy calculations for each preset dimension and mutual information calculations for each associated dimension pair are performed sequentially. The updated global confidence discrete entropy is obtained by subtracting the sum of mutual information from the sum of independent entropy results. This updated global confidence discrete entropy represents the latest uncertainty of the overall intent space after one round of user interaction. Based on the heatmap of intent determination, the process of determining the update target confidence level corresponding to each key dimension is as follows: Each preset dimension within the heatmap is traversed, and the deterministic score with the largest value under each preset dimension is extracted as the dimension confidence level for that dimension. Then, the dimension confidence levels corresponding to all key dimensions are selected from all dimension confidence levels; this selection result is the update target confidence level. The process involves repeatedly executing the convergence condition verification logic by comparing the updated global confidence discrete entropy with the convergence threshold, and by comparing the confidence level of each update target with the peak score threshold. This process also verifies the uncertainty indicators at the global level and the confidence levels of each key dimension. When the global confidence discrete entropy is less than the convergence threshold and the confidence levels of all update targets are greater than the peak score threshold, the convergence condition is deemed met. The structured user intent vector is output based on the update intent deterministic heatmap. This involves reading the highest-scoring candidate value and its corresponding confidence value in each preset dimension of the update intent deterministic heatmap, assembling the candidate value identifier and confidence value into a machine-readable vector format according to the dimensional arrangement order specified in the multi-dimensional intent space. This vector encapsulates complete detection requirement parameters and can be directly input into the downstream solution recommendation engine to generate detection projects and quotation documents.

[0046] The entire interactive update process filters question directions based on information gain, reducing the overall uncertainty of the intent space with as few user interaction rounds as possible. It differentiates between positive and negative feedback and performs differentiated score adjustments. At the same time, it relies on the correlation conditional probability matrix to complete cross-dimensional confidence propagation, making full use of domain business constraints to synchronously correct the scores of correlation dimensions. This avoids excessive corrections caused by score mutations due to relying solely on single-round user feedback. The cyclic convergence verification mechanism ensures that results are only output when both global uncertainty and key dimension confidence meet the criteria. This effectively avoids the problem of high confidence in local dimensions but unclear overall needs. The output structured user intent vector has a standardized format, facilitating direct connection to downstream business processing modules.

[0047] In this embodiment of the invention, determining the target question based on each non-converged dimension and the global confidence discrete entropy includes: determining candidate question directions corresponding to each non-converged dimension and acquiring historical dialogue data; determining multiple candidate answers corresponding to each candidate question direction and the occurrence probability corresponding to each candidate answer based on the historical dialogue data; calculating the expected information gain corresponding to each candidate question direction based on the occurrence probability of each candidate answer for each candidate question direction; selecting the candidate question direction with the highest expected information gain as the target question direction; acquiring a preset question granularity adaptation rule, and determining the target question based on the question granularity adaptation rule, the target question direction, and the global confidence discrete entropy.

[0048] In this embodiment, candidate question directions corresponding to each non-converged dimension are determined, and historical dialogue data is obtained. For each key dimension determined to be non-converged, question directions for information collection from users are generated. Each non-converged dimension corresponds to a set of candidate question directions. Historical dialogue data comes from the pre-sales interaction records of material testing business stored in the system, containing all user response samples for various business issues in historical scenarios, used to statistically analyze the prior probability of user response behavior. Candidate question directions represent the business topic to be inquired about for a specific non-converged dimension. For example, question directions generating inquiries about test target information for the test target dimension belong to one type of candidate question direction. Historical dialogue data provides a statistical sample source for the entire information gain calculation. Determining multiple candidate answers for each candidate question direction based on historical dialogue data involves statistically analyzing the historical dialogue data to extract various business responses that the user has given in response to that candidate question direction. The candidate answer is the business value that the user may output for that question topic, corresponding to each candidate value in that dimension within the multi-dimensional intent space. The probability of occurrence of each candidate answer is the conditional probability of the user outputting that candidate answer under the current intent state, obtained by relying on historical dialogue statistics. The system maintains a response conditional probability table to store this type of probability. When the business system is cold-started or the business sample data is sparse, and it is impossible to obtain a reliable response conditional probability by relying on historical dialogue data statistics, a uniform distribution is used as the prior probability, that is, it is assumed that the probability of occurrence of all candidate answers under that candidate question direction is equal. This ensures that the calculation process of expected information gain can be executed normally and avoids algorithm failure due to missing samples.

[0049] In this embodiment, for each candidate question direction, the expected information gain corresponding to that candidate question direction is calculated based on the occurrence probability of each candidate answer. This involves traversing all candidate answers under that candidate question direction, combining the occurrence probability of each candidate answer, simulating the decrease in the global confidence discrete entropy of the overall intent space after the user selects that candidate answer, and weighting and summing the uncertainty reduction brought by different responses according to their corresponding occurrence probabilities. Finally, the expected information gain corresponding to that candidate question direction is obtained. The expected information gain can characterize the expected reduction in the overall intent space uncertainty after executing the question. The uncertainty reduction is calculated as follows: after simulating the selection of the candidate answer, the certainty score of the dimension corresponding to the candidate answer is first set to the highest confidence level. The highest confidence level is a value of 1 in the range of 0 to 1, representing that the candidate answer is completely certain. The scores of other related dimensions are adjusted synchronously according to the preset dimension association rules to obtain the simulated updated complete intent space score distribution. Then, the global confidence discrete entropy is recalculated based on the simulated updated complete intent space score distribution. Finally, the original global confidence discrete entropy before the question is subtracted from the updated global confidence discrete entropy, and the difference is the uncertainty reduction brought by the candidate answer. The probability of a candidate answer appearing is obtained by normalizing the deterministic score of that candidate value in the current multidimensional intent space. The candidate question direction with the highest expected information gain is selected as the target question direction. This is achieved by comparing the expected information gain values ​​calculated from all candidate question directions and choosing the question direction that minimizes overall intent uncertainty as the priority for this interaction. The target question direction is the business dimension theme corresponding to the information to be collected in this interaction.

[0050] In this embodiment, the preset question granularity adaptation rule is a set of business configuration rules that control the output format of questions. These rules are set by business developers based on user experience in material testing interaction scenarios. Internally, the rules establish a mapping relationship between the global confidence discrete entropy value range and the question format, used to select the matching question format based on the overall ambiguity of the intent. The specific process of determining the target question based on the question granularity adaptation rule, the target question direction, and the global confidence discrete entropy involves reading the currently calculated global confidence discrete entropy H value and comparing it with the value range within the question granularity adaptation rule to determine the ambiguity level of the overall intent. For example, the question granularity adaptation rule can be set as follows: when H is greater than 0.6, it is determined to be a high ambiguity level, generating an open-ended question. Open-ended questions do not have preset optional answers, allowing users to freely input their business needs, such as "What aspect of the material's performance do you mainly want to test?"; when H is greater than 0.3 and less than or equal to 0.6, it is determined to be a medium certainty level, generating a closed-ended question, providing users with limited business options to choose from. For example, "Are you more concerned about mechanical properties or component analysis?"; when H is less than or equal to 0.3, it is judged as a low fuzziness but lacking detail level, generating fill-in-the-blank follow-up questions to confirm the remaining few unclear business details, such as "Do you need to follow GB / T or ASTM standards?". The business theme corresponding to the target question direction is filled into the corresponding sentence template, and the final target question output to the user is assembled. The sentence template is a set of question frames with placeholders that are pre-stored by the system, configured according to open-ended options, closed-ended options, and fill-in-the-blank follow-up questions. The sentence template is set with placeholders for filling in business themes; the business theme is the business dimension information to be collected corresponding to the target question direction. The entire processing flow utilizes historical dialogue data to obtain the prior statistical distribution of user responses. By optimizing the expected information gain, it selects the optimal question direction and prioritizes interactions based on dimensions that minimize uncertainty. This effectively reduces the number of human-computer interaction rounds. Furthermore, it uses numerical adaptive matching of question formats based on global confidence discrete entropy. In the high ambiguity stage (H > 0.6), open-ended questions are used to fully collect user information. In the moderately certain stage (H between 0.3 and 0.6), closed-ended questions are used to narrow down the selection range. In the low ambiguity stage (H ≤ 0.3), fill-in-the-blank follow-up questions are used to complete the remaining business details. This approach balances information collection efficiency and user interaction experience, providing reliable interactive input for subsequent user feedback and iterative updates to the heatmap score.

[0051] In this embodiment of the invention, updating the initial intent deterministic heatmap based on the feedback information to obtain an updated intent deterministic heatmap includes: taking the non-converged dimension corresponding to the feedback information as the dimension to be updated, and determining the feedback type of the feedback information; when the feedback type is a positive information type, determining the positive candidate value corresponding to the feedback information based on the dimension to be updated, and adjusting the initial deterministic score corresponding to the positive candidate value based on a preset score adjustment rule to obtain an adjusted score; when the feedback type is a negative information type, determining the negative candidate value corresponding to the feedback information and the negative candidate value based on the dimension to be updated. The mutually exclusive candidate values ​​corresponding to the negative candidate values ​​are adjusted based on a preset contradiction rule. The initial deterministic scores corresponding to the negative candidate values ​​are adjusted downwards, and the initial deterministic scores corresponding to the mutually exclusive candidate values ​​are adjusted upwards, resulting in an adjusted score set. Based on the multidimensional intent space, preset dimension association rules are obtained, and other preset dimensions that are associated with the dimension to be updated are determined according to the dimension association rules. Based on the adjusted score set or the adjusted score set, combined with a preset association conditional probability matrix and a propagation strength factor, confidence propagation processing is performed on the other preset dimensions to update the initial intent deterministic heatmap, resulting in the updated intent deterministic heatmap.

[0052] In this embodiment, the non-converged dimension corresponding to the feedback information is designated as the dimension to be updated. Determining the feedback type of the feedback information involves locating the non-converged dimension to which the current user feedback information belongs; this dimension is the dimension to be updated. The dimension to be updated is the business dimension whose score needs to be corrected after receiving user feedback. The feedback type is a categorization of the user feedback content, used to select different subsequent score correction logics. A positive information type represents feedback content where the user confirms or supplements business requirements, used to confirm the validity of a candidate value under the dimension to be updated. Determining the positive candidate value corresponding to the feedback information based on the dimension to be updated involves matching the value with the same business meaning as the user feedback content from all candidate values ​​of the dimension to be updated. The positive candidate value is the dimension candidate value confirmed as valid by the user feedback. The preset score adjustment rule is a pre-configured score adjustment strategy, set according to the business scenario and score value range. It is used to limit the magnitude and upper limit of the score adjustment for affirmative candidate values, preventing the score from exceeding the legal value range. The process of adjusting the initial deterministic score corresponding to the affirmative candidate value based on the preset score adjustment rule involves reading the current initial deterministic score of the affirmative candidate value, raising the score according to the adjustment magnitude given by the score adjustment rule, while constraining the score to not exceed the upper limit. The score result after the adjustment is the adjusted score. The negative information type represents the feedback content of the user's denial or correction, used to indicate that a certain candidate value in the dimension to be updated is not true. Based on the dimension to be updated, the negative candidate value and the mutually exclusive candidate value corresponding to the feedback information are determined. This involves matching the value that the user explicitly denies from the candidate values ​​of the dimension to be updated as the negative candidate value, and then finding the value that cannot be true at the same time as the negative candidate value in business logic as the mutually exclusive candidate value. The negative candidate value is the dimension candidate value that the user explicitly denies in the feedback, and the mutually exclusive candidate value is the candidate value that is mutually exclusive with the negative candidate value in business logic, and the two cannot be true at the same time. The preset contradiction rules are a set of rules that record the mutual exclusion constraints between candidate values ​​within a dimension. They are constructed and set based on the business experience of domain experts. The process of adjusting the initial deterministic scores corresponding to negative candidate values ​​and adjusting the initial deterministic scores corresponding to mutually exclusive candidate values ​​based on the preset contradiction rules involves reading the initial deterministic scores of both negative and mutually exclusive candidate values, reducing the score of negative candidate values ​​according to the adjustment range set by the contradiction rules, and simultaneously increasing the score of mutually exclusive candidate values. This process rewrites the scores of multiple candidate values ​​within a single dimension. After this step, all the modified score data are obtained, forming the adjusted score set.

[0053] In this embodiment, obtaining preset dimension association rules based on the multidimensional intent space involves reading the business association constraints between different business dimensions stored in the multidimensional intent space, and determining other preset dimensions that are associated with the dimension to be updated according to the dimension association rules. This involves traversing all business dimensions and filtering out dimensions that are coupled with the dimension to be updated in terms of business logic. The dimensions obtained from this filtering are the other preset dimensions.

[0054] Based on the score adjustment or score set adjustment, combined with a preset correlation conditional probability matrix and propagation strength factor, the confidence propagation process is applied to other preset dimensions. This process occurs after the score of the dimension to be updated has been modified, leveraging the business relationships between dimensions to diffuse the score change to other related preset dimensions, achieving synchronous cross-dimensional score correction. The preset correlation conditional probability matrix is ​​constructed based on domain expert experience and historical testing commission data statistics. The matrix internally records the conditional probability between any two candidate values ​​from different dimensions. , representing the candidate values ​​of the dimension to be updated When established, another candidate value for the associated dimension (other preset dimensions) corresponding to the dimension to be updated. Probability of occurrence, propagation intensity factor This is used to control the correction magnitude for cross-dimensional confidence propagation. The default value is 0.3, but it can also be dynamically adjusted based on the updated confidence level. When the candidate values ​​of the dimension to be updated... When the initial deterministic score is increased or decreased, in another preset dimension corresponding to the dimension to be updated... The score update formula is: , in, Represents candidate values ​​for other preset dimensions The updated deterministic score, Represents candidate values ​​for other preset dimensions The initial deterministic score, Indicates the propagation intensity factor. This represents the conditional probability recorded in the associated conditional probability matrix. The above score update process is the confidence propagation process. This formula uses the current old score as a basis, combined with the propagation strength factor, to smoothly migrate towards the target value corresponding to the conditional probability. The greater the difference between the current score of the candidate value and the target value of the conditional probability, the greater the magnitude of this correction. This corresponds to the score being raised when the business supports the candidate value, and the score being suppressed when the business does not support the candidate value, avoiding excessive score correction caused by a single update. The updated scores of all candidate values ​​obtained through this formula are written into the initial intent deterministic heatmap to complete the heatmap update and obtain the updated intent deterministic heatmap.

[0055] This processing flow distinguishes between two sets of single-dimensional score correction logics for positive and negative information types, adapting to different forms of user feedback input. It leverages contradictory rules to achieve coordinated adjustments of mutually exclusive candidate values ​​within each dimension, while simultaneously achieving smooth cross-dimensional confidence propagation through the associated conditional probability matrix and propagation strength factor. It not only modifies the dimensions to be updated based on direct user feedback but also synchronously corrects the scores of candidate values ​​in other associated dimensions using business association constraints. This ensures that confidence changes in each dimension conform to business logic. The propagation strength factor constrains the magnitude of each cross-dimensional correction, preventing sudden score changes caused by a single user feedback, thus improving the stability of the overall intent recognition results. The updated intent deterministic heatmap fully preserves the deterministic scores after all dimensions have been updated, providing a complete data foundation for subsequent global entropy recalculation and convergence condition verification.

[0056] The user demand intent recognition method in the embodiments of the present invention has been described above. The user demand intent recognition device in the embodiments of the present invention will be described below. Please refer to [link / reference]. Figure 2 One embodiment of the user demand intent recognition device in this invention includes: Data acquisition module 201: used to acquire the user's initial input text, a pre-constructed multidimensional intent space, a preset convergence threshold and a peak score threshold, wherein the multidimensional intent space includes multiple key dimensions; Semantic matching module 202: used to perform semantic matching processing on the user's initial input text based on the multidimensional intent space, so as to construct an initial intent deterministic heatmap; Confidence determination module 203: used to determine the target confidence level corresponding to each of the key dimensions based on the initial intent deterministic heatmap; Structured entropy calculation module 204: used to perform structured entropy calculation on the initial intention deterministic heatmap to obtain global confidence discrete entropy; Comparison module 205: used to compare the global confidence discrete entropy with the convergence threshold, and to compare each target confidence level with the peak score threshold respectively, so as to determine whether the preset convergence condition is met; Output module 206: When the convergence condition is met, output a structured user intent vector based on the initial intent deterministic heatmap.

[0057] Based on the same ideas as the methods in the above embodiments, the apparatus provided in this application can implement the methods in the above embodiments.

[0058] above Figure 2 The user intent recognition device in this embodiment of the invention is described in detail from the perspective of modular functional entities. The user intent recognition device in this embodiment of the invention is described in detail below from the perspective of hardware processing.

[0059] Figure 3 This is a schematic diagram of the structure of a user intent recognition device 300 provided in an embodiment of the present invention. The user intent recognition device 300 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors) and a memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) for storing application programs 333 or data 332. The memory 320 and storage media 330 can be temporary or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the user intent recognition device 300. Furthermore, the processor 310 may be configured to communicate with the storage media 330 and execute a series of instruction operations on the storage media 330 on the user intent recognition device 300 to implement the steps of the user intent recognition method provided in the above-described method embodiments.

[0060] The user intent recognition device 300 may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 3 The illustrated user intent recognition device structure does not constitute a limitation on the user intent recognition device, and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0061] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the user demand intent recognition method.

[0062] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0063] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0064] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for recognizing user intent, characterized in that, include: The system acquires the user's initial input text, a pre-constructed multidimensional intent space, a preset convergence threshold, and a peak score threshold. The multidimensional intent space includes multiple key dimensions. Based on the multidimensional intent space, semantic matching processing is performed on the user's initial input text to construct an initial intent deterministic heatmap; Based on the initial intent deterministic heatmap, determine the target confidence level corresponding to each of the key dimensions; The initial intention deterministic heatmap is processed by structured entropy calculation to obtain global confidence discrete entropy; The global confidence discrete entropy is compared with the convergence threshold, and each target confidence level is compared with the peak score threshold to determine whether the preset convergence condition is met. When the convergence condition is met, a structured user intent vector is output based on the initial intent deterministic heatmap.

2. The user demand intent recognition method according to claim 1, characterized in that, The multidimensional intent space also includes multiple non-critical dimensions, with each non-critical dimension and each critical dimension respectively serving as a preset dimension, and each preset dimension including multiple candidate values; the semantic matching processing of the user's initial input text based on the multidimensional intent space to construct an initial intent deterministic heatmap includes: The user's initial input text and all the candidate values ​​are respectively input into a pre-trained language model for vector encoding processing to obtain semantic vectors and multiple candidate vectors corresponding to each preset dimension; Calculate the matching score between the semantic vector and each candidate vector under the same preset dimension to obtain multiple matching scores corresponding to each preset dimension; Normalize each matching score under the same preset dimension to obtain multiple initial deterministic scores corresponding to each preset dimension; Based on the dimensional organization structure of the multidimensional intent space, heatmap assembly processing is performed on all the initial deterministic scores to construct the initial intent deterministic heatmap.

3. The user demand intent recognition method according to claim 2, characterized in that, The step of determining the target confidence level corresponding to each of the key dimensions based on the initial intent deterministic heatmap includes: Based on the initial intent deterministic heatmap, multiple initial deterministic scores corresponding to each of the preset dimensions are obtained; The highest initial certainty score under each preset dimension is taken as the dimension confidence level corresponding to that preset dimension. Extract the dimension confidence scores corresponding to each key dimension from the dimension confidence scores corresponding to all preset dimensions, and obtain the target confidence scores corresponding to each key dimension.

4. The user demand intent recognition method according to claim 1, characterized in that, The multidimensional intent space also includes multiple non-critical dimensions. Each non-critical dimension and each critical dimension are respectively used as preset dimensions. The process of performing structured entropy calculation on the initial intent deterministic heatmap to obtain global confidence discrete entropy includes: Based on the multidimensional intent space, preset dimension association rules are obtained, and dimension pairs with direct relationships are determined from all preset dimensions according to the dimension association rules, resulting in multiple associated dimension pairs. Based on the initial intent deterministic heatmap, multiple initial deterministic scores corresponding to each of the preset dimensions are obtained; Based on the initial deterministic score, independent entropy calculation is performed on the corresponding preset dimensions to obtain the independent entropy corresponding to each preset dimension; Based on the initial deterministic score, mutual information is calculated for the corresponding association dimension pairs to obtain the mutual information corresponding to each association dimension pair; The global confidence discrete entropy is obtained by performing aggregation operations on all the independent entropies and all the mutual information.

5. The user demand intent recognition method according to claim 1, characterized in that, After comparing the global confidence discrete entropy with the convergence threshold, and comparing each target confidence level with the peak score threshold to determine whether the preset convergence condition is met, the method further includes: When the convergence condition is not met, the key dimension whose target confidence is less than or equal to the peak score threshold is regarded as the non-converged dimension. The target question is determined based on each of the non-converged dimensions and the global confidence discrete entropy. Obtain feedback information from the user's question input based on the target, and update the initial intent deterministic heatmap based on the feedback information to obtain an updated intent deterministic heatmap; The structured entropy calculation is performed on the updated intention deterministic heatmap to obtain the updated global confidence discrete entropy; Based on the aforementioned heatmap of update intent, determine the confidence level of the update target corresponding to each of the key dimensions; The updated global confidence discrete entropy is compared with the convergence threshold, and each updated target confidence is compared with the peak score threshold. When the convergence condition is met, a structured user intent vector is output based on the updated intent deterministic heatmap.

6. The user demand intent recognition method according to claim 5, characterized in that, The process of determining the target question based on each of the non-converged dimensions and the global confidence discrete entropy includes: Determine the candidate question direction corresponding to each of the aforementioned non-converged dimensions, and obtain historical dialogue data; Based on the historical dialogue data, multiple candidate answers corresponding to each candidate question direction are determined, as well as the probability of occurrence corresponding to each candidate answer; For each candidate question direction, the expected information gain corresponding to that candidate question direction is calculated based on the occurrence probability of each candidate answer. The candidate questioning direction with the highest expected information gain value is taken as the target questioning direction; Obtain preset question granularity adaptation rules, and determine the target question based on the question granularity adaptation rules, the target question direction, and the global confidence discrete entropy.

7. The user demand intent recognition method according to claim 5, characterized in that, The step of updating the initial intent deterministic heatmap based on the feedback information to obtain an updated intent deterministic heatmap includes: The non-converged dimension corresponding to the feedback information is taken as the dimension to be updated, and the feedback type of the feedback information is determined. When the feedback type is positive information type, a positive candidate value corresponding to the feedback information is determined based on the dimension to be updated, and the initial deterministic score corresponding to the positive candidate value is adjusted upward based on the preset score adjustment rule to obtain the adjusted score. When the feedback type is a negative information type, the negative candidate value corresponding to the feedback information and the mutually exclusive candidate value corresponding to the negative candidate value are determined based on the dimension to be updated. The initial deterministic score corresponding to the negative candidate value is down-adjusted based on the preset contradiction rule, and the initial deterministic score corresponding to the mutually exclusive candidate value is up-adjusted to obtain the adjusted score set. Based on the multidimensional intent space, preset dimension association rules are obtained, and other preset dimensions that are associated with the dimension to be updated are determined according to the dimension association rules. Based on the upward score or the set of adjusted scores, combined with the preset association conditional probability matrix and propagation strength factor, confidence propagation processing is performed on the other preset dimensions to update the initial intent deterministic heatmap, thereby obtaining the updated intent deterministic heatmap.

8. A user demand intent recognition device, characterized in that, include: Data acquisition module: used to acquire the user's initial input text, a pre-constructed multidimensional intent space, a preset convergence threshold and a peak score threshold, wherein the multidimensional intent space includes multiple key dimensions; Semantic matching module: used to perform semantic matching processing on the user's initial input text based on the multidimensional intent space, so as to construct an initial intent deterministic heatmap; Confidence determination module: used to determine the target confidence level corresponding to each of the key dimensions based on the initial intent deterministic heatmap; Structured entropy calculation module: used to perform structured entropy calculation on the initial intention deterministic heatmap to obtain global confidence discrete entropy; The comparison module is used to compare the global confidence discrete entropy with the convergence threshold, and to compare each target confidence level with the peak score threshold to determine whether the preset convergence condition is met. Output module: When the convergence condition is met, output a structured user intent vector based on the initial intent deterministic heatmap.

9. A user demand intent recognition device, characterized in that, The user demand intent recognition device includes: a memory and at least one processor, wherein the memory stores instructions; At least one of the processors invokes the instructions in the memory to cause the user demand intent recognition device to perform the steps of the user demand intent recognition method as described in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the steps of the user demand intent recognition method as described in any one of claims 1-7.

Images