A self-evolving process entry routing closed-loop method and system based on user feedback

Abstract

The present application relates to the technical field of intelligent information processing, and particularly relates to a self-evolution process entry routing closed loop method and system based on user feedback. The method constructs a two-dimensional dynamic word library containing process keywords and fuzzy synonyms; extracts core keywords from user natural language; respectively calculates the first and second hit rates of the core keywords and the two-dimensional word library, and obtains a comprehensive matching score by weighted fusion based on two-dimensional weight coefficients; returns a recommended result to the user according to the score and a sorting threshold value, and obtains interactive feedback data; and finally dynamically updates the underlying word library according to the feedback data, and adaptively adjusts the weight coefficients and the sorting threshold value in combination with business working condition demand data. The present application breaks the barrier of traditional single precise matching, realizes self-evolution of the routing mechanism without manual intervention through data closed loop, and significantly improves the routing accuracy under complex context and dynamic business working conditions.

Description

Technical Field

[0001] This invention relates to the field of intelligent information processing technology, and in particular to a self-evolving process entry routing closed-loop method and system based on user feedback. Background Technology

[0002] With the deepening of enterprise digital transformation, the number of business processes carried within modern management information systems is growing exponentially. Faced with a massive number of process nodes, enabling users to quickly and accurately find the entry point of the target business process (i.e., process routing) through natural language is key to improving system usability and office efficiency. Currently, traditional process entry point location mechanisms mainly rely on fixed hierarchical navigation or keyword-based precise matching retrieval methods, which are insufficient to meet users' diverse needs for natural language descriptions and intelligent interactive experiences.

[0003] Meanwhile, some related technical solutions have been developed in the fields of semantic understanding and intelligent matching. A Chinese patent application with application number 2021115303721 and publication number CN114238568A, entitled "A Method, System, and Terminal Device for Acquiring Teacher Resources," discloses a method for acquiring teacher resources. This method constructs a demand synonym tree and a resource synonym tree, and calculates the total matching degree based on the matching degree and preset weights to achieve resource recommendation. The Chinese patent application number is 202410538343.7, the publication number is CN118468881A, and the invention title is "A Semantic Retrieval Method and System for Automatically Extracting Keywords". It proposes a semantic retrieval method for automatically extracting keywords. By preprocessing and segmenting the text, combining the word frequency-inverse text frequency value to calculate the comprehensive score, and expanding the synonyms, the retrieval accuracy is improved. The Chinese patent application number is 202110019975.9, the publication number is CN112699663A, and the invention title is "A Semantic Understanding System Based on the Combination of Multiple Algorithms". It describes a semantic understanding system based on the combination of multiple algorithms. The system combines semantic grammar algorithms with deep learning algorithms and realizes semantic matching by calculating similarity scores. The Chinese patent application number is 202510431046.7, the publication number is CN119938902B, and the invention title is "An Automatic Recommendation Method for Standard Terms". It discloses an automatic recommendation method for standard terms, which uses a knowledge graph and gated attention fusion mechanism to process user input and searches for candidate recommended terms through a graph sampling algorithm. The Chinese patent application number is 202010758817.0, the publication number is CN111899829B, and the invention title is "A Full-Text Search Matching Engine Based on ICD9 / 10 Word Segmentation Thesaurus". It provides a full-text search matching engine based on ICD9 / 10 word segmentation thesaurus, which forms a professional thesaurus through natural language processing methods such as keyword extraction and part-of-speech tagging, and configures a personalized analyzer to improve the matching accuracy.

[0004] However, existing process entry routing methods still have the following technical shortcomings: First, it is unavoidable that different users' descriptions are highly diverse and ambiguous. Users from different departments and positions have significant differences in their natural language descriptions (especially colloquial expressions) of the same business process. Traditional routing mechanisms mainly rely on precise keyword matching, which is difficult to handle semantic ambiguity. This results in extremely low hit and recall rates when faced with users' colloquial search requests. Second, existing routing thesauruses are mostly statically configured. When the system encounters mismatches or recommendation failures, there is a lack of an effective user feedback loop mechanism, and the system cannot automatically recover from errors. Learning and correcting matching rules by mistake has resulted in the routing dictionary being updated in a way that lags far behind changes in business language, making it highly dependent on regular manual review and hard-coded maintenance. Finally, the existing matching rules are rigid and cannot adapt to dynamic business needs, especially since the business process system has different "business condition requirements" for routing response speed and accuracy at different times (such as peak weekday periods and end-of-month non-working day closing periods). The existing method uses a fixed matching threshold and a fixed weight scoring formula, which cannot be adaptively and dynamically adjusted according to real-time system operating indicators and user input habits, making it difficult to achieve the optimal performance balance under complex business conditions. Summary of the Invention

[0005] To address the technical problems of existing process entry routing technologies, such as low matching rates due to differences in user descriptions, lack of closed-loop feedback leading to inability to self-evolve, and rigid matching rules that cannot adapt to dynamic working conditions, this invention provides a self-evolving process entry routing closed-loop method and system based on user feedback.

[0006] One method for closed-loop routing of a self-evolving process entry point based on user feedback includes the following steps: S1. Based on the working condition requirements data of the target business process, construct a process database and a dynamic thesaurus. The dynamic thesaurus includes a list of process keywords and a list of process fuzzy synonyms corresponding to each candidate process in the process database. S2. Obtain the natural language description data input by the user, and perform semantic segmentation and intent recognition on the natural language description data to extract the core keyword set; S3. Based on the core keyword set, calculate the first hit rate of the core keyword set and the process keyword list of each candidate process in the process database, and the second hit rate of the core keyword set and the process fuzzy synonym list of each candidate process in the process database. S4. Based on a preset first weight coefficient for the list of keywords for the process and a preset second weight coefficient for the list of fuzzy synonyms for the process, the first hit rate and the second hit rate are weighted and fused to obtain the comprehensive matching score of each candidate process. S5. Based on the comprehensive matching score of each candidate process and the preset sorting threshold, the candidate processes are filtered and sorted, and the recommendation results are returned to the user. The user click-through rate, the accuracy of matching user input words with the recommendation results are obtained as user feedback data. S6. Dynamically update the list content of each candidate process in the dynamic thesaurus based on the feedback data, and adaptively adjust the first weight coefficient, the second weight coefficient, and the sorting threshold based on the feedback data and the working condition requirement data of the target business process.

[0007] Furthermore, step S2 involves extracting the core keyword set by processing the natural language description through a word segmentation component and using the TF-IDF algorithm to extract words with semantic features to form the core keyword set.

[0008] Furthermore, step S3 includes the following sub-steps: S301. Based on the core keyword set, use the edit distance algorithm or string similarity algorithm to calculate the similarity between the input words in the core keyword set and the preset words in the dynamic vocabulary; S302. The first hit rate and the second hit rate are calculated using the hit rate calculation formula; the hit rate calculation formula is expressed as follows: ; in, This represents the first hit rate or the second hit rate calculated for a single candidate process. Represents the first in the set of core keywords One input word, This represents the total number of preset words in the list of process keywords or the list of process fuzzy synonyms to be matched. Represents the first in the corresponding list One keyword, Represents the similarity function. This represents the total number of word segments in the core keyword set.

[0009] Furthermore, step S4 involves weighted fusion of the first hit rate and the second hit rate to obtain the comprehensive matching score for each candidate process, the calculation formula of which is expressed as follows: ; ; in, This represents the overall matching score. Indicates the first hit rate. Indicates the second hit rate. This represents the first weighting coefficient. This represents the second weighting coefficient. The activation function is used to map and normalize the weighted calculation result to the (0,1) interval. The activation function is either the Sigmoid function or the ReLU function.

[0010] Furthermore, step S6, which involves dynamically updating the list of each candidate process in the dynamic lexicon based on the feedback data, includes: The feedback data is analyzed. If the feedback data indicates that the user finally confirmed that the target process in the recommendation results was selected, it is determined to be a positive sample and positive sample is updated. The input words in the feedback data are added to the process keyword list or the process fuzzy synonym list of the target process, and the association weight parameter between the input words and the target process is increased. If the feedback data indicates that the user did not select the recommended process ranked first in the recommendation results, but instead selected other subsequent actual processes, it is determined as a negative sample and a negative sample update is performed. The association weight parameter between the input word in the feedback data and the recommended process ranked first in the recommendation results is reduced, and the input word is directly added to the fuzzy synonym list of the actual process.

[0011] Furthermore, based on the accumulated feedback data within a fixed time window, a pre-trained semantic representation model is used to perform cluster analysis on the accumulated feedback data to obtain words in the accumulated feedback data that have potential semantic association features with existing words in the dynamic lexicon, and these words are marked as mismatched words. The statistical number of times negative sample updates are triggered within the corresponding fixed time window is monitored. If the statistical number reaches a preset anti-oscillation threshold, the mismatched words within the current fixed time window are determined to be valid words, and the list of fuzzy synonyms in the process is updated, while the corresponding weight parameters are reset. The pre-trained semantic representation model is either a Word2Vec model or a BERT model.

[0012] Furthermore, step S6, which involves adaptively adjusting the first weight coefficient, the second weight coefficient, and the ranking threshold, includes: updating the target business process's operational requirements data by real-time monitoring of user click-through rate and recommendation result matching accuracy; dynamically adjusting the ranking threshold based on the updated operational requirements data; and determining the degree of colloquial input based on user input words. If the degree of colloquial input is determined to be too high, the second weight coefficient is increased and the first weight coefficient is correspondingly decreased; otherwise, the first weight coefficient is increased and the second weight coefficient is correspondingly decreased.

[0013] The present invention also provides a self-evolving process entry routing closed-loop system based on user feedback. The system is implemented based on any of the self-evolving process entry routing closed-loop methods based on user feedback described above. It includes a lexicon construction module, which is used to construct a process database and a dynamic lexicon containing information on multiple candidate processes, and to assign a corresponding process keyword list and a process fuzzy synonym list to each candidate process. The intent recognition module is used to receive natural language descriptions input by users, perform semantic word segmentation and intent recognition algorithms, and output a set of core keywords after filtering out stop words; The hit rate calculation module is used to receive the core keyword set, and call the similarity algorithm to calculate the first hit rate between the core keyword set and the process keyword list, and the second hit rate between the core keyword set and the process fuzzy synonym list. The weighted fusion module is used to obtain the set first weight coefficient and second weight coefficient, and to combine and weight the first hit rate and the second hit rate to calculate the comprehensive matching score of the corresponding candidate process. The ranking and recommendation module is used to compare the comprehensive matching scores of all candidate processes with a preset ranking threshold, perform filtering and descending sorting, and output the final recommendation results to the user. The adaptive evolution module is used to capture user feedback behavior data, trigger the closed-loop update logic of the list content of the lexicon construction module, and, in combination with system operation indicators, send an instruction to the weighted fusion module to adjust the ratio of the first weight coefficient and the second weight coefficient, and send an instruction to the ranking recommendation module to modify the ranking threshold.

[0014] A computer-readable storage medium for storing a computer program that, when run on a computer, causes the computer to execute any of the above-mentioned self-evolving process entry routing closed-loop methods based on user feedback.

[0015] An electronic device includes: a memory for storing a computer program; and a processor for executing the computer program to implement a user feedback-based self-evolving process entry routing closed-loop method as described above.

[0016] Compared with the prior art, the present invention has the following advantages: This invention constructs a dual-dimensional dynamic lexicon of process keywords and fuzzy synonyms, combined with word segmentation technology and a multi-dimensional weighted fusion scoring mechanism based on first and second hit rates. This effectively balances matching accuracy with the tolerance for colloquial expressions, significantly improving the accuracy of process routing in complex contexts. Secondly, by constructing a complete positive and negative sample association mechanism based on user feedback loops, the lexicon and weights are updated in real time. Furthermore, a pre-trained semantic representation model combined with a threshold for anti-oscillation counts within a time window is introduced, enabling the system to possess a self-learning ability that improves accuracy with use, automatically adapting to massive daily data. By drawing inferences from frequent interactions, the system can uncover potential synonymous business terms and update the thesaurus in batches, achieving self-learning and avoiding the cost of manual maintenance required by traditional static thesauruses. On the other hand, this invention introduces an adaptive adjustment mechanism based on working condition demand data (such as concurrency pressure and response requirements) and user click-through rate and accuracy feedback. This enables the system to dynamically adjust the sorting threshold according to different working conditions during peak or off-peak periods, and can intelligently identify the user's colloquialism level to dynamically allocate the weight coefficients of the two-dimensional dynamic thesaurus, ensuring that the system can output the optimal routing recommendation strategy under any working condition and user habits. Attached Figure Description

[0017] Figure 1 This is a flowchart of a self-evolving process entry routing closed-loop method based on user feedback proposed in an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of the structure of a self-evolving process entry routing closed-loop system based on user feedback proposed in an embodiment of the present invention.

[0019] Figure 3 This is a schematic diagram of the terminal device structure of a self-evolving process entry routing closed-loop method based on user feedback proposed in an embodiment of the present invention.

[0020] Figure 4 This is a computer-readable storage medium structure diagram of a self-evolving process entry routing closed-loop method based on user feedback proposed in an embodiment of the present invention.

[0021] In the diagram, 200 is the terminal device, 210 is the memory, 211 is the RAM, 212 is the cache memory, 213 is the ROM, 214 is the program / utility, 215 is the program module, 220 is the processor, 230 is the bus, 240 is the external device, 250 is the I / O interface, 260 is the network adapter, and 300 is the program product. Detailed Implementation

[0022] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, and not all of them. The components of the embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0024] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention. It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0025] Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or machine that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or machine. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or machine that includes said element.

[0026] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0027] Example 1: like Figure 1 As shown, this embodiment provides a self-evolving process entry route closed-loop method based on user feedback, including the following steps: S1. Based on the working condition requirements data of the target business process, construct a process database and a dynamic thesaurus. The dynamic thesaurus includes a list of process keywords and a list of process fuzzy synonyms corresponding to each candidate process in the process database. Specifically, this step addresses the technical shortcomings of traditional process routing systems, such as rigid thesaurus and overly coarse matching granularity, by constructing a process data thesaurus based on actual business conditions at the system's underlying layer. Existing technologies mostly rely on single, static, precise matching thesaurus, resulting in extremely low recall rates and difficulty in maintenance when faced with diverse expressions. This method breaks down the precise matching barrier of traditional information retrieval through a dual-dimensional dynamic thesaurus architecture of precise keywords + fuzzy synonyms, building a highly scalable and inclusive underlying data foundation for subsequent fault-tolerant semantic matching and the system's self-evolution.

[0028] Specifically, the implementation principle described above is as follows: First, the system retrieves the operational requirements data of the target business process (such as the importance of the process, historical concurrency frequency, etc.) as an initialization benchmark, and registers the metadata of all candidate processes in the database. Then, for each candidate process, its relevant professional terms are entered into a process keyword list, and common colloquial expressions and historical slang terms are entered into a process fuzzy synonym list. Furthermore, it should be noted that because the processes in modern enterprise management systems are complex and user language varies from person to person, by performing a dual-list dimensionality reduction split on the thesaurus at the underlying level, not only is structural decoupling of precise semantics and fuzzy semantics achieved at the physical storage level, but also the necessary data structure support is provided at the logical level for subsequently assigning differentiated weight coefficients and achieving fine-grained, targeted self-evolutionary updates.

[0029] S2. Obtain the natural language description data input by the user, and perform semantic segmentation and intent recognition on the natural language description data to extract the core keyword set; Furthermore, step S2 involves extracting the core keyword set by processing the natural language description through a word segmentation component and using the TF-IDF algorithm to extract words with semantic features to form the core keyword set.

[0030] Specifically, this step addresses the large amount of irrelevant interjections and redundant information in user input commands by deploying a natural language processing engine at the system front end to purify information and reduce the dimensionality of features from the original unstructured long text. Existing technologies often directly compare the entire text when processing user searches, which not only consumes enormous computing power but is also highly susceptible to interference from redundant information, leading to false matches. This method, by introducing a word segmentation component combined with the term frequency-inverse document frequency (TF-IDF) algorithm, effectively filters out business-irrelevant noise, accurately identifies the user's core needs, and delivers a high signal-to-noise ratio core feature vector set to the downstream similarity calculation module.

[0031] Specifically, the implementation principle described above is as follows: First, by calling HanLP and other word segmentation components, fine-grained lexical segmentation is performed on the acquired natural language description data to obtain a basic vocabulary sequence. Then, the TF-IDF algorithm is used to traverse this vocabulary sequence, calculating the term frequency (TF) of each word in the current input and its inverse document frequency (IDF) in the global business corpus. Generalized stop words whose TF-IDF product is lower than a set threshold are removed, and the set of words with the highest scores constitutes the core keyword set. Furthermore, it should be noted that the core advantage of the TF-IDF algorithm lies in its ability to quantitatively evaluate the contribution of a word to the information entropy of the current sentence. Extracting core words with high semantic features through this algorithm is equivalent to performing a high-precision semantic distillation before matching calculations, greatly reducing the computational complexity of subsequent multi-dimensional matrix comparisons, while ensuring the accuracy of intent recognition from the source.

[0032] S3. Based on the core keyword set, calculate the first hit rate of the core keyword set and the process keyword list of each candidate process in the process database, and the second hit rate of the core keyword set and the process fuzzy synonym list of each candidate process in the process database. Furthermore, step S3 includes the following sub-steps: S301. Based on the core keyword set, use the edit distance algorithm or string similarity algorithm to calculate the similarity between the input words in the core keyword set and the preset words in the dynamic vocabulary; S302. The first hit rate and the second hit rate are calculated using the hit rate calculation formula; the hit rate calculation formula is expressed as follows: ; in, This represents the first hit rate or the second hit rate calculated for a single candidate process. Represents the first in the set of core keywords One input word, This represents the total number of preset words in the list of process keywords or the list of process fuzzy synonyms to be matched. Represents the first in the corresponding list One keyword, Represents the similarity function. This represents the total number of word segments in the core keyword set.

[0033] Specifically, this step addresses the fuzzy boundaries of natural language semantics by performing omnidirectional cross-comparison matrix operations at the computational level. Similarity scores are calculated independently for each of the two thesaurus lists in the candidate process. Existing single similarity calculation methods struggle to find a balance between rigorous business terminology and generalized colloquialisms. This method, based on edit distance or string similarity algorithms and combined with a double-layer nested summation model, achieves multi-dimensional cross-boundary comparison from literal spelling to semantic space, constructing a two-dimensional index system reflecting both precise matching and fuzzy tolerance, providing a refined quantitative basis for the system.

[0034] Specifically, the implementation principle described above is as follows: First, the algorithm module iterates through each input word in the core keyword set. Call the similarity function Calculate the relationship between each keyword in the process keyword list or fuzzy synonym list. The similarity score of all cross-pairs is summed and then divided by the total number of input words. Mean normalization is performed, and independent first and second hit rates are output separately. Furthermore, it should be noted that the hit rate calculation formula is essentially a mean mapping of the full permutation similarity energy pool. This ensures that even if the user inputs a complex long sentence composed of multiple scattered words, the system can still assess the semantic convergence between the current core keyword set and the overall target flow by traversing all preset words, thus mathematically preventing overall misjudgment caused by a single word mismatch.

[0035] S4. Based on a preset first weight coefficient for the list of keywords for the process and a preset second weight coefficient for the list of fuzzy synonyms for the process, the first hit rate and the second hit rate are weighted and fused to obtain the comprehensive matching score of each candidate process. Furthermore, step S4 involves weighted fusion of the first hit rate and the second hit rate to obtain the comprehensive matching score for each candidate process, the calculation formula of which is expressed as follows: ; ; in, This represents the overall matching score. Indicates the first hit rate. Indicates the second hit rate. This represents the first weighting coefficient. This represents the second weighting coefficient. The activation function is used to map and normalize the weighted calculation result to the (0,1) interval. The activation function is either the Sigmoid function or the ReLU function.

[0036] Specifically, this step addresses the differences in the contribution of different dimensional thesauri to the final routing decision by constructing a scoring mechanism that combines linear weighting and nonlinear activation. Traditional equal-weight scoring mechanisms are prone to failure in highly specialized or divergent business scenarios. This method introduces a first and a second weight coefficient and relies on Sigmoid or ReLU activation functions for boundary convergence, constructing a comprehensive evaluation mechanism that is both highly flexible and robust, ensuring global comparability of the final scores of candidate processes.

[0037] Specifically, the implementation principle described above is as follows: First, the system obtains the first weight coefficient set in the current state. With the second weighting coefficient Compare each with the calculated first hit rate Second hit rate Perform scalar multiplication and summation to complete linear dimension feature fusion; then, feed this linear weighted sum as an input parameter to the activation function. In this process, excessively large or small values ​​are smoothed and suppressed through mapping, and the output is normalized to a comprehensive matching score within the (0,1) interval. Furthermore, it should be noted that this is based on conditional constraints. This ensures energy conservation during the multi-dimensional feature fusion process, ensuring that scores under different configurations always fluctuate within the same dimension. The introduction of the activation function, on the other hand, endows the entire linear scoring system with non-linear expressive capabilities, effectively avoiding the gradient explosion phenomenon caused by the occasional extremely high similarity of a synonym, thus guaranteeing the stability of the scoring system.

[0038] S5. Based on the comprehensive matching score of each candidate process and the preset sorting threshold, the candidate processes are filtered and sorted, and the recommendation results are returned to the user. The user click-through rate, the accuracy of matching user input words with the recommendation results are obtained as user feedback data. Specifically, existing retrieval systems often only handle one-way information output, lacking in-depth tracking of subsequent user behavior. This method introduces a ranking threshold and constructs a full-scale information collection channel from the front end of the business process to the back end of the data, obtaining real interactive feedback containing implicit user intent, thus providing more data sources for subsequent self-learning mechanisms.

[0039] Specifically, the implementation principle described above is as follows: First, by comparing the comprehensive matching score of each candidate process with the current preset ranking threshold, low-relevance redundant processes with scores below the threshold are intercepted and eliminated. The remaining candidate processes are then sorted in descending order of score and pushed to the user interface. Simultaneously, the click-through rate and matching accuracy of the user in the current query session are collected in real time and mapped to their original input word data, generating a structured feedback data package. Furthermore, it should be noted that the mechanism of obtaining user click-through rate and matching accuracy is essentially equivalent to manually annotating and verifying the recommendation matching logic of the current process through user feedback. This constitutes the core prerequisite for achieving supervised learning in an unsupervised environment within the entire closed-loop method.

[0040] S6. Dynamically update the list content of each candidate process in the dynamic thesaurus based on the feedback data, and adaptively adjust the first weight coefficient, the second weight coefficient, and the sorting threshold based on the feedback data and the working condition requirement data of the target business process.

[0041] Furthermore, step S6, which involves dynamically updating the list of each candidate process in the dynamic lexicon based on the feedback data, includes: The feedback data is analyzed. If the feedback data indicates that the user finally confirmed that the target process in the recommendation results was selected, it is determined to be a positive sample and positive sample is updated. The input words in the feedback data are added to the process keyword list or the process fuzzy synonym list of the target process, and the association weight parameter between the input words and the target process is increased. If the feedback data indicates that the user did not select the recommended process ranked first in the recommendation results, but instead selected other subsequent actual processes, it is determined as a negative sample and a negative sample update is performed. The association weight parameter between the input word in the feedback data and the recommended process ranked first in the recommendation results is reduced, and the input word is directly added to the fuzzy synonym list of the actual process.

[0042] Furthermore, based on the accumulated feedback data within a fixed time window, a pre-trained semantic representation model is used to perform cluster analysis on the accumulated feedback data to obtain words in the accumulated feedback data that have potential semantic association features with existing words in the dynamic lexicon, and these words are marked as mismatched words. The statistical number of times negative sample updates are triggered within the corresponding fixed time window is monitored. If the statistical number reaches a preset anti-oscillation threshold, the mismatched words within the current fixed time window are determined to be valid words, and the list of fuzzy synonyms for the process is updated, while the corresponding weight parameters are reset. The pre-trained semantic representation model is a Word2Vec model or a BERT model.

[0043] Specifically, this step addresses the technical bottlenecks of traditional systems that cannot self-evolve and heavily rely on manual maintenance. It constructs a self-learning model that integrates micro-level reinforcement learning, macro-level AI deep mining, and adaptive adjustment based on operational conditions. Existing matching rules, once fixed, cannot flexibly adapt to changes in enterprise business terminology and complex, ever-changing concurrency pressures. This method, through automated recombination of positive and negative samples and latent semantic mining in high-dimensional vector spaces, enables the system to become increasingly accurate with use. Simultaneously, by dynamically adjusting weights and thresholds based on operational feedback, it effectively improves the matching accuracy for business processes.

[0044] Specifically, the implementation principle described above is as follows: First, in terms of dynamic lexicon updates, the system breaks down the feedback data. If a match is successful, it is recorded as a positive sample, and the input words are assimilated into the target process lexicon and their association weight is increased. If a mismatch occurs, it is recorded as a negative sample, and the input words are merged into the fuzzy synonym lexicon of the user's actual selected process. Simultaneously, within a fixed time window, the accumulated feedback words are mapped to a high-dimensional dense vector space by calling the Word2Vec or BERT model, and cluster analysis is performed to obtain words in the accumulated feedback data that have potential semantic association features with existing words in the dynamic lexicon, and these are marked as mismatched words. Once the number of times a mismatched word triggers negative feedback transfer within the time window exceeds the anti-oscillation threshold, it is determined to be non-random noise, and the mismatched words within the current fixed time window are considered valid words, thereby enabling batch expansion of the lexicon and system-level correction.

[0045] Secondly, in terms of adaptive parameter adjustment, the system monitors changes in operational demand data in real time. By increasing the sorting threshold, it accelerates algorithm convergence and reduces server load. Simultaneously, it analyzes the colloquialism of input words; if colloquialisms or slang are detected, the second weight coefficient is increased while the first weight coefficient is decreased, and vice versa. Furthermore, it should be noted that the combination of deep clustering in the AI ​​large-scale model and the anti-oscillation threshold ensures high accuracy for new business vocabulary while effectively preventing dictionary crashes caused by malicious clicks. The adaptive bidirectional parameter adjustment mechanism achieves dynamic coupling between computing resources and business conditions, as well as self-learning of the routing system.

[0046] Example 2

[0047] like Figure 2 As shown, as a preferred embodiment of the above embodiments, a self-evolving process entry routing closed-loop system based on user feedback is provided. This system is implemented based on any of the self-evolving process entry routing closed-loop methods based on user feedback described above, and includes: The thesaurus construction module is used to build a process database and dynamic thesaurus containing information on multiple candidate processes, and to assign a corresponding list of process keywords and a list of process fuzzy synonyms to each candidate process. The intent recognition module is used to receive natural language descriptions input by users, perform semantic word segmentation and intent recognition algorithms, and output a set of core keywords after filtering out stop words; The hit rate calculation module is used to receive the core keyword set, and call the similarity algorithm to calculate the first hit rate between the core keyword set and the process keyword list, and the second hit rate between the core keyword set and the process fuzzy synonym list. The weighted fusion module is used to obtain the set first weight coefficient and second weight coefficient, and to combine and weight the first hit rate and the second hit rate to calculate the comprehensive matching score of the corresponding candidate process. The ranking and recommendation module is used to compare the comprehensive matching scores of all candidate processes with a preset ranking threshold, perform filtering and descending sorting, and output the final recommendation results to the user. The adaptive evolution module is used to capture user feedback behavior data, trigger the closed-loop update logic of the list content of the lexicon construction module, and, in combination with system operation indicators, send an instruction to the weighted fusion module to adjust the ratio of the first weight coefficient and the second weight coefficient, and send an instruction to the ranking recommendation module to modify the ranking threshold.

[0048] Specifically, the implementation principle of the above embodiments is as follows: First, the lexicon building module initializes the underlying database based on the operational requirements data of the target business process, establishing a dynamic lexicon base that includes a bidirectional decoupled list of process keywords and a list of fuzzy synonyms for the process. When the system front end receives the user's unstructured query command, the intent recognition module immediately connects and uses the built-in semantic word segmentation component and TF-IDF algorithm to reduce the dimensionality and remove noise from the natural language description data, accurately extracting the core keyword set. Then, after receiving the core keyword set, the hit rate calculation module starts a dual-track parallel comparison mechanism at the operation level. By calling the edit distance or string similarity algorithm, it calculates the cross space matrix similarity between the input word array and the process keyword list and process fuzzy synonym list of each candidate process in the underlying dynamic word library, and independently parses and outputs the first hit rate reflecting hard and precise constraints and the second hit rate reflecting flexible semantic fault tolerance. Next, the weighted fusion module, based on the first and second weight coefficients obtained from the current business scenario, feeds the aforementioned two-dimensional hit rate into the mathematical model of the embedded activation function for nonlinear feature stitching and normalization smoothing processing, generating a comprehensive matching score for each candidate process that is globally comparable. Immediately afterward, the ranking and recommendation module takes over the data stream, performs noise filtering and descending reordering on all candidate processes based on the preset ranking threshold, and pushes the high-fidelity recommendation results to the front-end user interface. At the same time, the hidden data probe embedded in this module is activated synchronously, monitoring the user's final click point at all times, and capturing real interactive feedback data including user click rate, original input words, and matching accuracy. Finally, after capturing the feedback data from the above-mentioned return flow, the adaptive evolution module drives the core autonomous closed-loop learning logic to immediately perform positive sample association enhancement and negative sample cross-library word transfer operations for single-point deviations. At the same time, the module calls the embedded BERT or Word2Vec pre-trained semantic representation engine to perform high-dimensional vector mapping and cluster mining on the discrete feedback data within a fixed time window. Once the discovered potential synonymous business terms reach the built-in anti-oscillation number threshold safety red line, they are immediately judged as valid mismatch patterns, and batch correction instructions are issued to the underlying lexicon construction module to complete the intelligent topology update of the underlying lexicon structure without manual intervention. Furthermore, throughout the entire system operation, the adaptive evolution module continuously monitors real-time system operation indicators that characterize the target business process requirements. It can also sense fluctuations in instantaneous concurrent access pressure and characteristic shifts in the colloquial input habits of user groups. Based on this, it dynamically generates adaptive adjustment signals, sends fine-tuning instructions for the first and second weight coefficients to the weighted fusion module, and sends instructions to increase or decrease the ranking threshold to the ranking recommendation module. This completely breaks down the rigid barriers of traditional routing rules in the overall architecture, ensuring that the intelligent system can respond to business conditions and complex speech evolution with millisecond-level response and dynamic global optimal collaboration.

[0049] Example 3

[0050] like Figure 3 As shown, this embodiment proposes a terminal device for a self-evolving process entry routing closed-loop method based on user feedback. The terminal device includes at least one memory, at least one processor, and a bus connecting different platform systems.

[0051] The memory may include readable media in the form of volatile memory, such as RAM 211 and / or cache memory, and may further include ROM 213.

[0052] The memory also stores a computer program that can be executed by a processor, causing the processor to execute any of the user feedback-based self-evolving process entry routing closed-loop methods described in the embodiments of this application. The specific implementation and technical effects are consistent with those described in the embodiments of the above methods, and some details will not be repeated here. The memory may also include a program / utility having a set (at least one) of program modules, including but not limited to: an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment.

[0053] Accordingly, the processor can execute the aforementioned computer programs, as well as executable programs / utilities.

[0054] A bus can represent one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus that uses any of the various bus structures.

[0055] The terminal device can also communicate with one or more external devices, such as keyboards, pointing devices, Bluetooth devices, etc., and with one or more devices capable of interacting with the terminal device, and / or with any device that enables the terminal device to communicate with one or more other computing devices (e.g., routers, modems, etc.). This communication can be performed through I / O interfaces. Furthermore, the terminal device can communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via a network adapter. The network adapter can communicate with other modules of the terminal device via a bus. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with the terminal device, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, tape drives, and data backup storage platforms.

[0056] Example 4

[0057] like Figure 4 As shown, based on Embodiment 1, this embodiment proposes a computer-readable storage medium for a user-feedback-based self-evolving process entry route closed-loop method. The computer-readable storage medium stores instructions that, when executed by a processor, implement any of the aforementioned user-feedback-based self-evolving process entry route closed-loop methods. Its specific implementation method is consistent with the implementation methods and achieved technical effects described in the embodiments above, and some details will not be repeated.

[0058] This embodiment provides a program product for implementing the above-described method, which may employ a portable compact disc read-only memory (CD-ROM) and include program code, and may run on a terminal device, such as a personal computer. However, the program product of the present invention is not limited thereto. In this embodiment, the readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, apparatus, or device. The program product may employ any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0059] Computer-readable storage media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting a program for use by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof. Program code for performing operations of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on a user computing device, partially on a user device, as a standalone software package, partially on a user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing devices can be connected to user computing devices via any type of network, including local area networks (LANs) or wide area networks (WANs), or they can be connected to external computing devices (e.g., via the Internet through an Internet service provider).

[0060] This invention is described from the perspectives of its intended use, effectiveness, progress, and novelty. Its practical and progressive features meet the functional enhancement and use requirements emphasized by the Patent Law. The above description and drawings are merely preferred embodiments of this application and are not intended to limit this application. Therefore, all structures, devices, features, etc., that are similar to or identical to those of this application, i.e., all equivalent substitutions or modifications made in accordance with the scope of this patent application, shall fall within the scope of protection of this patent application.

[0061] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A self-evolving process entry routing closed-loop method based on user feedback, characterized in that, Includes the following steps: S1. Based on the working condition requirements data of the target business process, construct a process database and a dynamic thesaurus. The dynamic thesaurus includes a list of process keywords and a list of process fuzzy synonyms corresponding to each candidate process in the process database. S2. Obtain the natural language description data input by the user, and perform semantic segmentation and intent recognition on the natural language description data to extract the core keyword set; S3. Based on the core keyword set, calculate the first hit rate of the core keyword set and the process keyword list of each candidate process in the process database, and the second hit rate of the core keyword set and the process fuzzy synonym list of each candidate process in the process database. S4. Based on a preset first weight coefficient for the list of keywords for the process and a preset second weight coefficient for the list of fuzzy synonyms for the process, the first hit rate and the second hit rate are weighted and fused to obtain the comprehensive matching score of each candidate process. S5. Based on the comprehensive matching score of each candidate process and the preset sorting threshold, the candidate processes are filtered and sorted, and the recommendation results are returned to the user. The user click-through rate, the accuracy of matching user input words with the recommendation results are obtained as user feedback data. S6. Dynamically update the list content of each candidate process in the dynamic thesaurus based on the feedback data, and adaptively adjust the first weight coefficient, the second weight coefficient, and the sorting threshold based on the feedback data and the working condition requirement data of the target business process; The step S6, which dynamically updates the list of each candidate process in the dynamic lexicon based on the feedback data, includes: The feedback data is analyzed. If the feedback data indicates that the user finally confirmed that the target process in the recommendation results was selected, it is determined to be a positive sample and positive sample is updated. The input words in the feedback data are added to the process keyword list or the process fuzzy synonym list of the target process, and the association weight parameter between the input words and the target process is increased. If the feedback data indicates that the user did not select the recommended process ranked first in the recommendation results, but instead selected other subsequent actual processes, it is determined as a negative sample and a negative sample update is performed. The association weight parameter between the input word in the feedback data and the recommended process ranked first in the recommendation results is reduced, and the input word is directly added to the fuzzy synonym list of the actual process. Based on the accumulated feedback data within a fixed time window, a pre-trained semantic representation model is used to perform cluster analysis on the accumulated feedback data to obtain words in the accumulated feedback data that have potential semantic association features with existing words in the dynamic lexicon, and these words are marked as mismatched words. The statistical number of times negative sample updates are triggered within the corresponding fixed time window is monitored. If the statistical number reaches a preset anti-oscillation threshold, the mismatched words in the current fixed time window are determined to be valid words, and the list of fuzzy synonyms for the process is updated, while the corresponding weight parameters are reset. The pre-trained semantic representation model is a Word2Vec model or a BERT model.

2. The self-evolving process entry routing closed-loop method based on user feedback according to claim 1, characterized in that, Step S2 involves extracting the core keyword set by processing the natural language description through a word segmentation component and using the TF-IDF algorithm to extract words with semantic features to form the core keyword set.

3. The self-evolving process entry routing closed-loop method based on user feedback according to claim 1, characterized in that, Step S3 includes the following sub-steps: S301. Based on the core keyword set, use the edit distance algorithm or string similarity algorithm to calculate the similarity between the input words in the core keyword set and the preset words in the dynamic vocabulary; S302. The first hit rate and the second hit rate are calculated using the hit rate calculation formula; the hit rate calculation formula is expressed as follows: ； in, This represents the first hit rate or the second hit rate calculated for a single candidate process. Represents the first in the set of core keywords One input word, This represents the total number of preset words in the list of process keywords or the list of process fuzzy synonyms to be matched. Represents the number in the corresponding list One keyword, Represents the similarity function. This represents the total number of word segments in the core keyword set.

4. The self-evolving process entry routing closed-loop method based on user feedback according to claim 1, characterized in that, Step S4 involves weighted fusion of the first hit rate and the second hit rate to obtain the comprehensive matching score for each candidate process. The calculation formula is as follows: ； ； in, This represents the overall matching score. Indicates the first hit rate. Indicates the second hit rate. This represents the first weighting coefficient. This represents the second weighting coefficient. The activation function is used to map and normalize the weighted calculation result to the (0,1) interval. The activation function is either the Sigmoid function or the ReLU function.

5. The self-evolving process entry routing closed-loop method based on user feedback according to claim 1, characterized in that, Step S6, which involves adaptively adjusting the first weight coefficient, the second weight coefficient, and the ranking threshold, includes: updating the target business process's operational requirements data by real-time monitoring of user click-through rate and recommendation result matching accuracy; dynamically adjusting the ranking threshold based on the updated operational requirements data; and determining the degree of colloquial input based on user input words. If the degree of colloquial input is determined to be too high, the second weight coefficient is increased and the first weight coefficient is correspondingly decreased; otherwise, the first weight coefficient is increased and the second weight coefficient is correspondingly decreased.

6. A self-evolving process entry routing closed-loop system based on user feedback, the system being implemented based on the self-evolving process entry routing closed-loop method based on user feedback as described in any one of claims 1-5, characterized in that, include: The thesaurus building module is used to build a process database and dynamic thesaurus containing information on multiple candidate processes, and to assign a corresponding list of process keywords and a list of process fuzzy synonyms to each candidate process. The intent recognition module is used to receive natural language descriptions input by users, perform semantic word segmentation and intent recognition algorithms, and output a set of core keywords after filtering out stop words; The hit rate calculation module is used to receive the core keyword set, and call the similarity algorithm to calculate the first hit rate between the core keyword set and the process keyword list, and the second hit rate between the core keyword set and the process fuzzy synonym list. The weighted fusion module is used to obtain the set first weight coefficient and second weight coefficient, and to combine and weight the first hit rate and the second hit rate to calculate the comprehensive matching score of the corresponding candidate process. The ranking and recommendation module is used to compare the comprehensive matching scores of all candidate processes with a preset ranking threshold, perform filtering and descending sorting, and output the final recommendation results to the user. The adaptive evolution module is used to capture user feedback behavior data, trigger the closed-loop update logic of the list content of the lexicon construction module, and, in combination with system operation indicators, send an instruction to the weighted fusion module to adjust the ratio of the first weight coefficient and the second weight coefficient, and send an instruction to the ranking recommendation module to modify the ranking threshold.

Images