A method for constructing a government service convenience hotline knowledge base based on large model technology

Abstract

This application provides a method for constructing a knowledge base for a government service hotline based on large model technology, comprising: for the semantic vector representation, using an LSTM network to perform temporal modeling on a multi-turn dialogue sequence, and determining contextual dependencies by fusing vector associations between previous and subsequent turns; selecting high-priority fragment expression regions and merging them with the overall semantic vector through the integrated priority list, and using an attention mechanism to weightedly fuse the region vectors to determine the completeness of intent recognition; if the completeness of intent recognition is insufficient, generating a supplementary query template, and matching relevant templates from a preset follow-up question guidance library to obtain follow-up questions for missing information; for the updated structured demand representation, mapping it to a government service classification model, and determining the final service assignment path by comparing the similarity between the demand representation and the service conditions.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method for constructing a knowledge base for government service hotlines based on large-scale modeling technology. Background Technology

[0002] In the field of government services, hotlines serve as an important bridge connecting citizens and the government, bearing the key responsibility of listening to public opinion and resolving demands. Their level of intelligence is directly related to the efficiency of public services and public satisfaction.

[0003] As citizens' demands become increasingly complex, the hotline system needs to more accurately understand and respond to diverse expressions, which has become an important direction that cannot be ignored in improving service quality.

[0004] However, current methods for receiving and processing hotline calls often struggle to cope with citizens' colloquial and fragmented expressions. In particular, during multi-round conversations, the system's ability to understand contextual relationships is insufficient, leading to frequent errors in identifying the intent of the call and affecting the accurate allocation of subsequent services.

[0005] Looking further, existing methods generally lack adaptability to dynamic interactive contexts when dealing with citizens' demands.

[0006] Many systems can only make simple judgments based on the content of a single conversation, lacking the ability to integrate the overall semantics of the preceding and following conversations, and are also unable to effectively deal with situations where citizens express themselves unclearly or lack information.

[0007] This limitation makes it difficult for the system to accurately grasp the real needs of citizens when faced with complex requests, and even more difficult to transform fragmented spoken content into clear service conditions.

[0008] Focusing on the technical challenges, the core challenge lies in how to construct a complete semantic association network from multi-turn dialogues.

[0009] Citizens often express their demands in multiple, fragmented ways during hotline communications. The system needs to connect these fragmented messages to form a comprehensive intent map.

[0010] If the dynamic integration of semantic fragments cannot be achieved, key information will be missed, such as the type of service or specific matter mentioned by the citizen being ignored.

[0011] More complicated by this lack of integration, the system also fails to proactively guide citizens to supplement necessary information. For example, when inquiring about a service, if a citizen does not mention the required documents, the system may fail to ask in a timely manner, ultimately causing the service process to be interrupted.

[0012] Therefore, how to dynamically integrate citizens' semantic fragments in multi-round dialogues and generate effective follow-up questions to guide responses to unclear or incomplete information, transforming fragmented oral requests into structured service needs, has become a key issue in improving the intelligence level of government hotlines.

[0013] The above analysis clearly shows the progressive relationship from the importance of hotline services to the limitations of existing methods, and then to specific technical challenges. The root of the problem lies in the lack of semantic integration and interactive guidance, which is precisely the business pain point that urgently needs to be addressed. Summary of the Invention

[0014] This invention provides a method for constructing a knowledge base for government service hotlines based on large model technology, mainly including:

[0015] Obtain the multi-turn dialogue text sequence between citizens and the hotline system, extract the spoken expression content of each turn, and process the spoken expression content through a pre-trained BERT model to obtain a semantic vector representation; For the semantic vector representation, an LSTM network is used to perform temporal modeling on the multi-turn dialogue sequence, and the contextual dependencies are determined by fusing the vector associations between previous and subsequent turns. Based on the contextual dependencies, the semantic fragment connections in the current dialogue sequence are analyzed. If the fragment connection strength is lower than a preset threshold, it is marked as a fragment expression region. An integration priority list is obtained by calculating the similarity between fragments. By using the integrated priority list, high-priority fragment expression regions are selected and merged with the overall semantic vector. An attention mechanism is used to weighted fuse the region vectors to determine the completeness of intent recognition. If the completeness of the intent recognition is insufficient, a supplementary query template is generated, and relevant templates are matched from the preset follow-up question guidance library to obtain follow-up questions for missing information; Based on the new dialogue sequence obtained after the follow-up question, the aforementioned semantic vector extraction and context dependency determination process is repeated, and the updated structured requirement representation is obtained by iteratively fusing the new sequence fragments; The updated structured demand representation is mapped into the government service classification model, and the final service assignment path is determined by comparing the similarity between the demand representation and the service conditions.

[0016] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0017] This invention discloses a method for constructing a knowledge base for government service hotlines based on large-scale modeling technology. Addressing the unique business scenario problems of fragmented dialogue, incomplete intent recognition, and inaccurate demand matching in citizen-hotline interactions, this invention utilizes deep learning and semantic analysis techniques to achieve the transformation from spoken expression to structured demands and the precise mapping of service paths. First, this invention extracts dialogue semantic vectors through a pre-trained model, combines temporal modeling to capture contextual dependencies, identifies and integrates fragmented expression regions, and then improves the completeness of intent recognition through an attention mechanism. For cases with insufficient recognition, follow-up questions are generated to iteratively update the demand representation, and finally, service assignment is completed through similarity comparison. This invention effectively solves the problems of information loss and semantic fragmentation in multi-turn dialogues, significantly improving the response efficiency and matching accuracy of government services, and providing citizens with a more intelligent and convenient service experience. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a method for constructing a knowledge base for a government service hotline based on large-scale modeling technology, according to the present invention. Figure 2 This is a schematic diagram illustrating a method for constructing a knowledge base for a government service hotline based on large-scale modeling technology according to the present invention. Figure 3 This is another schematic diagram of a method for constructing a knowledge base for a government service hotline based on large model technology according to the present invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0020] like Figure 1-3 This embodiment of a method for constructing a knowledge base for a government service hotline based on large model technology may specifically include:

[0021] Step S101: Obtain the multi-turn dialogue text sequence between citizens and the hotline system, extract the spoken expression content of each turn, and process the spoken expression content through a pre-trained BERT model to obtain a semantic vector representation.

[0022] Multi-turn text data of citizen conversations is obtained from the hotline system. This multi-turn text is segmented to determine independent dialogue units for each turn. For each dialogue unit, spoken expressions are extracted, and irrelevant characters and noise are removed using pre-defined text cleaning rules to obtain clean spoken text data. This clean spoken text data is then input into a pre-trained BERT model, which processes it to generate corresponding semantic vector representations, thus determining the semantic features of each dialogue turn.

[0023] Furthermore, based on the generated semantic vector representations, semantic association analysis is performed on the multi-turn dialogue sequences. If the semantic vector of a certain turn of dialogue differs from other turns by more than a preset threshold, it is marked as an abnormal dialogue unit. For the marked abnormal dialogue units, contextual text data is obtained from the system interaction records. By comparing the contextual semantic vectors, it is determined whether the spoken expression content needs to be re-extracted. The overall content of the citizen dialogues is classified using semantic vector representations to determine the topic distribution of different turns of dialogue and generate classification result data. Based on the classification result data, the dialogue sequences of the hotline system are semantically structured and stored, obtaining the structured dialogue data and completing the semantic organization of the multi-turn text.

[0024] In one possible implementation, multi-turn text data of citizen dialogues is obtained from a hotline system. This multi-turn text is then segmented, either by interaction role or timestamp-assisted segmentation, to determine independent dialogue units for each turn. Specifically, segmentation by interaction role: From the original text sequence obtained from the hotline system, role labels (such as "citizen," "customer service") are used for alternating identification. Each time the speaker switches, it is marked as a new independent dialogue unit. Segmentation assisted by timestamp: Using timestamps in the interaction records, if the time interval between two inputs exceeds a preset threshold, they are considered different dialogue segments. After determining the independent dialogue units, the following operations are performed to obtain clean spoken text data: 1. Redundant character removal: Modal particles (such as "ah," "ne," "that"), repeated words, and irrelevant characters automatically generated by the system are deleted from spoken expressions. 2. Core spoken language extraction: Keywords and phrases reflecting user intent are retained. The cleaned text is then input into a pre-trained BERT model to generate semantic vectors, and the semantic differences between the current turn and previous / next turns are calculated.

[0025] The model employs a pre-trained BERT model as its foundation, which means it utilizes a deep bidirectional transformer representation technique trained on large-scale general text. The model receives cleaned, spoken text data as input and processes the spoken expressions, transforming them into high-dimensional semantic vector representations. This process converts discrete text into a machine-understandable numerical feature space, serving as the basis for subsequent temporal modeling.

[0026] In one possible implementation, based on the generated semantic vector representation, semantic association analysis is performed on the multi-turn dialogue sequence. If the semantic vector of a certain turn of dialogue differs from other turns by more than a preset threshold, it is marked as an abnormal dialogue unit. The system performs semantic association analysis on the multi-turn dialogue sequence, that is, it calculates the distance or similarity between the current turn's vector and the vectors of other turns in the sequence (especially context turns) using a cosine similarity algorithm. This is done by comparing the semantic vector of the current turn with the vectors of other turns. If the content of a certain turn of dialogue deviates from the overall topic of the dialogue in the semantic space (e.g., a citizen suddenly switches to a completely unrelated topic, or a large amount of noise in the recognition result causes vector drift), it will be identified as having too large a difference. The preset threshold is a pre-set value used to measure the degree of difference between semantic vectors. The purpose of setting this threshold is to filter out invalid information caused by excessive colloquialism, recognition errors, or semantic breaks. Once marked as abnormal, the system retrieves contextual text data from the interaction record to determine whether the spoken content of that turn needs to be re-extracted.

[0027] In one possible implementation, for a marked abnormal dialogue unit, the system first retrieves the contextual text data before and after the abnormal unit from the interaction log. This data serves as a semantic baseline to verify whether the abnormal unit's logical break is due to the citizen's expression style (e.g., excessive colloquialism, speech recognition errors) or a genuine topic shift. Then, the system quantitatively compares the semantic vector of the abnormal unit with the semantic vectors of each round in the context. By calculating the correlation between the abnormal vector and the context vector sequence, it determines whether the colloquial expression content needs to be re-extracted, such as:

[0028] Situations requiring re-extraction: If the comparison reveals that the semantic vector of the unit has extremely low relevance to the context, and it is determined that the difference is caused by recognition noise or non-substantial character interference (i.e., not an expression of the citizen's true intention), the system will trigger a re-extraction process. By adjusting the cleaning rules or reprocessing the original speech / text, a cleaner spoken content that better matches the context can be obtained.

[0029] In cases where re-extraction is not required: If the comparison confirms that the "anomaly" is due to a service jump or topic switch initiated by a citizen, and its semantics are reasonable in the local context, then the content is retained, and subsequent topic distribution generation and structured storage continue.

[0030] Step S102: For the semantic vector representation, an LSTM network is used to perform temporal modeling on the multi-turn dialogue sequence, and the contextual dependencies are determined by fusing the vector associations of previous and subsequent turns.

[0031] For multi-turn dialogue data, a Long Short-Term Memory (LSTM) network is used to perform temporal modeling of the dialogue sequence, processing information from consecutive turns to obtain a temporal feature representation of the dialogue content. Based on this temporal feature representation, the vector information associations between consecutive turns are analyzed, and the similarity of semantic vectors between turns is calculated to determine the association strength between each turn. According to the association strength, the vector information from relevant turns is fused to construct a holistic representation of contextual dependencies, obtaining the dependencies between different parts of the dialogue sequence. For the constructed dependencies, if the dependency strength of a certain turn is lower than a preset threshold, it is fused with information from other turns to supplement it, resulting in a complete contextual dependency representation.

[0032] Furthermore, by representing the complete contextual dependencies, key information in the dialogue sequence is extracted to determine the role and contribution of each turn's content to the overall dialogue. Based on the extracted key information and role assessments, a structured representation of the dialogue sequence is generated, determining the logical flow and content coherence of the multi-turn dialogue. This structured representation is then used for subsequent processing of the dialogue sequence to obtain the final semantic understanding result, achieving comprehensive capture of contextual dependencies.

[0033] In one possible implementation, for the multi-turn dialogue data, a Long Short-Term Memory (LSTM) network is used to perform temporal modeling of the dialogue sequence, processing the information of successive turns in the sequence to obtain a temporal feature representation of the dialogue content. Specifically, the extracted multi-turn dialogue semantic vectors are arranged in chronological order to form the dialogue sequence input. The LSTM network's long short-term memory characteristic is utilized to process the information of each turn in the sequence. The network uses internal gating mechanisms (input gate, forget gate, output gate) to determine which historical semantics need to be retained and which current information needs to be extracted. Through recursive calculation of successive turns in the sequence, the original independent vectors are transformed into a temporal feature representation containing temporal flow information.

[0034] In one possible implementation, the fusion of vector information from relevant rounds based on association strength is achieved through a dynamic iterative fusion mechanism: 1. Association strength-based fusion: The system first calculates the similarity between semantic vectors from preceding and following rounds to determine the association strength. 2. Low-strength supplementary fusion: If the dependency strength of a certain round is lower than a preset threshold, the system will fuse it with information from other rounds to supplement it, thereby obtaining a complete contextual dependency representation. 3. Iterative fusion process: When processing new dialogue sequences in subsequent steps, the system adopts a cyclic update mechanism to handle the combination of newly added fragments with existing structured representations.

[0035] The construction of a context-dependent overall representation involves obtaining the dependencies between different parts of the dialogue sequence. Specifically, LSTM is used to process the sequential information of each turn in the sequence to obtain the temporal feature representation of the dialogue content. The similarity between vectors is analyzed using these temporal features to determine the strength of the association. Vector information from relevant turns is then fused to construct an overall model with context-dependent features. Through this complete context-dependent representation, key information in the dialogue is extracted, and the role and contribution of each turn's content (i.e., each part) in the overall dialogue are determined, such as which are supplementary information and which are core requests. Based on the key information and role judgments, the logical structure of the multi-turn dialogue is determined, thereby clarifying the connections and dependencies between different parts.

[0036] In one possible implementation, key information in the dialogue sequence is extracted through a complete context-dependent representation. Specifically, this involves using temporal feature representations generated by an LSTM network to identify vector segments with high weights or high relevance from the dialogue sequence. The complete context-dependent representation is then transformed into a structured representation of the dialogue sequence, and key information is extracted by parsing internal fields and labels. An attention mechanism is used to calculate the importance weights of region vectors, and regions with higher weights are extracted as a set of key regions. Finally, the role and contribution of each round's content to the overall dialogue are determined, primarily based on the following three dimensions:

[0037] Association strength assessment: By calculating the similarity of semantic vectors between each round, the connection strength within the logical context is determined. The higher the association strength, the greater the contribution of that round to understanding the context.

[0038] Attention weight allocation: An attention mechanism is used to weight and fuse region vectors. The system calculates the contribution of each region vector to the overall semantics, and the contribution directly reflects the value of the information in that round to the final intent recognition.

[0039] Intent completeness verification: The completeness of intent recognition is determined based on the fusion result vector. If the information from a certain round can significantly improve the completeness (to reach the predetermined standard), then that round is determined to have a core contribution; otherwise, if the contribution is low and the completeness is insufficient, the system will re-filter or trigger a follow-up questioning process.

[0040] In one possible implementation, the step of generating a structured representation of the dialogue sequence based on extracted key information and function judgments, and determining the logical flow and content connection of multi-turn dialogues, involves a structured representation that is a composite data model containing semantic tags, fields, and logical relationships. For example, it may include the following structural elements:

[0041] Semantic Index / Label: Assign topic distribution labels to each round of dialogue (e.g., inquiry, complaint, supplementary explanation). Key Information Fields: Extracted core elements (e.g., name, service type, specific request). Temporal Feature Vector: Temporal feature representation with information from previous and subsequent rounds, obtained by LSTM processing. Dependency Mapping: Data recording the logical relationships between segments (e.g., segment A is a prerequisite for segment B).

[0042] Step S103: Based on the context dependency, analyze the semantic fragment connections in the current dialogue sequence. If the fragment connection strength is lower than a preset threshold, it is marked as a fragment expression region. The integration priority list is obtained by calculating the similarity between fragments.

[0043] Semantic fragments are extracted from the dialogue sequence, and the text content is split using a word segmentation tool to obtain a preliminary set of semantic fragments. For this preliminary set, the connection strength between fragments is calculated, and the semantic distance between fragments is obtained using a semantic vector model to determine the quantification of connection strength. If the connection strength is below a preset threshold, the corresponding fragment is marked as a fragment expression region, and the distribution of these marked regions is obtained to identify potential semantic breakpoints. Based on the marked fragment expression regions, the similarity value between fragments within each region is calculated, and the fragment content is compared using a cosine similarity algorithm to obtain a similarity ranking. An integration priority list is constructed based on the similarity ranking, and fragments with higher rankings are prioritized for semantic recombination to determine the recombined fragment set. The recombined fragment set is then compared with the original context of the dialogue sequence to determine the fit between the recombined fragments and the overall semantics, resulting in the final semantic integration result. Based on the final semantic integration result, the adjusted dialogue sequence structure is output, and the remaining fragments in the integration priority list are iteratively processed to determine the complete semantic connection framework.

[0044] In one possible implementation, for the set of semantic segments, the connection strength between any two semantic segments is calculated. Specifically, firstly, each semantic segment is converted into a high-dimensional semantic vector using a pre-trained semantic vector model (e.g., Sentence-BERT); then, the cosine similarity between each pair of semantic vectors is calculated as a measure of their semantic similarity; finally, the cosine similarity value is normalized to obtain a quantified connection strength result ranging from [0, 1]. A larger connection strength value indicates a stronger semantic association between the two semantic segments.

[0045] In one possible implementation, if the connection strength is lower than a preset threshold, the corresponding segment is marked as a fragment expression region, the distribution of the marked regions is obtained, and potential semantic breakpoints are determined. Specifically, the connection strength is calculated based on the similarity between segments. The system uses a cosine similarity algorithm to quantify the distance between segment content. When the quantification result calculated by cosine similarity is lower than this preset threshold, the system determines that the current two segments are semantically incoherent and insufficient to constitute a continuous intention expression. The determination of potential semantic breakpoints specifically involves the following three steps: First, the system uses a word segmentation tool to split the dialogue sequence and obtain independent semantic segments. Subsequently, using the temporal feature representation extracted by S102, the connection strength (i.e., similarity distance) between adjacent or related segments is calculated. Second, the connection strength is compared with a preset threshold. Segments with a strength lower than the threshold are marked as fragment expression regions by the system. The system obtains the spatial distribution of these marked regions in the entire dialogue stream (i.e., where they appear in the dialogue and how long they last). Third, based on the distribution of the marked regions, the system locks the breakpoints in the following way:

[0046] Jump detection: If, in a continuous dialogue stream, the labeling state of two consecutive regions changes significantly (e.g., from a strong connection to a fragmented label), the switching point is identified as a potential semantic breakpoint.

[0047] Dependency missing determination: Based on the contextual dependencies in S102, if there is a gap in the structured path that cannot be connected, the gap is the breakpoint.

[0048] In one possible implementation, the process involves obtaining a set of reconstructed fragments, comparing them with the original context of the dialogue sequence, determining the degree of fit between the reconstructed fragments and the overall semantics, and obtaining the final semantic integration result. Specifically, the system converts the set of reconstructed fragments into semantic vectors and maps them into a semantic space constructed from the original context. The consistency between the reconstructed vectors and the overall semantic vectors is quantified at the feature level by calculating the spatial distance between them (usually using cosine similarity). The reconstructed fragments are compared with the context flow of the original dialogue sequence to determine whether they fill the previously identified semantic breakpoints. If the reconstructed fragments smooth the logical connection of the context (i.e., the connection strength is increased and exceeds a preset threshold), the fit is considered high. An attention mechanism is used to calculate the contribution weight of the reconstructed fragments to the overall semantics. If the key information fields contained in the fragment (such as the requesting subject or core issue) effectively improve the certainty of intent recognition, it is judged as having a high fit.

[0049] Ultimately, segments deemed highly relevant (logically consistent) are integrated into the final semantic integration result, serving as the core input for determining the completeness of intent in segment S104. Segments deemed lowly relevant (conflicting or irrelevant) are identified as interference noise or secondary information, and their weight is reduced in the integration result. The resulting semantic integration is an enhanced structured semantic representation. It includes not only the original dialogue vector but also verified and reconstructed fragmented information. This result eliminates the semantic fragmentation problem caused by colloquial expressions, providing a high-quality, coherent data foundation for the next step of determining the completeness of intent recognition.

[0050] Step S104: Select high-priority fragment expression regions through the integrated priority list and merge them with the overall semantic vector. Use an attention mechanism to weighted fuse the region vectors to determine the completeness of intent recognition.

[0051] Data is initially filtered using a priority list to identify the distribution of high-priority fragment regions and determine the key region set. Based on this set, corresponding fragment region data is extracted and preliminarily matched with the overall semantic vector to obtain a preliminary combined vector representation. An attention mechanism is then used to process this preliminary combined vector representation, calculating the importance weights of the region vectors to determine the contribution of each region to the overall semantics. A fusion operation is performed on the weighted region vectors to generate a fusion result vector. If the completeness of the fusion result vector is below a preset threshold, the low-contribution region vectors are re-filtered to obtain supplementary data, resulting in an updated fusion vector. The updated fusion vector is then analyzed to assess the completeness of intent recognition and determine if it meets a predetermined standard. If the completeness still does not meet the standard, the semantic vector is locally adjusted, and combined with the supplementary data, the final intent recognition result is obtained.

[0052] In one possible implementation, based on the key region set, corresponding fragment region data is extracted and initially matched with the overall semantic vector to obtain a preliminary combined vector representation. Specifically, the system transforms the key region set determined in S103 (i.e., high-priority fragment fragments) into local semantic vectors and maps them to the same high-dimensional feature space as the overall semantic vector generated in S102. The correlation between these fragment vectors and the overall semantic vector is calculated (usually through inner product or cosine distance). This matching process is to determine the logical position that fragment information should occupy in the overall intent. During the matching process, the system will remove noise vectors that, although in the set, conflict with the overall semantics or have extremely low relevance, retaining only effective features that can supplement the main idea graph. Further, the selected fragment region vectors are initially mathematically fused with the original overall semantic vector (e.g., vector concatenation followed by a fully connected layer, or simple weighted addition). The output of the initial fusion is called the preliminary combined vector representation. It is an intermediate vector that retains the logic of the main dialogue flow (the overall vector modeled by LSTM) while absorbing the scattered information identified as high-value (fragment region data).

[0053] In one possible implementation, an attention mechanism is used to process the initially combined vector representations, calculate the importance weights of the region vectors, and determine the contribution of each region to the overall semantics. Specifically, the system compares each region vector (i.e., the fragment information vectors identified by S103) with the global semantic features. By calculating the correlation score between the two (usually using dot product or multilayer perceptron mapping), the degree of association of each region in the current context is obtained. The above scores are normalized using the Softmax function. The purpose of this is to map the weights of all regions to between 0 and 1, and the sum of all weights is 1. The final generated value is the importance weight. The larger the value, the greater the amount of information carried by the fragment region (such as a key service item added by a citizen) in the overall intent understanding. The determination of contribution is a process of evaluating the weights using business logic.

[0054] Contribution quantification: The system directly measures the ability of a region vector to shape the overall intent representation based on the calculated importance weights. Regions with high weights are considered to have a high contribution to the overall semantics.

[0055] Semantic value assessment:

[0056] Core contribution: If a region contains the key elements required for government services (such as business categories, specific administrative divisions, and certificate names), its contribution will be judged as high.

[0057] Auxiliary / noise contribution: If a region consists only of colloquial particles or repetitive descriptions, its weight is extremely low and its contribution is considered negligible.

[0058] Completeness Impact Analysis: The system uses the weighted fusion result vector to determine the completeness of intent recognition. If the intent features become clearer and more definite after fusing these high-contribution regions, it indicates that these fragments have successfully contributed to the missing semantics.

[0059] In one possible implementation, the fusion operation performed on the weighted region vectors to generate a fused result vector involves deeply stitching local fragment information with global semantic features. Specifically, the system uses importance weights previously calculated through an attention mechanism to weight each region vector. This means that vectors contributing significantly to intent (such as core appeal fragments) are amplified, while noise vectors are suppressed. These weighted region vectors are then superimposed or concatenated with the overall semantic vector (global vector) generated in S102. The fused vector is then input into a fully connected layer or a specific feature extraction module, which uses nonlinear transformations to map information from different dimensions into a unified feature space, thereby generating the fused result vector.

[0060] In one possible implementation, if the completeness of the fusion result vector is lower than a preset threshold (e.g., the threshold could be set at 0.85 or a specific probability value), then the low-contribution region vectors are re-filtered to obtain supplementary data. The completeness of the fusion result vector refers to the degree of clustering of the vector in the feature space after the system obtains the fusion result vector. If the vector clearly points to a specific government service category, the completeness score is high. The system checks whether the fused structured representation has filled all the key fields according to the necessary information list required for government services (such as subject, item, location, etc.). A quantified intent recognition completeness score is generated by calculating the predicted probability generated by the fusion vector.

[0061] In one possible implementation, by analyzing the distribution of the updated fusion vector in the semantic space, if the vector is very close to the center of a specific government service category and sufficiently far from other interfering categories, it is determined that the intent is clear and meets the standard. If the initial score is slightly lower than the standard, the system will perform a local adjustment, combining supplementary data (such as citizens' historical request records) to correct the semantic vector. After adjustment, the score is recalculated. If the score reaches the standard at this time, the final intent recognition result is directly output. If the score still does not reach the predetermined standard after adjustment, it is formally determined that the intent recognition completeness is insufficient, triggering the follow-up questioning mechanism in S105. Here, the local adjustment is a process of correcting the vector direction by introducing external compensation data. The specific steps are: the system retrieves historical interaction records related to the current citizen, the citizen's personal service profile, or related entries in the government service knowledge base. For example, if a citizen says "I want to apply for social security" but does not specify which district, the system will retrieve information about the citizen's place of residence or areas previously consulted. This supplementary data is converted into a supplementary vector, and the difference between it and the current fusion result vector is calculated. In this way, the system identifies the ambiguous or deviated parts in the semantic vector. For slots (fields) with uncertainties in the initial structure, supplementary data is used to enhance the weights of relevant dimensions. In this way, features that originally had low confidence (such as the location of the business) become clearer due to the addition of background knowledge.

[0062] Step S105: If the completeness of the intent recognition is insufficient, a supplementary query template is generated, and relevant templates are matched from the preset follow-up question guidance library to obtain follow-up questions for missing information.

[0063] By analyzing the intent recognition results, it is determined whether the completeness reaches a preset threshold. If not, a supplementary query process is triggered to obtain the initial missing information category. Based on the obtained missing information category, the corresponding query template is retrieved from a preset guidance library. Using template construction, a supplementary query framework suitable for the current scenario is determined. For the determined supplementary query framework, specific fields related to the missing information are obtained. A template matching mechanism is used to determine the fit between the fields and the template, generating an initial follow-up question structure. Through semantic analysis of the initial follow-up question structure, the context data required for statement generation is obtained. If the context data matches the missing information category, the expression logic of the statement is further optimized to determine the final statement content. Based on the final statement content, preset statement generation rules are used to handle the syntax and semantic consistency of the statement, resulting in a follow-up question that meets the user's interaction needs. By comparing the generated follow-up question with the initial intent recognition data, it is determined whether there is a semantic deviation. If a deviation exists, a backup template is retrieved from the guidance library to generate new follow-up question content.

[0064] In one possible implementation, the system analyzes the intent recognition results to determine if the completeness reaches a preset threshold. If not, a supplementary query process is triggered to obtain a preliminary missing information category. For example, if the confidence score is below 0.85, the system considers the information incomplete and cannot determine a unique service path. Besides the total score, the system also considers the coverage rate of key elements. For instance, if a service requires four core fields, and only two are currently extracted, even with a high semantic score, a supplementary query will be triggered because the element threshold is not met. The missing information category refers to key semantic elements necessary to complete a government service but not yet provided in the current dialogue. The system automatically compares the preliminary structure generated in S104 with the standard service template for this type of intent in the government knowledge base, checking the required fields in the standard template. The system scans the vector weights in the preliminary structure and marks the vector components corresponding to "service location" or "identity certificate" as missing if they are empty or have extremely low weights. The missing spaces are mapped to a preset labeling system to obtain the missing information category.

[0065] In one possible implementation, based on the obtained missing information category, a corresponding query template is retrieved from a preset guidance library. A template-based approach is used to determine a supplementary query framework suitable for the current scenario. The construction process of the preset guidance library involves: extracting excellent follow-up question examples from historical human scripts used by government service hotlines; categorizing and labeling these examples according to the "missing information category"; converting specific human scripts into general, variable templates; and having business experts review the templates to ensure they conform to policy specifications and grammatical logic. After the system goes live, based on feedback from S106 (i.e., whether citizens can understand and effectively respond), the priority of templates in the library is dynamically adjusted or their expression modified.

[0066] In one possible implementation, for a given supplementary query framework, specific fields related to missing information are obtained. A template matching mechanism is used to determine the fit between the fields and the template, generating a preliminary follow-up question structure. Specifically, the system checks whether the data type of the missing field (e.g., numeric, location entity, enumerated option) matches the response type required by the template. For example, if the missing field is "location," the fit algorithm filters out templates asking "when" or "how much," prioritizing templates with the [Location] attribute. The system then calculates the cosine similarity between the candidate template vector and the overall semantic vector of the current dialogue sequence. A high fit score is achieved if the template's expression (e.g., tone, use of technical terms) closely resembles the citizen's previous expression style. Finally, the system rehearses the filled-in sentence structure to check for logical conflicts. If the filled template forms a complete and unambiguous query framework that can loop through to the missing field, the fit is considered satisfactory.

[0067] Furthermore, after the compatibility determination is completed, the system performs the following operations to obtain the preliminary structure:

[0068] 1. Select the highest-scoring template: Choose the template with the highest relevance score.

[0069] 2. Construct the main body of the statement: Determine the questioning method for follow-up questions (e.g., confirmatory question "Are you handling...?" or open-ended question "May I ask what your... is?").

[0070] 3. Pre-defined logical nodes: Reserve "semantic cores" in the statement structure that need to be compared in subsequent S106 iterations.

[0071] Furthermore, the template matching mechanism is a retrieval and filling process based on semantic tag indexing. Specifically, the system enters the corresponding logical branch in the preset guidance library based on the "missing information category" (e.g., missing location, missing time, missing materials) identified in the previous step S105. Each missing category is associated with a set of preset templates. The matching mechanism activates the most relevant candidate templates based on the intent type of the current dialogue (e.g., consultation vs. complaint). The matching mechanism does not directly extract rigid sentences, but rather extracts frames with variables. The system injects the contextual information it has acquired (e.g., business name, processing steps) into the template placeholders to complete the initial structural assembly.

[0072] In one possible implementation, the semantic analysis of the initial question structure is used to obtain the context data required for sentence generation. If the context data matches the information missing category, the expression logic of the sentence is further optimized to determine the final sentence content. The specific optimization methods are as follows:

[0073] 1. Semantic Cohesion Optimization: The system utilizes known information from the context to modify the question, ensuring a natural transition in the statement. For example, if "social security card" has already been mentioned, the optimized statement will use "it" or omit the subject altogether, instead of repeating the lengthy full name of the business. Furthermore, based on the atmosphere of the conversation, it automatically adds conjunctions such as "that," "in addition," and "regarding what you mentioned..." before follow-up questions.

[0074] 2. Logical Premise Elimination: The system checks whether there is a logical inclusion relationship between the missing field and the known context. For example, if the missing category is "detailed address", but the context data already mentions that the citizen "lives in Xihu District", the optimization logic will refine the follow-up question from "Where is your address?" to "What is your specific street and house number in Xihu District?".

[0075] 3. Dynamic matching of predicate and verb: The system adjusts the tense and voice of follow-up questions based on the action state in the context. For example, if a citizen says "I have submitted my application," and the missing category is "location," the system will optimize "[processing] location" in the template to "Where did you [submit]?", ensuring verb consistency.

[0076] 4. Adaptive adjustment of expression granularity: Optimize the tone of the inquiry based on the nature of the missing information. For example, if the context indicates that the citizen's expression ability is weak or that multiple answers are off-topic, the optimization logic will refine "What is your request?" to the more guiding "Do you want to inquire about services or file a complaint?".

[0077] 5. Consistency and Deviation Comparison (Final Guarantee): After optimization, the system executes a crucial verification logic: checking whether the optimized statements conform to grammatical rules and whether they conflict with the intent goals defined in S104. The optimized follow-up questions are cross-referenced with the initial dialogue data. If a misinterpretation is found during optimization (e.g., mistaking a citizen's stated A for the missing B), the system will revert the optimization logic and re-build using a backup template.

[0078] In one possible implementation, based on the final sentence content, a preset sentence generation rule is used to process the grammatical and semantic consistency of the sentence, resulting in a follow-up question that meets the user's interaction needs. This preset sentence generation rule is a set of binding instructions designed to transform structured templates into natural and fluent text. It mainly includes the following three levels of rules: 1. Referential resolution and substitution rules: When the context clearly defines the subject, pronouns (such as "it" or "you") are forced to be used, or the subject is omitted to avoid repetition. Rule example: If Context_Subject == Current_Subject, then Omit(Subject). 2. Voice consistency conversion rules: Ensure that the tone of the question matches the tone of the citizen's expression. Rule example: If the citizen uses formal spoken language, the rule will restrict the use of auxiliary words in the template, adopting more standardized government terminology. 3. Slot completion and modification rules: Automatically add necessary prepositions or conjunctions before and after template variables. Rule example: Automatically complete "at" or "located" before the "location" variable; complete "during..." before the "time" variable.

[0079] The grammatical consistency processing primarily focuses on the structural correctness of sentences; specifically...

[0080] Subject-verb-object verification: The system performs rapid dependency parsing on the generated sentences. If it finds a subject-verb mismatch due to template filling (e.g., the subject is "materials" and the verb is "residence"), the system will trigger automatic correction and readjust the verb.

[0081] Tense and auxiliary word correction: According to the state described by the citizen (already occurred, occurring, about to occur), proofread the modal particles (such as "le", "guo", "ne") at the end of the sentence.

[0082] Smoothing optimization: Eliminate the extra pauses caused by mechanical splicing. For example, optimize "Excuse me, your, handling location, where is it?" to "Excuse me, where is your handling location?".

[0083] The semantic consistency processing focuses on logical coherence and closed-loop. Specifically,

[0084] Logical contradiction detection: The system will convert the generated follow-up question sentence back into a vector and compare it with the initial dialogue data. For example, if it is found that the content of the follow-up question has a high-weight feature record in the initial data (that is, the citizen has actually said it), it is determined as "semantic conflict".

[0085] Semantic deviation comparison: Calculate the correlation degree between the final sentence and the preliminary structure of the current intent recognition. If the comparison finds that the follow-up direction deviates from the core intent (for example, the intent is "social security" and the follow-up question becomes "medical insurance"), the system will immediately discard the currently generated sentence, roll back to the matching template stage, and retrieve the backup template to reconstruct.

[0086] Information increment verification: Ensure that each follow-up question can lead to new information. If the processing result shows that the follow-up question does not contribute to filling the missing information category, it will be regenerated.

[0087] Step S106, according to the new dialogue sequence obtained after the feedback of the follow-up question sentence, repeat the aforementioned semantic vector extraction and context dependence determination process, and obtain an updated structured requirement representation through iterative fusion of new sequence segments.

[0088] Initial semantic vector data is obtained through dialogue sequence processing. Word segmentation and embedding techniques are used to obtain preliminary semantic representation results. Based on these preliminary semantic representation results, contextual dependencies are analyzed, and a pre-established model is used to determine the logical relationships between segments. For the determined logical relationships, data for new segments is acquired. If the correlation between a new segment and an existing sequence is higher than a preset threshold, it is included in the iterative fusion process to determine if it meets the fusion conditions. Through the iterative fusion process, the combination of new segments and existing structured representations is processed, using a cyclic update mechanism to obtain adjusted intermediate representation data. Based on the intermediate representation data, the overall consistency of sequence analysis is analyzed. If the consistency is lower than a preset standard, a supplementary analysis step is triggered to determine the final structured representation form. The final structured representation form is obtained, and data mapping and storage operations are performed based on content update requirements to obtain complete updated content results. Through the updated content results, the matching degree between the semantic vectors and contextual dependencies is verified, and a record-keeping mechanism is used to determine the final output data of the business processing.

[0089] In one possible implementation, initial semantic vector data is obtained through dialogue sequence processing. Continuous sentences are then segmented into the smallest units (words or phrases) with independent semantics using word segmentation technology, i.e., determining what words were said. Embedding technology is then used to map the segmented results into a continuous, low-dimensional, and dense vector space, i.e., determining the mathematical meaning of these words, and outputting preliminary semantic representation results.

[0090] In one possible implementation, the step of analyzing contextual dependencies based on preliminary semantic representation results and determining the logical relationships between segments using a pre-established model typically refers to a trained deep learning semantic association model. In a government affairs scenario, this model specifically includes the following technical components:

[0091] Bidirectional Long Short-Term Memory (Bi-LSTM) or Transformer encoders: These models excel at processing sequential data and can capture long-distance dependencies between contexts in text.

[0092] Semantic matching model: Specifically designed to calculate whether two text fragments are related by "semantic implication", "semantic conflict", or "semantic irrelevance".

[0093] Government Knowledge Graph: The model embeds the logical framework of government business (e.g., it knows that there is a strong business relationship between "replacement" and "ID card").

[0094] Determining the logical relationships between the segments is a process from vector calculation to logical reasoning, and the specific steps are as follows:

[0095] A. Semantic Similarity Calculation: The model first compares the new fragment vector with the existing structured representation vectors. Using cosine similarity or the dot product operator, it determines whether the two are close in the semantic space. If the distance is very close, it indicates that the new fragment directly supplements the existing information.

[0096] B. Temporal Dependency Analysis: The model analyzes the position of new segments within the dialogue sequence. Through follow-up question-answer alignment, it determines whether the citizen's new answer directly responds to the follow-up question posed by S105. Additionally, the model identifies which specific entity (e.g., "social security card") in the preceding text each pronoun (such as "that" or "it") in the new segment refers to, thus establishing a logical connection.

[0097] C. Logical Relationship Classification: The model outputs a probability distribution, classifying the relationships between fragments as follows: Supplementary Relationship: The new fragment provides missing slot information (e.g., providing a specific service location). Corrective Relationship: The new fragment negates or modifies the previous semantics (e.g., a citizen says, "No, I'm not applying for social security, I'm applying for medical insurance"). Parallel / Jump Relationship: The citizen starts a new topic.

[0098] In one possible implementation, for a given logical correlation, data of newly added segments is acquired. If the correlation between the newly added segment and the existing sequence is higher than a preset threshold, such as 0.7 or 0.75, it is included in the iterative fusion process to determine whether it meets the fusion conditions. The iterative fusion process is an incremental update process, with the following specific steps: Based on the attention mechanism mentioned in S104, weights are assigned to the newly added segments (usually, newly added information has higher weights). The newly added segment vector is mathematically fused with the original structured representation vector (e.g., weighted summation or concatenation followed by a fully connected layer). Using a mechanism similar to GRU (Gated Recurrent Unit) or LSTM, the new segment is used as the input at the current time step, combined with the hidden state (original sequence representation) from the previous time step, to generate an updated global semantic vector. The slot information in the initial structure is refreshed in real time, and the newly acquired data is filled into the corresponding spaces. The fusion conditions include:

[0099] Logical non-conflict condition: The new fragment cannot create an irreconcilable contradiction with the existing established information (for example: if the existing information is confirmed to be personal business, but the new fragment says it is corporate business, if it is determined to be a conflict, it will not be directly merged, but may trigger a reconfirmation).

[0100] Semantic increment condition: The added fragment must provide new information or effectively correct old information. If it is just a simple repetition, it will not enter deep fusion.

[0101] Syntactic coherence condition: Determine whether the overall sequence's linguistic logic still conforms to the input requirements of the government affairs classification model after the addition of new segments.

[0102] In one possible implementation, the system iteratively fuses the newly added fragments with existing structured representations, employing a loop update mechanism to obtain adjusted intermediate representation data. Specifically, the system aligns the semantic vectors generated from the newly added fragments with the vectors of the existing structured representations in their feature spaces, ensuring that their dimensions are consistent. By comparing the two vectors, the system calculates the information increments brought by the new fragment. For example, if the "location" slot in the old vector is empty, but the new fragment includes "West Lake District," the weight of that dimension will be significantly activated. The system does not simply add the two vectors together; instead, it uses a fusion operator (usually a multilayer perceptron (MLP) or a gating mechanism) to cross-combine the new and old features, preserving effective contextual dependencies and suppressing redundant or conflicting information.

[0103] The cyclic update mechanism is key to the continuous iteration capability of S106. The system typically employs a logical architecture similar to GRU (Gated Recurrent Unit) or LSTM.

[0104] Hidden state propagation: Treating the existing structured representation as the hidden state of the model in the previous time step ( The newly added data segment is used as the input for the current time step. ).

[0105] Gating control: Forget gate: Automatically identifies and weakens outdated, citizen-corrected, or erroneous semantics in existing representations. Update gate: Controls how much new information can be written into the current understanding model.

[0106] State refresh: Through the above gating operation, the model calculates and outputs the new state at the current moment. ), which is the adjusted intermediate representation data.

[0107] In one possible implementation, the overall consistency of the sequence analysis is analyzed based on the intermediate representation data. If the consistency is lower than a preset standard (typically a confidence score between 0.8 and 0.9), a supplementary analysis is triggered to determine the final structured representation. The overall consistency analysis involves logical conflict detection throughout the entire dialogue sequence. The system transforms the intermediate representation data into high-dimensional spatial features and performs the following operations: First, it analyzes whether there is a "A contains B" or "A modifies B" relationship between the new segment and the historical sequence. If there is an irreconcilable logical gap between the old and new information (e.g., the previous text stated the processing location was in area A, but the new reply says it's in area B without explanation), the consistency score will decrease. Then, the movement trajectory of the intermediate representation vector during the dialogue process is evaluated. If the vector direction swings drastically (suddenly jumping from social security to transportation without logical connection), the consistency is considered low. Finally, it compares the data with the government knowledge base. For example, does the policy logic match the service materials mentioned by the citizen and their expressed business intentions?

[0108] When consistency fails to meet the standard, the system initiates supplementary analysis to prevent the output of incorrect intents. Specifically, the system uses attention weights to backtrack and locate conflicting segments that lead to reduced consistency. It then retrieves deeper-level government knowledge rules to attempt to resolve the conflict, such as determining whether two different locations mentioned by a citizen belong to the same administrative jurisdiction. If the subsequent text explicitly corrects the preceding text, the subsequent text takes precedence. If the semantic vector features of a certain piece of information are more significant, that information is retained. If the current idea graph path logic is flawed, the system attempts to map intermediate representation data to the second-highest probabilities candidate intent and observes whether consistency improves.

[0109] After consistency verification (or supplementary analysis and correction), the system performs dimensionality reduction mapping to fix the final form: projecting the high-dimensional intermediate representation data into predefined government service slots. The numerical strength of each dimension determines the content to be filled in the corresponding slot (e.g., name, item, region, request). The filled slot information is then encapsulated into a standard format (e.g., JSON or structured matrix). Example: [Intent: Social Security Re-application, Region: Xihu District, Subject: Individual, Status: Completed]

[0110] In one possible implementation, the process of verifying the matching degree between the semantic vector and the context dependency by updating the content result typically employs the following techniques:

[0111] A. Similarity quantification: Calculate the distance between the retrieved vector and the query vector using mathematical formulas, such as cosine similarity or Euclidean distance.

[0112] B. Validation of recall accuracy:

[0113] Top-K hit rate: Checks whether the top K results returned by the system contain the known standard answer.

[0114] RAGAS Evaluation Framework: This is a tool specifically designed for RAG systems. It checks fidelity, i.e., whether the generated content truly originates from the retrieved context. Relevance, i.e., whether the retrieved context actually answers the posed question.

[0115] C. Semantic drift test: Verify whether the updated content has caused confusion in the semantic space.

[0116] D. Manual annotation / gold standard comparison: Establish a test set (containing the questions and corresponding standard text). After each data or model update, run the test set and compare the degree of overlap between the results retrieved by the system and the standard answers.

[0117] Step S107: The updated structured demand representation is mapped into the government service classification model, and the final service assignment path is determined by comparing the similarity between the demand representation and the service conditions.

[0118] The process begins by acquiring structured demand data and extracting key information by parsing its internal fields and tags. For the extracted demand representations, a pre-established e-government model is used for preliminary classification, yielding matching results related to service classification. From these matching results, constraint rules related to service conditions are extracted. If the constraint rules match the field content of the demand representation, a preliminary service allocation direction is determined. The similarity value between the demand representation and the service conditions is calculated, and it is determined whether the similarity value reaches a preset threshold. If the threshold is reached, the accuracy of the condition matching is confirmed. Based on the condition matching results and the mapping process of the e-government model, the final allocation path is determined, and a corresponding path identifier is generated. For the generated allocation path, the association rules between the path identifier and service allocation are obtained, and the rationality of the path determination is judged by comparing the rule content. From the path determination results, the final service allocation scheme is extracted, and the entire allocation process is tracked by recording allocation logs.

[0119] In one possible implementation, the extracted demand representation is categorized according to a pre-established government service model based on a government service catalog. For example: primary classification by service recipient (e.g., personal services, business services); secondary classification by business area (e.g., human resources, health, public security); and tertiary classification by specific matter (e.g., passport application, marriage registration, medical insurance reimbursement). Matching results related to the service categories are then obtained. The construction steps of the pre-established government service model include: collecting massive amounts of government service guides, historical case data, and policies and regulations; extracting specialized terms in the government service field (e.g., "one-stop online service," "notification and commitment system") and establishing a word vector library; and training the model on a labeled government service dataset using a deep learning architecture (e.g., BERT or Convolutional Neural Network (CNN)) to enable the model to recognize specific business logic.

[0120] In one possible implementation, the constraint rules related to service conditions obtained from the matching results are like "thresholds" in the e-government system, used to determine whether a user is eligible to handle a certain service. Obtaining these rules from the matching results is for the purpose of automatically verifying the user's eligibility. Common constraint rules include:

[0121] Eligibility requirements: Applicants must be at least 18 years old; must have local household registration; and the company's registered capital must exceed 1 million yuan.

[0122] Eligibility requirements: Original ID card must be provided; proof of social security contributions for the past 6 months must be submitted.

[0123] Logical constraint: Only after completing the "settlement" process can one apply for an "ID card".

[0124] Time limit: The matter must be reported within 30 days from the date of occurrence.

[0125] In one possible implementation, the step of calculating the similarity value between the demand representation and the service conditions, determining whether the similarity value reaches a preset threshold, and confirming the accuracy of the condition matching if the threshold is reached, refers to the system verifying, through a logical closed loop, whether the match can truly be converted into a service action after reaching the threshold. There are typically three methods:

[0126] A. Reverse Element Verification: The system queries the user in reverse based on the matched service conditions. For example, if the system matches "maternity allowance," it will automatically prompt the user with the question: "Have you obtained a birth certificate?" (If the user answers "yes," the accuracy of the match is further confirmed.)

[0127] B. Logical Consistency Check: This checks whether the matched result violates features known to the user. For example, the system recommends "applying for a senior citizen's preferential treatment card" to the user based on semantic matching. The backend retrieves the user's real-name data and finds that the user is only 25 years old. In this case, even if the semantic similarity reaches 0.9, the match will be deemed inaccurate due to a logical constraint violation.

[0128] C. Human Feedback / User Clicks (Closed-Loop Feedback): In the interactive interface, subsequent user actions validate the feedback: if a user clicks on a recommended service and begins the process, the system records positive feedback. If a user repeatedly skips or clicks "irrelevant," the system automatically reduces the weight of that semantic match.

[0129] In one possible implementation, the process of determining the final assignment path and generating a corresponding path identifier based on the result of condition matching and the mapping process of the government affairs model typically includes three core components:

[0130] Organizational structure model: Defines the departmental tree within the government (e.g., municipal bureau -> district bureau -> street), the list of powers and responsibilities of each department, and the authority of approval personnel.

[0131] Business domain ontology: Defines the attributes of government affairs. For example, the "One-Child Certificate" belongs to the "Health and Wellness" domain, and its attributes include "Household Registration" and "Age".

[0132] Decision tree / graph model: a pre-defined logical path. For example: "If the person is a foreign national applying for a work permit, then follow the foreign-related approval process."

[0133] Determining the allocation path is a multi-factor decision-making process, typically employing the following logic:

[0134] A. Static attribute routing: Directly determined based on the ownership of the matter. If the matter identifier is "housing provident fund withdrawal" and the region is "Area A", then the assignment path is "Area A Housing Provident Fund Management Center".

[0135] B. Dynamic Load Balancing Strategy: If multiple windows or departments can handle the task, the system will assign tasks based on real-time status. Factors considered include: the number of people queuing at each window, the number of pending applications, and the online status of the approvers. The logic is to assign the task to the path with the fewest pending applications or the shortest average processing time.

[0136] C. Complex Logic Decisions: Redirecting applications based on the specific circumstances of the user. Example: Ordinary users are assigned via the normal dispatch path. Special groups (disabled persons, active military personnel) are assigned via the green channel dispatch path. Incomplete applications but meeting the credit commitment system are assigned via the incomplete application acceptance dispatch path.

[0137] In one possible implementation, for the generated assignment path, the association rules between the path identifier and service allocation are obtained. By comparing the rule content, the rationality of the path determination is judged. Specifically, for the proposed assignment path ID, all association rules corresponding to that ID are searched, and it is checked whether the current case characteristics violate any rule. If there is no conflict, it is confirmed as reasonable, and a formal assignment instruction is generated; if there is a conflict, it is marked as a path anomaly, and the process is rolled back to the previous step for recalculation or manual intervention is required.

[0138] The association rules are typically retrieved from the power list database, extracted from process definition documents, or automatically inferred from the government knowledge graph based on ontology reasoning. These association rules generally include: 1. Consistency of Authority Rules: Whether a certain type of matter (e.g., "construction permit") is under the jurisdiction of a certain level of department (e.g., "Housing and Construction Bureau"). 2. Authority Boundary Rules: Cases accepted in Zone A cannot be assigned to the approval system in Zone B. 3. Role Matching Rules: Specific path identifiers must correspond to approval personnel with specific qualifications (e.g., only personnel with administrative law enforcement qualifications can appear in the penalty path). 4. Timeliness Matching Rules: Matters promised to be "handled immediately" cannot be assigned to offline approval paths with an average processing time of 3 days.

[0139] The method for determining the reasonableness of the determined path is as follows:

[0140] Attribute collision method: Extract the department attribute pointed to by the path identifier and compare it with the main attribute in the requirement representation. Example: The requirement is "company deregistration", but the path is assigned to "Civil Affairs Bureau". Judgment: Unreasonable (should be Market Supervision Administration).

[0141] Path Coverage Verification: This checks whether the determined path fully covers all necessary steps in the process (preliminary review, acceptance, verification, and decision). Example: If a matter requiring on-site investigation is assigned to a purely online automated approval path, the result is: Unreasonable.

[0142] Compliance score calculation: The system will score the path. If the score is lower than the preset "reasonableness threshold" (e.g., 0.95), an exception will be triggered and the process will be transferred to manual pre-review.

[0143] The above embodiments are merely one of the preferred embodiments of the present invention and should not be used to limit the scope of protection of the present invention. Any modifications or refinements made to the main design concept and spirit of the present invention that are not of substantial significance, but solve the same technical problem as the present invention, should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a knowledge base for a government service hotline based on large-scale modeling technology, characterized in that, The method includes: Obtain the multi-turn dialogue text sequence between citizens and the hotline system, extract the spoken expression content of each turn, and process the spoken expression content through a pre-trained BERT model to obtain a semantic vector representation; For the semantic vector representation, an LSTM network is used to perform temporal modeling on the multi-turn dialogue sequence, and the contextual dependencies are determined by fusing the vector associations between previous and subsequent turns. Based on the contextual dependencies, the semantic fragment connections in the current dialogue sequence are analyzed. If the fragment connection strength is lower than a preset threshold, it is marked as a fragment expression region. An integration priority list is obtained by calculating the similarity between fragments. By using the integrated priority list, high-priority fragment expression regions are selected and merged with the overall semantic vector. An attention mechanism is used to weighted fuse the region vectors to determine the completeness of intent recognition. If the completeness of the intent recognition is insufficient, a supplementary query template is generated, and relevant templates are matched from the preset follow-up question guidance library to obtain follow-up questions for missing information; Based on the new dialogue sequence obtained after the follow-up question, the aforementioned semantic vector extraction and context dependency determination process is repeated, and the updated structured requirement representation is obtained by iteratively fusing the new sequence fragments; The updated structured demand representation is mapped into the government service classification model, and the final service assignment path is determined by comparing the similarity between the demand representation and the service conditions.

2. The method for constructing a knowledge base for a government service hotline based on large-scale modeling technology according to claim 1, characterized in that, The process of obtaining multi-turn dialogue text sequences between citizens and the hotline system, extracting spoken expressions from each turn, and processing these spoken expressions using a pre-trained BERT model to obtain semantic vector representations includes: The system obtains multi-turn text data of citizen dialogues from the hotline system, segments the multi-turn text, and determines the independent dialogue unit for each turn. For each round of dialogue, the spoken expressions are extracted, and irrelevant characters and noise are removed using preset text cleaning rules to obtain clean spoken text data. Pure spoken text data is input into a pre-trained BERT model, processed, and then used to generate corresponding semantic vector representations to determine the semantic features of each round of dialogue.

3. The method for constructing a knowledge base for a government service hotline based on large-scale modeling technology according to claim 1, characterized in that, The semantic vector representation employs an LSTM network for temporal modeling of multi-turn dialogue sequences, determining contextual dependencies by fusing vector associations from preceding and following turns, including: For multi-turn dialogue data, a long short-term memory network is used to perform temporal modeling of the dialogue sequence, process the information of the previous and next turns in the sequence, and obtain the temporal feature representation of the dialogue content. By using temporal feature representation, we analyze the vector information association between successive rounds and determine the association strength between each round by calculating the similarity of semantic vectors between successive rounds. Based on the association strength, the vector information of relevant rounds is fused to construct an overall representation of contextual dependencies and obtain the dependencies between different parts of the dialogue sequence; For the constructed dependencies, if the dependency strength of a certain round is lower than a preset threshold, it is then fused with information from other rounds to supplement it, resulting in a complete context dependency representation.

4. The method for constructing a knowledge base for a government service hotline based on large-scale modeling technology according to claim 1, characterized in that, The process involves analyzing semantic fragment connections in the current dialogue sequence based on the contextual dependencies. If the fragment connection strength is lower than a preset threshold, it is marked as a fragment expression region. An integration priority list is obtained by calculating the similarity between fragments, including: Semantic fragments are extracted from the dialogue sequence, and the text content is split using a word segmentation tool to obtain a preliminary set of semantic fragments; For the initial set of semantic segments, the connection strength between each segment is calculated, the semantic distance between segments is obtained through a semantic vector model, and the quantification result of the connection strength is determined. If the connection strength is lower than a preset threshold, the corresponding segment is marked as a fragment expression region, the distribution of the marked region is obtained, and potential semantic breakpoints are determined. Based on the labeled fragment representation regions, the similarity values ​​between fragments within each region are calculated, and the cosine similarity algorithm is used to compare the fragment content to obtain a similarity ranking. Construct an integration priority list by sorting by similarity.

5. The method for constructing a knowledge base for a government service hotline based on large-scale modeling technology according to claim 1, characterized in that, The step involves selecting high-priority fragment representation regions from the integrated priority list and merging them with the overall semantic vector. An attention mechanism is then used to weightedly fuse these region vectors to determine the completeness of intent recognition. This includes: The data is initially filtered using a priority list to obtain the distribution of high-priority fragment areas and determine the set of key areas. Based on the key region set, the corresponding fragment region data is extracted and initially matched with the overall semantic vector to obtain a preliminary combined vector representation; An attention mechanism is used to process the initially combined vector representations, calculate the importance weights of the region vectors, and determine the contribution of each region to the overall semantics. For the weighted region vectors, perform a fusion operation to generate a fusion result vector; If the completeness of the fusion result vector is lower than the preset threshold, the vectors of regions with low contribution will be re-filtered, supplementary data will be obtained, and the updated fusion vector will be obtained. The updated fusion vector is used to analyze the completeness of intent recognition and determine whether it meets the predetermined standards. If the completeness still does not meet the standard, the semantic vector is locally adjusted and combined with supplementary data to obtain the final intent recognition result.

6. The method for constructing a knowledge base for a government service hotline based on large-scale modeling technology according to claim 1, characterized in that, If the completeness of the intent recognition is insufficient, a supplementary query template is generated. This template is matched with relevant templates from a preset follow-up questioning library to obtain follow-up questions addressing the missing information, including: By analyzing the results of intent recognition, it is determined whether the completeness has reached a preset threshold. If it has not, a supplementary query process is triggered to obtain the preliminary missing information category. Based on the category of missing information obtained, the corresponding query template is obtained from the preset guidance library, and a supplementary query framework suitable for the current scenario is determined by template construction. For the established supplementary query framework, specific fields related to missing information are obtained, and a matching template mechanism is used to determine the degree of fit between the fields and the template, generating a preliminary follow-up question structure. By semantically analyzing the structure of the initial inquiry statement, the contextual data required for statement generation is obtained. If the contextual data matches the information missing category, the expression logic of the statement is further optimized to determine the final statement content. Based on the final statement content, the preset statement generation rules are used to process the syntax and semantic consistency of the statement, resulting in follow-up questions that meet the user's interaction needs. By comparing the generated follow-up questions with the initial data of intent recognition, it is determined whether there is a semantic deviation. If there is a deviation, a backup template is retrieved from the bootstrapping library to generate new follow-up question content.

7. The method for constructing a knowledge base for a government service hotline based on large-scale modeling technology according to claim 1, characterized in that, The process of obtaining a new dialogue sequence based on the feedback from the follow-up question, repeating the aforementioned semantic vector extraction and context dependency determination process, and obtaining an updated structured requirement representation by iteratively fusing the new sequence fragments includes: By processing the dialogue sequence, initial semantic vector data is obtained, and word segmentation and embedding techniques are used to obtain preliminary semantic representation results. Based on the preliminary semantic representation results, the contextual dependencies are analyzed, and the logical relationships between the segments are determined using a pre-established model. For the established logical correlation, the data of the new segment is obtained. If the correlation between the new segment and the existing sequence is higher than the preset threshold, it is included in the iterative fusion process to determine whether it meets the fusion conditions. Through an iterative fusion process, the combination of new fragments and existing structured representations is processed, and a cyclic update mechanism is used to obtain the adjusted intermediate representation data. Based on the intermediate representation data, the overall consistency of the sequence analysis is analyzed. If the consistency is lower than the preset standard, a supplementary analysis step is triggered to determine the final structured representation. Obtain the final structured representation, combine it with the content update requirements, perform data mapping and storage operations to obtain the complete updated content result; By updating the content results, the matching degree between the semantic vector and the context dependency is verified, and a record-keeping mechanism is adopted to determine the final output data of the business processing.

8. The method for constructing a knowledge base for a government service hotline based on large-scale modeling technology according to claim 1, characterized in that, The updated structured demand representation is mapped to a government service classification model, and the final service assignment path is determined by comparing the similarity between the demand representation and the service conditions, including: Obtain the data content of structured requirements, and extract the key information representing the requirements by parsing its internal fields and tags; Based on the extracted demand representations, a pre-established government affairs model is used for preliminary classification to obtain matching results related to service classification; From the matching results, obtain the constraint rules related to the service conditions. If the constraint rules are consistent with the field content of the requirement, determine the initial service allocation direction. By calculating the similarity value between the demand representation and the service conditions, it is determined whether the similarity value reaches a preset threshold. If the threshold is reached, the accuracy of the condition matching is confirmed. Based on the results of condition matching and the mapping process of the government affairs model, the final assignment path is determined and the corresponding path identifier is generated. For the generated assignment path, obtain the association rules between the path identifier and service allocation, and judge the rationality of the path determination by comparing the rule content; From the determined path, the final service allocation scheme is extracted, and the entire allocation process is tracked by recording allocation logs.

Images