Intelligent customer service information retrieval method and system based on large model

Abstract

The invention belongs to the technical field of semantic retrieval, and particularly relates to an intelligent customer service information retrieval method and system based on a large model. Performing retrieval judgment on a differential hierarchical index tree-shaped standard semantic question and answer information base of each main body node in the distributed isolation mapping security chain, returning retrieved standard question and answer information according to key entity frequency sorting, and generating an answer by utilizing a pre-training reasoning generation model and combining with the information base when the retrieved standard question and answer information is not retrieved; when the retrieval semantics comprise a reservation demand, generating a cross-system reservation parameter set by means of a distributed reservation decision tree and standard question and answer information, outputting recommended reservation information through a bidirectional mapping dictionary and a matching recommendation algorithm, and dynamically updating the bidirectional mapping dictionary and retrieval index information by collecting the consistency frequency ratio of recommendation to real reservation in real time. And the semantic retrieval precision and the cross-system reservation accuracy are effectively improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of semantic retrieval, and particularly relates to an intelligent customer service information retrieval method and system based on a large model. BACKGROUND

[0002] In the field of medical intelligent customer service, with the development of smart medical care, users have an increasing demand for personalized and accurate services, but traditional systems have technical bottlenecks. On the one hand, medical knowledge has strong hospital individualization characteristics, and there are obvious differences in department names and diagnosis and treatment processes among different hospitals. However, the traditional system lacks a dynamic adaptive knowledge base management mechanism, and it is difficult to flexibly maintain the individualized data of different hospitals. On the other hand, when the intelligent customer service is connected with the appointment registration system, the data mapping accuracy is insufficient, resulting in a cumbersome jump process from the user's query to the registration, and poor accuracy of department association. At the same time, the traditional retrieval algorithm cannot effectively handle the semantic ambiguity problems such as synonyms, polysemes and different expressions of different departments in the medical field, resulting in low semantic retrieval accuracy and insufficient answer accuracy. These problems collectively reflect the technical defects of traditional systems in individualized knowledge base management, accurate mapping across systems and medical semantic understanding, which affects the application effect of the medical intelligent customer service system. SUMMARY

[0003] In view of the deficiencies of the prior art, the present application provides an intelligent customer service information retrieval method and system based on a large model. The method retrieves and judges the difference classification index tree standard semantic question and answer information database of each subject node in the distributed isolated mapping security chain by obtaining the user retrieval semantic association vector, combining the preset standard key entity mapping table and the three-level index mapping table, and returns the retrieved standard question and answer information in the order of key entity frequency. When no information is retrieved, the answer is generated by using the pre-trained reasoning generation model in combination with the information database. When the retrieval semantics contains an appointment requirement, the distributed appointment decision tree and the standard question and answer information are used to generate a cross-system appointment parameter set, and the recommended appointment information is output by using a bidirectional mapping dictionary and a matching recommendation algorithm. The consistency frequency ratio of real-time collection of recommendations and real appointments is used to dynamically update the bidirectional mapping dictionary and the retrieval index information, effectively improving the semantic retrieval accuracy and the cross-system appointment accuracy.

[0004] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0005] The intelligent customer service information retrieval method based on a large model comprises the following steps:

[0006] In response to the user retrieval information, the user retrieval semantic association vector and the context association vector are obtained.

[0007] based on the search semantic association vector and the context association vector, combining a preset standard key entity mapping table and a three-level index mapping table built in a distributed isolation mapping security chain, performing standard question and answer search information retrieval judgment on a hierarchical index tree standard semantic question and answer information library configured for each subject node of the distributed isolation mapping security chain;

[0008] If the search is successful, the standard question and answer search information retrieved is returned according to the frequency priority of the search key entity in the search semantic association vector;

[0009] If the search is unsuccessful, according to the user search semantic association vector, through a pre-trained inference generation model combining a hierarchical index tree standard semantic question and answer information library, a question and answer search information sequence and a corresponding evaluation score are generated, and the information with the maximum evaluation score among all generated question and answer search information is fed back to the corresponding user.

[0010] Specifically, the intelligent customer service information retrieval method based on a large model further comprises:

[0011] When the search semantic association vector contains a reservation requirement, according to the user reservation requirement, through a distributed reservation decision tree corresponding to each subject node in the distributed isolation mapping security chain, cross-system reservation matching is performed on the standard question and answer search information, and a cross-system reservation parameter set is obtained;

[0012] Based on the cross-system reservation parameter set, through a preset cross-system bidirectional mapping dictionary combining a matching recommendation algorithm, recommended reservation information is obtained;

[0013] The frequency of reservation and the frequency of real reservation of the corresponding user according to the recommended reservation information are collected in real time, and a reservation consistency frequency ratio is obtained. If the consistency frequency ratio is less than a preset reservation consistency threshold, then according to the real-time reservation information parameters, the recommended mapping information in the bidirectional mapping dictionary and the retrieval index information in the three-level index mapping table are constructed and updated until the reservation consistency threshold is met.

[0014] Specifically, the construction process of the standard key entity mapping table further comprises:

[0015] Obtain the historical text record information and the field standardized text information corresponding to each type of subject containing all sub-subjects, and perform word segmentation preprocessing through a pre-trained professional word segmentation model to obtain the text information space and the field standard text information space corresponding to all sub-subjects under each subject;

[0016] Based on the text information space and the field standard text information space corresponding to all sub-subjects under each subject, through an entity relationship extraction algorithm, obtain the exclusive label key entity vector and the general label key entity vector corresponding to each sub-subject;

[0017] With each general label key entity vector as the cluster center, the clustering distance between each cluster center and all exclusive label key entity vectors is obtained through the cosine formula;

[0018] Based on all the clustering distances corresponding to each cluster center, the truncated distance corresponding to each cluster center is determined, and all the truncated distances corresponding to each cluster center and the clustering distance are input into the DBSCAN algorithm to obtain the exclusive label clustering mapping table corresponding to each general label; the exclusive label clustering mapping table includes the general label and the exclusive label sequence under the corresponding different sub-subjects.

[0019] Specifically, the construction process of the standard key entity mapping table further includes:

[0020] Based on the exclusive label clustering mapping table corresponding to all general labels and the domain standard text information space, the causal correlation degree between different exclusive labels under different general labels is obtained through a causal correlation analysis model;

[0021] Based on the causal correlation degree between different exclusive labels under different general labels, a bidirectional mapping connection between the general label key entity and the corresponding exclusive label key entity is constructed, and based on the bidirectional mapping connection and the exclusive label clustering mapping table corresponding to each general label, a standard key entity mapping table is obtained;

[0022] Real-time collection of historical text record information corresponding to all sub-subjects corresponding to each type of subject and revised domain standardized text information, repeating the construction process of the exclusive label clustering mapping table corresponding to each general label and the bidirectional mapping connection, and updating the standard key entity mapping table through an incremental update model combined with the real-time collected information.

[0023] Specifically, the construction process of the distributed isolated mapping security chain includes:

[0024] Based on the domain information in the domain standardized text information, a root node is constructed, and based on the domain standardized text information, a knowledge graph and a graph embedding algorithm are combined to obtain a domain classification tree and embed the domain classification tree into the root node;

[0025] Based on the subject, a first-level node is constructed, and based on the sub-subject, a second-level node is constructed, and the text information space, historical question and answer information and configured standard reply information corresponding to each sub-subject are combined with a tree database to construct a third-level difference database node;

[0026] Based on the domain type corresponding to each subject, a clustering algorithm is used to obtain a similar subject cluster, and based on the domain type corresponding to the similar subject cluster, a first-level index connection between each root node and the first-level node in the similar subject cluster is constructed.

[0027] Specifically, the construction process of the distributed isolated mapping security chain also includes:

[0028] According to the clustering similarity between the primary nodes in the similar subject clustering cluster, cross-subject connections are constructed, and the information security protection level, subject label code and historical cross-subject search access permission information between subjects are combined, and based on the hash algorithm and the encryption permission control algorithm, a hierarchical encryption permission hash mapping sequence is obtained, and the hierarchical private key permission hash mapping sequence in the hierarchical encryption permission hash mapping sequence is mapped to the cross-subject connection between the corresponding subjects, and a cross-subject security mapping connection sequence is obtained;

[0029] Based on the information of all secondary nodes corresponding to each primary node and the subfield type information of the function information corresponding to each secondary node, a secondary index connection is constructed, and through the function information corresponding to all secondary nodes, a similar sub-subject clustering cluster corresponding to similar function sub-subjects and a causal correlation degree between sub-subjects with causal relationship under each subject are obtained through clustering algorithm and causal analysis algorithm;

[0030] Using the clustering similarity between each sub-subject in the similar sub-subject clustering cluster, cross-subject connections between each sub-subject are constructed, and the obtained causal correlation degree is mapped into the cross-subject connection corresponding to the causal relationship, and the hierarchical public key permission hash mapping sequence in the hierarchical encryption permission hash mapping sequence is mapped into the cross-subject connection corresponding to the different subjects, and a cross-subject causal permission connection is obtained.

[0031] Specifically, the construction process of the distributed isolated mapping security chain also includes:

[0032] With the historical search information frequency of each type of key entity stored in the nodes in the third-level difference database corresponding to each secondary node, the evaluation score of each key entity corresponding to the search feedback question and answer information, and the auxiliary pre-appointment consistency of the search feedback question and answer information, a third-level index connection is constructed through a hash algorithm;

[0033] Based on the root node, the primary node, the secondary node, the third-level difference database node, the first-level index connection, the second-level index connection, and the third-level index connection, and combined with the graph neural network algorithm, a hierarchical index tree standard semantic question and answer information library corresponding to each subject is obtained;

[0034] Based on the hierarchical index tree standard semantic question and answer information library corresponding to each subject combined with the cross-subject security mapping connection and the cross-subject causal permission connection, a distributed isolated mapping security chain is constructed through a blockchain algorithm and a distributed framework;

[0035] Based on the distributed isolated mapping security chain, the general label key entity in each sub-field type in the standard key entity mapping table is mapped to the cross-subject security mapping connection of the corresponding sub-field type, the bidirectional mapping connection is mapped to the cross-sub-subject causal authority connection with the exclusive label clustering mapping table, and the exclusive label standardized distributed isolated mapping security chain is obtained.

[0036] Specifically, the process of recommending appointment information includes:

[0037] Based on the exclusive label standardized distributed isolated mapping security chain, a distributed security appointment chain with the same structure is constructed, and each hierarchical index tree standard semantic question and answer information library corresponding to the three-level difference database node under the distributed isolated mapping security chain is replaced by a three-level appointment information library node constructed by the work staff function type, skill proficiency, occupation level, and work scheduling information and the corresponding work staff current load information under the same sub-subject, to obtain a distributed appointment decision tree corresponding to each subject under the distributed security appointment chain;

[0038] According to the search key entity semantics and type information, subject code, sub-subject name, sub-subject code, and work staff function type, skill proficiency, occupation level, and each work staff historical appointment recommendation consistency frequency contained in the secondary nodes between the distributed appointment decision tree corresponding to each subject and the hierarchical index tree standard semantic question and answer information library, a bidirectional mapping dictionary is constructed by a hash algorithm, and a bidirectional balanced mapping dictionary is obtained by using work scheduling information and corresponding work staff current load information, through reinforcement learning and load balancing algorithm.

[0039] Based on the bidirectional balanced mapping dictionary, a cross-system appointment auxiliary connection between the distributed appointment decision tree and the hierarchical index tree standard semantic question and answer information library corresponding to each subject is constructed, and all secondary nodes corresponding to the hierarchical index tree standard semantic question and answer information library corresponding to each subject under the distributed isolated mapping security chain and the distributed appointment decision tree corresponding to the same subject under the distributed security appointment chain are connected by the cross-system appointment auxiliary connection, to obtain a dual search-appointment security chain.

[0040] Specifically, the process of recommending appointment information includes:

[0041] Obtain the user search information, obtain the user search semantic association vector and context association vector through the intent analysis model, and determine whether there is appointment demand information;

[0042] If not, the user retrieval semantic association vector and the context association vector are input to the standard key entity mapping table in the double retrieval-reservation security chain for retrieval key entity standardization, combined with the distributed isolation mapping security chain according to the standardized retrieval semantic association vector and the context association vector, information retrieval is performed, and the retrieved information is fed back;

[0043] If there is a reservation requirement, the standardized retrieval semantic association vector and the context association vector are synchronously input to the distributed isolation mapping security chain and the distributed security reservation chain for retrieval and reservation, and the retrieved information is fed back to the corresponding user at the same time, the retrieved information is fed back to the distributed security reservation chain through the cross-system reservation auxiliary connection to adjust the reservation result, when the information needs to be reserved different sub-subjects, cross-subject association reservation is performed through the cross-subject causal authority connection in the distributed security reservation chain, and the association reservation result is fed back to the corresponding user through the recommendation algorithm;

[0044] Based on the information retrieval and reservation or association reservation process of the corresponding double retrieval-reservation security chain, retrieval and reservation cross-system synchronous simulation training is performed through simulation algorithm, until the evaluation score and reservation consistency frequency ratio of the question and answer retrieval information meet the corresponding threshold value at the same time, the trained double retrieval-reservation security chain is obtained.

[0045] The intelligent customer service information retrieval system based on a large model comprises a response module, a retrieval discrimination module, a discrimination result module, a reservation parameter module, a recommendation module, and a discrimination updating module.

[0046] The response module is used to respond to user retrieval information and obtain user retrieval semantic association vectors and context association vectors.

[0047] The retrieval discrimination module is based on the retrieval semantic association vector and the context association vector combined with the pre-set standard key entity mapping table and the built-in three-level index mapping table in the pre-set distributed isolation mapping security chain, and performs standard question and answer retrieval information retrieval judgment on the hierarchical index tree standard semantic question and answer information library configured in each main node of the distributed isolation mapping security chain.

[0048] The discrimination result module is based on the retrieval discrimination result, returns the standard question and answer retrieval information according to the frequency priority of the retrieval key entity in the retrieval semantic association vector, and when no retrieval is performed, generates a question and answer retrieval information sequence and a corresponding evaluation score through a pre-trained inference generation model combined with the hierarchical index tree standard semantic question and answer information library according to the user retrieval semantic association vector, and feeds back the information with the maximum evaluation score in all generated question and answer retrieval information to the corresponding user.

[0049] The reservation parameter module is used for, when the semantic association vector contains reservation requirements, obtaining a cross-system reservation parameter set by combining standard question and answer retrieval information through a distributed reservation decision tree corresponding to each subject node in a distributed isolated mapping security chain according to user reservation requirements;

[0050] The recommendation module obtains recommended reservation information by combining a matching recommendation algorithm through a preset cross-system bidirectional mapping dictionary based on the cross-system reservation parameter set;

[0051] The discrimination updating module is used for collecting the frequency of reservation according to the recommended reservation information and the frequency of real reservation corresponding to the user in real time, and obtaining a reservation consistency frequency ratio, and if the consistency frequency ratio is less than a preset reservation consistency threshold, then the recommended mapping information in the bidirectional mapping dictionary and the retrieval index information in the three-level index mapping table are constructed and updated according to real-time reservation information parameters until the reservation consistency threshold is met.

[0052] Compared with the prior art, the beneficial effects of the present application are:

[0053] The present application realizes multi-dimensional technical breakthrough by constructing a double-database three-core architecture and a distributed isolated mapping security chain. The semantic vector retrieval and large model inference are used to solve the problem of low semantic retrieval accuracy of traditional systems. The four-level tree information database and three-level index mechanism are used to improve the cross-subject retrieval efficiency. The distributed reservation decision tree and bidirectional mapping dictionary are used to optimize the cross-system reservation process. The causal association analysis and graph neural network are introduced to strengthen the semantic ambiguity resolution and dynamic adaptation capability. The blockchain and encryption permission control are relied on to protect the multi-subject data security and permission management. The scheme forms a closed loop from semantic understanding, intelligent retrieval to accurate reservation, effectively improves the personalized service, cross-system collaboration and semantic understanding ability of medical intelligent customer service, and breaks through the technical bottleneck of traditional systems. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 The present application provides an embodiment 1: a large model-based intelligent customer service information retrieval method flowchart;

[0055] Figure 2 The present application provides an embodiment 2: a large model-based intelligent customer service information retrieval system module diagram. DETAILED DESCRIPTION

[0056] Embodiment 1

[0057] Please refer to Figure 1 The present application provides an embodiment: a large model-based intelligent customer service information retrieval method, applied to medical information question and answer, including the following steps:

[0058] S1, in response to user retrieval information, obtaining user retrieval semantic association vector and context association vector;

[0059] It should be noted that the user retrieval information in the embodiment is medical related question and answer information; it needs to be further explained that the user retrieval semantic association vector in the embodiment is the key entity related to retrieval extracted from the user input text, such as the key text description of disease name, symptom manifestation, etc. After vectorization processing, a high-dimensional vector is formed. Its function is to convert the user natural language question into a machine computable semantic representation, match with the standard label vector in the knowledge base, realize accurate information retrieval and semantic understanding, for example, when the user inputs that diabetes causes vision loss, extract the key entities such as diabetes and vision loss into semantic vectors, which are used to match the knowledge base content of related departments such as endocrinology department and ophthalmology department, and then recommend appropriate diagnosis and treatment information or registration department, improve the retrieval efficiency and answer accuracy of intelligent customer service system;

[0060] It needs to be further explained that the user retrieval semantic association vector and the context association vector in the embodiment are preferably pre-trained medical word segmentation model and Med-BERT pre-trained language model. These models are pre-trained by collecting historical medical question and answer text information before use, so that the corresponding model can better extract the corresponding retrieval semantic entity and the corresponding context vector from the user input text information; it needs to be further explained that the context association vector in the embodiment is a high-dimensional feature vector generated by combining user conversation history and personalized context of target hospital, which is obtained by extracting conversation history, encoding context features, fusing hospital dimension prior knowledge and dynamically adjusting weight; Its function is to capture the context semantic relationship of user question and the hospital exclusive context features, such as the use habit of department alias, and then eliminate the semantic ambiguity across departments and hospitals, so that the system can accurately understand the user's intention according to the personalized terminology and inquiry mode of different hospitals, such as distinguishing the difference of corresponding department of "stroke" in different hospitals, improving the accuracy of complex disease retrieval and cross system reservation; The context semantic relationship includes but is not limited to symptom evolution path and corresponding symptom professional entity information;

[0061] S2, based on the retrieval semantic association vector and the context association vector, combining the pre-set standard key entity mapping table and the three-level index mapping table built in the pre-set distributed isolation mapping security chain, judging the standard question and answer retrieval information retrieval of the hierarchical index tree standard semantic question and answer information database configured in each main node of the distributed isolation mapping security chain;

[0062] S3, if the retrieval is found, the standard question and answer retrieval information is returned according to the frequency priority of the retrieval key entity in the retrieval semantic association vector;

[0063] In this embodiment, the key entities such as disease names and symptoms in the retrieved semantic association vector are extracted, the frequency of each key entity in the retrieved text is counted, such as "diabetes" appearing twice and "visual decline" appearing once, and the standard question and answer information retrieved is frequency matched and prioritized based on the association between the key entities and the question and answer information in the three-level index mapping table. The specific technical means are as follows: first, the frequency of each key entity in the retrieved semantic association vector is calculated using the TF algorithm (term frequency algorithm) to generate a frequency priority list; second, the question and answer items containing the corresponding key entities in the hierarchical index tree library are quickly located through the hash index, and the question and answer items are sorted in descending order according to the matching frequency of the key entities (such as a question and answer containing "diabetes" and "visual decline" and appearing more frequently), and finally the results are returned according to the priority; the role is to ensure that when the user searches, the system can prioritize the question and answer information with higher matching degree based on the frequency weight of the key entities, for example, when the user asks "the reason why diabetes causes visual decline", the standardized answer that simultaneously mentions the key entities such as "diabetes" and "retinal disease" with high frequency is returned first, improving the accuracy of information feedback and the degree of user demand.

[0064] S4, if no information is retrieved, the question and answer retrieval information sequence and the corresponding evaluation score are generated based on the user retrieval semantic association vector through the pre-trained inference generation model combined with the hierarchical index tree standard semantic question and answer information library, and the information with the maximum evaluation score among all the generated question and answer retrieval information is fed back to the corresponding user; it needs to be further explained that the pre-trained inference generation model in this embodiment is trained by combining the pre-trained Chinese Bert model with the index information constructed by the historical medical question and answer information and the historical retrieval in this embodiment, which will not be described here;

[0065] S5, when the retrieval semantic association vector contains reservation requirements, cross-system reservation matching is performed based on the user reservation requirements through the distributed reservation decision tree corresponding to each subject node in the distributed isolation mapping security chain combined with the standard question and answer retrieval information, and the cross-system reservation parameter set is obtained;

[0066] S6, based on the cross-system reservation parameter set, the recommended reservation information is obtained through the pre-set cross-system bidirectional mapping dictionary combined with the matching recommendation algorithm;

[0067] S7, the frequency of the corresponding user making a reservation according to the recommended reservation information and the frequency of the real reservation are collected in real time, and the reservation consistency frequency ratio is obtained, if the consistency frequency ratio is less than the pre-set reservation consistency threshold, the recommended mapping information in the bidirectional mapping dictionary and the retrieval index information in the three-level index mapping table are constructed and updated according to the real-time reservation information parameters until the reservation consistency threshold is met.

[0068] It needs to be further explained that the construction process of the standard key entity mapping table in the embodiment includes:

[0069] The historical text record information and the field standardized text information corresponding to all sub-agents of each type of agent are obtained, and the pre-trained professional word segmentation model is used for word segmentation preprocessing to obtain the text information space and the field standard text information space corresponding to all sub-agents under each agent;

[0070] It needs to be further explained that the agent in the embodiment refers to a hospital, and the sub-agent refers to the corresponding departments in the hospital;

[0071] It needs to be further explained that the process of obtaining the text information space and the field standard text information space corresponding to all sub-agents under each agent in the embodiment includes:

[0072] The historical question and answer text record information of each type of agent and its sub-agent is collected, including but not limited to the department diagnosis and treatment process of the hospital, disease question and answer data, various disease symptoms and diagnosis information, etc., and the standardized text information in the industry field is obtained, including but not limited to the standard department classification and professional department name in the medical industry, disease diagnosis specification and standard disease professional name, etc.;

[0073] Based on the collected information, the pre-trained medical word segmentation model is used for word segmentation preprocessing of these texts to remove stop words, special symbols and other irrelevant contents, and extract key terms; the pre-trained medical word segmentation model is constructed by the pre-trained model disclosed in the art, which will not be described here;

[0074] Taking the hospital agent in the medical field as an example, assuming that the agent A is a hospital, and its sub-agents include the department of cerebral disease and the department of cardiovascular disease; for the historical question and answer text record of the department of cerebral disease in the hospital A, such as the main diagnosis and treatment of diseases such as stroke and headache in the department of cerebral disease, the word segmentation obtains the words such as department of cerebral disease, diagnosis and treatment, stroke and headache, forming the text information space of the sub-agent; and the description of the department of neurology in the industry standard text information is that the department of neurology mainly diagnoses and treats cerebrovascular diseases and headache, etc., and the word segmentation obtains the department of neurology, diagnosis and treatment, cerebrovascular disease, headache, etc., to constitute the field standard text information space.

[0075] Based on the text information space and the field standard text information space corresponding to all sub-agents under each agent, the exclusive label key entity vector and the general label key entity vector corresponding to each sub-agent are obtained through the entity relationship extraction algorithm; it needs to be further explained that the entity relationship extraction algorithm in the embodiment is constructed by combining the Chinese pre-trained Bert model with BILSTM and CRF (conditional random field) algorithm;

[0076] It needs to be further explained that the acquisition process of the exclusive label key entity vector and the general label key entity vector corresponding to each sub-subject in the embodiment includes:

[0077] First, input the text in the text information space and the text in the domain standard text information space into the Chinese pre-trained Bert model, extract the semantic feature representation of the text, then input these features into the BILSTM network, use its ability to process sequence data in both directions to capture long-distance dependencies in the text, then perform sequence labeling on the results output by BILSTM through the CRF layer, identify the entities and the relationship between entities in the text, and then obtain the exclusive label key entity vector and the general label key entity vector corresponding to each sub-subject. Illustratively, for the stroke disease in the diagnosis and treatment of the brain disease department in the brain disease department text information space of the A hospital brain disease department, after processing by the algorithm, the exclusive label key entity vector of the brain disease department and the general label key entity vector of the neurology department can be extracted, and the entity relationship of "diagnosis and treatment" can be identified;

[0078] Taking each general label key entity vector as the clustering center, the clustering distance between each clustering center and all exclusive label key entity vectors is obtained through the cosine formula;

[0079] Based on all clustering distances corresponding to each clustering center, determine the truncated distance corresponding to each clustering center, input all truncated distances corresponding to each clustering center and clustering distance into the DBSCAN algorithm, and obtain the exclusive label clustering mapping table corresponding to each general label; The exclusive label clustering mapping table includes the general label and the exclusive label sequence under the corresponding different sub-subjects.

[0080] It needs to be further explained that the acquisition process of the exclusive label clustering mapping table corresponding to each general label in the embodiment includes:

[0081] With each general label key entity vector as a clustering center, the clustering distance between each clustering center and all exclusive label key entity vectors is calculated by a cosine formula; based on all clustering distances corresponding to each clustering center, a k-distance graph is generated by calculating the Euclidean distance from each exclusive label key entity vector to the k-th nearest neighbor vector and observing the elbow point position of the curve slope mutation in the graph, and the distance value corresponding to the position is determined as the cutoff distance; then the cutoff distance and the clustering distance corresponding to each clustering center are input into the DBSCAN algorithm, and the core points and boundary points are identified by density peak clustering to obtain the exclusive label clustering mapping table corresponding to each general label. For example, taking the "neurology" general label key entity vector as the clustering center, the clustering distances of the "neurology" general label key entity vector with the "encephalopathy department" of A hospital and the "brain department" of B hospital are calculated, and after the cutoff distance is determined by the k-distance graph, the DBSCAN algorithm is input to generate the clustering mapping table corresponding to the "neurology" general label, which contains exclusive labels such as "encephalopathy department" and "brain department". It needs to be further explained that in this embodiment, the role of the cutoff distance is to serve as a threshold for density calculation. When the clustering distance between the exclusive label and the general label is less than or equal to the cutoff distance, it is considered that the semantics of the two are similar and they are included in the same cluster, otherwise they are considered to be different clusters, so as to ensure that the DBSCAN algorithm can accurately identify the core clustering area and generate a clustering mapping table containing the corresponding relationship between the general label and different sub-subject exclusive labels, thereby realizing the consistency mapping of cross-subject semantics.

[0082] Based on the exclusive label clustering mapping table corresponding to all general labels and the field standard text information space, the causal correlation degree between different exclusive labels under different general labels is obtained by combining an expert experience-based causal correlation analysis model;

[0083] It needs to be further explained that the specific process of obtaining the causal correlation degree between different key entities under different general labels in the medical field includes:

[0084] By marking the causal relationship of the exclusive label entity by experts, a labeled dataset containing the trigger-triggered relationship between diseases is constructed. For example, the label "high blood pressure" can trigger "arteriosclerosis"; secondly, the labeled data and the field standard text are input into the causal correlation analysis model constructed by the BERT pre-training model combined with the graph attention network. The semantic vector of the key entity is extracted by BERT, the node features are propagated in the causal relationship graph through the graph attention network, and the causal weight between entities is learned. At the same time, the expert pre-set causal rule library is introduced as prior knowledge to regularize and constrain the correlation degree output by the model, and finally the causal correlation degree between key entities under different general labels is obtained. For example, based on the exclusive label clustering mapping table of the "diabetes" general label, including but not limited to "diabetes" and "abnormal glucose metabolism", combined with the description of "diabetes can cause retinopathy" in the field standard text, the expert marks the causal relationship and inputs the model to obtain the causal correlation degree of "diabetes" and "retinopathy". When the user only describes "visual decline and suffering from diabetes", the system can associate the key entities of "diabetic retinopathy" through causal reasoning; through this causal reasoning process, even if the user cannot describe the real disease name that he wants to express, but through the description of the associated disease name and the external symptoms of the associated disease, the corresponding professional disease information can be accurately matched, and more accurate question and answer information and recommendation results can be given.

[0085] The causal rule library is constructed by those skilled in the art according to the mutual influence relationship between different diseases combined with the correlation algorithm; for example, the strong correlation rule that diabetes causes retinopathy;

[0086] Based on the causal correlation degree between different exclusive labels under different general labels, a bidirectional mapping connection between the key entities of different general labels and the corresponding exclusive label key entities is constructed, and based on the bidirectional mapping connection combined with the exclusive label clustering mapping table corresponding to each general label, a standard key entity mapping table is obtained; it needs to be further explained that the bidirectional mapping connection in the embodiment plays a role in establishing a semantic intercommunication bridge between general labels and exclusive labels, realizing semantic consistency mapping and causal reasoning across subjects and sub-fields; specifically, it realizes bidirectional association between industry standardized terms (such as "neurology") and self-defined terms of each subject (such as "encephalopathy department" and "brain department") by fusing the causal correlation degree of exclusive labels under different general labels, which not only supports mapping exclusive labels (such as "diabetes") searched by the user to general labels (such as "diabetes") to match standardized knowledge, but also can reversely map the associated knowledge of general labels (such as "diabetes causes retinopathy") to exclusive label scenarios based on causal relationship, so that the standard key entity mapping table can not only eliminate semantic ambiguity, but also support complex retrieval based on causal logic, and improve the accuracy of cross-subject knowledge linkage and intelligent reasoning.

[0087] The history text record information corresponding to all sub-main bodies and the revised field standardized text information corresponding to each type of main body added in real time are collected, and the above-mentioned process of constructing the exclusive label clustering mapping table corresponding to each general label and the bidirectional mapping connection is repeated, and the standard key entity mapping table is updated by combining the incremental update model with the information collected in real time.

[0088] It should be further explained that the specific process corresponding to the incremental update model implemented in the embodiment includes:

[0089] After the newly added main body text record and the revised industry standard text are collected in real time, the incremental update model is based on the dynamic vector interpolation algorithm and the hierarchical clustering optimization mechanism. First, the exclusive label and the general label key entity vector are generated by preprocessing the newly added data, then the distance between the newly added vector and the existing clustering center is calculated, if the distance is less than the cutoff distance, the clustering boundary is updated by vector interpolation; exemplary, the weighted average of the newly added vector and the adjacent vector is taken as the new clustering center, otherwise a new cluster is created. At the same time, the bidirectional mapping connection is updated by using the incremental graph neural network, the newly added entity relationship is integrated into the knowledge graph, and the causal association degree is recalculated, finally the dynamic update of the standard key entity mapping table is realized; exemplary, when a system imports "Tangluo disease" as a custom label of "diabetes" through Excel, the system generates its semantic vector after parsing, and calculates the distance with the existing "diabetes" clustering center, if it meets the threshold, the clustering is updated by interpolation algorithm, so that "Tangluo disease" is added to the clustering mapping table, and the causal association weight with entities such as "retinopathy" is updated.

[0090] The process converts the hospital department personalized text and the industry standard text into key entity vectors through the pre-trained word segmentation model and the entity relation extraction algorithm, effectively eliminates the semantic gap between different hospital self-defined terms and industry standards, enables exclusive labels such as "brain disease department" and "brain department" to accurately correspond to general labels such as "neurology department", and ensures the consistency of cross-hospital retrieval. Secondly, based on the clustering algorithm and the cut-off distance, the exclusive label clustering mapping table is determined, the semantic similarity area is accurately divided by density calculation threshold, the core clustering is identified by DBSCAN algorithm, the accuracy and reliability of label classification are enhanced, and the foundation for consistent mapping of cross-subject semantics is laid. In addition, the process combines expert experience and causal relationship analysis model to mine the causal relationship of key entities, integrates the expert rule base through BERT and graph attention network, realizes the reverse reasoning from symptom to cause, supports users to obtain accurate disease question and answer results under fuzzy description, for example, when the user only mentions "visual decline and suffering from diabetes", the system can be associated to "diabetic retinopathy" through causal reasoning. Finally, the incremental update model uses dynamic vector interpolation and incremental graph neural network to respond to the addition of new data in hospitals and the revision of industry standards in real time, automatically adjusts the clustering boundary and causal weight, ensures that the mapping table continuously adapts to the dynamic evolution of medical knowledge, so that the system can automatically optimize the mapping relationship when facing new terms or knowledge updates, and maintain the accuracy of semantic understanding and retrieval.

[0091] It should be further explained that the construction process of the distributed isolation mapping security chain in the embodiment includes:

[0092] The root node is constructed based on the domain information in the domain standardized text information, and the domain classification tree is obtained based on the domain standardized text information, the knowledge graph and the graph embedding algorithm, and the domain classification tree is embedded into the root node;

[0093] The domain classification tree is obtained by expert experience algorithm decomposition according to industry domain information, each sub-domain is constructed according to the level size marked by expert experience and combined with the knowledge graph and the graph embedding algorithm, including the corresponding sub-node of each sub-domain and the corresponding domain annotation information;

[0094] It should be further explained that the process of constructing the domain classification tree in the embodiment includes:

[0095] Based on the expert experience algorithm combined with industry field standardized text, the field information is decomposed into sub-fields. For example, the medical field is decomposed into internal medicine, surgery, etc., and each sub-field is labeled with a level size. For example, internal medicine is a first-level sub-field, and cardiovascular internal medicine is a second-level sub-field. Secondly, based on the decomposed sub-fields and labeled information, a knowledge graph containing parent-child node relationships is constructed. For example, the "medical field" root node points to the internal medicine and surgery sub-nodes. Thirdly, the graph embedding algorithm is used to map the nodes and relationships in the knowledge graph to a low-dimensional vector space, retaining the field hierarchical structure and semantic association. Finally, the embedded field classification tree is integrated into the root node to form a field classification tree containing sub-field nodes and field labeling information. For example, in the medical field, the expert decomposes the first-level sub-fields of internal medicine, surgery, etc., and the second-level sub-fields of cardiovascular internal medicine, neurology, etc. After constructing the knowledge graph, the TransR (knowledge graph embedding) algorithm is used to embed the vector space to obtain the field classification tree embedded in the root node, where the "neurology" node contains its level label as a second-level sub-field and the semantic association vector with concepts such as "neurological diseases".

[0096] Based on the subject, a first-level node is constructed, and a second-level node is constructed by a sub-subject. The corresponding text information space, historical question and answer information, and configured standard reply information under each sub-subject combine a tree-shaped database to construct a third-level difference database node.

[0097] It should be further explained that the construction process of each level node in the present embodiment includes:

[0098] First-level nodes are constructed based on the basic information of the subject, including but not limited to subject code and name, and stored in a distributed database. Second-level nodes are constructed based on the classification information of sub-subjects, establishing parent-child relationships with the first-level nodes; for example, hospitals and departments. Next, for each sub-subject, its text information space, historical question-and-answer information (including but not limited to user question-standard answer pairs), and configured standard response information (i.e., pre-stored predefined response templates) are structured and processed. Third-level nodes are constructed in a tree-structured database using a B+ tree indexing algorithm. It should be further noted that the third-level nodes in this embodiment include the data partition index of the corresponding subject, the symptom text hash value corresponding to each sub-subject under the current subject, and multi-dimensional tags. For example, the multi-dimensional tags are: department-disease-symptom, etc., three-level association tags. For instance, when constructing the third-level nodes of a hospital, the first-level node is the hospital's main... The system is structured with two levels: Level 2 nodes represent the neurology sub-sub-subject, and Level 3 nodes store text information for the department (including but not limited to word segmentation results of stroke diagnosis and treatment procedures), historical Q&A (including but not limited to standard answers regarding stroke symptoms), and standard response templates. A B+ tree index enables rapid retrieval and hierarchical management, ensuring data isolation and efficient access across different departments. Data partitioning indexing refers to the system's use of a storage architecture combining hospital IDs and data partitions to create an independent knowledge base space for each hospital. For example, when deployed in a particular hospital, the system automatically generates a dedicated namespace to store its departmental classifications and user Q&A data for different symptoms within each department, ensuring complete isolation from data from other hospitals. The backend management interface supports parallel login from multiple hospital accounts, allowing administrators to switch hospital dimensions for targeted maintenance of the knowledge base. For instance, a hospital can update its neurology diagnosis and treatment procedures independently without affecting data from other hospitals.

[0099] Based on the domain type corresponding to each subject, clustering algorithms are used to obtain clusters of similar subjects. Then, based on the domain type corresponding to the clusters of similar subjects, a first-level index connection is constructed between each root node and the first-level node in the corresponding cluster of similar subjects.

[0100] Simultaneously, based on the cluster similarity between first-level nodes in similar subject clusters, cross-subject connections are constructed. Combining the information security protection level, subject tag encoding, and historical cross-subject retrieval access permission information between subjects, a hierarchical encrypted access permission hash mapping sequence is obtained based on hash algorithms and encrypted access control algorithms. The hierarchical private key access permission hash mapping sequence in the hierarchical encrypted access permission hash mapping sequence is then mapped to the corresponding cross-subject connections between subjects to obtain a cross-subject secure mapping connection sequence.

[0101] It should be further explained that this embodiment uses a domain clustering algorithm to aggregate subjects into similar subject clusters according to domain type, and constructs a first-level index connection from the root node to the first-level node to achieve rapid location and hierarchical retrieval of domain knowledge. At the same time, it constructs cross-subject connections based on cluster similarity, and integrates information such as security protection level and tag encoding. It generates a hierarchical permission hash sequence through hash and encryption algorithms, and maps it to the connection edge to form a cross-subject secure mapping connection. This ensures the semantic rationality of the connection through domain relevance, and realizes the secure isolation and on-demand sharing of data between subjects through hierarchical permission control. Finally, it constructs a cross-subject association system that combines retrieval efficiency and security protection, which is suitable for scenarios such as medical care with high requirements for domain aggregation and data security.

[0102] It should be further explained that the specific implementation process of the hash algorithm and the encryption access control algorithm in this embodiment includes:

[0103] The DBSCAN clustering algorithm is used to cluster subject domain type vectors to obtain similar subject clusters. Next, based on the similarity of subject domain vectors within a cluster, a first-level index connection from the root node to the first-level node is constructed using cosine distance calculation. In constructing cross-subject connections, the cluster similarity between first-level nodes within similar subject clusters is first calculated to generate cross-subject connection edges. Then, combining the information security protection level, subject tag encoding, and historical access permission records of each subject, the SHA-256 hash algorithm is used to hash the permission parameters, and the AES encryption permission control algorithm is used to perform level-based encryption on the hash values. First, encrypted hash sequences containing different permission levels are generated. Then, the hash sequences encrypted with different private keys are mapped to cross-subject connection edges to form a secure cross-subject mapping connection. For example, in a hospital cluster, if the domain vector similarity between hospitals A and B is 'a', when constructing a cross-subject connection, based on hospital A's protection level (Level 3) and hospital B's level (Level 2), and the subject label encodings being H123 and H456 respectively, the permission parameter hash value is calculated using SHA-256, then encrypted using AES-128 to generate a low-permission hash sequence. After mapping to the connection edge, read-only secure access from hospital A to hospital B's data is achieved. The historical access permission levels are specifically set by those skilled in the art based on the sensitivity of the corresponding data.

[0104] Based on the information of all second-level nodes corresponding to each first-level node and the subdomain type information of the functional information corresponding to each second-level node, a second-level index connection is constructed. Through the functional information corresponding to all second-level nodes, clustering and causal analysis algorithms are used to obtain the clusters of similar sub-sub ...

[0105] It should be further explained that, in this embodiment, the process of constructing a secondary index connection and obtaining causal correlation based on the information of all secondary nodes corresponding to each primary node and the subdomain type to which the functional information of the secondary nodes belongs is as follows: First, extract the functional information of all secondary nodes under the primary node, such as the scope of diagnosis and treatment of ophthalmology, otolaryngology, stomatology, nasal surgery, and neurosurgery. Map their respective subdomain types through a domain classification tree, and use an inverted index algorithm to construct a secondary index connection of subdomain-department to achieve fast retrieval of relevant departments by subdomain; Second, for the functional information of all secondary nodes, first use the DBSCAN clustering algorithm to generate similar functional sub-masters based on the functional vector similarity. The system first identifies clusters of data. Then, using a causal analysis algorithm constructed by combining a BERT pre-trained model with a causal convolutional network, the functional text is input into BERT to extract semantic features. These features are then captured through causal convolutional layers to capture long-distance causal dependencies and calculate the causal correlation between sub-subjects. For example, it analyzes the causal correlation between endocrinology and ophthalmology due to diabetic retinopathy. For instance, under a primary node of a hospital, the functional information of the secondary node ophthalmology is mapped to the ENT sub-domain to construct an index. Simultaneously, clustering reveals that ophthalmology and endocrinology belong to the same metabolic eye disease treatment cluster. The causal analysis algorithm calculates the causal correlation between the two departments due to vision impairment caused by diabetes, which is used for cross-departmental association retrieval and collaborative diagnosis and treatment.

[0106] By utilizing the cluster similarity between sub-sub ...

[0107] It should be further explained that this embodiment uses a medical scenario as an example. When a patient has diabetes, the endocrinology department is responsible for basic blood glucose control, while the ophthalmology department may need to handle diabetic retinopathy. By calculating the cluster similarity between the endocrinology and ophthalmology departments, a cross-sub-entity connection is constructed to clarify the association between the two departments in the context of treating diabetic complications. The causal relationship of "diabetes leading to retinopathy" is mapped to this connection, enabling the system to quickly identify the causal relationship between the two departments when processing patient-related inquiries, achieving cross-departmental knowledge linkage and providing more comprehensive treatment suggestions. Furthermore, a hierarchical public key permission hash mapping sequence is mapped to the connection to ensure that endocrinologists, after obtaining authorization, can securely access ophthalmology's treatment data on diabetic retinopathy, and ophthalmologists can also view the endocrinology department's blood glucose control plan for patients. This ensures the secure isolation of data from different departments while enabling cross-departmental data collaboration and permission sharing when necessary, optimizing the patient's treatment process and information interaction efficiency, thereby obtaining a cross-sub-entity causal permission connection that combines causal association and secure permission control.

[0108] The three-level index is constructed using a hash algorithm, based on the frequency of historical retrieval information for each type of key entity stored in the three-level differential database corresponding to each secondary node, the evaluation score of retrieval feedback question and answer information for each key entity, and the auxiliary appointment consistency of retrieval feedback question and answer information. The auxiliary appointment consistency of retrieval feedback question and answer information refers to the consistency between the appointment recommendation information corrected by the retrieval information and the user's actual appointment information. For example, the system recommends an appointment with Dr. C in ophthalmology, but the user actually makes an appointment with Dr. D in ophthalmology.

[0109] Based on the root node, first-level node, second-level node, and third-level differential database node, combined with first-level index connection, second-level index connection, and third-level index connection, and combined with graph neural network algorithm, a hierarchical index tree-structured standard semantic question-answering information database corresponding to each subject is obtained;

[0110] It should be further explained that this embodiment extracts the retrieval frequency, feedback evaluation score, and auxiliary appointment consistency data of key entities from the secondary nodes corresponding to the tertiary database nodes. After normalization and weighted summation, a tertiary index connection is constructed using a hash algorithm to achieve rapid location of key entities with high frequency, high evaluation, and high appointment consistency. Then, the root node to the tertiary nodes and the index connections at each level are integrated, and a graph neural network algorithm is used to learn the semantic associations and hierarchical relationships between nodes. Domain knowledge and subject information are integrated to generate a hierarchical index tree-structured standard semantic question-and-answer information database. Its function is to improve the accuracy of retrieval by constructing an index through multi-dimensional data, and to achieve semantic fusion and hierarchical management by combining graph neural networks. The final information database can support users to quickly call high-frequency and high-quality question-and-answer data when consulting. At the same time, the semantic associations between knowledge are strengthened through index connections and graph structures, improving the efficiency of information retrieval and the adaptability of answers in complex scenarios.

[0111] It should be further explained that the process of obtaining the hierarchical indexed tree-structured standard semantic question-and-answer information database in this embodiment includes:

[0112] First, for each secondary node corresponding to a tertiary differential database node, the frequency of historical key entity retrieval information is extracted, i.e., the number of times each key entity is retrieved is counted. Then, using sentiment analysis algorithms in natural language processing, sentiment polarity and scoring are calculated on the retrieval feedback information to obtain an evaluation score for the retrieval feedback question-and-answer information corresponding to each key entity. When calculating the auxiliary reservation consistency of the retrieval feedback question-and-answer information, a sequence matching algorithm is used to compare the reservation recommendation information corrected by the retrieval information with the user's actual reservation information to determine the degree of matching. After normalizing the data from the above three dimensions, a weighted summation algorithm is used to generate a comprehensive weight. Then, the SHA-256 hash algorithm is used to perform a hash operation on the comprehensive weight and the key entity information to construct a three-level index connection, enabling rapid location of high-frequency, high-evaluation, and highly consistent key entity information.

[0113] Secondly, the existing root node, first-level node, second-level node, and third-level differential database node are integrated with the first-level index connection, second-level index connection, and third-level index connection. Then, a graph neural network algorithm is used to learn the semantic associations and hierarchical relationships between nodes through node feature aggregation and message passing mechanisms. Domain knowledge, subject information, sub-subject functions, and key entity question and answer data are fused to generate an initial hierarchical index tree-shaped standard semantic question and answer information database corresponding to each subject.

[0114] For example, under a hospital, the ophthalmology department serves as a second-level node. Its third-level differential database node stores retrieval records for key entities such as cataract surgery. Statistical analysis reveals a high retrieval frequency for cataract surgery, and sentiment analysis shows positive user feedback. Furthermore, the appointment recommendations and actual appointments show strong consistency. A third-level index connection is constructed using weighted summation and hash operations. The ophthalmology node is then connected to the hospital's first-level node, the root node of the medical field, and the various levels of indexes, and input into a graph neural network algorithm. This ultimately forms a hierarchical, tree-structured standard semantic question-and-answer information database containing both ophthalmology-specific knowledge and industry-wide general knowledge. This database allows users to quickly access high-frequency, high-quality, and relevant question-and-answer information when consulting about cataract treatment.

[0115] Based on the hierarchical index tree-structured standard semantic question-answering information database corresponding to each subject, combined with cross-subject secure mapping connections and cross-subject causal permission connections, a distributed isolated mapping security chain is constructed through blockchain algorithms and a distributed framework.

[0116] Based on the distributed isolation mapping security chain, the common label key entity of each subdomain type in the standard key entity mapping table is mapped to the cross-subject security mapping connection of the corresponding subdomain type. The bidirectional mapping connection and the exclusive label clustering mapping table are mapped to the cross-subject causal permission connection to perform cross-subject exclusive label association mapping, thus obtaining a distributed isolation mapping security chain with exclusive labels.

[0117] It should be further explained that this embodiment integrates a hierarchical indexed tree-structured standard semantic question-and-answer information database, cross-subject secure mapping connections, and cross-subject causal permission connections into a distributed isolated mapping security chain through blockchain algorithms and a distributed framework, thereby achieving distributed storage and secure isolation of data. It maps general-label key entities in the standard key entity mapping table to cross-subject secure mapping connections, ensuring consistency in standardized retrieval among different subjects. Through bidirectional mapping connections and the association mapping of exclusive label clustering mapping tables in cross-subject causal permission connections, it strengthens the semantic and causal relationships between exclusive labels. The resulting distributed isolated mapping security chain with standardized exclusive labels not only ensures secure isolation of data between different hospitals and departments but also achieves efficient semantic association and retrieval, supporting cross-subject and cross-subdomain knowledge linkage and secure sharing.

[0118] It should be further explained that the distributed isolation mapping security chain in this embodiment is used to achieve secure isolation and semantic association of data between different subjects and sub-subjects. Based on blockchain and distributed algorithms, it integrates a hierarchical index tree-structured standard semantic question-and-answer information database, cross-subject secure mapping connections, and cross-subject causal permission connections into a distributed storage structure. Secure node links are achieved through hash pointers, and the RSA algorithm is used to ensure data integrity and immutability. The specific workflow includes: first, mapping the general-label key entities in the standard key entity mapping table to the cross-subject secure mapping connection through hash value matching to achieve standardized retrieval between different subjects; then, using the graph database attribute graph model, transforming the causal relationships and semantic associations between the bidirectional mapping connection and the exclusive label clustering mapping table into edge attributes and node attributes of the cross-subject causal permission connection, embedding the connection structure through a graph traversal algorithm, and updating the edge weights based on the causal association degree to strengthen the exclusive label association. For example, "diabetes" is used as a general label to achieve secure sharing between hospitals through hash matching, while specific labels such as "diabetes" are associated with cross-departmental causal edges through graph traversal, ultimately forming a distributed chain structure that combines secure isolation and semantic association, supporting efficient retrieval and knowledge linkage across subjects and sub-subjects.

[0119] It should be further explained that the specific process of mapping the bidirectional mapping connection and the exclusive label clustering mapping table to the cross-sub-subject causal permission connection for cross-sub-subject exclusive label association mapping in this embodiment includes:

[0120] First, using blockchain algorithms and Distributed Hash Table (DHT) algorithms, a hierarchical indexed tree-structured standard semantic question-and-answer information database for each entity is distributed and stored, with each node storing a partial data shard. Simultaneously, cross-entity secure mapping connections and cross-sub-entity causal permission connections are used as on-chain data structures, with hash pointers enabling secure links between nodes, forming a distributed, isolated mapping security chain. During this process, the RSA algorithm is used to sign and verify on-chain data, ensuring data integrity and immutability.

[0121] It should be further explained that in this embodiment, when constructing the distributed isolation mapping security chain, the hash pointer achieves secure links between nodes through the following process: First, each data node (such as a third-level node in a four-level tree database) generates a unique hash value (such as a SHA-256 hash) during distributed storage. This hash value contains information such as node data content, timestamp, and permission identifier. Second, the cross-subject secure mapping connection and the cross-subject causal permission connection serve as edge structures on the chain, and their hash pointers point to the hash value of the target node, forming a chain structure of source node hash, edge hash, and target node hash. When data is updated, only the hash value of the corresponding node needs to be updated, and the data is synchronized to other nodes through the PBFT algorithm. The chain association of hash pointers automatically verifies the legality of the data change. For example, when the data of the neurology node in Hospital A changes, a new hash value H1 is generated. The pointer of the cross-subject connection is updated from the original hash H0 to H1. Other nodes verify the association between H1 and H0 and the RSA signature to confirm that the data has not been tampered with, thereby realizing secure links and dynamic verification between nodes based on hash pointers.

[0122] Secondly, for the common tag key entities in the standard key entity mapping table, the corresponding cross-entity secure mapping connection is located through a lookup algorithm. Specifically, the hash value of the common tag key entity is matched with the node identifier in the cross-entity secure mapping connection, and the automatic mapping is achieved using smart contracts to ensure standardized retrieval of common tags across different entities.

[0123] Third, the bidirectional mapping connection and the exclusive label clustering mapping table are mapped to the cross-sub-subject causal permission connection for cross-sub-subject exclusive label association mapping. First, using the attribute graph model of the graph database, the causal relationship in the bidirectional mapping connection and the semantic association in the exclusive label clustering mapping table are transformed into the edge attributes and node attributes of the cross-sub-subject causal permission connection. Through graph traversal algorithm, nodes and edges with similarity greater than the similarity threshold with exclusive labels are found in the cross-sub-subject causal permission connection, and the corresponding bidirectional mapping relationship and clustering information are embedded into the connection structure. At the same time, combined with the causal association degree between sub-subjects with causal relationships under each subject, the weight of the connection edge is updated through a weighted average algorithm to strengthen the association between exclusive labels with strong causal relationships.

[0124] For example, in the medical field, diabetes, as a general-purpose key entity, is mapped to a cross-subject secure mapping connection between hospital A and hospital B through hash matching, enabling secure data sharing. Meanwhile, specific tags such as diabetes mellitus and diabetic retinopathy are associated with causal relationship edges of diabetic retinopathy in the cross-subject causal permission connection between endocrinology and ophthalmology through graph traversal algorithms, and the weight of the edges is adjusted according to the degree of causal association. Ultimately, a standardized distributed isolated mapping security chain with specific tags is formed, which not only ensures the secure isolation of data between different hospitals and departments, but also achieves efficient semantic association and retrieval.

[0125] This process, through multi-dimensional technology integration and algorithm application, constructs a hierarchical, secure, and efficient semantic association and data management system. First, at the domain knowledge structuring level, a domain classification tree is constructed based on expert experience and graph embedding algorithms, and the root node is embedded. This system decomposes industry domain information into a hierarchical subdomain structure, enabling different subjects and sub-subjects to perform knowledge association and integration within a unified domain framework.

[0126] In terms of data node construction and index connection, by constructing first- to third-level nodes and corresponding index connections, refined management from the main body to sub-subjects and then to specific data nodes is achieved; the B+ tree index algorithm is used to construct three levels of differentiated database nodes, ensuring efficient data storage and fast retrieval; the hash algorithm is used to construct three levels of index connections, and combined with multi-dimensional information such as retrieval frequency, feedback score and appointment consistency, high-quality question and answer data can be accurately located, improving the quality and efficiency of users obtaining information. Meanwhile, the establishment of index connections at all levels enables close semantic relationships between data at different levels, supporting multi-dimensional rapid retrieval and knowledge linkage. In terms of security and access control, hash algorithms and encrypted access control algorithms generate cross-subject secure mapping connections and cross-subject causal access connections, ensuring secure data isolation between different subjects and sub-subjects while enabling secure data sharing and collaborative access under authorized conditions. Combining causal relationships with enhanced knowledge linkage between sub-subjects ensures that in scenarios such as healthcare, relevant departments can achieve knowledge complementarity and cross-departmental collaborative diagnosis and treatment based on causal relationships. Finally, a distributed isolated mapping security chain is constructed using blockchain algorithms and a distributed framework, incorporating standard key entity mapping tables and exclusive tag information, achieving distributed data storage and secure sharing while ensuring standardized mapping of exclusive tags between different subjects and sub-subjects. This process not only ensures data integrity and immutability but also significantly improves the system's semantic understanding and knowledge reasoning capabilities in complex scenarios by strengthening causal relationships and semantic connections, providing users with more accurate and comprehensive question-and-answer services and data support.

[0127] It should be further explained that the process of recommending appointment information in this embodiment includes:

[0128] Based on the standardized distributed isolation mapping security chain with exclusive labels, a distributed security reservation chain with the same structure is constructed. The three-level differential database node corresponding to each hierarchical index tree standard semantic question and answer information database under the distributed isolation mapping security chain is replaced with the three-level reservation information database node under the same sub-subject, which is constructed by the staff's job type, skill proficiency, professional level and work schedule information and the corresponding staff's current load information. The distributed reservation decision tree corresponding to each subject under the distributed security reservation chain is obtained.

[0129] It should be further explained that the distributed secure appointment chain in this embodiment aims to achieve efficient, intelligent and secure appointment decisions. It is based on a standardized distributed isolated mapping security chain with exclusive labels. It constructs a distributed appointment decision tree with the same structure as the hierarchical index tree standard semantic question-and-answer information database. It also constructs three-level appointment information database nodes based on information such as staff job types and skill proficiency to achieve digital management of medical resources. The specific workflow is as follows: First, multi-dimensional information of staff under each sub-entity is collected. After data cleaning and feature engineering, a three-level node is constructed using the B+ tree indexing algorithm to achieve fast retrieval. Next, a bidirectional mapping dictionary is constructed using a hash algorithm to link the appointment decision tree with the semantic question-answering information database. A bidirectional balanced mapping dictionary is generated through reinforcement learning and load balancing algorithms to optimize staff load distribution. Then, based on the bidirectional balanced mapping dictionary, a cross-system appointment auxiliary connection is constructed between the distributed appointment decision tree under each entity and the corresponding second-level node of the hierarchical index tree standard semantic question-answering information database to obtain a dual retrieval-appointment security chain. After obtaining user retrieval information, the intent parsing model determines the appointment demand. If there is no appointment demand, information retrieval feedback is provided. If there is an appointment demand, retrieval and appointment are performed simultaneously. The appointment results are adjusted with the help of the cross-system appointment auxiliary connection. Cross-sub-entity associated appointments are achieved through cross-sub-entity causal permission connections. Finally, the results are fed back to the user through a recommendation algorithm to ensure reasonable resource allocation and accurate appointments. The intent parsing model is preferably a pre-trained Chinese BERT model combined with a fine-tuned model based on the user's historical retrieval text information.

[0130] It should be further explained that the construction process of the three-level reservation information database node in this embodiment includes:

[0131] First, using the hierarchical index tree-like standard semantic question-and-answer information database of each subject in the standardized distributed isolation mapping security chain with exclusive labels as a template, the hierarchical structure of the root node, first-level node, and second-level node is retained to construct a distributed reservation decision tree.

[0132] When constructing the three-tiered appointment information database nodes, the following steps are taken: First, the functional type (e.g., attending physician, nurse), skill proficiency (assessed through historical medical data combined with machine learning algorithms), professional level (professional title evaluation information), and work schedule information (structured data from calendar schedules) of staff under each sub-entity are collected. Simultaneously, the current workload information of staff is acquired in real time, such as the number of patients awaiting treatment and the progress of patients currently seeing patients. Second, data cleaning algorithms are used to remove redundant and erroneous data. Feature engineering algorithms are used to transform unstructured or semi-structured data into structured feature vectors, such as converting skill proficiency into a quantified value in the 0-1 range. Third, a B+ tree index algorithm is used, with the staff ID as the primary key, and information such as functional type, skill proficiency, professional level, work schedule, and current workload as index fields to construct the three-tiered appointment information database nodes. Each node contains a data partition index, personnel information hash values, and multi-dimensional tags, including department, function, and skill level, enabling rapid location and efficient retrieval.

[0133] For example, under the surgical sub-entity of a hospital, information such as each doctor's job type (e.g., orthopedic surgeon), skill proficiency (assessed based on past orthopedic surgery success rates), professional level (chief physician), work schedule (Monday to Friday mornings), and current workload (3 patients to be seen) is collected. After data cleaning and feature engineering, a three-level appointment information database node is constructed using the B+ tree index algorithm. This enables the system to quickly filter out doctors who meet the needs of patients and have a reasonable workload, ultimately forming a complete distributed secure appointment chain that supports efficient, intelligent, and secure appointment decisions.

[0134] Based on the semantic and type information of key entities, subject codes, sub-subject names, sub-subject codes, and staff function types, skill proficiency, professional levels, and historical appointment recommendation consistency frequency of each staff member, contained in the secondary nodes of the distributed appointment decision tree and hierarchical index tree standard semantic question-answering information database corresponding to each subject, a bidirectional mapping dictionary is constructed using a hash algorithm. Furthermore, using work schedule information and the current load information of the corresponding staff, a bidirectional balanced mapping dictionary is obtained through reinforcement learning and load balancing algorithms. It should be further noted that this embodiment defines a bidirectional mapping dictionary containing fields such as hospital code, custom department name, and standard department code. When the system connects to a new hospital, the large model first parses the semantics of the custom department name, matches it with the standard department code, and then the administrator confirms and writes it into the mapping dictionary.

[0135] It should be further explained that the construction process of the bidirectional balanced mapping dictionary in this embodiment includes:

[0136] First, for each entity, key entity information, entity code (e.g., hospital number), sub-entity name (e.g., neurology department), sub-entity code (unique department identifier), and staff job type, skill proficiency, and professional level are extracted from the second-level nodes of the distributed appointment decision tree and hierarchical indexed tree standard semantic question-answering information database. This information is then combined and concatenated, and a hash value is generated using the SHA-256 hash algorithm. A bidirectional mapping dictionary, indexed by the hash value and linking the appointment decision tree and semantic question-answering information database data, is constructed to enable rapid bidirectional information retrieval.

[0137] When constructing the bidirectional balanced mapping dictionary, the work schedule information (such as schedule cycle and work period) and current load information (such as remaining patient volume and average patient duration) of staff are first obtained from the three-level appointment information database nodes of the distributed appointment decision tree. A load balancing algorithm based on dynamic weights is adopted, assigning initial load weights to staff according to their skill proficiency and professional level, with higher weights given to staff with higher skill proficiency and professional levels. Combining the work schedule information, the schedule period is subdivided into time slices, and the theoretical patient volume that staff can receive in each time slice is calculated. Current load information is monitored in real time; when a staff member's remaining patient volume is below a threshold, their load weight is reduced; conversely, it is increased. By continuously adjusting the weights and using the Hungarian algorithm to optimize task allocation, the load of staff within the same time period tends to be balanced, generating a bidirectional balanced mapping dictionary that includes the relationship between personnel, time, and load balance.

[0138] For example, in a hospital, the neurology department acts as a secondary node. A bidirectional mapping dictionary is constructed by hashing its disease retrieval information, department code, and data such as the functions and skills of the department's doctors. For the doctors in this department, based on the schedule and real-time patient reception, dynamic weights and a Hungarian algorithm are used to adjust the doctors' patient reception tasks at different times. For instance, 2 of the 5 appointments originally concentrated on Doctor C between 10:00 and 11:00 AM are rationally allocated to Doctor D, whose workload is lower. Ultimately, this forms a bidirectional balanced mapping dictionary that ensures reasonable resource allocation and improves service efficiency.

[0139] It should be further explained that this embodiment uses a distributed isolated mapping security chain as a template to construct a distributed appointment decision tree, replacing the third-level nodes with information such as staff functions and skills. Combined with a B+ tree indexing algorithm, it enables rapid retrieval and management of personnel information. A bidirectional mapping dictionary constructed using a hash algorithm links the appointment decision tree with semantic question-and-answer information database data, eliminating semantic ambiguity between hospital-defined department names and standard department codes, and supporting cross-system department semantic conversion. A bidirectional balanced mapping dictionary constructed based on dynamic weights and the Hungarian algorithm dynamically adjusts appointment allocation according to staff scheduling and workload, achieving balanced optimization of personnel load within the same time period. The resulting distributed secure appointment chain and bidirectional balanced mapping system not only realizes the full-process digitization from disease knowledge retrieval to doctor resource matching, but also improves the utilization rate of medical resources through load balancing algorithms, ensuring users receive accurate and efficient appointment recommendations, and optimizing resource allocation and user experience in medical services.

[0140] Based on a bidirectional balanced mapping dictionary, a cross-system reservation auxiliary connection is constructed between the distributed reservation decision tree under each subject and the corresponding second-level nodes of the hierarchical index tree standard semantic question-answering information database. The cross-system reservation auxiliary connection is then used to perform many-to-many connections between the hierarchical index tree standard semantic question-answering information database corresponding to each subject under the distributed isolated mapping security chain and all second-level nodes corresponding to the distributed reservation decision tree corresponding to the same subject under the distributed security reservation chain, thus obtaining a dual retrieval-reservation security chain.

[0141] Obtain user search information, use the intent parsing model to obtain the semantic association vector and contextual association vector of the user search, and determine whether there is any reservation request information;

[0142] If it does not exist, the user's retrieval semantic association vector and contextual association vector are input into the standard key entity mapping table in the dual retrieval-reservation security chain to standardize the retrieval key entities. Based on the standardized retrieval semantic association vector and contextual association vector combined with the distributed isolation mapping security chain, information retrieval is performed, and the retrieved information is fed back.

[0143] If an appointment is needed, the standardized retrieval semantic association vector and contextual association vector are simultaneously input into the distributed isolated mapping security chain and the distributed secure appointment chain for retrieval and appointment. The retrieved information is fed back to the corresponding user, and the information is also fed back to the distributed secure appointment chain through a cross-system appointment auxiliary connection to adjust the appointment results. When the retrieved information requires appointments with different sub-subjects, cross-sub-subject causal permission connections are used in the distributed secure appointment chain to perform cross-sub-subject associated appointments, and the associated appointment results are fed back to the corresponding user through a recommendation algorithm. It should be noted that in this embodiment, cross-subject associated appointments can achieve the following: when a user asks how to treat diabetes and hypertension, the system parses keywords to match the primary department, retrieves a combined treatment plan or generates a comprehensive answer through a large model, recommends appointments with multiple departments, and prompts the user to consult multiple doctors. Further explanation is needed. According to this process, when a user inquires about stroke treatment at Hospital A, the system recommends its neurology department and redirects the user to the registration system via a two-way mapping dictionary, carrying parameters such as the hospital code, standard department code, and problem semantic tags. The registration system then retrieves the hospital's distributed appointment decision tree based on the parameters, converts it into a local department ID, displays the corresponding doctor based on the semantic tags, and feeds back the corresponding doctor's information to the user for appointment decision-making.

[0144] If a user mentions persistent postoperative pain after an orthopedic consultation, the system identifies the semantic tags of pain, automatically links to the pain management department, and pushes a referral prompt, asking if the user needs to be redirected to make an appointment.

[0145] Based on the information retrieval and reservation or associated reservation process corresponding to the dual retrieval-reservation security chain, the retrieval and reservation are synchronously simulated and trained across systems through a simulation algorithm until the evaluation score corresponding to the question-answer retrieval information and the reservation consistency frequency ratio simultaneously meet the corresponding thresholds, thus obtaining the trained dual retrieval-reservation security chain.

[0146] This embodiment achieves deep integration and secure collaboration between semantic retrieval and intelligent appointment by constructing a dual retrieval-appointment security chain, demonstrating multi-dimensional technical advantages and application value in fields such as healthcare. First, a distributed appointment decision tree is constructed based on a hierarchical indexed tree-structured standard semantic question-and-answer information database. The third-level nodes are replaced with information such as staff functions, skills, schedules, and workloads, forming a distributed secure appointment chain. This enables end-to-end digital management from knowledge retrieval to resource scheduling, providing users with accurate appointment decision support.

[0147] In terms of information association and mapping, a bidirectional mapping dictionary is constructed using a hash algorithm, enabling rapid bidirectional retrieval of semantic question-and-answer information and appointment decision data. This ensures that the symptom information retrieved by users can be accurately associated with staff possessing the corresponding skills. Furthermore, the bidirectional balanced mapping dictionary, built based on dynamic weights and the Hungarian algorithm, can intelligently schedule staff according to their shift schedules and real-time load, effectively balancing workload and improving resource utilization and service efficiency.

[0148] When processing user retrieval information, an intent parsing model is used to determine whether there is an appointment requirement, enabling intelligent routing of retrieval and appointment. For users who only need information retrieval, standardized retrieval is performed using a standard key entity mapping table and a distributed isolated mapping security chain to ensure the accuracy and consistency of information. For users with appointment requirements, retrieval and appointment are performed simultaneously using a distributed isolated mapping security chain and a distributed secure appointment chain. Information interaction and appointment result adjustment are achieved through cross-system appointment auxiliary connections, supporting cross-sub-entity associated appointments to meet the multi-departmental collaboration needs in complex medical scenarios.

[0149] Finally, a cross-system synchronous simulation training of the dual retrieval-appointment security chain was conducted using simulation algorithms to continuously optimize the evaluation score of question-answering retrieval and appointment consistency. This ensures that the system can consistently provide high-quality retrieval and appointment services in practical applications, improving user satisfaction and the system's intelligence level. This complete technical system not only guarantees data security and isolation but also achieves efficient collaboration between knowledge retrieval and resource scheduling, providing strong technical support for intelligent services in fields such as healthcare.

[0150] Example 2

[0151] Please see Figure 2 Another embodiment of the present invention provides: an intelligent customer service information retrieval system based on a large model, comprising: a response module, a retrieval and discrimination module, a discrimination result module, an appointment parameter module, a recommendation module, and a discrimination and update module;

[0152] The response module is used to respond to user search information and obtain the semantic association vector and contextual association vector of the user search.

[0153] The retrieval and discrimination module, based on the retrieval semantic association vector and contextual association vector, combined with the preset standard key entity mapping table and the preset three-level index mapping table built into the distributed isolation mapping security chain, performs standard question and answer retrieval information retrieval judgment on the hierarchical index tree-structured standard semantic question and answer information database configured for each main node of the distributed isolation mapping security chain;

[0154] The discrimination result module, based on the retrieval discrimination result, returns the retrieved standard question and answer retrieval information according to the frequency priority of the key entities appearing in the retrieval semantic association vector. When no retrieval is found, the module generates a question and answer retrieval information sequence and corresponding evaluation scores based on the user's retrieval semantic association vector, through a pre-trained reasoning generation model combined with a hierarchical index tree-structured standard semantic question and answer information database. The module then feeds back the information with the highest evaluation score among all the generated question and answer retrieval information to the corresponding user.

[0155] The reservation parameter module is used to obtain a cross-system reservation parameter set when the retrieved semantic association vector contains reservation requirements. It combines the distributed reservation decision tree corresponding to each subject node in the distributed isolation mapping security chain with standard question and answer retrieval information to perform cross-system reservation matching based on the user's reservation requirements.

[0156] The recommendation module, based on a cross-system reservation parameter set, obtains recommended reservation information by combining a preset cross-system bidirectional mapping dictionary with a matching recommendation algorithm;

[0157] The determination and update module is used to collect in real time the frequency of the corresponding user making a reservation based on the recommended reservation information and the frequency of the actual reservation, and obtain the reservation consistency frequency ratio. If the consistency frequency ratio is less than the preset reservation consistency threshold, the module constructs and updates the recommended mapping information in the bidirectional mapping dictionary and the retrieval index information in the three-level index mapping table based on the real-time reservation information parameters, until the reservation consistency threshold is met.

[0158] Example 3

[0159] An electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement an intelligent customer service information retrieval method based on a large model.

[0160] A computer-readable storage medium storing computer instructions that, when executed, perform an intelligent customer service information retrieval method based on a large model.

[0161] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the claims. All of these variations are within the protection scope of the present invention.

[0162] If the technical solution disclosed herein involves personal information, the product using this technical solution has clearly informed the user of the personal information processing rules and obtained the user's voluntary consent before processing the personal information. If the technical solution disclosed herein involves sensitive personal information, the product using this technical solution has obtained the user's separate consent before processing the sensitive personal information, and also meets the requirement of "express consent". For example, at personal information collection devices such as cameras, clear and prominent signs are set up to inform users that they have entered the scope of personal information collection and that personal information will be collected. If an individual voluntarily enters the collection scope, it is deemed that they have agreed to the collection of their personal information; or on the personal information processing device, with clear signs / information informing users of the personal information processing rules, authorization is obtained from the individual through pop-up information or by asking the individual to upload their personal information; wherein, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the types of personal information processed.

Claims

1. A method for intelligent customer service information retrieval based on a large model, characterized in that, include: In response to user search information, obtain the semantic association vector and contextual association vector of the user search; Based on the combination of retrieval semantic association vectors and contextual association vectors, a pre-set standard key entity mapping table and a pre-set three-level index mapping table built into the distributed isolation mapping security chain are used to perform standard question and answer retrieval information retrieval judgment on the hierarchical index tree-structured standard semantic question and answer information database configured for each main node of the distributed isolation mapping security chain. If found, the retrieved standard question and answer retrieval information will be returned according to the frequency priority of the key entities appearing in the retrieval semantic association vector. If no results are found, the system uses a pre-trained reasoning generation model combined with a hierarchical index tree-structured standard semantic question-and-answer information database to generate a sequence of question-and-answer retrieval information and corresponding evaluation scores based on the user's retrieval semantic association vector. The system then feeds back the information with the highest evaluation score among all generated question-and-answer retrieval information to the corresponding user.

2. The intelligent customer service information retrieval method based on a large model as described in claim 1, characterized in that, The method further includes: When the retrieved semantic association vector contains reservation requests, cross-system reservation matching is performed based on the user's reservation requests by combining the distributed reservation decision tree corresponding to each subject node in the distributed isolation mapping security chain with standard question and answer retrieval information to obtain the cross-system reservation parameter set. Based on the cross-system reservation parameter set, recommended reservation information is obtained by combining a pre-set cross-system bidirectional mapping dictionary with a matching recommendation algorithm; The frequency of reservations made by the corresponding user based on the recommended reservation information and the frequency of actual reservations are collected in real time, and the reservation consistency frequency ratio is obtained. If the consistency frequency ratio is less than the preset reservation consistency threshold, the recommended mapping information in the bidirectional mapping dictionary and the retrieval index information in the three-level index mapping table are constructed and updated according to the real-time reservation information parameters until the reservation consistency threshold is met.

3. The intelligent customer service information retrieval method based on a large model as described in claim 2, characterized in that, The process of constructing the standard key entity mapping table includes: We obtain historical text records and domain-standardized text information corresponding to all sub-sub ... Based on the text information space corresponding to all sub-sub ... Using each general tag key entity vector as the cluster center, the cluster distance between each cluster center and all exclusive tag key entity vectors is obtained through the cosine formula; Based on all cluster distances corresponding to each cluster center, the cutoff distance corresponding to each cluster center is determined. All cutoff distances and cluster distances corresponding to each cluster center are input into the DBSCAN algorithm to obtain a dedicated label clustering mapping table corresponding to each general label. The dedicated label clustering mapping table includes the general label and the dedicated label sequence under the corresponding different sub-sub-subjects.

4. The intelligent customer service information retrieval method based on a large model as described in claim 3, characterized in that, The process of constructing the standard key entity mapping table also includes: Based on the clustering mapping table of exclusive tags corresponding to all general tags and the domain standard text information space, the degree of causal relationship between different exclusive tags under different general tags is obtained through the causal relationship analysis model. Based on the degree of causal relationship between different exclusive tags under different general tags, a bidirectional mapping connection is constructed between key entities of different general tags and key entities of corresponding exclusive tags. Based on the bidirectional mapping connection and combined with the exclusive tag clustering mapping table corresponding to each general tag, a standard key entity mapping table is obtained. The system collects historical text records and revised domain-standardized text information for each newly added subject type and all its sub-subjects in real time. It repeats the process of constructing the exclusive tag clustering mapping table and bidirectional mapping connection for each general tag. The system then updates the standard key entity mapping table using an incremental update model combined with the information collected in real time.

5. The intelligent customer service information retrieval method based on a large model as described in claim 4, characterized in that, The construction process of the distributed isolation mapping security chain includes: A root node is constructed based on the domain information in the domain-standardized text information, and a domain classification tree is obtained by combining the domain-standardized text information with knowledge graphs and graph embedding algorithms, and the domain classification tree is embedded into the root node. Based on the main body, construct first-level nodes, construct second-level nodes for sub-main bodies, and construct third-level differential database nodes by combining the corresponding text information space, historical question and answer information and configured standard response information under each sub-main body with a tree-structured database. Based on the domain type corresponding to each subject, clustering algorithms are used to obtain clusters of similar subjects. Then, according to the domain type corresponding to the clusters of similar subjects, a first-level index connection is constructed between each root node and the first-level node in the corresponding cluster of similar subjects.

6. The intelligent customer service information retrieval method based on a large model as described in claim 5, characterized in that, The construction process of the distributed isolation mapping security chain also includes: Based on the cluster similarity between first-level nodes in similar subject clusters, cross-subject connections are constructed. Combining the information security protection level, subject tag encoding, and historical cross-subject retrieval access permission information of each subject, a hierarchical encrypted access permission hash mapping sequence is obtained based on hash algorithms and encrypted access control algorithms. The hierarchical private key access permission hash mapping sequence in the hierarchical encrypted access permission hash mapping sequence is then mapped to the corresponding cross-subject connections to obtain a cross-subject secure mapping connection sequence. Based on the information of all second-level nodes corresponding to each first-level node and the subdomain type information of the functional information corresponding to each second-level node, a second-level index connection is constructed. Through the functional information corresponding to all second-level nodes, clustering algorithms and causal analysis algorithms are used to obtain the clusters of similar sub-sub ... By utilizing the cluster similarity between sub-sub ...

7. The intelligent customer service information retrieval method based on a large model as described in claim 6, characterized in that, The construction process of the distributed isolation mapping security chain also includes: Based on the frequency of historical retrieval information for each type of key entity stored in the tertiary differential database corresponding to each secondary node, the evaluation score of retrieval feedback question and answer information corresponding to each key entity, and the auxiliary reservation consistency of retrieval feedback question and answer information, a three-level index connection is constructed using a hash algorithm. Based on the root node, first-level node, second-level node, and third-level differential database node, combined with first-level index connection, second-level index connection, and third-level index connection, and combined with graph neural network algorithm, a hierarchical index tree-structured standard semantic question-answering information database corresponding to each subject is obtained; Based on the hierarchical index tree-structured standard semantic question-answering information database corresponding to each subject, combined with cross-subject secure mapping connections and cross-subject causal permission connections, a distributed isolated mapping security chain is constructed through blockchain algorithms and a distributed framework. Based on the distributed isolation mapping security chain, the general label key entity of each subdomain type in the standard key entity mapping table is mapped to the cross-subject security mapping connection of the corresponding subdomain type. The bidirectional mapping connection and the exclusive label clustering mapping table are mapped to the cross-subject causal permission connection to perform cross-subject exclusive label association mapping, thereby obtaining a distributed isolation mapping security chain with exclusive labels.

8. The intelligent customer service information retrieval method based on a large model as described in claim 7, characterized in that, The process of recommending appointment information includes: Based on the standardized distributed isolation mapping security chain with exclusive labels, a distributed security reservation chain with the same structure is constructed. The three-level differential database node corresponding to each hierarchical index tree standard semantic question and answer information database under the distributed isolation mapping security chain is replaced with the three-level reservation information database node under the same sub-subject, which is constructed by the staff's job type, skill proficiency, professional level and work schedule information and the corresponding staff's current load information. The distributed reservation decision tree corresponding to each subject under the distributed security reservation chain is obtained. Based on the key entity semantic and type information, subject code, sub-subject name, sub-subject code, and staff function type, skill proficiency, professional level, and historical appointment recommendation consistency frequency of each staff member contained in the second-level nodes between the distributed appointment decision tree and the hierarchical index tree standard semantic question-answering information base corresponding to each subject, a bidirectional mapping dictionary is constructed through a hash algorithm. Then, using work scheduling information and the current load information of the corresponding staff member, a bidirectional balanced mapping dictionary is obtained through reinforcement learning and load balancing algorithms. Based on a bidirectional balanced mapping dictionary, a cross-system reservation auxiliary connection is constructed between the distributed reservation decision tree and the corresponding second-level nodes of the hierarchical index tree standard semantic question-answering information database under each subject. The cross-system reservation auxiliary connection is then used to perform many-to-many connections between the hierarchical index tree standard semantic question-answering information database corresponding to each subject under the distributed isolated mapping security chain and all second-level nodes corresponding to the distributed reservation decision tree of the same subject under the distributed security reservation chain, thus obtaining a dual retrieval-reservation security chain.

9. The intelligent customer service information retrieval method based on a large model as described in claim 8, characterized in that, The process of recommending appointment information also includes: Obtain user search information, use the intent parsing model to obtain the semantic association vector and contextual association vector of the user search, and determine whether there is any reservation request information; If it does not exist, the user's retrieval semantic association vector and contextual association vector are input into the standard key entity mapping table in the dual retrieval-reservation security chain to standardize the retrieval key entities. Based on the standardized retrieval semantic association vector and contextual association vector combined with the distributed isolation mapping security chain, information retrieval is performed, and the retrieved information is fed back. If there is a reservation requirement, the standardized retrieval semantic association vector and contextual association vector are synchronously input into the distributed isolated mapping security chain and the distributed secure reservation chain for synchronous retrieval and reservation. The retrieved information is fed back to the corresponding user. At the same time, the retrieved information is fed back to the distributed secure reservation chain through cross-system reservation auxiliary connection to adjust the reservation result. When the retrieved information requires reservation for different sub-subjects, cross-subject causal permission connection in the distributed secure reservation chain is used to perform cross-subject association reservation, and the association reservation result is fed back to the corresponding user through recommendation algorithm. Based on the information retrieval and reservation or associated reservation process corresponding to the dual retrieval-reservation security chain, the retrieval and reservation are synchronously simulated and trained across systems through a simulation algorithm until the evaluation score corresponding to the question-answer retrieval information and the reservation consistency frequency ratio simultaneously meet the corresponding thresholds, thus obtaining the trained dual retrieval-reservation security chain.

10. A large-model-based intelligent customer service information retrieval system, used to implement the large-model-based intelligent customer service information retrieval method according to any one of claims 1-9, characterized in that, include: Response module, retrieval and discrimination module, discrimination result module, reservation parameter module, recommendation module, and discrimination update module; The response module is used to respond to user search information and obtain the semantic association vector and contextual association vector of the user search. The retrieval and discrimination module, based on the retrieval semantic association vector and contextual association vector, combined with the preset standard key entity mapping table and the preset three-level index mapping table built into the distributed isolation mapping security chain, performs standard question and answer retrieval information retrieval judgment on the hierarchical index tree standard semantic question and answer information database configured for each main node of the distributed isolation mapping security chain; The discrimination result module, based on the retrieval discrimination result, returns the retrieved standard question and answer retrieval information according to the frequency priority of the key entities appearing in the retrieval semantic association vector. When no retrieval is found, the module generates a question and answer retrieval information sequence and corresponding evaluation scores based on the user's retrieval semantic association vector, through a pre-trained reasoning generation model combined with a hierarchical index tree-structured standard semantic question and answer information database. The module then feeds back the information with the highest evaluation score among all the generated question and answer retrieval information to the corresponding user. The reservation parameter module is used to obtain a cross-system reservation parameter set by combining the distributed reservation decision tree corresponding to each subject node in the distributed isolation mapping security chain with standard question and answer retrieval information when the retrieved semantic association vector contains reservation requirements. The recommendation module obtains recommended reservation information based on a cross-system reservation parameter set, using a preset cross-system bidirectional mapping dictionary combined with a matching recommendation algorithm. The discrimination and update module is used to collect in real time the frequency of the corresponding user making reservations based on the recommended reservation information and the frequency of actual reservations, and obtain the reservation consistency frequency ratio. If the consistency frequency ratio is less than the preset reservation consistency threshold, the module constructs and updates the recommended mapping information in the bidirectional mapping dictionary and the retrieval index information in the three-level index mapping table based on the real-time reservation information parameters, until the reservation consistency threshold is met.

Images