Campus service handling method and system based on AI large model

Abstract

The invention discloses a campus service handling method and system based on an AI large model, and relates to the technical field of artificial intelligence data processing. The method comprises the following steps: acquiring a campus service network search log information set, and preprocessing to obtain a new log set; term and document vectors are constructed, historical query recommendation scores are calculated, and query context groups are formed; constructing a word and text matrix in combination with the vocabulary, obtaining a semantic matrix after decomposition, and calculating the cohesion weight of the term and the candidate term; a final query is obtained based on weight and new query screening expansion, the campus service system outputs a result according to a user habit format, and semantic remarks are attached; the problem that the retrieval efficiency is low due to the fact that the query context is ignored and the semantic gap between the query and the document cannot be made up is solved, and the user satisfaction and the query result accuracy are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence data processing technology, specifically relating to a campus service processing method and system based on an AI large model. Background Technology

[0002] More and more universities are now demanding full coverage of "one-stop online services" and "unified data source," breaking down barriers to cross-departmental collaboration through digital tools, connecting with the construction of smart campus benchmarks, and promoting service upgrades from functional integration to experience optimization. Technically, campus informatization has moved from a "centralized portal" to an "intelligent hub." Relying on basic capabilities such as unified identity authentication and process engines, new features include "data asset governance" for standardized cleaning and tagging management of academic affairs, personnel, and financial data, and "scenario-based intelligent empowerment" to push customized services based on teacher and student behavior. Simultaneously, low-code platforms enable rapid iteration of micro-applications, and AI large models support semantic retrieval and intelligent business guidance.

[0003] The existing publication number: CN118467679A discloses a text big data query system and method, including obtaining the semantics of the text big data, determining the keywords of the text big data, performing type analysis on the keywords of the text big data, determining the type of text big data and the text big data query data, obtaining user feedback on the text big data query data based on the text big data query data, and performing query control on the text big data query data based on the user feedback, and intelligently controlling the text big data query data.

[0004] While the above methods improve the efficiency of querying large amounts of text data through intelligent control, they fail to bridge the semantic gap between the query and the document, resulting in poor retrieval results. For example, if a user needs to "repair the computer in the lab" on campus but does not use the standard category "Repair of Information Technology Equipment (Computer Category)," the system may return irrelevant service entries such as "Office Furniture Repair" or "Water and Electricity Repair." Such situations in campus service processing can lead to semantic ambiguity and reduced retrieval efficiency, which may seriously affect the normal operation of campus services and cause unnecessary economic losses. Summary of the Invention

[0005] The purpose of this invention is to solve the problem of low retrieval efficiency in campus service processing due to the inability to bridge the semantic gap between the query and the document by ignoring the query context, and to propose a campus service processing method and system based on an AI big model.

[0006] In a first aspect of this invention, a method for processing campus services based on an AI large-scale model is first proposed, the method comprising: Obtain a log information set from the campus service website search engine, and preprocess various types of information in the log information set to obtain a new log information set; the new log information set includes new query documents and a glossary. Term vectors and document vectors are constructed based on the new log information set. Recommendation scores for historical queries are calculated based on the term vectors and document vectors. Query context groups are constructed using the recommendation scores and historical query documents. A word matrix is ​​constructed based on the query context group and vocabulary, and the word matrix is ​​decomposed to obtain a semantic matrix. The cohesion weight is calculated for the semantic vector of each term in the semantic matrix and the semantic vector of each candidate term. By aggregating weights and filtering and expanding new query documents, the final expanded query is obtained. The campus service system outputs the final query results in a user-preferred format with semantic annotations.

[0007] Optionally, term vectors and document vectors are constructed based on the new log information set, including: Historical query documents and new query documents are judged based on the vocabulary. If a term belongs to the vocabulary, the value of the corresponding position in the vector is 1; if a term does not belong to the vocabulary, the value of the corresponding position in the vector is 0, thus obtaining the term vector of the historical query document and the term vector of the new query document. The historical query document is determined by the set of clicked document IDs in the historical query. If the clicked document is a clicked document in the historical query, the value of the corresponding position in the vector is 1. If the document is not a clicked document in the historical query, the value of the corresponding position in the vector is 0, thus obtaining the document vector of the historical query document. The clicked document refers to the document that the user actively clicks to open. The new query document is judged based on the set of document IDs in the test set. The top N documents after the initial retrieval of the new query document are taken and defined as the retrieval document. If the document belongs to the retrieval document, the corresponding position value is 1; if the document does not belong to the retrieval document, the corresponding position value is 0, thus obtaining the document vector of the new query document.

[0008] Optionally, a query context group is constructed using the recommendation score and the new query document, including: For new and historical query documents, the frequency of occurrence of terms in the query is statistically analyzed, and the probability scores of the terms in the language models of historical and new query documents are calculated; the probability scores include the probability in the language model of historical query documents and the probability in the language model of new query documents. The similarity of terms is calculated based on probability scores, and then normalized to output a recommendation score for each historical query document relative to the new query document.

[0009] Optionally, a lexical matrix is ​​constructed based on the query context group and the vocabulary, including: Sort all historical query documents in descending order of recommendation score to obtain the mapping relationship between historical query documents and recommendation scores. Based on the mapping relationship, determine the number of optimal recommended query documents and select the top K historical query documents as recommended query documents. The query context group is obtained by collecting a combination of new query documents, retrieved documents, and recommended query documents; Define the number of times all terms appear in the query context group as the row dimension, and define the number of documents contained in the query context group plus 1 as the column dimension. Construct an empty matrix based on the row dimension and the column dimension. The weight of each term in the real document is calculated using TF-IDF as the first weight. The new query document is treated as a virtual document, and the weight of each term in the new query document is calculated as the second weight. The real document is a combination of the retrieved document and the recommended query document after deduplication. Fill the corresponding positions of the empty matrix with the first and second weights to obtain the word matrix.

[0010] Optionally, the final expanded query is obtained by filtering and expanding the aggregated weights and new query text, including: The terminology matrix is ​​decomposed using linear algebra to obtain the term semantic matrix, singular value diagonal matrix, and document semantic matrix. The term semantic matrix, singular value diagonal matrix, and document semantic matrix are then truncated and dimensionality reduced to obtain the dimensionality-reduced terminology matrix. Extract the semantic vectors of each term in the new query document from the semantic matrix as the first semantic vector, and extract the semantic vectors of all candidate terms in the query context group as the second semantic vector; The cohesion weights between candidate terms and the new query document are obtained by operating on the first semantic vector and the second semantic vector. All candidate terms are sorted in descending order according to the cohesion weights, and terms that are duplicates of terms in the new query document are excluded. Based on a preset number M of extended terms, the first M terms are selected from a multi-source text set as extended terms; the multi-source text set includes the text of the new query document, the text of the retrieved document, the text of the recommended query document, and the text of the clicked document of the recommended query document; The extended terms are added to the text of the new query document to form the final extended query.

[0011] In a second aspect of this invention, a campus service processing system based on an AI large-scale model is proposed, comprising a preprocessing module, a construction module, a matrix generation module, and an extension module, wherein: The preprocessing module is used to obtain a log information set from the campus service website search engine, and preprocess various types of information in the log information set to obtain a new log information set; the new log information set includes new query documents and a glossary. The construction module is used to construct term vectors and document vectors based on the new log information set, calculate the recommendation score of historical queries based on the term vectors and document vectors, and construct a query context group by combining the recommendation score with the historical query documents. The weight calculation module is used to construct a word matrix based on the query context group and vocabulary, decompose the word matrix to obtain a semantic matrix, and calculate the cohesive weight for the semantic vector of each term and the semantic vector of each candidate term in the semantic matrix. The extension module is used to filter and expand the final extended query by aggregating weights and new query documents. The campus service system outputs the final query results in a user-preferred format with semantic annotations.

[0012] Optionally, the preprocessing module includes a term vector module, a first document vector module, and a second document vector module, wherein: The term vector module is used to determine the terms in historical and new query documents based on the vocabulary. If a term belongs to the vocabulary, the value of the corresponding position in the vector is 1; if a term does not belong to the vocabulary, the value of the corresponding position in the vector is 0, thus obtaining the term vectors of historical and new query documents. The first document vector module is used to determine the historical query document based on the set of clicked document IDs of the historical query. If the clicked document is a clicked document of the historical query document, the value of the corresponding position of the vector is 1; if the document is not a clicked document of the historical query document, the value of the corresponding position of the vector is 0, thus obtaining the document vector of the historical query document. The clicked document refers to the document that the user actively clicks to open. The second document vector module is used to judge the new query document based on the test set document ID set as the dimension, take the top N documents after the initial retrieval of the new query document, and define them as the retrieval documents. If the document belongs to the retrieval documents, the corresponding position value is 1, and if the document does not belong to the retrieval documents, the corresponding position value is 0, thus obtaining the document vector of the new query document.

[0013] Optionally, the construction module includes a probability score calculation module and a recommendation score calculation module, wherein: The probability score calculation module is used to statistically analyze the frequency of occurrence of terms in queries for new and historical query documents, and to calculate the probability score of the terms in the language models of historical and new query documents; the probability score includes the probability in the language model of historical query documents and the probability in the language model of new query documents. The recommendation score calculation module is used to calculate the similarity of terms based on probability scores, and then normalize and output the recommendation score of each historical query document relative to the new query document.

[0014] Optionally, the matrix generation module includes a mapping module, a combination module, a dimension module, and a weight calculation and filling module, wherein: The mapping module is used to sort all historical query documents in descending order of recommendation score to obtain the mapping relationship between historical query documents and recommendation scores, determine the number of optimal recommended query documents through the mapping relationship, and select the top K historical query documents as recommended query documents. The combination module is used to obtain a query context group by collecting a combination of new query documents, retrieved documents, and recommended query documents; The dimension module is used to define the number of times all terms appear in the query context group as the row dimension, and the number of documents contained in the query context group plus 1 as the column dimension, and to construct an empty matrix based on the row dimension and the column dimension. The weight calculation is used to calculate the weight of each term in the real document using TF-IDF as the first weight, treat the new query document as a virtual document, and calculate the weight of each term in the new query document as the second weight; the real document is a combination of the retrieved document and the recommended query document after deduplication. The filling module is used to fill the first weight and the second weight into the corresponding positions of the empty matrix to obtain the word matrix.

[0015] Optionally, the extension module includes a dimensionality reduction module, a semantic vector module, an agglomerative weighting module, a filtering module, and an adding module, wherein: The dimensionality reduction module is used to decompose the word matrix according to linear algebra to obtain the term semantic matrix, the singular value diagonal matrix, and the document semantic matrix, and to truncate and reduce the term semantic matrix, the singular value diagonal matrix, and the document semantic matrix to obtain the dimensionality-reduced word matrix. The semantic vector module is used to extract the semantic vectors of each term in the new query document from the term semantic matrix as the first semantic vector, and to extract the semantic vectors of all candidate terms in the query context group as the second semantic vector. The cohesion weight module is used to calculate the cohesion weight between candidate terms and the new query document based on the first semantic vector and the second semantic vector, sort all candidate terms in descending order according to the cohesion weight, and exclude terms that are duplicates of terms in the new query document. The filtering module is used to filter out the top M terms as extended terms from a multi-source text set according to a preset number M of extended terms; the multi-source text set includes the text of the new query document, the text of the retrieved document, the text of the recommended query document, and the text of the clicked document of the recommended query document; The adding module is used to add extended terms to the text of a new query document to form the final extended query.

[0016] The beneficial effects of this invention are: This invention proposes a campus service processing method based on an AI large model. By preprocessing log information sets and decomposing semantic matrices within the campus service processing system, it achieves intelligent query expansion and contextual recommendation. This addresses the problems of insufficient keywords in short queries and ambiguity in long queries, as well as the poor retrieval results caused by users ignoring the semantic gap between queries and documents in existing keyword matching systems. Furthermore, by calculating recommendation scores based on both terminology and document vectors, it effectively improves search relevance, user satisfaction, and the accuracy of query results, thereby enhancing the overall efficiency and user experience of the campus service processing system. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 A flowchart illustrating a campus service processing method based on an AI large model, provided as an embodiment of the present invention; Figure 2 This is a framework diagram of a campus service processing system based on an AI large model, provided as an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The term "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and B can represent: A alone, A and B simultaneously, and B alone. Furthermore, descriptions involving "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" can explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0020] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This invention provides a method for processing campus services based on an AI-powered large-scale model. See also... Figure 1 , Figure 1 This is a flowchart illustrating a campus service processing method based on an AI large-scale model, provided as an embodiment of the present invention. The method includes the following steps: S101. Obtain the log information set in the search engine of the campus service official website, and preprocess various types of information in the log information set to obtain a new log information set; S102. Construct term vectors and document vectors based on the new log information set, calculate the recommendation scores of historical queries according to the term vectors and document vectors, and construct a query context group through the recommendation scores and historical query documents; S103. Construct a word-document matrix based on the query context group and the vocabulary, decompose the word-document matrix to obtain a semantic matrix, and calculate the cohesion weights for the semantic vectors of each term and the semantic vectors of each candidate term in the semantic matrix; S104. Screen and expand through the cohesion weights and the new query document to obtain the final expanded query, and the campus service system outputs the final query result in the user's habitual format and attaches semantic remarks.

[0022] The new log information set includes new query documents and a vocabulary.

[0023] Based on a campus service handling method based on an AI large model provided by an embodiment of the present invention, through the campus service handling system, through log information set preprocessing and semantic matrix decomposition, intelligent expansion and context recommendation of queries are realized, solving the problems of lack of short query keywords, ambiguity in long queries, and low retrieval efficiency of existing keyword matching systems due to ignoring query context and being unable to bridge the semantic gap between queries and documents; effectively improving search relevance, user satisfaction, and the accuracy of query results, thereby improving the overall efficiency and user experience of the campus service handling system.

[0024] In one implementation, the log information set includes historical query documents and the user click document IDs corresponding to each historical query, and a global term source (the set of terms in all historical query documents).

[0025] In one implementation, preprocessing various types of information in the log information set includes: Eliminating invalid data and garbled text from the log information set; the invalid data includes empty queries and duplicate queries; Organizing the query logs in the log information set into the key-value pair format of historical query document → click document ID set; Performing word segmentation, stop word removal (such as "de, di, de, zhi, suo"), and stemming (such as "computer" → "computer", "AI" → "artificial intelligence") on all terms, and summarizing all unique terms after standardization to provide a unified term set for subsequent vector representation and probability calculation.

[0026] In one embodiment, constructing term vectors and text vectors based on the new log information set includes: Historical query documents and new query documents are judged based on the vocabulary. If a term belongs to the vocabulary, the value of the corresponding position in the vector is 1; if a term does not belong to the vocabulary, the value of the corresponding position in the vector is 0, thus obtaining the term vector of the historical query document and the term vector of the new query document. The historical query document is determined by the set of clicked document IDs in the historical query. If the clicked document is a clicked document in the historical query, the value of the corresponding position in the vector is 1. If the document is not a clicked document in the historical query, the value of the corresponding position in the vector is 0, thus obtaining the document vector of the historical query document. The clicked document refers to the document that the user actively clicks to open. The new query document is judged based on the set of document IDs in the test set. The top N documents after the initial retrieval of the new query document are taken and defined as the retrieval document. If the document belongs to the retrieval document, the corresponding position value is 1; if the document does not belong to the retrieval document, the corresponding position value is 0, thus obtaining the document vector of the new query document.

[0027] In one implementation, the relevance score between each document and the new query document is calculated using the BM25 algorithm, and the top N documents are sorted from high to low according to the score as the top N documents. N is set to 10 in the experiment. By constructing term vectors and document vectors, we provide dual-dimensional feature support for subsequent query recommendations, namely "term semantics + document association." Term vectors accurately capture the literal term composition of the query, providing a foundation for calculating "term-level similarity" and avoiding semantic ambiguity caused by relying solely on text matching. Document vectors, on the other hand, capture the document information actually associated with the query through user clicks on documents in historical queries (real interaction feedback) or document searches (pseudo-relevance feedback), supplementing the basis for calculating "document-level similarity."

[0028] In one implementation, the query representation goes beyond just terms, better aligning with the document association needs in actual retrieval scenarios. Both serve as inputs to the recommendation score, ensuring that the recommended historical queries match both the semantics of the terms and the document association preferences of the queries.

[0029] In one embodiment, a query context group is constructed using recommendation scores and new query documents, including: For new and historical query documents, the frequency of occurrence of terms in the query is statistically analyzed, and the probability scores of terms in the language models of historical and new query documents are calculated. The probability scores include the probability in the language model of historical query documents and the probability in the language model of new query documents. The similarity of terms is calculated based on probability scores, and then normalized to output a recommendation score for each historical query document relative to the new query document.

[0030] In one implementation, the calculation process for probability score, similarity score, and recommendation score is as follows:

[0031] in, This indicates the recommendation score for new and historical query documents. Let f(t) represent the generation probability of term t in the language model of query document Q, and let f(t) represent the frequency of occurrence of term t in query Q. For new query documents and historical document search The similarity, where V represents the vocabulary. Statistics on the probability of terms being generated in new query documents. Statistics on the probability of terms being generated in historical query documents; The normalization parameter is used to balance the weights of the term vector and the document vector. It is a similarity score calculated based on term vectors. It is a similarity score calculated based on document vectors.

[0032] In one implementation, the purpose of calculating the recommendation score is to quantify the similarity between historical query documents and new query documents in two dimensions: "term semantics + document association", so as to select the most relevant recommended queries, construct high-quality query context, and provide materials for subsequent query expansion. Compared with existing technologies, by integrating term-vector-based similarity and document-vector-based similarity, semantic bias caused by single-dimensional evaluation is avoided, and the true relationship between queries is captured more accurately, thereby significantly improving the effectiveness of information retrieval systems.

[0033] In one embodiment, a lexical matrix is ​​constructed based on the query context group and the vocabulary, including: Sort all historical query documents in descending order of recommendation score to obtain the mapping relationship between historical query documents and recommendation scores. Based on the mapping relationship, determine the number of optimal recommended query documents and select the top K historical query documents as recommended query documents. The query context group is obtained by collecting a combination of new query documents, retrieved documents, and recommended query documents; Define the number of times all terms appear in the query context group as the row dimension, and define the number of documents contained in the query context group plus 1 as the column dimension. Construct an empty matrix based on the row dimension and the column dimension. The weight of each term in the real document is calculated using TF-IDF as the first weight. The new query document is treated as a virtual document, and the weight of each term in the new query document is calculated as the second weight. The real document is a combination of the retrieved document and the recommended query document after deduplication. Fill the corresponding positions of the empty matrix with the first and second weights to obtain the word matrix.

[0034] In one implementation, the calculation process for the first and second weights is as follows:

[0035] Where TF−TF(t,D) represents the first weight of term t in real document D, and TF(t,D) is the frequency percentage of term t in real document D; IDF(t) represents the scarcity of term t in real documents of the query context group, which is used to measure the global distinguishability of term t; S is the number of real documents, and df(t,S) is the number of real documents S containing term t. The second weight is calculated in the same way as the first weight.

[0036] In one implementation, a query context group is constructed by filtering recommended queries. The purpose is to supplement the new query document with a semantic background composed of similar historical queries and related documents, so as to solve the problem of incomplete or ambiguous semantics of the new query document. The purpose of constructing the word matrix is ​​to quantify the association between the terms in the query context and the text (real document + virtual document), and to provide mathematical input for subsequent extraction of latent semantics and filtering of extended terms. Compared with existing technologies, it avoids the lack of context caused by relying on a single document or isolated query, and achieves accurate semantic association between terms and multi-source text, thereby improving the effectiveness of information retrieval systems.

[0037] In one embodiment, the final expanded query is obtained by filtering and expanding the aggregated weights and new query text, including: The terminology matrix is ​​decomposed using linear algebra to obtain the term semantic matrix, singular value diagonal matrix, and document semantic matrix. The term semantic matrix, singular value diagonal matrix, and document semantic matrix are then truncated and dimensionality reduced to obtain the dimensionality-reduced terminology matrix. Extract the semantic vectors of each term in the new query document from the semantic matrix as the first semantic vector, and extract the semantic vectors of all candidate terms in the query context group as the second semantic vector; The cohesion weights between candidate terms and the new query document are obtained by operating on the first semantic vector and the second semantic vector. All candidate terms are sorted in descending order according to the cohesion weights, and terms that are duplicates of terms in the new query document are excluded. Based on the preset number of extended terms M, the top M terms are selected from the multi-source text set as extended terms; the multi-source text set includes the text of the new query document, the text of the retrieved document, the text of the recommended query document, and the text of the clicked document of the recommended query document; The extended terms are added to the text of the new query document to form the final extended query.

[0038] In one implementation, the cohesive weights between candidate terms and the new query document are obtained by operating on the first semantic vector and the second semantic vector. The process is as follows:

[0039] in, This indicates the similarity between the first semantic vector and the second semantic vector. This indicates the cohesion weight between candidate terms and the new query document, reflecting the strength of their semantic association. This represents the semantic vector of the new queried document. The semantic vector of the candidate term; The L2 norm of a vector is used to normalize the dot product result and avoid the interference of vector length on similarity.

[0040] In one implementation, extended terms are selected by calculating term similarity and cohesion weight to form extended queries. The purpose is to supplement new query documents with semantically relevant and noise-free terms, solve the problems of lack of keywords in short queries and semantic ambiguity in long queries, and bridge the semantic gap between queries and documents to improve retrieval accuracy. Compared with existing technologies, this technology avoids semantic bias caused by relying solely on a single document or isolated terms for filtering. By combining multi-source query context and dual-dimensional weight calculation, the expanded terms are more closely aligned with the actual needs of new query documents, significantly improving the effectiveness of the information retrieval system.

[0041] Based on the same inventive concept, this invention also provides a campus service processing system based on an AI large-scale model. See also Figure 2 , Figure 2 This invention provides a framework diagram of a campus service processing system based on an AI large model. The system includes a preprocessing module, a construction module, a matrix generation module, and an extension module, wherein: The preprocessing module is used to obtain the log information set from the campus service website search engine, and preprocess various types of information in the log information set to obtain a new log information set; the new log information set includes new query documents and a glossary. The building module is used to construct term vectors and document vectors based on the new log information set, calculate the recommendation score of historical queries based on the term vectors and document vectors, and construct a query context group by combining the recommendation score with the historical query documents; The weight calculation module is used to construct a word matrix based on the query context group and vocabulary, decompose the word matrix to obtain a semantic matrix, and calculate the cohesive weight of the semantic vector of each term and the semantic vector of each candidate term in the semantic matrix. The extension module is used to filter and expand the final extended query by aggregating weights and new query documents. The campus service system outputs the final query results in a user-preferred format with semantic comments.

[0042] This invention provides a campus service processing system based on an AI-powered large model. By preprocessing log information sets and decomposing semantic matrices, the system achieves intelligent expansion and contextual recommendation of queries. This addresses the problems of insufficient keywords in short queries, ambiguity in long queries, and low retrieval efficiency caused by traditional keyword matching systems ignoring query context and failing to bridge the semantic gap between queries and documents. The system improves search relevance, user satisfaction, and the accuracy of query results, thereby enhancing the overall efficiency and user experience of the campus service processing system.

[0043] In one embodiment, the preprocessing module includes a term vector module, a first document vector module, and a second document vector module, wherein: The term vector module is used to determine the terms in historical and new query documents based on the vocabulary. If a term belongs to the vocabulary, the corresponding position of the vector is 1; if a term does not belong to the vocabulary, the corresponding position of the vector is 0, thus obtaining the term vectors for historical and new query documents. The first document vector module is used to determine the historical query document based on the set of clicked document IDs of the historical query. If the clicked document is a clicked document of the historical query document, the value of the corresponding position of the vector is 1; if the document is not a clicked document of the historical query document, the value of the corresponding position of the vector is 0, thus obtaining the document vector of the historical query document. A clicked document refers to a document that the user actively clicks to open. The second document vector module is used to determine the new query document based on the test set document ID set as the dimension. It takes the top N documents after the initial retrieval of the new query document and defines them as the retrieval documents. If the document belongs to the retrieval documents, the corresponding position value is 1; if the document does not belong to the retrieval documents, the corresponding position value is 0, thus obtaining the document vector of the new query document.

[0044] In one embodiment, the building module includes a probability score calculation module and a recommendation score calculation module, wherein: The probability score calculation module is used to statistically analyze the frequency of terms in queries for new and historical query documents, and calculate the probability score of each term in the language model of both historical and new query documents. The probability score includes the probability in the language model of historical query documents and the probability in the language model of new query documents. The recommendation score calculation module is used to calculate the similarity of terms based on probability scores, and then normalize and output the recommendation score of each historical query document relative to the new query document.

[0045] In one embodiment, the matrix generation module includes a mapping module, a combination module, a dimension module, and a weight calculation and filling module, wherein: The mapping module is used to sort all historical query documents in descending order of recommendation score to obtain the mapping relationship between historical query documents and recommendation scores. The mapping relationship is used to determine the number of optimal recommended query documents and filter out the top K historical query documents as recommended query documents. The combination module is used to obtain a query context group by combining new query documents, retrieved documents, and recommended query documents; The Dimension module is used to define the number of times all terms appear in the query context group as the row dimension, and the number of documents contained in the query context group plus 1 as the column dimension. An empty matrix is ​​constructed based on the row dimension and the column dimension. Weight calculation is used to calculate the weight of each term in the real document using TF-IDF as the first weight. The new query document is treated as a virtual document, and the weight of each term in the new query document is calculated as the second weight. The real document is a combination of the retrieved document and the recommended query document after deduplication. The filling module is used to fill the first and second weights into the corresponding positions of the empty matrix to obtain the word matrix.

[0046] In one embodiment, the extension module includes a dimensionality reduction module, a semantic vector module, an agglomerative weight module, a filtering module, and an adding module, wherein: The dimensionality reduction module is used to decompose the word matrix according to linear algebra to obtain the term semantic matrix, singular value diagonal matrix, and document semantic matrix, and to truncate and reduce the dimensionality of the term semantic matrix, singular value diagonal matrix, and document semantic matrix to obtain the dimensionality-reduced word matrix. The semantic vector module is used to extract the semantic vectors of each term in the new query document from the term semantic matrix as the first semantic vector, and to extract the semantic vectors of all candidate terms in the query context group as the second semantic vector. The cohesion weight module is used to calculate the cohesion weight between candidate terms and the new query document based on the first semantic vector and the second semantic vector, sort all candidate terms in descending order according to the cohesion weight, and exclude terms that are duplicates of terms in the new query document. The filtering module is used to select the top M terms as extended terms from a multi-source text set based on a preset number M of extended terms; the multi-source text set includes the text of the new query document, the retrieved document, the text of the recommended query document, and the text of the clicked document of the recommended query document; Add a module to add extended terms to the text of a new query document to form the final extended query.

[0047] The foregoing has described one embodiment of the present invention in detail, but this content is merely a preferred embodiment and should not be considered as limiting the scope of the present invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the scope of the claims of this invention.

Claims

1. A campus service processing method based on an AI large-scale model, characterized in that, The method includes: Obtain a log information set from the campus service website search engine, and preprocess various types of information in the log information set to obtain a new log information set; the new log information set includes new query documents and a glossary. Term vectors and document vectors are constructed based on the new log information set. Recommendation scores for historical queries are calculated based on the term vectors and document vectors. Query context groups are constructed using the recommendation scores and historical query documents. A word matrix is ​​constructed based on the query context group and vocabulary, and the word matrix is ​​decomposed to obtain a semantic matrix. The cohesion weight is calculated for the semantic vector of each term in the semantic matrix and the semantic vector of each candidate term. By aggregating weights and filtering and expanding new query documents, the final expanded query is obtained. The campus service system outputs the final query results in a user-preferred format with semantic annotations.

2. The campus service processing method based on an AI large model according to claim 1, characterized in that, Term vectors and document vectors are constructed based on the new log information set, including: Historical query documents and new query documents are judged based on the vocabulary. If a term belongs to the vocabulary, the value of the corresponding position in the vector is 1; if a term does not belong to the vocabulary, the value of the corresponding position in the vector is 0, thus obtaining the term vector of the historical query document and the term vector of the new query document. The historical query document is determined by the set of clicked document IDs in the historical query. If the clicked document is a clicked document in the historical query, the value of the corresponding position in the vector is 1. If the document is not a clicked document in the historical query, the value of the corresponding position in the vector is 0, thus obtaining the document vector of the historical query document. The clicked document refers to the document that the user actively clicks to open. The new query document is judged based on the set of document IDs in the test set. The top N documents after the initial retrieval of the new query document are taken and defined as the retrieval document. If the document belongs to the retrieval document, the corresponding position value is 1; if the document does not belong to the retrieval document, the corresponding position value is 0, thus obtaining the document vector of the new query document.

3. The campus service processing method based on an AI large model according to claim 1, characterized in that, Recommendation scores for historical queries are calculated based on term vectors and document vectors, including: For new and historical query documents, the frequency of occurrence of terms in the query is statistically analyzed, and the probability scores of the terms in the language models of historical and new query documents are calculated; the probability scores include the probability in the language model of historical query documents and the probability in the language model of new query documents. The similarity of terms is calculated based on probability scores, and then normalized to output a recommendation score for each historical query document relative to the new query document.

4. The campus service processing method based on an AI large model according to claim 2, characterized in that, A word matrix is ​​constructed based on the query context group and vocabulary, including: Sort all historical query documents in descending order of recommendation score to obtain the mapping relationship between historical query documents and recommendation scores. Based on the mapping relationship, determine the number of optimal recommended query documents and select the top K historical query documents as recommended query documents. The query context group is obtained by collecting a combination of new query documents, retrieved documents, and recommended query documents; Define the number of times all terms appear in the query context group as the row dimension, and define the number of documents contained in the query context group plus 1 as the column dimension. Construct an empty matrix based on the row dimension and the column dimension. The weight of each term in the real document is calculated using TF-IDF as the first weight. The new query document is treated as a virtual document, and the weight of each term in the new query document is calculated as the second weight. The real document is a combination of the retrieved document and the recommended query document after deduplication. Fill the corresponding positions of the empty matrix with the first and second weights to obtain the word matrix.

5. The campus service processing method based on an AI large model according to claim 4, characterized in that, The final expanded query is obtained by filtering and expanding the aggregated weights and new query text, including: The terminology matrix is ​​decomposed using linear algebra to obtain the term semantic matrix, singular value diagonal matrix, and document semantic matrix. The term semantic matrix, singular value diagonal matrix, and document semantic matrix are then truncated and dimensionality reduced to obtain the dimensionality-reduced terminology matrix. Extract the semantic vectors of each term in the new query document from the semantic matrix as the first semantic vector, and extract the semantic vectors of all candidate terms in the query context group as the second semantic vector; The cohesion weights between candidate terms and the new query document are obtained by operating on the first semantic vector and the second semantic vector. All candidate terms are sorted in descending order according to the cohesion weights, and terms that are duplicates of terms in the new query document are excluded. Based on a preset number M of extended terms, the first M terms are selected from a multi-source text set as extended terms; the multi-source text set includes the text of the new query document, the text of the retrieved document, the text of the recommended query document, and the text of the clicked document of the recommended query document; The extended terms are added to the text of the new query document to form the final extended query.

6. A campus service processing system based on an AI large-scale model, characterized in that, The system includes a preprocessing module, a construction module, a matrix generation module, and an extension module, wherein: The preprocessing module is used to obtain a log information set from the campus service website search engine, and preprocess various types of information in the log information set to obtain a new log information set; the new log information set includes new query documents and a glossary. The construction module is used to construct term vectors and document vectors based on the new log information set, calculate the recommendation score of historical queries based on the term vectors and document vectors, and construct a query context group by combining the recommendation score with the historical query documents. The weight calculation module is used to construct a word matrix based on the query context group and vocabulary, decompose the word matrix to obtain a semantic matrix, and calculate the cohesive weight for the semantic vector of each term and the semantic vector of each candidate term in the semantic matrix. The extension module is used to filter and expand the final extended query by aggregating weights and new query documents. The campus service system outputs the final query results in a user-preferred format with semantic annotations.

7. A campus service processing system based on an AI large model according to claim 6, characterized in that, The preprocessing module includes a term vector module, a first document vector module, and a second document vector module, wherein: The term vector module is used to determine the terms in historical and new query documents based on the vocabulary. If a term belongs to the vocabulary, the value of the corresponding position in the vector is 1; if a term does not belong to the vocabulary, the value of the corresponding position in the vector is 0, thus obtaining the term vectors of historical and new query documents. The first document vector module is used to determine the historical query document based on the set of clicked document IDs of the historical query. If the clicked document is a clicked document of the historical query document, the value of the corresponding position of the vector is 1; if the document is not a clicked document of the historical query document, the value of the corresponding position of the vector is 0, thus obtaining the document vector of the historical query document. The clicked document refers to the document that the user actively clicks to open. The second document vector module is used to judge the new query document based on the test set document ID set as the dimension, take the top N documents after the initial retrieval of the new query document, and define them as the retrieval documents. If the document belongs to the retrieval documents, the corresponding position value is 1, and if the document does not belong to the retrieval documents, the corresponding position value is 0, thus obtaining the document vector of the new query document.

8. A campus service processing system based on an AI large model according to claim 6, characterized in that, The construction module includes a probability score calculation module and a recommendation score calculation module, wherein: The probability score calculation module is used to statistically analyze the frequency of occurrence of terms in queries for new and historical query documents, and to calculate the probability score of the terms in the language models of historical and new query documents; the probability score includes the probability in the language model of historical query documents and the probability in the language model of new query documents. The recommendation score calculation module is used to calculate the similarity of terms based on probability scores, and then normalize and output the recommendation score of each historical query document relative to the new query document.

9. A campus service processing system based on an AI large model according to claim 7, characterized in that, The matrix generation module includes a mapping module, a combination module, a dimension module, and a weight calculation and filling module, wherein: The mapping module is used to sort all historical query documents in descending order of recommendation score to obtain the mapping relationship between historical query documents and recommendation scores, determine the number of optimal recommended query documents through the mapping relationship, and select the top K historical query documents as recommended query documents. The combination module is used to obtain a query context group by collecting a combination of new query documents, retrieved documents, and recommended query documents; The dimension module is used to define the number of times all terms appear in the query context group as the row dimension, and the number of documents contained in the query context group plus 1 as the column dimension, and to construct an empty matrix based on the row dimension and the column dimension. The weight calculation is used to calculate the weight of each term in the real document using TF-IDF as the first weight, treat the new query document as a virtual document, and calculate the weight of each term in the new query document as the second weight; the real document is a combination of the retrieved document and the recommended query document after deduplication. The filling module is used to fill the first weight and the second weight into the corresponding positions of the empty matrix to obtain the word matrix.

10. A campus service processing system based on an AI large model according to claim 9, characterized in that, The extended module includes a dimensionality reduction module, a semantic vector module, an aggregated weighting module, a filtering module, and an adding module, wherein: The dimensionality reduction module is used to decompose the word matrix according to linear algebra to obtain the term semantic matrix, the singular value diagonal matrix, and the document semantic matrix, and to truncate and reduce the term semantic matrix, the singular value diagonal matrix, and the document semantic matrix to obtain the dimensionality-reduced word matrix. The semantic vector module is used to extract the semantic vectors of each term in the new query document from the term semantic matrix as the first semantic vector, and to extract the semantic vectors of all candidate terms in the query context group as the second semantic vector. The cohesion weight module is used to calculate the cohesion weight between candidate terms and the new query document based on the first semantic vector and the second semantic vector, sort all candidate terms in descending order according to the cohesion weight, and exclude terms that are duplicates of terms in the new query document. The filtering module is used to filter out the top M terms as extended terms from a multi-source text set according to a preset number M of extended terms; the multi-source text set includes the text of the new query document, the text of the retrieved document, the text of the recommended query document, and the text of the clicked document of the recommended query document; The adding module is used to add extended terms to the text of a new query document to form the final extended query.

Images