A method for realizing operation and maintenance knowledge search based on embedding vector

Through the method of embedding vectors, the problem of insufficient semantic analysis capabilities in the existing technology is solved, and more efficient and accurate operation and maintenance knowledge retrieval is achieved, which improves operation and maintenance efficiency and effect.

CN118428481BActive Publication Date: 2024-11-19QINGDAO HISENSE INFORMATION TECH CO LTD
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202410895576.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-05
Publication Date
2024-11-19
Estimated Expiration
2044-07-05

AI Technical Summary

Technical Problem

The existing operation and maintenance knowledge retrieval technology has insufficient semantic analysis capabilities and retrieval accuracy, resulting in low retrieval efficiency and accuracy, which cannot meet the increasing retrieval needs.

Method used

Using the method based on embedded vectors, the initial text data is processed through preset preprocessing methods, the data is mapped into vector representations using the preset language model, and similar analysis is performed in combination with the preset algorithm to improve semantic understanding ability and retrieval accuracy, and use multiple methods to The parallel analysis results of these models are sorted to improve the accuracy of the search results.

Benefits of technology

It improves the accuracy and search efficiency of operation and maintenance knowledge retrieval, and can accurately search related content based on semantic similarity without fully matching keywords, which enhances operation and maintenance efficiency and effect.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118428481B_ABST
    Figure CN118428481B_ABST
Patent Text Reader

Abstract

The present invention provides a method for realizing operation and maintenance knowledge search based on embedded vectors, which relates to the field of text retrieval technology, and includes: step 1: obtaining initial text data, and preprocessing the initial text data using a preset preprocessing method to obtain first data; step 2: combining a preset vector database and mapping the first data using a preset language model to obtain a corresponding first vector; step 3: combining a preset knowledge base with a preset algorithm to perform similarity analysis on the first vector, and outputting a vector analysis result; step 4: using a preset sorting method to sort the documents corresponding to the vector analysis result in the preset knowledge base, and selecting to obtain a retrieval result based on a preset selection instruction. The present invention can improve the accuracy of the operation and maintenance knowledge retrieval results, and improve the retrieval efficiency, and ensure the subsequent operation and maintenance effect and efficiency.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of text retrieval technology, and in particular to a method for realizing operation and maintenance knowledge search based on embedded vectors. Background Art

[0002] Operation and maintenance knowledge search is of great significance in the modern information technology environment. It can not only quickly retrieve relevant solutions and save time for troubleshooting and solving operation and maintenance problems, but also automate some operation and maintenance tasks, thereby improving operation and maintenance efficiency. Therefore, the accuracy of operation and maintenance knowledge retrieval will directly affect the efficiency and effectiveness of operation and maintenance.

[0003] At present, operation and maintenance knowledge retrieval is usually achieved through text retrieval. Most text retrieval technologies on the market are based on the Elasticsearch (ES) algorithm. Although they perform well in indexing and querying structured or semi-structured data, they have disadvantages in some specific functions and scenarios. For example, they only use traditional retrieval methods such as keyword matching and lack the ability to conduct deep semantic analysis of operation and maintenance query statements. Therefore, their retrieval accuracy and efficiency are low, and they are gradually unable to meet the increasing retrieval needs.

[0004] Therefore, the present invention provides a method for realizing operation and maintenance knowledge search based on embedded vectors. Summary of the invention

[0005] The present invention provides a method for realizing operation and maintenance knowledge search based on embedded vectors, which is used to improve the accuracy of operation and maintenance knowledge retrieval results, enhance the retrieval efficiency, and ensure the subsequent operation and maintenance effect and efficiency.

[0006] The present invention provides a method for realizing operation and maintenance knowledge search based on embedded vectors, comprising:

[0007] Step 1: obtaining initial text data, and preprocessing the initial text data using a preset preprocessing method to obtain first data;

[0008] Step 2: Mapping the first data by combining with a preset vector database and using a preset language model to obtain a corresponding first vector;

[0009] Step 3: Perform similarity analysis on the first vector using a preset algorithm in combination with a preset knowledge base, and output a vector analysis result;

[0010] Step 4: Sort the documents corresponding to the vector analysis results in the preset knowledge base using a preset sorting method, and select the retrieval results based on the preset selection instructions.

[0011] Preferably, step 1 includes:

[0012] Obtain the search instructions input by the system in real time, identify the corresponding text data and data format, and output the initial text data;

[0013] Based on the initial text data, a corresponding preset preprocessing method is acquired, and the initial text data is preprocessed using the corresponding preset preprocessing method to obtain first data.

[0014] Preferably, based on the initial text data, a corresponding preset preprocessing method is obtained, including:

[0015] Performing keyword recognition and extraction on the initial text data to construct a keyword recognition result;

[0016] At the same time, feature extraction is performed on the initial text data, and a corresponding preset preprocessing method is selected from a method database in combination with the keyword recognition result and the extracted features.

[0017] Preferably, step 2 includes:

[0018] Acquire a target conversion format, and in combination with a data format corresponding to the first data, acquire a corresponding preset language model;

[0019] Based on the preset language model, the first data is converted into a vector representation to obtain a first vector. At the same time, an index between the first data and the first vector is established, and the new index relationship therein is synchronously updated to the preset vector database.

[0020] Preferably, establishing an index between the first data and the first vector, and synchronously updating the new index relationship therein to a preset vector database, includes:

[0021] Constructing an index relationship list based on a vector space mapping relationship between the first data and the first vector;

[0022] The index relationship list is analyzed for differences with the index data in the preset vector database, and new index relationships that meet the first preset condition are synchronized to the preset vector database.

[0023] Preferably, step 3 includes:

[0024] Parsing all the generated first vectors to obtain first sub-vectors corresponding to the words, and performing similarity analysis between the first sub-vectors and each vector in the first vector set in the preset knowledge base to obtain a first result;

[0025] Based on the first result, merging the first sub-vectors corresponding to adjacent words within a preset range to obtain a total vocabulary vector set;

[0026] Based on a plurality of preset sentence structures selected from a sentence database, the total vocabulary vector set is divided to construct a heterogeneous sentence vector set and a homogeneous sentence vector set;

[0027] Inputting the heterogeneous sentence vector set into a preset knowledge base and performing similarity analysis with the corresponding second vector set to obtain a second result;

[0028] At the same time, the isomorphic sentence vector set is subjected to similarity analysis with a second vector set in a preset knowledge base to obtain an isomorphic sentence analysis result;

[0029] A consensus is formed on all the isomorphic sentence analysis results under the same isomorphic sentence vector set to obtain a third result;

[0030] Summarizing all the first results, the second results, and the third results to obtain a first analysis result;

[0031] The first analysis result is subjected to context analysis and context analysis by a preset context-context analysis module, and the first analysis result is adjusted based on the context analysis result and the context analysis result to obtain a vector analysis result.

[0032] Preferably, step 4 includes:

[0033] Acquire a search requirement corresponding to the initial text data, and output the requirement information;

[0034] Based on the demand information, at least two different preset language models are selected from a model database to construct a model library;

[0035] Based on the multiple preset language models in the model library, the first data is mapped and similarity analyzed to obtain parallel analysis results corresponding to the preset language models;

[0036] At the same time, the model identification information corresponding to each preset language model is annotated to the corresponding parallel analysis result to obtain the model analysis result;

[0037] Obtaining format information of the model analysis results corresponding to each preset language model, and performing format conversion in combination with a preset standardization method to obtain standardized analysis results corresponding to each preset language model;

[0038] Performing feature recognition and feature extraction on the standardized analysis results to obtain feature sets under each preset language model;

[0039] Based on the feature set and in combination with a preset feature-coefficient comparison table, a first weight coefficient corresponding to each of the standardized analysis results is selected from a weight database;

[0040] At the same time, combining the demand information, obtaining historical search records and heat maps corresponding to each of the feature sets in the historical database;

[0041] Based on the historical search records and the heat map, the first weight coefficient is modified to obtain a second weight coefficient;

[0042] Based on the standardized analysis results under each preset language model and the corresponding second weight coefficient, a preset analysis method is used to obtain the relevance corresponding to each preset language model;

[0043] Based on the relevance, the documents corresponding to each of the standardized analysis results are sorted according to a preset sorting method to obtain a search result sorting table;

[0044] The data in the search result sorting table is selected based on the preset selection instruction, and the final search result is output.

[0045] Preferably, based on the standardized analysis results under each preset language model and the corresponding second weight coefficient, a preset analysis method is used to obtain the relevance corresponding to each preset language model, including:

[0046] Perform a content analysis on the standardized analysis results corresponding to each preset language model, and perform cluster analysis on the analyzed content to obtain the character vector set, word vector set, sentence vector set, and paragraph vector set under each preset language model;

[0047] Based on the mapping relationship between the character vector set, the word vector set, the sentence vector set and the paragraph vector set, a character-word-sentence-paragraph embedding vector set corresponding to each preset language model is constructed;

[0048] A first preset analysis method is selected from a method database based on a feature set under each preset language model, and a comparison vector set corresponding to a character-word-sentence-paragraph embedding vector set corresponding to each preset language model is selected from a preset vector database based on the feature set under each preset language model;

[0049] Dimensionally matching the character-word-sentence-paragraph embedding vector set with corresponding vectors in the comparison vector set;

[0050] Based on the first preset analysis method and the second weight coefficient, the character-word-sentence-paragraph embedding vector set whose matching results meet the preset threshold condition is analyzed with the corresponding vector in the corresponding comparison vector set to obtain the first similarity corresponding to each preset language model;

[0051] ;

[0052] in, represents the first similarity corresponding to the i-th preset language model; Represents the autocorrelation adjustment factor of the character vector set, word vector set, sentence vector set and paragraph vector set under the i-th preset language model; represents the mutual correlation adjustment factor between the character vector set, word vector set, sentence vector set and paragraph vector set under the i-th preset language model; e represents the exponential function; m represents the dimension corresponding to the two vectors to be compared; represents the coordinate value of each vector in the character-word-sentence-paragraph vector set under the i-th preset language model in the j-th dimension in the vector space; represents the coordinate value of each corresponding vector in the comparison vector set under the i-th preset language model located in the j-th dimension in the vector space;

[0053] At the same time, selecting from a method database at least one second preset analysis method different from the first preset analysis method, and obtaining a second similarity based on the second preset analysis method;

[0054] A consensus is reached on the first similarities and the second similarities under the same model, and the relevance of each preset language model is obtained based on the consensus result.

[0055] The present invention provides a method for realizing operation and maintenance knowledge search based on embedded vectors. By using a preset preprocessing method to preprocess the acquired initial text data, the accuracy of the operation and maintenance retrieval data is improved, which facilitates the subsequent operation and maintenance retrieval. Then, the corresponding embedded vector is obtained through a preset language model, and similarity analysis is performed in combination with a preset algorithm. The semantic understanding ability of the operation and maintenance retrieval information is improved through the vector analysis method. In the case where there is no completely matching keyword between the query and the document, relevant content can also be accurately retrieved based on semantic similarity, thereby improving the accuracy of the retrieval. The present invention can also combine the parallel vector analysis results of multiple models to further improve the accuracy of the retrieval results. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0057] Figure 1 It is a flowchart of a method for realizing operation and maintenance knowledge search based on embedded vectors provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] like Figure 1 As shown, a method for realizing operation and maintenance knowledge search based on embedded vectors provided by an embodiment of the present invention includes:

[0060] Step 1: obtaining initial text data, and preprocessing the initial text data using a preset preprocessing method to obtain first data;

[0061] Step 2: Mapping the first data using a preset language model in combination with a preset vector database to obtain a corresponding first vector;

[0062] Step 3: Perform similarity analysis on the first vector using a preset algorithm in combination with a preset knowledge base, and output a vector analysis result;

[0063] Step 4: Sort the documents corresponding to the vector analysis results in the preset knowledge base using a preset sorting method, and select the retrieval results based on the preset selection instructions.

[0064] In this embodiment, the initial text data, i.e., the text data containing the operation and maintenance search information, is the original data directly obtained;

[0065] In this embodiment, the preset preprocessing method: the method used to preprocess the initial text data is preset, for example, data cleaning, etc.;

[0066] In this embodiment, the first data: that is, the initial text data after preprocessing, the data quality of which is improved compared with the initial text data;

[0067] In this embodiment, the preset vector database: a database containing vector representations corresponding to various operation and maintenance knowledge, which is preset;

[0068] In this embodiment, the preset language model: the language model used to obtain the embedding vector corresponding to the first data is preset, for example, BERT, GPT, etc.;

[0069] In this embodiment, the first vector: that is, an embedding vector corresponding to the first data obtained by a preset language model;

[0070] In this embodiment, the preset knowledge base: a database containing operation and maintenance documents, operation and maintenance problems and corresponding solutions and other operation and maintenance related knowledge, is pre-set;

[0071] In this embodiment, the preset algorithm: the algorithm used for similarity analysis of the first vector is pre-set, for example, cosine similarity, Euclidean distance and other algorithms;

[0072] In this embodiment, the vector analysis result: that is, the analysis result obtained by performing similarity analysis on the first vector;

[0073] In this embodiment, the preset sorting method is a method for sorting corresponding documents in a preset knowledge base based on the vector analysis results. In the present invention, the parallel vector analysis results of multiple models can be comprehensively considered to sort the documents, thereby reducing the error caused by a single model, thereby improving the accuracy of the sorting results;

[0074] In this embodiment, the preset selection instruction: the instruction for selecting documents from the sorting results is preset and is usually dynamically adjusted according to the retrieval requirements.

[0075] The implementation principle and beneficial effects of this embodiment: The present invention improves the accuracy of operation and maintenance retrieval data by preprocessing the acquired initial text data using a preset preprocessing method, providing convenience for subsequent operation and maintenance retrieval, and then obtains the corresponding embedding vector through a preset language model, and performs similarity analysis in combination with a preset algorithm. The semantic understanding ability of operation and maintenance retrieval information is improved through the vector analysis method. In the case where there is no complete matching keyword between the query and the document, relevant content can also be accurately retrieved based on semantic similarity, thereby improving the accuracy of the retrieval. The present invention can also combine the parallel vector analysis results of multiple models to further improve the accuracy of the retrieval results.

[0076] An embodiment of the present invention provides a method for realizing operation and maintenance knowledge search based on an embedded vector, wherein step 1 includes:

[0077] Obtain the search instructions input by the system in real time, identify the corresponding text data and data format, and output the initial text data;

[0078] Based on the initial text data, a corresponding preset preprocessing method is acquired, and the initial text data is preprocessed using the corresponding preset preprocessing method to obtain first data.

[0079] In this embodiment, the search instruction is an instruction for searching the operation and maintenance knowledge in the operation and maintenance knowledge base, so as to quickly find relevant operation and maintenance documents, solutions, etc., so as to quickly solve the operation and maintenance problems with the help of the search results, so as to improve the operation and maintenance effect and efficiency;

[0080] In this embodiment, text data: that is, text information corresponding to the search instruction;

[0081] In this embodiment, the data format refers to the format corresponding to the text data, for example, the txt format of plain text, the HTML format of web page text, and the XML format of extensible markup language.

[0082] The implementation principle and beneficial effects of this embodiment: The present invention can identify the search instruction, identify the text data contained in the search instruction and the corresponding data format, convert it into initial text data, and then select the corresponding preset preprocessing method according to the format of the text data and the data characteristics, thereby improving the matching degree between the preprocessing method and the data to be preprocessed, thereby ensuring the data processing effect and efficiency of the preprocessing, and improving the data quality of the first data.

[0083] The embodiment of the present invention provides a method for realizing operation and maintenance knowledge search based on embedded vectors, which obtains a corresponding preset preprocessing method based on initial text data, including:

[0084] Perform keyword recognition and extraction on the initial text data to construct a keyword recognition result;

[0085] At the same time, feature extraction is performed on the initial text data, and the corresponding preset preprocessing method is selected in the method database in combination with the keyword recognition result and the extracted features.

[0086] In this embodiment, the keyword recognition result includes the recognition results of the operation and maintenance keywords recognized and extracted from the initial text data, such as fault codes, identity authentication information, etc.;

[0087] In this embodiment, the method database: a database containing a variety of preprocessing methods, which is used to obtain corresponding preprocessing methods according to the keyword recognition results and extracted features to improve the preprocessing effect.

[0088] The implementation principle and beneficial effects of this embodiment: The present invention combines the keyword recognition results obtained by keyword recognition and the extracted features obtained by feature extraction, and selects a method from the method database for preprocessing each data in the initial text data, thereby improving the adaptation effect between the method and the data, and thereby ensuring the preprocessing effect.

[0089] An embodiment of the present invention provides a method for realizing operation and maintenance knowledge search based on an embedded vector, wherein step 2 includes:

[0090] Obtaining a target conversion format, and combining it with a data format corresponding to the first data to obtain a corresponding preset language model;

[0091] Based on a preset language model, the first data is converted into a vector representation to obtain a first vector. At the same time, an index between the first data and the first vector is established, and the new index relationship therein is synchronously updated to a preset vector database.

[0092] In this embodiment, the target conversion format is a target format for vector mapping, which is used to select an adapted language model in combination with the data format corresponding to the first data, so as to improve the matching degree between the model and the data;

[0093] In this embodiment, the index is a mapping relationship between the first vector and the first data, which facilitates subsequent retrieval or vector conversion and improves retrieval efficiency and acquisition speed of embedded vectors.

[0094] The implementation principles and beneficial effects of this embodiment: The present invention obtains the corresponding preset language model by combining the target conversion with the corresponding data format in the first data, thereby achieving a high degree of matching between the language model and the data, and then maps the first data based on the selected language model to obtain the first vector, and synchronizes the new index relationship to the preset vector database, gradually improving the richness of the vectors in the preset vector database, and providing convenience for subsequent vector mapping and operation and maintenance retrieval.

[0095] An embodiment of the present invention provides a method for realizing operation and maintenance knowledge search based on embedded vectors, establishing an index between first data and a first vector, and synchronously updating the new index relationship therein to a preset vector database, including:

[0096] Constructing an index relationship list based on a vector space mapping relationship between the first data and the first vector;

[0097] The index relationship list is analyzed for differences with the index data in the preset vector database, and new index relationships that meet the first preset condition are synchronized to the preset vector database.

[0098] In this embodiment, the index relationship list: that is, a data table containing mapping relationships of various vector spaces, which facilitates querying and comparing all index relationships in the first data;

[0099] In this embodiment, difference analysis is used to identify newly established index relationships and update the new index relationships to a preset vector database, so as to gradually improve the diversity and richness of index data in the vector database, thereby facilitating subsequent operation and maintenance retrieval.

[0100] In this embodiment, the first preset condition is a determination condition for determining whether the index relationship belongs to a new index relationship.

[0101] The implementation principle and beneficial effects of this embodiment: The present invention constructs an index relationship list according to the vector space mapping relationship corresponding to each data in the first data. This not only can intuitively understand the vector mapping relationship of each data in the first data and facilitate the search for abnormal mapping data, but also facilitates the subsequent updating of the index data in the preset vector database, continuously improving the diversity of the vector database, and laying a data foundation for subsequent operation and maintenance retrieval.

[0102] An embodiment of the present invention provides a method for realizing operation and maintenance knowledge search based on an embedded vector, wherein step 3 includes:

[0103] Parsing all generated first vectors to obtain first sub-vectors corresponding to the words, and performing similarity analysis between the first sub-vectors and each vector in the first vector set in the preset knowledge base to obtain a first result;

[0104] Based on the first result, first sub-vectors corresponding to adjacent words within a preset range are merged to obtain a total vocabulary vector set;

[0105] Combined with multiple preset sentence structures selected from the sentence database, the total vocabulary vector set is divided to construct a heterogeneous sentence vector set and a homogeneous sentence vector set;

[0106] Inputting the heterogeneous sentence vector set into a preset knowledge base and performing similarity analysis with the corresponding second vector set to obtain a second result;

[0107] At the same time, the isomorphic sentence vector set is similarly analyzed with the second vector set in the preset knowledge base to obtain the isomorphic sentence analysis result;

[0108] The third result is obtained by consensus of all isomorphic sentence analysis results under the same isomorphic sentence vector set;

[0109] Summarize all the first results, the second results, and the third results to obtain a first analysis result;

[0110] The first analysis result is subjected to context analysis and context analysis by a preset context-context analysis module, and the first analysis result is adjusted based on the context analysis result and the context analysis result to obtain a vector analysis result.

[0111] In this embodiment, the first sub-vector: that is, all first vectors about a word, can also be called a word embedding vector;

[0112] In this embodiment, the first vector set: that is, a word vector data set in a preset knowledge base, is used for comparison with the first sub-vector;

[0113] In this embodiment, the first result: a result obtained by performing similarity analysis on the first sub-vector and each vector in the first vector set;

[0114] In this embodiment, the preset range is a pre-set length range containing multiple adjacent words, which is used for subsequent sentence construction and semantic analysis;

[0115] In this embodiment, the total vocabulary vector set is a set obtained by summarizing the first sub-vectors corresponding to all words within a preset range;

[0116] In this embodiment, the preset sentence structure: the structure used for sentence construction is preset, for example, the sentence structure of subject + predicate, subject + predicate + object, etc.;

[0117] In this embodiment, the heterogeneous sentence vector set is a set of vectors corresponding to sentences of different structures obtained by constructing sentences with the same words or different words according to different sentence structures;

[0118] In this embodiment, the isomorphic sentence vector set is a set of vectors corresponding to sentences of the same structure obtained by constructing sentences with the same words or different words according to the same sentence structure;

[0119] In this embodiment, the second vector set: that is, the sentence vector set in the preset knowledge base, is used for comparative analysis with the isomorphic sentence vector set and the heteromorphic sentence vector set;

[0120] In this embodiment, the second result: the analysis result obtained by performing similarity analysis on the heterogeneous sentence vector set and the second vector set;

[0121] In this embodiment, the isomorphic sentence analysis result: the analysis result obtained by performing similarity analysis on the isomorphic sentence vector set and the second vector set;

[0122] In this embodiment, the third result: a consensus result obtained by consensus of all isomorphic sentence analysis results under the same isomorphic sentence vector set;

[0123] In this embodiment, the first analysis result: that is, the summary data of the first result, the second result and the third result;

[0124] In this embodiment, a preset context-context analysis module: a module used to perform context analysis and context analysis on the first analysis result, is pre-trained through big data.

[0125] The implementation principle and beneficial effects of this embodiment: The present invention parses all first vectors to obtain the first sub-vector about the word, and performs similarity analysis with the first vector set in the preset knowledge base to obtain a first result, and then summarizes and merges the first sub-vectors in combination with the preset sentence structure to construct a heterogeneous sentence vector set and a homogeneous sentence vector set, and performs similarity analysis on them to obtain a second result and a third result, and then combines the first result to obtain a first analysis result and performs context analysis and context analysis to obtain a vector analysis result. It not only realizes vector mapping and similarity analysis from the two aspects of words and sentences, but also takes into account the impact of different sentence structures on the meaning of sentences, greatly improving the deep semantic understanding ability of the operation and maintenance knowledge retrieval system for sentences, thereby ensuring the accuracy and efficiency of the retrieval results.

[0126] An embodiment of the present invention provides a method for realizing operation and maintenance knowledge search based on an embedded vector, wherein step 4 includes:

[0127] Obtaining the search requirements corresponding to the initial text data and outputting the requirement information;

[0128] Based on the demand information, at least two different preset language models are selected from the model database to construct a model library;

[0129] Based on multiple preset language models in the model library, the first data is mapped and similarity analyzed to obtain parallel analysis results corresponding to each preset language model;

[0130] At the same time, the model identification information corresponding to each preset language model is annotated to the corresponding parallel analysis result to obtain the model analysis result;

[0131] Obtaining format information of model analysis results corresponding to each preset language model, and performing format conversion in combination with a preset standardization method to obtain standardized analysis results corresponding to each preset language model;

[0132] Perform feature recognition and feature extraction on the standardized analysis results to obtain feature sets under each preset language model;

[0133] Based on the feature set and in combination with a preset feature-coefficient comparison table, a first weight coefficient corresponding to each standardized analysis result is selected from a weight database;

[0134] At the same time, historical search records and heat maps corresponding to each feature set are obtained in the historical database in combination with the demand information;

[0135] Based on the historical search records and the heat map, the first weight coefficient is modified to obtain a second weight coefficient;

[0136] Based on the standardized analysis results under each preset language model and the corresponding second weight coefficient, a preset analysis method is used to obtain the relevance corresponding to each preset language model;

[0137] Based on the relevance, the documents corresponding to each standardized analysis result are sorted according to a preset sorting method to obtain a search result sorting table;

[0138] Select the data in the search result sorting table based on the preset selection instructions and output the final search results.

[0139] In this embodiment, the demand information: that is, information including the operation and maintenance knowledge retrieval requirements, for example, performance requirements such as accuracy and recall rate;

[0140] In this embodiment, the model database includes a database of multiple preset language models for vector mapping, including but not limited to Word2Vec, FastText, BERT and other models;

[0141] In this embodiment, the model library is a database composed of multiple selected preset language models;

[0142] In this embodiment, the parallel analysis result: that is, the analysis result obtained by performing similar analysis on each preset language model at the same time;

[0143] In this embodiment, the model identification information includes information of significant identification corresponding to each preset language model, which is used to characterize the source of each analysis result;

[0144] In this embodiment, the model analysis result: that is, the parallel analysis result including identification information of each preset language model;

[0145] In this embodiment, the preset standardization method: the method for standardizing the analysis results corresponding to each preset language model is preset;

[0146] In this embodiment, the standardized analysis result refers to the model analysis result obtained after the standardized processing;

[0147] In this embodiment, the feature set is a set of data features obtained by extracting features from the standardized analysis results;

[0148] In this embodiment, the preset feature-coefficient comparison table: a comparison table containing mapping relationships between data features and weight coefficients, which is preset and used to obtain the corresponding first weight coefficient according to the input data features;

[0149] In this embodiment, the weight database: includes a database of first weight coefficients, which is used to select and obtain weight coefficients corresponding to data features according to the comparison table;

[0150] In this embodiment, the first weight coefficient is a weight coefficient selected from a weight database;

[0151] In this embodiment, the historical search record: that is, the historical search information of each data feature;

[0152] In this embodiment, the heat map is a chart containing the search frequency of each data feature and the correlation between each feature, and different shades of colors are usually used to represent the search frequency and correlation degree of the feature;

[0153] In this embodiment, the second weight coefficient is the first weight coefficient after correction;

[0154] In this embodiment, the preset analysis method: the analysis method used to calculate the relevance of each preset language model is preset;

[0155] In this embodiment, relevance is a parameter used to characterize the strength of the association between the analysis results corresponding to each preset language model and the search object;

[0156] In this embodiment, the search result ranking table is a table obtained by ranking the documents corresponding to each standardized analysis result according to a preset ranking method, and is usually arranged in descending order of relevance.

[0157] The implementation principle and beneficial effects of this embodiment: The present invention obtains demand information and selects multiple adapted preset language models in the model database, and then obtains parallel analysis results under each preset language model, reducing the system error caused by similar analysis of a single model, and then performs standardized analysis on the model analysis results under multiple models, and obtains the relevance under each preset language model in combination with the obtained second weight coefficient, and obtains a retrieval result sorting table according to the preset sorting method. Not only can the model information corresponding to each retrieval result be intuitively obtained through the sorting table, but also the accuracy of the retrieval result is guaranteed.

[0158] An embodiment of the present invention provides a method for realizing operation and maintenance knowledge search based on an embedded vector, which obtains the relevance corresponding to each preset language model using a preset analysis method based on the standardized analysis results under each preset language model and the corresponding second weight coefficient, including:

[0159] Perform a content analysis on the standardized analysis results corresponding to each preset language model, and perform cluster analysis on the analyzed content to obtain the character vector set, word vector set, sentence vector set, and paragraph vector set under each preset language model;

[0160] Based on the mapping relationship between the character vector set, the word vector set, the sentence vector set and the paragraph vector set, a character-word-sentence-paragraph embedding vector set corresponding to each preset language model is constructed;

[0161] A first preset analysis method is selected from a method database based on a feature set under each preset language model, and a comparison vector set corresponding to a character-word-sentence-paragraph embedding vector set corresponding to each preset language model is selected from a preset vector database based on a feature set under each preset language model;

[0162] Dimensionally match the character-word-sentence-paragraph embedding vector set with the corresponding vector in the comparison vector set;

[0163] Based on the first preset analysis method and the second weight coefficient, the character-word-sentence-paragraph embedding vector set whose matching results meet the preset threshold condition is analyzed with the corresponding vector in the corresponding comparison vector set to obtain the first similarity corresponding to each preset language model;

[0164] ;

[0165] in, represents the first similarity corresponding to the i-th preset language model; Represents the autocorrelation adjustment factor of the character vector set, word vector set, sentence vector set and paragraph vector set under the i-th preset language model; represents the mutual correlation adjustment factor between the character vector set, word vector set, sentence vector set and paragraph vector set under the i-th preset language model; e represents the exponential function; m represents the dimension corresponding to the two vectors to be compared; represents the coordinate value of each vector in the character-word-sentence-paragraph vector set under the i-th preset language model in the j-th dimension in the vector space; represents the coordinate value of each corresponding vector in the comparison vector set under the i-th preset language model in the j-th dimension in the vector space;

[0166] At the same time, selecting from the method database at least one second preset analysis method different from the first preset analysis method, and obtaining a second similarity based on the second preset analysis method;

[0167] A consensus is reached on each first similarity and second similarity under the same model, and the relevance of each preset language model is obtained based on the consensus result.

[0168] In this embodiment, the character vector set is a set including the embedding vectors corresponding to each character;

[0169] In this embodiment, the word vector set is a set of embedding vectors corresponding to each word or phrase;

[0170] In this embodiment, the sentence vector set: that is, a set of embedding vectors corresponding to each sentence pattern;

[0171] In this embodiment, the paragraph vector set: that is, a set including the embedding vectors corresponding to each paragraph;

[0172] In this embodiment, the character-word-sentence-paragraph embedding vector set includes a character vector set, a word vector set, a sentence vector set, a paragraph vector set, and a vector set of mapping relationships between the vector sets;

[0173] In this embodiment, the first preset analysis method: the method for calculating the first similarity of each preset language model is preset, and the present invention adopts an improved algorithm of cosine similarity;

[0174] In this embodiment, the comparison vector set is a vector set corresponding to the character-word-sentence-paragraph embedding vector set selected from a preset vector database and used for comparison analysis;

[0175] In this embodiment, the preset threshold condition is used to determine whether the dimension matching result meets the threshold condition of subsequent analysis and calculation. It is pre-set. In the present invention, the vectors of the dimensions of the two vectors are the same and have calculation value. Therefore, the preset threshold condition is that the dimensions of the two vectors are the same.

[0176] In this embodiment, the first similarity: that is, the similarity calculation results under each preset language model are obtained by using the first preset analysis method;

[0177] In this embodiment, the second preset analysis method: that is, another vector analysis method different from the first preset analysis method, for example, Euclidean distance, etc.;

[0178] In this embodiment, the second similarity is the similarity calculation result under each preset language model calculated by the second preset method.

[0179] The implementation principle and beneficial effects of this embodiment: The present invention parses the character vector set, word vector set, sentence vector set and paragraph vector set under each model, and constructs a character-word-sentence-paragraph embedding vector set based on the mapping relationship between the sets. Subsequently, the similarity under each preset language model is calculated by obtaining multiple preset analysis methods, and the errors caused by the analysis method and model are taken into account, thereby improving the accuracy of the operation and maintenance retrieval results.

[0180] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for realizing operation and maintenance knowledge search based on embedding vectors, It is characterized in that include: Step 1: obtaining initial text data, and preprocessing the initial text data using a preset preprocessing method to obtain first data; Step 2: Mapping the first data by combining with a preset vector database and using a preset language model to obtain a corresponding first vector; Step 3: Perform similarity analysis on the first vector using a preset algorithm in combination with a preset knowledge base, and output a vector analysis result; Step 4: sorting the documents corresponding to the vector analysis results in the preset knowledge base using a preset sorting method, and selecting the retrieval results based on the preset selection instructions; Wherein, step 4 includes: Acquire a search requirement corresponding to the initial text data, and output the requirement information; Based on the demand information, at least two different preset language models are selected from a model database to construct a model library; Based on the multiple preset language models in the model library, the first data is mapped and similarity analyzed to obtain parallel analysis results corresponding to the preset language models; At the same time, the model identification information corresponding to each preset language model is annotated to the corresponding parallel analysis result to obtain the model analysis result; Obtaining format information of the model analysis results corresponding to each preset language model, and performing format conversion in combination with a preset standardization method to obtain standardized analysis results corresponding to each preset language model; Performing feature recognition and feature extraction on the standardized analysis results to obtain feature sets under each preset language model; Based on the feature set and in combination with a preset feature-coefficient comparison table, a first weight coefficient corresponding to each of the standardized analysis results is selected from a weight database; At the same time, combining the demand information, obtaining historical search records and heat maps corresponding to each of the feature sets in the historical database; Based on the historical search records and the heat map, the first weight coefficient is modified to obtain a second weight coefficient; Based on the standardized analysis results under each preset language model and the corresponding second weight coefficient, a preset analysis method is used to obtain the relevance corresponding to each preset language model; Based on the relevance, the documents corresponding to each of the standardized analysis results are sorted according to a preset sorting method to obtain a search result sorting table; Selecting data from the search result ranking table based on a preset selection instruction, and outputting a final search result; Wherein, based on the standardized analysis results under each preset language model and the corresponding second weight coefficient, the corresponding relevance of each preset language model is obtained by using a preset analysis method, including: Perform a content analysis on the standardized analysis results corresponding to each preset language model, and perform cluster analysis on the analyzed content to obtain the character vector set, word vector set, sentence vector set, and paragraph vector set under each preset language model; Based on the mapping relationship between the character vector set, the word vector set, the sentence vector set and the paragraph vector set, a character-word-sentence-paragraph embedding vector set corresponding to each preset language model is constructed; A first preset analysis method is selected from a method database based on a feature set under each preset language model, and a comparison vector set corresponding to a character-word-sentence-paragraph embedding vector set corresponding to each preset language model is selected from a preset vector database based on the feature set under each preset language model; Dimensionally matching the character-word-sentence-paragraph embedding vector set with corresponding vectors in the comparison vector set; Based on the first preset analysis method and the second weight coefficient, the character-word-sentence-paragraph embedding vector set whose matching results meet the preset threshold condition is analyzed with the corresponding vector in the corresponding comparison vector set to obtain the first similarity corresponding to each preset language model; ; in, represents the first similarity corresponding to the i-th preset language model; Represents the autocorrelation adjustment factor of the character vector set, word vector set, sentence vector set and paragraph vector set under the i-th preset language model; represents the mutual correlation adjustment factor between the character vector set, word vector set, sentence vector set and paragraph vector set under the i-th preset language model; e represents the exponential function; m represents the dimension corresponding to the two vectors to be compared; represents the coordinate value of each vector in the character-word-sentence-paragraph vector set under the i-th preset language model in the j-th dimension in the vector space; represents the coordinate value of each corresponding vector in the comparison vector set under the i-th preset language model located in the j-th dimension in the vector space; At the same time, selecting from a method database at least one second preset analysis method different from the first preset analysis method, and obtaining a second similarity based on the second preset analysis method; A consensus is reached on the first similarities and the second similarities under the same model, and the relevance of each preset language model is obtained based on the consensus result.

2. According to claim 1, a method for realizing operation and maintenance knowledge search based on embedded vectors, It is characterized in that Step 1 includes: Acquire the search instructions input by the system in real time, identify the corresponding text data and data format, and output the initial text data; Based on the initial text data, a corresponding preset preprocessing method is acquired, and the initial text data is preprocessed using the corresponding preset preprocessing method to obtain first data.

3. According to claim 2, a method for realizing operation and maintenance knowledge search based on embedded vectors, It is characterized in that Based on the initial text data, a corresponding preset preprocessing method is obtained, including: Performing keyword recognition and extraction on the initial text data to construct a keyword recognition result; At the same time, feature extraction is performed on the initial text data, and a corresponding preset preprocessing method is selected from a method database in combination with the keyword recognition result and the extracted features.

4. According to claim 1, a method for realizing operation and maintenance knowledge search based on embedded vectors, It is characterized in that Step 2 includes: Acquire a target conversion format, and in combination with a data format corresponding to the first data, acquire a corresponding preset language model; Based on the preset language model, the first data is converted into a vector representation to obtain a first vector. At the same time, an index between the first data and the first vector is established, and the new index relationship therein is synchronously updated to the preset vector database.

5. According to claim 4, a method for realizing operation and maintenance knowledge search based on embedded vectors, It is characterized in that Establishing an index between the first data and the first vector, and synchronously updating the new index relationship therein to a preset vector database, including: Constructing an index relationship list based on a vector space mapping relationship between the first data and the first vector; The index relationship list is analyzed for differences with the index data in the preset vector database, and new index relationships that meet the first preset condition are synchronized to the preset vector database.

6. According to claim 1, a method for realizing operation and maintenance knowledge search based on embedded vectors, It is characterized in that Step 3 includes: Parsing all the generated first vectors to obtain first sub-vectors corresponding to the words, and performing similarity analysis between the first sub-vectors and each vector in the first vector set in the preset knowledge base to obtain a first result; Based on the first result, merging the first sub-vectors corresponding to adjacent words within a preset range to obtain a total vocabulary vector set; Based on a plurality of preset sentence structures selected from a sentence database, the total vocabulary vector set is divided to construct a heterogeneous sentence vector set and a homogeneous sentence vector set; Inputting the heterogeneous sentence vector set into a preset knowledge base and performing similarity analysis with the corresponding second vector set to obtain a second result; At the same time, the isomorphic sentence vector set is subjected to similarity analysis with a second vector set in a preset knowledge base to obtain an isomorphic sentence analysis result; A consensus is formed on all the isomorphic sentence analysis results under the same isomorphic sentence vector set to obtain a third result; Summarizing all the first results, the second results, and the third results to obtain a first analysis result; The first analysis result is subjected to context analysis and context analysis by a preset context-context analysis module, and the first analysis result is adjusted based on the context analysis result and the context analysis result to obtain a vector analysis result.

Citation Information

Patent Citations

  • RAG knowledge question-answering method and device based on fusion vector and keyword retrieval

    CN117951274A

  • Knowledge retrieval method based on large language model and knowledge base management platform

    CN118210875A

  • Question and answer method and apparatus for machine reading comprehension, computer device, and storage medium

    WO2022227165A1