Knowledge graph-based retrieval model recommendation method and device

Abstract

The application discloses a retrieval model recommendation method and device based on a knowledge graph, and relates to the technical field of artificial intelligence.The application solves the technical problem that retrieval models are relatively single, mostly rely on simple keyword matching, and result in uneven quality of retrieval results.The application collects multi-source heterogeneous data, uses entity recognition and relationship extraction technology, and constructs a knowledge graph with rigorous structure and rich information.This makes the relationship between data clear, provides more comprehensive and accurate data support for subsequent retrieval models, and through a series of processes such as labeling, retrieval operation, result analysis and sorting, the best model can be selected according to different retrieval problems.This comprehensive use and scientific selection of multiple models change the limitations of traditional single model retrieval and significantly improve the adaptability of retrieval models and retrieval problems.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and specifically to a retrieval model recommendation method and device based on knowledge graphs. Background Art

[0002] With the development of artificial intelligence technology, retrieval models are widely used in various fields, such as search engines and natural language processing. The use of retrieval models can search and analyze large amounts of data quickly and efficiently.

[0003] According to the application with publication number CN117540062A, a retrieval model recommendation method and device based on knowledge graph are disclosed, which includes: obtaining multiple retrieval models and a knowledge graph corresponding to a target platform; based on the knowledge graph, analyzing the pre-acquired question information to obtain a question analysis result corresponding to the question information; for each retrieval model, searching for the question information based on the retrieval model to obtain a retrieval result corresponding to the retrieval model; and determining the target retrieval model corresponding to the target platform based on the question analysis result and the retrieval results corresponding to all retrieval models.

[0004] In the traditional search field, retrieval models are relatively simple, mostly relying on simple keyword matching, which makes it difficult to accurately understand the complex semantic information and potential technical connections in the text. Furthermore, faced with multi-source heterogeneous data, such as text, illustrations, and review opinions, there is a lack of effective integration and deep mining methods, making it impossible to fully realize the value of this data. Furthermore, there is no systematic classification and adaptation mechanism for different types of search questions, resulting in uneven quality of search results, making it difficult to meet the growing search needs. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides a retrieval model recommendation method and device based on knowledge graph, which solves the problem that the retrieval model is relatively simple and mostly relies on simple keyword matching, resulting in uneven quality of retrieval results.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: a retrieval model recommendation method based on knowledge graph, which specifically includes the following steps:

[0007] Build a knowledge graph, obtain all retrieval models and corresponding retrieval questions, and determine the question type corresponding to the retrieval question;

[0008] The search question is split to obtain split information, and the split information is searched using different search models to obtain corresponding search results. Then, the content matching degree between the search question and the search results is analyzed, and the search results are screened based on the content matching degree to obtain pre-selected results. At the same time, the search models are ranked according to the number of pre-selected results to generate model ranking information;

[0009] Obtain model ranking information, filter the search question to obtain keywords, and calculate the frequency of occurrence of keywords. Then calculate the semantic similarity between the search question and the search results. At the same time, calculate the sum of the frequency of occurrence and the semantic similarity to obtain the association value of the pre-selected results. Based on discrete value analysis, obtain the results to be analyzed, and calculate the average sum of the association values ​​of the results to be analyzed to obtain the association index.

[0010] The relevance index of the retrieval model is compared with the threshold to obtain the pre-selected model, and the rank assignment of the pre-selected model is determined based on the retrieval time. At the same time, the sum of the value of the pre-selected model and the relevance index is calculated to obtain the model benefit value. The pre-selected model with the largest model benefit value is selected to generate the recommended model, and the recommendation information is generated in combination with the retrieval type of the retrieval question.

[0011] As a further solution of the present invention, the specific method of splitting the search question to obtain split information is:

[0012] Obtain the search question and use the bag-of-words model to split the search question. The bag-of-words model will decompose the search question into independent and domain-meaningful words or phrases during splitting to obtain corresponding split information.

[0013] As a further solution of the present invention, the specific method of generating model ranking information is:

[0014] Get all retrieval models and label them as i, where i = 1, 2, ..., j, where j represents the number of retrieval models. Get the retrieval result Ci corresponding to retrieval model i, get the split information, and calculate the proportion of the number of split information in the retrieval result Ci;

[0015] And compare it with the preset value, and the search results with a screening quantity greater than the preset value are recorded as pre-selected results;

[0016] Similarly, all pre-selected results corresponding to the retrieval model i are screened and sorted from large to small according to the number of screened results to generate model ranking information.

[0017] As a further solution of the present invention, the specific method of screening the search question to obtain keywords is:

[0018] Obtain a search question, and use natural language processing tools to analyze the search question according to lexical rules, and combine the stop word filtering method to comprehensively process the search question, screen out the keywords corresponding to the search question, and label them as n, where n=1, 2, ..., m, where m represents the number of keywords.

[0019] As a further solution of the present invention, the specific method of calculating the frequency of occurrence of keywords is:

[0020] Perform sentence processing on the pre-selected result a to obtain sentence information, and calculate the number of times the keyword n ​​appears in the sentence information, and calculate the ratio of the number of appearances to the total number of sentence information;

[0021] Similarly, calculate the proportion of keyword n ​​in all sentence information, and then sum the calculated proportions to obtain the frequency of occurrence of keyword n.

[0022] As a further solution of the present invention, the specific method of calculating the sum of the correlation values ​​of the results to be analyzed to obtain the correlation index is:

[0023] Using the word vector model, the semantic similarity between the search question and the search results is calculated. The sum of the semantic similarity and the frequency of occurrence is calculated as the pre-selected result association value. The pre-selected results are then filtered according to the discrete value to obtain the results to be analyzed. At the same time, the mean of the sum of the association values ​​of the results to be analyzed is calculated and recorded as the association index between the model to be analyzed and the search question.

[0024] Similarly, the relevance indexes of all retrieval models are calculated and sorted from large to small according to the relevance index.

[0025] As a further solution of the present invention, the specific method of obtaining the model benefit value is:

[0026] Obtain the correlation index corresponding to the retrieval model and compare it with the threshold. Filter the retrieval model with a correlation index greater than the threshold and record it as the pre-selected model. Then obtain the retrieval time corresponding to the pre-selected model, and divide the retrieval time into level intervals based on big data analysis, and perform corresponding assignment processing. At the same time, match the retrieval time of the pre-selected model with the level interval, determine the retrieval time level corresponding to the pre-selected model, and obtain the corresponding level assignment. Then calculate the sum of the level assignment and the correlation index, and record it as the model benefit value.

[0027] As a further solution of the present invention, the specific method of generating recommendation information in combination with the search type of the search question is:

[0028] Select the pre-selected model with the largest model benefit value as the standard, generate the recommended model, obtain the corresponding search question type, and combine it with the recommended model to generate recommendation information;

[0029] Similarly, historical data is obtained, and the search types corresponding to all deceleration models are obtained to generate corresponding recommendation information.

[0030] The retrieval model recommendation device based on the knowledge graph includes:

[0031] Information acquisition module, used to obtain multiple retrieval models and corresponding knowledge graphs;

[0032] An information analysis module, configured to analyze the pre-acquired search questions based on the knowledge graph and obtain search results corresponding to the search questions;

[0033] A retrieval analysis module, configured to perform a search for each retrieval model based on the retrieval model and the retrieval question, and obtain a retrieval result corresponding to the retrieval model;

[0034] The retrieval determination module is used to determine the corresponding target retrieval model based on the retrieval results corresponding to all the retrieval models and generate recommendation information.

[0035] Beneficial effects

[0036] The present invention provides a method and device for recommending retrieval models based on knowledge graphs. Compared with the existing technology, it has the following advantages:

[0037] This method collects heterogeneous data from multiple sources and applies entity recognition and relationship extraction techniques to construct a rigorous and information-rich knowledge graph. This allows for clear visualization of relationships between data, providing more comprehensive and accurate data support for subsequent retrieval models. Compared to traditional approaches that rely on a single data source or simple data processing, this approach enhances the value and depth of data utilization.

[0038] This encompasses a variety of retrieval models based on keyword matching, deep learning for semantic understanding, and knowledge graph structures. Through a series of processes, including labeling retrieval models, performing retrieval operations, analyzing and sorting results, we can select the best performing models for different search questions. This comprehensive application of multiple models and scientific screening approach overcomes the limitations of traditional single-model retrieval and significantly improves the adaptability of retrieval models to search questions.

[0039] During the search process, the recommendation model is ultimately determined through operations such as query splitting, keyword screening, and semantic similarity calculation, combined with multi-dimensional analysis of the search results, such as calculating quantity percentages and relevance indices, and grading and assigning values ​​based on search time. This series of innovative methods significantly improves the matching degree between search results and query questions, significantly increases search efficiency, and effectively controls search time. Compared with traditional search methods, it can provide searchers with faster and more accurate search services, helping them obtain the required information more efficiently. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a diagram of the steps and methods of the present invention;

[0041] Figure 2 This is a module diagram of the device of the present invention. DETAILED DESCRIPTION

[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only 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 making creative efforts are within the scope of protection of the present invention.

[0043] For example 1, please refer to Figure 1 , this application provides a retrieval model recommendation method based on knowledge graph, which specifically includes the following steps:

[0044] Step 1.

[0045] First, we set about building a knowledge graph. By collecting multi-source heterogeneous data, including text, illustration information, examination opinions, etc., and using entity recognition technology, we can accurately extract various entities from these data, such as the name of the invention, technical field, classification number and other entities. Then, we use relationship extraction technology to determine the association between entities, such as the "invention name-belonging-technical field" relationship. With the help of knowledge fusion methods, we will integrate the same entity information obtained from different data sources to eliminate data redundancy and conflict, and finally build a knowledge graph with rigorous structure and rich information. For example, when constructing a knowledge graph in the field of electronic communications, we extract entities such as "5G communication technology" and "chip manufacturing process" from a large amount of electronic communications literature, as well as the relationship between them "5G communication technology-application-chip manufacturing process".

[0046] After the knowledge graph is built, all retrieval models applicable to retrieval are fully acquired. These retrieval models cover a variety of types, including traditional retrieval models based on keyword matching, which perform retrieval by searching for content that exactly or fuzzily matches the input keyword within the entity and relationship descriptions of the knowledge graph; deep learning retrieval models based on semantic understanding, which utilize deep learning algorithms to deeply mine and understand the semantic information within the knowledge graph and are capable of handling retrieval problems with complex semantics; and graph retrieval models based on the structure of the knowledge graph, which perform retrieval based on the structural features of nodes and edges within the knowledge graph, for example, by finding the adjacent nodes of a specific node or a specific path to obtain relevant information.

[0047] Step 2,

[0048] First, we comprehensively sort out and collect various retrieval models suitable for retrieval. These models may include traditional retrieval models based on keyword matching, deep learning retrieval models based on semantic understanding, and specialized retrieval models designed for text structure characteristics. All collected retrieval models are labeled in sequence, denoted as i, and i=1, 2, ..., j, where j represents the number of retrieval models. For example, if we have collected three models, namely, a model based on title keyword retrieval, a model based on full-text semantic retrieval, and a model based on classification number retrieval, then j=3, and the corresponding i are 1, 2, and 3 respectively.

[0049] For each retrieval model i, use it to perform a retrieval operation on the input retrieval question to obtain the corresponding retrieval result Ci. For example, retrieval model 1 is a model based on title keyword retrieval. When the retrieval question is "a new type of energy-saving air conditioner", its retrieval result C1 may be a list of documents whose titles contain keywords related to "energy-saving air conditioner". Next, the retrieval question is split, and the bag-of-words model is used here. Due to the professionalism and standardization of the text, the bag-of-words model will decompose the retrieval question into independent and domain-meaningful words or phrases when splitting, ignoring the order relationship between them. For example, for the retrieval question "Looking for a smartphone with an efficient heat dissipation structure", after splitting by the bag-of-words model, the split information that may be obtained is the four groups of "looking for", "efficient heat dissipation structure", "smartphone", and "".

[0050] Then, for each search result Ci, calculate the percentage of the corresponding split information. Specifically, count the number of split information contained in the search result Ci and divide it by the total number of split information. Assuming there are five groups of split information after splitting, and a group of search results Ci corresponds to three groups of split information, then the percentage of the corresponding number of split information for that search result is 3 / 5. The obtained percentage is then compared with the preset value.

[0051] The specific value of the preset value is set by the search operator based on actual needs such as the accuracy requirements of the search target, the search time limit, and past experience. For example, the operator sets the preset value to 0.6 based on the higher accuracy expected for this search. The search results with a quantity ratio greater than the preset value are screened out and recorded as pre-selected results. In this way, the search results corresponding to all search models i are processed and screened in turn. For example, for the search model 1 in the above example, if the calculated quantity ratio is = 0.6, which is exactly equal to the preset value, if the preset value is set to a strictly greater than relationship, then the search result C1 will not be screened as a pre-selected result; if the preset value is set to a greater than or equal to relationship, then the search result C1 will be screened as a pre-selected result.

[0052] This process is repeated for all retrieval models i (from 1 to j) to complete the screening of the preselected results corresponding to all retrieval models. Finally, the screened preselected results are sorted in descending order of quantity to generate model sorting information. For example, if the number of preselected results screened by retrieval model 2 is 100, the number of preselected results screened by retrieval model 3 is 80, and the number of preselected results screened by retrieval model 1 is 50, then the generated model sorting information is that retrieval model 2 ranks first, retrieval model 3 ranks second, and retrieval model 1 ranks last. Through such sorting, the performance of different retrieval models in this retrieval problem can be intuitively understood, providing a strong basis for subsequent optimization of the retrieval strategy or selection of the retrieval model.

[0053] Step3、

[0054] Obtain the corresponding sorting information, and take one group of retrieval models as the analysis object for analysis. Obtain the preselected results corresponding to the analysis object and label them as a, where a = 1, 2,..., b, and b represents the number of preselected results of the analysis object. At the same time, obtain the retrieval time corresponding to the analysis object, and here the retrieval time represents the duration from the start of the retrieval according to the retrieval problem to obtaining the retrieval results. Then, perform keyword screening on the retrieval problem, and the specific screening method is as follows:

[0055] Obtain the retrieval problem, and at the same time, with the help of natural language processing tools, analyze the retrieval problem according to the lexical rules. For example, use a词性标注工具 (POS tagging tool) to identify words of different词性 (POS) such as nouns, verbs, and adjectives. Usually, nouns as entity concepts and verbs representing actions or behaviors are relatively key information and can be preferentially considered as keywords. Combine the停用词过滤方法 (stop word filtering method) to comprehensively process the retrieval problem. Stop words are words that frequently appear in the text but have little effect on expressing the core meaning, such as "的" (de), "是" (shi), "在" (zai), "如何" (ruhe), etc. By establishing a stop word list, filter out the stop words in the retrieval problem. The remaining words often better reflect the key content of the problem. Screen the keywords corresponding to the retrieval problem and label them as n, where n = 1, 2,..., m, and m represents the number of keywords. Then, analyze the number of occurrences of keyword n in the preselected results a, and calculate the occurrence frequency of keyword n in the preselected results a, denoted as Pa. The calculation method of the occurrence frequency is as follows: perform sentence splitting on the preselected results a to obtain sentence splitting information, calculate the number of occurrences of keyword n in the sentence splitting information, and at the same time calculate the ratio of the number of occurrences to the total number of sentence splitting information. By analogy, calculate the ratio corresponding to keyword n in all sentence splitting information, and then sum up the calculated ratios to obtain the occurrence frequency of keyword n;

[0056] It should be noted that the Chinese term "词性标注工具" and "停用词过滤方法" in the original text are left untranslated as they might be specific tool or method names in Chinese that are more accurately understood in the context without a direct equivalent in English. If there are specific English terms for them, they can be replaced accordingly.For example, first obtain the search question, such as "Looking for lithium-ion battery manufacturing processes with high safety and long life." With the help of professional natural language processing tools, analyze it according to lexical rules. Taking the part-of-speech tagging tool as an example, the analysis can identify words of different parts of speech. In the field, nouns often represent key entity concepts, such as "lithium-ion battery" and "manufacturing process"; verbs reflect key actions or behaviors, such as "search." These words can be given priority as keywords. At the same time, combined with the stop word filtering method, stop words are words that appear frequently in the text but have little effect on expressing the core meaning, such as "of", "is", "in", "how", "about", etc. that are common in the text. By pre-establishing a stop word list, the stop words in the search question are removed. After this series of processing, the keywords corresponding to the search question are screened out.

[0057] Then, natural language processing techniques, such as word vector models and semantic role labeling, are used to calculate the semantic similarity between the retrieval question and the retrieval results. For example, by converting the sentences in the question and the result into vector space representations, and then using methods such as cosine similarity to measure the semantic distance between them, the closer the distance, the higher the association strength, and the sum of the numerical values ​​of semantic similarity and frequency of occurrence is calculated and recorded as the association value of the pre-selected result. Then, the pre-selected results are filtered according to the discrete value to obtain the results to be analyzed. Here, the pre-selected results whose association values ​​are lower than the overall discrete value are eliminated, and the remaining pre-selected results are marked as the results to be analyzed. The specific calculation of the discrete value is an existing technology and will not be described in detail here. At the same time, the mean of the sum of the numerical values ​​of the association values ​​of the results to be analyzed is calculated and recorded as the association index between the model to be analyzed and the retrieval question.

[0058] Similarly, the relevance indexes of all retrieval models are calculated and sorted from large to small according to the relevance index;

[0059] Step 4.

[0060] Next, the relevance index of each retrieval model is compared with a threshold set by the operator based on actual needs. For example, considering the high accuracy requirements of the retrieval results for this retrieval task, the operator sets a threshold of 1.2. Retrieval models with relevance indices greater than this threshold are screened out and marked as preselected models. Suppose, among the five existing retrieval models, the relevance index of the model based on full-text semantic retrieval is 1.5, and the relevance index of the model based on combined classification number and keyword retrieval is 1.3. These are both greater than the threshold of 1.2, so these two models are identified as preselected models.

[0061] For each preselected model, we need to obtain its corresponding search time. Search time refers to the length of time from entering the search question to obtaining the search results. For example, the above model based on full-text semantic search has a search time of 4 seconds when searching for the question "Miniaturized flexible battery technology for smart wearable devices." The model based on combined classification number and keyword search has a search time of 2 seconds when searching for the same question.

[0062] Then, based on big data analysis, we categorize search times into different levels and assign corresponding values. Based on historical search data, we assume three levels: 0-2 seconds is "fast," assigned a value of 3; 3-5 seconds is "medium," assigned a value of 2; and greater than 5 seconds is "slow," assigned a value of 1.

[0063] The search time of each pre-selected model is matched with the above-mentioned grade intervals to determine its corresponding search time grade and obtain the corresponding grade assignment. For example, the search time of the model based on the combined search of classification number and keywords is 2 seconds, which falls in the "fast" grade interval and is assigned a grade of 3; the search time of the model based on full-text semantic search is 4 seconds, which falls in the "medium" grade interval and is assigned a grade of 2.

[0064] Next, calculate the sum of the rank assignment and the relevance index of each pre-selected model, and record this sum as the model benefit value. For example, the model based on the joint retrieval of classification number and keyword has a relevance index of 1.3, a rank assignment of 3, and a model benefit value of 13.+3=4.3. The model based on full-text semantic retrieval has a relevance index of 1.5, a rank assignment of 2, and a model benefit value of 1.5+2=3.5.

[0065] Among all pre-selected models, the one with the highest benefit value is selected as the criterion and determined as the recommended model. In the above example, the benefit value of the model based on combined classification number and keyword search (4.3) is greater than the benefit value of the model based on full-text semantic search (3.5), so the model based on combined classification number and keyword search is determined as the recommended model.

[0066] Next, the retrieval question type corresponding to the recommendation model is obtained. For example, this recommendation model is primarily applicable to retrieval questions with clearly defined technical features and classification numbers. The retrieval question type and the recommendation model are combined to generate recommendation information. For example, if the retrieval question type is "retrieval based on specific technical features and classification numbers," the recommendation information could be expressed as "Recommended model based on combined retrieval of classification numbers and keywords, applicable to retrieval questions based on specific technical features and classification numbers."

[0067] Similarly, we obtain historical search data and comprehensively identify the search question types corresponding to all search models in the historical data. We then generate corresponding recommendations for each search question type according to the above process. This allows searchers to quickly select the most appropriate search model based on the generated recommendations for different search questions, greatly improving search efficiency and accuracy.

[0068] For example 2, please refer to Figure 2 This application provides a retrieval model recommendation device based on knowledge graph, including: information acquisition module, information analysis module, retrieval analysis module and retrieval determination module, and combined with Figure 2 It can be known that the functional modules are electrically connected in a unidirectional manner.

[0069] Information acquisition module, used to obtain multiple retrieval models and corresponding knowledge graphs;

[0070] An information analysis module is used to analyze a pre-acquired search question based on the knowledge graph, obtain search results corresponding to the search question, split the search question to obtain split information, and search the split information using different search models to obtain corresponding search results. Then, the module analyzes the content matching degree between the search question and the search results, and filters the search results based on the content matching degree to obtain pre-selected results. At the same time, the module ranks the search models according to the number of pre-selected results to generate model ranking information.

[0071] A retrieval analysis module is used to search for the retrieval question based on each retrieval model, obtain retrieval results corresponding to the retrieval model, obtain model ranking information, and screen the retrieval question to obtain keywords, while calculating the frequency of occurrence of the keywords, then calculating the semantic similarity between the retrieval question and the retrieval results, and calculating the sum of the numerical values ​​of the frequency of occurrence and the semantic similarity to obtain a pre-selected result association value, and obtain the result to be analyzed based on discrete value analysis, and calculate the average value of the sum of the numerical values ​​of the association values ​​of the result to be analyzed to obtain a association index;

[0072] The retrieval determination module compares the relevance index of the retrieval model with the threshold to obtain the pre-selected model, and determines the grade assignment of the pre-selected model based on the retrieval time. At the same time, it calculates the sum of the value of the pre-selected model and the relevance index to obtain the model benefit value, and selects the pre-selected model with the largest model benefit value to generate a recommended model, and generates recommendation information based on the retrieval type of the retrieval question.

[0073] Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0074] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. The retrieval model recommendation method based on knowledge graph is characterized by: The method specifically comprises the following steps: Build a knowledge graph, obtain all retrieval models and corresponding retrieval questions, and determine the question type corresponding to the retrieval question; The search question is split to obtain split information, and the split information is searched using different search models to obtain corresponding search results. Then, the content matching degree between the search question and the search results is analyzed, and the search results are screened based on the content matching degree to obtain pre-selected results. At the same time, the search models are ranked according to the number of pre-selected results to generate model ranking information; Obtain model ranking information, filter the search question to obtain keywords, and calculate the frequency of occurrence of keywords. Then calculate the semantic similarity between the search question and the search results. At the same time, calculate the sum of the frequency of occurrence and the semantic similarity to obtain the association value of the pre-selected results. Based on discrete value analysis, obtain the results to be analyzed, and calculate the average sum of the association values ​​of the results to be analyzed to obtain the association index. Compare the relevance index of the retrieval model with the threshold to obtain the pre-selected model, and determine the rank assignment of the pre-selected model based on the retrieval time. At the same time, calculate the sum of the rank assignment and the relevance index to obtain the model benefit value. Then select the pre-selected model with the largest model benefit value to generate the recommended model, and generate the recommended information based on the retrieval type of the retrieval question. Using the word vector model, the semantic similarity between the search question and the search results is calculated. The sum of the semantic similarity and the frequency of occurrence is calculated as the pre-selected result association value. The pre-selected results are then filtered according to the discrete value to obtain the results to be analyzed. At the same time, the mean of the sum of the association values ​​of the results to be analyzed is calculated and recorded as the association index between the model to be analyzed and the search question. Similarly, the relevance indexes of all retrieval models are calculated and sorted from large to small according to the relevance index; Obtain the correlation index corresponding to the retrieval model and compare it with the threshold. Filter the retrieval model with a correlation index greater than the threshold and record it as the pre-selected model. Then obtain the retrieval time corresponding to the pre-selected model, and divide the retrieval time into level intervals based on big data analysis, and perform corresponding assignment processing. At the same time, match the retrieval time of the pre-selected model with the level interval, determine the retrieval time level corresponding to the pre-selected model, and obtain the corresponding level assignment. Then calculate the sum of the level assignment and the correlation index, and record it as the model benefit value.

2. The retrieval model recommendation method based on knowledge graph according to claim 1 is characterized in that: The specific method of splitting the search question to obtain split information is: Obtain the search question and use the bag-of-words model to split the search question. The bag-of-words model will decompose the search question into independent and domain-meaningful words or phrases during splitting to obtain corresponding split information.

3. The retrieval model recommendation method based on knowledge graph according to claim 1 is characterized in that: The specific method of generating model ranking information is: Get all retrieval models and label them as i, where i = 1, 2, ..., j, where j represents the number of retrieval models. Get the retrieval result Ci corresponding to retrieval model i, get the split information, and calculate the proportion of the number of split information in the retrieval result Ci; And compare it with the preset value, and the search results with a screening quantity greater than the preset value are recorded as pre-selected results; Similarly, all pre-selected results corresponding to the retrieval model i are screened and sorted from large to small according to the number of screened results to generate model ranking information.

4. The retrieval model recommendation method based on knowledge graph according to claim 1 is characterized in that: The specific method of screening the search question to obtain keywords is: Obtain a search question, and use natural language processing tools to analyze the search question according to lexical rules, and combine the stop word filtering method to comprehensively process the search question, screen out the keywords corresponding to the search question, and label them as n, where n=1, 2, ..., m, where m represents the number of keywords.

5. The retrieval model recommendation method based on knowledge graph according to claim 1 is characterized in that: The specific method of calculating the occurrence frequency of keywords is: Perform sentence processing on the pre-selected result a to obtain sentence information, and calculate the number of times the keyword n ​​appears in the sentence information, and calculate the ratio of the number of appearances to the total number of sentence information; Similarly, calculate the proportion of keyword n ​​in all sentence information, and then sum the calculated proportions to obtain the frequency of occurrence of keyword n.

6. The retrieval model recommendation method based on knowledge graph according to claim 1 is characterized in that: The specific method of generating recommendation information based on the search type of the search question is: Select the pre-selected model with the largest model benefit value as the standard, generate the recommended model, obtain the corresponding search question type, and combine it with the recommended model to generate recommendation information; Similarly, historical data is obtained, and the search types corresponding to all deceleration models are obtained to generate corresponding recommendation information.

7. A knowledge graph-based retrieval model recommendation device, configured to execute the knowledge graph-based retrieval model recommendation method according to any one of claims 1 to 6, characterized in that: include: Information acquisition module, used to obtain multiple retrieval models and corresponding knowledge graphs; An information analysis module, configured to analyze the pre-acquired search questions based on the knowledge graph and obtain search results corresponding to the search questions; A retrieval analysis module, configured to perform a search for each retrieval model based on the retrieval model and the retrieval question, and obtain a retrieval result corresponding to the retrieval model; The retrieval determination module is used to determine the corresponding target retrieval model based on the retrieval results corresponding to all the retrieval models and generate recommendation information.

Images