A vertical search engine ranking method and system based on multi-dimensional features

Abstract

The present invention relates to the field of computer software technology, and in particular to a vertical search engine precision ranking method and system based on multi-dimensional features. The method comprises obtaining recalled resource information, performing feature extraction based on the obtained resource information, including extracting document features, resource heat features, column features, and user behavior features of the resource information; weighting a classification model according to the extracted resource information features, and weighting resource heat according to the extracted resource information features, including setting resource heat hyperparameters for adjusting the weight of resource heat; integrating the classification model weighted results and resource heat weighted results into a model, performing precision ranking scoring using the weights after model integration, and sorting the recalled resources according to the final scoring results. The present invention reasonably weights the search results by comprehensively considering multi-dimensional features, thereby achieving accurate search result display.

Description

Technical Field

[0001] The present invention relates to the field of computer software technology, and in particular to a vertical search engine ranking method and system based on multi-dimensional features. Background Art

[0002] Vertical search is a search engine technology specifically tailored to specific fields or topics. Unlike general search engines (such as Google and Baidu), vertical search engines focus on indexing and retrieving specific types of information, such as real estate, travel, shopping, government services, or data from specific industries. With the rapid development of information technology, search engines play a vital role in how people access information. However, traditional general search engines often fail to provide sufficiently accurate and relevant search results for searches in specific fields or topics. To address this problem, vertical search engines have emerged. These search engines focus on indexing and retrieving information in specific fields or topics, better meeting user search needs in these areas. For example, in areas such as real estate, travel, shopping, and government services, vertical search engines can provide more accurate and targeted search results.

[0003] In vertical search engines, the BM25 algorithm is a commonly used information retrieval and web page ranking algorithm. It evaluates the relevance of a document to a user query by considering the frequency of the query term appearing in the document and the frequency of its appearance in all documents. However, the traditional BM25 algorithm has some limitations in vertical search. For example, it does not consider factors such as resource popularity, column weight, and business focus. In addition, feature extraction is a key link in search engines. Effective feature extraction can capture the key information and potential semantics of a document, but existing feature extraction methods may not fully cover multi-dimensional features, resulting in the relevance and accuracy of search results being affected. At this stage, a vertical search engine ranking method and system based on multi-dimensional features is needed. Summary of the Invention

[0004] To address the issues of poor performance and low search accuracy in vertical search engines, the present invention provides a vertical search engine ranking method and system based on multi-dimensional features. This method comprehensively considers resource popularity, column weight, user behavior, and user profile factors, deeply mining document features to provide more accurate and personalized search results.

[0005] In a first aspect, the present invention provides a vertical search engine ranking method based on multi-dimensional features, which adopts the following technical solutions:

[0006] A vertical search engine ranking method based on multi-dimensional features, comprising:

[0007] Obtaining recalled resource information, including document information, resource popularity, column information, and user behavior data;

[0008] Perform feature extraction based on the acquired resource information, including extracting document features, resource popularity features, column features, and user behavior features of the resource information;

[0009] Performing classification model weighting based on the extracted resource information features, including calculating the classification model weighting by combining chi-square feature extraction and Bayesian classification;

[0010] Weighting resource heat based on the extracted resource information features, including setting resource heat hyperparameters to adjust the weight of resource heat;

[0011] The weighted results of the classification model and the weighted results of resource heat are integrated into the model, including using the weighted results of the classification model as the basic weight and combining them with the weighted results of resource heat through multiplication;

[0012] The weights after model integration are used for precise ranking and scoring, and the recalled resources are sorted according to the final scoring results.

[0013] Furthermore, the classification model is weighted based on the extracted resource information features, including calculating the frequency of occurrence and chi-square statistics of each feature in different categories, screening out features with discriminative ability, calculating the prior probability and posterior probability of each word in different categories through Bayesian classification, and selecting the category with the largest posterior probability as the category of the document. The calculation formula of the Bayesian classification is:

[0014] ,

[0015] in, Represented as keywords, Represented as categories, Represented as category The prior probability of Represented as in category Next, the word appears in the document The joint probability of Represents the word that appears in the document The posterior probability that the document belongs to the category in the case of .

[0016] Furthermore, the weighting of the classification model based on the extracted resource information features also includes setting a column coefficient based on the importance of the column information to the business, and obtaining a preliminary weight for each document by multiplying the Bayesian posterior probability with the column coefficient.

[0017] Furthermore, the resource heat weighting is performed according to the extracted resource information features, including obtaining the search log statistics of the recent resource visits, and setting the resource heat hyperparameters to adjust the weight of the resource heat, and calculating the resource heat weight based on the recent resource visits and the heat hyperparameters. The calculation formula is:

[0018] ,

[0019] in, Expressed as a resource heat hyperparameter, Represents the number of recent visits to the resource. Indicates the weight after resource heat is weighted.

[0020] Furthermore, the classification model weighted results and resource heat weighted results are integrated into a model, including the initial weight obtained by the classification model weighting As the basic weight, the result of multiplication combined with resource heat weighting The final weight after integration is obtained, and the final weight is expressed as:

[0021] ,

[0022] in, Represented as the weight of the classification model after weighting, is represented as the final weight.

[0023] Furthermore, the weights after model integration are used to perform fine ranking and scoring, including calculating the scores in the rough ranking results according to the matching rules, calculating the final fine ranking scores using the weights after model integration, and sorting the recalled resources according to the final scoring results. The formula for calculating the final fine ranking score is:

[0024] ,

[0025] in, It is expressed as the score in the rough sorting result. Indicates the final refined ranking score.

[0026] Furthermore, the score in the rough ranking results is calculated according to the matching rules, including selecting the matching rules and determining the matching score according to the matching rules, converting the documents and keywords into vectors using word frequency-inverse document frequency, calculating the relevance score through cosine similarity, determining the score in the rough ranking results according to the matching score and the relevance score, setting a score threshold and eliminating resource information with a score less than the threshold.

[0027] In the second aspect, a vertical search engine ranking system based on multi-dimensional features includes:

[0028] The data acquisition module is configured to: acquire user resource information, including document information, resource popularity, column information and user behavior data;

[0029] The feature extraction module is configured to: perform feature extraction based on the acquired resource information, including extracting document features, resource popularity features, column features, and user behavior features of the resource information;

[0030] The preliminary weighting module is configured to: perform classification model weighting according to the extracted resource information features, including calculating the classification model weighting by combining chi-square feature extraction and Bayesian classification;

[0031] A heat weighting module is configured to: perform resource heat weighting based on the extracted resource information features, including setting resource heat hyperparameters to adjust the weight of resource heat;

[0032] The integration module is configured to: perform model integration on the weighted results of the classification model and the weighted results of the resource heat, including using the weighted results of the classification model as a basic weight and combining the results of the resource heat weight through a multiplication operation;

[0033] The refined ranking module is configured to use the weights after model integration to perform refined ranking scoring, and sort the recalled resources according to the final scoring results.

[0034] In a third aspect, the present invention provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, for a vertical search engine ranking method based on multi-dimensional features.

[0035] In a fourth aspect, the present invention provides a terminal device comprising a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded and executed by the processor as a vertical search engine ranking method based on multi-dimensional features.

[0036] In summary, the present invention has the following beneficial technical effects:

[0037] 1. The present invention can more accurately determine the category of a document by calculating the frequency of occurrence of features, chi-square statistics and Bayesian classification, thereby improving the accuracy and reliability of classification. By setting a column coefficient and multiplying it with the Bayesian posterior probability, the importance of the column to the business can be fully considered, making the classification results more in line with actual business needs. Based on the search log statistics of recent resource visits and heat weighting, popular resources can be given more advantages in sorting and meet users' needs for popular information.

[0038] 2. The present invention reasonably weights the search results by comprehensively considering document features, resource popularity features, column features and user behavior features, thereby achieving accurate search result display and avoiding the problem of low search accuracy caused by the inability to fully cover multi-dimensional features. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a schematic diagram of the overall process of a vertical search engine ranking method based on multi-dimensional features according to an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The present invention will be further described in detail below with reference to the accompanying drawings.

[0041] Example 1

[0042] Reference Figure 1 The embodiment of the present invention provides a vertical search engine ranking method based on multi-dimensional features, including:

[0043] Obtaining recalled resource information, including document information, resource popularity, column information, and user behavior data;

[0044] Perform feature extraction based on the acquired resource information, including extracting document features, resource popularity features, column features, and user behavior features of the resource information;

[0045] Performing classification model weighting based on the extracted resource information features, including calculating the classification model weighting by combining chi-square feature extraction and Bayesian classification;

[0046] Weighting resource heat based on the extracted resource information features, including setting resource heat hyperparameters to adjust the weight of resource heat;

[0047] The weighted results of the classification model and the weighted results of resource heat are integrated into the model, including using the weighted results of the classification model as the basic weight and combining them with the weighted results of resource heat through multiplication;

[0048] The weights after model integration are used for precise ranking and scoring, and the recalled resources are sorted according to the final scoring results.

[0049] Specifically, a vertical search engine ranking method based on multi-dimensional features in this embodiment includes the following steps:

[0050] S1. Obtain the recalled resource information, including document information, resource popularity, column information, and user behavior data;

[0051] Based on the user's search intent and related conditions, determine the scope of resources to be searched, such as a specific website, database, or index. Extract keywords from the user's search query. These keywords will be used for matching and retrieval in the resources. Use the extracted keywords to search within the determined search scope to find resources related to the keywords. According to preset conditions and rules, filter the retrieved resources to remove irrelevant or non-compliant resources. For the filtered resources, collect relevant information, including document content, resource popularity, column information, user behavior data, and user portrait data.

[0052] S2. Extracting features based on the acquired resource information, including extracting document features, resource popularity features, column features, and user behavior features of the resource information;

[0053] Among them, document features include key words or phrases in the document, resource popularity features include visit and click volume, column features include column categories such as news, entertainment, technology, etc., the level of the column hierarchy, whether it is a first-level column or a second-level or third-level column, etc., user behavior features include search keywords, browsing history and preferences, etc.

[0054] S3. performing classification model weighting based on the extracted resource information features, including calculating the classification model weighting by combining chi-square feature extraction and Bayesian classification;

[0055] First, we need to classify the documents and determine different categories. Then, for each feature, we need to count the number of times it appears in each category. That is, we need to calculate the frequency of occurrence of each feature in different categories. We need to calculate the chi-square statistic for each feature according to the chi-square statistic calculation formula. The chi-square statistic calculation formula is:

[0056] ,

[0057] in, is the observed frequency, that is, the number of times the characteristic actually occurs in the category; is the expected frequency, which is the number of times the feature is expected to appear in the category based on the overall distribution. The expected frequency can be calculated by the total sample size, row sum, and column sum:

[0058] ,

[0059] The row sum is the sum of the observed frequencies for a specific row (e.g., a category), and the column sum is the sum of the observed frequencies for a specific column (e.g., a feature). Features with high discriminative power are selected based on the size of the chi-square statistic. A larger chi-square statistic indicates a stronger ability of the feature to distinguish between different categories. By selecting these features, the accuracy and effectiveness of subsequent Bayesian classification can be improved.

[0060] For example, in a vertical search engine, suppose we have two categories: technology and entertainment. For a feature called "artificial intelligence," we count its occurrence in both technology and entertainment documents to obtain the observed frequency. Then, based on the total sample size and the number of documents in each category, we calculate the expected frequency. By calculating the chi-square statistic, we can determine the ability of the feature "artificial intelligence" to distinguish between technology and entertainment documents. A large chi-square statistic indicates a significant difference in its distribution between the two categories, indicating that it is a highly discriminatory feature and can be used for subsequent classification.

[0061] Based on the above-screened discriminative features, the prior probability and posterior probability of each word with discriminative features in different categories are calculated through Bayesian classification, and the category with the largest posterior probability is selected as the category of the document. The calculation formula of the Bayesian classification is:

[0062] ,

[0063] in, Represented as keywords, Represented as categories, Represented as category The prior probability of Represented as in category Next, the word appears in the document The joint probability of Represents the word that appears in the document In the case of , the posterior probability that the document belongs to the category, Represented as category The prior probability of a category can be estimated by counting the frequency of its occurrence in all documents. For example, if 40% of all documents are technology documents and 60% are entertainment documents, then the prior probabilities are P(technology) = 0.4 and P(entertainment) = 0.6. For the words "artificial intelligence" and "algorithm" in a new document, the conditional probabilities in the technology category are calculated to be P(artificial intelligence | technology) = 0.7 and P(algorithm | technology) = 0.6, and in the entertainment category are calculated to be P(artificial intelligence | entertainment) = 0.2 and P(algorithm | entertainment) = 0.1. The posterior probabilities in the technology category are: P(technology | artificial intelligence, algorithm). P(Technology) P(Artificial Intelligence | Technology) P(Algorithm|Technology) 0.4 0.7 0.6. Similarly, calculate the posterior distribution of entertainment. Obviously, the technology category is larger than the entertainment category, so the document is classified as technology.

[0064] Calculate the Bayesian posterior probability, which represents the probability that a document belongs to a certain category and reflects the probability distribution of the document in each category. Then, set the column coefficient based on the importance of the column to the business. Different columns may have different degrees of importance in the business. By setting the column coefficient, this difference in importance can be reflected. The column coefficient is set by collecting document data, classifying and labeling this data, recording the column where each document is located, and determining the indicators for evaluating the importance of the column based on business goals and actual needs. A certain score range is set for each indicator, that is, the indicator is quantified. The quantified indicator is normalized to obtain the column coefficient, and then weighted by calculating the classification model:

[0065] ,

[0066] in, represents the probability that a document belongs to a certain category, Expressed as column coefficient, This represents the initial weight of each document, which takes into account both its classification probability and the column coefficient. The classification probability indicates its likelihood of belonging to different categories, while the column coefficient reflects the importance of the column to the business. By multiplying the two, the resulting initial weight better reflects the document's importance within the overall classification system, providing a foundation for subsequent weighting. In subsequent steps, such as resource popularity weighting, this initial weight is further adjusted to achieve more accurate search result sorting and recommendations.

[0067] S4. Weighting the resource heat according to the extracted resource information features, including setting resource heat hyperparameters to adjust the weight of the resource heat;

[0068] First, the recent resource visits are counted through search logs. This means that the system will record the number of times or frequency that each resource is accessed by users within a period of time (such as the last week or month). This statistic can intuitively reflect the user's attention to and demand for different resources. Then, a resource heat hyperparameter is set to control the importance of resource heat in the entire weight calculation. By analyzing the resource access situation and user behavior in historical data, the hyperparameter value range is determined. For example, if it is found that the resource visits are usually concentrated within a certain range, the hyperparameter is set within this range. By adjusting the value of this hyperparameter to adapt to different application scenarios and business needs, the resource heat weighted calculation is performed based on the recent resource visits and the heat hyperparameter. The calculation formula is:

[0069] ,

[0070] in, Expressed as a resource heat hyperparameter, Represents the number of recent visits to the resource. This is the weighted value of the resource's popularity. As the input value (i.e., the resource's recent visits) increases, the output value (i.e., the weight) increases more slowly. This means that popular resources (i.e., those with high visits) will be given a relatively higher weight, but without causing the weight to increase too dramatically, thus preventing a small number of extremely popular resources from completely dominating the ranking.

[0071] S5. Integrate the classification model weighted results and resource heat weighted results into a model, including using the classification model weighted results as the basic weights and combining them with the resource heat weighted results through multiplication operations;

[0072] Combine the results of resource popularity weighting with the results of classification model weighting. Classification model weighting is usually calculated based on the resource category, attributes and other information. After combining the two, the weight of each document takes into account the classification characteristics of the resource and its popularity. Such comprehensive consideration makes it more likely that popular resources will be ranked at the top in the sorting, because they may not only be related to user needs in terms of classification, but also show a high degree of popularity in actual user access behavior. The preliminary weight obtained by the classification model weighting is As the basic weight, the result of multiplication combined with resource heat weighting The final weight after integration is obtained, and the final weight is expressed as:

[0073] ,

[0074] in, Represented as the weight of the classification model after weighting, is represented as the final weight.

[0075] For example, suppose a document receives a weight of 0.6 from the classification model weighting and a weight of 0.8 from the resource popularity weighting. Combining the two, the new weight might be 0.6 × 0.8 = 0.48. If another document receives a weight of 0.5 from the classification model weighting and 0.5 from the resource popularity weighting, the new weight would be 0.5 × 0.5 = 0.25. During ranking, the first document would be ranked higher due to its higher combined weight. Even though it doesn't necessarily have an advantage over the second document based on the classification model weighting, its higher popularity ultimately gives it a better ranking position.

[0076] S6. Use the weights after model integration to perform precise ranking and scoring, and sort the recalled resources according to the final scoring results.

[0077] Select matching rules and determine the matching score based on the matching rules, including exact matching, fuzzy matching, synonym matching, and considering the position and part of speech of the keyword in the document. If the keyword completely matches a piece of text in the document, it is recorded as a complete match score. , using the edit distance algorithm to measure similarity, when the similarity threshold is 0.8, the score , there is a synonym match, score , keywords appear in the title, position score , appears at the beginning of the document, position score , the noun weight is 1.5, the verb weight is 1.2, and the document and keyword are converted into vectors using word frequency-inverse document frequency, where word frequency-inverse document frequency is expressed as:

[0078] Term Frequency - Inverse Document Frequency ,

[0079] ,

[0080] in, is the frequency of word j in document i, is the inverse document frequency, N is the total number of documents, is the number of documents containing word j. In the bag-of-words model, each element of the document vector represents the frequency of the corresponding word in the document. The term frequency-inverse document frequency (TF-IDF) focuses more on the ability to distinguish words. The term frequency (TF) calculates the frequency of a word in a document, while the inverse document frequency (IDF) measures the prevalence of a word in the entire document collection. For example, a common word that appears in many documents has a low IDF value; while a word that appears in only a few documents has a high IDF value. The TF-IDF value is obtained by multiplying TF and IDF to construct a document vector. The construction of the keyword vector is the same as that of the document vector, both using term frequency-inverse document frequency. The relevance score of the keyword vector and the document vector is calculated through cosine similarity. The score in the rough sorting result is obtained by adding the matching score and the relevance score. , set a score threshold and remove resource information that is less than the threshold. The score threshold is all The average of the scores is used to calculate the final refined ranking score using the weights integrated in S5, and the recalled resources are sorted according to the final scoring result. The formula for calculating the final refined ranking score is:

[0081] ,

[0082] in, It is expressed as the score in the rough sorting result. Indicates the final refined ranking score, based on the final scoring result , the recalled resources are sorted, and the resources with higher scores are considered to have a higher degree of matching with user needs, and are placed at the front of the search results.

[0083] Example 2

[0084] The difference between this embodiment and embodiment 1 is that this embodiment provides a vertical search engine ranking system based on multi-dimensional features.

[0085] The data acquisition module is configured to: acquire user resource information, including document information, resource popularity, column information and user behavior data;

[0086] The feature extraction module is configured to: perform feature extraction based on the acquired resource information, including extracting document features, resource popularity features, column features, and user behavior features of the resource information;

[0087] The preliminary weighting module is configured to: perform classification model weighting according to the extracted resource information features, including calculating the classification model weighting by combining chi-square feature extraction and Bayesian classification;

[0088] A heat weighting module is configured to: perform resource heat weighting based on the extracted resource information features, including setting resource heat hyperparameters to adjust the weight of resource heat;

[0089] The integration module is configured to: perform model integration on the weighted results of the classification model and the weighted results of the resource heat, including using the weighted results of the classification model as a basic weight and combining the results of the resource heat weight through a multiplication operation;

[0090] The refined ranking module is configured to use the weights after model integration to perform refined ranking scoring, and sort the recalled resources according to the final scoring results.

[0091] A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, for a vertical search engine precise ranking method based on multi-dimensional features.

[0092] A terminal device includes a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, wherein the instructions are suitable for being loaded and executed by the processor, a vertical search engine ranking method based on multi-dimensional features.

[0093] The above are all preferred embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A vertical search engine ranking method based on multi-dimensional features, characterized in that: include: Obtaining recalled resource information, including document information, resource popularity, column information, and user behavior data; Perform feature extraction based on the acquired resource information, including extracting document features, resource popularity features, column features, and user behavior features of the resource information; Performing classification model weighting based on the extracted resource information features, including calculating the classification model weighting by combining chi-square feature extraction and Bayesian classification; Weighting resource heat based on the extracted resource information features, including setting resource heat hyperparameters to adjust the weight of resource heat; The weighted results of the classification model and the weighted results of resource heat are integrated into the model, including using the weighted results of the classification model as the basic weight and combining them with the weighted results of resource heat through multiplication; Use the weights after model integration to perform precise ranking and scoring, and sort the recalled resources according to the final scoring results; The classification model is weighted based on the extracted resource information features, including calculating the frequency of occurrence and chi-square statistics of each feature in different categories, screening out features with discriminative ability, calculating the prior probability and posterior probability of each word in different categories through Bayesian classification, and selecting the category with the largest posterior probability as the category of the document. The calculation formula of the Bayesian classification is: ， in, Represented as keywords, Represented as categories, Represented as category The prior probability of Represented as in category Next, the word appears in the document The joint probability of Represents the word that appears in the document , the posterior probability that the document belongs to the category; The weighting of the classification model according to the extracted resource information features also includes setting a column coefficient according to the importance of the column information to the business, and obtaining a preliminary weight for each document by multiplying the Bayesian posterior probability and the column coefficient; The resource heat weighting according to the extracted resource information features includes obtaining the search log statistics of the recent resource visits and setting the resource heat hyperparameters to adjust the resource heat weight. The resource heat weighting is calculated based on the recent resource visits and the heat hyperparameters. The calculation formula is: ， in, Expressed as a resource heat hyperparameter, Represents the number of recent visits to the resource. It is expressed as the weight after resource heat is weighted; The classification model weighted results and resource heat weighted results are integrated into the model, including the initial weight obtained by the classification model weighting As the basic weight, the result of multiplication combined with resource heat weighting The final weight after integration is obtained, and the final weight is expressed as: , in, Represented as the weight of the classification model after weighting, Expressed as the final weight; The refined ranking score is performed using the weights after model integration, including calculating the scores in the rough ranking results according to the matching rules, calculating the final refined ranking score using the weights after model integration, and sorting the recalled resources according to the final scoring results. The formula for calculating the final refined ranking score is: , in, It is expressed as the score in the rough sorting result. Indicates the final refined ranking score; The method of calculating the scores in the rough ranking results according to the matching rules includes selecting the matching rules and determining the matching scores according to the matching rules, converting the documents and keywords into vectors using word frequency-inverse document frequency, calculating the relevance scores through cosine similarity, determining the scores in the rough ranking results according to the matching scores and the relevance scores, setting a score threshold and eliminating resource information with a score less than the threshold.

2. A vertical search engine ranking system based on multi-dimensional features, characterized by: include: The data acquisition module is configured to: obtain user resource information, including document information, resource popularity, column information and user behavior data; The feature extraction module is configured to: perform feature extraction based on the acquired resource information, including extracting document features, resource popularity features, column features, and user behavior features of the resource information; The preliminary weighting module is configured to: perform classification model weighting according to the extracted resource information features, including calculating the classification model weighting by combining chi-square feature extraction and Bayesian classification; A heat weighting module is configured to: perform resource heat weighting based on the extracted resource information features, including setting resource heat hyperparameters to adjust the weight of resource heat; The integration module is configured to: perform model integration on the weighted results of the classification model and the weighted results of the resource heat, including using the weighted results of the classification model as a basic weight and combining the results of the resource heat weight through a multiplication operation; The refined ranking module is configured to use the weights after model integration to perform refined ranking scoring and sort the recalled resources according to the final scoring results; The classification model is weighted based on the extracted resource information features, including calculating the frequency of occurrence and chi-square statistics of each feature in different categories, screening out features with discriminative ability, calculating the prior probability and posterior probability of each word in different categories through Bayesian classification, and selecting the category with the largest posterior probability as the category of the document. The calculation formula of the Bayesian classification is: ， in, Represented as keywords, Represented as categories, Represented as category The prior probability of Represented as in category Next, the word appears in the document The joint probability of Represents the word that appears in the document , the posterior probability that the document belongs to the category; The weighting of the classification model according to the extracted resource information features also includes setting a column coefficient according to the importance of the column information to the business, and obtaining a preliminary weight for each document by multiplying the Bayesian posterior probability and the column coefficient; The resource heat weighting according to the extracted resource information features includes obtaining the search log statistics of the recent resource visits and setting the resource heat hyperparameters to adjust the resource heat weight. The resource heat weighting is calculated based on the recent resource visits and the heat hyperparameters. The calculation formula is: ， in, Expressed as a resource heat hyperparameter, Represents the number of recent visits to the resource. It is expressed as the weight after resource heat is weighted; The classification model weighted results and resource heat weighted results are integrated into the model, including the initial weight obtained by the classification model weighting As the basic weight, the result of multiplication combined with resource heat weighting The final weight after integration is obtained, and the final weight is expressed as: , in, Represented as the weight of the classification model after weighting, Expressed as the final weight; The refined ranking score is performed using the weights after model integration, including calculating the scores in the rough ranking results according to the matching rules, calculating the final refined ranking score using the weights after model integration, and sorting the recalled resources according to the final scoring results. The formula for calculating the final refined ranking score is: , in, It is expressed as the score in the rough sorting result. Indicates the final refined ranking score; The method of calculating the scores in the rough ranking results according to the matching rules includes selecting the matching rules and determining the matching scores according to the matching rules, converting the documents and keywords into vectors using word frequency-inverse document frequency, calculating the relevance scores through cosine similarity, determining the scores in the rough ranking results according to the matching scores and the relevance scores, setting a score threshold and eliminating resource information with a score less than the threshold.

3. A computer-readable storage medium storing a plurality of instructions, characterized in that: The instructions are suitable for being loaded and executed by a processor of a terminal device, a vertical search engine ranking method based on multi-dimensional features as claimed in claim 1.

4. A terminal device comprising a processor and a computer-readable storage medium, wherein the processor is configured to implement various instructions; and the computer-readable storage medium is configured to store a plurality of instructions, wherein: The instructions are suitable for being loaded by a processor and executed by a vertical search engine ranking method based on multi-dimensional features as claimed in claim 1.

Images