Search engine optimization method, device, equipment and storage medium

Abstract

The application belongs to the technical field of search engines, and discloses a search engine optimization method, device, equipment and storage medium, the method obtains a preset query statement and historical non-displayed content of a search engine; according to a preset language model after training, feature extraction is performed on the historical non-displayed content and the preset query statement to obtain the correlation between the historical non-displayed content and the preset query statement; a query statement candidate set is generated according to the query statement corresponding to the historical non-displayed content; the query statement candidate set is filtered according to the correlation to obtain a target query statement, and the search engine is optimized according to the target query statement. In the application, the target query statement is generated for the historical non-displayed content in the search engine, the score of the historical non-displayed content in the search system retrieval recall stage is improved, the recall probability of the content can be improved, the retrieval result is more accurate, and the search engine is optimized.

Description

Technical Field

[0001] This invention relates to the field of search engine technology, and in particular to a search engine optimization method, apparatus, device, and storage medium. Background Technology

[0002] General search engines should cover as much content as possible across the entire internet. Based on the current design logic of search systems, some content is never retrieved. This un-displayed content will not be clicked by users, resulting in lower scores for this content during the search system's retrieval and recall phase, making it even more difficult to retrieve. Existing methods for improving recall rates have issues such as generating ambiguity and forcibly injecting additional information. These problems can lead to incorrect recalls and inaccurate search results.

[0003] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main objective of this invention is to provide a search engine optimization method, apparatus, device, and storage medium, which aims to solve the technical problems of low recall probability of undisplayed content in existing search engines and inaccurate search results caused by existing methods to improve recall rate.

[0005] To achieve the above objectives, the present invention provides a search engine optimization method, the method comprising the following steps:

[0006] Retrieve the search engine's default query statements and historical content that has not been displayed.

[0007] Based on the trained preset language model, feature extraction is performed on the historical undisplayed content and the preset query statement to obtain the relevance between the historical undisplayed content and the preset query statement;

[0008] Generate a candidate set of query statements based on the query statements corresponding to the historical content that was not displayed.

[0009] The candidate query statement is filtered based on the relevance to obtain the target query statement, and the search engine is optimized based on the target query statement.

[0010] Optionally, before the step of extracting features from the historically undisplayed content and the preset query statement based on the trained preset language model to obtain the relevance between the historically undisplayed content and the preset query statement, the method further includes:

[0011] Build a preprocessed dataset;

[0012] The preset language model is trained using a ternary loss function and the preprocessed dataset to obtain the trained preset language model.

[0013] Optionally, the step of constructing the preprocessed dataset includes:

[0014] Obtain the preset display content to be trained and the corresponding positive sample;

[0015] Negative example samples corresponding to the preset display content to be trained are obtained according to a preset random algorithm;

[0016] A preprocessed dataset is constructed based on the preset display content to be trained, the positive examples, and the negative examples.

[0017] Optionally, before the step of training the preset language model using a ternary loss function and the preprocessed dataset to obtain the trained preset language model, the method further includes:

[0018] Construct sample triples based on the preprocessed dataset;

[0019] Construct a ternary loss function based on the sample triples.

[0020] Optionally, the step of constructing a ternary loss function based on the sample triples includes:

[0021] The training preset display content, the positive sample, and the negative sample are obtained based on the sample triples;

[0022] Determine the positive distance between the preset display content to be trained and the positive sample;

[0023] Determine the negative example distance between the preset display content to be trained and the negative example samples;

[0024] The ternary loss function is determined based on the positive example distance, the negative example distance, and the preset interval.

[0025] Optionally, before the step of obtaining the search engine's preset query statement and historically undisplayed content, the method further includes:

[0026] Obtain the preset corpus of the search engine, and obtain the initial query statement based on the preset corpus;

[0027] The initial query statement is deduplicated to obtain the deduplicated query statement;

[0028] The deduplicated query statement is used as the preset query statement.

[0029] Optionally, the step of deduplicating the initial query statement to obtain a deduplicated query statement includes:

[0030] Obtain the string distance between the initial query statements respectively;

[0031] Determine whether the distance between the strings is less than a preset first distance threshold;

[0032] The initial query statements corresponding to strings whose distance is less than the preset first distance threshold are deleted to obtain duplicate query statements.

[0033] Optionally, the step of deduplicating the initial query statement to obtain a deduplicated query statement includes:

[0034] Obtain the edit distance between the initial query statements respectively;

[0035] Determine whether the edit distance is less than a preset second distance threshold;

[0036] The initial query statements corresponding to the edit distance being less than the preset second distance threshold are deleted to obtain the deduplicated query statements.

[0037] Optionally, the step of extracting features from the historically undisplayed content and the preset query statement based on the trained preset language model to obtain the relevance between the historically undisplayed content and the preset query statement includes:

[0038] Based on the trained preset language model, feature extraction is performed on the historical undisplayed content and the preset query statement to obtain the content feature vector of the historical undisplayed content and the semantic feature vector of the preset query statement;

[0039] The relevance between the historical undisplayed content and the preset query statement is determined based on the content feature vector and the semantic feature vector.

[0040] Optionally, the step of extracting features from the historically undisplayed content and the preset query statement based on the trained preset language model to obtain the content feature vector of the historically undisplayed content and the semantic feature vector of the preset query statement includes:

[0041] Based on the trained preset language model, feature extraction is performed on the historically undisplayed content to obtain the content feature vector of the historically undisplayed content;

[0042] Based on the trained preset language model, feature extraction is performed on the preset query statement to obtain the feature vectors of each word segment and each word group in the preset query statement.

[0043] The semantic feature vector of the preset query statement is constructed based on the feature vectors of each word segment and the feature vectors of each word group.

[0044] Optionally, the step of constructing the semantic feature vector of the preset query statement based on the word segmentation feature vector and the phrase feature vector includes:

[0045] The feature vectors of each word segmentation and each word group are summed to obtain the feature vector sum.

[0046] The semantic feature vector of the preset query statement is constructed based on the average value of the feature vectors.

[0047] Optionally, the step of filtering the candidate query statement set according to the relevance to obtain the target query statement, and optimizing the search engine based on the target query statement, includes:

[0048] The candidate query statement is filtered based on the relevance to obtain the target query statement;

[0049] Update the query statement candidate set according to the target query statement to obtain the target query statement set corresponding to the historical undisplayed content;

[0050] Optimize the search engine based on the target query set.

[0051] Optionally, the step of filtering the candidate query statement set according to the relevance to obtain the target query statement includes:

[0052] Construct the current semantic feature vector of the query statements in the query statement candidate set;

[0053] Calculate the similarity between the semantic feature vector of the preset query statement and the current semantic feature vector;

[0054] The target query statement is selected from the candidate query statement set based on the relevance and similarity.

[0055] Optionally, the step of filtering the target query statement from the candidate query statement set based on the relevance and the similarity includes:

[0056] Determine whether the similarity is greater than the relevance;

[0057] When the similarity is greater than the relevance, the candidate query statements in the query statement candidate set corresponding to the similarity are determined;

[0058] The candidate query statement is used as the target query statement.

[0059] Furthermore, to achieve the above objectives, the present invention also proposes a search engine optimization device, the search engine optimization device comprising:

[0060] The acquisition module is used to acquire the search engine's preset query statements and historical content that has not been displayed.

[0061] The extraction module is used to extract features from the historical undisplayed content and the preset query statement based on the trained preset language model, so as to obtain the relevance between the historical undisplayed content and the preset query statement;

[0062] The generation module is used to generate a candidate set of query statements based on the query statements corresponding to the historically undisplayed content;

[0063] The filtering module is used to filter the candidate set of query statements according to the relevance to obtain the target query statement, and optimize the search engine according to the target query statement.

[0064] Optionally, the search engine optimization device further includes: a training module;

[0065] The training module is used to construct the preprocessed dataset;

[0066] The preset language model is trained using a ternary loss function and the preprocessed dataset to obtain the trained preset language model.

[0067] Optionally, the training module is further configured to obtain the preset display content to be trained and the positive sample corresponding to the preset display content to be trained;

[0068] Negative example samples corresponding to the preset display content to be trained are obtained according to a preset random algorithm;

[0069] A preprocessed dataset is constructed based on the preset display content to be trained, the positive examples, and the negative examples.

[0070] Optionally, the training module is further configured to construct sample triples based on the preprocessed dataset;

[0071] Construct a ternary loss function based on the sample triples.

[0072] Furthermore, to achieve the above objectives, the present invention also proposes a search engine optimization device, which includes: a memory, a processor, and a search engine optimization program stored in the memory and executable on the processor, wherein the search engine optimization program is configured to implement the search engine optimization method described above.

[0073] Furthermore, to achieve the above objectives, the present invention also proposes a storage medium storing a search engine optimization program, which, when executed by a processor, implements the steps of the search engine optimization method described above.

[0074] This invention obtains preset query statements and historical undisplayed content from a search engine; extracts features from the historical undisplayed content and the preset query statements using a trained preset language model to obtain the relevance between them; generates a candidate set of query statements based on the query statements corresponding to the historical undisplayed content; filters the candidate set of query statements based on the relevance to obtain the target query statement; and optimizes the search engine based on the target query statement. In this invention, generating target query statements from historical undisplayed content in the search engine improves the score of this content during the search system's retrieval and recall phase, increasing the probability of content being recalled and making search results more accurate. This achieves search engine optimization, thus solving the technical problems of low recall probability of undisplayed content in existing search engines and inaccurate search results caused by existing methods of improving recall rates. Attached Figure Description

[0075] Figure 1 This is a schematic diagram of the structure of a search engine optimization device for the hardware operating environment involved in the embodiments of the present invention;

[0076] Figure 2 This is a flowchart illustrating the first embodiment of the search engine optimization method of the present invention;

[0077] Figure 3 This is a flowchart illustrating the second embodiment of the search engine optimization method of the present invention;

[0078] Figure 4 This is a schematic diagram illustrating the training of a preset language model using a ternary loss function according to an embodiment of the search engine optimization method of the present invention.

[0079] Figure 5 This is a flowchart illustrating the third embodiment of the search engine optimization method of the present invention;

[0080] Figure 6 This is a structural block diagram of the first embodiment of the search engine optimization device of the present invention.

[0081] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0082] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0083] Reference Figure 1 , Figure 1 This is a schematic diagram of the hardware operating environment of the search engine optimization device involved in the embodiment of the present invention.

[0084] like Figure 1As shown, the search engine optimization device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen and an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0085] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on search engine optimization equipment and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0086] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a network communication module, a user interface module, and a search engine optimization program.

[0087] exist Figure 1 In the search engine optimization device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the search engine optimization device of the present invention can be set in the search engine optimization device, and the search engine optimization device calls the search engine optimization program stored in the memory 1005 through the processor 1001 and executes the search engine optimization method provided in the embodiment of the present invention.

[0088] This invention provides a search engine optimization method, referring to... Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of a search engine optimization method according to the present invention.

[0089] In this embodiment, the search engine optimization method includes the following steps:

[0090] Step S10: Obtain the search engine's default query statement and historical undisplayed content.

[0091] It should be noted that the execution entity in this embodiment is the search engine optimization device, which can be an electronic device such as a personal computer or server; this embodiment does not limit this. The system obtains the search engine's preset query statements and historically undisplayed content. Regarding the historically undisplayed content, general search engines should cover as much content as possible across the entire network. Based on the design logic of current search engine systems, some content has never been retrieved. This undisplayed content will not be clicked by users, resulting in lower scores during the search system's retrieval and recall phase, making it even more difficult to retrieve. Therefore, content that has never been retrieved can be considered as historically undisplayed content in this embodiment.

[0092] It's easy to understand that the default corpus of a search engine contains a large number of repetitive or highly similar statements due to the inclusion of various forms and contents. These statements are essentially of the same type, or even variations of the same statement, with little difference in essence. When a statement is retrieved as similar to the query statement, theoretically, all these highly similar statements will be retrieved simultaneously. This results in retrieved statements that are nearly identical to the query statement, but with a high workload because the same retrieval procedure must be used to retrieve each of these nearly identical statements. This situation only increases the workload for similar statement retrieval and does not make the search results richer or more comprehensive. To improve this, this embodiment, after obtaining the initial query statement from the search engine's default corpus, performs deduplication on the initial query statement and uses the deduplicated query statement as the default query statement, making the search results richer and more comprehensive.

[0093] Specifically, the steps for obtaining a preset query statement can be as follows: obtaining a preset corpus of the search engine, and obtaining an initial query statement based on the preset corpus; performing deduplication on the initial query statement to obtain a deduplicated query statement; and using the deduplicated query statement as the preset query statement.

[0094] It should be understood that the deduplication process can be performed in any of the following two ways, or in other ways, and this embodiment does not limit this.

[0095] The first exemplary implementation involves obtaining the string distance between the initial query statements; determining whether the string distance is less than a preset first distance threshold; and removing the initial query statements whose string distance is less than the preset first distance threshold to obtain deduplicated query statements. Here, string distance refers to Dice distance, which measures the similarity between two sets. The strings contained in the statements can be understood as a set of characters, and Dice distance is also used to measure string similarity.

[0096] The second exemplary implementation involves obtaining the edit distance between the initial query statements; determining whether the edit distance is less than a preset second distance threshold; and deleting the initial query statements whose edit distance is less than the preset second distance threshold to obtain deduplicated query statements. Here, edit distance represents the minimum number of edits required to transform one string into another. In this embodiment, editing involves replacing one character in a string with another, or inserting or deleting characters. Edit distance directly measures the similarity between two texts. When comparing two statements, it considers word order and word overlap, effectively identifying statements that differ by only one character or punctuation mark.

[0097] It should be noted that the above deduplication process can be performed when building the preset corpus of the search engine, or when constructing the semantic feature vectors of each statement in the preset corpus of the search engine. This application does not limit the specific timing of the above deduplication process.

[0098] Step S20: Extract features from the historical undisplayed content and the preset query statement based on the trained preset language model to obtain the relevance between the historical undisplayed content and the preset query statement.

[0099] The process is straightforward: Features are extracted from the historically undisplayed content using a trained pre-defined language model to obtain content feature vectors; features are also extracted from the pre-defined query statement using the same model to obtain word segmentation feature vectors and phrase feature vectors; a semantic feature vector is constructed based on these word segmentation and phrase feature vectors; and the relevance between the historically undisplayed content and the pre-defined query statement is determined using the content feature vectors and semantic feature vectors.

[0100] It should be understood that the pre-trained language model can be the trained BERT model. The trained BERT model is used to extract features from the historically undisplayed content and the deduplicated query statements, respectively, to obtain the content feature vector of the historically undisplayed content, and to obtain the feature vectors of each word segment and each phrase in the deduplicated query statement. Based on the feature vectors of each word segment and each phrase, the semantic feature vector of the deduplicated query statement is constructed, and then the relevance between the historically undisplayed content and the deduplicated query statement is determined based on the content feature vector and the semantic feature vector.

[0101] Specifically, a preprocessed dataset is constructed. A pre-processed language model is trained using a ternary loss function and the preprocessed dataset to obtain the trained pre-processed language model. The trained pre-processed language model can be a trained BERT model, or other models; this embodiment does not impose any limitations on this. The preprocessed dataset may include pre-displayed content to be trained, positive examples, and negative examples. The pre-displayed content to be trained may be previously undisplayed content. Positive examples may come from query statements that have clicked on this previously undisplayed content, and negative examples may be randomly selected from query statements containing other content.

[0102] Step S30: Generate a candidate set of query statements based on the query statements corresponding to the historically undisplayed content.

[0103] It should be noted that general search engines should cover as much content as possible across the entire internet. Based on the design logic of current search engine systems, some content has never been retrieved. This content, which has never been displayed, will not be clicked by users, resulting in a lower score for this content during the retrieval and recall phase of the search system, making it even more difficult to retrieve. In this embodiment, the content that has never been retrieved can be regarded as historical undisplayed content. A candidate set of query statements is generated based on the query statements corresponding to the historical undisplayed content. Then, the target query statements for the historical undisplayed content are selected from the candidate set of query statements based on the relevance obtained by the trained BERT model, thereby improving the recall rate of the historical undisplayed content.

[0104] Step S40: Filter the candidate query statement set according to the relevance to obtain the target query statement, and optimize the search engine according to the target query statement.

[0105] It is easy to understand that the candidate set of query statements is filtered based on the relevance to obtain the target query statement; the candidate set of query statements is updated based on the target query statement to obtain the set of target query statements corresponding to the historically undisplayed content; and the search engine is optimized based on the set of target query statements. Specifically, in the stage of filtering the candidate set of query statements, keywords, named entities, etc., can be used to filter the candidate set of query statements, and the final target query statement is appended to the historically undisplayed content and sorted simultaneously with the historically undisplayed content, thereby improving the recall rate of the historically undisplayed content.

[0106] Specifically, the process of filtering the candidate query statement set based on the relevance can be as follows: constructing the current semantic feature vector of the query statements in the candidate query statement set; calculating the similarity between the semantic feature vector of the preset query statement and the current semantic feature vector; and selecting the target query statement from the candidate query statement set based on the relevance and similarity. The process of selecting the target query statement from the candidate query statement set based on relevance and similarity can be as follows: determining whether the similarity is greater than the relevance; when the similarity is greater than the relevance, determining the candidate query statement in the candidate query statement set corresponding to the similarity; and selecting the candidate query statement as the target query statement. By selecting the target query statement from the candidate query statement set based on the relevance obtained from the trained BERT model, the recall rate of previously undisplayed content is improved, making the search results more accurate and achieving search engine optimization.

[0107] This embodiment obtains preset query statements and historically undisplayed content from a search engine; it extracts features from the historically undisplayed content and the preset query statements using a trained preset language model to obtain the relevance between them; it generates a candidate set of query statements based on the query statements corresponding to the historically undisplayed content; it filters the candidate set of query statements based on the relevance to obtain the target query statement; and it optimizes the search engine based on the target query statement. In this embodiment, generating target query statements from historically undisplayed content in the search engine improves the score of this content during the search system's retrieval and recall phase, increasing the probability of content being recalled and making search results more accurate. This achieves search engine optimization, thus solving the technical problems of low recall probability for undisplayed content in existing search engines and inaccurate search results caused by existing methods of improving recall rates.

[0108] refer to Figure 3 , Figure 3 This is a flowchart illustrating a second embodiment of a search engine optimization method according to the present invention. Based on the first embodiment described above, the search engine optimization method of this embodiment further includes, before step S20:

[0109] Step S201: Construct a preprocessed dataset.

[0110] It is easy to understand that the process of constructing a preprocessed dataset can be as follows: obtain the preset display content to be trained and the positive sample corresponding to the preset display content to be trained; obtain the negative sample corresponding to the preset display content to be trained according to a preset random algorithm; and construct a preprocessed dataset according to the preset display content to be trained, the positive sample, and the negative sample.

[0111] Specifically, the preprocessed dataset may include the preset display content to be trained, positive examples, and negative examples. The preset display content to be trained may be historically undisplayed content, positive examples may come from query statements that have clicked on the historically undisplayed content, and negative examples may be randomly selected from query statements of other content.

[0112] Step S202: Train the preset language model using a ternary loss function and the preprocessed dataset to obtain the trained preset language model.

[0113] It should be noted that before training the preset language model using the ternary loss function and the preprocessed dataset, a ternary loss function needs to be constructed. The process of constructing the ternary loss function can be as follows: constructing sample triplets based on the preprocessed dataset; constructing the ternary loss function based on the sample triplets. Specifically, the preset display content to be trained, the positive examples, and the negative examples are obtained from the sample triplets; the positive distance between the preset display content to be trained and the positive examples is determined; the negative distance between the preset display content to be trained and the negative examples is determined; and the ternary loss function is determined based on the positive distance, the negative distance, and a preset interval. (Refer to...) Figure 4 , Figure 4 This is a schematic diagram of the training of a preset language model using the ternary loss function in this embodiment.

[0114] In practical implementation, the ternary loss function is determined using formula (1):

[0115] L tri =D p +D n +margin formula(1)

[0116] Among them, L tri For the ternary loss function, D p For positive example distances, D n The distance is the negative example distance, and the margin is the preset interval. In this embodiment, the margin can be 0.1.

[0117] This embodiment constructs a preprocessed dataset; a pre-defined language model is trained using a ternary loss function and the preprocessed dataset to obtain the trained pre-defined language model. In this embodiment, the pre-defined language model is trained using a ternary loss function, and the trained BERT model is used to extract text features, obtain the relevance between historically undisplayed content and pre-defined query statements, and then filter the query statement candidate set to obtain the target query statement. Target query statements are generated for historically undisplayed content in the search engine, improving the score of this content in the search system's retrieval recall stage. This increases the probability of content being recalled, making the search results more accurate, and achieving search engine optimization. This solves the technical problems of low recall probability for undisplayed content in existing search engines and inaccurate search results caused by existing methods of improving recall rates.

[0118] refer to Figure 5 , Figure 5 This is a flowchart illustrating a third embodiment of a search engine optimization method according to the present invention. Based on the first embodiment described above, the search engine optimization method of this embodiment includes, in step S20:

[0119] Step S21: Extract features from the historical undisplayed content and the preset query statement based on the trained preset language model to obtain the content feature vector of the historical undisplayed content and the semantic feature vector of the preset query statement.

[0120] It should be noted that, based on the trained preset language model, features are extracted from the historically undisplayed content to obtain the content feature vector of the historically undisplayed content; features are extracted from the preset query statement based on the trained preset language model to obtain the feature vectors of each word segment and each phrase in the preset query statement; and the semantic feature vector of the preset query statement is constructed based on the feature vectors of each word segment and the feature vectors of each phrase.

[0121] One optional implementation is that, in this embodiment of the application, the feature vectors of each word segment and each phrase of the preset query statement are summed and averaged to obtain the semantic feature vector of the preset query statement. That is, the feature vectors of each word segment and each phrase of the preset query statement are first summed, and then averaged to obtain the vector as the semantic feature vector of the preset query statement. Specifically, the process of constructing the semantic feature vector of the preset query statement based on the feature vectors of each word segment and each phrase can be as follows: summing the feature vectors of each word segment and each phrase to obtain a feature vector sum; and constructing the semantic feature vector of the preset query statement based on the average of the feature vector sums.

[0122] Specifically, in this embodiment, after obtaining the preset query statement, the preset query statement is first segmented into words to obtain each word of the preset query statement. An exemplary implementation is to use commonly used word segmentation software to perform the segmentation of the preset query statement, such as using the jieba word segmenter. After determining each word of the preset query statement, this embodiment constructs a feature vector for each word, that is, represents the word in vector form. It should be noted that in this embodiment, when constructing feature vectors for word segments, the semantic information of the word segments is retained in the word vectors. Therefore, the feature vectors of semantically similar words have a high correlation, meaning their word vectors are similar. For example, the words "microphone" and "microphone" have the same meaning, so their feature vectors are very similar. After constructing feature vectors for each word of the preset query statement, this embodiment further combines the word segments of the query statement to obtain word groups, and then constructs feature vectors for the word groups to obtain the feature vectors of each word segment and each word group of the preset query statement.

[0123] It is easy to understand that the above process of constructing the semantic feature vector of a preset query statement by using the feature vectors of each word segment and each phrase of the preset query statement may lose the order information of each word in the statement and the core word information in the statement. When searching for short text statements in search engines, the statements will not be too long, so the lost word order information will be less. In addition, although the processing of this embodiment will lose core word information, it can ensure that each word in the statement is equally important. For example, taking "bullet screen launch preparation" as an example, the three words are "bullet screen", "launch", and "prepare". Since each word is equally important, the similar statements obtained may be "bullet screen is so exciting", "Gatling launch", "bullet screen preparation", making the set of similar statements more diverse and avoiding the semantic concentration of the set of similar statements too much.

[0124] Step S22: Determine the relevance between the historical undisplayed content and the preset query statement based on the content feature vector and the semantic feature vector.

[0125] It should be understood that the technical solution of this application embodiment is most suitable for Chinese sentence retrieval. Therefore, in this application embodiment, the preset query statement is preferably a Chinese sentence, which includes at least one Chinese character. In this application embodiment, after obtaining the content feature vector of historically undisplayed content and the semantic feature vector of the preset query statement, the relevance between the semantic feature vector of the preset query statement and the content feature vector of historically undisplayed content is calculated respectively. The relevance between the feature vectors of sentences directly represents the relevance between sentences. Therefore, the relevance between the semantic feature vector of the preset query statement and the content feature vector of historically undisplayed content represents the semantic similarity between the preset query statement and historically undisplayed content.

[0126] Specifically, one exemplary implementation is that, in this application embodiment, the cosine value between the content feature vector of the previously undisplayed content and the semantic feature vector of the preset query statement is calculated respectively to measure the relevance between the previously undisplayed content and the preset query statement. Changes in the absolute value of the vectors (the size of the vectors) have little impact on the cosine value; using the cosine value reflects the difference in the direction of the two vectors, thus it is suitable for measuring the relevance between the two statements. It can be understood that the magnitude of the cosine value between the content feature vector and the semantic feature vector represents the relevance between the previously undisplayed content and the preset query statement. The smaller the cosine value between the content feature vector and the semantic feature vector, the lower the relevance between the previously undisplayed content and the preset query statement; the larger the cosine value between the content feature vector and the semantic feature vector, the higher the relevance between the previously undisplayed content and the preset query statement. Following the method for calculating the cosine value of the two feature vectors, this application embodiment calculates the cosine value between the content feature vector of the previously undisplayed content and the semantic feature vector of the preset query statement respectively, as the relevance between the previously undisplayed content and the preset query statement.

[0127] This embodiment extracts features from the historically undisplayed content and the preset query statement using a trained preset language model to obtain content feature vectors for the historically undisplayed content and semantic feature vectors for the preset query statement. The relevance between the historically undisplayed content and the preset query statement is then determined based on these feature vectors. In this embodiment, a trained BERT model is used to extract text features, obtain the relevance between historically undisplayed content and the preset query statement, and then filter the candidate query statement set to obtain the target query statement. This generates target query statements for historically undisplayed content in the search engine, improving the score of this content during the search system's retrieval and recall phase. This increases the probability of content being recalled, making the search results more accurate and achieving search engine optimization. This solves the technical problems of low recall probability for undisplayed content in existing search engines and inaccurate search results caused by existing methods of improving recall rates.

[0128] Furthermore, embodiments of the present invention also propose a storage medium storing a search engine optimization program, wherein the search engine optimization program is executed by a processor using the steps of the search engine optimization method described above.

[0129] Since this storage medium adopts all the technical solutions of all the above embodiments, it has at least all the beneficial effects brought about by the technical solutions of the above embodiments, which will not be repeated here.

[0130] Reference Figure 6 , Figure 6 This is a structural block diagram of the first embodiment of the search engine optimization device of the present invention.

[0131] like Figure 6 As shown, the search engine optimization device of this embodiment includes:

[0132] Module 10 is used to obtain the search engine's preset query statements and historical undisplayed content.

[0133] It should be noted that the preset query statements and historical undisplayed content of the search engine are obtained. Among them, the historical undisplayed content should be covered by the general search engine as much as possible across the entire network. Based on the design logic of the current search engine system, some content has never been retrieved. This content that has never been displayed will not be clicked by users, which will cause the score of this content to be lower in the retrieval and recall stage of the search system, making it more difficult to be retrieved. Therefore, the content that has never been retrieved can be regarded as the historical undisplayed content in this embodiment.

[0134] It's easy to understand that the default corpus of a search engine contains a large number of repetitive or highly similar statements due to the inclusion of various forms and contents. These statements are essentially of the same type, or even variations of the same statement, with little difference in essence. When a statement is retrieved as similar to the query statement, theoretically, all these highly similar statements will be retrieved simultaneously. This results in retrieved statements that are nearly identical to the query statement, but with a high workload because the same retrieval procedure must be used to retrieve each of these nearly identical statements. This situation only increases the workload for similar statement retrieval and does not make the search results richer or more comprehensive. To improve this, this embodiment, after obtaining the initial query statement from the search engine's default corpus, performs deduplication on the initial query statement and uses the deduplicated query statement as the default query statement, making the search results richer and more comprehensive.

[0135] Specifically, the steps for obtaining a preset query statement can be as follows: obtaining a preset corpus of the search engine, and obtaining an initial query statement based on the preset corpus; performing deduplication on the initial query statement to obtain a deduplicated query statement; and using the deduplicated query statement as the preset query statement.

[0136] It should be understood that the deduplication process can be performed in any of the following two ways, or in other ways, and this embodiment does not limit this.

[0137] The first exemplary implementation involves obtaining the string distance between the initial query statements; determining whether the string distance is less than a preset first distance threshold; and removing the initial query statements whose string distance is less than the preset first distance threshold to obtain deduplicated query statements. Here, string distance refers to Dice distance, which measures the similarity between two sets. The strings contained in the statements can be understood as a set of characters, and Dice distance is also used to measure string similarity.

[0138] The second exemplary implementation involves obtaining the edit distance between the initial query statements; determining whether the edit distance is less than a preset second distance threshold; and deleting the initial query statements whose edit distance is less than the preset second distance threshold to obtain deduplicated query statements. Here, edit distance represents the minimum number of edits required to transform one string into another. In this embodiment, editing involves replacing one character in a string with another, or inserting or deleting characters. Edit distance directly measures the similarity between two texts. When comparing two statements, it considers word order and word overlap, effectively identifying statements that differ by only one character or punctuation mark.

[0139] It should be noted that the above deduplication process can be performed when building the preset corpus of the search engine, or when constructing the semantic feature vectors of each statement in the preset corpus of the search engine. This application does not limit the specific timing of the above deduplication process.

[0140] The extraction module 20 is used to extract features from the historical undisplayed content and the preset query statement based on the trained preset language model, so as to obtain the relevance between the historical undisplayed content and the preset query statement.

[0141] The process is straightforward: Features are extracted from the historically undisplayed content using a trained pre-defined language model to obtain content feature vectors; features are also extracted from the pre-defined query statement using the same model to obtain word segmentation feature vectors and phrase feature vectors; a semantic feature vector is constructed based on these word segmentation and phrase feature vectors; and the relevance between the historically undisplayed content and the pre-defined query statement is determined using the content feature vectors and semantic feature vectors.

[0142] It should be understood that the pre-trained language model can be the trained BERT model. The trained BERT model is used to extract features from the historically undisplayed content and the deduplicated query statements, respectively, to obtain the content feature vector of the historically undisplayed content, and to obtain the feature vectors of each word segment and each phrase in the deduplicated query statement. Based on the feature vectors of each word segment and each phrase, the semantic feature vector of the deduplicated query statement is constructed, and then the relevance between the historically undisplayed content and the deduplicated query statement is determined based on the content feature vector and the semantic feature vector.

[0143] Specifically, a preprocessed dataset is constructed. A pre-processed language model is trained using a ternary loss function and the preprocessed dataset to obtain the trained pre-processed language model. The trained pre-processed language model can be a trained BERT model, or other models; this embodiment does not impose any limitations on this. The preprocessed dataset may include pre-displayed content to be trained, positive examples, and negative examples. The pre-displayed content to be trained may be previously undisplayed content. Positive examples may come from query statements that have clicked on this previously undisplayed content, and negative examples may be randomly selected from query statements containing other content.

[0144] The generation module 30 is used to generate a candidate set of query statements based on the query statements corresponding to the historically undisplayed content.

[0145] It should be noted that general search engines should cover as much content as possible across the entire internet. Based on the design logic of current search engine systems, some content has never been retrieved. This content, which has never been displayed, will not be clicked by users, resulting in a lower score for this content during the retrieval and recall phase of the search system, making it even more difficult to retrieve. In this embodiment, the content that has never been retrieved can be regarded as historical undisplayed content. A candidate set of query statements is generated based on the query statements corresponding to the historical undisplayed content. Then, the target query statements for the historical undisplayed content are selected from the candidate set of query statements based on the relevance obtained by the trained BERT model, thereby improving the recall rate of the historical undisplayed content.

[0146] The filtering module 40 is used to filter the candidate set of query statements according to the relevance to obtain the target query statement, and optimize the search engine according to the target query statement.

[0147] It is easy to understand that the candidate set of query statements is filtered based on the relevance to obtain the target query statement; the candidate set of query statements is updated based on the target query statement to obtain the set of target query statements corresponding to the historically undisplayed content; and the search engine is optimized based on the set of target query statements. Specifically, in the stage of filtering the candidate set of query statements, keywords, named entities, etc., can be used to filter the candidate set of query statements, and the final target query statement is appended to the historically undisplayed content and sorted simultaneously with the historically undisplayed content, thereby improving the recall rate of the historically undisplayed content.

[0148] Specifically, the process of filtering the candidate query statement set based on the relevance can be as follows: constructing the current semantic feature vector of the query statements in the candidate query statement set; calculating the similarity between the semantic feature vector of the preset query statement and the current semantic feature vector; and selecting the target query statement from the candidate query statement set based on the relevance and similarity. The process of selecting the target query statement from the candidate query statement set based on relevance and similarity can be as follows: determining whether the similarity is greater than the relevance; when the similarity is greater than the relevance, determining the candidate query statement in the candidate query statement set corresponding to the similarity; and selecting the candidate query statement as the target query statement. By selecting the target query statement from the candidate query statement set based on the relevance obtained from the trained BERT model, the recall rate of previously undisplayed content is improved, making the search results more accurate and achieving search engine optimization.

[0149] This embodiment of the search engine optimization device includes: an acquisition module 10, used to acquire preset query statements and historically undisplayed content from the search engine; an extraction module 20, used to extract features from the historically undisplayed content and the preset query statements based on a trained preset language model to obtain the relevance between the historically undisplayed content and the preset query statements; a generation module 30, used to generate a candidate set of query statements based on the query statements corresponding to the historically undisplayed content; and a filtering module 40, used to filter the candidate set of query statements based on the relevance to obtain a target query statement, and optimize the search engine based on the target query statement. In this embodiment, generating target query statements from historically undisplayed content in the search engine improves the score of this historically undisplayed content in the search system's retrieval recall stage, which can increase the probability of content being recalled, making the search results more accurate, and achieving search engine optimization. This solves the technical problems of low recall probability of undisplayed content in existing search engines and inaccurate search results caused by existing methods of improving recall rate.

[0150] In one embodiment, the search engine optimization device further includes: a training module;

[0151] The training module is used to construct the preprocessed dataset;

[0152] The preset language model is trained using a ternary loss function and the preprocessed dataset to obtain the trained preset language model.

[0153] In one embodiment, the training module is further configured to obtain the preset display content to be trained and the positive sample corresponding to the preset display content to be trained;

[0154] Negative example samples corresponding to the preset display content to be trained are obtained according to a preset random algorithm;

[0155] A preprocessed dataset is constructed based on the preset display content to be trained, the positive examples, and the negative examples.

[0156] In one embodiment, the training module is further configured to construct sample triples based on the preprocessed dataset;

[0157] Construct a ternary loss function based on the sample triples.

[0158] In one embodiment, the training module is further configured to obtain the preset display content to be trained, the positive sample, and the negative sample based on the sample triples;

[0159] Determine the positive distance between the preset display content to be trained and the positive sample;

[0160] Determine the negative example distance between the preset display content to be trained and the negative example samples;

[0161] The ternary loss function is determined based on the positive example distance, the negative example distance, and the preset interval.

[0162] In one embodiment, the acquisition module 10 is further configured to acquire a preset corpus of the search engine and acquire an initial query statement based on the preset corpus;

[0163] The initial query statement is deduplicated to obtain the deduplicated query statement;

[0164] The deduplicated query statement is used as the preset query statement.

[0165] In one embodiment, the acquisition module 10 is further configured to acquire the string distance between the initial query statements respectively;

[0166] Determine whether the distance between the strings is less than a preset first distance threshold;

[0167] The initial query statements corresponding to strings whose distance is less than the preset first distance threshold are deleted to obtain duplicate query statements.

[0168] In one embodiment, the acquisition module 10 is further configured to acquire the edit distance between the initial query statements respectively;

[0169] Determine whether the edit distance is less than a preset second distance threshold;

[0170] The initial query statements corresponding to the edit distance being less than the preset second distance threshold are deleted to obtain the deduplicated query statements.

[0171] In one embodiment, the extraction module 20 is further configured to perform feature extraction on the historically undisplayed content and the preset query statement based on the trained preset language model, so as to obtain the content feature vector of the historically undisplayed content and the semantic feature vector of the preset query statement;

[0172] The relevance between the historical undisplayed content and the preset query statement is determined based on the content feature vector and the semantic feature vector.

[0173] In one embodiment, the extraction module 20 is further configured to extract features from the historically undisplayed content based on a trained preset language model, so as to obtain the content feature vector of the historically undisplayed content;

[0174] Based on the trained preset language model, feature extraction is performed on the preset query statement to obtain the feature vectors of each word segment and each word group in the preset query statement.

[0175] The semantic feature vector of the preset query statement is constructed based on the feature vectors of each word segment and the feature vectors of each word group.

[0176] In one embodiment, the extraction module 20 is further configured to sum the feature vectors of each word segmentation and the feature vectors of each word group to obtain a feature vector sum;

[0177] The semantic feature vector of the preset query statement is constructed based on the average value of the feature vectors.

[0178] In one embodiment, the filtering module 40 is further configured to filter the candidate set of query statements based on the relevance to obtain the target query statement;

[0179] Update the query statement candidate set according to the target query statement to obtain the target query statement set corresponding to the historical undisplayed content;

[0180] Optimize the search engine based on the target query set.

[0181] In one embodiment, the filtering module 40 is further configured to construct the current semantic feature vector of the query statements in the query statement candidate set;

[0182] Calculate the similarity between the semantic feature vector of the preset query statement and the current semantic feature vector;

[0183] The target query statement is selected from the candidate query statement set based on the relevance and similarity.

[0184] In one embodiment, the filtering module 40 is further configured to determine whether the similarity is greater than the relevance;

[0185] When the similarity is greater than the relevance, the candidate query statements in the query statement candidate set corresponding to the similarity are determined;

[0186] The candidate query statement is used as the target query statement.

[0187] Other embodiments or specific implementations of the search engine optimization device described in this invention can be found in the above-described embodiments of search engine optimization methods, and will not be repeated here.

[0188] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solutions of the present invention. In specific applications, those skilled in the art can make settings as needed, and the present invention does not impose any restrictions on this.

[0189] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this invention. In practical applications, those skilled in the art can select some or all of the workflow to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0190] In addition, for technical details not described in detail in this embodiment, please refer to the search engine optimization method provided in any embodiment of the present invention, which will not be repeated here.

[0191] Furthermore, it should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0192] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0193] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0194] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A search engine optimization method, characterized in that, The search engine optimization method includes: Retrieve the search engine's default query statements and historical content that has not been displayed. Based on the trained preset language model, feature extraction is performed on the historical undisplayed content and the preset query statement to obtain the relevance between the historical undisplayed content and the preset query statement; Generate a candidate set of query statements based on the query statements corresponding to the historical content that was not displayed. The candidate query statement is filtered based on the relevance to obtain the target query statement, and the search engine is optimized based on the target query statement.

2. The search engine optimization method as described in claim 1, characterized in that, Before the step of extracting features from the historically undisplayed content and the preset query statement based on the trained preset language model to obtain the relevance between the historically undisplayed content and the preset query statement, the method further includes: Build a preprocessed dataset; The preset language model is trained using a ternary loss function and the preprocessed dataset to obtain the trained preset language model.

3. The search engine optimization method as described in claim 2, characterized in that, The steps for constructing the preprocessed dataset include: Obtain the preset display content to be trained and the corresponding positive sample; Negative example samples corresponding to the preset display content to be trained are obtained according to a preset random algorithm; A preprocessed dataset is constructed based on the preset display content to be trained, the positive examples, and the negative examples.

4. The search engine optimization method as described in claim 3, characterized in that, Before the step of training the preset language model using a ternary loss function and the preprocessed dataset to obtain the trained preset language model, the method further includes: Construct sample triples based on the preprocessed dataset; Construct a ternary loss function based on the sample triples.

5. The search engine optimization method as described in claim 4, characterized in that, The step of constructing a ternary loss function based on the sample triples includes: The training preset display content, the positive sample, and the negative sample are obtained based on the sample triples; Determine the positive distance between the preset display content to be trained and the positive sample; Determine the negative example distance between the preset display content to be trained and the negative example samples; The ternary loss function is determined based on the positive example distance, the negative example distance, and the preset interval.

6. The search engine optimization method as described in claim 1, characterized in that, Before the step of obtaining the search engine's preset query statement and historical undisplayed content, the method further includes: Obtain the preset corpus of the search engine, and obtain the initial query statement based on the preset corpus; The initial query statement is deduplicated to obtain the deduplicated query statement; The deduplicated query statement is used as the preset query statement.

7. The search engine optimization method as described in claim 6, characterized in that, The step of deduplicating the initial query statement to obtain the deduplicated query statement includes: Obtain the string distance between the initial query statements respectively; Determine whether the distance between the strings is less than a preset first distance threshold; The initial query statements corresponding to strings whose distance is less than the preset first distance threshold are deleted to obtain duplicate query statements.

8. The search engine optimization method as described in claim 6, characterized in that, The step of deduplicating the initial query statement to obtain the deduplicated query statement includes: Obtain the edit distance between the initial query statements respectively; Determine whether the edit distance is less than a preset second distance threshold; The initial query statements corresponding to the edit distance being less than the preset second distance threshold are deleted to obtain the deduplicated query statements.

9. The search engine optimization method according to any one of claims 1 to 8, characterized in that, The step of extracting features from the historically undisplayed content and the preset query statement based on the trained preset language model to obtain the relevance between the historically undisplayed content and the preset query statement includes: Based on the trained preset language model, feature extraction is performed on the historical undisplayed content and the preset query statement to obtain the content feature vector of the historical undisplayed content and the semantic feature vector of the preset query statement; The relevance between the historical undisplayed content and the preset query statement is determined based on the content feature vector and the semantic feature vector.

10. The search engine optimization method as described in claim 9, characterized in that, The step of extracting features from the historically undisplayed content and the preset query statement based on the trained preset language model to obtain the content feature vector of the historically undisplayed content and the semantic feature vector of the preset query statement includes: Based on the trained preset language model, feature extraction is performed on the historically undisplayed content to obtain the content feature vector of the historically undisplayed content; Based on the trained preset language model, feature extraction is performed on the preset query statement to obtain the feature vectors of each word segment and each word group in the preset query statement. The semantic feature vector of the preset query statement is constructed based on the feature vectors of each word segment and the feature vectors of each word group.

11. The search engine optimization method as described in claim 10, characterized in that, The step of constructing the semantic feature vector of the preset query statement based on the word segmentation feature vector and the phrase feature vector includes: The feature vectors of each word segmentation and each word group are summed to obtain the feature vector sum. The semantic feature vector of the preset query statement is constructed based on the average value of the feature vectors.

12. The search engine optimization method according to any one of claims 1 to 8, characterized in that, The step of filtering the candidate query statement set according to the relevance to obtain the target query statement, and optimizing the search engine according to the target query statement, includes: The candidate query statement is filtered based on the relevance to obtain the target query statement; Update the query statement candidate set according to the target query statement to obtain the target query statement set corresponding to the historical undisplayed content; Optimize the search engine based on the target query set.

13. The search engine optimization method as described in claim 12, characterized in that, The step of filtering the candidate query statement set according to the relevance to obtain the target query statement includes: Construct the current semantic feature vector of the query statements in the query statement candidate set; Calculate the similarity between the semantic feature vector of the preset query statement and the current semantic feature vector; The target query statement is selected from the candidate query statement set based on the relevance and similarity.

14. The search engine optimization method as described in claim 13, characterized in that, The step of filtering target query statements from the candidate query statement set based on the relevance and the similarity includes: Determine whether the similarity is greater than the relevance; When the similarity is greater than the relevance, the candidate query statements in the query statement candidate set corresponding to the similarity are determined; The candidate query statement is used as the target query statement.

15. A search engine optimization device, characterized in that, The search engine optimization device includes: The acquisition module is used to acquire the search engine's preset query statements and historical content that has not been displayed. The extraction module is used to extract features from the historical undisplayed content and the preset query statement based on the trained preset language model, so as to obtain the relevance between the historical undisplayed content and the preset query statement; The generation module is used to generate a candidate set of query statements based on the query statements corresponding to the historically undisplayed content; The filtering module is used to filter the candidate set of query statements according to the relevance to obtain the target query statement, and optimize the search engine according to the target query statement.

16. The search engine optimization apparatus as described in claim 15, characterized in that, The search engine optimization device further includes: a training module; The training module is used to construct the preprocessed dataset; The preset language model is trained using a ternary loss function and the preprocessed dataset to obtain the trained preset language model.

17. The search engine optimization apparatus as described in claim 16, characterized in that, The training module is also used to obtain the preset display content to be trained and the positive sample corresponding to the preset display content to be trained; Negative example samples corresponding to the preset display content to be trained are obtained according to a preset random algorithm; A preprocessed dataset is constructed based on the preset display content to be trained, the positive examples, and the negative examples.

18. The search engine optimization apparatus as described in claim 17, characterized in that, The training module is also used to construct sample triples based on the preprocessed dataset; Construct a ternary loss function based on the sample triples.

19. A search engine optimization device, characterized in that, The search engine optimization device includes: a memory, a processor, and a search engine optimization program stored in the memory and executable on the processor, the search engine optimization program being configured to implement the search engine optimization method as described in any one of claims 1 to 14.

20. A storage medium, characterized in that, The storage medium stores a search engine optimization program, which, when executed by a processor, implements the steps of the search engine optimization method as described in any one of claims 1 to 14.

Images