Data processing method and apparatus

Abstract

This specification provides a data processing method and apparatus, wherein the data processing includes: obtaining semantic identifier sequences corresponding to each initial keyword, and clustering each initial keyword according to each semantic identifier sequence to obtain at least one keyword cluster, wherein each keyword cluster includes at least one initial keyword; generating virtual keywords corresponding to each keyword cluster, and determining the keyword correspondence between each initial keyword and each virtual keyword; generating a target keyword index based on the initial keyword index and the keyword correspondence; and determining the target keyword object corresponding to the query keyword based on the target keyword index. This method can reduce a large number of redundant initial keywords to a small number of virtual keywords, greatly reducing the computational overhead in the recall retrieval stage. Simultaneously, this method can also reduce synonymous keywords, thereby increasing keyword matching accuracy and improving the accuracy of subsequent keyword object determination.

Description

Technical Field

[0001] This specification relates to the field of computer technology, and in particular to data processing methods. This specification also relates to a data processing apparatus, a computing device, a computer-readable storage medium, and a computer program product. Background Technology

[0002] With the development of internet technology, more and more users are shopping online, and correspondingly, more and more advertisers are placing ads online. In search advertising systems, keywords are the core bridge connecting user intent and advertiser placement needs. Advertisers trigger ad display by setting keywords, and the advertising system achieves rapid ad recall by building keyword indexes. With the growth of business scale, the number of keywords is growing explosively.

[0003] Current advertising systems suffer from a significant problem of keyword synonym duplication. This means that keywords representing the same content are frequently repeated, such as "sports shoes," "running shoes," and "sports running shoes"—keywords with highly overlapping meanings. The abundance of synonyms results in a massive number of entries in the keyword index, containing a large amount of redundant data. This leads to a significant waste of storage resources, which in turn increases the computational overhead during the recall phase, severely impacting system response latency. Furthermore, the large number of synonyms increases the uncertainty between query terms and recalled keywords, and reduces matching accuracy. Summary of the Invention

[0004] In view of this, embodiments of this specification provide a data processing method. This specification also relates to a data processing apparatus, a computing device, a computer-readable storage medium, and a computer program product, to address the aforementioned problems existing in the prior art.

[0005] According to a first aspect of the embodiments of this specification, a data processing method is provided, comprising: Obtain the semantic identifier sequence corresponding to each initial keyword, and cluster each initial keyword according to each semantic identifier sequence to obtain at least one keyword cluster, wherein each keyword cluster includes at least one initial keyword; Generate virtual keywords corresponding to each keyword cluster, and determine the keyword correspondence between each initial keyword and each virtual keyword; A target keyword index is generated based on the initial keyword index and the keyword correspondence, wherein the target keyword index includes virtual keywords and keyword objects corresponding to the virtual keywords; The target keyword object corresponding to the query keyword is determined based on the target keyword index.

[0006] According to a second aspect of the embodiments of this specification, a data processing apparatus is provided, comprising: The acquisition module is configured to acquire the semantic identifier sequence corresponding to each initial keyword, and cluster each initial keyword according to each semantic identifier sequence to obtain at least one keyword cluster, wherein each keyword cluster includes at least one initial keyword; The generation module is configured to generate virtual keywords corresponding to each keyword cluster and determine the keyword correspondence between each initial keyword and each virtual keyword; The index building module is configured to generate a target keyword index based on the initial keyword index and the keyword correspondence, wherein the target keyword index includes virtual keywords and keyword objects corresponding to the virtual keywords; The query module is configured to determine the target keyword object corresponding to the query keyword based on the target keyword index.

[0007] According to a third aspect of the embodiments of this specification, a computing device is provided, comprising: Memory and processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions, which, when executed by the processor, implement the steps of the above-described data processing method.

[0008] According to a fourth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores a computer program / instructions that, when executed by a processor, implement the steps of the data processing method described above.

[0009] According to a fifth aspect of the embodiments of this specification, a computer program product is provided, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described data processing method.

[0010] The data processing method provided in this specification involves obtaining semantic identifier sequences corresponding to each initial keyword, clustering each initial keyword according to each semantic identifier sequence to obtain at least one keyword cluster, wherein each keyword cluster includes at least one initial keyword; generating virtual keywords corresponding to each keyword cluster, and determining the keyword correspondence between each initial keyword and each virtual keyword; generating a target keyword index according to the initial keyword index and the keyword correspondence, wherein the target keyword index includes virtual keywords and keyword objects corresponding to the virtual keywords; and determining the target keyword object corresponding to the query keyword according to the target keyword index.

[0011] One embodiment of this specification converts initial keywords into semantic identifier sequences and clusters them, facilitating clustering at different semantic granularities. Then, corresponding virtual keywords are generated for each cluster, establishing a correspondence between the initial keywords and virtual keywords. This reduces multiple redundant and semantically similar initial keywords into a small number of virtual keywords. Based on this correspondence, index information between virtual word objects and virtual keywords is generated, enabling subsequent queries to find the corresponding keyword object. This method reduces a large number of redundant initial keywords to a small number of virtual keywords, significantly reducing the computational overhead of the recall retrieval stage. Furthermore, this method reduces synonymous keywords, thereby increasing keyword matching accuracy and improving the accuracy of subsequent keyword object identification. Attached Figure Description

[0012] Figure 1 This is a flowchart of a data processing method provided in one embodiment of this specification; Figure 2 This is a schematic diagram illustrating the generation of keyword clusters provided in one embodiment of this specification; Figure 3 This is a flowchart illustrating a data processing method for an advertising recall scenario provided in one embodiment of this specification. Figure 4 This is a schematic diagram of the structure of a data processing device provided in one embodiment of this specification; Figure 5 This is an architecture diagram of a data processing system provided in one embodiment of this specification; Figure 6 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0013] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0014] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0015] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0016] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this manual are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0017] First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0018] Virtual keywords: A type of keyword proposed in this specification. After clustering the keywords placed by advertisers, virtual keywords are extracted from the clustering results and can represent the clusters. They are used to replace the original keywords in the advertising indexing and retrieval process.

[0019] Semantic tag clustering: This method uses a pre-trained encoding model to encode keywords into semantic vectors, then discretizes the continuous vectors into semantic tag sequences through residual quantization, and finally performs hierarchical aggregation of keywords based on prefix matching of semantic tags.

[0020] Residual quantization is a multi-stage, hierarchical data compression technique that achieves progressively more accurate compression of the original data by repeatedly quantizing the errors (residuals) generated in the previous step. It is widely used in large model compression, audio and video processing, and vector retrieval, significantly reducing storage and computational costs while maintaining high accuracy.

[0021] ANN retrieval: Approximate Nearest Neighbor Retrieval, is a technique for quickly finding candidate vectors that are most similar to a given query vector in a high-dimensional vector space. It is often used for similarity matching of large-scale text vectors.

[0022] Bidword: A bid term, a keyword purchased by advertisers on a search advertising platform to trigger ad impressions. When a user's search term matches the bid term, the corresponding ad enters the candidate set for bidding.

[0023] Query: Search terms, query terms. The search terms entered by the user in the search box. The system needs to match them with the advertiser's bid terms to complete the ad recall.

[0024] The technical embodiments described in this specification are specifically applied to the core keyword recall stage in the search advertising system architecture. They are an indispensable pre-installation module in the search advertising launch process and directly determine the overall effectiveness and accuracy of subsequent ad candidate screening and bidding ranking.

[0025] In typical commercial search advertising scenarios, the basic trigger logic for ad placement relies on a keyword system. Advertisers independently select, purchase, and configure initial keywords in batches in the ad placement backend based on their core needs such as their products / services, target audience, and placement scenarios. They also set corresponding placement parameters such as bids, placement periods, audience targeting, and landing pages. This serves as the core basis for whether their ads can be triggered and displayed by user searches. Only user search behavior matching the corresponding keywords can potentially activate the advertiser's ad resources.

[0026] When an end user enters any search query into the search box of a search engine and initiates a search request, the system immediately initiates a full-link advertising matching process. First, it performs basic preprocessing operations on the user's original query, including text parsing, word segmentation, redundancy removal, and semantic normalization, to accurately extract the user's true search intent and core search needs. Then, the system retrieves the entire database of advertisers' purchased keywords stored in the backend and performs multi-dimensional, precise matching and relevance verification between the preprocessed user query and the massive amount of keywords targeted by advertisers.

[0027] Through a keyword matching and filtering mechanism, the system can quickly filter out all advertiser resources whose keywords match the user's search query and are suitable for the user's search needs, accurately identifying advertisers that fit the current user's search scenario. Based on this, it further aggregates all advertising materials under the corresponding advertiser that meet the matching criteria and are in an effective delivery state, and incorporates them into the system's ad candidate set.

[0028] The collected ad candidate set will serve as the core material pool for subsequent real-time bidding, weight ranking, quality score calculation, traffic distribution, and page display of ads. It will provide effective, compliant, and accurate ad resource support for the entire search ad delivery process, ensuring that search ads not only meet users' search needs but also satisfy advertisers' exposure and conversion requirements.

[0029] In the current mainstream search advertising keyword configuration system, the industry-standard initial keyword configuration methods are mainly divided into two categories: advertisers manually configuring keywords and platform systems intelligently recommending keywords. Advertisers can use one method alone or combine the two methods to complete the overall configuration of initial keywords for ad placement, providing basic keyword support for ad triggering and recall.

[0030] The manual keyword setting mode relies entirely on the advertiser's independent operation. Advertisers, based on their own product attributes, service content, business scenarios, target customer search habits, and overall advertising strategy, independently select, add, edit, and delete keywords in the advertising backend. The selected keywords are all highly relevant to their promotional content and can accurately match their business positioning. This is a core keyword pool that advertisers proactively build based on their own business understanding.

[0031] The system's intelligent keyword recommendation is an automated configuration mode, eliminating the need for advertisers to manually select keywords. The system leverages massive amounts of industry data accumulated in the backend, basic information about the advertiser's store and products, historical keyword data, past ad impressions and clicks, conversion data, and other multi-dimensional information. Combined with the platform's keyword mining algorithm and semantic matching model, it automatically analyzes the advertiser's business characteristics and suitable advertising scenarios, batch-mining and pushing high-quality keywords tailored to their advertising needs, thus filling the gaps and shortcomings of advertisers manually selecting keywords.

[0032] In actual campaign operations, most advertisers use a combination of manual keyword selection and intelligent keyword selection. Manual keyword selection ensures accurate coverage of core business keywords, while intelligent keyword recommendations broaden the reach of long-tail and potential traffic. These two configuration methods complement each other, resulting in a massive and diverse collection of initial keyword data accumulated in the search advertising system's backend over time. This provides a vast keyword database for subsequent keyword retrieval, matching, bidding, and other full-process advertising services. However, it also places higher demands on the system's keyword selection, filtering, and accurate matching capabilities. Specifically: 1. The intelligent keyword recommendation generates synonym duplication. Through analysis of the keywords automatically recommended by the system, it was found that most ad units have semantic duplication between the recommended keywords and the advertiser's free keywords, such as synonyms such as "sports", "running shoes" and "sports running shoes".

[0033] 2. Repetition in mixed advertising scenarios: In a mixed advertising model where advertisers manually purchase keywords and the system intelligently recommends keywords, most ad units exhibit semantic repetition of keywords.

[0034] 3. The index size has expanded. Due to the large number of synonyms, the current keyword index (advertiser-keyword) has reached billions of entries, which contain a large amount of redundant data, resulting in a significant waste of storage resources.

[0035] The above issues lead to wasted index storage space, increased retrieval computation overhead during the recall phase, and impact on system response latency. A large number of synonymous keywords increase the uncertainty of whether the query terms can be matched with the bid term support, reducing matching accuracy.

[0036] This specification provides a data processing method, and also relates to a data processing apparatus, a computing device, a computer-readable storage medium, and a computer program product, which are described in detail in the following embodiments.

[0037] Figure 1 A flowchart of a data processing method according to an embodiment of this specification is shown, which specifically includes the following steps: Step 102: Obtain the semantic identifier sequence corresponding to each initial keyword, and cluster each initial keyword according to each semantic identifier sequence to obtain at least one keyword cluster, wherein each keyword cluster includes at least one initial keyword.

[0038] In the actual operation of search advertising, advertisers often use a large number of similar keywords with the same meaning and target audience to increase the chances of their ads being successfully matched by users' search behavior. Advertisers hope to expand the reach of their ads by increasing the number of similar keywords, thereby increasing their exposure opportunities in the keyword matching process. This practice of repeatedly using similar keywords in bulk continuously increases the amount of data stored in the system's index, causing the index to swell. This not only consumes more system storage resources but also increases the computational pressure on subsequent keyword retrieval and matching, impacting the overall efficiency of the business process.

[0039] To address the aforementioned problems, the embodiments described in this specification propose corresponding solutions. This solution first conducts a unified semantic analysis and summarization of the various bid terms collected within the system, identifying the common or highly similar semantic implications behind multiple bid terms. Based on this, corresponding virtual keywords are generated using the integrated semantic information, and these single virtual keywords replace multiple original bid terms with similar semantic meanings. This word merging and replacement method effectively reduces the number of keyword entries actually stored in the index, alleviating the data bloat caused by a large number of duplicate bid terms, thereby optimizing the storage pressure on the index and the system's retrieval performance.

[0040] Initial keywords can be understood as the bid terms that advertisers use on advertising platforms. These initial keywords are stored in a keyword database. Semantic identifier sequences specifically refer to sequences of semantic IDs, which are essentially sequences of discrete tokens compressed from a continuous high-dimensional vector.

[0041] After extracting and transforming the semantic identifier sequences of each initial keyword, the system can perform batch clustering operations on all initial keywords based on the generated standardized semantic identifier sequences, ultimately dividing them into multiple independent keyword clusters. This clustering process covers all initial keywords in the database, and each generated keyword cluster contains at least one initial keyword, ensuring the completeness and effectiveness of the clustering results.

[0042] The core clustering logic of this scheme uses semantic similarity as the criterion. All initial keywords belonging to the same keyword cluster possess highly similar semantic information. The quantitative criterion is that the semantic identifier sequences corresponding to all initial keywords within the same cluster have completely identical word content at a preset alignment number, thereby achieving accurate aggregation in the semantic dimension and avoiding the problem of insufficient accuracy in traditional text matching clustering.

[0043] To clearly illustrate the rules for determining clustering granularity, let's take the structure of semantic identifier sequences as an example: Assume a single semantic identifier sequence contains a fixed 5 tokens. The system can flexibly configure the clustering granularity according to business needs. When the clustering comparison granularity is set to the first 3 tokens, the system will use the first three tokens of the semantic identifier sequence as the clustering basis. All initial keywords with identical first 3 tokens will be aggregated into the same keyword cluster. When the clustering comparison granularity is adjusted to the first 4 tokens, the first four tokens of the semantic identifier sequence will be used as the matching standard. Only initial keywords with identical first 4 tokens will be classified into the same cluster. By adjusting the number of tokens compared, the fineness of keyword clustering and semantic matching accuracy can be flexibly controlled.

[0044] In the actual operation of search advertising systems, the initial number of keywords accumulated by the platform is enormous. If the semantic similarity between each keyword's original vector were calculated directly to complete the clustering process, the entire process would generate a huge amount of computation. This would not only significantly consume the system's computing resources, causing unnecessary resource waste, but also lengthen the overall computation time, slow down the keyword clustering process, and affect the normal progress of subsequent business processes.

[0045] To effectively address this issue and comprehensively improve overall computational efficiency, this solution optimizes the processing flow. The system first converts all initial keywords into standardized semantic identifier sequences, and then uses these sequences to perform keyword clustering. This approach significantly reduces the computational complexity of the clustering process, substantially decreases overall computational overhead, and effectively shortens task execution time. Simultaneously, relying on semantic identifier sequences for clustering stably preserves the core semantic features of the keywords, ensuring the accuracy of the final clustering results while improving operational efficiency, thus balancing performance and business effectiveness.

[0046] In one specific embodiment provided in this specification, obtaining the semantic identifier sequence corresponding to each initial keyword includes: Each initial keyword is embedded to obtain the initial keyword vector corresponding to each initial keyword; Each initial keyword vector is converted into a semantic tag sequence with a preset number of tags.

[0047] In this embodiment, the method for obtaining the semantic identifier sequence corresponding to each initial keyword is explained in detail. In the overall processing flow, the first step is to perform embedding processing on all initial keywords in the system. Through this standardized transformation step, a unique initial keyword vector is generated for each initial keyword, which is also the prerequisite step for generating the semantic identifier sequence.

[0048] Specifically, a fully trained, pre-defined language model with fixed parameters can be used to perform semantic encoding. The BERT pre-trained language model, commonly used in the industry, is one of the preferred options. During execution, the system iterates through all initial keywords in the database, inputting each keyword sequentially into the model. The model then performs text parsing, semantic extraction, and encoding operations, ultimately outputting initial keyword vectors with fixed dimensions. These encoded initial keyword vectors can deeply mine and accurately capture the deep semantic features hidden behind the keywords, going beyond the literal text itself. Based on the distribution characteristics of the vector space, keywords with highly similar semantic meanings and expressing the same type of content will have significantly closer relative distances in their corresponding vectors within the entire vector space; conversely, keywords with large semantic differences will have larger vector distances. This characteristic provides reliable data support for subsequent semantic-based keyword clustering and matching operations.

[0049] In the keyword processing workflow of a search advertising system, the initial keyword vectors are characterized by high dimensionality and massive data volume. If these raw vectors are used directly for clustering operations, the system needs to perform a large amount of similarity comparison, numerical calculation, and grouping determination under high-dimensional data dimensions, which will generate extremely large computational overhead. On the one hand, it will continuously consume a large amount of server computing resources, causing excessive consumption of hardware resources; on the other hand, it will also significantly increase the overall task's computation time, slow down the processing pace of the keyword clustering stage, and even negatively impact the operational efficiency of subsequent upstream and downstream business processes such as ad recall and matching.

[0050] To optimize the computational process and comprehensively improve overall processing efficiency, the method proposed in the embodiments of this specification adds a dedicated data format conversion step. The system further processes each encoded initial keyword vector, uniformly converting it into a semantic identifier sequence with a fixed preset number of identifiers. This sequence replaces the original high-dimensional vector in subsequent clustering operations. The preset number of identifiers is a core parameter pre-configured in this embodiment, specifically used to define the structure of the semantic identifier sequence. This parameter directly specifies the total number of semantic units contained in a single semantic identifier sequence, and all keywords are converted according to this unified standard. For example, when the preset number of identifiers is set to 4 in the technical solution, the semantic identifier sequences generated after the conversion of all initial keywords will be internally structured as a sequence of 4 independent semantic units, thereby achieving data format standardization and lightweighting, laying the foundation for efficient subsequent clustering.

[0051] The initial keyword vector can be converted into a semantic identifier sequence by using residual quantization to discretize the initial keyword vector, gradually obtaining semantic identifiers until the number of semantic identifiers reaches the preset number. Residual quantization (RQ) is a multi-level recursive discretization method. Its core is to decompose continuous vectors and values ​​into multiple residuals, quantize them step by step into discrete codewords, and use the superposition from coarse-grained to fine-grained to approximate the original value, achieving high-precision discrete identifiers at low bit depths.

[0052] Specifically, codebooks corresponding to a preset number of identifiers are pre-trained. Each codebook contains a batch of standard reference vectors. The entire process is divided into multiple layers, with each layer corresponding to a codebook. The more layers there are, the higher the final reconstruction accuracy.

[0053] In the vector hierarchical quantization encoding process of this embodiment, the initial keyword vector to be processed is first directly used as the initial residual vector for the first round of iteration calculation. The core meaning of the residual vector is: the vector difference part that has not yet been matched and fitted by the reference vector, still retains valid semantic information, and needs to be processed through subsequent hierarchical processes. It is the core processing object of each layer encoding iteration.

[0054] The overall encoding process will perform standardized iterative processing on the initial keyword vector layer by layer and step by step according to the number of network layers pre-configured by the system. Vector compression and identifier extraction will be completed through multi-layer progressive fitting. The specific execution logic for each layer is as follows: First, the dedicated codebook file corresponding to the current encoding layer is accurately retrieved. The codebook contains a set of all reference vectors that can be used for vector matching at this layer. Then, all reference vectors in the codebook are traversed, and the reference vector with the highest matching degree and the largest similarity to the current input residual vector is selected by the vector similarity calculation rules. The unique position label of the optimal reference vector in the current codebook is recorded, and this position label is the semantic identifier corresponding to the current layer.

[0055] After extracting the identifiers for the current level, the optimal reference vector obtained through filtering is subtracted from the original residual vector input in this round, and a new residual vector is generated through vector difference operation. This new residual vector represents that the current level has fitted and matched some of the core semantic information in the original residual vector through the reference vector. The remaining unfitted difference semantic part will be handed over to the next layer of the encoding process for further processing.

[0056] After completing the entire calculation process for a single layer, the newly generated residual vector is passed as input to the next layer's encoding module, completely replicating the entire set of processing logic for matching, calculating, and updating the residual vector. Through multi-layered iterative calculations, semantic identifiers at each level are continuously extracted and residual vectors are updated until the cumulative number of semantic identifiers reaches the system's preset identifier count standard. This completes the encoding of the entire initial keyword vector, generating a complete and compliant semantic identifier sequence.

[0057] For example, taking "sports shoes" as the initial keyword, we first use the BERT semantic model to encode it semantically to obtain the initial keyword vector E0. Then, we perform residual quantization on the initial keyword vector E0. The preset number of labels is 5. After processing, we can obtain the semantic label sequence [12, 45, 78, 72, 56] corresponding to the initial keyword "sports shoes".

[0058] The above is a schematic illustration of obtaining the semantic identifier sequence for "sports shoes". In practical applications, the same operation can be performed on each initial keyword to obtain the semantic identifier sequence corresponding to each initial keyword.

[0059] In one specific embodiment provided in this specification, each initial keyword is clustered according to each semantic identifier sequence to obtain at least one keyword cluster, including: Determine the clustering semantic identifier information, and cluster each initial keyword according to the clustering semantic identifier information to obtain at least one initial keyword cluster; Based on the keyword settings of each keyword object, each initial keyword cluster is modified to obtain at least one keyword cluster.

[0060] After completing the encoding and conversion process for all initial keywords and generating their corresponding discrete semantic identifier sequences, the keyword clustering stage can begin. This scheme no longer relies on complex calculations based on high-dimensional vector similarity, but instead directly uses structured and standardized semantic identifier sequences as the basis for clustering, greatly simplifying the clustering matching logic and improving overall processing efficiency.

[0061] Specifically, all initial keywords with the same semantic identifier prefix will be automatically grouped and classified into the same keyword cluster, thereby achieving batch aggregation of semantically similar keywords. Among them, the cluster semantic identifier information used for clustering determination essentially refers to the character length of the semantic identifier prefix participating in the matching comparison. This length parameter is the core controllable parameter that regulates the overall clustering effect and determines the fine granularity of the clusters.

[0062] In actual clustering, the length of the semantic identifier prefix directly affects the clustering granularity and the semantic similarity of keywords within a cluster. When the selected semantic identifier prefix is ​​shorter, the number of identifier features participating in the matching constraints is less, the number of initial keywords satisfying the prefix consistency condition is more, and the corresponding clustering granularity is coarser. The final generated single cluster will accommodate more initial keywords with broad semantic matching, and the clustering coverage is wider, but the semantic fit of keywords within the cluster is relatively loose.

[0063] Correspondingly, the longer the selected semantic identifier prefix, the more identifier feature dimensions participate in the matching constraint, the stricter the judgment conditions, the significantly reduced number of initial keywords that can completely match the same prefix, the finer the corresponding clustering granularity, the smaller the final cluster size, the highly consistent semantic features of all initial keywords within the cluster, the higher the matching accuracy, and the stronger the semantic tightness between keywords.

[0064] Technical personnel can flexibly adjust the length of semantic identifier prefixes according to the actual needs of business scenarios, and dynamically adjust the fineness of clustering, thereby achieving the optimal balance between cluster coverage granularity and semantic density of keywords within clusters. This can avoid the problems of excessive clustering and data fragmentation caused by excessively fine granularity, and also prevent the defects of semantic mixing within clusters and insufficient clustering accuracy caused by excessively coarse granularity, thus adapting to different advertising recall and retrieval matching business needs.

[0065] For example, consider the semantic identifier sequence of athletic shoes as [12, 45, 78, 72, 56], running shoes as [12, 45, 78, 75, 56], and running shoes as [12, 45, 78, 75, 20]. If the cluster semantic identifier information is 3, then the first three semantic identifiers of these three initial keywords are all [12, 45, 78], and these three initial keywords can be in the same cluster. If the cluster semantic identifier information is 4, then the prefix [12, 45, 78, 72] of athletic shoes is different from the prefix [12, 45, 78, 75] of running shoes, so athletic shoes are in one cluster, and running shoes are in another cluster.

[0066] In the keyword clustering processing scheme proposed in the embodiments of this specification, a preliminary clustering operation is first performed on all initial keywords after encoding, based on preset clustering semantic identifier information. After standardized prefix matching and aggregation processing, one or more initial keyword clusters are finally generated. The initial keyword clusters produced in this stage are preliminary clustering results obtained purely based on semantic identifier sequence prefix matching rules. The entire clustering process only refers to the semantic feature information of the keywords and does not introduce auxiliary judgment conditions of other dimensions. The clustering logic is simple and fixed.

[0067] In search advertising scenarios, relying solely on semantic dimensions for clustering has limitations. Keywords with highly similar literal meanings and identical semantic prefixes may differ significantly in actual targeting attributes, settings, and business affiliations. Clustering based solely on semantic features is prone to mis-clustering and misjudgment, leading to keywords with different business attributes being incorrectly grouped into the same cluster, impacting the accuracy of subsequent keyword merging, index compression, and ad recall. To effectively compensate for the shortcomings of pure semantic clustering, correct biases in initial clustering results, and improve overall clustering accuracy and business adaptability, this solution adds a secondary correction process based on the initial semantic clustering. The system will further retrieve the complete keyword settings information corresponding to each keyword object, covering multiple dimensions such as advertiser's targeting, keyword matching mode, bidding strategy, placement scenario, and the category of the promoted goods and services. Combining these actual business setting characteristics, the system will verify, split, and reorganize all the generated initial keyword clusters one by one, remove abnormal keywords within the cluster that are semantically similar but do not match the business attributes, optimize the clustering structure, and finally generate a more accurate final keyword cluster that fits the actual placement business.

[0068] In the methods provided in the embodiments of this specification, the keyword object can be understood as the advertiser. That is, further modifications are needed based on the keyword settings information set in the advertiser. In a specific embodiment provided in this specification, each initial keyword cluster is modified according to the keyword settings information of each keyword object to obtain at least one keyword cluster, including: Determine co-occurrence constraint information based on the keyword settings information of each keyword object; Based on the co-occurrence constraint information, each initial keyword cluster is modified to obtain at least one keyword cluster.

[0069] Co-occurrence constraint information serves as the basis for determining the relationships between elements. Specifically, it refers to hard constraints that require two or more different elements to appear together or prohibit their simultaneous occurrence. It also includes combination preference characteristics derived from historical data statistics. In data processing and content combination, this information can effectively restrict the combination of various elements, automatically filtering out illogical or non-business-logical combinations. This optimizes the output content from a rule-based perspective, improving the overall rationality and practical value of the final processing results.

[0070] In the search advertising business scenario, the co-occurrence relationship between keywords has strong commercial reference value. The platform analyzes advertisers' historical campaign data and current campaign rules. When it discovers two or more initial keywords frequently used together by different advertisers in their campaign configurations and appearing together in the same campaign strategy, it can be determined that these keywords have a high degree of relevance at the commercial semantic level. These keywords are used to promote similar products, services, or target traffic, and their target user groups and business needs highly overlap. According to clustering logic, they should be uniformly classified into the same keyword cluster.

[0071] In the past, clustering based solely on semantic identifiers often focused only on the textual semantic features of keywords, easily overlooking the relationships formed by keywords in actual commercial advertising scenarios. This resulted in some commercially relevant keywords being split into different clusters, affecting subsequent index optimization and ad recall. Introducing co-occurrence constraints as supplementary criteria effectively compensates for the shortcomings of pure semantic clustering, fully exploring the commercial relevance behind keywords, and reasonably correcting the initial clustering results. This ensures that the final keyword clusters not only meet the requirement of semantic similarity but also align with the actual business logic of ad placement.

[0072] See Figure 2 , Figure 2 This diagram illustrates a method for generating keyword clusters according to an embodiment of this specification. Figure 2 As shown, taking running shoes, running shoes, and athletic shoes as examples, the semantic vector is obtained after semantic encoding. Then, after residual quantization of the semantic vector, a semantic identifier sequence for each initial keyword is obtained. With the cluster semantic identifier information set to 4, two initial clusters can be formed: cluster A for athletic shoes and cluster B for running shoes and running shoes. In practical applications, advertisers often group athletic shoes and running shoes together when placing ads. Therefore, the co-occurrence constraint information is "running shoes and athletic shoes are similar." Based on this co-occurrence constraint information, the initial clusters are modified to group athletic shoes and running shoes into the same cluster, resulting in cluster A, which includes athletic shoes, running shoes, and running shoes.

[0073] Through the clustering and correction operations performed in this step, the system can efficiently complete the automated clustering and aggregation of massive amounts of initial keywords by relying on standardized semantic identifier prefix matching rules combined with commercial co-occurrence constraint information. Compared with the traditional clustering method that calculates vector similarity one by one, this significantly reduces the overall computational complexity and enables fast and efficient clustering of large batches of keywords. The entire clustering mechanism uses semantic features as the core criterion and strictly divides clustering units according to the semantic identifier prefix matching rules. This fundamentally ensures the semantic similarity of all initial keywords within each keyword cluster, ensuring that keywords grouped in the same cluster have highly overlapping semantic features and effectively avoiding the problem of keywords with large semantic differences being mixed into the same cluster. At the same time, the clustering results after co-occurrence constraint correction take into account both text semantic similarity and advertising commercial relevance characteristics, making the clustering results more consistent with the actual business logic of search advertising. After accurate clustering, the originally scattered, massive, and disordered initial keywords were integrated into multiple regular, independent, and characteristically consistent keyword clusters. This provided a regular data foundation for subsequent core processing steps such as keyword merging, virtual word generation, index simplification, and ad matching and recall. Staff can perform batch and standardized processing on individual clusters with consistent characteristics without having to operate on individual keywords one by one, which greatly simplifies the subsequent data processing process and improves the overall system's operational efficiency and business processing consistency.

[0074] Step 104: Generate virtual keywords corresponding to each keyword cluster and determine the keyword correspondence between each initial keyword and each virtual keyword.

[0075] The core objective of this step is to automatically generate a representative virtual keyword that fully covers the core semantics of each keyword cluster obtained after the system's clustering correction, using a unified rule-based operation and feature integration method. Since all initial keywords within the same cluster possess highly similar textual semantics and commercial placement relevance, the generated virtual keyword can accurately replace and represent all initial keywords within that cluster. It can fully map the matching features and triggering capabilities of all original bid terms, effectively replacing a massive number of similar initial keywords in subsequent indexing, storage, retrieval, matching, and ad recall processes.

[0076] After determining the corresponding virtual keywords for each keyword cluster in batches using algorithms, the system automatically builds and solidifies a one-to-one mapping relationship between each keyword cluster and its corresponding virtual keyword, binding and archiving the overall characteristics of the cluster with the virtual keywords. Based on this mapping relationship between clusters and virtual words, it is possible to further break down and deduce to subdivided dimensions, accurately establishing a unique keyword mapping relationship between each initial keyword and its corresponding virtual keyword. This allows the system to accurately locate the cluster unit to which any original ad placement keyword belongs and its corresponding alternative virtual keyword. The establishment of this mapping relationship enables the normalized and consolidated management of massive amounts of original initial keywords, providing a stable and standardized data mapping foundation for subsequent simplification of the index database, replacement of duplicate semantic keywords, reduction of system retrieval computation, and optimization of the ad keyword matching link.

[0077] For example, taking keyword cluster A with initial keyword 1, initial keyword 2, and initial keyword 3 as an example, after analyzing keyword cluster A, we determine that its corresponding virtual keyword is virtual keyword 1. Therefore, the keyword correspondence can be defined as: "Initial keyword 1 corresponds to virtual keyword 1, initial keyword 2 corresponds to virtual keyword 1, and initial keyword 3 corresponds to virtual keyword 1." Establishing this keyword correspondence between initial keywords and virtual keywords facilitates subsequent reduction of index information, decreasing the number of indexes and improving data retrieval efficiency.

[0078] In one specific embodiment provided in this specification, virtual keywords corresponding to each keyword cluster are generated, including: Determine the keyword cluster to be processed, where the keyword cluster to be processed is any one of the keyword clusters; Each initial keyword in the cluster of keywords to be processed is input into the keyword generation model to obtain virtual keywords generated by the keyword generation model for the cluster of keywords to be processed.

[0079] In this embodiment, a more detailed and complete explanation of the principles and processes of the virtual keyword generation method corresponding to the keyword clusters will be provided. In the overall technical solution, all keyword clusters obtained by the system after semantic clustering and commercial co-occurrence constraint correction have completely unified virtual keyword generation logic, operation rules, and processing flows. The execution standards for all clusters remain highly consistent, with no differentiated processing rules. To avoid repetition and simplify the explanation of the principles, while clearly and intuitively demonstrating the complete virtual keyword generation steps, this embodiment selects any one of the many keyword clusters as an example, defining it as the keyword cluster to be processed. This selected keyword cluster to be processed has universality, fully representing the processing logic of all clusters. The entire generation process and technical principles explained based on this sample cluster are equally applicable to all other keyword clusters obtained by the system.

[0080] After identifying the keyword clusters to be processed, the initial keywords within each cluster can be input into the keyword generation model. The model will then be guided by corresponding virtual keyword generation prompts to generate a virtual keyword that best summarizes the initial keywords within the cluster. The virtual keyword generation prompts will include multiple examples demonstrating how to extract virtual words from a set of synonyms, helping the keyword generation model understand the rules for generating virtual keywords. Inputting the initial keywords into the keyword generation model can be done by assembling them into an initial keyword text and directly inputting it into the model.

[0081] In practical applications, due to limitations in cluster granularity and keyword popularity, the number of initial keywords in different keyword clusters varies. For keyword clusters with a small number of keywords, they can be directly input into the keyword generation model. However, if the number of initial keywords in the keyword cluster to be processed is large, problems such as increased computational latency and truncation in the keyword generation model will occur. Therefore, in the specific implementation provided in this specification, each initial keyword in the keyword cluster to be processed is input into the keyword generation model, including: Count the number of keywords in each initial keyword within the cluster of keywords to be processed; If the number of keywords exceeds a preset threshold, the initial keywords in the cluster of keywords to be processed are sorted according to a preset sorting rule, and a preset number of initial keywords to be processed are obtained based on the sorting results. Input the initial keywords to be processed and the preset virtual keyword generation prompts into the keyword generation model.

[0082] In this embodiment, before inputting each initial keyword in the keyword cluster to be processed into the keyword generation model for processing, the number of keywords in each initial keyword in the keyword cluster to be processed is counted. If the number of keywords is less than or equal to a preset threshold, all the initial keywords are concatenated together and input into the keyword generation model for processing to obtain virtual keywords.

[0083] In actual business processing, the number of initial keywords contained in different keyword clusters varies significantly. Some clusters with broad semantics and wide commercial relevance may collect a massive number of initial keywords. When the total number of initial keywords within a single keyword cluster exceeds the preset threshold configured by the system, the excessive keyword data within the cluster will significantly increase the computational load on the keyword generation model. This will not only increase the model's computational cost and processing time but may also interfere with the model's training and inference logic due to redundant and low-quality keyword data, reducing the accuracy and representativeness of the final generated virtual keywords. To avoid these problems and ensure the stable, efficient, and accurate operation of the keyword generation model, it is necessary to perform targeted screening and filtering of all initial keywords in the keyword clusters to be processed before generating virtual keywords, eliminating invalid and low-value keyword samples.

[0084] During the screening process, the system uses standardized preset sorting rules to classify keywords into hierarchical levels. This solution includes several core indicators for ranking, primarily including the historical frequency of each initial keyword on the platform, the number of advertisers covered by a single keyword, and the traffic volume covered by keyword exposure. These can be flexibly selected based on actual business needs. Based on the selected sorting rules, the system quantitatively scores each initial keyword within the cluster, accurately calculating the business value and representative weight of each keyword. Subsequently, it strictly sorts the keywords from highest to lowest according to the final score, achieving hierarchical differentiation of keywords within the cluster.

[0085] After sorting, the system selects a fixed number of top-scoring initial keywords from the ordered keyword list as the final sample of keywords to be processed. The retained keywords are high-quality keywords with strong business representativeness, high placement weight, and core semantic features within the cluster. Finally, all the selected initial keywords are concatenated and integrated to form standardized input data, which is then batch-fed into the keyword generation model for subsequent processing. This reduces the amount of input data to the model while maximizing the preservation of the core semantics and business features of the clusters.

[0086] For example, taking a cluster of 1340 initial keywords as an example, with a preset threshold of 1000, the 1340 initial keywords are scored and sorted in descending order of scores. The top 1000 initial keywords are selected from the sorted results as the initial keywords to be processed, and these 1000 initial keywords are concatenated and input into the keyword generation model for processing.

[0087] This step, involving the screening, scoring, ranking, and sample selection of initial keywords within clusters, effectively filters out initial keywords with low placement value, weak semantic contribution, and data redundancy. It maximizes the retention of high-quality keyword samples with core representativeness and high business weight, comprehensively ensuring the overall quality and sample validity of the initial keywords input into the keyword generation model from the source. In conventional model inference mechanisms, the input text sequence has a fixed length limit. If a massive number of initial keywords from a cluster are input at once, it is very easy to exceed the character limit of the model's text input. The system will automatically trigger a text truncation mechanism, randomly discarding some keyword text content, which can lead to problems such as loss of core semantics, incomplete features, and information shift, seriously affecting the generation accuracy of virtual keywords and the clustering representation effect. After the pre-screening process, the number of input samples is greatly reduced while the core semantics and commercial advertising features of the clusters are fully preserved. This completely avoids the text truncation problem caused by too many initial keywords and too long input text. The model can perform calculations and analyses based on complete, high-quality, and unmissing keyword samples. The resulting virtual keywords can more accurately and comprehensively represent the overall semantic features and business attributes of the corresponding clusters.

[0088] In one specific embodiment provided in this specification, obtaining virtual keywords generated by the keyword generation model for the keyword cluster to be processed includes: The initial virtual keywords generated by the keyword generation model for the cluster of keywords to be processed are obtained; Calculate the semantic similarity between the initial virtual keyword and each initial keyword in the keyword cluster to be processed; If the semantic similarity is greater than or equal to the preset similarity threshold, the initial virtual keywords are determined as virtual keywords.

[0089] The keyword generation model employed in this solution is based on a mature large language model as its underlying architecture. Leveraging the powerful semantic understanding, text summarization, and content generation capabilities of this model, it is adapted to keyword aggregation and virtual word generation tasks in search advertising scenarios. During actual operation, the initial keywords from the keyword clusters to be processed, along with preset virtual keyword generation prompts, are input into the model. Through deep analysis and learning of the prompt instructions, the model accurately grasps the specific generation rules, semantic summarization logic, and standardized output paradigms adapted to the advertising keyword scenario. It clarifies that virtual keywords must meet core generation requirements such as semantic generalization, business adaptability, and concise representativeness. After completing the learning and adaptation of the generation rules, the model, based on the high-quality initial keyword sample data selected and assembled in the early stages, and combined with the learned scenario-based generation rules, performs overall semantic refinement, feature summarization, and content summarization on the batch of similar initial keywords. Ultimately, it automatically generates new virtual keywords that can accurately replace the entire cluster of keywords, possessing both semantic uniformity and commercial representativeness, achieving lightweight normalization processing of massive amounts of similar initial keywords.

[0090] In the method provided in the embodiments of this specification, the keyword generation model is a model based on a large language model, which has the ability to understand natural language. However, there are cases where the quality of the generated virtual keywords is poor. In order to ensure the quality of the generated virtual keywords, the method provided in the embodiments of this specification can also perform quality verification on the virtual keywords generated by the keyword generation model. Only virtual keywords that pass the verification can be used as the final virtual keywords.

[0091] Specifically, when each initial keyword is input into the keyword generation model, the keyword generation model can first generate initial virtual keywords for the keyword clusters to be processed. The initial virtual keywords are virtual keywords generated by the keyword generation model but have not yet been verified.

[0092] Next, the quality of the initial virtual keywords needs to be verified. Specifically, the semantic similarity between the initial virtual keywords and each initial keyword in the keyword cluster to be processed is calculated. If each semantic similarity is greater than or equal to a preset similarity threshold, it means that the initial virtual keyword meets the quality verification, and can then be used as the virtual keyword corresponding to the keyword cluster to be processed.

[0093] It is important to note that this involves calculating the semantic similarity with all initial keywords in the keyword cluster to be processed. If the initial keywords are generated from some of the initial keywords in the keyword cluster to be processed, then the initial keywords also need to have their semantic similarity calculated with all the initial keywords in the keyword cluster to be processed, thus ensuring that the virtual keywords can represent each initial keyword in the keyword cluster to be processed.

[0094] For example, continuing with the previous example, let's take 1340 initial keywords in the keyword cluster to be processed as an example for explanation. In the example above, 1000 initial keywords were selected and input into the keyword model for processing to obtain the initial virtual keyword 'a'. At this time, it is necessary to calculate the semantic similarity between the initial virtual keyword and the 1340 initial keywords respectively. If the semantic similarity of these 1340 keywords is greater than or equal to the preset similarity threshold, the initial virtual keyword 'a' is determined to be the virtual keyword of the keyword cluster to be processed.

[0095] In another specific embodiment provided in this specification, the method further includes: If there are initial keywords with semantic similarity less than the preset similarity threshold, the step of inputting each initial keyword in the keyword cluster to be processed into the keyword generation model is repeated.

[0096] In this embodiment, if the semantic similarity is less than the preset similarity threshold after calculating the semantic similarity between the initial virtual keywords and each initial keyword, it means that the initial virtual keywords at this time cannot represent each initial keyword in the keyword cluster to be processed. In this case, it is necessary to extract the virtual keywords of the keyword cluster to be processed again, that is, repeat the above steps of inputting each initial keyword into the keyword generation model.

[0097] The virtual keywords obtained through the clustering screening and model generation methods provided in the embodiments of this specification can highly match the core semantic features of all initial keywords within the corresponding keyword cluster, maintaining a high degree of semantic similarity and business adaptability with each original initial keyword within the cluster. These virtual keywords are not simple replicas of individual initial keywords, but rather comprehensive keywords generated by the model after comprehensively extracting, summarizing, and integrating the semantic commonalities, core features, and commercial placement attributes of all high-quality initial keywords within the cluster. They can fully cover the mainstream semantic expressions and traffic triggering characteristics of various initial keywords within the cluster. Therefore, these virtual keywords can accurately and uniformly represent all initial keywords in the entire cluster, fully possessing the ability to replace massive amounts of homogeneous initial keywords in subsequent business processes such as system index storage, keyword matching, user search retrieval, and advertising bidding display. While significantly reducing the volume of system keywords, it fully ensures that the semantic coverage effect and advertising triggering capability of the original keywords are not lost.

[0098] In another specific embodiment provided in this specification, determining the keyword correspondence between each initial keyword and each virtual keyword includes: Determine the initial keyword to be matched, wherein the initial keyword to be matched is any one of the initial keywords; Among the virtual keywords, at least two candidate virtual keywords are identified that correspond to the initial keyword to be matched. The initial keyword to be matched and each candidate virtual keyword are input into the keyword matching model to obtain the target virtual keyword determined by the keyword matching model from each candidate virtual keyword; Establish a keyword correspondence between the initial keyword to be matched and the target virtual keyword.

[0099] In the methods provided in the embodiments of this specification, although corresponding virtual keywords have been generated for each initial keyword through the above steps, since the virtual keywords are generated based on keyword clusters, the grouping may not be accurate enough due to the clustering. In order to ensure a more accurate correspondence between the initial keywords and the virtual keywords, more suitable virtual keywords can be selected from multiple virtual keywords based on the initial keywords.

[0100] For example, taking 100 initial keywords as an example, after clustering, 12 clusters are obtained. A corresponding virtual keyword is generated for each cluster. At this point, there are 100 initial keywords and 12 virtual keywords, with each initial keyword corresponding to one virtual keyword. For instance, initial keyword 1 is in cluster a, and it corresponds to virtual keyword 1. However, during the clustering process, if there are issues with the clustering granularity, initial keyword 1 might be a better match for virtual keyword 2. To solve this problem, we can use initial keyword 1 as a base and then find a more closely matching virtual keyword from the 12 virtual keywords to establish the keyword correspondence.

[0101] Based on this, a keyword correspondence is established between the initial keywords and each virtual keyword, with the initial keywords serving as the primary basis for filtering. To further illustrate this, the initial keyword to be matched is used as an example to explain the establishment of the keyword matching relationship. The initial keyword to be matched can be any one of the initial keywords.

[0102] After determining the initial keyword to be matched, at least two candidate virtual keywords corresponding to the initial keyword are selected from the virtual keywords. Candidate virtual keywords can be understood as virtual keywords that are semantically similar to the initial keyword to be matched.

[0103] The initial keyword to be matched and each candidate virtual keyword are then concatenated and input into the keyword matching model. The keyword matching model selects the target virtual keyword corresponding to the initial keyword to be matched from the candidate virtual keywords. This establishes the keyword correspondence between the initial keyword to be matched and the target virtual keyword.

[0104] Keyword matching models are also natural language processing models based on large language models. After determining multiple candidate virtual keywords for the initial keyword to be matched, the semantic understanding and analysis capabilities of large language models can be used to select virtual keywords that are more relevant to the initial keyword to be matched from the candidate virtual keywords.

[0105] In the method provided in the embodiments of this specification, determining at least two candidate virtual keywords corresponding to the initial keyword to be matched from each virtual keyword includes: Obtain the initial keyword vector of the initial keyword to be matched, and obtain the virtual keyword vector of each virtual keyword; The approximate vector similarity is calculated based on the initial keyword vector to be matched and each virtual keyword vector, and at least two candidate virtual keywords are determined based on the calculation results.

[0106] In this implementation, to find candidate virtual keywords corresponding to the initial keywords to be matched, both the initial keywords and the virtual keywords can be vector-encoded. Specifically, the initial keywords to be matched are input into a pre-trained BERT model for processing, encoding them into dense vectors of initial keywords to be matched. Similarly, the same processing is performed on each virtual keyword to obtain its corresponding virtual keyword vector. Converting keywords into their respective vectors ensures that semantically similar keywords have a high similarity in the vector space.

[0107] Subsequently, the similarity between the initial keyword vector to be matched and each virtual keyword vector is calculated, and a preset number of virtual keywords are selected as candidate virtual keywords for the initial keyword to be matched based on the similarity.

[0108] In practical applications, due to the large number of initial keywords, the number of corresponding keyword clusters will also be large, resulting in a large number of virtual keywords. If vector calculation is performed for each initial keyword, the subsequent calculation of vector similarity will be extremely time-consuming.

[0109] In the similarity matching retrieval process between the initial keyword vector and the virtual keyword vector, to balance retrieval efficiency and matching accuracy, this solution employs ANN (Approximate Nearest Neighbor) technology to complete batch retrieval processing. In traditional traversal retrieval methods, the system needs to compare and calculate each vector in the virtual keyword vector set one by one. When the virtual keyword vector library is large, this incurs extremely high computational overhead, significantly increasing retrieval time and failing to meet the performance requirements of real-time retrieval in advertising systems. ANN (Approximate Nearest Neighbor) technology, however, can optimize and accelerate the high-dimensional vector retrieval process within a controllable accuracy error range, significantly reducing the computational complexity of the retrieval.

[0110] During execution, the system retrieves a pre-constructed set of virtual keyword vectors, using the initial keyword vectors to be matched as the search targets. Leveraging an ANN retrieval algorithm, it quickly calculates the approximate similarity between the target vector and all virtual keyword vectors in the set, accurately locating virtual keyword vectors that are close in distance and have similar features in the semantic space. After completing a full similarity comparison and filtering, the system sorts the vectors according to their similarity scores from highest to lowest, selecting a predetermined number of highly matched virtual keywords as candidate virtual keywords for this vector matching retrieval. This provides accurate and effective candidate materials for subsequent keyword mapping verification and ad matching recall.

[0111] In the method provided in the embodiments of this specification, matching virtual keywords are selected for the initial keywords through a two-stage screening process. The first stage employs a coarse-grained screening method, using ANN technology to select at least two candidate virtual keywords from multiple virtual keywords. The second stage employs a fine-grained screening method, inputting the initial keywords and each candidate virtual keyword into a keyword matching model to obtain virtual keywords that better match the initial keywords, selected by the keyword matching model from the candidate virtual keywords. ANN technology, while sacrificing a small amount of accuracy, quickly selects candidate virtual keywords, greatly compressing retrieval time and meeting latency requirements. Furthermore, virtual keywords are selected from the candidate virtual keywords through a large language model re-ranking method, which has stronger semantic discriminative ability compared to methods that only use vector similarity calculations.

[0112] Step 106: Generate a target keyword index based on the initial keyword index and the keyword correspondence, wherein the target keyword index includes virtual keywords and keyword objects corresponding to the virtual keywords.

[0113] The above steps have established a keyword correspondence between the initial keywords and the virtual keywords. To reduce the storage size of the initial keyword index and improve indexing efficiency, a new target keyword index needs to be generated based on the keyword correspondence and the initial keyword index. The target keyword index stores the correspondence between the virtual keywords and the keyword objects.

[0114] The initial keyword index is the original keyword index stored in the current system. For example, in an advertising system, the initial keyword index is an index of the correspondence between advertisers and their bid keywords. Taking 10 advertisers as an example, if each advertiser bids for 11 keywords, then the initial keyword index will have 110 records, recording the 11 bid keywords bid by each of the 10 advertisers. As the number of bid keywords placed by advertisers increases, the size of the initial keyword index will also increase, thus slowing down data retrieval efficiency.

[0115] In practical applications, advertising systems only need to know which advertisers can match the user's query. Therefore, a new target keyword index can be created based on the initial keyword index and keyword correspondences. This target keyword index stores the correspondence between advertisers and virtual keywords. If a virtual keyword is matched, the corresponding advertiser can be found through this target keyword index. The number of virtual keywords is much smaller than the number of initial keywords; therefore, this step reduces the amount of data in the index, thereby improving data retrieval efficiency.

[0116] In one specific embodiment provided in this specification, generating a target keyword index based on an initial keyword index and the keyword correspondence includes: Obtain the initial keyword index, which includes the initial keyword and the keyword object corresponding to the initial keyword; Based on the correspondence between the initial keywords and virtual keywords in the keyword correspondence, update the initial keywords in the initial keyword index to generate the target keyword index.

[0117] In this step, the initial keyword index is first obtained, which includes the initial keyword and the keyword object corresponding to the initial keyword. Then, the initial keyword in the initial keyword index is replaced by the correspondence between the initial keyword and the virtual keyword in the keyword correspondence relationship to generate the target keyword index.

[0118] For example, Table 1 illustrates the initial keyword index. Table 2 illustrates the keyword correspondence.

[0119] Table 1

[0120] Table 2

[0121] By merging Table 1 and Table 2, a schematic illustration of the target keyword index can be generated as shown in Table 3 below.

[0122] Table 3

[0123] Table 2 shows the keyword correspondence. Initial keyword 1, initial keyword 3, and initial keyword 4 all correspond to virtual keyword 1. Therefore, virtual keyword 1 can be used to replace initial keyword 1, initial keyword 3, and initial keyword 4, simplifying the three records in the initial keyword index into one condition record in the target keyword index.

[0124] The target keyword index and the initial keyword index can also be decoupled from each other. The data structure of the target keyword index and the initial keyword index is consistent, but the terms in the index are replaced with virtual keywords instead of the initial keywords. The target keyword index can be built, updated, and maintained independently of the initial keyword index, reducing the coupling of the system.

[0125] In subsequent maintenance, incremental updates can be performed according to preset time intervals (such as one hour, twelve hours, one day, three days, seven days, fifteen days, etc.). That is, the method provided in the embodiments of this specification is executed on the initial keywords added within the preset time interval to generate virtual keywords corresponding to the newly added initial keywords and update the target keyword index to ensure the timeliness and effectiveness of the data.

[0126] Step 108: Determine the target keyword object corresponding to the query keyword based on the target keyword index.

[0127] Once the target keyword index is obtained, it can be deployed online to the advertising system. The target keyword index can be used to find the target keyword objects that match the query keywords entered by the user.

[0128] In current advertising systems, users input query keywords, and the system searches for these keywords in an initial keyword index to find matching keyword objects (i.e., advertisers). Due to the large number of keywords in the initial keyword index, this search is time-consuming and resource-intensive. The method provided in the embodiments of this specification allows for keyword matching through a target keyword index. Specifically, after receiving the user's input query keywords, the system searches for these keywords in the target keyword index to determine the target keyword objects that match the query keywords.

[0129] In one specific embodiment provided in this specification, determining the target keyword object corresponding to the query keyword based on the target keyword index includes: Receive query keywords and determine the corresponding virtual query keywords based on the query keywords and the keyword correspondence; Based on the query virtual keywords, the target keyword object is determined by searching in the target keyword index.

[0130] In practical applications, the above steps generate corresponding virtual keywords for each initial keyword in order to ensure consistent query results.

[0131] Specifically, after receiving the query keywords, the corresponding virtual query keywords can be determined based on the query keywords and the keyword correspondence.

[0132] To ensure search effectiveness, you can first expand the search keywords to generate multiple similar search keywords with similar semantics. Then, generate search keyword clusters from the search keywords and similar search keywords, and finally generate virtual search keywords from the search keyword clusters.

[0133] Then, based on the query virtual keywords, a search is performed in the target keyword index to determine the target keyword object.

[0134] The method provided in the embodiments of this specification, by clustering the initial keywords and generating corresponding virtual keywords for each cluster, can initially compress a large number of initial keywords into a smaller number of virtual keywords, thereby reducing the number of keywords, reducing subsequent data processing time, and improving data processing efficiency. A keyword correspondence is established between the initial keywords and virtual keywords, and the initial keyword index is updated based on this correspondence. A target keyword index, decoupled from the initial keyword index, is generated. In subsequent retrieval processes, the new target keyword index is used, improving indexing efficiency by reducing the number of index entries. The decoupling and separate updating and maintenance of the initial keyword index and the target keyword index reduces system coupling and improves system stability.

[0135] Secondly, in the process of clustering initial keywords to generate virtual keywords, semantic identifier sequence clustering and a large language model were used to construct virtual keywords. First, the initial keywords were discretized into semantic vectors through residual quantization, generating multi-layer semantic ID sequences and performing hierarchical clustering based on prefix matching. Then, clustering was corrected by incorporating co-occurrence information of advertiser-purchased keywords. Finally, a large language model was used to generate virtual keywords from the clusters, achieving semantic deduplication and aggregation of large-scale synonymous keywords. This significantly reduced the index size while increasing the collision probability between query terms and bid terms, thereby improving the ad recall effect.

[0136] Finally, in the ad recall phase, a target keyword index was constructed that was decoupled from the initial keyword index. This was combined with a distributed cache to achieve online mapping of query keywords and bid keywords to virtual keywords. A two-stage method of "ANN vector coarse screening + large language model Reranker fine ranking" was adopted to support the incremental mapping of new keywords. This achieved a smooth migration and efficient recall of the index system without interrupting existing services.

[0137] The following is in conjunction with the appendix Figure 3 Taking the application of the data processing method provided in this specification in an advertising recall scenario as an example, the data processing method will be further explained. Among other things, Figure 3 This specification illustrates a data processing method for an advertising recall scenario according to an embodiment of the present invention, which specifically includes the following steps: Step 302: Obtain the initial keyword index, which includes the advertiser and the advertiser's corresponding initial keywords.

[0138] Step 304: Embed the initial keywords to obtain the initial keyword vectors corresponding to each initial keyword, and convert each initial keyword vector into a semantic tag sequence with a preset number of tags.

[0139] Step 306: Determine the clustering semantic identifier information, and cluster each initial keyword according to the clustering semantic identifier information to obtain at least one initial keyword cluster.

[0140] Step 308: Correct each initial keyword cluster according to the keyword setting information of each keyword object to obtain at least one keyword cluster.

[0141] Step 310: Input each initial keyword in each keyword cluster into the keyword generation model to obtain the initial virtual keywords generated by the keyword generation model for each keyword cluster.

[0142] Step 312: Calculate the semantic similarity between the initial virtual keywords and each initial keyword in each keyword cluster.

[0143] Step 314: If the semantic similarity is greater than or equal to the preset similarity threshold, determine the initial virtual keywords as the virtual keywords corresponding to the keyword clusters.

[0144] Step 316: Determine at least two candidate virtual keywords for each initial keyword.

[0145] Step 318: Input the initial keyword and each candidate virtual keyword into the keyword matching model to obtain the target virtual keyword determined by the keyword matching model for the initial keyword from each candidate virtual keyword.

[0146] Step 320: Construct keyword correspondence based on each initial keyword and the corresponding target virtual keyword.

[0147] Step 322: Generate a target keyword index based on the keyword correspondence and the initial keyword index, and determine the target advertiser corresponding to the query keyword based on the target keyword index.

[0148] The method provided in the embodiments of this specification, by clustering the initial keywords and generating corresponding virtual keywords for each cluster, can initially compress a large number of initial keywords into a smaller number of virtual keywords, thereby reducing the number of keywords, reducing subsequent data processing time, and improving data processing efficiency. A keyword correspondence is established between the initial keywords and virtual keywords, and the initial keyword index is updated based on this correspondence. A target keyword index, decoupled from the initial keyword index, is generated. In subsequent retrieval processes, the new target keyword index is used, improving indexing efficiency by reducing the number of index entries. The decoupling and separate updating and maintenance of the initial keyword index and the target keyword index reduces system coupling and improves system stability.

[0149] Secondly, in the process of clustering initial keywords to generate virtual keywords, semantic identifier sequence clustering and a large language model were used to construct virtual keywords. First, the initial keywords were discretized into semantic vectors through residual quantization, generating multi-layer semantic ID sequences and performing hierarchical clustering based on prefix matching. Then, clustering was corrected by incorporating co-occurrence information of advertiser-purchased keywords. Finally, a large language model was used to generate virtual keywords from the clusters, achieving semantic deduplication and aggregation of large-scale synonymous keywords. This significantly reduced the index size while increasing the collision probability between query terms and bid terms, thereby improving the ad recall effect.

[0150] Finally, in the ad recall phase, a target keyword index was constructed that was decoupled from the initial keyword index. This was combined with a distributed cache to achieve online mapping of query keywords and bid keywords to virtual keywords. A two-stage method of "ANN vector coarse screening + large language model Reranker fine ranking" was adopted to support the incremental mapping of new keywords. This achieved a smooth migration and efficient recall of the index system without interrupting existing services.

[0151] Corresponding to the above method embodiments, this specification also provides data processing apparatus embodiments. Figure 4 A schematic diagram of the structure of a data processing apparatus provided in one embodiment of this specification is shown. For example... Figure 4 As shown, the device includes: The acquisition module 402 is configured to acquire the semantic identifier sequence corresponding to each initial keyword, and cluster each initial keyword according to each semantic identifier sequence to obtain at least one keyword cluster, wherein each keyword cluster includes at least one initial keyword; The generation module 404 is configured to generate virtual keywords corresponding to each keyword cluster and determine the keyword correspondence between each initial keyword and each virtual keyword; The index building module 406 is configured to generate a target keyword index based on the initial keyword index and the keyword correspondence, wherein the target keyword index includes virtual keywords and keyword objects corresponding to the virtual keywords; The query module 408 is configured to determine the target keyword object corresponding to the query keyword based on the target keyword index.

[0152] In a specific embodiment provided in this specification, the acquisition module 402 is further configured as follows: Each initial keyword is embedded to obtain the initial keyword vector corresponding to each initial keyword; Each initial keyword vector is converted into a semantic tag sequence with a preset number of tags.

[0153] In a specific embodiment provided in this specification, the acquisition module 402 is further configured as follows: Determine the clustering semantic identifier information, and cluster each initial keyword according to the clustering semantic identifier information to obtain at least one initial keyword cluster; Based on the keyword settings of each keyword object, each initial keyword cluster is modified to obtain at least one keyword cluster.

[0154] In a specific embodiment provided in this specification, the acquisition module 402 is further configured as follows: Determine co-occurrence constraint information based on the keyword settings information of each keyword object; Based on the co-occurrence constraint information, each initial keyword cluster is modified to obtain at least one keyword cluster.

[0155] In a specific embodiment provided in this specification, the generation module 404 is further configured as follows: Determine the keyword cluster to be processed, where the keyword cluster to be processed is any one of the keyword clusters; Each initial keyword in the cluster of keywords to be processed is input into the keyword generation model to obtain virtual keywords generated by the keyword generation model for the cluster of keywords to be processed.

[0156] In a specific embodiment provided in this specification, the generation module 404 is further configured as follows: Count the number of keywords in each initial keyword within the cluster of keywords to be processed; If the number of keywords exceeds a preset threshold, the initial keywords in the cluster of keywords to be processed are sorted according to a preset sorting rule, and a preset number of initial keywords to be processed are obtained based on the sorting results. Input the initial keywords to be processed and the preset virtual keyword generation prompts into the keyword generation model.

[0157] In a specific embodiment provided in this specification, the generation module 404 is further configured as follows: The initial virtual keywords generated by the keyword generation model for the cluster of keywords to be processed are obtained; Calculate the semantic similarity between the initial virtual keyword and each initial keyword in the keyword cluster to be processed; If the semantic similarity is greater than or equal to the preset similarity threshold, the initial virtual keywords are determined as virtual keywords.

[0158] In a specific embodiment provided in this specification, the generation module 404 is further configured to: If there are initial keywords with semantic similarity less than the preset similarity threshold, the step of inputting each initial keyword in the keyword cluster to be processed into the keyword generation model is repeated.

[0159] In a specific embodiment provided in this specification, the generation module 404 is further configured as follows: Determine the initial keyword to be matched, wherein the initial keyword to be matched is any one of the initial keywords; Among the virtual keywords, at least two candidate virtual keywords are identified that correspond to the initial keyword to be matched. The initial keyword to be matched and each candidate virtual keyword are input into the keyword matching model to obtain the target virtual keyword determined by the keyword matching model from each candidate virtual keyword; Establish a keyword correspondence between the initial keyword to be matched and the target virtual keyword.

[0160] In a specific embodiment provided in this specification, the generation module 404 is further configured as follows: Obtain the initial keyword vector of the initial keyword to be matched, and obtain the virtual keyword vector of each virtual keyword; The approximate vector similarity is calculated based on the initial keyword vector to be matched and each virtual keyword vector, and at least two candidate virtual keywords are determined based on the calculation results.

[0161] In a specific embodiment provided in this specification, the index building module 406 is further configured as follows: Obtain the initial keyword index, which includes the initial keyword and the keyword object corresponding to the initial keyword; Based on the correspondence between the initial keywords and virtual keywords in the keyword correspondence, update the initial keywords in the initial keyword index to generate the target keyword index.

[0162] In a specific embodiment provided in this specification, the query module 408 is further configured as follows: Receive query keywords and determine the corresponding virtual query keywords based on the query keywords and the keyword correspondence; Based on the query virtual keywords, the target keyword object is determined by searching in the target keyword index.

[0163] The apparatus provided in the embodiments of this specification, by clustering the initial keywords and generating corresponding virtual keywords for each cluster, can initially compress a large number of initial keywords into a smaller number of virtual keywords, thereby reducing the number of keywords, shortening the processing time for subsequent data processing, and improving data processing efficiency. A keyword correspondence is established between the initial keywords and virtual keywords, and the initial keyword index is updated based on this correspondence. A target keyword index, decoupled from the initial keyword index, is generated. In subsequent retrieval processes, the new target keyword index is used for searching, improving indexing efficiency by reducing the number of index entries. The decoupling and separate updating and maintenance of the initial keyword index and the target keyword index reduces system coupling and improves system stability.

[0164] Secondly, in the process of clustering initial keywords to generate virtual keywords, semantic identifier sequence clustering and a large language model were used to construct virtual keywords. First, the initial keywords were discretized into semantic vectors through residual quantization, generating multi-layer semantic ID sequences and performing hierarchical clustering based on prefix matching. Then, clustering was corrected by incorporating co-occurrence information of advertiser-purchased keywords. Finally, a large language model was used to generate virtual keywords from the clusters, achieving semantic deduplication and aggregation of large-scale synonymous keywords. This significantly reduced the index size while increasing the collision probability between query terms and bid terms, thereby improving the ad recall effect.

[0165] Finally, in the ad recall phase, a target keyword index was constructed that was decoupled from the initial keyword index. This was combined with a distributed cache to achieve online mapping of query keywords and bid keywords to virtual keywords. A two-stage method of "ANN vector coarse screening + large language model Reranker fine ranking" was adopted to support the incremental mapping of new keywords. This achieved a smooth migration and efficient recall of the index system without interrupting existing services.

[0166] The above is an illustrative scheme of a data processing apparatus according to this embodiment. It should be noted that the technical solution of this data processing apparatus and the technical solution of the data processing method described above belong to the same concept. For details not described in detail in the technical solution of the data processing apparatus, please refer to the description of the technical solution of the data processing method described above.

[0167] See Figure 5 , Figure 5 This specification illustrates an architecture diagram of a data processing system according to one embodiment of the present specification. The data processing system may include a client 100 and a server 200. Client 100 is used to send an initial keyword index to server 200, wherein the initial keyword index includes an initial keyword and the keyword object corresponding to the initial keyword; Server 200 is used to obtain the semantic identifier sequence corresponding to each initial keyword, and cluster each initial keyword according to each semantic identifier sequence to obtain at least one keyword cluster, wherein each keyword cluster includes at least one initial keyword; generate virtual keywords corresponding to each keyword cluster, and determine the keyword correspondence between each initial keyword and each virtual keyword; generate a target keyword index according to the initial keyword index and the keyword correspondence, wherein the target keyword index includes virtual keywords and keyword objects corresponding to the virtual keywords; and send the target keyword index to client 100. Client 100 is also used to receive the target keyword index sent by server 200 and determine the target keyword object corresponding to the query keyword based on the target keyword index.

[0168] The data processing system may include multiple clients 100 and a server 200. Clients 100 can be referred to as edge devices, and server 200 can be referred to as cloud devices. Multiple clients 100 can establish communication connections through server 200. In an index update scenario, server 200 is used to provide index update services among multiple clients 100. Each client 100 can act as either a sender or a receiver, communicating through server 200.

[0169] Users can interact with server 200 through client 100 to receive data sent by other clients 100, or send data to other clients 100, etc. In the index update scenario, users can publish data streams to server 200 through client 100, server 200 can generate target keyword indexes based on the data streams, and push the target keyword indexes to other clients that have established communication.

[0170] In this system, client 100 and server 200 establish a connection via a network. The network provides the medium for communication between client 100 and server 200. The network can include various connection types, such as wired or wireless communication links or fiber optic cables. Data transmitted by client 100 may need to undergo encoding, transcoding, compression, or other processing before being published to server 200.

[0171] Client 100 can be a browser, an app (application), a web application such as an H5 (HyperText Markup Language 5) application, a lightweight application (also known as a mini-program), or a cloud application. Client 100 can be developed based on the software development kit (SDK) of the corresponding service provided by server 200, such as a real-time communication (RTC) SDK. Client 100 can be deployed on a computing device and depends on the device or certain apps on the device to run. The computing device may have a display screen and support information browsing, such as a personal mobile terminal like a mobile phone, tablet, or personal computer. Various other types of applications can also be configured on the computing device, such as human-computer interaction applications, model training applications, text processing applications, web browser applications, shopping applications, search applications, instant messaging tools, email clients, and social media platform software.

[0172] Server 200 may include servers providing various services, such as servers providing communication services to multiple clients, servers supporting backend training of models used on clients, and servers processing data sent by clients. It should be noted that server 200 can be implemented as a distributed server cluster composed of multiple servers, or as a single server. The server can also be a server in a distributed system, or a server integrated with blockchain. The server can also be a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms, or an intelligent cloud computing server or intelligent cloud host with artificial intelligence technology.

[0173] It is worth noting that the data processing methods provided in the embodiments of this specification are generally executed by the server. However, in other embodiments of this specification, the client may also have similar functions to the server, thereby executing the data processing methods provided in the embodiments of this specification. In other embodiments, the data processing methods provided in the embodiments of this specification may also be executed jointly by the client and the server.

[0174] Figure 6 A structural block diagram of a computing device 600 according to an embodiment of this specification is shown. The components of the computing device 600 include, but are not limited to, a memory 610 and a processor 620. The processor 620 is connected to the memory 610 via a bus 630, and a database 650 is used to store data.

[0175] The computing device 600 also includes an access device 640, which enables the computing device 600 to communicate via one or more networks 660. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 640 may include one or more of any type of wired or wireless network interface (e.g., a network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, a Near Field Communication (NFC) interface, and so on.

[0176] In one embodiment of this specification, the above-described components of the computing device 600 and Figure 6 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 6 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0177] The computing device 600 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 600 can also be a mobile or stationary server.

[0178] The processor 620 is used to execute the following computer program / instructions, which, when executed by the processor, implement the steps of the above-described data processing method.

[0179] The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solution of the data processing method described above belong to the same concept. For details not described in detail in the technical solution of the computing device, please refer to the description of the technical solution of the data processing method described above.

[0180] An embodiment of this specification also provides a computer-readable storage medium storing a computer program / instructions that, when executed by a processor, implement the steps of the above-described data processing method.

[0181] The above is an illustrative embodiment of a computer-readable storage medium. It should be noted that the technical solution of this storage medium and the technical solution of the data processing method described above belong to the same concept. Details not described in detail in the technical solution of the storage medium can be found in the description of the technical solution of the data processing method described above.

[0182] An embodiment of this specification also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described data processing method.

[0183] The above is an illustrative scheme of a computer program product according to this embodiment. It should be noted that the technical solution of this computer program product and the technical solution of the data processing method described above belong to the same concept. For details not described in detail in the technical solution of the computer program product, please refer to the description of the technical solution of the data processing method described above.

[0184] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0185] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0186] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this specification is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this specification. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this specification.

[0187] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0188] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. These embodiments have been selected and specifically described in this specification to better explain the principles and practical applications of this specification, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A data processing method, characterized in that, include: Obtain the semantic identifier sequence corresponding to each initial keyword, and cluster each initial keyword according to each semantic identifier sequence to obtain at least one keyword cluster, wherein each keyword cluster includes at least one initial keyword; Generate virtual keywords corresponding to each keyword cluster, and determine the keyword correspondence between each initial keyword and each virtual keyword; A target keyword index is generated based on the initial keyword index and the keyword correspondence, wherein the target keyword index includes virtual keywords and keyword objects corresponding to the virtual keywords; The target keyword object corresponding to the query keyword is determined based on the target keyword index.

2. The method as described in claim 1, characterized in that, Obtain the semantic identifier sequence corresponding to each initial keyword, including: Each initial keyword is embedded to obtain the initial keyword vector corresponding to each initial keyword; Each initial keyword vector is converted into a semantic tag sequence with a preset number of tags.

3. The method as described in claim 1, characterized in that, Clustering of each initial keyword based on each semantic identifier sequence yields at least one keyword cluster, including: Determine the clustering semantic identifier information, and cluster each initial keyword according to the clustering semantic identifier information to obtain at least one initial keyword cluster; Based on the keyword settings of each keyword object, each initial keyword cluster is modified to obtain at least one keyword cluster.

4. The method as described in claim 3, characterized in that, Based on the keyword settings information of each keyword object, each initial keyword cluster is corrected to obtain at least one keyword cluster, including: Determine co-occurrence constraint information based on the keyword settings information of each keyword object; Based on the co-occurrence constraint information, each initial keyword cluster is modified to obtain at least one keyword cluster.

5. The method as described in claim 1, characterized in that, Generate virtual keywords corresponding to each keyword cluster, including: Determine the keyword cluster to be processed, where the keyword cluster to be processed is any one of the keyword clusters; Each initial keyword in the cluster of keywords to be processed is input into the keyword generation model to obtain virtual keywords generated by the keyword generation model for the cluster of keywords to be processed.

6. The method as described in claim 5, characterized in that, Inputting each initial keyword from the keyword cluster to be processed into the keyword generation model includes: Count the number of keywords in each initial keyword within the cluster of keywords to be processed; If the number of keywords exceeds a preset threshold, the initial keywords in the cluster of keywords to be processed are sorted according to a preset sorting rule, and a preset number of initial keywords to be processed are obtained based on the sorting results. Input the initial keywords to be processed and the preset virtual keyword generation prompts into the keyword generation model.

7. The method as described in claim 5, characterized in that, The virtual keywords generated by the keyword generation model for the cluster of keywords to be processed include: The initial virtual keywords generated by the keyword generation model for the cluster of keywords to be processed are obtained; Calculate the semantic similarity between the initial virtual keyword and each initial keyword in the keyword cluster to be processed; If the semantic similarity is greater than or equal to the preset similarity threshold, the initial virtual keywords are determined as virtual keywords.

8. The method as described in claim 7, characterized in that, The method further includes: If there are initial keywords with semantic similarity less than the preset similarity threshold, the step of inputting each initial keyword in the keyword cluster to be processed into the keyword generation model is repeated.

9. The method as described in claim 1, characterized in that, Determine the keyword correspondence between each initial keyword and each virtual keyword, including: Determine the initial keyword to be matched, wherein the initial keyword to be matched is any one of the initial keywords; Among the virtual keywords, at least two candidate virtual keywords are identified that correspond to the initial keyword to be matched. The initial keyword to be matched and each candidate virtual keyword are input into the keyword matching model to obtain the target virtual keyword determined by the keyword matching model from each candidate virtual keyword; Establish a keyword correspondence between the initial keyword to be matched and the target virtual keyword.

10. The method as described in claim 9, characterized in that, Among the virtual keywords, at least two candidate virtual keywords corresponding to the initial keyword to be matched are determined, including: Obtain the initial keyword vector of the initial keyword to be matched, and obtain the virtual keyword vector of each virtual keyword; The approximate vector similarity is calculated based on the initial keyword vector to be matched and each virtual keyword vector, and at least two candidate virtual keywords are determined based on the calculation results.

11. The method as described in claim 1, characterized in that, Generate a target keyword index based on the initial keyword index and the keyword correspondence, including: Obtain the initial keyword index, which includes the initial keyword and the keyword object corresponding to the initial keyword; Based on the correspondence between the initial keywords and virtual keywords in the keyword correspondence, update the initial keywords in the initial keyword index to generate the target keyword index.

12. The method as described in claim 1, characterized in that, The target keyword object corresponding to the query keyword is determined based on the target keyword index, including: Receive query keywords and determine the corresponding virtual query keywords based on the query keywords and the keyword correspondence; Based on the query virtual keywords, the target keyword object is determined by searching in the target keyword index.

13. A data processing apparatus, characterized in that, include: The acquisition module is configured to acquire the semantic identifier sequence corresponding to each initial keyword, and cluster each initial keyword according to each semantic identifier sequence to obtain at least one keyword cluster, wherein each keyword cluster includes at least one initial keyword; The generation module is configured to generate virtual keywords corresponding to each keyword cluster and determine the keyword correspondence between each initial keyword and each virtual keyword; The index building module is configured to generate a target keyword index based on the initial keyword index and the keyword correspondence, wherein the target keyword index includes virtual keywords and keyword objects corresponding to the virtual keywords; The query module is configured to determine the target keyword object corresponding to the query keyword based on the target keyword index.

14. A computing device, characterized in that, include: Memory and processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions, which, when executed by the processor, implement the steps of the method according to any one of claims 1 to 12.

15. A computer-readable storage medium storing a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 12.

16. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 12.

Images