A method, device and computing device for determining keyword weight

Abstract

The present invention discloses a method, device and computing equipment for determining keyword weights. The method comprises: determining the actual number of searches for each keyword in a keyword set according to a search log; calculating a first ratio between the actual number of searches for a keyword and a set number of searches for the keyword; and determining the weight of the keyword according to the first ratio, wherein the weight is negatively correlated with the first ratio.

Description

Technical Field

[0001] The present invention relates to the field of Internet search technology, and in particular to a method, device and computing equipment for determining keyword weights for query rewriting. Background Art

[0002] Query rewrite (QR) refers to rewriting the search term entered by the user into one or more keywords similar to the search term, and then the search engine searches based on the rewritten keywords. Currently, many machine learning methods (such as deep semantic models, etc.) can rewrite dozens or hundreds of similar keywords. Due to performance constraints, only a few words can be selected as the final rewriting results.

[0003] For example, in the e-commerce search advertising scenario, advertisers purchase keywords for their advertising products on the e-commerce platform, which are also called bid-words. The e-commerce platform creates an inverted index for the bid-words. When the keywords (search terms) searched by users completely match the bid-words, the advertising products can be recalled by searching the inverted index.

[0004] However, advertisers do not know which search terms can bring them traffic. In many cases, the bid terms they purchased cannot trigger recall completely. Therefore, they need to rewrite the search terms into similar bid terms through query rewriting, and then recall the advertising products according to the rewritten bid terms. Since there are a large number of bid terms similar to search terms, in order to make the overall monetization efficiency of advertising products higher, it is also necessary to find the best set from many bid terms. Summary of the invention

[0005] In view of the above problems, the present invention is proposed to provide a method, apparatus and computing device for determining keyword weights that overcome the above problems or at least partially solve the above problems.

[0006] According to one aspect of the present invention, a method for determining keyword weights is provided, comprising:

[0007] Determine the actual number of searches for each keyword in the keyword set based on the search log;

[0008] Calculating a first ratio of an actual number of searches for a keyword to a set number of searches for the keyword;

[0009] The weight of the keyword is determined according to the first ratio, wherein the weight is negatively correlated with the first ratio.

[0010] Optionally, in the keyword weight determination method according to the present invention, determining the keyword weight based on the first ratio includes: when the first ratio is less than 1, updating the keyword weight to the quotient of the keyword's weight in the previous cycle and the first ratio; when the first ratio is equal to 1, updating the keyword weight to the product of the keyword's weight in the previous cycle and the proportion of the keyword's online time, wherein the proportion of online time is the ratio of the keyword's online time to the total time, and the online time is the time corresponding to the actual number of searches for the keyword reaching the set number of searches.

[0011] Optionally, in the keyword weight determination method according to the present invention, when the updated weight is greater than 1, the weight is set to 1.

[0012] According to another aspect of the present invention, a method for generating ranking weights of bidding words is also provided, comprising:

[0013] Determine the expected revenue and budget usage data of each bidding word in the bidding word set based on the search log;

[0014] Determine the ranking weight of the bidding word based on the expected revenue and budget usage data of the bidding word, where the ranking weight of the bidding word is positively correlated with the expected revenue and negatively correlated with the budget usage rate;

[0015] The bidding words and ranking weights are associated and stored as a bidding word weight table.

[0016] Optionally, in the method for generating ranking weights of bidding words according to the present invention, the expected revenue of the bidding word is the ratio of the total amount spent by all advertisers on the bidding word to the amount of traffic allocated to the advertising product associated with the bidding word.

[0017] Optionally, in the method for generating ranking weights of bidding words according to the present invention, determining the ranking weights of bidding words according to the expected revenue and budget usage data of the bidding words includes: maintaining a first coefficient for each bidding word; determining the budget usage rate and online time ratio of the bidding word according to the budget usage data of the bidding word, the budget usage rate being the ratio of the total amount of consumption of all advertisers on the bidding word to the total budget amount, the online time ratio being the ratio of the total online time of all advertisers on the bidding word to the total time, and the online time being the time corresponding to the consumption of the budget amount of the advertiser on the bidding word; when the online time ratio of the bidding word is less than 1, updating the first coefficient of the bidding word to the product of the first coefficient of the bidding word in the previous cycle and the online time ratio; when the budget usage rate of the bidding word is less than 1, updating the first coefficient of the bidding word to the quotient of the first coefficient of the bidding word in the previous cycle and the budget usage rate; determining the ranking weight of the bidding word as the product of the expected revenue of the bidding word and the first coefficient.

[0018] Optionally, in the method for generating ranking weights of bidding words according to the present invention, when the updated first coefficient is greater than 1, the first coefficient is set to 1.

[0019] According to another aspect of the present invention, there is also provided a query method, comprising:

[0020] Obtain search terms from the user's query request;

[0021] Obtain a subset of bidding terms related to the search term from the bidding term set;

[0022] Obtaining the ranking weight of each bidding word in the bidding word subset from the bidding word weight table, and selecting a predetermined number of bidding words from the bidding word subset in descending order of ranking weight as candidate bidding words corresponding to the search word;

[0023] Matching advertising products based on candidate bidding terms and search terms, and displaying the matched advertising products;

[0024] The bidding word weight table is generated according to the above-mentioned bidding word ranking weight generation method.

[0025] Optionally, in the query method according to the present invention, a bidding term subset related to the search term is obtained from the bidding term set according to one of the following methods: a deep semantic method, a synonym rewriting method, or a search term co-occurrence method.

[0026] According to another aspect of the present invention, there is also provided a device for generating ranking weights of bidding words, comprising:

[0027] A statistical unit, adapted to determine expected revenue and budget usage data of each bidding word in the bidding word set according to the search log;

[0028] A weight calculation unit, adapted to determine the ranking weight of the bidding word according to the expected revenue and budget usage data of the bidding word, wherein the ranking weight of the bidding word is positively correlated with the expected revenue and negatively correlated with the budget usage rate;

[0029] The bidding word weight table is suitable for storing the association between bidding words and ranking weights.

[0030] According to another aspect of the present invention, there is also provided a query device, comprising:

[0031] A search term acquisition unit, adapted to acquire a search term from a user's query request;

[0032] A bidding word acquisition unit, adapted to acquire a bidding word subset related to the search word from the bidding word set;

[0033] The query rewriting unit is adapted to obtain the ranking weight of each bidding word in the bidding word subset from the bidding word weight table, and select a predetermined number of bidding words from the bidding word subset in descending order of the ranking weight as candidate bidding words corresponding to the search word;

[0034] A search unit, adapted to match advertising products according to candidate bidding terms and search terms, and display the matched advertising products;

[0035] Wherein, the bidding word weight table is generated by the bidding word ranking weight generating device of the present invention.

[0036] According to yet another aspect of the present invention, there is further provided a computing device, comprising:

[0037] one or more processors;

[0038] Memory; and

[0039] One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs comprising instructions for executing any of the methods described above.

[0040] According to another aspect of the present invention, there is also provided a computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions, which, when executed by a computing device, enable the computing device to perform any of the above methods.

[0041] According to the keyword weight determination scheme of the embodiment of the present invention, the first ratio of the actual search times of the keyword and the set search times of the keyword is calculated, and the keyword weight is determined according to the first ratio. In this way, when performing query rewriting, a predetermined number of keywords can be screened out from multiple keywords similar to the query word according to the weight, and the search engine searches according to the screened keywords, which can improve the coverage of the search results.

[0042] In addition, when the keyword is a bidding word in an e-commerce search advertising scenario, the ranking weight generation scheme for the bidding words according to an embodiment of the present invention takes into account the expected return of the bidding words and the budget usage behind the bidding words, and uses the idea of ​​linear programming for derivation and solution, and sorts according to the derivation and solution results of linear programming. The budget usage information can be integrated into the sorting strategy, and the candidate bidding words tend to be bidding words with higher expected returns and lower budget utilization rates, making the sorting results more reasonable, thereby improving the overall monetization efficiency of the advertising products.

[0043] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0045] Figure 1 A flowchart of a method 100 for generating ranking weights of bidding words according to an embodiment of the present invention is shown;

[0046] Figure 2 A flowchart of a query rewriting method 200 according to an embodiment of the present invention is shown;

[0047] Figure 3 A flowchart of a query method 300 according to an embodiment of the present invention is shown;

[0048] Figure 4 A block diagram of a computing device 400 according to an embodiment of the present invention is shown;

[0049] Figure 5 A structural diagram of a ranking weight generating device 500 according to an embodiment of the present invention is shown;

[0050] Figure 6 A structural diagram of a query rewriting device 600 according to an embodiment of the present invention is shown;

[0051] Figure 7 A flowchart of a keyword weight setting method 700 according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0052] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0053] In order to improve the coverage of search results of a search engine, an embodiment of the present invention provides a keyword weight determination method, by which a weight is set for each keyword in a keyword set. When performing query rewriting, a predetermined number of keywords can be screened out from multiple keywords similar to the query term according to the weight, and then the search engine performs a search based on the screened keywords.

[0054] Figure 7 FIG. 7 is a flowchart of a method 700 for setting keyword weights according to an embodiment of the present invention. Figure 7 , method 700 starts at step S710. In step S710, the actual number of searches for each keyword in the keyword set is determined based on the search log. The keyword set is a set of keywords maintained by a search engine, and the search engine will create an index of keywords and data content that matches the keywords. The actual number of searches refers to the number of times the search engine searches using the keyword, and can be counted according to a predetermined period, for example, once a day. Accordingly, the calculation of the first ratio and the weight of the keyword in the subsequent steps are also updated according to a predetermined period, for example, once a day.

[0055] In step S720, a first ratio of the actual search times of the keyword to the set search times of the keyword is calculated. Here, the set search times is a search times upper limit set in advance for each keyword in the keyword set, that is, for a certain keyword, when the number of searches using the keyword reaches the search times upper limit, the search engine will no longer use the keyword for searching. The set search times of each keyword can be the same or different, and those skilled in the art can reasonably set them according to specific needs.

[0056] In step S730, the weight of the keyword is determined according to the first ratio, wherein the weight is negatively correlated with the first ratio, that is, the larger the first ratio is, the smaller the weight is, and the smaller the first ratio is, the larger the weight is.

[0057] Specifically, the weight of keywords can be determined as follows:

[0058] When the first ratio is less than 1, the weight of the keyword is updated to the quotient of the weight of the keyword in the previous cycle and the first ratio;

[0059] When the first ratio is equal to 1, the weight of the keyword is updated to the product of the weight of the keyword in the previous period and the proportion of the online time of the keyword, wherein the online time proportion is the ratio of the online time of the keyword to the total time, the online time is the time corresponding to the actual number of searches for the keyword reaching the set number of searches, and the total time is the time corresponding to the predetermined period.

[0060] In the above process, when the updated weight is greater than 1, the weight is set to 1.

[0061] According to the above method, the expected number of searches for a keyword in the current cycle can be made as close as possible to the set number of searches, and the online time of the keyword can be made as long as possible, thereby generally improving the coverage of the search results of the search engine.

[0062] When the keyword is a bidding word in an e-commerce search advertising scenario, there are many schemes for rewriting the query of the bidding word. General machine learning methods can rewrite the original search word (Query) into multiple bidding words. In order to make the overall monetization efficiency of the advertising product higher, it is necessary to find the optimal set from the numerous bidding words. In other words, it is necessary to sort the multiple bidding words obtained by rewriting, and select a predetermined number of bidding words as candidate bidding words according to the sorting results.

[0063] One way to implement bidding word ranking is to count the number of products purchased for each bidding word (the purchase depth of the product) and sort them in descending order according to the purchase depth. This method only considers the number of products behind the bidding word, but does not consider the revenue, budget and other factors behind the bidding word, so the product realization efficiency is not high enough.

[0064] Another way to implement bidding word ranking is to consider the revenue of the advertisements behind the bidding words, train them through machine learning (Learn To Rank), and then rank them. However, machine learning requires a large amount of labeled training data and consumes a lot of computing power.

[0065] To this end, an embodiment of the present invention also provides a bidding word ranking weight generation scheme for query rewriting, the principle of which is: from the perspective of constrained budget utilization rate, the bidding word ranking problem is viewed, the ranking problem is defined and abstracted, and a linear programming approach is used to derive and solve it. The resulting ranking formula reflects the bidding word dimension and budget utilization rate, and the ranking result is more reasonable.

[0066] The following introduces the specific process of deriving the sorting formula using linear programming ideas.

[0067] 1) The ranking problem of bidding words can be defined as: maximizing the advertising revenue of a traffic platform (such as an e-commerce advertising platform) while satisfying the budget utilization rate of the bidding words. The definition is as follows:

[0068]

[0069] Among them, RPM ijrepresents the expected revenue of bidding word j under traffic i. Here, one query request can be regarded as one traffic. Generally speaking, for bidding word j, the expected revenue corresponding to the bidding word under each traffic can be set to the same, that is, RPM ij =RPM j How to obtain the expected revenue RPM of bidding word j j , see the description below. ij Indicates whether traffic i is allocated to bidding word j, 0 means no allocation, 1 means allocation. The so-called allocation of traffic to bidding words can be understood as when receiving a query request, the advertising platform displays (exposes) the advertising products associated with the bidding word. M is the total traffic number, and N is the total number of bidding words in the bidding word set. Here, the total traffic number can be the total number of query requests in a predetermined statistical period (for example, within 1 day). Budget j Represents the budget of bid word j, that is, the sum of the budget amounts of all advertisers who have purchased the bid word on the bid word. In the field of e-commerce advertising, advertisers will purchase bid words at a certain price, and will also set a budget amount for the bid word, for example, setting the daily expenditure (consumption) on the bid word not to exceed the corresponding budget amount. When an advertiser has used up the budget amount for a bid word, the advertising platform will no longer expose the advertising products of the advertiser associated with the bid word.

[0070] 2) The above model can be solved by linear programming. By solving the dual problem of the original problem, the following formula can be obtained:

[0071]

[0072] Among them, α j , β i are variables related to bidding word j and traffic i, respectively. Then, by solving the dual problem, we can conclude that bidding word j should be calculated according to RPM for each traffic i. ij (1-α j ) are sorted from large to small. From the above, we can see that: RPM ij (1-α j ) = RPM j (1-α j ), that is, the ranking weight of bidding word j is RPM j (1-α j ).

[0073] 3) For RPM j , which can be obtained by counting the search logs (see the description below). Here is how to solve α j . j Substituting into the above dual formula, minimizing the objective G(α) can be obtained:

[0074]

[0075] 4) The objective function in 3) can be solved by the gradient descent method, and α can be updated:

[0076]

[0077]

[0078] According to the above iterative formula, the calculation is repeated until the objective function converges, or when the number of iterations reaches the predetermined number, the iteration is stopped and α is output. j Among them, η is the learning rate, which can be set as needed.

[0079] 5) Considering For each bidding word, the difference is large, so it is easy for α to be updated too quickly and difficult to control; the embodiment of the present invention also proposes a replacement method, which can prevent the α update speed from being too fast or too slow.

[0080] 6) Introduce two variable factors for bidding word granularity, DurationRatio j and BudgetRatio j , respectively represent the online time proportion and budget utilization rate of bidding word j.

[0081] Here, the budget utilization rate of bidding word j is the ratio of the total amount of consumption of all advertisers on the bidding word to the total budget amount, the online duration ratio is the ratio of the total online duration of all advertisers on the bidding word to the total duration, and the online duration is the duration corresponding to the advertiser's budget amount on the bidding word being consumed. For example, taking one day (24 hours) as the statistical cycle and starting from 0 o'clock, if an advertiser's budget amount on the bidding word is consumed at 8 o'clock, then the advertiser's online duration on the bidding word is 8 hours, and the total duration of all advertisers is 24*X, where X is the number of advertisers who purchased the bidding word.

[0082] Specifically, the definitions of these two variable factors are as follows:

[0083]

[0084]

[0085] Where c represents the set of advertisers who purchase bidding word j, ∑ c buys j Duration c represents the total online time of advertisers who purchase bidding word j, ∑ c buys j 24h represents the total duration of the advertiser purchasing bidding word j, ∑c buys j Cost c represents the total amount of money spent by advertisers who purchased bid word j on the bid word, ∑ c buys j Budget c Represents the total budget amount of advertisers who purchase bidding word j on the bidding word.

[0086] 7) Maintain a coefficient γ for each bidding word j j , update according to the following formula;

[0087]

[0088] Where T represents the statistical period, for example, 1 day, that is, γ j It can be iterated every day, so that represents the coefficient of the previous day, Indicates the coefficient of the current day. When the budget utilization rate of bidding word j is less than 1, it means that the bidding word has a budget and the ranking weight of the bidding word should be increased. j The update value will become larger; when the online time of bidding word j accounts for less than 1, it means that the budget has been spent, γ j Update values ​​will become smaller.

[0089] It should be noted that if the budget utilization rate is equal to 1, it means that the budget has been spent, but the granularity may be different. For example, some bidding words are triggered at the beginning of the day (that is, the corresponding budget has been spent), and some words are triggered at 18:00. Therefore, the embodiment of the present invention introduces the concept of online time proportion to indicate the specific triggering degree of each word to adjust the coefficient value.

[0090] where γ j The initialization value of can be 1, that is, in the first iteration, since there is no data from the previous day, the coefficient of the previous day can be set to 1.

[0091] 8) Get the coefficient γ j After that, the ranking weight of bidding word j can be calculated according to RPM. j ×γ j conduct.

[0092] According to the derivation results of the above bidding word ranking principle, when rewriting a query, the candidate bidding words tend to be those with higher expected returns and lower budget utilization. The following introduces a bidding word ranking weight generation method for query rewriting based on this principle.

[0093] Figure 1 FIG. 1 is a flow chart showing a method 100 for generating ranking weights of bidding words according to an embodiment of the present invention. Figure 1, method 100 begins at step S110, in which the expected revenue and budget usage data of each bidding word in the bidding word set are determined based on the search log. A traffic platform (such as an e-commerce advertising platform) usually establishes and maintains a bidding word set, which includes multiple bidding words. After receiving a user's query request, the search word is extracted from the query request, and the search word is rewritten into one or more bidding words in the bidding word set, which is query rewriting.

[0094] Since advertisers often add or delete words, in order to accurately obtain the situation of advertisers purchasing bidding words, in one implementation of the present invention, the traffic platform can obtain a bidding word set based on the log of advertisers purchasing bidding words. In this way, the bidding words in the obtained bidding word set are the most recently effective bidding words, and query rewriting based on the bidding word set can improve the accuracy of query rewriting.

[0095] The following is an example of a log:

[0096] Company ID Advertisement product id Bidding words Bidding word price 10001 5000001 Mp3player 3.3 10002 6000003 Wedding dress 4.2 … … … …

[0097] According to the above log statistics, the effective bidding word set can be obtained as follows:

[0098] Bidding words Mp3 player Wedding dress …

[0099] Then, the expected revenue of each bidding word can be obtained by statistically analyzing the user's search log data. Here, the expected revenue of a bidding word refers to the revenue brought to the traffic platform by the display (exposure) of the advertising product associated with the bidding word, which corresponds to the advertiser's consumption on the bidding word, that is, the fee paid by the advertiser to the traffic platform. In one implementation, the expected revenue is characterized by revenue per thousand impressions (Revenue Per Mile, RPM). Of course, the expected revenue can also be characterized in other ways, and the embodiments of the present invention are not limited to this. In general, the expected revenue of a bidding word can be the ratio of the sum of the consumption amounts of all advertisers on the bidding word to the number of traffic allocated to the advertising product associated with the bidding word.

[0100] Taking RPM as an example, the expected return of bidding word j can be calculated as PV j represents the total traffic of the advertising product associated with bidding word j, Cost j Represents the total amount spent by all advertisers on bidding word j.

[0101] The following is an example of a search log:

[0102]

[0103] It should be noted that when the expected return is calculated based on the search log, it is usually calculated based on a predetermined period, that is, the log data used for the calculation is data within the predetermined period. For example, if the predetermined period is 1 day, the expected return is calculated on a daily basis.

[0104] Then, the budget usage data within a predetermined period (e.g., 1 day) can also be determined based on the search log. The budget usage data of the bidding word may include: the budget usage rate and online time ratio of the bidding word, the budget usage rate is the ratio of the total amount of consumption of all advertisers on the bidding word to the total budget amount, the online time ratio is the ratio of the total online time of all advertisers on the bidding word to the total time, and the online time is the time corresponding to the budget amount of the advertiser on the bidding word is consumed.

[0105] The online duration ratio and budget utilization rate of bidding word j can be expressed as DurationRatio j and BudgetRatio j , please see above for the specific calculation formula.

[0106] After determining the expected revenue and budget usage data of the bidding word, method 100 enters step S120. In step S120, the ranking weight of the bidding word is determined according to the expected revenue and budget usage data of the bidding word. According to the ranking principle derived above, the ranking weight of the bidding word is positively correlated with the expected revenue, that is, the greater the expected revenue, the greater the ranking weight; the ranking weight of the bidding word is negatively correlated with the budget utilization rate, that is, the greater the budget utilization rate, the smaller the ranking weight.

[0107] Based on the principle that the ranking weight of the bidding words is positively correlated with the expected revenue and negatively correlated with the budget utilization rate, those skilled in the art can design a variety of ranking weight calculation formulas. The embodiment of the present invention schematically provides two implementation methods.

[0108] Method 1, based on solving the linear programming model, is as follows:

[0109] First, a linear programming model is constructed to minimize the following objective function:

[0110]

[0111] Among them, N is the total number of bidding words, M is the total traffic, and Budget is i is the sum of the budgets of all advertisers on bidding word j, RPM j is the expected return of bidding word j.

[0112] Then, the predetermined algorithm is used to solve the linear programming model and the coefficient α is obtained. j .

[0113] There are many methods to solve linear programming models, one of which is the gradient descent method. In the gradient descent method, the iterative formula used is:

[0114] α j ←α j -η(Budget j -M×RPM j )

[0115] Among them, η is the learning rate, and its specific value can be set according to experiments or experience.

[0116] Finally, the ranking weight of bidding word j is determined as RPM j (1-α j ).

[0117] It should be noted that in method 1, the expected revenue RPM of the bidding word can be calculated based on the search log of the previous day. j And the sum of the budget amount of the previous day to calculate the sorting weight RPM j (1-α j ), the calculated ranking weight can be applied to the bidding word ranking of the current day.

[0118] Method 2: Based on Budget Ratio j and DurationRatio j The calculation is as follows:

[0119] Maintain a coefficient γ for each bidding word j j , update according to the following formula:

[0120]

[0121] Where T represents the statistical period, for example, 1 day, that is, γ j It can be iterated every day, so that represents the coefficient of the previous day, Indicates the coefficient of the current day. When the budget utilization rate of bidding word j is less than 1, it means that the bidding word has a budget and the ranking weight of the bidding word should be increased. j The update value will become larger; when the online time of bidding word j accounts for less than 1, it means that the budget has been spent, γ j Update values ​​will become smaller.

[0122] where γ j The initialization value of can be 1, that is, in the first iteration, since there is no data from the previous day, the coefficient of the previous day can be set to 1.

[0123] Get the coefficient γj After that, the ranking weight of bidding word j can be calculated according to RPM. j ×γ j conduct.

[0124] After determining the ranking weight of the bidding word, the method 100 enters step S130. In step S130, the bidding words and the ranking weights are associated and stored as a bidding word weight table, so that when rewriting the query, multiple bidding words can be ranked according to the bidding word weight table.

[0125] Figure 2 FIG. 2 is a flowchart of a query rewriting method 200 according to an embodiment of the present invention. Figure 2 The method 200 starts at step S210, in which a search term is obtained from a user's query request. When a user needs to query a commodity or product, the user enters the information to be queried in the client or browser, and the client or browser generates a query request and sends it to the server. After receiving the query request, the server extracts the query keyword, i.e., the search term, from the query request.

[0126] Then, in step S220, a subset of bidding words related to the search word is obtained from the bidding word set. In one implementation, the search word can be rewritten as a set of semantically similar words in the vector space through a deep semantic method. The deep semantic method mainly uses deep learning technology to represent the text as a word vector (embedding), calculates the semantic similarity between texts, and then obtains a class of methods that similar texts. Specifically, the similarity between the word vector of the search word and the word vector of each bidding word in the bidding word set can be calculated, and the bidding words with a similarity greater than a threshold can be screened out to form a subset of bidding words related to the search word.

[0127] In other implementations, the method of generating the bidding word subset may also be generated by synonym rewriting and replacement, search word (query) co-occurrence, etc. Query co-occurrence means that if query and query frequently co-occur in the same session, then the two queries can also be considered to have a close correlation, which is also the method for search engines to mine and generate related query words.

[0128] Then, in step S230, the ranking weight of each bidding word in the bidding word subset is obtained from the bidding word weight table, and a predetermined number of bidding words are selected from the bidding word subset in descending order of ranking weight as candidate bidding words corresponding to the search word, so as to match (recall) the advertising product based on the candidate bidding words and the search word. Here, the bidding word weight table can adopt the bidding word weight table generated in method 100.

[0129] Figure 3FIG. 3 is a flowchart of a query method 300 according to an embodiment of the present invention. Figure 3 , method 300 starts at step S310. In step S310, a user's query request is obtained. When a user needs to query a commodity or product, the user enters the information to be queried in the client or browser, and the client or browser generates a query request and sends it to the server. After receiving the query request, the server extracts the query keyword, i.e., the search word, from the query request.

[0130] Then, in step S320, the search term in the query request is rewritten to obtain a plurality of candidate bidding terms. Here, the query rewriting method may adopt method 200.

[0131] Then, in step S330, the advertising products are matched according to the multiple candidate bidding words and the search words, and the matched advertising products are displayed. An index file (such as an inverted index) of the bidding words is established in the server, and after the query is rewritten, the advertising products can be matched (recalled) by searching the inverted index.

[0132] The above methods 100, 200, 300 may be executed in a computing device, for example, Figure 4 The computing device 400 is shown in FIG. Figure 4 As shown, the computing device 400 includes an input device 401, an input interface 402, a central processing unit 403, a memory 404, an output interface 405, and an output device 406. The input interface 402, the central processing unit 403, the memory 404, and the output interface 405 are connected to each other via a bus 410, and the input device 401 and the output device 406 are connected to the bus 410 via the input interface 402 and the output interface 405, respectively, and then connected to other components of the computing device 400.

[0133] Specifically, the input device 401 receives input information from the outside and transmits the input information to the central processing unit 403 through the input interface 402; the central processing unit 403 processes the input information based on the computer executable instructions stored in the memory 404 to generate output information, temporarily or permanently stores the output information in the memory 404, and then transmits the output information to the output device 406 through the output interface 405; the output device 406 outputs the output information to the outside of the computing device 400 for use by the client.

[0134] That is to say, Figure 4 The computing device shown may also be implemented to include: a memory storing computer-executable instructions; and a processor, which can implement methods 100, 200, 300 when executing the computer-executable instructions.

[0135] Figure 5 FIG. 5 shows a structural diagram of a ranking weight generating device 500 according to an embodiment of the present invention. Figure 5 As shown, the device 500 includes:

[0136] A statistical unit 520, adapted to determine expected revenue and budget usage data of each bidding word in the bidding word set according to the search log 510;

[0137] The weight calculation unit 530 is adapted to determine the ranking weight of the bidding word according to the expected revenue and budget usage data of the bidding word, wherein the ranking weight of the bidding word is positively correlated with the expected revenue and negatively correlated with the budget usage rate;

[0138] The bidding word weight table 540 is suitable for storing the association between bidding words and ranking weights, so that when the query is rewritten, the bidding words are ranked according to the bidding word weight table.

[0139] It should be noted that the specific operations performed by each module of the device 500 can refer to the method 100 and will not be described in detail here.

[0140] Figure 6 FIG. 6 shows a structural diagram of a query rewriting device 600 according to an embodiment of the present invention. Figure 6 As shown, the device 600 includes:

[0141] A search term acquisition unit 610, adapted to acquire a search term from a user's query request;

[0142] The bidding word acquisition unit 620 is adapted to acquire a bidding word subset related to the search word from the bidding word set 640;

[0143] The query rewriting unit 630 is adapted to obtain the ranking weight of each bidding word in the bidding word subset from the bidding word weight table 650, and select a predetermined number of bidding words from the bidding word subset in descending order of ranking weight as candidate bidding words corresponding to the search word, so as to match the advertising product according to the candidate bidding words and the search word. The bidding word weight table 650 can be generated by the device 500.

[0144] It should be noted that the specific operations performed by each module of the device 600 can refer to the method 200 and will not be described in detail here.

[0145] In addition, an embodiment of the present invention also provides a query device, which includes a query rewriting device 600 and a search unit 660. The search unit 660 is suitable for matching advertising products according to the candidate bidding words and search words output by the query rewriting device 600 (for example, matching according to the index file 670), and displaying the matched advertising products.

[0146] It should be noted that the specific operations performed by each module of the query device can refer to method 300 and will not be described in detail here.

[0147] The algorithm and display provided herein are not inherently related to any particular computer, virtual system or other device. Various general purpose systems can also be used together with the teachings based on this. According to the above description, it is obvious that the structure required for constructing such systems. In addition, the present invention is not directed to any specific programming language either. It should be understood that various programming languages ​​can be utilized to realize the content of the present invention described herein, and the description of the above specific languages ​​is for disclosing the best mode of the present invention.

[0148] In the description provided herein, a large number of specific details are described. However, it is understood that embodiments of the present invention can be practiced without these specific details. In some instances, well-known methods, structures and techniques are not shown in detail so as not to obscure the understanding of this description.

Claims

1. A method for determining keyword weights, comprising: Determine the actual number of searches for each keyword in the keyword set based on the search log; Calculating a first ratio of an actual number of searches for a keyword to a set number of searches for the keyword; determining a weight of the keyword according to the first ratio, wherein the weight is negatively correlated with the first ratio; When the first ratio is less than 1, the weight of the keyword is updated to the quotient of the weight of the keyword in the previous cycle and the first ratio; When the first ratio is equal to 1, the weight of the keyword is updated to the product of the weight of the keyword in the previous cycle and the online time ratio of the keyword, wherein the online time ratio is the ratio of the online time of the keyword to the total time, and the online time is the time corresponding to the actual number of searches for the keyword reaching the set number of searches.

2. The method of claim 1, wherein: When the updated weight is greater than 1, the weight is set to 1.

3. A query method, comprising: Obtain search terms from the user's query request; Obtain a subset of bidding terms related to the search term from the bidding term set; According to the determined ranking weight of the bidding word stored in the bidding word weight table, the ranking weight of each bidding word in the bidding word subset is obtained, and a predetermined number of bidding words are selected from the bidding word subset in descending order of the ranking weight as candidate bidding words corresponding to the search word; Matching advertising products based on candidate bidding terms and search terms, and displaying the matched advertising products; The step of generating the bidding word weight table includes: Determine the expected revenue and budget usage data of each bidding word in the bidding word set based on the search log; The ranking weight of the bidding word is determined according to the expected revenue and budget usage data of the bidding word, wherein the ranking weight of the bidding word is positively correlated with the expected revenue and negatively correlated with the budget usage rate. The method of determining the ranking weight of the bidding word according to the expected revenue and budget usage data of the bidding word includes: maintaining a first coefficient for each bidding word; determining the budget usage rate and online time proportion of the bidding word according to the budget usage data of the bidding word, wherein the budget usage rate is the ratio of the total amount of consumption of all advertisers on the bidding word to the total budget amount, and the online time proportion is the total amount of consumption of all advertisers on the bidding word. The ratio of the sum of the online time of the advertiser on the bidding word to the total time, wherein the online time is the time corresponding to the consumption of the budget amount of the advertiser on the bidding word; when the online time ratio of the bidding word is less than 1, the first coefficient of the bidding word is updated to the product of the first coefficient of the bidding word in the previous cycle and the online time ratio; when the budget utilization rate of the bidding word is less than 1, the first coefficient of the bidding word is updated to the quotient of the first coefficient of the bidding word in the previous cycle and the budget utilization rate; the ranking weight of the bidding word is determined as the product of the expected return of the bidding word and the first coefficient; The bidding words and ranking weights are associated and stored as the bidding word weight table.

4. The method of claim 3, wherein: The expected revenue of the bidding word is the ratio of the total amount of money spent by all advertisers on the bidding word to the amount of traffic allocated to the advertising product associated with the bidding word.

5. The method of claim 3, wherein: When the updated first coefficient is greater than 1, the first coefficient is set to 1.

6. The method of claim 3, wherein: Use one of the following methods to obtain a subset of bid terms related to the search term from the bid term set: Deep semantic method, synonym rephrasing method or search term co-occurrence method.

7. A query device, comprising: A search term acquisition unit, adapted to acquire a search term from a user's query request; A bidding word acquisition unit, adapted to acquire a bidding word subset related to the search word from the bidding word set; a query rewriting unit adapted to obtain the ranking weight of each bidding word in the bidding word subset according to the ranking weight of the determined bidding word stored in the bidding word weight table, and select a predetermined number of bidding words from the bidding word subset in descending order of the ranking weight as candidate bidding words corresponding to the search word; A search unit, adapted to match advertising products according to candidate bidding terms and search terms, and display the matched advertising products; The step of generating the bidding word weight table includes: Determine the expected revenue and budget usage data of each bidding word in the bidding word set based on the search log; The ranking weight of the bidding word is determined according to the expected revenue and budget usage data of the bidding word, wherein the ranking weight of the bidding word is positively correlated with the expected revenue and negatively correlated with the budget usage rate. The method of determining the ranking weight of the bidding word according to the expected revenue and budget usage data of the bidding word includes: maintaining a first coefficient for each bidding word; determining the budget usage rate and online time proportion of the bidding word according to the budget usage data of the bidding word, wherein the budget usage rate is the ratio of the total amount of consumption of all advertisers on the bidding word to the total budget amount, and the online time proportion is the total amount of consumption of all advertisers on the bidding word. The ratio of the sum of the online time of the advertiser on the bidding word to the total time, wherein the online time is the time corresponding to the consumption of the budget amount of the advertiser on the bidding word; when the online time ratio of the bidding word is less than 1, the first coefficient of the bidding word is updated to the product of the first coefficient of the bidding word in the previous cycle and the online time ratio; when the budget utilization rate of the bidding word is less than 1, the first coefficient of the bidding word is updated to the quotient of the first coefficient of the bidding word in the previous cycle and the budget utilization rate; the ranking weight of the bidding word is determined as the product of the expected return of the bidding word and the first coefficient; The bidding words and ranking weights are associated and stored as the bidding word weight table.

8. A computing device comprising: one or more processors; Memory; as well as One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs comprising instructions for executing any of the methods according to claims 1-6.

9. A computer-readable storage medium storing one or more programs, the one or more programs comprising instructions, which when executed by a computing device, cause the computing device to perform any of the methods according to claims 1-6.

Images