Triggering way and device of promoting key words

A keyword and keyword library technology, applied in the field of search, can solve the problems of low accuracy and recall rate of triggering promotion keywords, and achieve the effect of improving the accuracy rate and the recall rate.

Active Publication Date: 2017-05-10
BAIDU ONLINE NETWORK TECH (BEIJIBG) CO LTD
10 Cites 10 Cited by

AI-Extracted Technical Summary

Problems solved by technology

[0005] However, most of the current promotion keyword triggering technologies are based o...
View more

Abstract

The invention provides a triggering way and a device of promoting key words, wherein the triggering way of promoting key words includes the following steps: obtaining a query inputted by users; inputting an input sequence corresponding to the query onto a translation model to obtain an output sequence; using the output sequence to ensure the promoting key words triggered by the query; the translation model is obtained by training in advance through the following steps: obtaining the query and a clicked title corresponding to the query from the users' clicking behavior logs as training data; using the query in the training data to obtain the input sequence, using the clicked title corresponding to the query to obtain a target sequence, and training a neural network model to obtain the translation model. The triggering way and the device of promoting key words can really obtain the promoting key words corresponding to the query in semantic, and improve accuracy and recall rate triggered by the key words.

Application Domain

Semantic analysisNeural architectures +1

Technology Topic

Network modelNerve network +4

Image

  • Triggering way and device of promoting key words
  • Triggering way and device of promoting key words
  • Triggering way and device of promoting key words

Examples

  • Experimental program(1)

Example Embodiment

[0031] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
[0032] The terms used in the embodiments of the present invention are only for the purpose of describing specific embodiments, and are not intended to limit the present invention. The singular forms of "a", "said" and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings.
[0033] It should be understood that the term "and/or" used in this article is only an association relationship describing associated objects, which means that there can be three relationships. For example, A and/or B can mean that there is A alone, and both A and B, there are three cases of B alone. In addition, the character "/" in this text generally indicates that the associated objects before and after are in an "or" relationship.
[0034] Depending on the context, the word "if" as used herein can be interpreted as "when" or "when" or "in response to determination" or "in response to detection". Similarly, depending on the context, the phrase "if determined" or "if detected (statement or event)" can be interpreted as "when determined" or "in response to determination" or "when detected (statement or event) )" or "in response to detection (statement or event)".
[0035] figure 1 Is a flowchart of a method provided by an embodiment of the present invention, such as figure 1 As shown, the method may mainly include the following steps:
[0036] In 101, obtain the query entered by the user.
[0037] In this step, the query input by the user can be in any form, and can be a single word or a sentence.
[0038] In 102, the input sequence corresponding to the obtained query is input to the translation model to obtain the output sequence.
[0039] In this step, word segmentation is performed on the obtained query, and the word sequence obtained after word segmentation is input into the translation model as an input sequence. The translation model is pre-trained, and the neural network model is trained through the acquired training data to obtain the translation model.
[0040] When training the neural network model, the training data used is: a two-tuple formed by the query obtained from the user's click behavior log and the corresponding clicked title. The clicked title is preferably the title of the clicked promotion result.
[0041] Among them, the user click behavior log comes from the log system of the search engine. The log system of the search engine records the user's click behavior, which reflects the user's search intent and the correlation between search results and promotion. In the log system of the search engine, there will be mature algorithms that identify the effectiveness of users and the effectiveness of user click behavior. Therefore, the training data used in training the translation model in this step is the effective click data generated by effective users. The training data can also be called high-quality click behavior data of the user. The translation model obtained by training with the training data will be more accurate and effective when translating the input query.
[0042] When using training data to train the neural network model to obtain the translation model, the query in the training data is segmented, the word sequence obtained after the word segmentation is used as the input sequence, and the clicked title corresponding to the query in the training data is used as the target sequence for training.
[0043] In this step, the neural network model used is preferably Recurrent Neural Network (RNN, RecurrentNeural Network), because the hidden nodes in the network structure of the recurrent neural network model are connected to form a ring, and its internal state is input The sequence is determined, so the network structure of the recurrent neural network model is very suitable for processing longer text sequences. In the following description, the RNN is used to represent the recurrent neural network model.
[0044] The RNN network structure is as figure 2 As shown, the structure is divided into 5 parts from bottom to top, completing the training process from the input sequence to the target sequence, and the training process is actually the translation process from the query to the clicked title corresponding to the query.
[0045] Among them, the first layer of the RNN network structure is the word vector layer on the input side. This layer represents each word of the query as a word vector. For example, the query in the training data is "how to cultivate good habits for children", which is obtained after word segmentation The words "how", "cultivation", "child", "of", "good", and "habit" are expressed in the form of word vectors. The number in the box represents the word vector of the word. The cosine similarity is used to measure the semantic similarity between words; the second layer of the RNN network structure is a bidirectional RNN layer, which reads the input word vector sequence from both positive and negative directions, and reads the word vector sequence in an iterative manner The sequence information is encoded into the internal state, because the query in the training data contains rich semantic information. After the query is expressed as a word vector in the first layer, different word vectors contain different semantic information, and the bidirectional RNN layer converts the word vector sequence Encode the semantic information in the RNN and encode the semantic information of the input sequence in the internal state of the RNN; the third layer of the RNN network structure is the alignment layer, which uses different weight values ​​to weight the internal state of the bidirectional RNN; RNN The fourth layer of the network structure is the RNN layer. This layer calculates the probability of each word in this iteration based on the output of the alignment layer, the internal state of the previous iteration and the translated word sequence in an iterative manner; RNN network The fifth layer of the structure is the word vector layer at the output end, which represents the word vector of each word in the target sequence translated from the input sequence.
[0046] For example, the query in the training data is "how to cultivate children's good habits", and the title of the query corresponding to the clicked is "children's good habits training". Use the RNN network structure to train the training data. The goal of training is to pass the RNN network structure as much as possible. When the query is "how to cultivate children's good habits", the output target sequence is "children's good habits training". That is to say, the purpose of training with the RNN network structure is to make the query in the training data after translation as much as possible to obtain the title of the query corresponding to the click.
[0047] From the above description of the RNN training network structure, it can be seen that the RNN network structure is mainly used in the translation model to encode and decode sequence data. The bidirectional RNN layer encodes the input sequence and encodes the semantic information of the input sequence in the translation model. In the internal state of the RNN; the RNN layer decodes the internal state containing semantic information into the target sequence to complete the translation process from the input sequence to the target sequence.
[0048] Therefore, in this step, after the input sequence corresponding to the obtained query is input to the translation model, the obtained output sequence contains the semantic information of the query.
[0049] In 103, the output sequence is used to determine the promotion keywords triggered by the query.
[0050] In this step, when the training data is used to train the translation model, all the queries and titles in the acquired training data are subjected to word segmentation processing, and the vocabulary obtained after word segmentation of all the training data forms a candidate translation character set.
[0051] After obtaining the semantic information of the input query through step 102, the translation model assigns a corresponding probability value to each character in the set of candidate translation characters according to the semantic information of the input query. That is to say, when the input query is different, the probability value assigned by the translation model to each character in the candidate translation character set is different. The input sequence is translated according to the probability value of each character in the candidate translation character set, combined with beam-search technology and a promotion keyword library, to obtain an output sequence.
[0052] The triggering of promotion keywords can be regarded as the translation process from query to promotion keywords, so the above translation model can be used to complete the triggering of query to promotion keywords. However, there is still a problem in this translation process, that is, the translation process using the translation model is free, and the output sequence is an open space, so there is no guarantee that the input query will trigger the advertiser after the translation model is passed. The purchased promotion keywords, and standard machine translation only requires an optimal solution to be returned, which will reduce the recall rate triggered by the promotion keywords.
[0053] In view of the above-mentioned problems, the present invention adopts an improved beam-search technology when using a translation model to translate the input query, that is, before translating the input query, the beam-size and the probability threshold Q are preset. Among them, the probability threshold Q determines the lowest probability value of candidate sequence selection, which is obtained through multiple experimental observations; beam-size determines the number of candidate sequences selected during translation of each layer and the size of the final candidate sequence set. The value needs to be determined by weighing the translation performance and the relevance of the translation result. Moreover, when the present invention uses the translation model to translate the input query, it also needs to be combined with a promotion keyword database, which contains all the promotion keywords purchased by the advertiser.
[0054] Therefore, each output sequence translated by the translation model satisfies the following conditions: the number of output sequences is within the beam-size, and the probability of each output sequence is greater than or equal to the preset probability threshold Q, and in the promotion keyword database There are promotion keywords that are consistent with the output sequence or there are promotion keywords prefixed by the output sequence.
[0055] In the process of translating the input query by the translation model, according to the semantic information of the input query, the translation model assigns the probability value of each character in the candidate translation character set. According to the probability value of each character in the candidate character set, The promotion keyword database obtains beam-size candidate characters, so it is ensured that words not in the promotion keyword database will not appear in the results obtained by the translation model.
[0056] And when the training data is used to train the translation model, a special mark character is automatically added after the query and its corresponding clicked title. The special mark character is used to indicate the end of the translation. Then, by training the training data, the translation model is You can learn when the output words will form a complete sentence and end the translation process.
[0057] Therefore, when a candidate character ends with the special mark character, it indicates that the translation process is over. At the end of the translation, beam-size candidate characters greater than the probability threshold Q are selected as the final candidate characters, and the final candidate characters are the promotion keywords obtained after translating the input query.
[0058] For example, suppose that the candidate translation character set used by the translation model is {a, b, c, d, EOF}, that is, each candidate character in the target sequence is selected from this set during the translation process. Where EOF represents the special mark character at the end of translation, beam-size=2, probability threshold Q=0.3, and the final candidate sequence set is denoted as R.
[0059] When translating the first character of the input query, the translation model gives the probability value of each character as the first character in the candidate translation character set according to the semantic information of the query, such as p(a) represents the probability of character a as the first character of the translation result Value, p(b) represents the probability value of character b as the first character of the translation result, and so on. Sort the probability value of the character from largest to smallest. If p(a) is the largest, first search for promotion keywords prefixed with a in the promotion keyword database. If so, use a as the candidate first character , If not, prune a; set p(b) second only to p(a), then search for b in the promotion keyword database, if so, use b as the candidate first character, if not, then Prune away b. According to this selection rule, the beam-size candidate first characters are finally selected, and the remaining characters are all pruned, and the probability value of each candidate first character of the beam-size must be greater than or equal to the preset probability threshold Q. For the probability value Candidate first characters smaller than Q are also pruned. Because the beam-size=2 is set, two characters are selected as the candidate first characters, assuming that the candidate first characters selected this time are a and b.
[0060] For the second character of the translation, the translation model will calculate the probability of all sequences beginning with a or b {aa,ab,ac,ad,aEOF,ba,bb,bc,bd,bEOF}, the same as the first character selection rule Similarly, sort these sequences in descending order according to the probability value, and search the promotion keyword database to see if there are promotion keywords prefixed with the sequence obtained, if there are any, select them, if not, prune them. Finally, beam-size candidate sequences with a probability value greater than or equal to Q are selected for continued translation.
[0061] Repeat the above translation process. When a candidate sequence ends in EOF, the translation process of the input sequence is considered to be over. If the probability value of the obtained candidate sequence is greater than the probability threshold Q, then the sequence is put into the final candidate sequence set R in. When the size of the set R is equal to beam-size, the translation ends.
[0062] For another example, the query entered by the user is "How to choose Beijing Flower Express". After the query is processed for word segmentation, "how", "choice", "Beijing", "flowers", and "express" are obtained. After the word segmentation is processed The obtained vocabulary is input to the translation model. After the semantics of the input query is known through the translation model, the translation model starts to translate.
[0063] The translation model first determines what the first word is output, and assigns a probability value to each character in the candidate translation character set according to the semantics of the query. The probability value represents the probability of each character in the candidate translation character being the first character. Suppose, the probability values ​​of the three words "how", "where" and "Beijing" are high, but no promotion keywords starting with "how" and "where" are retrieved in the promotion keyword database, but For promotion keywords that start with "Beijing", the first two words are discarded and "Beijing" is selected as the first character. Then the translation model determines what the second word is output, that is, which word is appropriate to output after the word "Beijing". The translation model also assigns a probability value to each character in the set of candidate translation characters. This probability value represents following " The possibility of outputting this word after “Beijing”. Assuming that the second character is “flowers” ​​and “flowers”, the probability value is higher, and searched in the promotion keyword database that there are "Beijing flowers" or "Beijing flowers". For the promotion keywords, select “flowers” ​​and “flowers” ​​as the second word, and so on. Assuming that the translation model has selected the three words "Beijing", "flowers", and "express", when the fourth word is selected, the translation model finds that the fourth character has the greatest probability of being "EOF", and the translation is ended process. At this time, the model will search for the promotion keyword in the promotion keyword database. If there is, and the probability of the promotion keyword is greater than the set probability threshold Q, this word will be used as a candidate promotion keyword Output.
[0064] After using the output sequence to determine the promotion keyword triggered by the query, the promotion keyword can be used to obtain the promotion corresponding to the promotion keyword, and the corresponding search result page after the user enters the query includes The promotion result corresponding to the promotion keyword triggered by the query.
[0065] Using the technical solution provided by the present invention, the query in the user's click behavior log and its corresponding clicked title are selected as training data, and the neural network model is trained to obtain the translation model. After the query input by the user is obtained, the translation obtained by pre-training is obtained The model obtains the promotion keywords triggered by the query. In this way, the promotion keywords that match the query semantically can be obtained. Compared with the way of literal transformation, the accuracy and recall rate of triggering promotion keywords have been greatly improved.
[0066] The following describes the device structure diagram provided by the embodiment of the present invention in detail. Such as image 3 As shown in the figure, the device mainly includes: an acquisition unit 31, a translation unit 32, a determination unit 33, a training unit 34, and a trigger unit 35.
[0067] The obtaining unit 31 is configured to obtain the query input by the user.
[0068] In this step, the query input by the user can be in any form, and can be a single word or a sentence.
[0069] The translation unit 32 is configured to input the input sequence corresponding to the obtained query to the translation model to obtain the output sequence.
[0070] The translation unit 32 needs to perform word segmentation processing on the obtained query, use the word sequence obtained after the word segmentation processing as an input sequence, and then input the input sequence into the translation model for translation.
[0071] The translation model used in the translation unit 32 is pre-trained by the training unit 34, and the training unit 34 trains the neural network model through the acquired training data to obtain the translation model.
[0072] The training data used by the training unit 34 to train the neural network model is: a two-tuple formed by the query obtained from the user's click behavior log and the corresponding clicked title. The clicked title is preferably the title of the clicked promotion result.
[0073] Among them, the user click behavior log comes from the log system of the search engine. The log system of the search engine records the user's click behavior, which reflects the user's search intent and the correlation between search results and promotion. In the log system of search engines, there will be mature algorithms that identify the effectiveness of users and the effectiveness of user click behavior. Therefore, the training data used in training the translation model in this step is the effective click data generated by effective users. The training data can also be called high-quality click behavior data of the user. The translation model obtained by training with the training data will be more accurate and effective when translating the input query.
[0074] When the training unit 34 uses the training data to train the neural network model to obtain the translation model, it needs to segment the query in the training data, use the word sequence obtained after the segmentation as the input sequence, and use the clicked title corresponding to the query in the training data as the target Sequence for training.
[0075] In this step, the neural network model used is preferably Recurrent Neural Network (RNN, RecurrentNeural Network), because the hidden nodes in the network structure of the recurrent neural network model are connected to form a ring, and its internal state is input The sequence is determined, so the network structure of the recurrent neural network model is very suitable for processing long text sequences.
[0076] The RNN network structure is as figure 2 As shown, the structure is mainly divided into 5 parts from bottom to top, completing the training process from the input sequence to the target sequence, and the training process is actually the translation process from the query to the clicked title corresponding to the query.
[0077] Among them, the first layer of the RNN network structure is the word vector layer on the input side. This layer represents each word of the query as a word vector. For example, the query in the training data is "how to cultivate good habits for children", which is obtained after word segmentation The words "how", "cultivation", "child", "of", "good", and "habit" are expressed in the form of word vectors. The number in the box represents the word vector of the word. The cosine similarity is used to measure the semantic similarity between words; the second layer of the RNN network structure is a bidirectional RNN layer, which reads the input word vector sequence from both positive and negative directions, and reads the word vector sequence in an iterative manner The sequence information is encoded into the internal state, because the query in the training data contains rich semantic information. After the query is expressed as a word vector in the first layer, different word vectors contain different semantic information, and the bidirectional RNN layer converts the word vector sequence Encode the semantic information in the RNN, and encode the semantic information of the input sequence in the internal state of the RNN; the third layer of the RNN network structure is the alignment layer, which uses different weight values ​​to weight the internal state of the bidirectional RNN; RNN The fourth layer of the network structure is the RNN layer. This layer calculates the probability of each word in this iteration based on the output of the alignment layer, the internal state of the previous iteration and the translated word sequence in an iterative manner; RNN network The fifth layer of the structure is the word vector layer at the output end, which represents the word vector of each word in the target sequence translated from the input sequence.
[0078] For example, the query in the training data is "How to cultivate children's good habits", and the title corresponding to the query being clicked is "Children's good habits training". Use the RNN network structure to train the training data. The goal of training is to pass the RNN network structure as much as possible. When the query is "how to cultivate children's good habits", the output target sequence is "children's good habits training". That is to say, the purpose of training with the RNN network structure is to make the query in the training data after translation as much as possible to obtain the title of the query corresponding to the click.
[0079] From the above description of the RNN training network structure, it can be seen that the RNN network structure is mainly used in the translation model to encode and decode sequence data. The bidirectional RNN layer encodes the input sequence, and encodes the semantic information of the input sequence in In the internal state of the RNN; and the RNN layer decodes the internal state containing semantic information into the target sequence to complete the translation process from the input sequence to the target sequence.
[0080] Therefore, in this step, the translation unit 32 inputs the obtained input sequence corresponding to the query into the translation model, and the obtained output sequence contains the semantic information of the query.
[0081] The determining unit 33 is configured to determine the promotion keywords triggered by the query by using the output sequence.
[0082] When training the translation model using training data, all the queries and titles in the acquired training data are segmented, and the vocabulary obtained after segmenting all the training data forms a candidate translation character set.
[0083] After the semantic information of the input query is obtained by the translation unit 32, the translation model in the translation unit 32 assigns a corresponding probability value to each character in the set of candidate translation characters according to the semantic information of the input query. In other words, when the input query is different, the probability value assigned by the translation model to each character in the candidate translation character set is different. The input sequence is translated according to the probability value of each character in the candidate translation character set, combined with beam-search technology and a promotion keyword library, to obtain an output sequence.
[0084] The triggering of promotion keywords can be regarded as the translation process from query to promotion keywords, so the above translation model can be used to complete the triggering of query to promotion keywords. However, there is still a problem in this translation process, that is, the translation process using the translation model is free, and the output sequence is an open space, so there is no guarantee that the input query will trigger the advertiser after the translation model is passed. The purchased promotion keywords, and standard machine translation only requires an optimal solution to be returned, which will reduce the recall rate triggered by the promotion keywords.
[0085] In view of the above-mentioned problems, the present invention adopts an improved beam-search technology when using a translation model to translate the input query, that is, before translating the input query, the beam-size and the probability threshold Q are preset. Among them, the probability threshold Q determines the lowest probability value of the candidate sequence selection, which is obtained through multiple experimental observations; beam-size determines the number of candidate sequences selected during translation of each layer and the size of the final candidate sequence set. The value needs to be determined by weighing the translation performance and the relevance of the translation result. Moreover, when the present invention uses the translation model to translate the input query, it also needs to be combined with a promotion keyword library, which contains all the promotion keywords purchased by the advertiser.
[0086] Therefore, each output sequence translated by the translation model satisfies the following conditions: the number of output sequences is within the beam-size, and the probability of each output sequence is greater than or equal to the preset probability threshold Q, and in the promotion keyword database There are promotion keywords that are consistent with the output sequence or there are promotion keywords prefixed by the output sequence.
[0087] In the process of translating the input query by the translation model, according to the semantic information of the input query, the translation model assigns the probability value of each character in the candidate translation character set. According to the probability value of each character in the candidate character set, The promotion keyword database obtains beam-size candidate characters, so it is ensured that words not in the promotion keyword database will not appear in the results obtained by the translation model.
[0088] And when the training data is used to train the translation model, a special mark character is automatically added after the query and its corresponding clicked title. The special mark character is used to indicate the end of the translation. Then, by training the training data, the translation model is You can learn when the output words will form a complete sentence and end the translation process.
[0089] Therefore, when a candidate character ends with the special mark character, it indicates that the translation process is over. At the end of the translation, select beam-size candidate characters greater than the probability threshold Q as the final candidate characters, and the final candidate characters are the promotion keywords obtained after translating the input query.
[0090] For example, suppose that the candidate translation character set used by the translation model is {a, b, c, d, EOF}, that is, each candidate character in the target sequence is selected from this set during the translation process. Where EOF represents the special mark character at the end of translation, beam-size=2, probability threshold Q=0.3, and the final candidate sequence set is denoted as R.
[0091] When translating the first character of the input query, the translation model gives the probability value of each character as the first character in the candidate translation character set according to the semantic information of the query, such as p(a) represents the probability of character a as the first character of the translation result Value, p(b) represents the probability value of character b as the first character of the translation result, and so on. Sort the probability value of the character from largest to smallest. If p(a) is the largest, first search for a promotion keyword prefixed with a in the promotion keyword database, if so, use a as the candidate first character , If not, prune a; set p(b) second only to p(a), then search for b in the promotion keyword database, if so, use b as the candidate first character, if not, then Prune away b. According to this selection rule, the beam-size candidate first character is finally selected, and the remaining characters are all pruned, and the probability value of each candidate first character of the beam-size must be greater than or equal to the preset probability threshold Q. For the probability value Candidate first characters smaller than Q are also pruned. Because the beam-size=2 is set, two characters are selected as the candidate first characters, assuming that the candidate first characters selected this time are a and b.
[0092] For the second character of the translation, the translation model will calculate the probability of all sequences beginning with a or b {aa, ab, ac, ad, aEOF, ba, bb, bc, bd, bEOF}, the same as the first character selection rule Similarly, sort these sequences in descending order according to the probability value, and search for promotion keywords in the promotion keyword database whether there are promotion keywords prefixed with the obtained sequence, if there are, select them, if not, prune them. Finally, beam-size candidate sequences with a probability value greater than or equal to Q are selected for continued translation.
[0093] Repeat the above translation process. When a candidate sequence ends in EOF, the translation process of the input sequence is considered to be over. If the probability value of the obtained candidate sequence is greater than the probability threshold Q, the sequence is put into the final candidate sequence set R in. When the size of the set R is equal to beam-size, the translation ends.
[0094] For another example, the query entered by the user is "How to choose Beijing Flower Express". After the query is processed for word segmentation, "how", "choice", "Beijing", "flowers", and "express" are obtained. After the word segmentation is processed The obtained vocabulary is input to the translation model. After the semantics of the input query is known through the translation model, the translation model starts to translate.
[0095] The translation model first determines what the first word is output, and assigns a probability value to each character in the set of candidate translation characters according to the semantics of the input query. The probability value represents the probability of each character in the candidate translation character being the first character. Suppose, the probability values ​​of the three words "how", "where" and "Beijing" are high, but no promotion keywords starting with "how" and "where" are retrieved in the promotion keyword database, but For promotion keywords that start with "Beijing", the first two words are discarded and "Beijing" is selected as the first character. Then the translation model determines what the second word is output, that is, which word is appropriate to output after the word "Beijing". The translation model also assigns a probability value to each character in the set of candidate translation characters. This probability value represents following " The possibility of outputting this word after “Beijing”. Assuming that the second character is “flowers” ​​and “flowers”, the probability value is higher, and searched in the promotion keyword database that there are "Beijing flowers" or "Beijing flowers". For the promotion keywords, select “flowers” ​​and “flowers” ​​as the second word, and so on. Assuming that the translation model has selected the three words "Beijing", "flowers", and "express", when the fourth word is selected, the translation model finds that the fourth character has the greatest probability of being "EOF", and the translation is ended process. At this time, the model will search for the promotion keyword in the promotion keyword database. If there is, and the probability of the promotion keyword is greater than the set probability threshold Q, this word will be used as a candidate promotion keyword Output.
[0096] After the determining unit 33 uses the output sequence to determine the promotion keyword triggered by the query, the promotion keyword can be used to obtain the promotion corresponding to the promotion keyword, and the trigger unit 35 then corresponds to the query after the user inputs the query The search results page of includes the promotion results corresponding to the promotion keywords triggered by the query.
[0097] The above-mentioned method and device provided by the embodiment of the present invention may be embodied by a computer program installed and running in the device. The device may include one or more processors, memory and one or more programs, such as Figure 4 Shown in. The one or more programs are stored in the memory and executed by the one or more processors to implement the method flow and/or the device operation shown in the foregoing embodiment of the present invention. For example, the method flow executed by the one or more processors may include:
[0098] Get the query entered by the user;
[0099] Input the input sequence corresponding to the obtained query to the translation model to obtain the output sequence;
[0100] Using the output sequence to determine the promotion keywords triggered by the query;
[0101] Wherein, the translation model is obtained by pre-training in the following manner:
[0102] Obtain query and its corresponding clicked title from user click behavior log as training data;
[0103] Use the query in the training data to get the input sequence, use the clicked title corresponding to the query to get the target sequence, train the neural network model, and get the translation model.
[0104] Using the technical solution provided by the present invention, the query in the user's click behavior log and its corresponding clicked title are selected as training data, and the neural network model is trained to obtain the translation model. After the query input by the user is obtained, the translation obtained by pre-training is obtained The model obtains the promotion keywords triggered by the query. In this way, the promotion keywords that match the query semantically can be obtained. Compared with the way of literal transformation, the accuracy and recall rate of triggering promotion keywords have been greatly improved.
[0105] The triggering technology of promotion keywords is one of the core technologies in the search promotion delivery system. By applying the method for triggering promotion keywords disclosed in the present invention to the search promotion delivery system, the promotion key can be triggered by the query entered by the user. The promotion keyword in the thesaurus that matches the semantics of the input query, and then the search promotion placement system will display the promotion corresponding to the promotion keyword, so that the displayed promotion is more in line with the user's search intent.
[0106] In the several embodiments provided by the present invention, it should be understood that the disclosed device and method may be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division, and there may be other division modes in actual implementation.
[0107] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the objectives of the solutions of the embodiments.
[0108] In addition, the functional units in the various embodiments of the present invention may be integrated into one processing unit, or each unit may exist alone physically, or two or more units may be integrated into one unit. The above-mentioned integrated unit can be realized in the form of hardware or in the form of hardware plus software functional unit.
[0109] The above-mentioned integrated unit implemented in the form of a software functional unit may be stored in a computer readable storage medium. The above-mentioned software functional unit is stored in a storage medium and includes several instructions to make a computer device (which may be a personal computer, a server, or a network device, etc.) or a processor execute the method described in the various embodiments of the present invention. Part of the steps. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), magnetic disk or optical disk and other media that can store program code .
[0110] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the present invention Within the scope of protection.

PUM

no PUM

Description & Claims & Application Information

We can also present the details of the Description, Claims and Application information to help users get a comprehensive understanding of the technical details of the patent, such as background art, summary of invention, brief description of drawings, description of embodiments, and other original content. On the other hand, users can also determine the specific scope of protection of the technology through the list of claims; as well as understand the changes in the life cycle of the technology with the presentation of the patent timeline. Login to view more.

Similar technology patents

Adaptive fault detection method for airplane rotation actuator driving device based on deep learning

InactiveCN104914851Aimprove accuracyReduce the false alarm rate of detection
Owner:BEIHANG UNIV

Video monitoring method and system

Owner:深圳辉锐天眼科技有限公司

Classification and recommendation of technical efficacy words

  • Improve recall
  • improve accuracy

Micro blog network opinion leader identification method

InactiveCN103136331AImprove recallimprove accuracy
Owner:NORTHWESTERN POLYTECHNICAL UNIV

Object control method and the device

InactiveCN107220867AImprove recallincrease confidence
Owner:BEIJING XIAODU INFORMATION TECH CO LTD

Remote sensing image small target detection method based on multi-scale characteristics

PendingCN109800755Aeffective detectionImprove recall
Owner:SHENZHEN INST OF ADVANCED TECH CHINESE ACAD OF SCI +1

Method and device for quickly realizing NL2SQL based on vectorized semantic rule

ActiveCN112001188AGood accuracy and generalization abilityImprove recall
Owner:北京智源人工智能研究院

Golf club head with adjustable vibration-absorbing capacity

InactiveUS20050277485A1improve grip comfortimprove accuracy
Owner:FUSHENG IND CO LTD

Stent delivery system with securement and deployment accuracy

ActiveUS7473271B2improve accuracyreduces occurrence and/or severity
Owner:BOSTON SCI SCIMED INC

Method for improving an HS-DSCH transport format allocation

InactiveUS20060089104A1improve accuracyincrease benefit
Owner:NOKIA SOLUTIONS & NETWORKS OY

Catheter systems

ActiveUS20120059255A1increase selectivityimprove accuracy
Owner:ST JUDE MEDICAL ATRIAL FIBRILLATION DIV

Gaming Machine And Gaming System Using Chips

ActiveUS20090075725A1improve accuracy
Owner:UNIVERSAL ENTERTAINMENT CORP
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products