CN109508394A - A kind of training method and device of multi-medium file search order models - Google Patents
A kind of training method and device of multi-medium file search order models Download PDFInfo
- Publication number
- CN109508394A CN109508394A CN201811214519.4A CN201811214519A CN109508394A CN 109508394 A CN109508394 A CN 109508394A CN 201811214519 A CN201811214519 A CN 201811214519A CN 109508394 A CN109508394 A CN 109508394A
- Authority
- CN
- China
- Prior art keywords
- search result
- keyword
- function value
- multimedia file
- sample data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Landscapes
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present invention relates to computer technologies, a kind of method and device of multi-medium file search order models training are disclosed, to improve the sequence accuracy of search result.This method is, after determining keyword and corresponding search result, calculate the sequence labelling function value of the corresponding search result of each keyword, and generate corresponding sample data, wherein, the type of theme correlation degree of the sequence labelling function value and multimedia file being clicked between number and the multimedia file is positively correlated, and filter out the sample data that sequence labelling function value meets preset condition, model training is carried out using preset algorithm, generates corresponding multi-medium file search order models.In this way, the consistent sample data of search intention of keyword and search result can be filtered out, the quality of sample data is improved, and then is capable of the sequence accuracy of the multi-medium file search order models of effective training for promotion acquisition, solves the problems, such as the sorting consistence of search result.
Description
Technical field
The present invention relates to computing technique more particularly to a kind of training methods and dress of multi-medium file search order models
It sets.
Background technique
Under prior art, user carries out multi-medium file search (e.g., film is searched for) on smart television, when, intelligence electricity
Depending on the keyword often inputted according to user, search result is returned to using list mode.In order to be provided subject to more for user
True search result needs to scan for model training based on existing historical search data, however, using existing history
Search data scan for model training and but have the following deficiencies:
1) after smart television receives the keyword that user inputs, all multimedia files comprising keyword can be made
It is presented for search result.
However, only the search result comprising keyword usually contains very high noise from the point of view of current feedback result
Data.
For example, it is assumed that the keyword of user's input is " superman ", then, after smart television scans for, can will include
The film of " superman " two words is used as search result to present, however, user only wants to see the film of superman's series,
Other include " superman " two word, and incoherent film is noise data to content, e.g., The Incredibles, however, " superman is total
Mobilize " etc. the poor search result of correlations may be because the higher reason of temperature, sort before " superman ", therefore, row
Sequence result does not meet the expectation of user.
Obviously, if the search result comprising the higher noise data of temperature is scanned for model as sample data
Training, will affect the sequence accuracy of search model.
2) it in practical application, when inputting keyword by remote controler due to user, in order to save the time, often uses
The first letter of pinyin of multimedia file is inputted as keyword.
However, being inputted first letter of pinyin as keyword from the point of view of current feedback result, will cause antistop list
Meaning inaccuracy can be enabled in search result to a certain extent comprising more noise datas.
For example, it is assumed that the keyword of user's input is " CR ", and in search result, other than " superman ", there are also " grass
A series of texts of people (example: taking life scarecrow by force) ", " adult (example: growing to manhood) ", " successor (example: Shaolin successor) " etc., this
The actual search intention of a little texts and user are far apart, and therefore, after being scanned for using these noise datas, acquisition is searched
It still can include much noise data in hitch fruit.
Similarly, if the search result comprising the higher noise data of temperature is scanned for model as sample data
Training, will affect the sequence accuracy of search model.
In view of this, need to design a kind of method for building up of new multi-medium file search order models, it is above-mentioned to overcome
Defect.
Summary of the invention
The object of the present invention is to provide the training methods and device of a kind of multi-medium file search order models, to mention
The sequence accuracy of high search result.
Specific technical solution provided in an embodiment of the present invention is as follows:
A kind of training method of multi-medium file search order models, comprising:
It determines keyword, and obtains the corresponding search result of each keyword, wherein include in a search result
At least one corresponding multimedia file of corresponding keyword；
The sequence labelling function value of the corresponding search result of each keyword is calculated, and is based on each search result
Sequence labelling function value generate corresponding sample data；Wherein, the sequence labelling function value of a search result, at least with
The type of theme association of the multimedia file that described search result includes being clicked between number and the multimedia file
Degree is positively correlated；
The sample data that sequence labelling function value meets preset condition is filtered out, model training is carried out using preset algorithm,
Generate corresponding multi-medium file search order models.
Optionally, the sequence labelling function value an of search result is calculated, comprising:
The accumulation for counting each multimedia file that one search result includes is clicked number；
Counting each multimedia file that one search result includes is specifying the section in duration to be clicked number
Change rate.
The topic type distribution probability vector of each multimedia file in one search result is counted, and is based on institute
Topic type distribution probability vector is stated, the type of theme correlation degree between each multimedia file is calculated；
The accumulation based on each multimedia file is clicked number, the section is clicked number change rate and institute
Type of theme correlation degree is stated, the sequence labelling function value of one search result is calculated.
Optionally, the accumulation based on each multimedia file is clicked number, the section is clicked number variation
Rate and the type of theme correlation degree before the sequence labelling function value for calculating one search result, further wrap
It includes:
The accumulation for including in one search result is clicked multimedia file that number is not zero as positive example,
And the accumulation for including in one search result is clicked multimedia file that number is zero as negative example.
For each positive example, the highest M of type of theme correlation degree other multimedia files are determined respectively, it is described
Other multimedia files are not included in one search result；
Other multimedia files of selection setting number are right as positive example from other each multimedia files obtained
The negative example is replaced, and the ratio of replaced positive example and negative example is enabled to reach setting ratio thresholding.
Optionally, the sample data that sequence labelling function value meets preset condition is filtered out, comprising:
Filter out the sequence highest N number of sample data of labelling function value, wherein N is default natural number；Alternatively,
Filter out the sample data that sequence labelling function value reaches setup parameter thresholding.
Optionally, it filters out before sequence labelling function value meets the sample data of preset condition, further executes following
Any one in operation or combination:
Filter out the sample data for meeting preset data scale；
Filter out the sample data that the multimedia file that the corresponding search result of keyword includes reaches given threshold；
The issuing time for filtering out the multimedia file that the corresponding search result of keyword includes reaches setting duration thresholding
Sample data；
Deleting keyword is forms data or/and the corresponding sample data of single-letter.
Optionally, model training is carried out using preset algorithm, generates corresponding multi-medium file search order models, wrapped
It includes:
Determine that the linked character of each sample data, the linked character of a sample data include at least keyword respectively
The correlative character between multimedia file that feature, keyword and corresponding search result include, search result include more
The correlative character between multimedia file that the attributive character and search result of media file include；
Sample data is divided into training set and test set；
Based on the training set and corresponding linked character, decision tree GBDT algorithm is promoted using distributed gradient, is carried out
The repetitive exercise of Multiple trees model obtains corresponding training pattern；
Based on the test set and corresponding linked character, the training pattern of generation is tested, generation is commented accordingly
Valence index；
Evaluation index adjusting training parameter based on acquisition optimizes training pattern, obtains finally more by repeatedly training
Media file search order models.
A kind of training device of multi-medium file search order models, comprising:
Acquiring unit, for determining keyword, and the corresponding search result of each keyword of acquisition, wherein one is searched
It include at least one corresponding multimedia file of corresponding keyword in hitch fruit；
Processing unit for calculating the sequence labelling function value of the corresponding search result of each keyword, and is based on
The sequence labelling function value of each search result generates corresponding sample data；Wherein, the sequence of a search result marks letter
Numerical value, the multimedia file at least including with described search result are clicked between number and the multimedia file
Type of theme correlation degree is positively correlated；
Training unit meets the sample data of preset condition for filtering out sequence labelling function value, using preset algorithm
Model training is carried out, corresponding multi-medium file search order models are generated.
Optionally, when calculating the sequence labelling function value an of search result, the processing unit is used for:
The accumulation for counting each multimedia file that one search result includes is clicked number；
Counting each multimedia file that one search result includes is specifying the section in duration to be clicked number
Change rate.
The topic type distribution probability vector of each multimedia file in one search result is counted, and is based on institute
Topic type distribution probability vector is stated, the type of theme correlation degree between each multimedia file is calculated；
The accumulation based on each multimedia file is clicked number, the section is clicked number change rate and institute
Type of theme correlation degree is stated, the sequence labelling function value of one search result is calculated.
Optionally, the accumulation based on each multimedia file is clicked number, the section is clicked number variation
Rate and the type of theme correlation degree, before the sequence labelling function value for calculating one search result, the processing is single
Member is further used for:
The accumulation for including in one search result is clicked multimedia file that number is not zero as positive example,
And the accumulation for including in one search result is clicked multimedia file that number is zero as negative example.
For each positive example, the highest M of type of theme correlation degree other multimedia files are determined respectively, it is described
Other multimedia files are not included in one search result；
Other multimedia files of selection setting number are right as positive example from other each multimedia files obtained
The negative example is replaced, and the ratio of replaced positive example and negative example is enabled to reach setting ratio thresholding.
Optionally, sequence labelling function value is filtered out when meeting the sample data of preset condition, and the training unit is used
In:
Filter out the sequence highest N number of sample data of labelling function value, wherein N is default natural number；Alternatively,
Filter out the sample data that sequence labelling function value reaches setup parameter thresholding.
Optionally, filter out sequence labelling function value meet the sample data of preset condition before, the training unit into
One step for performing the following operations in any one or combination:
Filter out the sample data for meeting preset data scale；
Filter out the sample data that the multimedia file that the corresponding search result of keyword includes reaches given threshold；
The issuing time for filtering out the multimedia file that the corresponding search result of keyword includes reaches setting duration thresholding
Sample data；
Deleting keyword is forms data or/and the corresponding sample data of single-letter.
Optionally, model training is carried out using preset algorithm, when generating corresponding multi-medium file search order models,
The training unit is used for:
Determine that the linked character of each sample data, the linked character of a sample data include at least keyword respectively
The correlative character between multimedia file that feature, keyword and corresponding search result include, search result include more
The correlative character between multimedia file that the attributive character and search result of media file include；
Sample data is divided into training set and test set；
Based on the training set and corresponding linked character, decision tree GBDT algorithm is promoted using distributed gradient, is carried out
The repetitive exercise of Multiple trees model obtains corresponding training pattern；
Based on the test set and corresponding linked character, the training pattern of generation is tested, generation is commented accordingly
Valence index；
Evaluation index adjusting training parameter based on acquisition optimizes training pattern, obtains finally more by repeatedly training
Media file search order models.
A kind of training device of multi-medium file search order models includes at least processor and memory, wherein
Processor executes following process for reading the program in memory:
It determines keyword, and obtains the corresponding search result of each keyword, wherein include in a search result
At least one corresponding multimedia file of corresponding keyword；
The sequence labelling function value of the corresponding search result of each keyword is calculated, and is based on each search result
Sequence labelling function value generate corresponding sample data；Wherein, the sequence labelling function value of a search result, at least with
The type of theme association of the multimedia file that described search result includes being clicked between number and the multimedia file
Degree is positively correlated；
The sample data that sequence labelling function value meets preset condition is filtered out, model training is carried out using preset algorithm,
Generate corresponding multi-medium file search order models.
Optionally, when calculating the sequence labelling function value an of search result, the processor is used for:
The accumulation for counting each multimedia file that one search result includes is clicked number；
Counting each multimedia file that one search result includes is specifying the section in duration to be clicked number
Change rate.
The topic type distribution probability vector of each multimedia file in one search result is counted, and is based on institute
Topic type distribution probability vector is stated, the type of theme correlation degree between each multimedia file is calculated；
The accumulation based on each multimedia file is clicked number, the section is clicked number change rate and institute
Type of theme correlation degree is stated, the sequence labelling function value of one search result is calculated.
Optionally, the accumulation based on each multimedia file is clicked number, the section is clicked number variation
Rate and the type of theme correlation degree, before the sequence labelling function value for calculating one search result, the processor
It is further used for:
The accumulation for including in one search result is clicked multimedia file that number is not zero as positive example,
And the accumulation for including in one search result is clicked multimedia file that number is zero as negative example.
For each positive example, the highest M of type of theme correlation degree other multimedia files are determined respectively, it is described
Other multimedia files are not included in one search result；
Other multimedia files of selection setting number are right as positive example from other each multimedia files obtained
The negative example is replaced, and the ratio of replaced positive example and negative example is enabled to reach setting ratio thresholding.
Optionally, sequence labelling function value is filtered out when meeting the sample data of preset condition, and the processor is used for:
Filter out the sequence highest N number of sample data of labelling function value, wherein N is default natural number；Alternatively,
Filter out the sample data that sequence labelling function value reaches setup parameter thresholding.
Optionally, it filters out before sequence labelling function value meets the sample data of preset condition, the processor is into one
Walk any one in for performing the following operations or combination:
Filter out the sample data for meeting preset data scale；
Filter out the sample data that the multimedia file that the corresponding search result of keyword includes reaches given threshold；
The issuing time for filtering out the multimedia file that the corresponding search result of keyword includes reaches setting duration thresholding
Sample data；
Deleting keyword is forms data or/and the corresponding sample data of single-letter.
Optionally, model training is carried out using preset algorithm, when generating corresponding multi-medium file search order models,
The processor is used for:
Determine that the linked character of each sample data, the linked character of a sample data include at least keyword respectively
The correlative character between multimedia file that feature, keyword and corresponding search result include, search result include more
The correlative character between multimedia file that the attributive character and search result of media file include；
Sample data is divided into training set and test set；
Based on the training set and corresponding linked character, decision tree GBDT algorithm is promoted using distributed gradient, is carried out
The repetitive exercise of Multiple trees model obtains corresponding training pattern；
Based on the test set and corresponding linked character, the training pattern of generation is tested, generation is commented accordingly
Valence index；
Evaluation index adjusting training parameter based on acquisition optimizes training pattern, obtains finally more by repeatedly training
Media file search order models.
A kind of storage medium is stored with the program of the training for realizing multi-medium file search order models, the journey
When sequence is run by processor, following steps are executed:
It determines keyword, and obtains the corresponding search result of each keyword, wherein include in a search result
At least one corresponding multimedia file of corresponding keyword；
The sequence labelling function value of the corresponding search result of each keyword is calculated, and is based on each search result
Sequence labelling function value generate corresponding sample data；Wherein, the sequence labelling function value of a search result, at least with
The type of theme association of the multimedia file that described search result includes being clicked between number and the multimedia file
Degree is positively correlated；
The sample data that sequence labelling function value meets preset condition is filtered out, model training is carried out using preset algorithm,
Generate corresponding multi-medium file search order models.
In the embodiment of the present invention, after determining keyword and corresponding search result, calculate that each keyword is corresponding to be searched
The sequence labelling function value of hitch fruit, and generate corresponding sample data, wherein the sequence labelling function value and multimedia
The type of theme correlation degree of file being clicked between number and the multimedia file is positively correlated, and the row of filtering out
Sequence labelling function value meets the sample data of preset condition, carries out model training using preset algorithm, generates corresponding multimedia
File search order models.In this way, the mark of sample data can be effectively completed, to filter out keyword and search knot
The consistent sample data of the search intention of fruit improves the quality of sample data, and then can effectively improve training acquisition
The sequence accuracy of multi-medium file search order models, solves the problems, such as the sorting consistence of search result.
Detailed description of the invention
Fig. 1 is the training flow diagram of multi-medium file search order models in the embodiment of the present invention；
Fig. 2 is the training device illustrative view of functional configuration of multi-medium file search order models in the embodiment of the present invention；
Fig. 3 is the training device entity structure schematic diagram of multi-medium file search order models in the embodiment of the present invention.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete
Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, is not whole embodiments.It is based on
Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts it is all its
His embodiment, shall fall within the protection scope of the present invention.
As shown in fig.1, in the embodiment of the present invention, establishing multi-medium file search order models in the embodiment of the present invention
Detailed process it is as follows:
Step 100: determining the keyword of setting, and obtain the corresponding search result of each keyword, wherein one
It include at least one corresponding multimedia file of corresponding keyword in search result.
Step 110: calculating the sequence labelling function value of the corresponding search result of each keyword, a search result
Sequence labelling function value, the multimedia file at least including with described search result is clicked number and more matchmakers
File type associations degree between body file is positively correlated.
Specifically exist, by taking a search result x as an example, calculate search result x sequence labelling function value when, include but
It is not limited to following operation:
Step A: the accumulation of each multimedia file is clicked number in statistics search result x.
In order to guarantee the reasonability and accuracy of sample data mark, each multimedia text in search result x can be counted
The accumulation of part is clicked number, wherein so-called accumulation is clicked number, after referring to multimedia file publication, arrives statistics
Total degree is clicked between current time.
For example, it is assumed that after one keyword of input, it is assumed that the keyword set Q={ q of acquisition1,…,qm, each pass
Keyword qi corresponds to a search result, includes film list in search result1≤i≤m,
And the record result of user's one click is characterized asWherein, the shadow being clicked
Piece is equal to 1, other films are equal to 0, specifically as shown in table 1, then, and in the search result obtained based on keyword qi, each film
(i.e. multimedia file) from the accumulation after publication by points number be characterized asWherein, tire out
Product is clicked that number is more, and keyword qi is relative to search resultBetween search intention it is stronger, i.e., correlation degree is got over
By force.
Table 1 (qi):
Step B: section of each multimedia file that statistics search result x includes in specified duration is clicked number
Change rate.
In practical application, the accumulation of multimedia file, which is clicked number, can not reflect the search result of nearest a period of time
Variation tendency it is more to measure number change rate therefore can be clicked according to section of the multimedia file in specified duration
Click situation of change of the media file in current period, for example, being clicked number based on section for new online film
Change rate can quickly determine ordering relation of the film in search result, avoid accumulating quilt caused by show time is too short
Number of clicks is insufficient and leads to film sequence rearward.
Optionally, section is clicked number change rate and can be indicated using following formula, wherein when Δ t indicates specified
Long, T indicates that section is clicked the start time of number change rate: as follows:
Step C: the topic type distribution probability vector for each multimedia file that statistics search result x includes, and be based on
The topic type distribution probability vector calculates the type of theme correlation degree between each multimedia file.
Optionally, (Latent Dirichlet Allocation, LDA) theme can be distributed using implicit Di Li Cray
Clustering algorithm calculates the theme distribution probability vector of each multimedia file
Wherein, Topic characterizes theme, P (Topic1) characterization master
The probability of topic 1, S characterize theme number.The distribution of theme distribution probability vector reflects different multimedia files in different themes
On preference, can be used to measure the type of theme correlation degree of two multimedia files.Therefore, for search result
The multimedia file that it is 0 that middle accumulation, which is clicked number not, can be based on the multimedia text that above-mentioned theme distribution probability vector characterizes
Part is clicked the ranking results of number from high to low, calculates separately the more matchmakers of each multimedia file and sequence before it
The cosine similarity of body file, so that the type of theme correlation degree between multimedia file two-by-two is obtained, specifically, calculating
Formula is as follows:
Step D: the accumulation based on each multimedia file is clicked number, the section is clicked number variation
Rate and the type of theme correlation degree calculate the sequence labelling function value of search result x.
Specifically, the sequence labelling function value of search result x can be calculated using the following equation:
Wherein, α and β is scale factor, it is ensured that each component is same magnitude, in addition, for steady long tail effect shape
At partial velocities, can also to the functional value of above-mentioned sequence labelling function value carry out logarithm floor operation with optimize sample mark,
Then the output result of final sample data can indicate are as follows:
Above-mentioned expression formula characterizes the corresponding search result obtained after input keyword qi, i.e. keyword qi+ search knot
Fruit and corresponding sequence labelling function value characterize one and mark the sample data finished.
Using aforesaid way, the mark of sample data can be carried out respectively for each keyword, to obtain carry out mould
Sample data after all marks of type training.
On the other hand, in the embodiment of the present invention, after issuing a period of time, accumulate by points number be still 0 more matchmakers
Body file, it is believed that it is completely uncorrelated between such multimedia file and corresponding keyword and the search intention of user,
Also the value of multimedia file sequence can not be embodied, it therefore, can be using such multimedia file as including in search result
Negative example, using the accumulation multimedia file that be clicked number not be 0 as the positive example for including in search result.Since search is clicked
Long tail effect, it is generally the case that even if accumulation long period (i.e. multimedia file has issued the long period), search result
In include negative number of cases amount also can much larger than positive example therefore, can be based on the theme of multimedia file in the embodiment of the present invention
Correlation is replaced the negative example in part in search result, flat so as to adjust the ratio of positive example and negative example in search result
Weighing apparatus, and then further increase the quality of sample data；Because introducing new positive example also will increase the multiplicity that search result is recalled
Property, Sorting space is expanded, the performance of Ranking Algorithm is effectively improved, to further increase the multimedia text of subsequent training
The sequence accuracy of part searching order model.
Optionally, still by taking the corresponding search result x of keyword qi as an example, in the sequence labelling function for calculating search result x
Before value, performed replacement step is as follows:
One, the accumulation for including in search result x is clicked multimedia file that number is not zero as positive example, and
The accumulation for including in search result x is clicked multimedia file that number is zero as negative example.
Two, it is directed to each positive example, determines the highest M of type of theme correlation degree other multimedia files, institute respectively
It states other multimedia files and described search result x and the keyword qi is not corresponding, M is default natural number.
Optionally, LDA Subject Clustering algorithm still can be used, by calculating theme distribution probability vector, is obtained and one
Other strongest collection of multimedia documents of the theme relevance of a positive example
Wherein, rlOther multimedia files are characterized,Characterize the theme distribution probability vector cosine similarity of the two.
For example, it is assumed that keyword qi is " superman ", then the positive example obtained is " superman 1 " etc., and the negative example obtained is
" variable body superman ", " prehistory superman " etc., it is assumed that the ratio of positive and negative example is 1:3.So, it is associated with the type of theme of " superman 1 "
Other a multimedia files of degree highest 2 (assuming that M=2) are " just person alliance " and " body of steel ".
In the present embodiment, only by taking a positive example as an example, in practical application, for each positive example, theme will be selected
Highest M other multimedia files of type association degree, such as, it is assumed that there are 5 positive examples, then needing to obtain sum is 5M's
Other multimedia files, and therefrom choose the positive side for replacing negative example.
Three, other multimedia files of setting number are chosen from other each multimedia files obtained as positive example
The negative example is replaced, the ratio of replaced positive example and negative example is enabled to reach setting ratio thresholding.
For example, can will be with other a multimedias of type of theme correlation degree highest 2 (assuming that M=2) of " superman 1 "
File " just person alliance " and " body of steel " are used as positive example, and replacing negative example is " variable body superman ", and " prehistory superman " enables most
The ratio of whole positive example and negative example, which reaches, is approximately equal to 1:1 or 2:1.
Of course, it is possible to pre-generate when search engine is established and indexed and save several theme relative sets
In this way, whichsoever multimedia file, can be at any time from corresponding main as positive side after some cycles are searched in accumulation
Inscribe relative setOther multimedia files needed for middle acquisition are replaced negative side as positive example.
After being replaced to the negative example in part, the positive example being newly added is possible to be clicked number and area there are no accumulation
Between be clicked number change rate, therefore, initial value can be labeled based on the cosine similarity of theme distribution probability vector and
Sequence after accumulating a period of time, then calculates that corresponding accumulation is clicked number, section is clicked number change rate and theme class
Type correlation degree, and final and original positive example is together, calculates the sequence labelling function value of search result x, and be converted into marking
Sample data after note.
Further, after accumulating the fixed period, if the accumulation for the positive example being newly added remains as 0 by points number,
As negative example, it is replaced, is will not be described in great detail again using same way.
Step 120: filtering out the sample data that sequence labelling function value meets preset condition, carried out using preset algorithm
Model training generates corresponding multi-medium file search order models.
Specifically, can filter out good sample data as far as possible after the sample data for obtaining the magnanimity that mark finishes and use
In subsequent model training, optionally, above-mentioned preset condition, which may is that, filters out the sequence highest N number of sample of labelling function value
Data (N is setting natural number), above-mentioned preset condition is also possible to: filtering out sequence labelling function value and reaches setup parameter door
The sample data of limit.
This is because sequence labelling function value is higher, illustrate the search intention between keyword and corresponding search result
Correspondence is more clear, and therefore, the higher sample data of sequence labelling function value is more high-quality, is more advantageous to and improves subsequent training
The sequence accuracy of multi-medium file search order models.
Further, in addition to carrying out sample data screening according to sequence labelling function value, in the embodiment of the present invention, in base
It, can also be one of in the following ways or any before sequence labelling function value obtains final good sample data
Combination carries out the prescreening of sample data:
Under big data environment, the generalization ability of model training can be enhanced based on the sample data set of magnanimity, because
The quantity of this sample data is the bigger the better, and the longer the better for the duration of accumulation, it is therefore possible to use mode including but not limited to:
Mode 1: the sample data for meeting preset data scale is filtered out.
Such as: the total quantity for the sample data for needing to filter out >=1,000,000.
Another example is: needing to filter out data volume >=200,000 of the sample data as test sample set.
Mode 2: the sample number that the multimedia file that the corresponding search result of keyword includes reaches given threshold is filtered out
According to.
Such as: need to filter out the sample data for multimedia file >=60 that search result includes.
Mode 3: when filtering out the issuing time of the multimedia file that the corresponding search result of keyword includes and reaching setting
The sample data of long thresholding.
Such as: need to filter out the sample data of publication duration >=15 day for the multimedia file that search result includes.
Mode 4: deleting keyword is forms data or/and the corresponding sample data of single-letter.
In search log, it often will appear the search of single digital, single-letter, search purpose is indefinite, but tired
It is long-pending to be clicked that number is very high, and the search result hit rate returned is also very big, this is typical bad sample data, is needed
It removes, to guarantee the search intention consistency between keyword and search result.
It is further, excellent required for filtering out in the embodiment of the present invention based on above-mentioned sample data screening strategy
After the sample data of matter, model training can be carried out using preset algorithm, generate corresponding multi-medium file search sequence mould
Type specifically includes:
1) determine that the linked character of each sample data, the linked character of a sample data include at least key respectively
The correlative character between multimedia file that word feature, keyword and corresponding search result include, what search result included
The correlative character between multimedia file that the attributive character and search result of multimedia file include.
Above-mentioned each category feature can obtain corresponding search knot by keyword using the method calculated in search
It extracts, can also be extracted during carrying out positive and negative example replacement when fruit, during sequence labelling function value can also be calculated
It extracts, can also be extracted after calculating sequence labelling function value, in the journal file that can be saved after extraction, carrying out model
When training, it can obtain and use from journal file, will not be described in great detail.
Specifically, the linked character of sample data can be used to the form table of feature vector refering to shown in table 2
Show, be denoted as: f=(fq,fq-d,fd,fd-d)；
Wherein, refering to shown in table 2, fq, keyword feature is indicated, at least by the length of keyword and keyword in history
The number composition occurred in search.
fq-d, the correlative character between multimedia file that expression keyword and corresponding search result include, at least
By accounting, the offset of keyword and multimedia file title, keyword and more matchmakers of the keyword in multimedia file title
The term frequency-inverse document word frequency (TF-IDF) of the text similarity (BM25) of body file, keyword and multimedia file forms.
fd, indicate the attributive character of multimedia file, at least illustrate multimedia file in the measurement of multiple dimensions.
fd-d, the correlative character between multimedia file is indicated, by each multimedia text at least in expression search result
Theme distribution probability similarity after part is clicked number sequence according to accumulation, between adjacent multimedia file.
Table 2
2) sample data is converted into specified format.
Optionally, sample data can be converted to meet Spark MLlib (Apache Spark be aim at it is extensive
Data processing and the computing engines of Universal-purpose quick designed, MLlib is library machine learning (ML) of Spark) sequence learn class
Type LabelPoint (a kind of basic data type of MLlib) format.
3) sample data after conversion is divided into training set and test set.
4) it is based on training set and corresponding linked character, decision tree (Gradient is promoted using distributed gradient
Boosting Decision Tree, GBDT) algorithm, the repetitive exercise of Multiple trees model is carried out, training pattern is obtained.
Specifically, can construct on the basis of the distributed GBDT algorithm of Spark MLLib based on Lambda algorithm
It is polynary to add regression tree (Lambda and Multiple Additive Regression Tree, LambdaMART) progress more
The repetitive exercise of decision-tree model.
Wherein, optionally, algorithm iteration number: 300~500 generations, the depth of MART tree: 3 layers, Learning Step: 0.05,
Loss function: L2loss function:Wherein, i is true value, and F (xi) is predicted value, and N is sample
Data number, Loss value is smaller, then it is better to characterize the training pattern.
5) it is based on test set and corresponding linked character, the training pattern of generation is tested, generates corresponding evaluation
Index.
It optionally, can be using degree of fitting and NDGG as evaluation index, wherein
Degree of fitting: LambdaMART algorithm is evaluated using root-mean-square error (Root Mean Squared Error, RMSE)
Degree of fitting on test set；
Normalization is lost storage gain (Normalized Discounted Cumulative Gain, NDCG): using
The sequence effect of NDCG evaluation training pattern.
6) the evaluation index adjusting training parameter based on acquisition optimizes training pattern, and repeatedly training obtains optimal more matchmakers
Body file search order models.
Optionally, can use Spark MLlib bottom tree construction (DecisionTreeRegressionMode) will most
Excellent multi-medium file search order models are converted to XML format, facilitate ElasticSearch (searching based on Lucene
Rope server provides the full-text search engine of a distributed multi-user ability) search engine load.
The optimal multi-medium file search order models of acquisition are stored in hadoop distribution in the form of an xml-file
In formula file system (Hadoop Distributed File System, HDFS), then, ElasticSearch phase is being updated
When closing index, optimal multi-medium file search order models are written to specified index field, finally, in search statement
Middle addition Ranking Algorithm (Learning to Rank, LTR) model calls, in this way, can be using optimal multimedia text
Part searching order model scans for the crux word being newly entered, and obtains corresponding search result, in described search result
Include current most accurate multimedia file, and presents current most accurate ranking results between multimedia file.
Certainly, above-mentioned multi-medium file search order models (i.e. LTR model) need to update according to the setting period: such as, pressing
Sample data is reselected and replaced according to the setting period, and re-starts model training, to ensure that multi-medium file search sorts
The timeliness of model.
Obviously, the linked character multi-medium file search order models obtained based on sample data, can be accurately
Guarantee the search intention consistency between the search result of keyword, effectively increases the accuracy of search result, and improve
The sequence for the multimedia file that search result includes is accurate, i.e. effective use Ranking Algorithm optimization multimedia file is searched
The ranking results of rope.
Based on the above embodiment, as shown in fig.2, in the embodiment of the present invention, a kind of multi-medium file search row is provided
The training device of sequence model, the training device include at least:
Acquiring unit 20, for determining keyword, and the corresponding search result of each keyword of acquisition, wherein one
It include at least one corresponding multimedia file of corresponding keyword in search result；
Processing unit 21, for calculating the sequence labelling function value of the corresponding search result of each keyword, Yi Jiji
Corresponding sample data is generated in the sequence labelling function value of each search result；Wherein, the sequence mark of a search result
Functional value, the multimedia file at least including with described search result are clicked between number and the multimedia file
Type of theme correlation degree be positively correlated；
Training unit 22 meets the sample data of preset condition for filtering out sequence labelling function value, using pre- imputation
Method carries out model training, generates corresponding multi-medium file search order models.
Optionally, when calculating the sequence labelling function value an of search result, processing unit 21 is used for:
The accumulation for counting each multimedia file that one search result includes is clicked number；
Counting each multimedia file that one search result includes is specifying the section in duration to be clicked number
Change rate.
The topic type distribution probability vector of each multimedia file in one search result is counted, and is based on institute
Topic type distribution probability vector is stated, the type of theme correlation degree between each multimedia file is calculated；
The accumulation based on each multimedia file is clicked number, the section is clicked number change rate and institute
Type of theme correlation degree is stated, the sequence labelling function value of one search result is calculated.
Optionally, the accumulation based on each multimedia file is clicked number, the section is clicked number variation
Rate and the type of theme correlation degree, before the sequence labelling function value for calculating one search result, processing unit 21
It is further used for:
The accumulation for including in one search result is clicked multimedia file that number is not zero as positive example,
And the accumulation for including in one search result is clicked multimedia file that number is zero as negative example.
For each positive example, the highest M of type of theme correlation degree other multimedia files are determined respectively, it is described
Other multimedia files are not included in one search result；
Other multimedia files of selection setting number are right as positive example from other each multimedia files obtained
The negative example is replaced, and the ratio of replaced positive example and negative example is enabled to reach setting ratio thresholding.
Optionally, sequence labelling function value is filtered out when meeting the sample data of preset condition, and training unit 22 is used for:
Filter out the sequence highest N number of sample data of labelling function value, wherein N is default natural number；Alternatively,
Filter out the sample data that sequence labelling function value reaches setup parameter thresholding.
Optionally, filter out sequence labelling function value meet the sample data of preset condition before, training unit 22 into one
Walk any one in for performing the following operations or combination:
Filter out the sample data for meeting preset data scale；
Filter out the sample data that the multimedia file that the corresponding search result of keyword includes reaches given threshold；
The issuing time for filtering out the multimedia file that the corresponding search result of keyword includes reaches setting duration thresholding
Sample data；
Deleting keyword is forms data or/and the corresponding sample data of single-letter.
Optionally, model training is carried out using preset algorithm, when generating corresponding multi-medium file search order models,
Training unit 22 is used for:
Determine that the linked character of each sample data, the linked character of a sample data include at least keyword respectively
The correlative character between multimedia file that feature, keyword and corresponding search result include, search result include more
The correlative character between multimedia file that the attributive character and search result of media file include；
Sample data is divided into training set and test set；
Based on the training set and corresponding linked character, decision tree GBDT algorithm is promoted using distributed gradient, is carried out
The repetitive exercise of Multiple trees model obtains corresponding training pattern；
Based on the test set and corresponding linked character, the training pattern of generation is tested, generation is commented accordingly
Valence index；
Evaluation index adjusting training parameter based on acquisition optimizes training pattern, obtains finally more by repeatedly training
Media file search order models.
Based on the above embodiment, as shown in fig.3, in the embodiment of the present invention, a kind of multi-medium file search row is provided
The training device of sequence model, the training device include at least:
Processor 300 executes following process for reading the program in memory 310:
It determines keyword, and obtains the corresponding search result of each keyword, wherein include in a search result
At least one corresponding multimedia file of corresponding keyword；
The sequence labelling function value of the corresponding search result of each keyword is calculated, and is based on each search result
Sequence labelling function value generate corresponding sample data；Wherein, the sequence labelling function value of a search result, at least with
The type of theme association of the multimedia file that described search result includes being clicked between number and the multimedia file
Degree is positively correlated；
The sample data that sequence labelling function value meets preset condition is filtered out, model training is carried out using preset algorithm,
Generate corresponding multi-medium file search order models.
Wherein, in Fig. 3, bus architecture may include the bus and bridge of any number of interconnection, specifically by processor
The various circuits for the memory that 300 one or more processors represented and memory 310 represent link together.Total coil holder
Structure can also link together various other circuits of such as peripheral equipment, voltage-stablizer and management circuit or the like, this
It is all a bit it is known in the art, therefore, it will not be further described herein.Bus interface provides interface.Transceiver
It can be multiple element, that is, include transmitter and receiver, provide for over a transmission medium being communicated with various other devices
Unit.For different user equipmenies, user interface, which can also be, external the interface for needing equipment is inscribed, and connection is set
Standby including but not limited to keypad, display, loudspeaker, microphone, control stick etc..
Processor 300, which is responsible for management bus architecture and common processing, memory 310, can store processor 300 and is holding
Used data when row operation.
Optionally, when calculating the sequence labelling function value an of search result, processor 300 is used for:
The accumulation for counting each multimedia file that one search result includes is clicked number；
Counting each multimedia file that one search result includes is specifying the section in duration to be clicked number
Change rate.
The topic type distribution probability vector of each multimedia file in one search result is counted, and is based on institute
Topic type distribution probability vector is stated, the type of theme correlation degree between each multimedia file is calculated；
The accumulation based on each multimedia file is clicked number, the section is clicked number change rate and institute
Type of theme correlation degree is stated, the sequence labelling function value of one search result is calculated.
Optionally, the accumulation based on each multimedia file is clicked number, the section is clicked number variation
Rate and the type of theme correlation degree, before the sequence labelling function value for calculating one search result, processor 300
It is further used for:
The accumulation for including in one search result is clicked multimedia file that number is not zero as positive example,
And the accumulation for including in one search result is clicked multimedia file that number is zero as negative example.
For each positive example, the highest M of type of theme correlation degree other multimedia files are determined respectively, it is described
Other multimedia files are not included in one search result；
Other multimedia files of selection setting number are right as positive example from other each multimedia files obtained
The negative example is replaced, and the ratio of replaced positive example and negative example is enabled to reach setting ratio thresholding.
Optionally, sequence labelling function value is filtered out when meeting the sample data of preset condition, and processor 300 is used for:
Filter out the sequence highest N number of sample data of labelling function value, wherein N is default natural number；Alternatively,
Filter out the sample data that sequence labelling function value reaches setup parameter thresholding.
Optionally, filter out sequence labelling function value meet the sample data of preset condition before, processor 300 into one
Walk any one in for performing the following operations or combination:
Filter out the sample data for meeting preset data scale；
Filter out the sample data that the multimedia file that the corresponding search result of keyword includes reaches given threshold；
The issuing time for filtering out the multimedia file that the corresponding search result of keyword includes reaches setting duration thresholding
Sample data；
Deleting keyword is forms data or/and the corresponding sample data of single-letter.
Optionally, model training is carried out using preset algorithm, when generating corresponding multi-medium file search order models,
Processor 300 is used for:
Determine that the linked character of each sample data, the linked character of a sample data include at least keyword respectively
The correlative character between multimedia file that feature, keyword and corresponding search result include, search result include more
The correlative character between multimedia file that the attributive character and search result of media file include；
Sample data is divided into training set and test set；
Based on the training set and corresponding linked character, decision tree GBDT algorithm is promoted using distributed gradient, is carried out
The repetitive exercise of Multiple trees model obtains corresponding training pattern；
Based on the test set and corresponding linked character, the training pattern of generation is tested, generation is commented accordingly
Valence index；
Evaluation index adjusting training parameter based on acquisition optimizes training pattern, obtains finally more by repeatedly training
Media file search order models.
A kind of storage medium is stored with the program of the training for realizing multi-medium file search order models, the journey
When sequence is run by processor, following steps are executed:
It determines keyword, and obtains the corresponding search result of each keyword, wherein include in a search result
At least one corresponding multimedia file of corresponding keyword；
The sequence labelling function value of the corresponding search result of each keyword is calculated, and is based on each search result
Sequence labelling function value generate corresponding sample data；Wherein, the sequence labelling function value of a search result, at least with
The type of theme association of the multimedia file that described search result includes being clicked between number and the multimedia file
Degree is positively correlated；
The sample data that sequence labelling function value meets preset condition is filtered out, model training is carried out using preset algorithm,
Generate corresponding multi-medium file search order models.
Based on the same inventive concept, a kind of storage medium is provided, is stored with for realizing multi-medium file search sequence mould
The program of the training of type when described program is run by processor, executes following steps:
It determines keyword, and obtains the corresponding search result of each keyword, wherein include in a search result
At least one corresponding multimedia file of corresponding keyword；
The sequence labelling function value of the corresponding search result of each keyword is calculated, and is based on each search result
Sequence labelling function value generate corresponding sample data；Wherein, the sequence labelling function value of a search result, at least with
The type of theme association of the multimedia file that described search result includes being clicked between number and the multimedia file
Degree is positively correlated；
The sample data that sequence labelling function value meets preset condition is filtered out, model training is carried out using preset algorithm,
Generate corresponding multi-medium file search order models.
Based on the above embodiment, it in the embodiment of the present invention, after determining keyword and corresponding search result, calculates each
The sequence labelling function value of the corresponding search result of a keyword, and generate corresponding sample data, wherein the sequence mark
Infuse the type of theme correlation degree positive of functional value and multimedia file being clicked between number and the multimedia file
It closes, and filters out the sample data that sequence labelling function value meets preset condition, model training is carried out using preset algorithm,
Generate corresponding multi-medium file search order models.In this way, the mark of sample data can be effectively completed, to screen
The consistent sample data of the search intention of keyword and search result out improves the quality of sample data, and then can be effective
The sequence accuracy of the multi-medium file search order models of training acquisition is improved, solves the sorting consistence of search result
Problem.
It should be understood by those skilled in the art that, the embodiment of the present invention can provide as method, system or computer journey
Sequence product.Therefore, complete hardware embodiment, complete software embodiment or combining software and hardware aspects can be used in the present invention
The form of embodiment.Moreover, it wherein includes the calculating of computer usable program code that the present invention, which can be used in one or more,
The computer program implemented in machine usable storage medium (including but not limited to magnetic disk storage, CD-ROM, optical memory etc.)
The form of product.
The present invention be referring to according to the method for the embodiment of the present invention, the process of equipment (system) and computer program product
Figure and/or block diagram describe.It should be understood that can be realized by computer program instructions each in flowchart and/or the block diagram
The combination of process and/or box in process and/or box and flowchart and/or the block diagram.It can provide these computers
Processor of the program instruction to general purpose computer, special purpose computer, Embedded Processor or other programmable data processing devices
To generate a machine, so that being generated by the instruction that computer or the processor of other programmable data processing devices execute
For realizing the function of being specified in one or more flows of the flowchart and/or one or more blocks of the block diagram
Device.
These computer program instructions, which may also be stored in, is able to guide computer or other programmable data processing devices with spy
Determine in the computer-readable memory that mode works, so that instruction stored in the computer readable memory generation includes
The manufacture of command device, the command device are realized in one box of one or more flows of the flowchart and/or block diagram
Or the function of being specified in multiple boxes.
These computer program instructions also can be loaded onto a computer or other programmable data processing device, so that
Series of operation steps are executed on computer or other programmable devices to generate computer implemented processing, thus calculating
The instruction executed on machine or other programmable devices is provided for realizing in one or more flows of the flowchart and/or side
The step of function of being specified in block diagram one box or multiple boxes.
Although preferred embodiments of the present invention have been described, once a person skilled in the art knows basic wounds
The property made concept, then additional changes and modifications may be made to these embodiments.It is wrapped so the following claims are intended to be interpreted as
It includes preferred embodiment and falls into all change and modification of the scope of the invention.
Obviously, those skilled in the art can carry out various modification and variations without departing from this to the embodiment of the present invention
The spirit and scope of inventive embodiments.If being wanted in this way, these modifications and variations of the embodiment of the present invention belong to right of the present invention
Ask and its equivalent technologies within the scope of, then the present invention is also intended to include these modifications and variations.
Claims (9)
1. a kind of training method of multi-medium file search order models characterized by comprising
It determines keyword, and obtains the corresponding search result of each keyword, wherein comprising corresponding in a search result
At least one corresponding multimedia file of keyword；
Calculate the sequence labelling function value of the corresponding search result of each keyword, and the sequence based on each search result
Labelling function value generates corresponding sample data；Wherein, the sequence labelling function value of a search result, at least with described search
As a result the type of theme correlation degree positive of the multimedia file for including being clicked between number and the multimedia file
It closes；
The sample data that sequence labelling function value meets preset condition is filtered out, model training is carried out using preset algorithm, is generated
Corresponding multi-medium file search order models.
2. the method as described in claim 1, which is characterized in that calculate the sequence labelling function value of a search result, comprising:
The accumulation for counting each multimedia file that one search result includes is clicked number；
Counting each multimedia file that one search result includes is specifying the section in duration to be clicked number variation
Rate；
The topic type distribution probability vector of each multimedia file in one search result is counted, and is based on the theme
Type distribution probability vector calculates the type of theme correlation degree between each multimedia file；
The accumulation based on each multimedia file is clicked number, the section is clicked number change rate and the theme
Type association degree calculates the sequence labelling function value of one search result.
3. method according to claim 2, which is characterized in that the accumulation based on each multimedia file is clicked secondary
Several, the described section is clicked number change rate and the type of theme correlation degree, calculates the sequence of one search result
Before labelling function value, further comprise:
The accumulation for including in one search result is clicked multimedia file that number is not zero as positive example, and will
The accumulation for including in one search result is clicked multimedia file that number is zero as negative example；
For each positive example, the highest M of type of theme correlation degree other multimedia files are determined respectively, described other are more
Media file is not included in one search result；
Other multimedia files of setting number are chosen from other each multimedia files obtained as positive example, to described negative
Example is replaced, and the ratio of replaced positive example and negative example is enabled to reach setting ratio thresholding.
4. the method as described in claim 1, which is characterized in that filter out the sample that sequence labelling function value meets preset condition
Data, comprising:
Filter out the sequence highest N number of sample data of labelling function value, wherein N is default natural number；Alternatively,
Filter out the sample data that sequence labelling function value reaches setup parameter thresholding.
5. method as claimed in claim 4, which is characterized in that filter out the sample that sequence labelling function value meets preset condition
Before data, any one in following operation or combination are further executed:
Filter out the sample data for meeting preset data scale；
Filter out the sample data that the multimedia file that the corresponding search result of keyword includes reaches given threshold；
The issuing time for filtering out the multimedia file that the corresponding search result of keyword includes reaches the sample of setting duration thresholding
Notebook data；
Deleting keyword is forms data or/and the corresponding sample data of single-letter.
6. the method according to claim 1 to 5, which is characterized in that carry out model training using preset algorithm, generate
Corresponding multi-medium file search order models, comprising:
Determine the linked character of each sample data respectively, the linked character of a sample data include at least keyword feature,
The correlative character between multimedia file that keyword and corresponding search result include, the multimedia text that search result includes
The correlative character between multimedia file that the attributive character and search result of part include；
Sample data is divided into training set and test set；
Based on the training set and corresponding linked character, decision tree GBDT algorithm is promoted using distributed gradient, carries out determine more
The repetitive exercise of plan tree-model obtains corresponding training pattern；
Based on the test set and corresponding linked character, the training pattern of generation is tested, corresponding evaluation is generated and refers to
Mark；
Evaluation index adjusting training parameter based on acquisition optimizes training pattern, obtains final multimedia by repeatedly training
File search order models.
7. a kind of training device of multi-medium file search order models characterized by comprising
Acquiring unit, for determining keyword, and the corresponding search result of each keyword of acquisition, wherein a search knot
It include at least one corresponding multimedia file of corresponding keyword in fruit；
Processing unit, for calculating the sequence labelling function value of the corresponding search result of each keyword, and based on each
The sequence labelling function value of search result generates corresponding sample data；Wherein, the sequence labelling function value of a search result,
The theme class of the multimedia file at least including with described search result being clicked between number and the multimedia file
Type correlation degree is positively correlated；
Training unit is met the sample data of preset condition for filtering out sequence labelling function value, is carried out using preset algorithm
Model training generates corresponding multi-medium file search order models.
8. a kind of training device of multi-medium file search order models, which is characterized in that processor and memory are included at least,
Wherein,
Processor executes following process for reading the program in memory:
It determines keyword, and obtains the corresponding search result of each keyword, wherein comprising corresponding in a search result
At least one corresponding multimedia file of keyword；
Calculate the sequence labelling function value of the corresponding search result of each keyword, and the sequence based on each search result
Labelling function value generates corresponding sample data；Wherein, the sequence labelling function value of a search result, at least with described search
As a result the type of theme correlation degree positive of the multimedia file for including being clicked between number and the multimedia file
It closes；
The sample data that sequence labelling function value meets preset condition is filtered out, model training is carried out using preset algorithm, is generated
Corresponding multi-medium file search order models.
9. a kind of storage medium, which is characterized in that be stored with the journey of the training for realizing multi-medium file search order models
Sequence when described program is run by processor, executes following steps:
It determines keyword, and obtains the corresponding search result of each keyword, wherein comprising corresponding in a search result
At least one corresponding multimedia file of keyword；
Calculate the sequence labelling function value of the corresponding search result of each keyword, and the sequence based on each search result
Labelling function value generates corresponding sample data；Wherein, the sequence labelling function value of a search result, at least with described search
As a result the type of theme correlation degree positive of the multimedia file for including being clicked between number and the multimedia file
It closes；
The sample data that sequence labelling function value meets preset condition is filtered out, model training is carried out using preset algorithm, is generated
Corresponding multi-medium file search order models.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811214519.4A CN109508394A (en) | 2018-10-18 | 2018-10-18 | A kind of training method and device of multi-medium file search order models |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811214519.4A CN109508394A (en) | 2018-10-18 | 2018-10-18 | A kind of training method and device of multi-medium file search order models |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109508394A true CN109508394A (en) | 2019-03-22 |
Family
ID=65746735
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811214519.4A Pending CN109508394A (en) | 2018-10-18 | 2018-10-18 | A kind of training method and device of multi-medium file search order models |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109508394A (en) |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110909182A (en) * | 2019-11-29 | 2020-03-24 | 北京达佳互联信息技术有限公司 | Multimedia resource searching method and device, computer equipment and storage medium |
| CN111221969A (en) * | 2019-12-31 | 2020-06-02 | 国网北京市电力公司 | Text difference recognition method and device |
| CN111444380A (en) * | 2020-03-26 | 2020-07-24 | 腾讯音乐娱乐科技（深圳）有限公司 | Music search sorting method, device, equipment and storage medium |
| CN111597469A (en) * | 2020-05-19 | 2020-08-28 | 北京字节跳动网络技术有限公司 | Display position determining method and device, electronic equipment and storage medium |
| CN111599219A (en) * | 2020-05-27 | 2020-08-28 | 中航信移动科技有限公司 | Multi-data-source flight takeoff time prediction method based on sequencing learning |
| CN111783452A (en) * | 2020-06-30 | 2020-10-16 | 北京百度网讯科技有限公司 | Model training method, information processing method, device, equipment and storage medium |
| CN111782950A (en) * | 2020-06-30 | 2020-10-16 | 北京三快在线科技有限公司 | Sample data set acquisition method, device, equipment and storage medium |
| CN112084435A (en) * | 2020-08-07 | 2020-12-15 | 北京三快在线科技有限公司 | Search ranking model training method and device and search ranking method and device |
| CN112364184A (en) * | 2020-11-12 | 2021-02-12 | 北京达佳互联信息技术有限公司 | Method, device, server and storage medium for ordering multimedia data |
| CN112650952A (en) * | 2020-12-25 | 2021-04-13 | 北京达佳互联信息技术有限公司 | Search sorting method and device |
| CN113076395A (en) * | 2021-03-25 | 2021-07-06 | 北京达佳互联信息技术有限公司 | Semantic model training, search and display method, device, equipment and storage medium |
| CN113326363A (en) * | 2021-05-27 | 2021-08-31 | 北京百度网讯科技有限公司 | Searching method and device, prediction model training method and device, and electronic device |
| CN113535990A (en) * | 2020-11-10 | 2021-10-22 | 腾讯科技（深圳）有限公司 | Method, device, storage medium and electronic equipment for determining multimedia content |
| CN113704507A (en) * | 2021-10-26 | 2021-11-26 | 腾讯科技（深圳）有限公司 | Data processing method, computer device and readable storage medium |
| CN113934872A (en) * | 2021-10-29 | 2022-01-14 | 北京达佳互联信息技术有限公司 | Search result sorting method, device, equipment and storage medium |
| CN114691906A (en) * | 2020-12-29 | 2022-07-01 | 北京达佳互联信息技术有限公司 | Media content processing method and device, electronic equipment and storage medium |
| CN115048587A (en) * | 2022-08-12 | 2022-09-13 | 中博信息技术研究院有限公司 | LambdaMart-based address book search intelligent sorting method |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20110029517A1 (en) * | 2009-07-31 | 2011-02-03 | Shihao Ji | Global and topical ranking of search results using user clicks |
| US20120030152A1 (en) * | 2010-07-30 | 2012-02-02 | Yahoo! Inc. | Ranking entity facets using user-click feedback |
| CN102637179A (en) * | 2011-02-14 | 2012-08-15 | 阿里巴巴集团控股有限公司 | Method and device for determining lexical item weighting functions and searching based on functions |
| CN102722501A (en) * | 2011-03-31 | 2012-10-10 | 北京百度网讯科技有限公司 | Search engine and realization method thereof |
| CN103914478A (en) * | 2013-01-06 | 2014-07-09 | 阿里巴巴集团控股有限公司 | Webpage training method and system and webpage prediction method and system |
| CN104899322A (en) * | 2015-06-18 | 2015-09-09 | 百度在线网络技术（北京）有限公司 | Search engine and implementation method thereof |
| CN106202294A (en) * | 2016-07-01 | 2016-12-07 | 北京奇虎科技有限公司 | The related news computational methods merged based on key word and topic model and device |
| CN107506402A (en) * | 2017-08-03 | 2017-12-22 | 北京百度网讯科技有限公司 | Sort method, device, equipment and the computer-readable recording medium of search result |
| CN108121736A (en) * | 2016-11-30 | 2018-06-05 | 北京搜狗科技发展有限公司 | A kind of descriptor determines the method for building up, device and electronic equipment of model |
| CN108280155A (en) * | 2018-01-11 | 2018-07-13 | 百度在线网络技术（北京）有限公司 | The problem of based on short-sighted frequency, retrieves feedback method, device and its equipment |
| CN108345702A (en) * | 2018-04-10 | 2018-07-31 | 北京百度网讯科技有限公司 | Entity recommends method and apparatus |
-
2018
- 2018-10-18 CN CN201811214519.4A patent/CN109508394A/en active Pending
Patent Citations (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20110029517A1 (en) * | 2009-07-31 | 2011-02-03 | Shihao Ji | Global and topical ranking of search results using user clicks |
| US20120030152A1 (en) * | 2010-07-30 | 2012-02-02 | Yahoo! Inc. | Ranking entity facets using user-click feedback |
| US9262532B2 (en) * | 2010-07-30 | 2016-02-16 | Yahoo! Inc. | Ranking entity facets using user-click feedback |
| CN102637179A (en) * | 2011-02-14 | 2012-08-15 | 阿里巴巴集团控股有限公司 | Method and device for determining lexical item weighting functions and searching based on functions |
| CN102722501A (en) * | 2011-03-31 | 2012-10-10 | 北京百度网讯科技有限公司 | Search engine and realization method thereof |
| CN103914478A (en) * | 2013-01-06 | 2014-07-09 | 阿里巴巴集团控股有限公司 | Webpage training method and system and webpage prediction method and system |
| CN104899322A (en) * | 2015-06-18 | 2015-09-09 | 百度在线网络技术（北京）有限公司 | Search engine and implementation method thereof |
| CN106202294A (en) * | 2016-07-01 | 2016-12-07 | 北京奇虎科技有限公司 | The related news computational methods merged based on key word and topic model and device |
| CN108121736A (en) * | 2016-11-30 | 2018-06-05 | 北京搜狗科技发展有限公司 | A kind of descriptor determines the method for building up, device and electronic equipment of model |
| CN107506402A (en) * | 2017-08-03 | 2017-12-22 | 北京百度网讯科技有限公司 | Sort method, device, equipment and the computer-readable recording medium of search result |
| CN108280155A (en) * | 2018-01-11 | 2018-07-13 | 百度在线网络技术（北京）有限公司 | The problem of based on short-sighted frequency, retrieves feedback method, device and its equipment |
| CN108345702A (en) * | 2018-04-10 | 2018-07-31 | 北京百度网讯科技有限公司 | Entity recommends method and apparatus |
Non-Patent Citations (2)
| Title |
|---|
| MICHAEL BENDERSKY, ET AL.: "Up Next: Retrieval Methods for Large Scale Related Video Suggestion", 《KDD "14: PROCEEDINGS OF THE 20TH ACM SIGKDD INTERNATIONAL CONFERENCE ON KNOWLEDGE DISCOVERY AND DATA MINING》 * |
| QIUSHA ZHU;MEI-LING SHYU;HAOHONG WANG: "VideoTopic: Content-Based Video Recommendation Using a Topic Model", 《2013 IEEE INTERNATIONAL SYMPOSIUM ON MULTIMEDIA》 * |
Cited By (26)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110909182B (en) * | 2019-11-29 | 2023-05-09 | 北京达佳互联信息技术有限公司 | Multimedia resource searching method, device, computer equipment and storage medium |
| CN110909182A (en) * | 2019-11-29 | 2020-03-24 | 北京达佳互联信息技术有限公司 | Multimedia resource searching method and device, computer equipment and storage medium |
| CN111221969A (en) * | 2019-12-31 | 2020-06-02 | 国网北京市电力公司 | Text difference recognition method and device |
| CN111444380B (en) * | 2020-03-26 | 2023-05-02 | 腾讯音乐娱乐科技（深圳）有限公司 | Music search ordering method, device, equipment and storage medium |
| CN111444380A (en) * | 2020-03-26 | 2020-07-24 | 腾讯音乐娱乐科技（深圳）有限公司 | Music search sorting method, device, equipment and storage medium |
| CN111597469A (en) * | 2020-05-19 | 2020-08-28 | 北京字节跳动网络技术有限公司 | Display position determining method and device, electronic equipment and storage medium |
| CN111599219A (en) * | 2020-05-27 | 2020-08-28 | 中航信移动科技有限公司 | Multi-data-source flight takeoff time prediction method based on sequencing learning |
| CN111782950A (en) * | 2020-06-30 | 2020-10-16 | 北京三快在线科技有限公司 | Sample data set acquisition method, device, equipment and storage medium |
| CN111783452B (en) * | 2020-06-30 | 2024-04-02 | 北京百度网讯科技有限公司 | Model training method, information processing method, device, equipment and storage medium |
| CN111783452A (en) * | 2020-06-30 | 2020-10-16 | 北京百度网讯科技有限公司 | Model training method, information processing method, device, equipment and storage medium |
| CN112084435A (en) * | 2020-08-07 | 2020-12-15 | 北京三快在线科技有限公司 | Search ranking model training method and device and search ranking method and device |
| CN113535990A (en) * | 2020-11-10 | 2021-10-22 | 腾讯科技（深圳）有限公司 | Method, device, storage medium and electronic equipment for determining multimedia content |
| CN113535990B (en) * | 2020-11-10 | 2023-12-15 | 腾讯科技（深圳）有限公司 | Method, device, storage medium and electronic equipment for determining multimedia content |
| CN112364184B (en) * | 2020-11-12 | 2024-04-30 | 北京达佳互联信息技术有限公司 | Method, device, server and storage medium for ordering multimedia data |
| CN112364184A (en) * | 2020-11-12 | 2021-02-12 | 北京达佳互联信息技术有限公司 | Method, device, server and storage medium for ordering multimedia data |
| CN112650952A (en) * | 2020-12-25 | 2021-04-13 | 北京达佳互联信息技术有限公司 | Search sorting method and device |
| CN112650952B (en) * | 2020-12-25 | 2023-07-25 | 北京达佳互联信息技术有限公司 | Search ordering method and device |
| CN114691906A (en) * | 2020-12-29 | 2022-07-01 | 北京达佳互联信息技术有限公司 | Media content processing method and device, electronic equipment and storage medium |
| CN113076395B (en) * | 2021-03-25 | 2024-03-26 | 北京达佳互联信息技术有限公司 | Semantic model training and search display method, device, equipment and storage medium |
| CN113076395A (en) * | 2021-03-25 | 2021-07-06 | 北京达佳互联信息技术有限公司 | Semantic model training, search and display method, device, equipment and storage medium |
| CN113326363B (en) * | 2021-05-27 | 2023-07-25 | 北京百度网讯科技有限公司 | Searching method and device, prediction model training method and device and electronic equipment |
| CN113326363A (en) * | 2021-05-27 | 2021-08-31 | 北京百度网讯科技有限公司 | Searching method and device, prediction model training method and device, and electronic device |
| CN113704507B (en) * | 2021-10-26 | 2022-02-11 | 腾讯科技（深圳）有限公司 | Data processing method, computer device and readable storage medium |
| CN113704507A (en) * | 2021-10-26 | 2021-11-26 | 腾讯科技（深圳）有限公司 | Data processing method, computer device and readable storage medium |
| CN113934872A (en) * | 2021-10-29 | 2022-01-14 | 北京达佳互联信息技术有限公司 | Search result sorting method, device, equipment and storage medium |
| CN115048587A (en) * | 2022-08-12 | 2022-09-13 | 中博信息技术研究院有限公司 | LambdaMart-based address book search intelligent sorting method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109508394A (en) | A kind of training method and device of multi-medium file search order models | |
| US9454599B2 (en) | Automatic definition of entity collections | |
| US11216503B1 (en) | Clustering search results | |
| CN103218436B (en) | A kind of Similar Problems search method and device that merges class of subscriber label | |
| Hayat et al. | Best concept selection in design process: An application of generalized intuitionistic fuzzy soft sets | |
| CN108717433A (en) | A kind of construction of knowledge base method and device of programming-oriented field question answering system | |
| CN105005589A (en) | Text classification method and text classification device | |
| CN104111925B (en) | Item recommendation method and device | |
| CN106168980A (en) | Multimedia resource recommends sort method and device | |
| US20150100605A1 (en) | Determining collection membership in a data graph | |
| CN110046298A (en) | Query word recommendation method and device, terminal device and computer readable medium | |
| CN109947902A (en) | A kind of data query method, apparatus and readable medium | |
| Bai et al. | Service recommendation for mashup creation based on time-aware collaborative domain regression | |
| CN116306923A (en) | Evaluation weight calculation method based on knowledge graph | |
| Wei et al. | Online education recommendation model based on user behavior data analysis | |
| Ma et al. | Content Feature Extraction-based Hybrid Recommendation for Mobile Application Services. | |
| CN113987161A (en) | Text sorting method and device | |
| Chen et al. | Time-aware collaborative poisson factorization for service recommendation | |
| Revindasari et al. | Traceability between business process and software component using Probabilistic Latent Semantic Analysis | |
| Wang et al. | A personalization-oriented academic literature recommendation method | |
| Baazaoui-Zghal et al. | A fuzzy-ontology-driven method for a personalized query reformulation | |
| CN115292167A (en) | Life cycle prediction model construction method, device, equipment and readable storage medium | |
| CN108536796A (en) | A kind of isomery Ontology Matching method and system based on figure | |
| CN114692978A (en) | Social media user behavior prediction method and system based on big data | |
| CN110413782B (en) | Automatic table theme classification method and device, computer equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |