CN107609960A - Rationale for the recommendation generation method and device - Google Patents
Rationale for the recommendation generation method and device Download PDFInfo
- Publication number
- CN107609960A CN107609960A CN201710978303.4A CN201710978303A CN107609960A CN 107609960 A CN107609960 A CN 107609960A CN 201710978303 A CN201710978303 A CN 201710978303A CN 107609960 A CN107609960 A CN 107609960A
- Authority
- CN
- China
- Prior art keywords
- recommended
- classification
- comment
- user
- recommendation
- 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)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a kind of rationale for the recommendation generation method and device, wherein method includes:Excavate the comment content of each object to be recommended;For the comment content of each object to be recommended, the categorical attribute of default dimension is established, comment classification prediction is carried out, obtains the comment sentence for belonging to each classification of each object to be recommended;According to the comment sentence for belonging to each classification of each object to be recommended, the rationale for the recommendation of object to be recommended is generated.Using the present invention program, the rationale for the recommendation using the comment content excavated from the comment of object to be recommended as the candidate of the object to be recommended, the diversity of rationale for the recommendation is added;The categorical attribute of the default dimension of foundation, chooses the comment content for belonging to various classification of object to be recommended as rationale for the recommendation, and then adds the specific aim of rationale for the recommendation, realizes the effect for attracting user and making user produce click or lower single act to the object to be recommended.
Description
Technical field
The present invention relates to field of computer technology, and in particular to a kind of rationale for the recommendation generation method and device.
Background technology
With the development of on-line off-line ecommerce (Online To Offline, abbreviation O2O) trade mode, how to make
User produces enough interest to the shop provided on O2O electric business transaction platforms, and guides user to produce click or lower single act
As the major issue of electric business platform development.In the prior art, the mode to solve the above problems mainly has two kinds:Mode one, is pressed
Urban area and shop classification count the sales volume in each shop, and using the ranking of sales volume list as rationale for the recommendation, for example, pushing away
It is the 1st, Xihu District chafing dish to recommend reason;Mode two, the footprint and the record of trading activity that user's history is browsed are managed as recommendation
By for example, rationale for the recommendation is you, in 3 days the browsed shop.
However, inventor is in implementing the present invention, it may, have found prior art, at least there are the following problems:Mode one is pressed
List is recommended so that each shop is fixed a kind of list, i.e.,:Urban area+classification list, it is difficult to extend to it
The list of his type, it is all the same so to result in the rationale for the recommendation that all users see, and for closely located
And similar purpose shop, homogeneity just occurs in the rationale for the recommendation of displaying, i.e., on the original list of shop, several continuous shops
Rationale for the recommendation it is identical;Mode two is recommended by user behavior, although which has very strong acceptance, Er Qieshi to user
Personalized displaying, but personally for, the part in shop, simply very little that user clicks on and merchandised, most of shop
It is no behavior, causes most of shop not show rationale for the recommendation, therefore coverage rate is relatively low.
The content of the invention
In view of the above problems, it is proposed that the present invention so as to provide one kind overcome above mentioned problem or at least in part solve on
State the rationale for the recommendation generation method and device, computing device, computer-readable storage medium of problem.
According to an aspect of the invention, there is provided a kind of rationale for the recommendation generation method, including:
Excavate the comment content of each object to be recommended;
For the comment content of each object to be recommended, the categorical attribute of default dimension is established, carries out comment classification prediction,
Obtain the comment sentence for belonging to each classification of each object to be recommended;
According to the comment sentence for belonging to each classification of each object to be recommended, the recommendation for generating object to be recommended is managed
By.
Alternatively, it is described obtain each object to be recommended belong to the comment sentence of each classification after, the side
Method also includes:
Store the first index relative of the object to be recommended and the comment sentence of each classification.
Alternatively, methods described also includes:
User action log data are excavated, obtain the user preference value that each user is directed to each classification, storage is used
Family and the second index relative of the user preference value of each classification.
Alternatively, the comment sentence for belonging to each classification of each object to be recommended of the basis, generation are to be recommended right
The rationale for the recommendation of elephant further comprises:
The object recommendation for carrying user's mark sent according to specific user is asked, and obtains recommended row to be presented
Table;
According to first index relative, inquire about in the recommended list to be presented corresponding to each object to be recommended
The comment sentence of each classification;According to second index relative, the user of each classification corresponding to the specific user is inquired about
Preference value;
The user preference value of each classification according to corresponding to the specific user, from each corresponding to each object to be recommended
At least one comment sentence, the rationale for the recommendation as the object to be recommended are chosen in the comment sentence of classification.
Alternatively, after the comment content for excavating each object to be recommended, methods described also includes:
Subordinate sentence processing is carried out to the comment content of each object to be recommended, obtains being directed to the comment of each object to be recommended
Sentence set.
Alternatively, it is described obtain being directed to the comment sentence set of each object to be recommended after, methods described is also wrapped
Include:
The emotion forecast model obtained using training, the comment sentence in comment sentence set to each object to be recommended
It is predicted, filters out the comment sentence in the comment sentence set with emotion negative sense attribute.
Alternatively, the categorical attribute for establishing default dimension, comment classification prediction is carried out, obtains each object to be recommended
The comment sentence for belonging to each classification further comprise:
The comment disaggregated model obtained using training, the comment sentence in comment sentence set to each object to be recommended
The probability for belonging to each classification is predicted, and determines to adhere to separately in the comment sentence set of each object to be recommended according to prediction result
In the comment sentence of each classification.
Alternatively, the comment sentence for belonging to each classification for obtaining each object to be recommended further comprises:
Train to obtain text generation using the comment sentence for belonging to each classification of each object to be recommended as training sample
Model, the new comment sentence of each classification is belonged to using the generation of text generation model;
Point of each object to be recommended is obtained using as the comment sentence of training sample and new comment sentence common combination
Belong to the comment sentence of each classification.
Alternatively, it is described to be trained using the comment sentence for belonging to each classification of each object to be recommended as training sample
To text generation model, the new comment sentence that each classification is belonged to using the generation of text generation model is further comprised:
For each classification, by the comment sentence for belonging to the classification of the object to be recommended with same object classification
As object classification and the training sample set of the classification;
Train to obtain corresponding text generation model for each training sample set;
Belong to the new comment sentence of the classification using the generation of text generation model;
Obtain that there is same object classification using as the comment sentence of training sample and new comment sentence common combination
The comment sentence for belonging to the classification of object to be recommended.
Alternatively, the new comment sentence that each classification is belonged to using the generation of text generation model is further wrapped
Include:
For each classification, every popular head-word for commenting on sentence under the classification is obtained, as source phrase Candidate Set;
Source phrase is chosen from the source phrase Candidate Set and obtains the input data of text generation model, input data is defeated
Enter into text generation model, output obtains object phrase;
Source phrase and object phrase are combined to the input data for obtaining the text generation model again, number will be inputted
According to input into text generation model, output again obtains new object phrase, by that analogy, until obtaining meeting preset length
It is required that source phrase and at least one object phrase combination as new comment sentence.
Alternatively, it is described obtain the input data of text generation model before, methods described also includes:If the source is short
The phrase sequence length of the combination of language and/or source phrase and object phrase is less than fixed length, then to the source phrase and/or source phrase
Combination with object phrase carries out mending long processing, so as to get the phrase sequence length of input data be fixed length.
Alternatively, the user preference value of the classification each according to corresponding to specific user, from each object pair to be recommended
At least one comment sentence is chosen in the comment sentence for each classification answered, the rationale for the recommendation as the object to be recommended is further
Including:
The order of the user preference value of each classification from high to low is arranged each classification according to corresponding to specific user
Sequence;
According to the ranking results of each classification, successively from the comment sentence of each classification corresponding to each object to be recommended
Choose at least one comment sentence;
Rationale for the recommendation of the comment sentence as object to be recommended is randomly selected from least one comment sentence.
Alternatively, the categorical attribute of the default dimension is specially the categorical attribute by theme dimension, by theme dimension
Each categorical attribute includes one or more of following categorical attribute:
Service, environment, cost performance, taste, quality, purpose and/or storekeeper.
According to another aspect of the present invention, there is provided a kind of rationale for the recommendation generating means, including:
Module is excavated, suitable for excavating the comment content of each object to be recommended;
Sort module, suitable for the comment content for each object to be recommended, the categorical attribute of default dimension is established, is carried out
Comment classification prediction, obtains the comment sentence for belonging to each classification of each object to be recommended;
Matching module, it is to be recommended suitable for the comment sentence for belonging to each classification according to each object to be recommended, generation
The rationale for the recommendation of object.
Alternatively, described device also includes:Memory module, suitable for storing the object to be recommended and the comment of each classification
First index relative of sentence.
Alternatively, the excavation module is further adapted for:User action log data are excavated, each user is obtained and is directed to institute
State the user preference value of each classification;
The memory module is further adapted for:Store user and the second index relative of the user preference value of each classification.
Alternatively, the matching module is further adapted for:
The object recommendation for carrying user's mark sent according to specific user is asked, and obtains recommended row to be presented
Table;
According to first index relative, inquire about in the recommended list to be presented corresponding to each object to be recommended
The comment sentence of each classification;According to second index relative, the user of each classification corresponding to the specific user is inquired about
Preference value;
The user preference value of each classification according to corresponding to the specific user, from each corresponding to each object to be recommended
At least one comment sentence, the rationale for the recommendation as the object to be recommended are chosen in the comment sentence of classification.
Alternatively, the sort module is further adapted for:
The comment disaggregated model obtained using training, the comment sentence in comment sentence set to each object to be recommended
The probability for belonging to each classification is predicted, and determines to adhere to separately in the comment sentence set of each object to be recommended according to prediction result
In the comment sentence of each classification.
Alternatively, described device also includes:Generation module, suitable for each classification that belongs to of each object to be recommended
Comment sentence is that training sample trains to obtain text generation model, belongs to the new of each classification using the generation of text generation model
Comment sentence;
The sort module is further adapted for:Using as the comment sentence of training sample and new comment sentence common combination
Obtain the comment sentence for belonging to each classification of each object to be recommended.
Alternatively, the generation module is further adapted for:For each classification, will wait to push away with same object classification
The comment sentence for belonging to the classification of object is recommended as object classification and the training sample set of the classification;
Train to obtain corresponding text generation model for each training sample set;
Belong to the new comment sentence of the classification using the generation of text generation model;
The sort module is further adapted for:Using as the comment sentence of training sample and new comment sentence common combination
The comment sentence for belonging to the classification for the object to be recommended for obtaining that there is same object classification.
Alternatively, the generation module is further adapted for:For each classification, every comments under the classification are obtained
The popular head-word of sentence, as source phrase Candidate Set;
Source phrase is chosen from the source phrase Candidate Set and obtains the input data of text generation model, input data is defeated
Enter into text generation model, output obtains object phrase;
Source phrase and object phrase are combined to the input data for obtaining the text generation model again, number will be inputted
According to input into text generation model, output again obtains new object phrase, by that analogy, until obtaining meeting preset length
It is required that source phrase and at least one object phrase combination as new comment sentence.
Alternatively, the generation module is further adapted for:If the combination of the source phrase and/or source phrase and object phrase
Phrase sequence length be less than fixed length, then the combination to the source phrase and/or source phrase and object phrase carries out mending long processing,
The phrase sequence length for the input data for making to obtain is fixed length.
Alternatively, the matching module is further adapted for:
The order of the user preference value of each classification from high to low is arranged each classification according to corresponding to specific user
Sequence;
According to the ranking results of each classification, successively from the comment sentence of each classification corresponding to each object to be recommended
Choose at least one comment sentence;
Rationale for the recommendation of the comment sentence as object to be recommended is randomly selected from least one comment sentence.
According to another aspect of the invention, there is provided a kind of computing device, including:Processor, memory, communication interface and
Communication bus, processor, memory and communication interface complete mutual communication by communication bus;
Memory is used to deposit an at least executable instruction, and executable instruction makes the above-mentioned rationale for the recommendation generation of computing device
Operated corresponding to method.
In accordance with a further aspect of the present invention, there is provided a kind of computer-readable storage medium, be stored with least one in storage medium
Executable instruction, executable instruction make computing device be operated as corresponding to above-mentioned rationale for the recommendation generation method.
The rationale for the recommendation generation method provided according to the present embodiment, first by excavating the comment content of object to be recommended,
Rationale for the recommendation using the comment content of these objects to be recommended as the candidate of the object to be recommended, due to the comment content of user
It can vary with each individual, and it is rich and varied, therefore the rationale for the recommendation of the candidate excavated by this way can greatly increase recommendation
The diversity of reason;For the comment content of each object to be recommended, the categorical attribute of default dimension is established, carries out comment classification
Prediction, the comment sentence for belonging to each classification of each object to be recommended is obtained, and then every of object to be recommended is commented on
Content is corresponded in the categorical attribute of the object to be recommended, so that can be according to be recommended when recommending the object to be recommended
The different classifications attribute of object is recommended so that rationale for the recommendation is more targeted, and recommends that difference can be shown every time
The rationale for the recommendation of categorical attribute;It is to be recommended right according to the comment sentence for belonging to each classification of each object to be recommended, generation
The rationale for the recommendation of elephant, to realize the advantageous information for the specific classification attribute that the recommended is shown to user, and then attract user
Click or lower single act are produced to the object to be recommended.The rationale for the recommendation generation method provided using the present embodiment, by increasing capacitance it is possible to increase
The diversity of rationale for the recommendation, and the comment content of the particular community feature of object to be recommended can be chosen as rationale for the recommendation,
And then the specific aim of rationale for the recommendation is added, attract user to produce click or lower single act to the object to be recommended to realize, carry
The order volume and conversion ratio of peaceful.
Described above is only the general introduction of technical solution of the present invention, in order to better understand the technological means of the present invention,
And can be practiced according to the content of specification, and in order to allow above and other objects of the present invention, feature and advantage can
Become apparent, below especially exemplified by the embodiment of the present invention.
Brief description of the drawings
By reading the detailed description of hereafter preferred embodiment, it is various other the advantages of and benefit it is common for this area
Technical staff will be clear understanding.Accompanying drawing is only used for showing the purpose of preferred embodiment, and is not considered as to the present invention
Limitation.And in whole accompanying drawing, identical part is denoted by the same reference numerals.In the accompanying drawings:
Fig. 1 shows user's global behavior flow chart when being done shopping on O2O electric business transaction platforms;
Fig. 2 shows the flow chart of the rationale for the recommendation generation method of one embodiment of the invention;
Fig. 3 shows the recommending data flow chart of rationale for the recommendation in the rationale for the recommendation generation method of the embodiment of the present invention;
Fig. 4 shows the flow chart of the rationale for the recommendation generation method of another embodiment of the present invention;
Fig. 5 shows the flow chart of the rationale for the recommendation generation method of another embodiment of the invention
Fig. 6, which is shown, utilizes the method flow of the new comment sentence of text generation formula model generation in the embodiment of the present invention
Figure;
Fig. 7 shows the stream of the method for the rationale for the recommendation of the selection recommended to be presented of one specific embodiment of the present invention
Cheng Tu;
Fig. 8 shows the functional block diagram of the rationale for the recommendation generating means of one embodiment of the invention;
Fig. 9 shows the functional block diagram of the rationale for the recommendation generating means of another embodiment of the present invention;
Figure 10 shows a kind of structural representation of the structural representation of computing device of the embodiment of the present invention.
Embodiment
The exemplary embodiment of the disclosure is more fully described below with reference to accompanying drawings.Although the disclosure is shown in accompanying drawing
Exemplary embodiment, it being understood, however, that may be realized in various forms the disclosure without should be by embodiments set forth here
Limited.On the contrary, these embodiments are provided to facilitate a more thoroughly understanding of the present invention, and can be by the scope of the present disclosure
Completely it is communicated to those skilled in the art.
Fig. 1 shows user's global behavior flow chart when being done shopping on O2O electric business transaction platforms.As shown in figure 1, with
After family logs in, by searching for related shop, and/or the shop of platform push is browsed, and the recommendation shown according to each shop
Reason determines to have a mind to the shop of consumption, click on the shop, understands shop details, and place an order purchase, wherein, rationale for the recommendation refers to
O2O shop search listing or the shop list area of the recommendation list page, the text shown is clicked on and merchandised for guiding user
This content, mark is what reason to recommend this shop for user by, usually in the short sentence text on shop forward direction favorable comment
Hold, such as:" vegetable is fresh, the awfully hot feelings of waiter " or " chafing dish city list first place ";After merchandising successfully, disappear for producing
The shop of expense behavior deliver shop comment, wherein, shop comment comment form include text, picture, popularity value, star and
Comment on label etc..The comment content of shop comment reflects user to the impression and suggestion in process of exchange, these comment contents
There are guiding and reference value well, especially positive favorable comment content user can be guided to click on immediately the user of potential transaction
Or place an order, lift the order volume and conversion ratio of platform.
Accordingly, it is considered to it can intuitively reflect store information and shop for the comment content that shop is filled in user
Quality information, comment content can vary with each individual, rich and varied, also, in order to carry out shop search and O2O electric business in user
Platform carries out in the scenes such as shop recommendation rationale for the recommendation being utilized to go the concern for attracting user to the shop of displaying, and promotes to use
Family produce click on or place an order wait behavior, and then lifting platform order volume and conversion ratio, the invention provides a kind of rationale for the recommendation
Generation method and device.
Fig. 2 shows the flow chart of the rationale for the recommendation generation method of one embodiment of the invention.As shown in Fig. 2 this method
Comprise the following steps:
Step S201, excavate the comment content of each object to be recommended.
Wherein, object to be recommended is included in that user scans for and O2O electric business platform carries out recommending etc. to recommend in scene
To objects such as the shop of user, sight spot, vegetable, masseur, photographers, in other words, object to be recommended can be O2O electric business platforms
Any form of object of user can be showed;The comment content of object to be recommended refers to:User is after consumption, for shop
The comment that the objects such as paving, sight spot, vegetable, masseur, photographer are delivered.In following examples, by the shop in O2O electric business platforms
As object to be recommended.
Specifically, for each object to be recommended, the comment content of object to be recommended is excavated from data source, wherein, number
It is the comment for object to be recommended that user delivers under the object to be recommended according to source, for example, being commented for what snack bar a was delivered
By:" snack bar waiter's attitude is fine, awfully hot feelings.It is and competitively priced cheap ";Also, excavate the shape of obtained comment content
Formula includes text, picture, popularity value, star and comment label, the rationale for the recommendation of the object to be recommended due to showing user
For content of text, therefore, when excavate obtained comment content for image content, popularity value, star content and comment label when,
Need that the form for commenting on content first is converted into content of text, that is, the comment sentence of textual form is converted into, for example, taste is commented into 5
The star content transformation of star is the content of text of " taste very rod ".
By this step, the comment sentence for each textual form of object to be recommended is finally given, i.e.,:For each
The comment content of object to be recommended.For example, 1000 comment contents for shop a are obtained, for sight spot b 500 comments
Content, content is commented on for 800 of photographer c.
Step S202, for the comment content of each object to be recommended, the categorical attribute of default dimension is established, is commented on
Classification prediction, obtains the comment sentence for belonging to each classification of each object to be recommended.
Because for same object to be recommended, different Consumer's Experiences is different, and comment is also just different, accordingly, comments on
Dimension also can be different, in the present embodiment, in order that repeatedly recommending the rationale for the recommendation energy of the same object to be recommended of user
The different attribute feature of the object to be recommended is enough embodied, establishes the categorical attribute of default dimension, comment classification prediction is carried out, obtains
The comment sentence for belonging to each classification of each object to be recommended.Wherein, default dimension includes but is not limited to personalized classification
Dimension, user often go ground dimension or theme dimension, and the categorical attribute of default dimension refers to object to be recommended to be had under default dimension
Some attributive character, for example, when default dimension is the theme dimension, shop a attributive character includes following at least one:Clothes
Business, environment, cost performance, taste, quality, purpose and storekeeper;And for example, photographer c attributive character includes following at least one:
Image, attitude, technology and service.
Accordingly, according to the attributive character of object to be recommended, i.e.,:The categorical attribute of default dimension, it is pre- to carry out comment classification
Survey.Specifically, each bar comment sentence of the object to be recommended to excavating carries out classification prediction by the categorical attribute of default dimension,
Predict the attributive character of the object to be recommended of each bar comment sentence comment, for example, default dimension is the theme during dimension, shop
Comment sentence " vegetable deal is big, competitively priced " corresponding to a is exactly the comment to shop a cost performance.It is emphasized that this
Embodiment is not defined to the method for classification prediction, and the method for every existing classification prediction that can realize comment sentence is all recognized
For in the scope of protection of the invention.
By this step, after the comment sentence for treating recommended carries out classification prediction, each object to be recommended is obtained
The comment sentence for belonging to each classification.
Above-mentioned steps S201 and step S202 is the mining process of rationale for the recommendation, and it is only the generation stream to rationale for the recommendation
Journey is prepared, and above-mentioned two step lower can be completed online, i.e., not against network, and does not influence the real-time deal of user.And
Following step S203 is then the process of the comment sentence generation rationale for the recommendation obtained according to excavating, and the step needs to hold on line
OK.
Step S203, according to the comment sentence for belonging to each classification of each object to be recommended, generate object to be recommended
Rationale for the recommendation.
When O2O e-commerce platforms need to recommend object to be recommended to user, O2O e-commerce platforms are for different
Scene, recommend corresponding object to be recommended to user.For example, in the scene that user carries out the search such as shop, service, O2O electricity
Sub- business platform recommends shop or the technician for meeting user's search condition to user;And for example, O2O ecommerce is opened in user to put down
In the scene of platform, O2O e-commerce platforms recommend the to be recommended of homepage displaying according to the recommendation rules that server is formulated to user
Object.
At the same time, in order to attract user to treat the concern of recommended, and user is promoted to treat recommended generation
Click on or lower single act, in the present embodiment, for each object generation to be recommended to should object rationale for the recommendation.Specifically, treat
The rationale for the recommendation of recommended is chosen from the comment sentence of the object to be recommended, further, to make to wait to push away for same
Recommending rationale for the recommendation that object shows can be more rich, in the interface of different users, and/or the multiple recommendation in same user
In, it can be selected from the comment sentence for belonging to different classifications of the object to be recommended, or, from the object to be recommended
Same classification under comment sentence in select different comment sentences to be shown.
Specifically, the mode that selection is carried out from different classification can be so that show object to be recommended when recommending every time
Different attribute feature, for example, being directed to same user, when recommending shop a for the first time, the comment sentence that selection belongs to Service Properties is made
For rationale for the recommendation, when recommending shop a for the second time, selection belongs to the comment sentence of environment attribute as rationale for the recommendation, it can be seen that,
Which can avoid repeatedly showing the comment sentence of same attributive character;Different comment sentences is selected to make from same classification
It is shown for rationale for the recommendation, can targetedly selects the comment sentence of a certain attributive character as rationale for the recommendation so that
The purpose for recommending the object to be recommended is clearer and more definite.In addition, when sentence is commented in selection, can randomly choose, can also be according to spy
Fixed rule is selected, and is not especially limited herein.
By this step, recommendation of the comment sentence for the object to be recommended classified by prediction as object to be recommended is chosen
Reason shows user, and the advantageous information of the particular community feature of the recommended is shown to user, and then attracts user to this
Object to be recommended produces click or lower single act.
The rationale for the recommendation generation method provided according to the present embodiment, digging flow is excavated by offline rationale for the recommendation first
The comment content of object to be recommended, managed the comment content of these objects to be recommended as the recommendation of the candidate of the object to be recommended
By, because the comment content of user can vary with each individual, and it is rich and varied, therefore the recommendation reason of the candidate excavated by this way
By the diversity that can greatly increase rationale for the recommendation;For the comment content of each object to be recommended, default dimension is established
Categorical attribute, comment classification prediction is carried out, obtain the comment sentence for belonging to each classification of each object to be recommended, and then will
Every comment content of object to be recommended is corresponded in the attributive character of the object to be recommended, so as to be carried out to the object to be recommended
It can be recommended during recommendation according to the different attribute feature of object to be recommended so that rationale for the recommendation is more targeted, and
Recommendation can show the rationale for the recommendation of different attribute feature every time;Commenting for each classification is belonged to according to each object to be recommended
The Analects of Confucius sentence, generate the rationale for the recommendation of object to be recommended, with realize to user show the recommended particular community feature it is excellent
Gesture information, and then attract user to produce click or lower single act to the object to be recommended.The recommendation provided using the present embodiment is managed
By generation method, by increasing capacitance it is possible to increase the diversity of rationale for the recommendation, and commenting for the particular community feature of object to be recommended can be chosen
By content as rationale for the recommendation, and then the specific aim of rationale for the recommendation is added, attract user to produce the object to be recommended to realize
Raw click or lower single act, lift the order volume and conversion ratio of platform.
On the other hand, on the basis of rationale for the recommendation generation method embodiment corresponding to the flow chart shown in Fig. 2, it is contemplated that
The preference of different user is different, i.e.,:The focus that different user treats recommended is different, for example, some users pay close attention to shop
The quality of commodity, other users pay close attention to the service that shop provides.And the preference of user can be embodied in user journal, wherein,
The behavioral datas such as the search for having user on O2O electric business platforms, click, purchase, collection are recorded in user journal, it can be seen that,
The feature and preference of user can be excavated using these behavioral datas in user journal.
To sum up, it is contemplated that user can intuitively reflect object to be recommended for the comment content that object to be recommended is filled in
The quality information of information and object to be recommended, and comment on content and can vary with each individual, it is rich and varied, and in view of different use
The preference at family is different.Therefore, in order to be scanned in user and O2O electric business platform recommend etc. in scene utilizing and pushed away
Recommend reason go attract user treat the concern of recommended, and promote user produce click on or place an order wait behavior, and then lifted put down
The order volume and conversion ratio of platform, the invention provides the rationale for the recommendation generation method of following embodiments.
Fig. 3 shows the recommending data flow chart of rationale for the recommendation in the rationale for the recommendation generation method of the embodiment of the present invention.Such as
Shown in Fig. 3, the recommending data flow of the rationale for the recommendation in following embodiments includes offline rationale for the recommendation excavation and pushed away with online
Recommend reason and match two main flows.
Wherein, rationale for the recommendation excavates flow and lower online can completed, i.e., the real-time of user is influenceed not against network and not
Transaction, it includes two sub-processes again:Sub-process one, carry out commenting on content mining and comments for each object to be recommended
Sentence generation, obtains the comment sentence that each object to be recommended corresponds to different classifications attribute, establishes object to be recommended and each classification
Comment sentence object reviews index to be recommended, in following embodiments of the corresponding present invention of the object reviews index to be recommended
First index relative;The row such as sub-process two, the user in user journal browses, user's click, user's collection, customer transaction
For data, preference of the user to each classification is calculated, and establishes user and the user preference index of the preference of each classification, the use
The second index relative in the corresponding following embodiments of the present invention of family preference index.
Rationale for the recommendation matching flow is completed on line, i.e., is matched simultaneously when O2O platforms show object to be recommended to user
Rationale for the recommendation, i.e.,:According to offline part utilize user journal establish user preference search index user preference, and according to from
The object reviews index to be recommended inquiry comment sentence that line part is established, it is final to be arrived according to the user preference matching inquiry inquired
Object to be recommended comment sentence, obtain meeting the comment sentence of user preference using the rationale for the recommendation exhibition as object to be recommended
Show to user.
Flow is excavated by above-mentioned offline rationale for the recommendation and online rationale for the recommendation match flow can will be to be recommended
The comment content excavated in object comment matches with user preference, and shows user as the rationale for the recommendation of object to be recommended,
And then user is established to the rationale for the recommendation of object to be recommended personalized rationale for the recommendation, the i.e. face of thousand people thousand.
Fig. 4 shows the flow chart of the rationale for the recommendation generation method of another embodiment of the present invention.As shown in figure 4, the party
Method comprises the following steps:
Step S401, the comment content of object to be recommended is excavated, point commented on according to the categorical attribute of default dimension
Class is predicted, and establishes the first index relative of object to be recommended and the comment sentence of each classification.
In this step, the comment content of each object to be recommended is excavated, for the comment content of each object to be recommended, root
According to the categorical attribute of default dimension, comment classification prediction is carried out, obtain each object to be recommended belongs to commenting for each classification
The Analects of Confucius sentence, store the first index relative of object to be recommended and the comment sentence of each classification.Specifically, for each to be recommended
Object, the comment content of object to be recommended is excavated from data source, wherein, data source is, for example, the pin that user delivers under a of shop
Comment to shop a, for example, " snack bar waiter's attitude is fine, awfully hot feelings.It is and competitively priced cheap ".Specifically, excavate
The form of obtained comment content includes text, picture, popularity value, star and comment label, due to showing treating for user
The rationale for the recommendation of recommended is content of text, therefore, when excavate obtained comment content for image content, star content and
When commenting on label, need that the form for commenting on content first is converted into content of text, that is, be converted into the comment sentence of textual form, example
Such as, taste is commented into the content of text that the star content transformation of 5 stars is " taste very rod ".
Because for same object to be recommended, different Consumer's Experiences is different, and comment is also just different, accordingly, comments on
Dimension also can be different, and in order to match the rationale for the recommendation of corresponding classification according to user preference, in the present embodiment, by default dimension
The categorical attribute of degree is classified to the comment sentence of each object to be recommended of excavation, you can with according to description object to be recommended
Attributive character rationale for the recommendation is classified, wherein, preset dimension categorical attribute include by personalized classification dimension carry out
Classification, ground dimension is often gone by user to be classified or classified by theme dimension, specifically, to be tieed up by theme in the present embodiment
The categorical attribute of degree is classified to illustrate to comment sentence.
In one particular embodiment of the present invention, the attributive character of object to be recommended on O2O electric business transaction platforms is obtained,
It is divided into following 7 attribute feature:Service, environment, cost performance, taste, quality, purpose, storekeeper, most of object comment to be recommended
In comment sentence evaluated both at least a kind of in this 7 class.Such as:" vegetable deal is big, competitively priced " is exactly
Treat the comment of the cost performance of recommended.
After to comment sentence carries out classification prediction corresponding to the comment content of the object to be recommended excavated, belonged to
The comment sentence of each classification of object to be recommended, by object to be recommended and the comment sentence pair of each classification of object to be recommended
Should store, such as the comment sentence for belonging to classification of service is stored under classification of service, with establish object to be recommended with it is to be recommended
First index relative of the comment sentence of each classification of object.So that object to be recommended is shop as an example, corresponding different shop marks
Know shop_id, storage respectively belongs to the comment sentence of service, environment, cost performance, taste, quality, purpose and storekeeper's classification,
Accordingly, in the first index relative, the comment sentence of each classification in shop is specially when showing the shop, for corresponding point
The rationale for the recommendation of candidate when class is recommended.
Step S402, User action log data are excavated, obtain the user preference value that each user is directed to each classification, deposit
Store up user and the second index relative of the user preference value of each classification.
Wherein, each classification under the i.e. default dimension of each classification, for example, the classification of service under theme dimension, environment point
Class etc..Specifically, the user behaviors log data of user include the search behavior data, navigation patterns data, click behavior number of user
It can reflect the focus and point of interest of user, example according to, splitting glass opaque data and lower single behavioral data, these behavioral datas
Such as, in lower single behavioral data of user, what the user selected consumption is the high-grade dining room of environment, then reflects the user to shop
The environment of paving is valued very much, thus by above-mentioned behavioral data analysis calculate can obtain user for each classification user it is inclined
Good value.
Specifically, analyzing the method for calculating includes the decision Tree algorithms (Gradient of machine learning method, statistic law or iteration
Boosting Decision Tree, abbreviation GBDT).
By to the behavioral data in user journal carry out analysis be calculated user for each classification user it is inclined
After good value, by user and user for each classification user preference value it is corresponding store, to establish user and each classification
Second index relative of user preference value, for example, corresponding different user identifies user_id, store respectively the user to service,
Environment, cost performance, taste, quality, purpose and the user preference value of storekeeper's classification.In second index relative, user is for each
The user preference value of individual classification is preference of the user to each classification, according to the user preference value in second index relative
The rationale for the recommendation that the user reflected carries out object to be recommended to the preference of each classification matches.
Above-mentioned steps S401 and step S402 corresponds to offline rationale for the recommendation and excavates flow, above-mentioned two step, only
It is that matching flow to rationale for the recommendation is prepared, step S403 to step S406 below is then according to the first index relative and
Two index relatives carry out the flow of the On-line matching of rationale for the recommendation.
Step S403, the object recommendation for carrying user's mark sent according to specific user are asked, and obtain to be presented push away
Recommend list object.
Wherein, object recommendation request includes but is not limited to:User opens the request of O2O e-commerce platforms, and user's point is opened
The request in specific special column in O2O e-commerce platforms, and the request of user's search.Asked in response to object recommendation, O2O electronics
Business platform is both needed to determine recommended list to be presented, specifically, for user open O2O e-commerce platforms request or
User's point opens the request in specific special column in O2O e-commerce platforms, need to determine the recommended to be presented row in homepage or special column
Table;For the request of user's search, need to determine to match the recommended list to be presented of search result.In the present embodiment, wait to open up
Show that recommended list can determine according to achievable mode of the prior art, be not specifically limited herein.For example, according to O2O
The recommendation rules of e-commerce platform server determine.
By the step, determine that the user of the specific user of sending object recommendation request identifies user_id, and according to object
Recommendation request obtains recommended list to be presented, wherein, each object to be recommended in recommended list to be presented
There is object identity to be recommended corresponding to one, for example, there is corresponding shop_id in shop to be presented..
Step S404, inquire about the comment sentence that recommended to be presented corresponds to each classification, and inquiry specific user couple
Answer the preference value of each classification.
Specifically, according to the first index relative, inquire about in recommended list to be presented corresponding to each object to be recommended
The comment sentence of each classification;According to the second index relative, the user preference value of each classification corresponding to inquiry specific user.Example
Such as, the comment sentence that each shop corresponds to each classification is inquired about in the first index relative according to store identification shop_id, that is, is waited
The rationale for the recommendation of choosing;According to the user of specific user identify user_id inquired about in the second index relative the user correspond to it is each
The user preference value of classification, in order to according to user preference value determine show the rationale for the recommendation of the user belonging to classification.
Step S405, the recommendation that the preference value selection recommended to be presented of each classification is corresponded to according to specific user are managed
By.
The user preference value of each classification according to corresponding to specific user, from each classification corresponding to each object to be recommended
Comment sentence in choose at least one comment sentence, the rationale for the recommendation as the object to be recommended.Specifically, according to specific use
The preference value that family corresponds to each classification determines to show the classification belonging to the rationale for the recommendation of the specific user;From recommendation pair to be presented
Rationale for the recommendation is chosen in classification as belonging to middle correspondence shows the rationale for the recommendation of the specific user.Further, show
The rationale for the recommendation of user is the comment that one classification of user preference value highest is corresponded in each classification of recommended to be presented
Sentence, or the comment sentence of at least two classification for corresponding user preference value from high to low;Also, show pushing away for user
Reason is recommended as a comment sentence under corresponding classification, or is at least two comment sentences under corresponding classification.
Step S406, the combination of the rationale for the recommendation of recommended list to be presented and each object to be recommended is sent to spy
Determine user.
By each object to be recommended in recommended list to be presented with to should object to be recommended rationale for the recommendation
Combination, and the combination is showed into specific user to reach the purpose for responding the object recommendation of the specific user and asking.
The flow of above-mentioned steps S403 to step S406 rationale for the recommendation On-line matching, directly established using offline part
First index relative and the second index relative can be achieved with the rationale for the recommendation and the user preference phase of the user that will show user
Matching, and then make it that the flow of On-line matching is simpler, the speed of matching is rapider, thus reduces response object recommendation
The time of request.
The rationale for the recommendation generation method provided according to the present embodiment, digging flow is excavated by offline rationale for the recommendation first
The comment content of object to be recommended, and classification prediction is carried out to comment content to establish the comment of object to be recommended and each classification
First index relative of sentence, by first index relative, can quick search correspond to each point to each object to be recommended
The rationale for the recommendation of the candidate of class, and the comment of the object to be recommended to be excavated adds the various of reason as rationale for the recommendation
Property, solve the problems, such as that rationale for the recommendation is single in the prior art;And User action log data are excavated, and to user behavior day
Will data carry out analysis calculating, obtain the user preference value that user is directed to each classification, and the user preference value has reflected user
To the degree of concern of each classification, and establish user and the second index relative of the user preference value of each classification;Again by
The object recommendation request of the rationale for the recommendation matching process response user of line, user's mark is carried with specific reference to what specific user sent
The object recommendation request of knowledge, obtains recommended list to be presented, in the second index relative, inquires about each corresponding to specific user
The user preference value of individual classification in the first index relative, inquires about recommendation to be presented to determine classification that the user most pays close attention to
The comment sentence for the classification that the user corresponding to each object to be recommended most pays close attention in list object, you can will be each to be recommended right
The preference of rationale for the recommendation and user during as displaying is to matching so that rationale for the recommendation has personalization, can farthest inhale
Family is quoted, and realizes the bandwagon effect of the rationale for the recommendation in the face of thousand people thousand.
Fig. 5 shows the flow chart of the rationale for the recommendation generation method of another embodiment of the invention.In the present embodiment, preset
The categorical attribute of dimension is specially the categorical attribute by theme dimension, as shown in figure 5, this method comprises the following steps:
Step S501, excavate the comment content of each object to be recommended.
For each object to be recommended, the comment content of object to be recommended is excavated from data source, wherein, data source is should
The comment that user delivers under object to be recommended, for example, " snack bar waiter's attitude is fine, awfully hot feelings.And it is competitively priced just
Preferably ".Specifically, the form for the comment content for excavating to obtain includes text, picture, star and comment label, due to showing
The rationale for the recommendation of the object to be recommended of user is content of text, therefore, when the comment content that excavation obtains is image content, star
When content and comment label, need that the form for commenting on content first is converted into content of text, that is, be converted into the comment of textual form
Sentence, for example, taste is commented into the content of text that the star content transformation of 5 stars is " taste very rod ".By the step, can obtain
To the comment sentence of the object to be recommended comment content of corresponding each object to be recommended.
Step S502, filtration treatment is carried out to the comment content of each object to be recommended.
Commented on due to object to be recommended in content and some low-quality comments be present, these low-quality comments are not suitable for making
User is showed for rationale for the recommendation, therefore, as an optional step of the present embodiment, is excavating each object to be recommended
Comment content after, filtration treatment is carried out to the comment content of each object to be recommended, filters low-quality comment content.Tool
Body, filtration treatment is carried out according to the information quality point of comment content, wherein, the evaluation factor for calculating information quality timesharing has:
Whether there is mess code in comment content, whole section is commented on whether content has punctuation mark, and whether the words and phrases for commenting on content are enumerated and piled up,
Whether the syntax for commenting on content clear and coherent, if containing semantic focus word, comment content whether and phase related to object to be recommended
Closing property height.By being scored for above-mentioned reference factor comment content, the information quality point of every comment content is obtained,
And filter out comment content of the information quality point less than preset quality point threshold value.
In one particular embodiment of the present invention, the information quality point of comment content is calculated using equation below (1):
In formula (1), comment_score is the information quality point of comment content, and n is the number of evaluation factor, wkTo comment
Divide the weight of factor, scorekThe quality point of evaluation factor is corresponded to for comment content, and the quality point more than or equal to 0 and is less than
Or equal to 1;Also,And then it ensure that the end value comment_score finally calculated is also greater than or equal to 0
And less than or equal to 1.
Step S503, subordinate sentence processing is carried out to the comment content of each object to be recommended, obtain being directed to each to be recommended
The comment sentence set of object.
Because the rationale for the recommendation in each object to be recommended shown in O2O e-commerce platforms is a simple sentence
Son, and the object to be recommended comment that user delivers is one section of content of text, therefore subordinate sentence is carried out to text, so that the exhibition of platform
Show that interface is more concise.Specifically, according to the limitation of punctuation mark and/or default sentence length to each object to be recommended
Comment on content of text corresponding to content and carry out subordinate sentence processing, so that division of teaching contents will be commented on as semantic complete short sentence.Wherein, punctuate
Symbol includes fullstop, exclamation mark, and branch etc..
In a specific embodiment, the comment content of excavation is subjected to subordinate sentence by punctuation mark, further determines that subordinate sentence
Whether the length of comment sentence afterwards is in the limitation of default sentence length, if so, subordinate sentence is disposed;If it is not, then by length
The comment sentence in the limitation of default sentence length does not press default sentence length and carries out secondary subordinate sentence processing, specifically, to protect
Card comments on the complete of statement semantics, and its length have to be larger than or the minimum length equal to default sentence length, meanwhile, short sentence length
The maximum length of default sentence length is should be less than, accordingly, secondary subordinate sentence processing includes:It is more than or equal to default sentence for length
The comment sentence of the maximum length of sub- length, the comment sentence less than maximum length is marked off from left to right, and remove length
Less than the comment sentence of the minimum length of default sentence length.
After being handled by subordinate sentence, each comment content is all divided into multiple comment sentences, and then, treated for each
The a plurality of comment content of recommended, it becomes possible to obtain a comment sentence set.
Step S504, the emotion forecast model obtained using training, in the comment sentence set to each object to be recommended
Comment sentence be predicted, filter out comment sentence set in have emotion negative sense attribute comment sentence.
Must be the advantageous properties information of object to be recommended, i.e. rationale for the recommendation table as the rationale for the recommendation for showing user
The emotion reached must be positive favorable comment, it is therefore desirable to which filtering out has emotion negative sense in the comment sentence set of each object to be recommended
The comment sentence that the comment sentence of attribute, i.e. negative sense difference are commented.Specifically, the emotion of comment sentence expression is analyzed, according to
Comment sentence of the analysis result prediction with emotion negative sense attribute, and filter out the comment sentence.Further, using training
To emotion forecast model to comment on sentence emotion attribute carry out analysis prediction, wherein emotion attribute includes emotion forward direction attribute
With emotion negative sense attribute.
In one particular embodiment of the present invention, emotion forecast model include but is not limited in following disaggregated model one
Kind:FastText classifies, maximum entropy Maxent classification, Logistic Regression classification and GBDT classification.
By taking the FastText classification in above-mentioned disaggregated model as an example, the classification of emotion attribute is carried out to comment sentence, specifically
Algorithm include model training and prediction two steps:
Step 1, model training.The historical review content of some users is chosen, and every comment content is carried out at subordinate sentence
Reason is so that as training sample, optionally, number of samples is more than 100,000;One kind in the following manner is determined in every comment
The emotion attribute of appearance, mode one, the emotion attribute of every comment content is manually marked out, i.e., artificial judgment every, which comments on content, is
With emotion forward direction attribute still with emotion negative sense attribute.Mode two, the comment is determined according to star corresponding to comment content
The emotion attribute of content, such as the comment content that will be above 3 stars will be less than or are equal to as the sample with emotion forward direction attribute
The comment content of 3 stars is as the sample with emotion negative sense attribute;The training of training sample and each training sample is being determined
After target, the feature of each training sample is further determined that, in order to be trained to obtain corresponding target according to these features,
Wherein, training objective refers to emotion attribute, and emotion forward direction attribute is designated as 1, and emotion negative sense attribute is designated as 0;Wherein, the spy of training sample
Sign refers to every after subordinate sentence is handled comment content carrying out word segmentation processing again, obtain with it is every after subordinate sentence is handled
At least one participle phrase corresponding to bar comment content.
The instantiation of one training sample is as follows:
A comment sentence, i.e. a training sample corresponding to the comment content that comment content obtains after subordinate sentence is handled
For:Waiter's reception is thorough, warm;
The feature of the training sample is:It is waiter, reception, thorough, enthusiasm, generous;
The target of the training sample is:With emotion forward direction attribute, 1 is designated as
It is trained to obtain being used for emotion using FastText sorting algorithms by all training samples after above-mentioned processing
The emotion forecast model of attribute forecast.
Step 2, prediction.Comment sentence in the comment sentence set of each object to be recommended is segmented, obtained pair
At least one participle phrase of sentence should be each commented on, is input to using at least one participle phrase as the feature of comment sentence
The emotion forecast model that above-mentioned training obtains, obtain probability and emotion negative sense that the comment sentence corresponds to emotion forward direction attribute respectively
The probability of attribute, the emotion attribute prediction result using the larger emotion attribute of probable value as the comment sentence, that is, comment on sentence
The emotion attribute of each comment sentence in set is argmax (p (SentimentType)), and wherein SentimentType is
Emotion forward direction attribute or emotion negative sense attribute.
Step S505, the comment disaggregated model obtained using training, in the comment sentence set to each object to be recommended
Comment sentence belong to the probability of each classification and be predicted, the comment sentence of each object to be recommended is determined according to prediction result
Belong to the comment sentence of each classification in set.
In the present embodiment, the categorical attribute for presetting dimension is specially the categorical attribute by theme dimension, specific at one
In embodiment, include one or more of following classification by each classification of theme dimension:Service, environment, cost performance, mouth
Taste, quality, purpose and/or storekeeper.Wherein, the comment sentence in the comment sentence set to each object to be recommended belongs to each
The probability of classification is predicted, i.e.,:For a classification under every comment statement matching theme dimension.
In one particular embodiment of the present invention, one that disaggregated model includes but is not limited in following disaggregated model is commented on
Kind:FastText classifies, maximum entropy Maxent classification, and model-naive Bayesian.
Still by taking the FastText classification in above-mentioned disaggregated model as an example, for one point under comment statement matching theme dimension
Class, specific algorithm is similar to the algorithm of the above-mentioned classification for carrying out emotion attribute to comment sentence, will not be repeated here, below only
Difference is illustrated:
Step 1, model training.The theme dimension described by every comment sentence is marked out by way of manually marking
Under classification, the training objective using the classification as training sample corresponding to the comment sentence.Because artificial mark needs largely
Manpower, the present invention another specific embodiment in, be labeled by the method for semi-artificial half machine learning, specifically
Step is:The corresponding dictionary for describing each classification under artificial accumulation theme dimension, for example, dictionary under " service " include enthusiasm,
The words such as waiter, attitude;According to the dictionary of each classification, preliminary classification is carried out to comment sentence using Statistical learning model, its
In, Statistical learning model includes Bayesian model or maximum entropy model Maxent;The result of preliminary classification is manually evaluated and tested
And confirmation, obtain the training objective of each training sample.
The instantiation of one training sample is as follows:
A comment sentence, i.e. a training sample corresponding to the comment content that comment content obtains after subordinate sentence is handled
For:Waiter's reception is thorough, warm;
The feature of the training sample is:It is waiter, reception, thorough, enthusiasm, generous;
The target of the training sample is:Service.
It is trained to obtain being used for theme using FastText sorting algorithms by all training samples after above-mentioned processing
The comment disaggregated model of classification prediction.
Step 2, prediction.Comment sentence in the comment sentence set of each object to be recommended is segmented, obtained pair
At least one participle phrase of sentence should be each commented on, is input to using at least one participle phrase as the feature of comment sentence
The comment disaggregated model that above-mentioned training obtains, obtain the general of each classification that each comment sentence is belonging respectively under theme dimension
Rate, the classification that the big classification of probable value is belonged to as the comment sentence, that is, comment on each comment sentence in sentence set
What is belonged to is categorized as argmax (p (Topic)), and wherein p (Topic) is each point that comment sentence belongs under theme dimension
The probability of class, such as service, environment, cost performance, taste, quality, purpose or storekeeper.
Step S506, new comment sentence is generated using text generation model.
After belonging to the comment sentence of each classification in the comment sentence set for obtaining each object to be recommended, in order that
Belong to that the comment sentence of each classification is more rich, as an optional step of the present embodiment, utilize text generation model
New comment sentence is generated, and then causes the rationale for the recommendation of the candidate under each classification to have more novelty, and adds and pushes away
Recommend the diversity of reason.Optionally, using the Seq2Seq methods based on LSTM, i.e.,:Text life based on shot and long term memory network
New comment sentence is generated into model.
Fig. 6 shows the method flow diagram of the comment sentence using the generation of text generation model newly in the embodiment of the present invention,
I.e.:Step S506 refined flow chart, wherein, this method generates new comment sentence with the Seq2Seq models based on LSTM.Such as
Shown in Fig. 6, this method comprises the following steps:
Step S601, train to obtain as training sample using the comment sentence for belonging to each classification of each object to be recommended
Text generation model.
After classification prediction is carried out, the comment sentence for belonging to each classification of each object to be recommended is obtained, with same
Comment sentence under same classification specifically, is carried out word segmentation processing, obtained by the comment sentence under one classification as training sample
To at least one participle short sentence of each comment sentence, and with least one participle of every comment sentence under same classification
The phrase book cooperation of phrase composition is training sample, is trained using training sample, obtains corresponding each object to be recommended
The text generation model of each classification.For example, training is used as using the comment sentence in the environment classification under shop a theme dimension
The text generation model of sample, the then trained environment classification for obtaining shop a.
Step S602, the new comment sentence of each classification is belonged to using the generation of text generation model.
Specifically, before new comment sentence is generated using text generation model, for each classification, this point is obtained
The popular head-word of every comment sentence under class, as source phrase Candidate Set, wherein, popular head-word is that each bar under the classification is commented
First hot word of The Analects of Confucius sentence;The acquisition of popular head-word is by manually marking, or algorithm counts to obtain, specifically by each classification
The popular first participle phrase of comment sentence be collected, obtain the source phrase candidate of each classification of object to be recommended
Collection, the source phrase Candidate Set are made up of the popular head-word of the comment sentence under corresponding classify.
During new comment sentence is generated using text generation model, source phrase is chosen from the phrase Candidate Set of source
The input data of text generation model is obtained, input data is inputted into text generation model, output obtains object phrase;Will
Source phrase and object phrase are combined the input data for obtaining text generation model again, and input data is inputted to text and given birth to
Into in model, output again obtains new object phrase, by that analogy, until obtain meeting preset length requirement source phrase and
The combination of at least one object phrase is as new comment sentence.In this process, the need to be only chosen from the phrase Candidate Set of source
One source phrase is input in text generation model, and text generation model can just meet by said process final output one
The phrase sequence of preset length requirement, i.e.,:The combination of source phrase and at least one object phrase, commented as new under the classification
The Analects of Confucius sentence.
In one particular embodiment of the present invention, before the input data of text generation model is obtained, it is also necessary to will
Phrase sequence as input data supplies the phrase sequence for fixed length.Specifically, if source phrase and/or source phrase and target are short
The phrase sequence length of the combination of language is less than fixed length, then the combination to source phrase and/or source phrase and object phrase carries out benefit length
Processing, so as to get the phrase sequence length of input data be fixed length.Generally, inside LSTM text generations model, often
Secondary input is the phrase sequence of fixed length, however, during utilizing the new comment sentence of model generation in the present embodiment,
The phrase number inputted every time is uncertain, for example, during generation first aim phrase, input is a source phrase, raw
During into second target phrase, input is source phrase and first aim phrase, therefore, when inputting phrase sequence, it is necessary to
Elongated phrase sequence is modified to the phrase sequence of fixed length, specifically, for the phrase sequence of insufficient fixed length, with specific word
Symbol supplies the phrase lacked, wherein specific character includes line feed character.So that the length of fixed length is 5 phrases as an example, then defeated
When entering source phrase, the deletion sites of preceding 4 phrases are supplied with line feed symbol;In input source phrase and first aim phrase,
The deletion sites of preceding 3 phrases are supplied with line feed symbol.
Step S603, it will be obtained as the comment sentence of training sample and new comment sentence common combination each to be recommended
The comment sentence for belonging to each classification of object.
Specifically, the new comment sentence of corresponding each classification is obtained using text generation model, by under same classification
New comment sentence and the comment sentence common combination for belonging to the same classification predicted by classification, are obtained each to be recommended
Object belongs to the comment sentence of each classification.For example, being directed to shop a, categorized prediction has 100 comment sentences to belong to theme
" service " classification under dimension, and have 900 using the new comment sentence of the classification of service of text generation model generation, then
The service that belongs to that the 100 comment sentences and 900 new comment sentence common combinations that classification is predicted obtain shop a divides
1000 comment sentences of class.
Further, since the limited amount of the comment sentence in a certain classification of an object to be recommended, as the present embodiment
A kind of optional mode, for each classification, the object to be recommended with same object classification is belonged into the classification
Comment sentence as object classification and the training sample set of the classification;Train to obtain correspondingly for each training sample set
Text generation model;Wherein, object classification can be divided according to the business of operation, such as Sichuan cuisine, Hunan cuisine;Given birth to using text
Belong to the new comment sentence of the classification into model generation;It is total to as the comment sentence of training sample and new comment sentence
The comment sentence for belonging to the classification for the object to be recommended for obtaining that there is same object classification is combined together.Wherein, sample will be trained
Comment sentence and new comment sentence common combination in this set obtain point of the object to be recommended with same object classification
The comment sentence for belonging to the classification is specially:New comment sentence is correspondingly given into same object class now each to be recommended right
The corresponding classification of elephant, and the comment sentence that each object classification to be recommended predicts still falls within the correspondence of original object to be recommended
Classification so that the comment sentence of each classification of each object to be recommended more really embodies the feature of object to be recommended.
Wherein, new comment sentence is correspondingly given to the corresponding mode bag classified of each object to be recommended of same object class now
Include:Corresponding by each to be recommended object of the new comment sentence average mark to same object class now is classified, or, will be new
Comment sentence gives the corresponding classification of each object to be recommended of same object class now simultaneously.
Step S507, store the first index relative of object to be recommended and the comment sentence of each classification.
According to classification results, object to be recommended and the first index for commenting on sentence of each classification under theme dimension are established
Relation, wherein, it is as shown in table 1 below, each classification under corresponding theme dimension, store several rationale for the recommendation, optionally, from point
Belong in the comment sentence of corresponding classification and select several comment sentence conducts of the probable value of emotion forward direction attribute from high to low
The rationale for the recommendation of candidate, and then enable the rationale for the recommendation of candidate farthest to reflect the advantage letter of object to be recommended
Breath, for example, being predicted by emotional semantic classification, the comment sentence for belonging to classification of service in shop a comment sentence set has 100
Bar, the probable value that this 100 comment sentences are exported by emotion forecast model is sorted from high to low, selects emotion forward direction attribute
Probable value arrange rationale for the recommendation of the comment sentence of first five as the candidate of classification of service.
The shop of table 1 and the first index relative of the comment sentence (rationale for the recommendation of candidate) of each classification under theme dimension
Step S508, User action log data are excavated, obtain the user preference value that each user is directed to each classification, deposit
Store up user and the second index relative of the user preference value of each classification.
Specifically, the user behaviors log data of user include the search behavior data, navigation patterns data, click behavior of user
Data, splitting glass opaque data and lower single behavioral data, these behavioral datas can reflect the focus and point of interest of user,
The user preference value that user can be obtained and be directed to each classification is calculated by the analysis of behavior data.Wherein, calculating is analyzed
Method including machine learning method, statistic law or iteration decision Tree algorithms (Gradient Boosting Decision Tree,
Abbreviation GBDT).It is and user's storage corresponding with the user preference value of each classification is inclined to establish user and the user of each classification
The second index relative being worth well.
In one particular embodiment of the present invention, it is inclined to the user of each classification using equation below (2) calculating user
Good value:
Wherein, user_topic_score (topic) is preference fraction of the user to each theme, and n is user behavior
Species, behavior include but is not limited to search behavior, navigation patterns, click behavior, splitting glass opaque and lower single act, act_cnt
The number occurred for behavior in preset time period, wkFor weight corresponding to behavior,It is the time decay factor of behavial factor,Number of days for behavior time of origin to current time is poor.
By above-mentioned calculating, the user preference value that user is directed to each classification is obtained, establishes the use of user and each classification
Second index relative of family preference value, in the second index relative, for each user, the user couple is stored with each classification
The fraction of the user preference value of corresponding classification, it is as shown in table 2 below:
The user of table 2 and the second index relative of the user preference value of each classification under theme dimension
Above-mentioned steps S501 and step S508 corresponds to offline rationale for the recommendation and excavates flow, and above-mentioned steps are only to pushing away
The matching flow for recommending reason is prepared, and step S509 to step S512 below is indexed according to the first index relative and second
Relation carries out the flow of the On-line matching of rationale for the recommendation.
Step S509, the object recommendation for carrying user's mark sent according to specific user are asked, and obtain to be presented push away
Recommend list object.
Wherein, object recommendation request includes the request that user opens O2O e-commerce platforms, and user's point opens O2O electronics business
The request in specific special column in business platform, and the request of user's search, for above-mentioned request, O2O e-commerce platforms are both needed to really
Fixed recommended list to be presented, specifically, open the request of O2O e-commerce platforms for user or user's point opens O2O electricity
The request in specific special column in sub- business platform, the recommended list to be presented in homepage or special column need to be determined;Searched for user
The request of rope, it need to determine to match the recommended list to be presented of search result.In the present embodiment, recommended list to be presented
Can be determined according to achievable mode of the prior art, be not specifically limited herein.For example, according to O2O e-commerce platforms
The recommendation rules of server determine.
By the step, user's mark of the specific user of sending object recommendation request is determined, and is asked according to object recommendation
Ask and obtain recommended list to be presented.
Step S510, inquire about the comment short sentence that recommended to be presented corresponds to each classification, and inquiry specific user couple
Answer the preference value of each classification.
Specifically, according to the first index relative, inquire about in recommended list to be presented corresponding to each object to be recommended
The comment short sentence of each classification;According to the second index relative, the user preference value of each classification corresponding to inquiry specific user.Example
Such as, each shop is inquired about according to the corresponding store identification in the first index relative of table 1 in shop to be presented and corresponds to each classification
Comment sentence, i.e. candidate rationale for the recommendation;According to the user of specific user mark, inquiry should in the second index relative of table 2
User corresponds to the user preference value of each classification, in order to be determined to show the rationale for the recommendation of the user according to user preference value
Classification.
Step S511, the recommendation that the preference value selection object to be recommended to be presented of each classification is corresponded to according to specific user are managed
By.
Specifically, the preference value that each classification is corresponded to according to specific user determines to show the rationale for the recommendation of the specific user
Affiliated classification;Correspondingly show to choose in the classification belonging to the rationale for the recommendation of the specific user from recommended to be presented and push away
Recommend reason.
Fig. 7 shows the stream of the method for the rationale for the recommendation of the selection recommended to be presented of one specific embodiment of the present invention
Cheng Tu, i.e.,:The refined flow charts of step S511 in a specific embodiment, the flow chart are given to choose recommendation to be presented
One in a classification under the theme dimension of object comments on the step of short sentence is as rationale for the recommendation.As shown in fig. 7, step
S511 further comprises following sub-step:
Step S701, the order of the user preference value of each classification from high to low is to each point according to corresponding to specific user
Class is ranked up.
The user preference value of each classification according to corresponding to the specific user inquired, by the specific user to each classification
Preference sort from high to low, in order to subsequently according to the sequence determine show the rationale for the recommendation of user belonging to point
Class.
Step S702, according to the ranking results of each classification, successively from each classification corresponding to each object to be recommended
At least one comment sentence is chosen in comment sentence.
According to the classification of user's most preference, corresponded to from each object to be recommended inquired in the rationale for the recommendation of each classification
The rationale for the recommendation of the classification of user's most preference is chosen, for example, preference value highest of the user to taste, then select corresponding each in table 1
5 rationale for the recommendation under the taste classification in individual shop to be presented.
Particularly, because the business of different object operation to be recommended is different, some objects to be recommended will occur and only have
Classes of attribute under theme dimension, accordingly, the namely only corresponding several classes of the rationale for the recommendation excavated from user comment,
For example, for a shop for managing clothes business, then the taste classification under corresponding theme dimension, the shop is not just recommended
Reason, in this case, the just classification according to corresponding to the sequence of user preference value, corresponding point of object to be recommended is searched successively
The rationale for the recommendation of class, untill classification corresponding to presence is found.
Step S703, recommendation reason of the comment sentence as object to be recommended is randomly selected from least one comment sentence
By.
A rationale for the recommendation is randomly selected from least one rationale for the recommendation of each recommended to be presented as displaying
To the rationale for the recommendation of the specific user, this mode randomly selected so that even if it is same to be recommended to be repeatedly pushed to user
Object can also show different rationale for the recommendation.
Step S512, the combination of the rationale for the recommendation of recommended list to be presented and each object to be recommended is sent to spy
Determine user.
By each object to be recommended in recommended list to be presented with to should object to be recommended rationale for the recommendation
Combination, and the combination is showed into specific user to reach the purpose for responding the object recommendation of the specific user and asking.
The flow of above-mentioned steps S508 to step S512 rationale for the recommendation On-line matching, directly established using offline part
First index relative and the second index relative can be achieved with the rationale for the recommendation and the user preference phase of the user that will show user
Matching, and then make it that the flow of On-line matching is simpler, the speed of matching is rapider, thus reduces response object recommendation
The time of request.
The rationale for the recommendation generation method provided by the present embodiment, digging flow is excavated by offline rationale for the recommendation first
The comment content of object to be recommended, filtration treatment is carried out to the comment content of each object to be recommended, it is low to filter out information quality point
Comment content, and then cause as Candidate Recommendation reason comment on sentence quality it is higher;And to each object to be recommended
Comment content carry out subordinate sentence processing, obtain being directed to the comment sentence set of each object to be recommended, so that the displaying of platform
Interface is more concise;And the emotion forecast model obtained using training, in the comment sentence set to each object to be recommended
Comment sentence be predicted, filter out the comment sentence in comment sentence set with emotion negative sense attribute, and then show use
The rationale for the recommendation at family can reflect the advantageous information of object to be recommended;The comment disaggregated model obtained using training, to comment
Sentence carries out classification prediction, and has obtained corresponding the new of each each classification of object to be recommended using text generation model and commented
The Analects of Confucius sentence, and then to show the rationale for the recommendation of user more abundant and diversified, it is each using each object to be recommended
The lower comment sentence of classification and the new comment sentence of generation establish comment on sentence the of object to be recommended and each classification
One index relative, by first index relative, can quick search the candidate of each classification is corresponded to each object to be recommended
Rationale for the recommendation, and add using the comment of object to be recommended as rationale for the recommendation the diversity of reason, solve existing skill
The problem of rationale for the recommendation is single in art;And User action log data are excavated, and User action log data are analyzed
Calculate, obtain the user preference value that user is directed to each classification, the user preference value has reflected pass of the user to each classification
Note degree, and establish user and the second index relative of the user preference value of each classification;Pass through online rationale for the recommendation again
Object recommendation request with process response user, the object recommendation for carrying user's mark sent with specific reference to specific user please
Ask, obtain recommended list to be presented, in the second index relative, the user for inquiring about each classification corresponding to specific user is inclined
Good value in the first index relative, is inquired about in recommended list to be presented each with determining classification that the user most pays close attention to
The comment sentence for the classification that the user corresponding to object to be recommended most pays close attention to, you can recommendation when each object to be recommended is shown
Reason and the preference of user are to matching so that rationale for the recommendation has personalization, can farthest attract user;Treated from each
Recommended is same as rationale for the recommendation to that should choose at least one comment sentence in the comment sentence of classification most paid close attention to of user
The object to be recommended is pushed to user together, can specifically randomly select at least one comment sentence, accordingly even when repeatedly push
Can also recommend different rationale for the recommendation to the same object to be recommended of user, user every time with regard to different rationale for the recommendation can be seen,
And realize the bandwagon effect of the rationale for the recommendation in the face of thousand people thousand..
Fig. 8 shows the functional block diagram of the rationale for the recommendation generating means of one embodiment of the invention.As shown in figure 8, the dress
Put including:Excavate module 801, sort module 802 and matching module 803.
Module 801 is excavated, suitable for excavating the comment content of each object to be recommended.
Sort module 802, suitable for the comment content for each object to be recommended, the categorical attribute for presetting dimension is established,
Comment classification prediction is carried out, obtains the comment sentence for belonging to each classification of each object to be recommended.
Matching module 803, suitable for the comment sentence for belonging to each classification according to each object to be recommended, generate and wait to push away
Recommend the rationale for the recommendation of object.
Fig. 9 shows the functional block diagram of the rationale for the recommendation generating means of another embodiment of the present invention.As shown in figure 9, should
Device on the basis of Fig. 8, in addition to:Memory module 901 and generation module 902.
Memory module 901, suitable for storing the first index relative of object to be recommended and the comment sentence of each classification.
Module 801 is excavated to be further adapted for:User action log data are excavated, obtain each user for each classification
User preference value;
Memory module 901 is further adapted for:Store user and the second index relative of the user preference value of each classification.
Matching module 803 is further adapted for:The object recommendation for carrying user's mark sent according to specific user is asked,
Obtain recommended list to be presented;
According to the first index relative, each each classification corresponding to object to be recommended in recommended list to be presented is inquired about
Comment sentence;According to the second index relative, the user preference value of each classification corresponding to inquiry specific user;
The user preference value of each classification according to corresponding to specific user, from each classification corresponding to each object to be recommended
Comment sentence in choose at least one comment sentence, the rationale for the recommendation as the object to be recommended.
Sort module 802 is further adapted for:The comment disaggregated model obtained using training, is commented each object to be recommended
The probability that comment sentence in The Analects of Confucius sentence set belongs to each classification is predicted, and it is each to be recommended right to be determined according to prediction result
Belong to the comment sentence of each classification in the comment sentence set of elephant.
Generation module 902, suitable for using the comment sentence for belonging to each classification of each object to be recommended as training sample
Training obtains text generation model, and the new comment sentence of each classification is belonged to using the generation of text generation model;
Sort module 802 is further adapted for:Using as the comment sentence of training sample and new comment sentence common combination
Obtain the comment sentence for belonging to each classification of each object to be recommended.
Generation module 902 is further adapted for:For each classification, by the object to be recommended with same object classification
Belong to the comment sentence of the classification as object classification and the training sample set of the classification;
Train to obtain corresponding text generation model for each training sample set;
Belong to the new comment sentence of the classification using the generation of text generation model;
Sort module 802 is further adapted for:Using as the comment sentence of training sample and new comment sentence common combination
The comment sentence for belonging to the classification for the object to be recommended for obtaining that there is same object classification.
Generation module 902 is further adapted for:For each classification, every hot topic for commenting on sentence under the classification is obtained
Head-word, as source phrase Candidate Set;
From the phrase Candidate Set of source choose source phrase obtain the input data of text generation model, by input data input to
In text generation model, output obtains object phrase;
Source phrase and object phrase are combined to the input data for obtaining text generation model again, input data is defeated
Enter into text generation model, output again obtains new object phrase, by that analogy, until obtaining meeting preset length requirement
Source phrase and at least one object phrase combination as new comment sentence.
Generation module 902 is further adapted for:If the phrase sequence length of the combination of source phrase and/or source phrase and object phrase
Degree be less than fixed length, then the combination to source phrase and/or source phrase and object phrase carries out mending long processing, so as to get input data
Phrase sequence length be fixed length.
Matching module 803 is further adapted for:
The order of the user preference value of each classification from high to low is arranged each classification according to corresponding to specific user
Sequence;
According to the ranking results of each classification, successively from the comment sentence of each classification corresponding to each object to be recommended
Choose at least one comment sentence;
Rationale for the recommendation of the comment sentence as object to be recommended is randomly selected from least one comment sentence.
The embodiment of the present application provides a kind of nonvolatile computer storage media, and computer-readable storage medium is stored with least
One executable instruction, the computer executable instructions can perform the rationale for the recommendation generation method in above-mentioned any means embodiment.
Figure 10 shows a kind of structural representation of the structural representation of computing device of the embodiment of the present invention, present invention tool
Body embodiment is not limited the specific implementation of computing device.
As shown in Figure 10, the computing device can include:Processor (processor) 112, communication interface
(Communications Interface) 114, memory (memory) 116 and communication bus 118.
Wherein:
Processor 112, communication interface 114 and memory 116 complete mutual communication by communication bus 118.
Communication interface 114, for being communicated with the network element of miscellaneous equipment such as client or other servers etc..
Processor 112, for configuration processor 110, it can specifically perform in above-mentioned rationale for the recommendation generation method embodiment
Correlation step.
Specifically, program 110 can include program code, and the program code includes computer-managed instruction.
Processor 112 is probably central processor CPU, or specific integrated circuit ASIC (Application
Specific Integrated Circuit), or it is arranged to implement the integrated electricity of one or more of the embodiment of the present invention
Road.The one or more processors that computing device includes, can be same type of processor, such as one or more CPU;Also may be used
To be different types of processor, such as one or more CPU and one or more ASIC.
Memory 116, for depositing program 110.Memory 116 may include high-speed RAM memory, it is also possible to also include
Nonvolatile memory (non-volatile memory), for example, at least a magnetic disk storage.
Program 110 specifically can be used for so that processor 112 performs following operation:
Excavate the comment content of each object to be recommended;
For the comment content of each object to be recommended, the categorical attribute of default dimension is established, carries out comment classification prediction,
Obtain the comment sentence for belonging to each classification of each object to be recommended;
According to the comment sentence for belonging to each classification of each object to be recommended, the recommendation for generating object to be recommended is managed
By.
Program 110 can specifically be further used for so that processor 112 performs following operation:
Store the first index relative of object to be recommended and the comment sentence of each classification.
Program 110 can specifically be further used for so that processor 112 performs following operation:
Excavate User action log data, obtain each user be directed to each classification user preference value, storage user with
Second index relative of the user preference value of each classification.
Program 110 can specifically be further used for so that processor 112 performs following operation:
The object recommendation for carrying user's mark sent according to specific user is asked, and obtains recommended row to be presented
Table;
According to the first index relative, each each classification corresponding to object to be recommended in recommended list to be presented is inquired about
Comment sentence;According to the second index relative, the user preference value of each classification corresponding to inquiry specific user;
The user preference value of each classification according to corresponding to specific user, from each classification corresponding to each object to be recommended
Comment sentence in choose at least one comment sentence, the rationale for the recommendation as the object to be recommended.
Program 110 can specifically be further used for so that processor 112 performs following operation:
Subordinate sentence processing is carried out to the comment content of each object to be recommended, obtains being directed to the comment of each object to be recommended
Sentence set.
Program 110 can specifically be further used for so that processor 112 performs following operation:
The emotion forecast model obtained using training, the comment sentence in comment sentence set to each object to be recommended
It is predicted, filters out the comment sentence in comment sentence set with emotion negative sense attribute.
Program 110 can specifically be further used for so that processor 112 performs following operation:
The comment disaggregated model obtained using training, the comment sentence in comment sentence set to each object to be recommended
The probability for belonging to each classification is predicted, and determines to adhere to separately in the comment sentence set of each object to be recommended according to prediction result
In the comment sentence of each classification.
Program 110 can specifically be further used for so that processor 112 performs following operation:
Train to obtain text generation using the comment sentence for belonging to each classification of each object to be recommended as training sample
Model, the new comment sentence of each classification is belonged to using the generation of text generation model;
Point of each object to be recommended is obtained using as the comment sentence of training sample and new comment sentence common combination
Belong to the comment sentence of each classification.
Program 110 can specifically be further used for so that processor 112 performs following operation:
For each classification, by the comment sentence for belonging to the classification of the object to be recommended with same object classification
As object classification and the training sample set of the classification;
Train to obtain corresponding text generation model for each training sample set;
Belong to the new comment sentence of the classification using the generation of text generation model;
Obtain that there is same object classification using as the comment sentence of training sample and new comment sentence common combination
The comment sentence for belonging to the classification of object to be recommended.
Program 110 can specifically be further used for so that processor 112 performs following operation:
For each classification, every popular head-word for commenting on sentence under the classification is obtained, as source phrase Candidate Set;
From the phrase Candidate Set of source choose source phrase obtain the input data of text generation model, by input data input to
In text generation model, output obtains object phrase;
Source phrase and object phrase are combined to the input data for obtaining text generation model again, input data is defeated
Enter into text generation model, output again obtains new object phrase, by that analogy, until obtaining meeting preset length requirement
Source phrase and at least one object phrase combination as new comment sentence.
Program 110 can specifically be further used for so that processor 112 performs following operation:
If the phrase sequence length of the combination of source phrase and/or source phrase and object phrase is less than fixed length, to source phrase
And/or the combination of source phrase and object phrase carries out mending long processing, so as to get the phrase sequence length of input data be fixed length.
Program 110 can specifically be further used for so that processor 112 performs following operation:
The order of the user preference value of each classification from high to low is arranged each classification according to corresponding to specific user
Sequence;
According to the ranking results of each classification, successively from the comment sentence of each classification corresponding to each object to be recommended
Choose at least one comment sentence;
Rationale for the recommendation of the comment sentence as object to be recommended is randomly selected from least one comment sentence.
Wherein, the categorical attribute for presetting dimension is specially the categorical attribute by theme dimension, by each point of theme dimension
Generic attribute includes one or more of following categorical attribute:
Service, environment, cost performance, taste, quality, purpose and/or storekeeper.
Algorithm and display be not inherently related to any certain computer, virtual system or miscellaneous equipment provided herein.
Various general-purpose systems can also be used together with teaching based on this.As described above, required by constructing this kind of system
Structure be obvious.In addition, the present invention is not also directed to any certain programmed language.It should be understood that it can utilize various
Programming language realizes the content of invention described herein, and the description done above to language-specific is to disclose this hair
Bright preferred forms.
In the specification that this place provides, numerous specific details are set forth.It is to be appreciated, however, that the implementation of the present invention
Example can be put into practice in the case of these no details.In some instances, known method, structure is not been shown in detail
And technology, so as not to obscure the understanding of this description.
Similarly, it will be appreciated that in order to simplify the disclosure and help to understand one or more of each inventive aspect,
Above in the description to the exemplary embodiment of the present invention, each feature of the invention is grouped together into single implementation sometimes
In example, figure or descriptions thereof.However, the method for the disclosure should be construed to reflect following intention:I.e. required guarantor
The application claims of shield features more more than the feature being expressly recited in each claim.It is more precisely, such as following
Claims reflect as, inventive aspect is all features less than single embodiment disclosed above.Therefore,
Thus the claims for following embodiment are expressly incorporated in the embodiment, wherein each claim is in itself
Separate embodiments all as the present invention.
Those skilled in the art, which are appreciated that, to be carried out adaptively to the module in the equipment in embodiment
Change and they are arranged in one or more equipment different from the embodiment.Can be the module or list in embodiment
Member or component be combined into a module or unit or component, and can be divided into addition multiple submodule or subelement or
Sub-component.In addition at least some in such feature and/or process or unit exclude each other, it can use any
Combination is disclosed to all features disclosed in this specification (including adjoint claim, summary and accompanying drawing) and so to appoint
Where all processes or unit of method or equipment are combined.Unless expressly stated otherwise, this specification (including adjoint power
Profit requires, summary and accompanying drawing) disclosed in each feature can be by providing the alternative features of identical, equivalent or similar purpose come generation
Replace.
In addition, it will be appreciated by those of skill in the art that although some embodiments in this include institute in other embodiments
Including some features rather than further feature, but the combination of the feature of different embodiments means to be in the scope of the present invention
Within and form different embodiments.For example, in the following claims, embodiment claimed it is any it
One mode can use in any combination.
The all parts embodiment of the present invention can be realized with hardware, or to be run on one or more processor
Software module realize, or realized with combinations thereof.It will be understood by those of skill in the art that it can use in practice
Microprocessor or digital signal processor (DSP) realize some in rationale for the recommendation generating means according to embodiments of the present invention
Or some or all functions of whole parts.The present invention be also implemented as perform method as described herein one
Partly or completely equipment or program of device (for example, computer program and computer program product).It is such to realize this
The program of invention can store on a computer-readable medium, or can have the form of one or more signal.So
Signal can download and obtain from internet website, either provide on carrier signal or provided in the form of any other.
It should be noted that the present invention will be described rather than limits the invention for above-described embodiment, and ability
Field technique personnel can design alternative embodiment without departing from the scope of the appended claims.In the claims,
Any reference symbol between bracket should not be configured to limitations on claims.Word "comprising" does not exclude the presence of not
Element or step listed in the claims.Word "a" or "an" before element does not exclude the presence of multiple such
Element.The present invention can be by means of including the hardware of some different elements and being come by means of properly programmed computer real
It is existing.In if the unit claim of equipment for drying is listed, several in these devices can be by same hardware branch
To embody.The use of word first, second, and third does not indicate that any order.These words can be explained and run after fame
Claim.
Claims (10)
1. a kind of rationale for the recommendation generation method, including:
Excavate the comment content of each object to be recommended;
For the comment content of each object to be recommended, the categorical attribute of default dimension is established, comment classification prediction is carried out, obtains
The comment sentence for belonging to each classification of each object to be recommended;
According to the comment sentence for belonging to each classification of each object to be recommended, the rationale for the recommendation of object to be recommended is generated.
2. according to the method for claim 1, wherein, each classification that belongs to of each object to be recommended is obtained described
After commenting on sentence, methods described also includes:
Store the first index relative of the object to be recommended and the comment sentence of each classification.
3. according to the method for claim 2, wherein, methods described also includes:
Excavate User action log data, obtain each user be directed to each classification user preference value, storage user with
Second index relative of the user preference value of each classification.
4. according to the method for claim 3, wherein, each object to be recommended belongs to commenting for each classification to the basis
The Analects of Confucius sentence, the rationale for the recommendation for generating object to be recommended further comprise:
The object recommendation for carrying user's mark sent according to specific user is asked, and obtains recommended list to be presented;
According to first index relative, inquire about each corresponding to each object to be recommended in the recommended list to be presented
The comment sentence of classification;According to second index relative, the user preference of each classification corresponding to the specific user is inquired about
Value;
The user preference value of each classification according to corresponding to the specific user, from each classification corresponding to each object to be recommended
Comment sentence in choose at least one comment sentence, the rationale for the recommendation as the object to be recommended.
5. according to the method described in claim any one of 1-4, wherein, in the comment content for excavating each object to be recommended
Afterwards, methods described also includes:
Subordinate sentence processing is carried out to the comment content of each object to be recommended, obtains being directed to the comment sentence of each object to be recommended
Set.
6. the method according to claim 11, wherein, in the comment sentence collection for obtaining being directed to each object to be recommended
After conjunction, methods described also includes:
The emotion forecast model obtained using training, the comment sentence in comment sentence set to each object to be recommended are carried out
Prediction, filter out the comment sentence in the comment sentence set with emotion negative sense attribute.
7. the method according to claim 5 or 6, wherein, the basis presets the categorical attribute of dimension, carries out comment classification
Prediction, the comment sentence for belonging to each classification for obtaining each object to be recommended further comprise:
The comment disaggregated model obtained using training, the comment sentence in comment sentence set to each object to be recommended are belonged to
The probability of each classification is predicted, and determines to belong to respectively in the comment sentence set of each object to be recommended according to prediction result
The comment sentence of individual classification.
8. a kind of rationale for the recommendation generating means, including:
Module is excavated, suitable for excavating the comment content of each object to be recommended;
Sort module, suitable for the comment content for each object to be recommended, the categorical attribute of default dimension is established, is commented on
Classification prediction, obtains the comment sentence for belonging to each classification of each object to be recommended;
Matching module, suitable for the comment sentence for belonging to each classification according to each object to be recommended, generate object to be recommended
Rationale for the recommendation.
9. a kind of computing device, including:Processor, memory, communication interface and communication bus, the processor, the storage
Device and the communication interface complete mutual communication by the communication bus;
The memory is used to deposit an at least executable instruction, and the executable instruction makes the computing device such as right will
Ask and operated corresponding to the rationale for the recommendation generation method any one of 1-7.
10. a kind of computer-readable storage medium, an at least executable instruction, the executable instruction are stored with the storage medium
Make operation corresponding to rationale for the recommendation generation method of the computing device as any one of claim 1-7.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710978303.4A CN107609960A (en) | 2017-10-18 | 2017-10-18 | Rationale for the recommendation generation method and device |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710978303.4A CN107609960A (en) | 2017-10-18 | 2017-10-18 | Rationale for the recommendation generation method and device |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107609960A true CN107609960A (en) | 2018-01-19 |
Family
ID=61077580
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710978303.4A Pending CN107609960A (en) | 2017-10-18 | 2017-10-18 | Rationale for the recommendation generation method and device |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107609960A (en) |
Cited By (31)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108537828A (en) * | 2018-04-16 | 2018-09-14 | 成都博锐智晟科技有限公司 | A kind of shop data analysing method and system |
CN108733842A (en) * | 2018-05-29 | 2018-11-02 | 北京奇艺世纪科技有限公司 | Video recommendation method and device |
CN109003131A (en) * | 2018-07-18 | 2018-12-14 | 口口相传(北京)网络技术有限公司 | Precision Marketing Method and device based on user's scene properties information |
CN109189892A (en) * | 2018-09-17 | 2019-01-11 | 北京点网聚科技有限公司 | A kind of recommended method and device based on article review |
CN109299460A (en) * | 2018-09-18 | 2019-02-01 | 北京三快在线科技有限公司 | Analyze method, apparatus, electronic equipment and the storage medium of the evaluation data in shop |
CN109886823A (en) * | 2019-02-25 | 2019-06-14 | 北京奇艺世纪科技有限公司 | A kind of recommended method and device of social circle |
CN109902229A (en) * | 2019-02-01 | 2019-06-18 | 中森云链(成都)科技有限责任公司 | A kind of interpretable recommended method based on comment |
CN110264315A (en) * | 2019-06-20 | 2019-09-20 | 北京百度网讯科技有限公司 | Recommended information generation method and device |
WO2019214236A1 (en) * | 2018-05-11 | 2019-11-14 | 北京三快在线科技有限公司 | User-generated content summary determining and user-generated content recommending |
CN110457452A (en) * | 2019-07-08 | 2019-11-15 | 汉海信息技术(上海)有限公司 | Rationale for the recommendation generation method, device, electronic equipment and readable storage medium storing program for executing |
CN110472995A (en) * | 2019-07-08 | 2019-11-19 | 汉海信息技术(上海)有限公司 | To shop prediction technique, device, readable storage medium storing program for executing and electronic equipment |
CN110706064A (en) * | 2019-09-20 | 2020-01-17 | 汉海信息技术(上海)有限公司 | Dish recommendation information generation method, device, equipment and storage medium |
CN110929138A (en) * | 2018-09-04 | 2020-03-27 | 阿里巴巴集团控股有限公司 | Recommendation information generation method, device, equipment and storage medium |
CN110929021A (en) * | 2018-08-31 | 2020-03-27 | 阿里巴巴集团控股有限公司 | Text information generating method and text information generating device |
JP2020052612A (en) * | 2018-09-26 | 2020-04-02 | 富士ゼロックス株式会社 | Information processing system |
CN110968786A (en) * | 2019-11-29 | 2020-04-07 | 百度在线网络技术(北京)有限公司 | Visual information recommendation method, device, equipment and storage medium |
CN110968670A (en) * | 2019-12-02 | 2020-04-07 | 名创优品(横琴)企业管理有限公司 | Method, device, equipment and storage medium for acquiring attributes of popular commodities |
CN110990600A (en) * | 2019-12-04 | 2020-04-10 | 腾讯科技(深圳)有限公司 | Multimedia file recommendation method, multimedia file recommendation device, multimedia file parameter adjustment device, multimedia file recommendation medium and electronic equipment |
CN111046138A (en) * | 2019-11-15 | 2020-04-21 | 北京三快在线科技有限公司 | Recommendation reason generation method and device, electronic device and storage medium |
CN111444345A (en) * | 2020-03-30 | 2020-07-24 | 拉扎斯网络科技(上海)有限公司 | Dish name classification method and device |
CN111680496A (en) * | 2019-03-11 | 2020-09-18 | 丰田自动车株式会社 | Recommended sentence generation device, recommended sentence generation method, and computer-readable recording medium |
CN112115300A (en) * | 2020-09-28 | 2020-12-22 | 北京奇艺世纪科技有限公司 | Text processing method and device, electronic equipment and readable storage medium |
US20210264480A1 (en) * | 2020-02-26 | 2021-08-26 | International Business Machines Corporation | Text processing based interface accelerating |
CN113553421A (en) * | 2021-06-22 | 2021-10-26 | 北京百度网讯科技有限公司 | Comment text generation method and device, electronic equipment and storage medium |
CN113592589A (en) * | 2021-07-27 | 2021-11-02 | 上海致景信息科技有限公司 | Textile raw material recommendation method and device and processor |
CN113657975A (en) * | 2021-09-03 | 2021-11-16 | 广州微行网络科技有限公司 | Marketing method and system based on Internet E-commerce live broadcast platform |
WO2022001896A1 (en) * | 2020-07-01 | 2022-01-06 | 北京沃东天骏信息技术有限公司 | Recommendation reason generation method and apparatus, and device and storage medium |
CN114938473A (en) * | 2022-05-16 | 2022-08-23 | 上海幻电信息科技有限公司 | Comment video generation method and device |
CN115080845A (en) * | 2022-05-27 | 2022-09-20 | 北京百度网讯科技有限公司 | Recommendation reason generation method and device, electronic device and readable storage medium |
WO2023207450A1 (en) * | 2022-04-28 | 2023-11-02 | 北京字节跳动网络技术有限公司 | Topic recommendation method and apparatus, computer device, and storage medium |
CN117436997A (en) * | 2023-12-21 | 2024-01-23 | 深圳须弥云图空间科技有限公司 | Recommendation reason generation method and device based on retrieval capability |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20140365207A1 (en) * | 2013-06-06 | 2014-12-11 | Xerox Corporation | Method and system for classifying reviewers' comments and recommending related actions in idea-generating social media platforms |
CN104462363A (en) * | 2014-12-08 | 2015-03-25 | 百度在线网络技术(北京)有限公司 | Aspect displaying method and device |
CN105005552A (en) * | 2014-04-22 | 2015-10-28 | 北京四维图新科技股份有限公司 | Information processing method and apparatus |
CN107038609A (en) * | 2017-04-24 | 2017-08-11 | 广州华企联信息科技有限公司 | A kind of Method of Commodity Recommendation and system based on deep learning |
CN107133224A (en) * | 2017-04-25 | 2017-09-05 | 中国人民大学 | A kind of language generation method based on descriptor |
-
2017
- 2017-10-18 CN CN201710978303.4A patent/CN107609960A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20140365207A1 (en) * | 2013-06-06 | 2014-12-11 | Xerox Corporation | Method and system for classifying reviewers' comments and recommending related actions in idea-generating social media platforms |
CN105005552A (en) * | 2014-04-22 | 2015-10-28 | 北京四维图新科技股份有限公司 | Information processing method and apparatus |
CN104462363A (en) * | 2014-12-08 | 2015-03-25 | 百度在线网络技术(北京)有限公司 | Aspect displaying method and device |
CN107038609A (en) * | 2017-04-24 | 2017-08-11 | 广州华企联信息科技有限公司 | A kind of Method of Commodity Recommendation and system based on deep learning |
CN107133224A (en) * | 2017-04-25 | 2017-09-05 | 中国人民大学 | A kind of language generation method based on descriptor |
Cited By (43)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108537828B (en) * | 2018-04-16 | 2020-12-04 | 成都博锐智晟科技有限公司 | Shop data analysis method and system |
CN108537828A (en) * | 2018-04-16 | 2018-09-14 | 成都博锐智晟科技有限公司 | A kind of shop data analysing method and system |
WO2019214236A1 (en) * | 2018-05-11 | 2019-11-14 | 北京三快在线科技有限公司 | User-generated content summary determining and user-generated content recommending |
CN108733842A (en) * | 2018-05-29 | 2018-11-02 | 北京奇艺世纪科技有限公司 | Video recommendation method and device |
CN108733842B (en) * | 2018-05-29 | 2021-06-22 | 北京奇艺世纪科技有限公司 | Video recommendation method and device |
CN109003131A (en) * | 2018-07-18 | 2018-12-14 | 口口相传(北京)网络技术有限公司 | Precision Marketing Method and device based on user's scene properties information |
CN110929021A (en) * | 2018-08-31 | 2020-03-27 | 阿里巴巴集团控股有限公司 | Text information generating method and text information generating device |
CN110929021B (en) * | 2018-08-31 | 2024-05-31 | 阿里巴巴集团控股有限公司 | Text information generation method and text information generation device |
CN110929138A (en) * | 2018-09-04 | 2020-03-27 | 阿里巴巴集团控股有限公司 | Recommendation information generation method, device, equipment and storage medium |
CN110929138B (en) * | 2018-09-04 | 2024-03-29 | 阿里巴巴集团控股有限公司 | Recommendation information generation method, device, equipment and storage medium |
CN109189892A (en) * | 2018-09-17 | 2019-01-11 | 北京点网聚科技有限公司 | A kind of recommended method and device based on article review |
CN109299460A (en) * | 2018-09-18 | 2019-02-01 | 北京三快在线科技有限公司 | Analyze method, apparatus, electronic equipment and the storage medium of the evaluation data in shop |
JP2020052612A (en) * | 2018-09-26 | 2020-04-02 | 富士ゼロックス株式会社 | Information processing system |
JP7172361B2 (en) | 2018-09-26 | 2022-11-16 | 富士フイルムビジネスイノベーション株式会社 | Information processing system |
CN109902229A (en) * | 2019-02-01 | 2019-06-18 | 中森云链(成都)科技有限责任公司 | A kind of interpretable recommended method based on comment |
CN109886823A (en) * | 2019-02-25 | 2019-06-14 | 北京奇艺世纪科技有限公司 | A kind of recommended method and device of social circle |
CN111680496A (en) * | 2019-03-11 | 2020-09-18 | 丰田自动车株式会社 | Recommended sentence generation device, recommended sentence generation method, and computer-readable recording medium |
CN110264315A (en) * | 2019-06-20 | 2019-09-20 | 北京百度网讯科技有限公司 | Recommended information generation method and device |
CN110472995A (en) * | 2019-07-08 | 2019-11-19 | 汉海信息技术(上海)有限公司 | To shop prediction technique, device, readable storage medium storing program for executing and electronic equipment |
CN110457452B (en) * | 2019-07-08 | 2022-06-14 | 汉海信息技术(上海)有限公司 | Recommendation reason generation method and device, electronic equipment and readable storage medium |
CN110457452A (en) * | 2019-07-08 | 2019-11-15 | 汉海信息技术(上海)有限公司 | Rationale for the recommendation generation method, device, electronic equipment and readable storage medium storing program for executing |
CN110706064A (en) * | 2019-09-20 | 2020-01-17 | 汉海信息技术(上海)有限公司 | Dish recommendation information generation method, device, equipment and storage medium |
CN111046138B (en) * | 2019-11-15 | 2023-06-27 | 北京三快在线科技有限公司 | Recommendation reason generation method and device, electronic equipment and storage medium |
CN111046138A (en) * | 2019-11-15 | 2020-04-21 | 北京三快在线科技有限公司 | Recommendation reason generation method and device, electronic device and storage medium |
CN110968786B (en) * | 2019-11-29 | 2023-10-17 | 百度在线网络技术(北京)有限公司 | Visual information recommendation method, device, equipment and storage medium |
CN110968786A (en) * | 2019-11-29 | 2020-04-07 | 百度在线网络技术(北京)有限公司 | Visual information recommendation method, device, equipment and storage medium |
CN110968670A (en) * | 2019-12-02 | 2020-04-07 | 名创优品(横琴)企业管理有限公司 | Method, device, equipment and storage medium for acquiring attributes of popular commodities |
CN110990600A (en) * | 2019-12-04 | 2020-04-10 | 腾讯科技(深圳)有限公司 | Multimedia file recommendation method, multimedia file recommendation device, multimedia file parameter adjustment device, multimedia file recommendation medium and electronic equipment |
CN110990600B (en) * | 2019-12-04 | 2023-04-18 | 腾讯科技(深圳)有限公司 | Multimedia file recommendation method, multimedia file recommendation device, multimedia file parameter adjustment device, multimedia file recommendation medium and electronic equipment |
US20210264480A1 (en) * | 2020-02-26 | 2021-08-26 | International Business Machines Corporation | Text processing based interface accelerating |
CN111444345A (en) * | 2020-03-30 | 2020-07-24 | 拉扎斯网络科技(上海)有限公司 | Dish name classification method and device |
WO2022001896A1 (en) * | 2020-07-01 | 2022-01-06 | 北京沃东天骏信息技术有限公司 | Recommendation reason generation method and apparatus, and device and storage medium |
CN112115300A (en) * | 2020-09-28 | 2020-12-22 | 北京奇艺世纪科技有限公司 | Text processing method and device, electronic equipment and readable storage medium |
CN113553421A (en) * | 2021-06-22 | 2021-10-26 | 北京百度网讯科技有限公司 | Comment text generation method and device, electronic equipment and storage medium |
CN113592589B (en) * | 2021-07-27 | 2024-03-29 | 上海致景信息科技有限公司 | Textile raw material recommendation method, device and processor |
CN113592589A (en) * | 2021-07-27 | 2021-11-02 | 上海致景信息科技有限公司 | Textile raw material recommendation method and device and processor |
CN113657975B (en) * | 2021-09-03 | 2024-03-26 | 西安稻叶山供应链管理有限公司 | Marketing method and system based on Internet E-commerce live broadcast platform |
CN113657975A (en) * | 2021-09-03 | 2021-11-16 | 广州微行网络科技有限公司 | Marketing method and system based on Internet E-commerce live broadcast platform |
WO2023207450A1 (en) * | 2022-04-28 | 2023-11-02 | 北京字节跳动网络技术有限公司 | Topic recommendation method and apparatus, computer device, and storage medium |
CN114938473B (en) * | 2022-05-16 | 2023-12-12 | 上海幻电信息科技有限公司 | Comment video generation method and comment video generation device |
CN114938473A (en) * | 2022-05-16 | 2022-08-23 | 上海幻电信息科技有限公司 | Comment video generation method and device |
CN115080845A (en) * | 2022-05-27 | 2022-09-20 | 北京百度网讯科技有限公司 | Recommendation reason generation method and device, electronic device and readable storage medium |
CN117436997A (en) * | 2023-12-21 | 2024-01-23 | 深圳须弥云图空间科技有限公司 | Recommendation reason generation method and device based on retrieval capability |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107609960A (en) | Rationale for the recommendation generation method and device | |
CN110633373B (en) | Automobile public opinion analysis method based on knowledge graph and deep learning | |
CN109299994B (en) | Recommendation method, device, equipment and readable storage medium | |
CN105378727B (en) | Inverse operator is used to inquire about on online social networks | |
CN107403388B (en) | Syntactic model for structured search inquiry | |
CN104281622B (en) | Information recommendation method and device in a kind of social media | |
Gozhyj et al. | Uniform Method of Operative Content Management in Web Systems. | |
CN107578292B (en) | User portrait construction system | |
CN109767318A (en) | Loan product recommended method, device, equipment and storage medium | |
US20150286937A1 (en) | Text-messaging based concierge services | |
CN107133345A (en) | Exchange method and device based on artificial intelligence | |
CN106970991B (en) | Similar application identification method and device, application search recommendation method and server | |
CN106105107A (en) | For analyzing and synthesize the system and method for social communication data | |
CN105144151A (en) | Natural-language rendering of structured search queries | |
CN104254852A (en) | Method and system for hybrid information query | |
CN102760128A (en) | Telecommunication field package recommending method based on intelligent customer service robot interaction | |
CN108932322A (en) | A kind of geographical semantics method for digging based on text big data | |
CN109213859A (en) | A kind of Method for text detection, apparatus and system | |
Dey et al. | Emtagger: a word embedding based novel method for hashtag recommendation on twitter | |
CN110298029A (en) | Friend recommendation method, apparatus, equipment and medium based on user's corpus | |
CN109740072B (en) | Hotel ordering method and system based on POI under OTA platform | |
CN105160545A (en) | Delivered information pattern determination method and device | |
CN107368509B (en) | Communication means, communication system and computer-readable non-transitory storage medium | |
CN110362663A (en) | Adaptive more perception similarity detections and parsing | |
CN108256970A (en) | A kind of method that Products Show is carried out based on shopping need |
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 | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180119 |