CN110019662B - Label reconstruction method and device - Google Patents
Label reconstruction method and device Download PDFInfo
- Publication number
- CN110019662B CN110019662B CN201710818990.3A CN201710818990A CN110019662B CN 110019662 B CN110019662 B CN 110019662B CN 201710818990 A CN201710818990 A CN 201710818990A CN 110019662 B CN110019662 B CN 110019662B
- Authority
- CN
- China
- Prior art keywords
- historical
- model component
- tags
- label
- reconstruction
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/33—Querying
- G06F16/3331—Query processing
- G06F16/334—Query execution
- G06F16/3344—Query execution using natural language analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q30/00—Commerce
- G06Q30/06—Buying, selling or leasing transactions
- G06Q30/0601—Electronic shopping [e-shopping]
- G06Q30/0621—Item configuration or customization
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Business, Economics & Management (AREA)
- Physics & Mathematics (AREA)
- Accounting & Taxation (AREA)
- Finance (AREA)
- General Physics & Mathematics (AREA)
- General Business, Economics & Management (AREA)
- Strategic Management (AREA)
- Marketing (AREA)
- Economics (AREA)
- Development Economics (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- General Engineering & Computer Science (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The embodiment of the application discloses a label rebuilding method and a label rebuilding device. The method comprises the following steps: obtaining a target product label; performing label reconstruction on the target product label by using a deep neural network model component to generate a reconstructed label of the target product label; and the deep neural network model component is obtained by training according to the corresponding relation between a plurality of historical original labels and a plurality of historical reconstruction labels. By utilizing the method and the device, the product label can be quickly rebuilt, and the display speed of the product rebuilt label is increased.
Description
Technical Field
The present application relates to the field of data processing technologies, and in particular, to a method and an apparatus for reconstructing a tag.
Background
In an e-commerce platform, in order to improve search recall indexes and exposure opportunities of products, a plurality of descriptors, such as modifiers, marketing words, product words and the like, are often stacked in displayed product labels. While an excessive number of descriptors can result in product labels that are too long and contain varying degrees of redundant information. Since the screen size of the user client device (mobile phone, tablet computer) is limited, the product tag with the fixed length is often displayed in the product search result display page, and therefore, the original overlong product tag needs to be compressed.
The method for reconstructing the product label in the prior art mainly comprises two steps: (1) Performing Named Entity Recognition (NER) and term weighting on a plurality of descriptors in a product tag; (2) Constraint optimization (e.g., integer linear programming given business rules and length constraints) is performed on the reconstructed descriptors selected from the multiple descriptors. When the product label is reconstructed by using the above steps, it is necessary to perform category identification on each descriptor included in the product label, and calculate a weight value of each descriptor. Then, based on the weight values of the descriptors, reconstructing descriptors are selected from the descriptors according to given business rules or length limits and other constraint conditions, and reconstructing labels of the product labels are generated according to the reconstructing descriptors.
According to the steps, before named entity recognition is carried out on the descriptors in the product label, the category marking of the descriptors needs to be carried out manually. However, thousands of product categories often exist in an e-commerce platform, and if descriptor category marking needs to be performed manually, high labor cost is consumed.
Therefore, a product label rebuilding method which can get rid of manual marking and can be processed efficiently is needed in the prior art.
Disclosure of Invention
The embodiment of the application aims to provide a label rebuilding method and a label rebuilding device, which can be used for quickly rebuilding a product label and quickening the display speed of the product rebuilt label.
The method and the device for reconstructing the label provided by the embodiment of the application are specifically realized as follows:
a method of tag reconstruction, the method comprising:
obtaining a target product label;
performing label reconstruction on the target product label by using a deep neural network model component to generate a reconstructed label of the target product label; and the deep neural network model component is obtained by training according to the corresponding relation between a plurality of historical original labels and a plurality of historical reconstruction labels.
A method of tag reconstruction, the method comprising:
obtaining a search word, and obtaining a product label of at least one product according to the search word;
performing label reconstruction on the product label of the at least one product by using a preset deep neural network model component to generate a reconstructed label of the product label of the at least one product;
displaying the rebuilt label of the at least one product.
A method of tag reconstruction, the method comprising:
acquiring description information of a target product;
label reconstruction is carried out on the target product description information by utilizing a preset deep neural network model component, and a reconstructed label of the target product description information is generated; and the preset deep neural network model component is obtained by training according to the corresponding relation of a plurality of historical product description information and a plurality of historical reconstruction labels.
A method of tag reconstruction, the method comprising:
acquiring a search word, and acquiring product description information of at least one product according to the search word;
performing label reconstruction on the product description information of the at least one product by using a deep neural network model component to generate a reconstructed label of the product description information of the at least one product;
displaying the rebuilt label of the at least one product.
A tag reconstruction device comprising a processor and a memory for storing processor-executable instructions that when executed by the processor implement:
obtaining a target product label;
performing label reconstruction on the target product label by using a deep neural network model component to generate a reconstructed label of the target product label; and the deep neural network model component is obtained by training according to the corresponding relation between a plurality of historical original labels and a plurality of historical reconstruction labels.
A tag reconstruction device comprising a processor and a memory for storing processor-executable instructions that when executed by the processor implement:
obtaining a search word, and obtaining a product label of at least one product according to the search word;
performing label reconstruction on the product label of the at least one product by using a preset deep neural network model component to generate a reconstructed label of the product label of the at least one product;
displaying the rebuilt label of the at least one product.
A tag reconstruction device comprising a processor and a memory for storing processor-executable instructions that when executed by the processor implement:
acquiring description information of a target product;
label reconstruction is carried out on the target product description information by utilizing a preset deep neural network model component, and a reconstructed label of the target product description information is generated; and the preset deep neural network model component is obtained by training according to the corresponding relation of the historical product description information and the historical reconstruction labels.
A tag reconstruction device comprising a processor and a memory for storing processor-executable instructions that when executed by the processor implement:
acquiring a search word, and acquiring product description information of at least one product according to the search word;
performing label reconstruction on the product description information of the at least one product by using a deep neural network model component to generate a reconstructed label of the product description information of the at least one product;
displaying the rebuilt label of the at least one product.
The label rebuilding method provided by the application can be used for deeply learning the corresponding relation between a plurality of historical original labels and a plurality of historical rebuilding labels to build a deep neural network model component. Therefore, the deep neural network model component can be directly utilized to quickly reconstruct the label of the target product, and compared with the prior art, the process of setting the category label of each product descriptor in advance in the prior art can be avoided, and a large amount of labor cost is saved. In addition, on one hand, for a user client, when a user provides search words to recall a plurality of products, and then when the product labels of the products are displayed, the deep neural network model component can be used for quickly reconstructing the original labels of the products, so that the display speed of the reconstructed labels of the products is increased, and on the other hand, through deep learning of a large amount of historical data, the reconstructed labels obtained by the deep neural network model component better meet the search requirements of the user, and the experience of the user is improved.
Drawings
In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below, it is obvious that the drawings in the following description are only some embodiments described in the present application, and for those skilled in the art, other drawings can be obtained according to the drawings without any creative effort.
FIG. 1 is a schematic diagram of deep learning of historical product tags using the solution of the present application;
FIG. 2 is a schematic diagram of deep learning of historical product tags using the solution of the present application;
FIG. 3 is a schematic diagram of the processing of a target product tag using a seq2seq provided in the present application;
FIG. 4 is a schematic diagram of the processing of a target product tag with a seq2seq as provided herein;
FIG. 5 is a method flow diagram of one embodiment of a tag reconstruction method provided herein;
FIG. 6 is a schematic diagram of deep learning of historical product tags using the solution of the present application;
fig. 7 is a schematic diagram of deep learning of historical product tags in the technical solution of the present application.
Detailed Description
In order to make those skilled in the art better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only a part of the embodiments of the present application, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.
In order to facilitate those skilled in the art to understand the technical solutions provided in the embodiments of the present application, a technical environment for implementing the technical solutions is described below.
The method for reconstructing the product label in the prior art is described below by an example, if the original product label is "dedicated for the gas of the pan of the XX brand frying pan, which has less oil smoke and is not adhered to the pan of the steak pan". According to the description of the tag reconstruction method in the prior art, the category of each descriptor in the original product tag and the weight value of each descriptor need to be obtained. For example, the category of the descriptor "XX" is a brand word, the weight value is 0.35, the category of the "frying pan" is a kitchen ware, the weight value is 0.2, and so on. In the process of selecting the reconstruction descriptor according to the category and the weight value, the reconstruction descriptor needs to be adjusted according to a constraint condition (for example, the number of words of the reconstruction tag is 10 words). As can be seen from the above process, the e-commerce platform needs to set a category label and a weight value for each descriptor in advance, and such work is often completed manually. However, in an e-commerce platform, there are often thousands of product categories, and if each descriptor needs to be classified accurately and labeled with a category, high labor cost is required.
Based on the technical requirements similar to those described above, the label reconstruction method provided by the application can get rid of the working link of manually setting and describing the word category label. And performing deep neural network learning on the plurality of historical original label and reconstructed label pairs based on the plurality of historical original label and reconstructed label pairs, and training to obtain a deep neural network model component. The product label is reconstructed by the deep neural network model component obtained by training, so that not only can the original product label be reconstructed, but also the reconstructed product label can better meet the preset requirement.
The following describes a specific implementation of the method according to this embodiment through a specific application scenario.
In this embodiment, a deep neural network model component may be used to perform deep learning on a plurality of historical original tags and historical reconstructed tags of the historical products, so as to continuously optimize the deep neural network model component, and finally make the deep neural network model component meet preset requirements (for example, a value of a loss function is not greater than a preset threshold value, etc.). Since the plurality of historical original tags and the historical rebuilt tags of the historical product and then tags contain certain requirements or rules, in some embodiments, the historical rebuilt tags of the historical original tags need to have a certain product access rate, that is, after the calcium element and the rebuilt tags are displayed to the user, more than a certain number of users access the product information, such as clicking, collecting, generating transactions, and the like. Table 1 is a comparison table of a part of the original historical labels and the historical rebuilt labels thereof, and as shown in table 1, the historical rebuilt label corresponding to the original product label "Z brand summer slimming bandage a-shaped doll dress sweet two colors" is a "sweet two-color a-shaped doll dress", and the original product label "Y brand 2017 new spring dress korean edition slimming silk dress a-shaped dress having large codes" is a "Y brand korean edition slimming silk dress women's dress". Certainly, in other embodiments, the data may be obtained from historical data of a platform such as an electronic commerce platform, and may also be compiled according to manual experience, and the source of the data is not limited herein.
TABLE 1 comparison table of original label and historical rebuilt label of product
Taking a Sequence to Sequence (seq 2 seq) model component in the deep neural network model component as an example without limitation, the following describes how to use the seq2seq model component to perform a deep learning process on the first product original tag and the historical reconstruction tag in table 1. In other embodiments, the data may also be deeply learned by using model components such as Pointer Network and Pointer Generator, which is not limited herein. As shown in fig. 1, the basic seq2seq model component has an encoder part and a decoder part, both of which can be composed of a Recurrent Neural Network (RNN). In the encoder part, the input sequence can be converted into a context semantic vector of fixed length. Correspondingly, in a decoder portion, the context semantic vector may be used as input data for the decoder portion to generate an output sequence. As shown in fig. 1, in the encoder portion of the seq2seq model component, the input data may include a sequence of { x1, x2, x3, \ 8230;, x8} consisting of the word vectors of the individual descriptors in "Z-brand summer thinning band a doll skirt two colors sweet", where x1 corresponds to the word vector of the descriptor "Z-brand", x2 corresponds to the word vector of the descriptor "summer", and so on. H1-h8 in fig. 1 can be represented as hidden vectors respectively corresponding to x1-x8, and training parameters are set in the hidden vectors. After obtaining the sequence { x1, x2, x3, \ 8230;, x8}, the encoder may process the sequence to generate a context vector c (not shown), which may be transformed from the last concealment vector (i.e., h 8) in the above-mentioned concealment vectors. Of course, in other embodiments, the context vector c may also be obtained by transforming any one or more of h1-h8, which is not limited in this application. As shown in fig. 1, each concealment vector has an association relationship with its previous concealment vector, so the last concealment vector h8 can contain the information of all previous concealment vectors, i.e. the context vector c can contain all the information in the sequence { x1, x2, x3, \ 8230;, x8 }.
As shown in fig. 1, in the decoder portion, the context vector c may be input to the decoder, and similarly, a recurrent neural network model component may be included in the decoder, and the recurrent neural network model component may include a plurality of hidden vectors, such as h '1, h'2, h '3, and h' 4. The decoder may transform the context vector c to generate an output sequence. Ideally, the output sequence of the decoder is a sequence formed by each descriptor word vector in the sweet two-color A-shaped doll skirt. However, in practical cases, the output sequence does not match the historical reconstruction tag, and therefore, the training parameters in the respective hidden vectors in the encoder and the decoder need to be adjusted until the output sequence matches the historical reconstruction tag.
As shown in FIG. 1, the seq2seq model component may also include an Attention mechanism (Attention mechanism). The value of the histogram at the upper left of the graph may represent the degree of association between a single vector in the output sequence and each descriptor word vector in the original label of the product. The histogram in fig. 1 may be used to indicate the association degree between the word vectors corresponding to the "two colors" output from the decoder and the respective description word vectors in the "Z-brand summer thinning band a-shaped doll skirt sweet two colors". As can be seen from fig. 1, in the distribution of the association degrees between the word vectors of the respective descriptors and the word vector of "two colors", the maximum association degree between the "a word" and the "two colors" can characterize that the probability that the next descriptor of the "two colors" is the "a word" is the maximum in the prediction. Therefore, the accuracy of the seq2seq model component can be improved by influencing the output of the decoder through the Attention mechanism.
The method of the embodiment of the present application is explained by another example shown in fig. 2. As shown in fig. 2, two sets of data can be deep-learned simultaneously by two seq2seq model components, and the two seq2seq model components can share the same encoder. In the model component, not only the historical original tags and the corresponding historical reconstruction tags can be deeply learned, but also the historical original tags and the search terms of the historical original tags can be deeply learned. In one embodiment, after the historical original tags are recalled by the search terms and displayed to the user, a certain number of users can access the product information, and the historical original tags can be considered to be more in line with the search will of the users. As shown in fig. 2, after the user provides search words such as "two colors", "baby skirt", "a word", and the like to search for the two colors of the dress "Z-brand summer slimming band a-word baby skirt sweet", the probability that the user accesses the dress is high, and thus it can be seen that the search words "two colors", "baby skirt", "a word" have a certain influence on the product label "Z-brand summer slimming band a-word baby skirt sweet two colors". Then, the historical original labels and the corresponding search terms may be trained in the same deep learning manner as described above, so as to continuously adjust the training parameters in each hidden vector.
It should be noted that, as shown in fig. 2, both seq2seq model components have an Attention mechanism. In both the decoder 1 and the decoder 2, descriptors such as "two colors", "a word", "baby skirt", and the like are present. Then, in the embodiment of the present application, the training parameters in the hidden vectors h1 to h8 may be adjusted, so that the association degrees corresponding to the same descriptors are as consistent as possible. Specifically, for example, in the process of processing the association degree of the descriptor "two colors", the association degrees between the descriptors in the encoder 1 and the history original tag and the descriptors in the "two colors" in the encoder 2 may be obtained respectively, and the descriptors with the largest association degree may be disassociated respectively. It can be found that the two descriptors with the highest degree of association are both "a words". Errors of the association degrees between the two colors and the A word are calculated, and the errors are smaller than a preset error threshold value by continuously adjusting training parameters in hidden vectors h1-h 8. Therefore, the consistency of the association degree between the reconstructed product label and the original product label and the association degree between the user search word and the original product label can be enhanced.
How to use the above trained seq2seq model component is further described by an example as shown in fig. 3. The user small M searches products on the E-commerce platform client, and the search word is 'ground mat'. After the client obtains the search word "floor mat", the client may obtain product information of a plurality of floor mat products from the client or a server of the e-commerce platform, where the product information may include original tags of the products. When the client or the server judges that the original label of the product needs to be reconstructed according to the size of the user equipment, if the screen size of the user equipment is small and the complete original label cannot be displayed, the reconstruction process of the original label can be immediately implemented. In this example, the client or the server is provided with the trained seq2seq model component in fig. 2, so that the target product label can be reconstructed by using the encoder and the decoder 1. As shown in fig. 4, the target product label is "european style floor mat bedroom anti-skid window mat bedside mat machine washable". Similar to the above process, at least one descriptor may be extracted from the target product label, and the descriptor may include "euro", "floor mat", "living room", "bedroom", "anti-skid", "bay window mat", "bedside mat", "machine washable". The descriptors are converted into word vectors and then input in a sequential manner into an encoder, which, after generating a context vector c, sends the context vector c to the decoder 1. As shown in fig. 4, the decoder 1 calculates a target product label "the machine washable anti-slip mat for the european style floor mat living room bedroom anti-slip bay window mat bedside mat" as a reconstructed label "the machine washable anti-slip mat" according to the trained hidden vector.
The label reconstruction method described in the present application is described in detail below with reference to fig. 5. Fig. 5 is a flowchart of a method of an embodiment of a tag reconstruction method provided in the present application. Although the present application provides method steps as shown in the following examples or figures, more or fewer steps may be included in the method based on conventional or non-inventive efforts. In the case of steps where no necessary causal relationship exists logically, the order of execution of the steps is not limited to that provided by the embodiments of the present application. The method can be executed in sequence or in parallel according to the method shown in the embodiment or the figures (for example, in the environment of a parallel processor or a multi-thread processing) in the actual tag reconstruction process or the device execution.
As shown in fig. 5, the tag reconstruction method may include the steps of:
s501: obtaining a target product label;
s503: performing label reconstruction on the target product label by using a preset deep neural network model component to generate a reconstructed label of the target product label; the preset deep neural network model component is obtained by training according to a plurality of historical original labels and historical reconstruction labels of the historical original labels.
In this embodiment, the historical data may be trained by using a deep neural network model component, which is exemplified by a Sequence to Sequence (seq 2 seq) model component in the deep neural network model component without limitation. In this embodiment, the historical data may include a plurality of historical original tags and historical reconstructed tags of the historical original tags. In an example, the historical reconstruction tags may include product tags selected from a large amount of historical data and having a high recall rate, that is, tags generated after reconstructing a certain historical original tag are displayed to obtain a high user access rate, where the user access rate may include, for example, a click rate, a collection rate, a transaction rate, and the like of a user. Of course, in other embodiments, the historical rebuild tags may include rebuild tags having other preset criteria, which is not limited herein.
In this embodiment, the seq2seq model component can convert one sequence into another sequence, the sequence can be represented by a vector, and in this embodiment, the sequence can be composed of a plurality of descriptors in a product tag. The seq2seq model component has an encoder portion and a decoder portion, which may each be comprised of a Recurrent Neural Network (RNN). In the encoder part, the input sequence can be converted into a context semantic vector of fixed length. Correspondingly, in the decoder part, the fixed length context semantic vector can be used as the input data of the decoder part to generate an output sequence. Based on the above theory, in one embodiment of the present application, at least one descriptor may be extracted from the historical original tags, and the at least one descriptor may be configured into an original descriptor sequence. In the process of constructing the original sequence of descriptors, the descriptors may be converted into word vectors, and the original sequence of descriptors is composed of the word vectors of the descriptors. For example, a descriptor is extracted from a historical original label "Y-brand 2017 new spring dress, women's dress, korean edition, slimming silk dress, and a large size" of a dress of a certain dress, and an original descriptor sequence consisting of word vectors of 12 descriptors, such as "Y-brand", "2017", "new style", "spring dress", "women's dress", "korean edition", "slimming silk", "one dress", "a dress", and "large size" is generated. According to the statistical discovery of historical data, after the historical original label is reconstructed into the 'Y-brand Korean-edition real silk one-piece dress suit' the one-piece dress obtains a higher user access rate, and if a plurality of users click the product information, collect the product information or generate transaction on the product. Therefore, the 'Y brand korean edition retouching real silk dress suit-dress' can be set to correspond to the history rebuilding label of 'Y brand 2017 new style spring dress costume korean edition retouching real silk dress-dress a-shaped dress has big size'. And extracting descriptors from the historical reconstruction tag to obtain a reconstruction descriptor sequence comprising 6 descriptor vectors such as ' Y ' -brand ', ' real silk ', ' one-piece dress ', ' women's dress ', ' Korean edition ' and ' shape.
In this embodiment, as shown in fig. 6, the original descriptor sequence "Y-brand", "2017", "new style", "spring dress", "women dress", "korean edition", "body-building", "thin", "silk", "one-piece dress", "a-shaped dress" and "big code" is used as input data of the seq2seq model component encoder. As shown in fig. 3, x1 to x12 may be respectively expressed as word vectors of 12 descriptors in the original descriptor sequence, and h1 to h12 may be respectively expressed as hidden vectors corresponding to x1 to x 12. As shown in fig. 3, the original sequence of descriptive words may be transformed into a context vector c. The context vector c may be transformed from the last concealment vector in the above concealment vectors (i.e., h 12). Of course, in other embodiments, the context vector c may also be obtained by transforming any one or more of h1-h12, which is not limited in this application.
In this embodiment, after obtaining the context vector c, as shown in fig. 7, the context vector c may be input into a decoder of the seq2seq model component. The decoder, upon retrieving the context vector c, may convert the context vector c using another RNN to generate an output sequence of y1-y 6. Ideally, the output sequence of Y1-Y6 is the corresponding word vector sequence of the reconstruction descriptor sequence "Y brand", "real silk", "one-piece dress", "women's dress", "korean edition" and "trimming". Of course, in practical situations, it is difficult to match the output sequence y1-y6 exactly to the sequence of reconstruction descriptors. At this point, the training parameters in the respective concealment vectors (including the concealment vectors in the encoder and decoder) may be adjusted. And repeating the steps until an output sequence which is completely matched with the reconstruction descriptor sequence can be output in the decoder.
The label rebuilding method provided by the application can be used for deeply learning a plurality of historical original labels and historical rebuilding labels of the historical original labels, and a deep neural network model component is built. Therefore, the deep neural network model component can be directly utilized to quickly reconstruct the label of the target product, and compared with the prior art, the process of setting the category label of each product descriptor in advance in the prior art can be avoided, and a large amount of labor cost is saved. In addition, on one hand, for a user client, when a user provides search words to recall a plurality of products, and then when the product labels of the products are displayed, the deep neural network model component can be used for quickly reconstructing the original labels of the products, so that the display speed of the reconstructed labels of the products is increased, and on the other hand, through deep learning of a large amount of historical data, the reconstructed labels obtained by the deep neural network model component better meet the search requirements of the user, and the experience of the user is improved.
In one embodiment of the present application, an Attention mechanism (Attention mechanism) may also be added to the seq2seq model component. In the Attention mechanism, the association degree between the output vector and each description word vector in the history product original description word sequence can be calculated every time one step is performed in the decoding process. In this embodiment, when the degree of association between a certain term vector and the output vector is greater, it may be considered that the probability that the next output vector predicted by the output vector is the term vector is greater. As shown in fig. 7, when Y2 is calculated to be "real silk", according to the Attention mechanism, the association degree between "real silk" and each descriptive word and word vector in the original descriptive word sequence "Y brand", "2017", "new style", "spring dress", "women dress", "korean edition", "body built", "thin", "real silk", "one-piece dress", "a-shaped dress", "big code", respectively, is calculated, where the association degree may include the distance between word vectors. According to the histogram shown in fig. 3, it can be found that the association degree between the "real silk" and the "one-piece dress" in the above descriptor is the greatest, that is, the probability that the next output vector (i.e. Y3) of the "real silk" is the vector of the descriptor "one-piece dress" is higher. By adding the association degree into the hidden vector, the output of the hidden vector can be influenced, namely the prediction probability of the description word vector with higher association degree is improved.
In another embodiment of the present application, another seq2seq model component may be further provided on the basis of the seq2seq model component in the foregoing embodiment, and the seq2seq model component in the foregoing embodiment may share the same encoder. In this embodiment, the seq2seq model component in the above embodiment may be set as a first seq2seq model component, and the new seq2seq model component in this embodiment is a second seq2seq model component. In this embodiment, the first seq2seq model component and the second seq2seq model component may share one encoder, but both may have their own decoders, respectively. In the second seq2seq model component, training may be performed with another type of historical data, which may be obtained from the user's search logs. On many e-commerce platforms, there is often a huge amount of historical data, such as a user's search logs. In a specific scenario, after a user inputs a search term on the e-commerce platform to request product information, the e-commerce platform may recall information of a plurality of products according to the search term of the user and display the information of the plurality of products to the user. The user can click, collect or trade one or more product information according to own wishes. If a certain product is recalled through a certain search term, the probability that the user accesses the product information is high, and the label of the product information can be considered to be more in line with the search intention of the user.
Based on the actual application scenario, in one embodiment of the present application, a one-to-many data relationship from a product tag to a search term may be established. For example, according to data statistics on the e-commerce platform, the product label "Y-brand 2017 new spring dress, korean edition, slimming silk dress and a-type dress has a large code, and after the product label is recalled and displayed to the user through three search terms such as" silk "," one-type dress "," a-type dress ", and the like, the probability that the user accesses the product information is high. Therefore, the user search words corresponding to the fact that the Y-brand 2017 new-style spring dress ladies 'Korean-version dress slimming silk dress A-shaped dress has large codes' can be set as 'silk', 'one-piece dress' and 'A-shaped dress'. Then, the second seq2seq model assembly can be trained in the same way as the above embodiment with "Y-brand 2017 new spring dress korean edition tailored thin silk dress a-type dress having big code" as input data and "silk", "dress", "a-type dress" as output data.
It should be noted that the first seq2seq model component and the second seq2seq model component may share the same encoder, that is, in this embodiment, the shared encoder may be trained by using two types of history data at the same time. Wherein the first seq2seq model component utilizes a plurality of historical original tags and historical reconstructed tags of the historical original tags, and the second seq2seq model component utilizes the plurality of historical original tags and user search terms of the historical original tags. Therefore, on one hand, the attention distribution of the historical reconstruction tag and each descriptor in the historical original tag by the user search word can be drawn, namely the generated reconstruction tag of the target product tag is more consistent with the attention distribution of historical data; on the other hand, due to the action of the second seq2seq model component, the generated reconstructed label of the target product label can better accord with the search intention of the user.
Of course, the deep neural network can not only perform deep learning on the original label and the historical reconstruction label of the product, but also perform deep learning on the description information and the historical reconstruction label of the product. The product description information may include product labels, product profiles, product detail descriptions, and the like. Product profiles and product detail descriptions often contain richer information than product labels during a particular process. In one example, the product description information of a certain decorative picture is' brand XX mapping, more than triple, picture core material including canvas, mounting mode including frame, outer frame material including metal, color classification including A style-Lianxiang leaf B style-sansevieria C style-sansevieria D style-specular grass E style-tortoise back leaf F style-phoenix tree leaf G style-gold star fern H style-banana leaf I style-silver edge round leaf south American ginseng J style-spruce leaf, style of simple modern, spray painting process: the method comprises the steps of setting a history reconstruction label corresponding to product description information of the decorative picture as a ' green plant northern Europe style decorative picture ' according to statistics of user history data, wherein the history reconstruction label is independent, and the picture form comprises a plane, a pattern, a plant flower, a size of 40 × 60cm 50 × 70cm 60 × 90cm, and an outer frame type of a light wood color aluminum alloy frame black aluminum alloy frame and a product number of 0739 '. Thereafter, the product description information and the historical rebuilt tags may be deeply learned in the same manner as the above embodiments. It should be noted that, in the process of extracting the descriptor from the description information of the product, redundant information in the product description information may be removed, and a keyword having an actual meaning, such as a brand word, a material descriptor, a core word, or the like, may be extracted from the product description information. For example, for the product description information of the above decorative painting, description words that can be extracted may include "triple," "canvas," "framed," "metal outer frame," "painting," "plane," "plant flower," "aluminum alloy," and the like.
Although the present application provides method steps as described in the examples or flowcharts, more or fewer steps may be included based on conventional or non-inventive means. The order of steps recited in the embodiments is merely one manner of performing the steps in a multitude of orders and does not represent the only order of execution. When an actual apparatus or client product executes, it may execute sequentially or in parallel (e.g., in the context of parallel processors or multi-threaded processing) according to the embodiments or methods shown in the figures.
Those skilled in the art will also appreciate that, in addition to implementing the controller as pure computer readable program code, the same functionality can be implemented by logically programming method steps such that the controller is in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers, embedded microcontrollers and the like. Such a controller may therefore be considered as a hardware component, and the means included therein for performing the various functions may also be considered as a structure within the hardware component. Or even means for performing the functions may be conceived to be both a software module implementing the method and a structure within a hardware component.
The application may be described in the general context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, classes, etc. that perform particular tasks or implement particular abstract data types. The application may also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media including memory storage devices.
From the above description of the embodiments, it is clear to those skilled in the art that the present application can be implemented by software plus necessary general hardware platform. Based on such understanding, the technical solutions of the present application or portions thereof that contribute to the prior art may be embodied in the form of a software product, which may be stored in a storage medium, such as a ROM/RAM, a magnetic disk, an optical disk, or the like, and includes several instructions for enabling a computer device (which may be a personal computer, a mobile terminal, a server, or a network device) to execute the method according to the embodiments or some portions of the embodiments of the present application.
The embodiments in the present specification are described in a progressive manner, and the same or similar parts among the embodiments are referred to each other, and each embodiment focuses on the differences from the other embodiments. The application is operational with numerous general purpose or special purpose computing system environments or configurations. For example: personal computers, server computers, hand-held or portable devices, tablet-type devices, multiprocessor systems, microprocessor-based systems, set top boxes, programmable electronic devices, network PCs, minicomputers, mainframe computers, distributed computing environments that include any of the above systems or devices, and the like.
While the present application has been described with embodiments, those skilled in the art will appreciate that there are numerous variations and permutations of the present application without departing from the spirit of the application, and it is intended that the appended claims encompass such variations and modifications without departing from the spirit of the application.
Claims (16)
1. A method for label reconstruction, the method comprising:
obtaining a target product label;
performing label reconstruction on the target product label by using a deep neural network model component to generate a reconstructed label of the target product label; the deep neural network model component is obtained by training according to the corresponding relation between a plurality of historical original labels and a plurality of historical reconstruction labels; the deep neural network model component comprises one of: sequence to Sequence (seq 2 seq) model components, sequence to Sequence and Attention mechanism (seq 2seq and Attention mechanism) model components; the deep neural network model component is arranged to be trained in the following manner: acquiring a plurality of historical original tags, historical reconstruction tags of the historical original tags and user search words; constructing a first seq2seq model component and a second seq2seq model component, wherein the first seq2seq model component and the second seq2seq model component share the same encoder; and training the first seq2seq model component by respectively utilizing the plurality of historical original tags and the historical reconstruction tags of the historical original tags, and training the second seq2seq model component by utilizing the plurality of historical original tags and the user search words of the historical original tags.
2. The method of claim 1, wherein the deep neural network model component is configured to be trained in the following manner:
acquiring a plurality of historical original tags and historical reconstructed tags of the historical original tags;
constructing a deep neural network model component, wherein training parameters are set in the deep neural network model component;
and training the deep neural network model component by respectively utilizing the corresponding relations between the plurality of historical original labels and the historical reconstruction labels, and adjusting the training parameters until the deep neural network model component meets the preset requirement.
3. The method of claim 2, wherein the training the deep neural network model component using the correspondence between the plurality of historical original labels and the historical reconstructed labels respectively comprises:
extracting at least one descriptor from the historical original tag to form an original descriptor sequence;
extracting at least one descriptor from the historical reconstruction tag to form a reconstruction descriptor sequence;
and taking the original description word sequence as input data of the neural network and the reconstructed description word sequence as output data, and adjusting the training parameters until the deep neural network model component meets the preset requirement.
4. The method of claim 1, wherein the training the first seq2seq model component using the plurality of historical original tags and the historical rebuilt tags of the historical original tags, respectively, and wherein the training the second seq2seq model component using the plurality of historical original tags and the user search terms of the historical original tags comprises:
extracting at least one descriptor from the historical original label to form an original descriptor sequence;
extracting at least one descriptor from the historical reconstruction tag to form a reconstruction descriptor sequence;
constructing the user search word into a search word sequence;
taking the original description word sequence as input data of the first seq2seq model assembly and the reconstructed description word sequence as output data, and adjusting training parameters in the first seq2seq model assembly until the first seq2seq model assembly meets preset requirements;
and taking the original description word sequence as input data of the second seq2seq model assembly and the search word sequence as output data, and adjusting training parameters in the second seq2seq model assembly until the second seq2seq model assembly meets preset requirements.
5. The method of claim 1 or 4, wherein the label reconstruction of the target product label using the pre-set deep neural network model component, and generating the reconstructed label of the target product label comprises:
inputting the target product label to the encoder;
the encoder encodes the target product label to generate a context vector;
inputting the context vector to a decoder of the first seq2seq model component, the decoder generating a reconstructed tag of the target product tag according to the context vector decoding.
6. A method for tag reconstruction, the method comprising:
obtaining a search word, and obtaining a product label of at least one product according to the search word;
performing label reconstruction on the product label of the at least one product by using a preset deep neural network model component to generate a reconstructed label of the product label of the at least one product; the deep neural network model component comprises one of: sequence to Sequence (seq 2 seq) model components, sequence to Sequence and Attention mechanism (seq 2seq and Attention mechanism) model components; the deep neural network model component is arranged to be trained in the following manner: acquiring a plurality of historical original tags, historical reconstruction tags of the historical original tags and user search words; constructing a first seq2seq model component and a second seq2seq model component, wherein the first seq2seq model component and the second seq2seq model component share the same encoder; respectively training the first seq2seq model component by using the plurality of historical original tags and the historical reconstruction tags of the historical original tags, and training the second seq2seq model component by using the plurality of historical original tags and the user search words of the historical original tags;
displaying the rebuilt label of the at least one product.
7. A method for label reconstruction, the method comprising:
acquiring description information of a target product;
label reconstruction is carried out on the target product description information by utilizing a preset deep neural network model component, and a reconstructed label of the target product description information is generated; the preset deep neural network model component is obtained by training according to the corresponding relation of a plurality of historical product description information and a plurality of historical reconstruction labels; the deep neural network model component comprises one of: sequence to Sequence (seq 2 seq) model components, sequence to Sequence and Attention mechanism (seq 2seq and Attention mechanism) model components; the deep neural network model component is arranged to be trained in the following manner: acquiring a plurality of historical original tags, historical reconstruction tags of the historical original tags and user search words; constructing a first seq2seq model component and a second seq2seq model component, wherein the first seq2seq model component and the second seq2seq model component share the same encoder; and training the first seq2seq model component by respectively utilizing the plurality of historical original tags and the historical reconstruction tags of the historical original tags, and training the second seq2seq model component by utilizing the plurality of historical original tags and the user search words of the historical original tags.
8. A method for label reconstruction, the method comprising:
acquiring a search word, and acquiring product description information of at least one product according to the search word;
performing label reconstruction on the product description information of the at least one product by using a deep neural network model component to generate a reconstructed label of the product description information of the at least one product; the deep neural network model component comprises one of: sequence to Sequence (seq 2 seq) model components, sequence to Sequence and Attention mechanism (seq 2seq and Attention mechanism) model components; the deep neural network model component is arranged to be trained in the following manner: acquiring a plurality of historical original tags, historical reconstruction tags of the historical original tags and user search words; constructing a first seq2seq model component and a second seq2seq model component, wherein the first seq2seq model component and the second seq2seq model component share the same encoder; respectively training the first seq2seq model component by using the plurality of historical original tags and the historical reconstruction tags of the historical original tags, and training the second seq2seq model component by using the plurality of historical original tags and the user search words of the historical original tags;
displaying the rebuilt label of the at least one product.
9. A tag reconstruction apparatus comprising a processor and a memory for storing processor-executable instructions, the processor when executing the instructions implementing:
obtaining a target product label;
performing label reconstruction on the target product label by using a deep neural network model component to generate a reconstructed label of the target product label; the deep neural network model component is obtained by training according to the corresponding relation between a plurality of historical original labels and a plurality of historical reconstruction labels; the deep neural network model component comprises one of: sequence to Sequence (seq 2 seq) model components, sequence to Sequence and Attention mechanism (seq 2seq and Attention mechanism) model components; the deep neural network model component is arranged to be trained in the following manner: acquiring a plurality of historical original tags, historical reconstruction tags of the historical original tags and user search words; constructing a first seq2seq model component and a second seq2seq model component, wherein the first seq2seq model component and the second seq2seq model component share the same encoder; and training the first seq2seq model component by respectively utilizing the plurality of historical original tags and the historical reconstruction tags of the historical original tags, and training the second seq2seq model component by utilizing the plurality of historical original tags and the user search words of the historical original tags.
10. The apparatus of claim 9, wherein the deep neural network model component is configured to be trained in the following manner:
acquiring a plurality of historical original tags and historical reconstruction tags of the historical original tags;
constructing a deep neural network model component, wherein training parameters are set in the deep neural network model component;
and training the deep neural network model component by respectively using the plurality of historical original labels and the historical reconstruction labels of the historical original labels, and adjusting the training parameters until the deep neural network model component meets the preset requirement.
11. The apparatus of claim 10, wherein the processor when implementing the step of training the deep neural network model component using the correspondence between the plurality of historical original labels and the historical reconstructed labels, respectively, comprises:
extracting at least one descriptor from the historical original label to form an original descriptor sequence;
extracting at least one descriptor from the historical reconstruction tag to form a reconstruction descriptor sequence;
and taking the original description word sequence as input data of the neural network and the reconstructed description word sequence as output data, and adjusting the training parameters until the deep neural network model component meets the preset requirement.
12. The apparatus of claim 9, wherein the processor, in performing the step of training the first seq2seq model component using the plurality of historical original tags and the historical rebuilt tags of the historical original tags, respectively, and in training the second seq2seq model component using the plurality of historical original tags and the user search terms of the historical original tags, comprises:
extracting at least one descriptor from the historical original tag to form an original descriptor sequence;
extracting at least one descriptor from the historical reconstruction tag to form a reconstruction descriptor sequence;
constructing the user search word into a search word sequence;
taking the original description word sequence as input data of the first seq2seq model assembly and the reconstructed description word sequence as output data, and adjusting training parameters in the first seq2seq model assembly until the first seq2seq model assembly meets preset requirements;
and taking the original description word sequence as input data of the second seq2seq model assembly and the search word sequence as output data, and adjusting the training parameters in the second seq2seq model assembly until the second seq2seq model assembly meets the preset requirements.
13. The apparatus of claim 9 or 12, wherein the processor, when implementing step, performs label reconstruction on the target product label by using a preset deep neural network model component, and generates a reconstructed label of the target product label, comprises:
inputting the target product label to the encoder;
the encoder encodes the target product label to generate a context vector;
inputting the context vector to a decoder of the first seq2seq model component, the decoder generating a reconstructed tag of the target product tag according to the context vector decoding.
14. A tag reconstruction apparatus comprising a processor and a memory for storing processor-executable instructions, the processor when executing the instructions implementing:
obtaining a search word, and obtaining a product label of at least one product according to the search word;
performing label reconstruction on the product label of the at least one product by using a preset deep neural network model component to generate a reconstructed label of the product label of the at least one product; the deep neural network model component comprises one of: sequence to Sequence (seq 2 seq) model components, sequence to Sequence and Attention mechanism (seq 2seq and Attention mechanism) model components; the deep neural network model component is arranged to be trained in the following manner: acquiring a plurality of historical original tags, historical reconstruction tags of the historical original tags and user search words; constructing a first seq2seq model component and a second seq2seq model component, wherein the first seq2seq model component and the second seq2seq model component share the same encoder; respectively training the first seq2seq model component by using the plurality of historical original tags and the historical reconstruction tags of the historical original tags, and training the second seq2seq model component by using the plurality of historical original tags and the user search words of the historical original tags;
displaying the rebuilt label of the at least one product.
15. A tag reconstruction device comprising a processor and a memory for storing processor-executable instructions, the processor when executing the instructions implementing:
acquiring description information of a target product;
label reconstruction is carried out on the description information of the target product by utilizing a preset deep neural network model component, and a reconstructed label of the description information of the target product is generated; the preset deep neural network model component is obtained by training according to the corresponding relation of a plurality of historical product description information and a plurality of historical reconstruction labels; the deep neural network model component comprises one of: sequence to Sequence (seq 2 seq) model components, sequence to Sequence and Attention mechanism (seq 2seq and Attention mechanism) model components; the deep neural network model component is arranged to be trained in the following manner: acquiring a plurality of historical original tags, historical reconstruction tags of the historical original tags and user search words; constructing a first seq2seq model component and a second seq2seq model component, wherein the first seq2seq model component and the second seq2seq model component share the same encoder; and training the first seq2seq model component by respectively utilizing the plurality of historical original tags and the historical reconstruction tags of the historical original tags, and training the second seq2seq model component by utilizing the plurality of historical original tags and the user search words of the historical original tags.
16. A tag reconstruction device comprising a processor and a memory for storing processor-executable instructions, the processor when executing the instructions implementing:
acquiring a search word, and acquiring product description information of at least one product according to the search word;
performing label reconstruction on the product description information of the at least one product by using a deep neural network model component to generate a reconstructed label of the product description information of the at least one product; the deep neural network model component comprises one of: sequence to Sequence (seq 2 seq) model components, sequence to Sequence and Attention mechanism (seq 2seq and Attention mechanism) model components; the deep neural network model component is arranged to be trained in the following manner: acquiring a plurality of historical original tags, historical reconstruction tags of the historical original tags and user search words; constructing a first seq2seq model component and a second seq2seq model component, wherein the first seq2seq model component and the second seq2seq model component share the same encoder; respectively training the first seq2seq model component by using the plurality of historical original tags and the historical reconstruction tags of the historical original tags, and training the second seq2seq model component by using the plurality of historical original tags and the user search words of the historical original tags;
displaying the rebuilt label of the at least one product.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710818990.3A CN110019662B (en) | 2017-09-12 | 2017-09-12 | Label reconstruction method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710818990.3A CN110019662B (en) | 2017-09-12 | 2017-09-12 | Label reconstruction method and device |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110019662A CN110019662A (en) | 2019-07-16 |
| CN110019662B true CN110019662B (en) | 2022-10-18 |
Family
ID=67186231
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710818990.3A Active CN110019662B (en) | 2017-09-12 | 2017-09-12 | Label reconstruction method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110019662B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111341309A (en) * | 2020-02-18 | 2020-06-26 | 百度在线网络技术（北京）有限公司 | Voice interaction method, device, equipment and computer storage medium |
| US11551277B2 (en) * | 2020-09-11 | 2023-01-10 | Beijing Wodong Tianjun Information Technology Co., Ltd. | System and method for automatic generation of knowledge-powered content planning |
| CN112231580B (en) * | 2020-11-10 | 2024-04-02 | 腾讯科技（深圳）有限公司 | Information recommendation method and device based on artificial intelligence, electronic equipment and storage medium |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104111933B (en) * | 2013-04-17 | 2017-08-04 | 阿里巴巴集团控股有限公司 | Obtain business object label, set up the method and device of training pattern |
| CN106708813A (en) * | 2015-07-14 | 2017-05-24 | 阿里巴巴集团控股有限公司 | Title processing method and equipment |
-
2017
- 2017-09-12 CN CN201710818990.3A patent/CN110019662B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN110019662A (en) | 2019-07-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109992764B (en) | File generation method and device | |
| CN110147483B (en) | Title reconstruction method and device | |
| Kruiger et al. | Graph Layouts by t‐SNE | |
| CN112035747B (en) | Information recommendation method and device | |
| CN106959966A (en) | A kind of information recommendation method and system | |
| CN110019662B (en) | Label reconstruction method and device | |
| CN111260448A (en) | Artificial intelligence-based medicine recommendation method and related equipment | |
| CN105468596B (en) | Picture retrieval method and device | |
| CN107280693A (en) | Psychoanalysis System and method based on VR interactive electronic sand tables | |
| CN108846695A (en) | The prediction technique and device of terminal replacement cycle | |
| CN111401339B (en) | Method and device for identifying age of person in face image and electronic equipment | |
| CN107046557A (en) | The intelligent medical calling inquiry system that dynamic Skyline is inquired about under mobile cloud computing environment | |
| CN111737473B (en) | Text classification method, device and equipment | |
| CN106169961A (en) | The network parameter processing method and processing device of neutral net based on artificial intelligence | |
| CN113282623A (en) | Data processing method and device | |
| CN116862454A (en) | Indoor building design management method and system | |
| CN104573726B (en) | Facial image recognition method based on the quartering and each ingredient reconstructed error optimum combination | |
| CN107644042A (en) | Software program clicking rate estimates sort method and server | |
| CN111009299A (en) | Similar medicine recommendation method and system, server and medium | |
| CN109948055A (en) | A kind of recommended method based on group's preference | |
| CN115860835A (en) | Advertisement recommendation method, device and equipment based on artificial intelligence and storage medium | |
| CN111127145B (en) | Sorting recommendation method and system based on combination of catboost algorithm and deep learning | |
| CN114610308A (en) | Application function layout adjusting method and device, electronic equipment and storage medium | |
| CN108089842A (en) | A kind of method and its system using artificial intelligence structure UI | |
| CN106777092A (en) | The intelligent medical calling querying method of dynamic Skyline inquiries under mobile cloud computing environment |
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 | ||
| REG | Reference to a national code |
Ref country code: HK Ref legal event code: DE Ref document number: 40010954 Country of ref document: HK |
|
| GR01 | Patent grant | ||
| GR01 | Patent grant |