WO2019072098A1 - 一种识别核心产品词的方法和系统 - Google Patents

一种识别核心产品词的方法和系统 Download PDF

Info

Publication number
WO2019072098A1
WO2019072098A1 PCT/CN2018/108230 CN2018108230W WO2019072098A1 WO 2019072098 A1 WO2019072098 A1 WO 2019072098A1 CN 2018108230 W CN2018108230 W CN 2018108230W WO 2019072098 A1 WO2019072098 A1 WO 2019072098A1
Authority
WO
WIPO (PCT)
Prior art keywords
product
image
training
word
words
Prior art date
Application number
PCT/CN2018/108230
Other languages
English (en)
French (fr)
Inventor
马超义
Original Assignee
北京京东尚科信息技术有限公司
北京京东世纪贸易有限公司
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 北京京东尚科信息技术有限公司, 北京京东世纪贸易有限公司 filed Critical 北京京东尚科信息技术有限公司
Priority to US16/754,737 priority Critical patent/US11741094B2/en
Publication of WO2019072098A1 publication Critical patent/WO2019072098A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/24Querying
    • G06F16/245Query processing
    • G06F16/2452Query translation
    • G06F16/24522Translation of natural language queries to structured queries
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION 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
    • G06Q30/00Commerce
    • G06Q30/06Buying, selling or leasing transactions
    • G06Q30/0601Electronic shopping [e-shopping]
    • G06Q30/0623Item investigation
    • G06Q30/0625Directed, with specific intent or strategy
    • G06Q30/0627Directed, with specific intent or strategy using item specifications
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33Querying
    • G06F16/3331Query processing
    • G06F16/334Query execution
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/31Indexing; Data structures therefor; Storage structures
    • G06F16/313Selection or weighting of terms for indexing
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/35Clustering; Classification
    • G06F16/355Class or cluster creation or modification
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION 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
    • G06Q30/00Commerce
    • G06Q30/06Buying, selling or leasing transactions
    • G06Q30/0601Electronic shopping [e-shopping]
    • G06Q30/0623Item investigation

Definitions

  • the present invention relates to the field of computer technologies, and in particular, to a method and system for identifying core product words.
  • the core product word in the product title refers to the product of the product described by the title. For example, the core product of the "Korean NY Yankees Men's and Women's Caps" is the "Cap”.
  • the methods for identifying core product words in a product title mainly include: a vocabulary-based rule method, a conditional random field-based sequence labeling method, and an LSTM-based deep learning method.
  • a vocabulary-based rule method refers to maintaining a mapped vocabulary. For example: air conditioning filter - air conditioning, when the "air conditioning" and "filter” appear in the title, then the "filter” is the core product word.
  • the conditional random field sequential annotation method and the LSTM-based deep learning method both use the machine learning model to train through a large number of corpora, and use the learned model to predict whether the product words in the new title are core product words.
  • the prior art methods for identifying core product words in product titles utilize only textual information.
  • the vocabulary method does not have generalization ability, that is, it cannot process new samples that are not included in the vocabulary, and there are many kinds of product titles, and it is impossible to include all the cases in the vocabulary.
  • the product word stacking phenomenon in the title is serious and it is difficult to use syntax information analysis, so the machine learning method can not very effectively identify the core product words in the product title.
  • the merchant's various and irregular headline writing forms make it difficult to judge the core product words in the title by simply using the text information.
  • the embodiments of the present invention provide a method and system for identifying a core product word, which can more effectively and intuitively and accurately determine a core product word of a product title.
  • the quality of the results returned by the user can be improved, the user experience can be improved, and the conversion can be improved.
  • a method of identifying a core product word is provided.
  • the method for identifying a core product word includes: obtaining a display image of a product, determining a plurality of candidate product words included in a title of the product, and determining a plurality of product image sets, wherein the plurality of product image sets are Each product image set is in one-to-one correspondence with each of the plurality of candidate product words; for each of the plurality of product image sets, according to the display image and the product image set And determining, by the respective images, the similarity of the candidate product words corresponding to the product image set in the display image, thereby obtaining a plurality of similarities; and comparing the plurality of similarities to a similarity greater than a preset threshold The corresponding candidate product words are determined as core product words.
  • the determining the similarity of the candidate product words corresponding to the product image set in the display image comprises: training the Siamese network based on the training data to obtain the trained Siamese network; Determining, in each of the plurality of product image sets, a product corresponding to the product image set in the display image according to the display image and each image in the product image set through the trained Siamese network The similarity of candidate product words, resulting in multiple similarities.
  • the method before training the Siamese network based on the training data to obtain the trained Siamese network, the method further includes: selecting a predetermined number of training product words; for each training product word, selecting more than two according to the product search log to include the training The search term of the product word; for each search term, according to the click volume and click rate of the product under the search term, the same quantity of the product and the display image of the product are selected; for each training product word, the training product word is The display images of the products belonging to the same search term constitute a pair of positive examples, and the display images of the products belonging to the different search words under the training product word constitute a pair of negative examples.
  • the step of training the Siamese network based on the training data comprises: clustering the display images of the training product words by using a clustering algorithm to obtain more than one for the display image of each training product word in the training data. a category center; using the one or more category centers as a product image set of the training product word; and saving the product image set of each training product word to the image feature library; the step of determining the plurality of product image sets includes Determining, according to the image feature library, a plurality of product image sets, each product image set of the plurality of product image sets being in one-to-one correspondence with each of the plurality of candidate product words.
  • conditional random field is used to determine a plurality of candidate product words included in the title of the item.
  • a system for identifying a core product word is provided.
  • the system for identifying a core product word in the embodiment of the present invention includes: an obtaining module, configured to acquire a display image of an item, determine a plurality of candidate product words included in a title of the product, and determine a plurality of product image sets, the plurality of Each product image set in the product image set is in one-to-one correspondence with each of the plurality of candidate product words; a similarity determining module is configured for each product image set in the plurality of product image sets Determining, according to the display image and each image in the product image set, a similarity of the candidate product words corresponding to the product image set in the display image, thereby obtaining a plurality of similarities; comparing the module, using The candidate product words corresponding to the similarities greater than the preset threshold among the plurality of similarities obtained by the similarity determining module are determined as core product words.
  • the similarity determining module includes: a model training unit, configured to train the Siamese network based on the training data to obtain a trained Siamese network; and a calculating unit, configured to determine, in the display image, by the trained Siamese network The similarity of the candidate product words corresponding to the product image set.
  • the similarity determining module further includes a training data acquiring unit, configured to select a predetermined number of training product words; for each training product word, select two or more search words including the training product word according to the product search log. For each search term, according to the click volume and click rate of the product under the search term, the same quantity of the product and the display image of the product are selected; for each training product word, the training product word belongs to the same search term.
  • the display image of the product constitutes a pair of positive examples, and the display images of the products belonging to the different search words under the training product word constitute a pair of negative examples.
  • the model training unit is further configured to perform a clustering process on the display image of the training product word by using a clustering algorithm to obtain a display image of each training product word in the training data to obtain more than one category center; More than one category center as a product image set of the training product word; and, the product image set of each training product word is saved to the image feature library; the obtaining module determines a plurality of product image sets according to the image feature library, Each product image set in the plurality of product image sets is in one-to-one correspondence with each of the plurality of candidate product words.
  • the obtaining module determines, by using a conditional random field, a plurality of candidate product words included in a title of the commodity.
  • an electronic device for identifying a core product word is provided.
  • An electronic device for identifying a core product word of an embodiment of the present invention includes: one or more processors; and storage means for storing one or more programs when the one or more programs are used by the one or more processors Executing to cause the one or more processors to implement the above method of identifying core product words.
  • a computer readable medium having stored thereon a computer program, the program being executed by a processor, implements the method of identifying a core product word.
  • One embodiment of the above invention has the following advantages or advantages: since the image information of the commodity is added to the judgment of the core product word, it is difficult to accurately determine the core product word in the product title caused by simply using the text information in the prior art. The problem, in turn, achieves the technical effect of improving the accuracy of the word recognition of the core product of the product by means of the information of the product image.
  • FIG. 1 is a schematic diagram of a main flow of a method of identifying a core product word according to an embodiment of the present invention
  • FIG. 2 is a schematic structural diagram of a Siamese network
  • FIG. 3 is a schematic diagram of main modules of a system for identifying core product words in accordance with an embodiment of the present invention
  • FIG. 4 is an exemplary system architecture diagram to which an embodiment of the present invention may be applied;
  • Figure 5 is a block diagram showing the structure of a computer system suitable for implementing a terminal device or server in accordance with an embodiment of the present invention.
  • FIG. 1 is a schematic diagram of a main process of a method for identifying a core product word according to an embodiment of the present invention. As shown in FIG. 1 , a method for identifying a core product word according to an embodiment of the present invention mainly includes:
  • Step S101 Acquire a display image of the product, determine a plurality of candidate product words included in the title of the product, and determine a plurality of product image sets, each product image set in the plurality of product image sets and each of the plurality of candidate product words
  • the candidate product words correspond one-to-one.
  • the display image of the product refers to an image displayed on the search page when searching for the product, and the display image is provided and set by the seller, which can fully reflect the product, so the display image according to the product can be more accurate and fast. Identify the core product words in the title.
  • a plurality of candidate product words included in the title of the item may be determined using a conditional random field (CRF).
  • CRF is a sequence labeling model, which can effectively and accurately distinguish between word segmentation and part-of-speech and name entity recognition.
  • Each candidate product word corresponds to a product image collection, and the image in the product image collection can reflect the candidate product word.
  • the product image set includes one or more images, and according to the image in the product image set, the product corresponding to the candidate product word can be determined. Therefore, when the similarity is determined in step S102, the similarity between the candidate product word and the product corresponding to the product image set may be determined by comparing the image in the product image set with the displayed image of the product, in the embodiment of the present invention.
  • the item in the product is the item of the core product word to be identified.
  • Step S102 determining, for each product image set in the plurality of product image sets, the similarity of the candidate product words corresponding to the product image in the display image according to the display image and each image in the product image set, thereby Get multiple similarities.
  • the similarity of the candidate product words corresponding to the product image and the product image set in the display image determined in this step may reflect the degree of similarity between the product of the core product word to be identified and the product represented by the candidate product word.
  • the product title is “Fashion Casual Commercial Spinning Bicycle Belt Commercial Dynamic Exercise Bike”, after determining the candidate product words “Belt”, “Bicycle”, “Exercise Car”, A product image set of the candidate product word “belt”, a product image set of the candidate product word “cycling”, and a product image set of the candidate product word "exercise car” are respectively obtained.
  • the product image set of the candidate product word "belt” includes 5 images, and the 5 images are respectively compared with the displayed image of the product A, and the similarity of the combination is 0.2, 0.5, 0.3, 0.1, 0.4, and the product is further
  • the similarity of the candidate product word "belt” corresponding to the image set to the item A is 0.5.
  • it is determined that the similarity between the commodity A and the candidate product word "cycling” is 0.9
  • the similarity between the commodity A and the candidate product word "exercise bicycle” is 0.85. Therefore, after the above process, three similarities are obtained, which are 0.5, 0.9, and 0.85, respectively.
  • the process of step S102 can be performed by the trained model. Therefore, the method for identifying a core product word in the embodiment of the present invention further includes: training the Siamese network based on the training data to obtain the trained Siamese network.
  • the Siamese network is a multi-branch parameter shared network structure, which is a measure of similarity.
  • the principle of this network uses the neural network to extract the description operator, obtain the feature vector, and then use the feature vectors of the two pictures to determine the similarity. The advantage is that it can distinguish the samples of the new untrained categories.
  • the Siamese network learns a measure of similarity from the data (the closer the two things are, the greater the measure of similarity, and the more alienated the two things, the smaller their measure of similarity).
  • the resulting metrics compare and match the samples of the new unknown category.
  • the method for identifying a core product word further includes acquiring training data.
  • selecting a predetermined number of training product words selecting, based on the product search log, two or more search words including the training product words for each training product word; according to the click volume and click of the product under the search word Rate, for each search term, select the same number of products and the display image of the product; for each training product word, the display image of the product belonging to the same search word under the training product word is composed of a pair of positive examples, The display images of the products belonging to different search terms under the product word constitute a pair of negative examples. For each training product word, a pair of positive pairs and a pair of negative pairs can be obtained.
  • the positive and negative pairs of all product words constitute training data. Therefore, based on the obtained positive and negative examples, the Siamese network is trained.
  • the Siamese network is trained based on the following assumptions: the same product words are similar under the product, and the products under different product terms are not similar. Although the products under different product terms may be the same, they are not similar in the case of large similarity in the training process, such as selecting 6000 training product words, "refrigerator” and different product words "air conditioning” similar products if There are five, and the possible error is only 5/6000. It is assumed that the same product word is similar under the product, and it is feasible to have different products under different product words.
  • each training product word select two search words containing the training product words, a total of 12,000; each search word selects a maximum of 20 items of clicks and clicks, and Correspondingly, the display image of the product was obtained, and a total of 240,000 images were obtained.
  • the display image of one product under each training product word the display image of another product under the same search term may be selected to form a pair of positive examples, and the display images of the products under other search words are selected to form a pair of negative examples. Yes, I finally got 480,000 pairs of training data.
  • the Siamese network is trained to obtain the trained Siamese network.
  • the trained Siamese network can be applied to the display images of all the merchandise in the merchandise library, and the core product words for each merchandise are determined. Furthermore, more than one product image set can be obtained, each product image set corresponding to one product word, and the product set contains more than one image or image features of the one or more images.
  • the embodiment of the present invention may be directed to a ambiguous product word (the product word has multiple understandings), and the display image of all product words may be processed by a clustering algorithm (DBSACN) to obtain a The above category centers, and more than one category center as the product image collection of the product word.
  • DBSACN is a density-based clustering algorithm that divides high-density intervals into clusters and can find arbitrary shapes in the spatial database of noise.
  • the product image collection for each product word is saved to the image feature library. Therefore, when determining a product image set corresponding to each candidate product word, a plurality of product image sets may be determined according to the image feature library.
  • the candidate corresponding to the product image and the product image set in the display image can be determined according to the display image and each image in the product image set through the trained Siamese network. The similarity of the product words, resulting in multiple similarities.
  • Step S103 Determine candidate product words corresponding to the similarities greater than the preset threshold among the plurality of similarities as core product words.
  • the similarity obtained in step S102 is the similarity of the candidate product words corresponding to the product and the product image set in the display image, and for each similarity, it corresponds to the candidate product word determining the similarity. Therefore, after a plurality of similarities are obtained, each similarity corresponds to one candidate product word. Moreover, each degree of similarity reflects the degree of similarity between the item represented by the candidate product word and the item of the core product word to be identified. After the similarity corresponding to each candidate product word is obtained, one or more candidate product words may be determined as core product words according to a preset threshold.
  • the candidate product words are “belt”, “cycling”, “exercise car”, through the product display picture and product image collection of candidate product words
  • the similarity between the product and the candidate product words “belt”, “bicycle” and “exercise bike” is 0.5, 0.9 and 0.85, then 0.5 corresponds to the candidate product word “belt”, and 0.9 corresponds to the candidate product word “cycling”, 0.85 Corresponding to the candidate product word "fitness car”. If the preset threshold is 0.8, select "Bicycle” and “Workout” as the core product words of the product.
  • FIG. 2 is a schematic structural diagram of a Siamese network.
  • the first five layers of the Siamese network are Alexnet.
  • Each layer of Alexnet includes a convolution function and a pooling function.
  • the last three layers of the Siamese network are fully connected layers.
  • the final loss function uses the contrast loss function.
  • Alexnet is a classic network structure in the image field, including a multi-layer convolution structure.
  • the Siamese network After acquiring the training data, the Siamese network is trained. Moreover, applying the trained Siamese network to the commodity library for image feature extraction can obtain a dimensionality reduction representation of each image. Dimensionality reduction means that low-dimensional vectors are used to represent high-dimensional vectors. If the original image is directly stored, performance will be affected by taking up too much resources.
  • the Siamese network can extract features for a full-scale billion-level commodity, and feature extraction of incremental million-level data every day.
  • the Siamese network After obtaining the image features of each commodity in the commodity library through the Siamese network after training, the Siamese network needs to summarize the product image features to obtain the feature representation of all product words contained in the commodity library, and the process generally includes product words. Under the high-confidence commodity acquisition and multi-meaning product word clustering two steps.
  • the user feedback data search log
  • the user feedback data can be used to aggregate the click logs of the past nine months in the feedback data, and extract the products with sufficient clicks and high click rates, for example, the product clicks are in the top 200, and Products with a click rate of the top 50.
  • each product word may have multiple meanings, such as "apple", which can be a mobile phone or a fruit.
  • Each meaning item may have a picture form, such as the word "underwear”. The picture may be a piece of clothing or a boxed underwear.
  • the image under the product word can be clustered, and the category center is taken for each category (the weighted average of the vectors in the category, the weight can be the click rate of the product).
  • the clustering method is DBSCAN (Density-Based Spatial Clustering of Applications with Noise), the distance between vectors is Euclidean distance, the minimum distance between classes is set to 0.3, and the minimum number of samples per class is 4.
  • each meaning item can get a category center, and each category center is represented as an image feature of the product word. For each product word, multiple category centers are retained as their final image feature representation, and the number of reserved category centers can be set.
  • the core product words are confirmed in the products to be recognized for the core product words.
  • the image feature representation of each candidate product word can be obtained by using the trained Siamese network.
  • the trained Siamese network extracting image feature representations of the displayed images of the products of the core product words to be identified, and the trained Siamese network displays the image features of each candidate product word and the displayed image of the product of the core product word to be identified.
  • the image feature representation is compared to determine the similarity between the product of the core product word to be identified and the product represented by each candidate product word, that is, the similarity of the candidate product words corresponding to the product image and the product image set in the display image,
  • the similarity is obtained by calculating the inner product of the commodity display image vector of the core product word to be identified and the candidate product word vector.
  • the trained Siamese network can be obtained by inputting the display image of the product of the core product word to be recognized and the product image set of each candidate product word into the trained Siamese network.
  • the similarity between the item of the core product word to be identified and the item represented by each candidate product word is output.
  • candidate product words whose similarity is greater than a preset threshold are determined as core product words. Therefore, training the Siamese network makes the method of identifying the core product words of the embodiment of the present invention more convenient to implement, and can also make the core product words in the product title more quickly recognized.
  • the first task is to identify the intent of the search user and return accurate results. If you do not consider the product word (what product is sold by the product), only considering the text recall method will bring a lot of wrong results to the search user. The wrong product word recognition result will directly lead to the correct product cannot be recalled or the wrong result. Return to the search user.
  • the image information can be added to the judgment of the core product word, which is not only intuitive but also effective. In some cases, simply using textual information to distinguish core product words is difficult for people, and image recognition tends to be more intuitive. Applying images to core product word recognition plays an important role in improving the quality of the returned results, improving the user experience, and improving conversion.
  • FIG. 3 is a schematic diagram of a main module of a system for identifying a core product word according to an embodiment of the present invention.
  • the system 300 for identifying a core product word according to an embodiment of the present invention mainly includes:
  • the obtaining module 301 is configured to acquire a display image of the product, determine a plurality of candidate product words included in the title of the product, and determine a plurality of product image sets, each product image set and the plurality of candidate product words in the plurality of product image sets Each of the candidate product words corresponds one-to-one.
  • the obtaining module 301 determines a plurality of candidate product words included in the title of the product by using the conditional random field, and can accurately and quickly determine a plurality of candidate product words included in the title of the product.
  • the similarity determining module 302 is configured to determine, for each product image set in the plurality of product image sets, candidate product words corresponding to the product image and the product image set in the display image according to the display image and each image in the product image set. The similarity, resulting in multiple similarities.
  • the similarity determination module 302 includes a model training unit and a calculation unit for training the Siamese network based on the training data to obtain the trained Siamese network; the calculation unit is configured to determine the products in the display image through the trained Siamese network.
  • the similarity determination module 302 also includes a training data acquisition unit for acquiring training data.
  • the step of acquiring the training data by the training data acquiring unit includes: selecting a predetermined number of training product words; for each training product word, selecting two or more search words including the training product words according to the product search log; for each search word, according to each search word The number of clicks and the click rate of the product under the search term, and the same quantity of the product and the display image of the product are selected; for each training product word, the display image of the product belonging to the same search word under the training product word is composed of a pair of positive For example, the display image of the product belonging to different search words under the training product word is composed of a pair of negative examples. Furthermore, for each training product word, a pair of positive pairs and a pair of negative examples can be obtained, and the positive and negative pairs of all product words constitute training data.
  • the model training unit is further used for displaying images of each training product word in the training data, and clustering the display images of the training product words by a clustering algorithm to obtain more than one category center; using more than one category center as a collection of product images of the training product word; and saving the product image collection for each training product word to the image feature library.
  • the obtaining module determines a plurality of product image sets according to the image feature library, wherein each product image set of the plurality of product image sets is in one-to-one correspondence with each of the plurality of candidate product words.
  • the comparison module 303 is configured to determine a candidate product word corresponding to the similarity greater than the preset threshold among the plurality of similarities obtained by the similarity determining module as the core product word.
  • the system for identifying core product words in the embodiment of the present invention trains the Siamese network through the model training unit of the similarity determination module to obtain the trained Siamese network.
  • the obtaining module obtains a display image of the product of the core product word to be identified, determines a plurality of candidate product words included in the product of the core product word to be identified, and determines a product image set corresponding to each candidate product word
  • the calculation unit of the similarity determination module determines the similarity of the commodity of the core product word to be identified to each candidate product word through the trained Siamese network.
  • the comparison module determines candidate product words corresponding to the similarities greater than the preset threshold among the plurality of similarities as the core product words.
  • the system for identifying core product words in the embodiment of the present invention adds the image information of the product to the core product word judgment, and overcomes the problem that it is difficult to accurately determine the core product word in the title caused by simply using the text information in the prior art, thereby achieving The information of the product image improves the technical effect of the word recognition accuracy of the core product of the product.
  • FIG. 4 illustrates an exemplary system architecture 400 of a method of identifying core product words or a system for identifying core product words to which embodiments of the present invention may be applied.
  • system architecture 400 can include terminal devices 401, 402, 403, network 404, and server 405.
  • Network 404 is used to provide a medium for communication links between terminal devices 401, 402, 403 and server 405.
  • Network 404 can include a variety of connection types, such as wired, wireless communication links, fiber optic cables, and the like.
  • the user can interact with the server 405 via the network 404 using the terminal devices 401, 402, 403 to receive or send messages and the like.
  • Various communication client applications such as a shopping application, a web browser application, a search application, an instant communication tool, a mailbox client, a social platform software, and the like can be installed on the terminal devices 401, 402, and 403 (for example only).
  • the terminal devices 401, 402, 403 may be various electronic devices having a display screen and supporting web browsing, including but not limited to smartphones, tablets, laptop portable computers, desktop computers, and the like.
  • the server 405 may be a server that provides various services, such as a background management server (for example only) that provides support for a shopping site browsed by the user using the terminal devices 401, 402, 403.
  • the background management server may analyze and process data such as the received product information query request, and feed back the processing result (for example, target push information, product information--only examples) to the terminal device.
  • the method for identifying the core product words provided by the embodiments of the present invention is generally performed by the server 405. Accordingly, the system for identifying the core product words is generally disposed in the server 405.
  • terminal devices, networks, and servers in FIG. 4 is merely illustrative. Depending on the implementation needs, there can be any number of terminal devices, networks, and servers.
  • FIG. 5 there is shown a block diagram of a computer system 500 suitable for use in implementing a terminal device in accordance with an embodiment of the present invention.
  • the terminal device shown in FIG. 5 is merely an example, and should not impose any limitation on the function and scope of use of the embodiments of the present invention.
  • computer system 500 includes a central processing unit (CPU) 501 that can be loaded into a program in random access memory (RAM) 503 according to a program stored in read only memory (ROM) 502 or from storage portion 508. And perform various appropriate actions and processes.
  • RAM random access memory
  • ROM read only memory
  • RAM 503 various programs and data required for the operation of the system 500 are also stored.
  • the CPU 501, the ROM 502, and the RAM 503 are connected to each other through a bus 504.
  • An input/output (I/O) interface 505 is also coupled to bus 504.
  • the following components are connected to the I/O interface 505: an input portion 506 including a keyboard, a mouse, etc.; an output portion 507 including, for example, a cathode ray tube (CRT), a liquid crystal display (LCD), and the like, and a storage portion 508 including a hard disk or the like. And a communication portion 509 including a network interface card such as a LAN card, a modem, or the like. The communication section 509 performs communication processing via a network such as the Internet.
  • Driver 510 is also coupled to I/O interface 505 as needed.
  • a removable medium 511 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory or the like is mounted on the drive 510 as needed so that a computer program read therefrom is installed into the storage portion 508 as needed.
  • embodiments of the present disclosure include a computer program product comprising a computer program embodied on a computer readable medium, the computer program comprising program code for executing the method illustrated in the flowchart.
  • the computer program can be downloaded and installed from the network via the communication portion 509, and/or installed from the removable medium 511.
  • CPU central processing unit
  • the computer readable medium shown in the present invention may be a computer readable signal medium or a computer readable storage medium or any combination of the two.
  • the computer readable storage medium can be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples of computer readable storage media may include, but are not limited to, electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read only memory (ROM), erasable Programmable read only memory (EPROM or flash memory), optical fiber, portable compact disk read only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the foregoing.
  • a computer readable storage medium may be any tangible medium that can contain or store a program, which can be used by or in connection with an instruction execution system, apparatus or device.
  • a computer readable signal medium may include a data signal that is propagated in the baseband or as part of a carrier, in which computer readable program code is carried. Such propagated data signals can take a variety of forms including, but not limited to, electromagnetic signals, optical signals, or any suitable combination of the foregoing.
  • the computer readable signal medium can also be any computer readable medium other than a computer readable storage medium, which can transmit, propagate, or transport a program for use by or in connection with the instruction execution system, apparatus, or device.
  • Program code embodied on a computer readable medium can be transmitted by any suitable medium, including but not limited to wireless, wire, fiber optic cable, RF, etc., or any suitable combination of the foregoing.
  • each block of the flowchart or block diagrams can represent a module, a program segment, or a portion of code that includes one or more Executable instructions.
  • the functions noted in the blocks may also occur in a different order than that illustrated in the drawings. For example, two successively represented blocks may in fact be executed substantially in parallel, and they may sometimes be executed in the reverse order, depending upon the functionality involved.
  • each block of the block diagrams or flowcharts, and combinations of blocks in the block diagrams or flowcharts can be implemented by a dedicated hardware-based system that performs the specified function or operation, or can be used A combination of dedicated hardware and computer instructions is implemented.
  • the modules involved in the embodiments of the present invention may be implemented by software or by hardware.
  • the described modules may also be provided in the processor, for example, as a processor including an acquisition module, a similarity determination module, and a comparison module.
  • the names of these modules do not constitute a limitation on the module itself under certain circumstances.
  • the comparison module may also be described as a module that determines a candidate product word whose similarity is greater than a preset threshold as a core product word. .
  • the present invention also provides a computer readable medium, which may be included in the apparatus described in the above embodiments, or may be separately present and not incorporated in the apparatus.
  • the computer readable medium carries one or more programs.
  • the device includes: obtaining a display image of the product, and determining a plurality of candidate product words included in the title of the product.
  • each product image set of the plurality of product image sets being in one-to-one correspondence with each of the plurality of candidate product words; for each product image set of the plurality of product image sets, Determining, according to the display image and each image in the product image set, a similarity of the candidate product words corresponding to the product image set in the display image, thereby obtaining a plurality of similarities; and comparing the plurality of similarities to be greater than a preset threshold
  • the candidate product words corresponding to the similarity are determined as core product words.
  • the image information can be added to the judgment of the core product word, which is not only intuitive but also effective. Applying images to core product word recognition is important for improving the quality of returned results, improving the user experience, and improving conversions.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • Finance (AREA)
  • Accounting & Taxation (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • General Business, Economics & Management (AREA)
  • Strategic Management (AREA)
  • Marketing (AREA)
  • Economics (AREA)
  • Development Economics (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Biophysics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Image Analysis (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

本发明公开了一种识别核心产品词的方法和系统,涉及计算机技术领域。该方法的一具体实施方式包括:获取商品的展示图像,确定所述商品的标题中包含的多个候选产品词,以及确定多个产品图像集合,多个产品图像集合中的各产品图像集合与多个候选产品词中的各候选产品词一一对应;对于所述多个产品图像集合中的每个产品图像集合,根据所述展示图像和该产品图像集合中的各个图像,确定所述展示图像中的商品与所述产品图像集合对应的候选产品词的相似度,从而得到多个相似度;将所述多个相似度中大于预设阈值的相似度所对应的候选产品词确定为核心产品词。该实施方式将图像信息加入核心产品词的识别,使得更直观且精准的确定出商品标题的核心产品词。

Description

一种识别核心产品词的方法和系统 技术领域
本发明涉及计算机技术领域,尤其涉及一种识别核心产品词的方法和系统。
背景技术
在电商领域中,准确的分析商品标题的成分,是用户意图识别、产品召回、个性化推荐等的基础。区别于一般自然语言分析,在电商场景中,需要区分商品标题中的品牌词、修饰词、产品词等成分。而且,在电商领域中,很多商品卖家为提高商品的点击量,会在商品标题中堆砌罗列产品词,甚至其中很多产品词并非是对应本商品。在卖家将商品上架之后,为了能将商品精准的提供给买家,需识别出商品标题中的核心产品词。其中,商品标题的核心产品词就是指标题描述的商品具体是什么产品。如:“韩版NY洋基队男女款鸭舌帽”的核心产品词是“鸭舌帽”。
由于商品标题具有语义不清晰、产品词堆砌等特点,所以,如何识别出商品标题中的核心产品词一直是一个难题。现有技术中,识别商品标题中的核心产品词的方法主要包括:基于词表的规则方法、基于条件随机场的序列标注方法和基于LSTM的深度学习方法。基于词表的规则方法是指维护一个映射的词表。例如:空调滤芯-空调,当标题中同时出现“空调”和“滤芯”时,则认为“滤芯”是核心产品词。基于条件随机场的序列标注方法和基于LSTM的深度学习方法都是利用机器学习模型,通过大量语料进行训练,用学到的模型对新的标题中的产品词是否为核心产品词进行预测。
但是,现有技术中识别商品标题中的核心产品词的方法都只利用 了文本信息。具体的,词表的方法不具备泛化能力,即无法处理不包含在词表中的新样本,而商品标题种类繁多,不可能将所有情况都包含进词表。在电商场景下,标题中产品词堆砌现象严重且难以利用句法信息解析,所以机器学习的方法也不能非常有效地识别出商品标题中的核心产品词。并且,商家多种多样不规范的标题书写形式,使得单纯利用文本信息,难以判断标题中的核心产品词。
发明内容
有鉴于此,本发明实施例提供一种识别核心产品词的方法和系统,能够更有效直观、精准的确定出商品标题的核心产品词。进而,可提高用户返回结果的质量,改进用户体验、提高转化。
为实现上述目的,根据本发明实施例的一个方面,提供了一种识别核心产品词的方法。
本发明实施例的识别核心产品词的方法包括:获取商品的展示图像,确定所述商品的标题中包含的多个候选产品词,以及确定多个产品图像集合,所述多个产品图像集合中的各产品图像集合与所述多个候选产品词中的各候选产品词一一对应;对于所述多个产品图像集合中的每个产品图像集合,根据所述展示图像和该产品图像集合中的各个图像,确定所述展示图像中的商品与所述产品图像集合对应的候选产品词的相似度,从而得到多个相似度;将所述多个相似度中大于预设阈值的相似度所对应的候选产品词确定为核心产品词。
可选地,所述确定所述展示图像中的商品与所述产品图像集合对应的候选产品词的相似度的步骤包括:基于训练数据训练Siamese网络,以得到训练后的Siamese网络;对于所述多个产品图像集合中的每个产品图像集合,通过训练后的Siamese网络根据所述展示图像和该产品图像集合中的各个图像,确定所述展示图像中的商品与所述产品图像集合对应的候选产品词的相似度,从而得到多个相似度。
可选地,在基于训练数据训练Siamese网络,以得到训练后的Siamese网络之前,还包括:选取预定数量的训练产品词;对于每个训练产品词,根据商品搜索日志选取两个以上包含该训练产品词的搜索词;对于每个搜索词,根据该搜索词下商品的点击量以及点击率,选取相同数量的商品以及该商品的展示图像;对于每个训练产品词,将该训练产品词下属于同一搜索词的商品的展示图像组成一对正例对,将该训练产品词下属于不同搜索词的商品的展示图像组成一对负例对。
可选地,所述基于训练数据训练Siamese网络的步骤包括:对于训练数据中每个训练产品词的展示图像,通过聚类算法对训练产品词的展示图像进行聚类处理,以获得一个以上的类别中心;将所述一个以上的类别中心作为该训练产品词的产品图像集合;并且,将每个训练产品词的产品图像集合保存至图像特征库;所述确定多个产品图像集合的步骤包括:根据所述图像特征库确定多个产品图像集合,所述多个产品图像集合中的各产品图像集合与所述多个候选产品词中的各候选产品词一一对应。
可选地,利用条件随机场确定所述商品的标题中包含的多个候选产品词。
为实现上述目的,根据本发明实施例的另一个方面,提供了一种识别核心产品词的系统。
本发明实施例的识别核心产品词的系统包括:获取模块,用于获取商品的展示图像,确定所述商品的标题中包含的多个候选产品词,以及确定多个产品图像集合,所述多个产品图像集合中的各产品图像集合与所述多个候选产品词中的各候选产品词一一对应;相似度确定模块,用于对于所述多个产品图像集合中的每个产品图像集合,根据 所述展示图像和该产品图像集合中的各个图像,确定所述展示图像中的商品与所述产品图像集合对应的候选产品词的相似度,从而得到多个相似度;比较模块,用于将所述相似度确定模块得到的多个相似度中大于预设阈值的相似度所对应的候选产品词确定为核心产品词。
可选地,所述相似度确定模块包括:模型训练单元,用于基于训练数据训练Siamese网络,以得到训练后的Siamese网络;计算单元,用于通过训练后的Siamese网络确定所述展示图像中的商品与所述产品图像集合对应的候选产品词的相似度。
可选地,所述相似度确定模块还包括训练数据获取单元,用于选取预定数量的训练产品词;对于每个训练产品词,根据商品搜索日志选取两个以上包含该训练产品词的搜索词;对于每个搜索词,根据该搜索词下商品的点击量以及点击率,选取相同数量的商品以及该商品的展示图像;对于每个训练产品词,将该训练产品词下属于同一搜索词的商品的展示图像组成一对正例对,将该训练产品词下属于不同搜索词的商品的展示图像组成一对负例对。
可选地,模型训练单元还用于对于训练数据中每个训练产品词的展示图像,通过聚类算法对训练产品词的展示图像进行聚类处理,以获得一个以上的类别中心;将所述一个以上的类别中心作为该训练产品词的产品图像集合;以及,将每个训练产品词的产品图像集合保存至图像特征库;所述获取模块根据所述图像特征库确定多个产品图像集合,所述多个产品图像集合中的各产品图像集合与所述多个候选产品词中的各候选产品词一一对应。
可选地,所述获取模块利用条件随机场确定所述商品的标题中包含的多个候选产品词。
为实现上述目的,根据本发明实施例的再一个方面,提供了一种 识别核心产品词的电子设备。
本发明实施例的识别核心产品词的电子设备包括:一个或多个处理器;存储装置,用于存储一个或多个程序,当所述一个或多个程序被所述一个或多个处理器执行,使得所述一个或多个处理器实现上述识别核心产品词的方法。
为实现上述目的,根据本发明实施例的再一个方面,提供了一种计算机可读介质,其上存储有计算机程序,所述程序被处理器执行时实现上述识别核心产品词的方法。
上述发明中的一个实施例具有如下优点或有益效果:因为将商品的图像信息加入核心产品词的判断,所以克服了现有技术中单纯利用文本信息导致的难以准确判断商品标题中的核心产品词的问题,进而达到借助商品图像的信息,提高商品核心产品词识别准确率的技术效果。
上述的非惯用的可选方式所具有的进一步效果将在下文中结合具体实施方式加以说明。
附图说明
附图用于更好地理解本发明,不构成对本发明的不当限定。其中:
图1是根据本发明实施例的识别核心产品词的方法的主要流程的示意图;
图2是Siamese网络的结构示意图;
图3是根据本发明实施例的识别核心产品词的系统的主要模块的示意图;
图4是本发明实施例可以应用于其中的示例性系统架构图;
图5是适于用来实现本发明实施例的终端设备或服务器的计算机系统的结构示意图。
具体实施方式
以下结合附图对本发明的示范性实施例做出说明,其中包括本发明实施例的各种细节以助于理解,应当将它们认为仅仅是示范性的。因此,本领域普通技术人员应当认识到,可以对这里描述的实施例做出各种改变和修改,而不会背离本发明的范围和精神。同样,为了清楚和简明,以下的描述中省略了对公知功能和结构的描述。
图1是根据本发明实施例的识别核心产品词的方法的主要流程的示意图,如图1所示,本发明实施例的识别核心产品词的方法主要包括:
步骤S101:获取商品的展示图像,确定商品的标题中包含的多个候选产品词,以及确定多个产品图像集合,多个产品图像集合中的各产品图像集合与多个候选产品词中的各候选产品词一一对应。其中,商品的展示图像是指在搜索该商品时,展示在搜索页面的图像,该展示图像是由卖家提供并设置的,可充分体现该商品,因此根据商品的展示图像能更精准、快速的识别出标题中的核心产品词。在步骤S101中,可利用条件随机场(CRF)确定商品的标题中包含的多个候选产品词。CRF是一种序列标注模型,可以有效准确的对分词、词性进行标注以及进行命名实体识别等。
每个候选产品词对应一个产品图像集合,该产品图像集合中的图像可反映出该候选产品词。并且,对于每一个候选产品词的产品图像集合,该产品图像集合中包含有一个以上的图像,根据该产品图像集合中的图像,可确定出该候选产品词对应的商品。因此,在步骤S102确定相似度时,可通过将该产品图像集合中的图像与商品的展示图像进行对比,确定出该产品图像集合对应的候选产品词与商品相似的相似度,本发明实施例中的商品即为待识别核心产品词的商品。
步骤S102:对于多个产品图像集合中的每个产品图像集合,根据 展示图像和该产品图像集合中的各个图像,确定展示图像中的商品与产品图像集合对应的候选产品词的相似度,从而得到多个相似度。本步骤中确定出的展示图像中的商品与产品图像集合对应的候选产品词的相似度,可体现出待识别核心产品词的商品与候选产品词所代表的产品的相似程度。
在确定展示图像中的商品与产品图像集合对应的候选产品词的相似度的过程中,由于一个产品图像集合中会存在很多图像,因此,该产品集合中的每一个图像与待识别核心产品词的商品的展示图像进行比较,都会得出一个相似度。但是,对于某一个候选产品词的产品图像集合而言,将该产品图像集合中所有图像与待识别核心产品词的商品的展示图像进行比较后得出的最大相似度,作为该候选产品词与待识别核心产品词的商品的相似度。例如,针对待识别核心产品词的商品A,其商品标题为“时尚休闲商用动感单车皮带商用动感健身车车”,确定出的候选产品词“皮带”、“单车”、“健身车”之后,分别获取候选产品词“皮带”的产品图像集合、候选产品词“单车”的产品图像集合、候选产品词“健身车”的产品图像集合。候选产品词“皮带”的产品图像集合中包括5个图像,将这5个图像分别与商品A的展示图像进行对比,得出合的相似度为0.2、0.5、0.3、0.1、0.4,进而该产品图像集合对应的候选产品词“皮带”与商品A的相似度为0.5。同样,确定出商品A与候选产品词“单车”的相似度为0.9,商品A与候选产品词“健身车”的相似度为0.85。所以,经过上述过程之后,得出三个相似度,分别为0.5、0.9和0.85。
为使得本发明实施例的识别核心产品词的方法在实际操作中更易于实施和便于操作,步骤S102的过程可通过训练后的模型进行。所以,本发明实施例的识别核心产品词的方法还包括:基于训练数据训练Siamese网络,以得到训练后的Siamese网络。其中,Siamese网络是一个多分支参数共享的网络结构,是一种相似性度量方法。这个网络的原理利用神经网络提取描述算子,得到特征向量,然后利用两个图 片的特征向量判断相似度,其优势是可以去区分那些新的没有经过训练的类别的样本。因为Siamese网络从数据中去学习一个相似性度量(两个事物越接近,它们的相似性度量也就越大,而两个事物越疏远,它们的相似性度量也就越小),用这个学习出来的度量去比较和匹配新的未知类别的样本。
本发明实施例的识别核心产品词的方法还包括获取训练数据。在获取训练数据的过程中,选取预定数量的训练产品词;基于商品搜索日志,为每个训练产品词选取两个以上包含该训练产品词的搜索词;根据搜索词下商品的点击量以及点击率,为每个搜索词选取相同数量的商品以及该商品的展示图像;对于每个训练产品词,将该训练产品词下属于同一搜索词的商品的展示图像组成一对正例对,将该产品词下属于不同搜索词的商品的展示图像组成一对负例对。对于每个训练产品词,都可获取到一对正例对和一对负例对,所有产品词的正例对和负例对构成了训练数据。因此,基于获取的正例对和负例对训练Siamese网络。基于以下假设对Siamese网络进行训练:同样的产品词下商品相似,不同产品词下商品不相似。虽然不同产品词下的商品可能相同,但是在训练过程中大相似度的情况下还是不相似,比如选定6000个训练产品词,“制冷机”和不同的产品词“空调”相似的商品如果有5个,那有可能的误差也只有5/6000,进而假设同样的产品词下商品相似,不同产品词下商品不相似是可行的。例如,选定6000个训练产品词;对于每个训练产品词,选取两个包含该训练产品词的搜索词,共12000个;每个搜索词下选取点击量及点击率最高20个商品,并对应地获取该商品的展示图像,共24万张图像。对于每个训练产品词下的某一个商品的展示图像,可选取同一搜索词下的另外一个商品的展示图像组成一对正例对,选取其它搜索词下的商品的展示图像组成一对负例对,最终得到48万对训练数据。
基于上述获取的训练数据,对Siamese网络进行训练得到训练后的Siamese网络。可将该训练后的Siamese网络应用于商品库中的所有商 品的展示图片,并且每个商品的核心产品词是确定的。进而,可得到一个以上的产品图像集合,每一个产品图像集合对应一个产品词,并且该产品集合中包含着一个以上的图像或者该一个以上图像的图像特征。
本发明实施例在对Siamese网络进行训练的过程中,可能会针对多义产品词(产品词有多个理解),可通过聚类算法(DBSACN)对所有产品词的展示图像进行处理,获得一个以上的类别中心,并将一个以上的类别中心作为该产品词的产品图像集合。其中,DBSACN是一种基于密度的聚类算法,将高密度区间划分为簇,并可在噪声的空间数据库中发现任意形状。在将一个以上的类别中心作为该产品词的产品图像集合之后,将每个产品词的产品图像集合保存至图像特征库。所以在确定每个候选产品词对应的产品图像集合时,可根据该图像特征库确定出多个产品图像集合。
而且,对于多个产品图像集合中的每个产品图像集合,通过训练后的Siamese网络根据展示图像和该产品图像集合中的各个图像,可确定出展示图像中的商品与产品图像集合对应的候选产品词的相似度,从而得到多个相似度。
步骤S103:将多个相似度中大于预设阈值的相似度所对应的候选产品词确定为核心产品词。通过步骤S102得到的相似度为展示图像中的商品与产品图像集合对应的候选产品词的相似度,对于每一个相似度,其都与确定该相似度的候选产品词相对应。因此,在得出多个相似度后,每个相似度对应着一个候选产品词。并且,每个相似度体现着一个候选产品词代表的商品与待识别核心产品词的商品的相似程度。在得出对应着每个候选产品词的相似度后,可根据预设阈值将一个或多个候选产品词确定为核心产品词。例如,针对商品标题“时尚休闲商用动感单车皮带商用动感健身车车”,候选产品词为“皮带”、“单车”、“健身车”,通过商品的展示图片以及候选产品词的产品 图像集合,得到该商品分别与候选产品词“皮带”、“单车”、“健身车”的相似度为0.5、0.9和0.85,则0.5对应候选产品词“皮带”,0.9对应候选产品词“单车”,0.85对应候选产品词“健身车”。如果预设阈值为0.8,则选取“单车”、“健身车”作为该商品的核心产品词。
图2是Siamese网络的结构示意图。由图2可知,Siamese网络的前五层为Alexnet,Alexnet的每层包括一个卷积函数和一个池化函数,Siamese网络的后三层为全连接层,最后损失函数选用对比损失函数。其中,Alexnet为图像领域的一种经典网络结构,包含多层卷积结构。
获取到训练数据之后,对Siamese网络进行训练。并且,将训练后的Siamese网络应用于商品库以进行图像特征提取,可得到每个图像的降维表示。降维表示就是用低维的向量去表示高维向量,如果直接存储原始图像,会因为占用资源太多影响性能。获取到商品库中每个商品的展示图像后,将展示图像维度转化为220*220*3后放入训练后的Siamese网络进行图像特征提取。该过程中的图像特征提取就是将原始图片输入训练后的Siamese网络,逐层计算,最终得到一个20维的向量作为其降维表示,原始图片维度为220*220*3=145200。该训练后Siamese网络可对全量亿级商品提取特征,并且可每天对增量的百万级数据进行特征提取。
通过训练后Siamese网络得到商品库中每个商品的图像特征后,该训练后Siamese网络需要将商品图像特征进行汇总,以得到商品库中包含的所有产品词的特征表示,该过程一般包括产品词下高置信商品获取和多义产品词聚类两步。
对于产品词下高置信商品获取,由于商品库中某一产品词的商品很多,所以需要获取确实与产品词一致的高置信商品来进行聚合。在该过程中,可利用用户反馈数据(搜索日志),聚合反馈数据中过去 九个月的点击日志,抽取点击量充分且点击率高的商品,例如在产品词下点击量位于前200,且点击率在前50的商品。
对于多义产品词聚类,是因为每个产品词可能有多种义项,如“苹果”即可以是一种手机,也可以是一种水果。每种义项可能有一种图片形式,如“内衣”这一产品词,图片可能为一件衣服,也可能为一个盒装内衣。针对产品词有多种义项的情况,可对产品词下图像进行聚类,对每个类别取类别中心(该类别内的向量加权平均,权重可为商品的点击率)。聚类方法为DBSCAN(Density-Based Spatial Clustering of Applications with Noise),向量间距离采用欧氏距离,类间最小距离设为0.3,每类最少样本数为4。聚类之后,每个义项可得到一个类别中心,将每个类别中心作为该产品词的图像特征表示。对于每个产品词,保留多个类别中心作为其最终的图像特征表示,保留的类别中心的个数可进行设置。
对Siamese网络训练,以及利用训练后的Siamese网络获取商品库中包含的产品词和每个产品词的图像特征表示(产品图像集合)后,在对待识别核心产品词的商品进行核心产品词进行确认时,通过CRF确定该商品标题中包含的候选产品词后,利用训练后的Siamese网络可获取每个候选产品词的图像特征表示。以及,通过训练后的Siamese网络提取待识别核心产品词的商品的展示图像的图像特征表示,训练后的Siamese网络将每个候选产品词的图像特征表示与待识别核心产品词的商品的展示图像的图像特征表示进行比对,确定出待识别核心产品词的商品与每个候选产品词代表的商品的相似度,即展示图像中的商品与产品图像集合对应的候选产品词的相似度,该相似度是通过计算待识别核心产品词的商品展示图像向量和候选产品词向量的内积得出的。
所以,基于训练数据对Siamese网络训练完成后,只要将待识别核心产品词的商品的展示图像与每个候选产品词的产品图像集合输入该 训练后的Siamese网络,该训练后的Siamese网络即可输出待识别核心产品词的商品与每个候选产品词代表的商品的相似度。进一步,将相似度大于预设阈值的候选产品词确定为核心产品词。因此,对Siamese网络进行训练,使得本发明实施例的识别核心产品词的方法更便于实施,也可使得更加快速的识别出商品标题中的核心产品词。
对于搜索而言,首要任务就是识别出搜索用户的意图,并返回准确的结果。如果不考虑产品词(商品所卖的是什么产品)只考虑文本的召回方式将给搜索用户带来大量的错误结果,错误的产品词识别结果也将直接导致正确商品不能被召回或者把错误结果返回给搜索用户。对于本发明实施例的识别核心产品词的方法,可将图像信息加入核心产品词的判断,该方式不仅直观而且有效。在一些情况下,如果单纯利用文本信息来区分核心产品词,对人来说都是很困难的,而图像识别往往更直观。将图像应用于核心产品词识别对提高返回结果的质量,改进用户体验、提高转化有重要作用。
图3是根据本发明实施例的识别核心产品词的系统的主要模块的示意图,如图3所示,本发明实施例的识别核心产品词的系统300主要包括:
获取模块301,用于获取商品的展示图像,确定商品的标题中包含的多个候选产品词,以及确定多个产品图像集合,多个产品图像集合中的各产品图像集合与多个候选产品词中的各候选产品词一一对应。获取模块301利用条件随机场确定商品的标题中包含的多个候选产品词,可准确快速的将商品的标题中包含的多个候选产品词确定出。
相似度确定模块302,用于对于多个产品图像集合中的每个产品图像集合,根据展示图像和该产品图像集合中的各个图像,确定展示图像中的商品与产品图像集合对应的候选产品词的相似度,从而得到多个相似度。相似度确定模块302包括模型训练单元和计算单元,该模型训练单元用于基于训练数据训练Siamese网络,以得到训练后的 Siamese网络;计算单元用于通过训练后的Siamese网络确定展示图像中的商品与产品图像集合对应的候选产品词的相似度。相似度确定模块302还包括用于获取训练数据的训练数据获取单元。训练数据获取单元获取训练数据的步骤包括:选取预定数量的训练产品词;对于每个训练产品词,根据商品搜索日志选取两个以上包含该训练产品词的搜索词;对于每个搜索词,根据该搜索词下商品的点击量以及点击率,选取相同数量的商品以及该商品的展示图像;对于每个训练产品词,将该训练产品词下属于同一搜索词的商品的展示图像组成一对正例对,将该训练产品词下属于不同搜索词的商品的展示图像组成一对负例对。进而,对于每个训练产品词,都可获取到一对正例对和一对负例对,所有产品词的正例对和负例对构成了训练数据。
模型训练单元还用于对于训练数据中每个训练产品词的展示图像,通过聚类算法对训练产品词的展示图像进行聚类处理,以获得一个以上的类别中心;将一个以上的类别中心作为该训练产品词的产品图像集合;以及,将每个训练产品词的产品图像集合保存至图像特征库。进而,获取模块根据该图像特征库确定多个产品图像集合,其中,多个产品图像集合中的各产品图像集合与多个候选产品词中的各候选产品词一一对应。
比较模块303,用于将相似度确定模块得到的多个相似度中大于预设阈值的相似度所对应的候选产品词确定为核心产品词。
本发明实施例的识别核心产品词的系统通过相似度确定模块的模型训练单元对Siamese网络进行训练,以得到训练后的Siamese网络。在获取模块获取到待识别核心产品词的商品的展示图片,确定出待识别核心产品词的商品的中包含的多个候选产品词,以及确定出每个候选产品词对应的产品图像集合后,相似度确定模块的计算单元通过训练后的Siamese网络确定待识别核心产品词的商品与每个候选产品词的相似度。比较模块将多个相似度中大于预设阈值的相似度所对应的 候选产品词确定为核心产品词。本发明实施例的识别核心产品词的系统将商品的图像信息加入核心产品词的判断,克服了现有技术中单纯利用文本信息导致的难以准确判断标题中的核心产品词的问题,进而达到借助商品图像的信息,提高商品核心产品词识别准确率的技术效果。
图4示出了可以应用本发明实施例的识别核心产品词的方法或识别核心产品词的系统的示例性系统架构400。
如图4所示,系统架构400可以包括终端设备401、402、403,网络404和服务器405。网络404用以在终端设备401、402、403和服务器405之间提供通信链路的介质。网络404可以包括各种连接类型,例如有线、无线通信链路或者光纤电缆等等。
用户可以使用终端设备401、402、403通过网络404与服务器405交互,以接收或发送消息等。终端设备401、402、403上可以安装有各种通讯客户端应用,例如购物类应用、网页浏览器应用、搜索类应用、即时通信工具、邮箱客户端、社交平台软件等(仅为示例)。
终端设备401、402、403可以是具有显示屏并且支持网页浏览的各种电子设备,包括但不限于智能手机、平板电脑、膝上型便携计算机和台式计算机等等。
服务器405可以是提供各种服务的服务器,例如对用户利用终端设备401、402、403所浏览的购物类网站提供支持的后台管理服务器(仅为示例)。后台管理服务器可以对接收到的产品信息查询请求等数据进行分析等处理,并将处理结果(例如目标推送信息、产品信息--仅为示例)反馈给终端设备。
需要说明的是,本发明实施例所提供的识别核心产品词的方法一 般由服务器405执行,相应地,识别核心产品词的系统一般设置于服务器405中。
应该理解,图4中的终端设备、网络和服务器的数目仅仅是示意性的。根据实现需要,可以具有任意数目的终端设备、网络和服务器。
下面参考图5,其示出了适于用来实现本发明实施例的终端设备的计算机系统500的结构示意图。图5示出的终端设备仅仅是一个示例,不应对本发明实施例的功能和使用范围带来任何限制。
如图5所示,计算机系统500包括中央处理单元(CPU)501,其可以根据存储在只读存储器(ROM)502中的程序或者从存储部分508加载到随机访问存储器(RAM)503中的程序而执行各种适当的动作和处理。在RAM 503中,还存储有系统500操作所需的各种程序和数据。CPU 501、ROM 502以及RAM 503通过总线504彼此相连。输入/输出(I/O)接口505也连接至总线504。
以下部件连接至I/O接口505:包括键盘、鼠标等的输入部分506;包括诸如阴极射线管(CRT)、液晶显示器(LCD)等以及扬声器等的输出部分507;包括硬盘等的存储部分508;以及包括诸如LAN卡、调制解调器等的网络接口卡的通信部分509。通信部分509经由诸如因特网的网络执行通信处理。驱动器510也根据需要连接至I/O接口505。可拆卸介质511,诸如磁盘、光盘、磁光盘、半导体存储器等等,根据需要安装在驱动器510上,以便于从其上读出的计算机程序根据需要被安装入存储部分508。
特别地,根据本发明公开的实施例,上文参考流程图描述的过程可以被实现为计算机软件程序。例如,本发明公开的实施例包括一种计算机程序产品,其包括承载在计算机可读介质上的计算机程序,该计算机程序包含用于执行流程图所示的方法的程序代码。在这样的实 施例中,该计算机程序可以通过通信部分509从网络上被下载和安装,和/或从可拆卸介质511被安装。在该计算机程序被中央处理单元(CPU)501执行时,执行本发明的系统中限定的上述功能。
需要说明的是,本发明所示的计算机可读介质可以是计算机可读信号介质或者计算机可读存储介质或者是上述两者的任意组合。计算机可读存储介质例如可以是——但不限于——电、磁、光、电磁、红外线、或半导体的系统、装置或器件,或者任意以上的组合。计算机可读存储介质的更具体的例子可以包括但不限于:具有一个或多个导线的电连接、便携式计算机磁盘、硬盘、随机访问存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(EPROM或闪存)、光纤、便携式紧凑磁盘只读存储器(CD-ROM)、光存储器件、磁存储器件、或者上述的任意合适的组合。在本发明中,计算机可读存储介质可以是任何包含或存储程序的有形介质,该程序可以被指令执行系统、装置或者器件使用或者与其结合使用。而在本发明中,计算机可读的信号介质可以包括在基带中或者作为载波一部分传播的数据信号,其中承载了计算机可读的程序代码。这种传播的数据信号可以采用多种形式,包括但不限于电磁信号、光信号或上述的任意合适的组合。计算机可读的信号介质还可以是计算机可读存储介质以外的任何计算机可读介质,该计算机可读介质可以发送、传播或者传输用于由指令执行系统、装置或者器件使用或者与其结合使用的程序。计算机可读介质上包含的程序代码可以用任何适当的介质传输,包括但不限于:无线、电线、光缆、RF等等,或者上述的任意合适的组合。
附图中的流程图和框图,图示了按照本发明各种实施例的系统、方法和计算机程序产品的可能实现的体系架构、功能和操作。在这点上,流程图或框图中的每个方框可以代表一个模块、程序段、或代码的一部分,上述模块、程序段、或代码的一部分包含一个或多个用于实现规定的逻辑功能的可执行指令。也应当注意,在有些作为替换的实现中,方框中所标注的功能也可以以不同于附图中所标注的顺序发 生。例如,两个接连地表示的方框实际上可以基本并行地执行,它们有时也可以按相反的顺序执行,这依所涉及的功能而定。也要注意的是,框图或流程图中的每个方框、以及框图或流程图中的方框的组合,可以用执行规定的功能或操作的专用的基于硬件的系统来实现,或者可以用专用硬件与计算机指令的组合来实现。
描述于本发明实施例中所涉及到的模块可以通过软件的方式实现,也可以通过硬件的方式来实现。所描述的模块也可以设置在处理器中,例如,可以描述为:一种处理器包括获取模块、相似度确定模块和比较模块。其中,这些模块的名称在某种情况下并不构成对该模块本身的限定,例如,比较模块还可以被描述为“将相似度大于预设阈值的候选产品词确定为核心产品词的模块”。
作为另一方面,本发明还提供了一种计算机可读介质,该计算机可读介质可以是上述实施例中描述的设备中所包含的;也可以是单独存在,而未装配入该设备中。上述计算机可读介质承载有一个或者多个程序,当上述一个或者多个程序被一个该设备执行时,使得该设备包括:获取商品的展示图像,确定商品的标题中包含的多个候选产品词,以及确定多个产品图像集合,多个产品图像集合中的各产品图像集合与多个候选产品词中的各候选产品词一一对应;对于多个产品图像集合中的每个产品图像集合,根据展示图像和该产品图像集合中的各个图像,确定展示图像中的商品与产品图像集合对应的候选产品词的相似度,从而得到多个相似度;将多个相似度中大于预设阈值的相似度所对应的候选产品词确定为核心产品词。
根据本发明实施例的技术方案,可将图像信息加入核心产品词的判断,该方式不仅直观而且有效。将图像应用于核心产品词识别对提高返回结果的质量、改进用户体验以及提高转化都有重要作用。
上述具体实施方式,并不构成对本发明保护范围的限制。本领域 技术人员应该明白的是,取决于设计要求和其他因素,可以发生各种各样的修改、组合、子组合和替代。任何在本发明的精神和原则之内所作的修改、等同替换和改进等,均应包含在本发明保护范围之内。

Claims (12)

  1. 一种识别核心产品词的方法,其特征在于,包括:
    获取商品的展示图像,确定所述商品的标题中包含的多个候选产品词,以及确定多个产品图像集合,所述多个产品图像集合中的各产品图像集合与所述多个候选产品词中的各候选产品词一一对应;
    对于所述多个产品图像集合中的每个产品图像集合,根据所述展示图像和该产品图像集合中的各个图像,确定所述展示图像中的商品与所述产品图像集合对应的候选产品词的相似度,从而得到多个相似度;
    将所述多个相似度中大于预设阈值的相似度所对应的候选产品词确定为核心产品词。
  2. 根据权利要求1所述的方法,其特征在于,所述确定所述展示图像中的商品与所述产品图像集合对应的候选产品词的相似度的步骤包括:
    基于训练数据训练Siamese网络,以得到训练后的Siamese网络;
    对于所述多个产品图像集合中的每个产品图像集合,通过训练后的Siamese网络根据所述展示图像和该产品图像集合中的各个图像,确定所述展示图像中的商品与所述产品图像集合对应的候选产品词的相似度,从而得到多个相似度。
  3. 根据权利要求2所述的方法,其特征在于,在基于训练数据训练Siamese网络,以得到训练后的Siamese网络之前,还包括:
    选取预定数量的训练产品词;
    对于每个训练产品词,根据商品搜索日志选取两个以上包含该训练产品词的搜索词;
    对于每个搜索词,根据该搜索词下商品的点击量以及点击率,选取相同数量的商品以及该商品的展示图像;
    对于每个训练产品词,将该训练产品词下属于同一搜索词的商品 的展示图像组成一对正例对,将该训练产品词下属于不同搜索词的商品的展示图像组成一对负例对。
  4. 根据权利要求2所述的方法,其特征在于,
    所述基于训练数据训练Siamese网络的步骤包括:对于训练数据中每个训练产品词的展示图像,通过聚类算法对训练产品词的展示图像进行聚类处理,以获得一个以上的类别中心;将所述一个以上的类别中心作为该训练产品词的产品图像集合;并且,将每个训练产品词的产品图像集合保存至图像特征库;
    所述确定多个产品图像集合的步骤包括:根据所述图像特征库确定多个产品图像集合,所述多个产品图像集合中的各产品图像集合与所述多个候选产品词中的各候选产品词一一对应。
  5. 根据权利要求1所述的方法,其特征在于,利用条件随机场确定所述商品的标题中包含的多个候选产品词。
  6. 一种识别核心产品词的系统,其特征在于,包括:
    获取模块,用于获取商品的展示图像,确定所述商品的标题中包含的多个候选产品词,以及确定多个产品图像集合,所述多个产品图像集合中的各产品图像集合与所述多个候选产品词中的各候选产品词一一对应;
    相似度确定模块,用于对于所述多个产品图像集合中的每个产品图像集合,根据所述展示图像和该产品图像集合中的各个图像,确定所述展示图像中的商品与所述产品图像集合对应的候选产品词的相似度,从而得到多个相似度;
    比较模块,用于将所述相似度确定模块得到的多个相似度中大于预设阈值的相似度所对应的候选产品词确定为核心产品词。
  7. 根据权利要求6所述的系统,其特征在于,所述相似度确定模块包括:
    模型训练单元,用于基于训练数据训练Siamese网络,以得到训练后的Siamese网络;
    计算单元,用于通过训练后的Siamese网络确定所述展示图像中的商品与所述产品图像集合对应的候选产品词的相似度。
  8. 根据权利要求7所述的系统,其特征在于,所述相似度确定模块还包括训练数据获取单元,用于选取预定数量的训练产品词;对于每个训练产品词,根据商品搜索日志选取两个以上包含该训练产品词的搜索词;对于每个搜索词,根据该搜索词下商品的点击量以及点击率,选取相同数量的商品以及该商品的展示图像;对于每个训练产品词,将该训练产品词下属于同一搜索词的商品的展示图像组成一对正例对,将该训练产品词下属于不同搜索词的商品的展示图像组成一对负例对。
  9. 根据权利要求7所述的系统,其特征在于,模型训练单元还用于对于训练数据中每个训练产品词的展示图像,通过聚类算法对训练产品词的展示图像进行聚类处理,以获得一个以上的类别中心;将所述一个以上的类别中心作为该训练产品词的产品图像集合;以及,将每个训练产品词的产品图像集合保存至图像特征库;
    所述获取模块根据所述图像特征库确定多个产品图像集合,所述多个产品图像集合中的各产品图像集合与所述多个候选产品词中的各候选产品词一一对应。
  10. 根据权利要求6所述的系统,其特征在于,所述获取模块利用条件随机场确定所述商品的标题中包含的多个候选产品词。
  11. 一种识别核心产品词的电子设备,其特征在于,包括:
    一个或多个处理器;
    存储装置,用于存储一个或多个程序,
    当所述一个或多个程序被所述一个或多个处理器执行,使得所述 一个或多个处理器实现如权利要求1-5中任一所述的方法。
  12. 一种计算机可读介质,其上存储有计算机程序,其特征在于,所述程序被处理器执行时实现如权利要求1-5中任一所述的方法。
PCT/CN2018/108230 2017-10-12 2018-09-28 一种识别核心产品词的方法和系统 WO2019072098A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US16/754,737 US11741094B2 (en) 2017-10-12 2018-09-28 Method and system for identifying core product terms

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201710946537.0A CN107832338B (zh) 2017-10-12 2017-10-12 一种识别核心产品词的方法和系统
CN201710946537.0 2017-10-12

Publications (1)

Publication Number Publication Date
WO2019072098A1 true WO2019072098A1 (zh) 2019-04-18

Family

ID=61647823

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/108230 WO2019072098A1 (zh) 2017-10-12 2018-09-28 一种识别核心产品词的方法和系统

Country Status (3)

Country Link
US (1) US11741094B2 (zh)
CN (1) CN107832338B (zh)
WO (1) WO2019072098A1 (zh)

Families Citing this family (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107832338B (zh) 2017-10-12 2020-02-07 北京京东尚科信息技术有限公司 一种识别核心产品词的方法和系统
CN109325163A (zh) * 2018-07-09 2019-02-12 深圳市吧圃创新科技有限公司 一种从设计到产品发布的高效方法
CN112835990A (zh) * 2019-11-22 2021-05-25 北京沃东天骏信息技术有限公司 一种识别方法和装置
CN111046180A (zh) * 2019-12-05 2020-04-21 竹间智能科技(上海)有限公司 一种基于文本数据的标签识别方法
CN113220980A (zh) * 2020-02-06 2021-08-06 北京沃东天骏信息技术有限公司 物品属性词识别方法、装置、设备及存储介质
CN111881674B (zh) * 2020-06-28 2023-07-25 百度在线网络技术(北京)有限公司 核心商品词挖掘方法、装置、电子设备及存储介质
CN115272302B (zh) * 2022-09-23 2023-03-14 杭州申昊科技股份有限公司 图像中零部件检测方法,零部件检测设备及系统

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110173528A1 (en) * 2004-09-22 2011-07-14 Yonatan Zunger Determining Semantically Distinct Regions of a Document
CN102270234A (zh) * 2011-08-01 2011-12-07 北京航空航天大学 一种图像搜索方法及其搜索引擎
CN103942693A (zh) * 2013-01-18 2014-07-23 阿里巴巴集团控股有限公司 识别杂货铺的方法、装置及搜索店铺的方法、系统
CN107832338A (zh) * 2017-10-12 2018-03-23 北京京东尚科信息技术有限公司 一种识别核心产品词的方法和系统

Family Cites Families (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8768050B2 (en) * 2011-06-13 2014-07-01 Microsoft Corporation Accurate text classification through selective use of image data
US8737728B2 (en) * 2011-09-30 2014-05-27 Ebay Inc. Complementary item recommendations using image feature data
US10114902B2 (en) * 2012-06-29 2018-10-30 Ebay Inc. Method for detecting and analyzing site quality
US9443164B2 (en) * 2014-12-02 2016-09-13 Xerox Corporation System and method for product identification
US9436895B1 (en) * 2015-04-03 2016-09-06 Mitsubishi Electric Research Laboratories, Inc. Method for determining similarity of objects represented in images
US20160350336A1 (en) * 2015-05-31 2016-12-01 Allyke, Inc. Automated image searching, exploration and discovery
US9965717B2 (en) * 2015-11-13 2018-05-08 Adobe Systems Incorporated Learning image representation by distilling from multi-task networks
US10007860B1 (en) * 2015-12-21 2018-06-26 Amazon Technologies, Inc. Identifying items in images using regions-of-interest
US11080918B2 (en) * 2016-05-25 2021-08-03 Metail Limited Method and system for predicting garment attributes using deep learning
US10109051B1 (en) * 2016-06-29 2018-10-23 A9.Com, Inc. Item recommendation based on feature match
CN106469299B (zh) * 2016-08-31 2019-07-19 北京邮电大学 一种车辆搜索方法及装置

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110173528A1 (en) * 2004-09-22 2011-07-14 Yonatan Zunger Determining Semantically Distinct Regions of a Document
CN102270234A (zh) * 2011-08-01 2011-12-07 北京航空航天大学 一种图像搜索方法及其搜索引擎
CN103942693A (zh) * 2013-01-18 2014-07-23 阿里巴巴集团控股有限公司 识别杂货铺的方法、装置及搜索店铺的方法、系统
CN107832338A (zh) * 2017-10-12 2018-03-23 北京京东尚科信息技术有限公司 一种识别核心产品词的方法和系统

Also Published As

Publication number Publication date
US11741094B2 (en) 2023-08-29
CN107832338A (zh) 2018-03-23
US20200311071A1 (en) 2020-10-01
CN107832338B (zh) 2020-02-07

Similar Documents

Publication Publication Date Title
CN107832338B (zh) 一种识别核心产品词的方法和系统
WO2018192491A1 (zh) 信息推送方法和装置
CN107330752B (zh) 识别品牌词的方法和装置
TWI631474B (zh) Method and device for product identification label and method for product navigation
US20190392258A1 (en) Method and apparatus for generating information
CN110232152A (zh) 内容推荐方法、装置、服务器以及存储介质
US10664888B2 (en) Method and system for attribute extraction from product titles using sequence labeling algorithms
CN110827112B (zh) 深度学习的商品推荐方法、装置、计算机设备及存储介质
US20230214895A1 (en) Methods and systems for product discovery in user generated content
CN110347908B (zh) 语音购物方法、装置、介质及电子设备
CN110084658B (zh) 物品匹配的方法和装置
TWI705411B (zh) 社交業務特徵用戶的識別方法和裝置
CN111444304A (zh) 搜索排序的方法和装置
CN110363206B (zh) 数据对象的聚类、数据处理及数据识别方法
US20230030560A1 (en) Methods and systems for tagged image generation
CN110827101B (zh) 一种店铺推荐的方法和装置
CN111368081A (zh) 一种精选文本内容确定方法和系统
CN113450172A (zh) 一种商品推荐方法和装置
CN111782850A (zh) 一种基于手绘图的物品搜索方法和装置
CN114092194A (zh) 一种产品推荐方法、装置、介质及设备
CN110110199B (zh) 信息输出方法和装置
CN113744011A (zh) 物品搭配方法和物品搭配装置
US20200226167A1 (en) Methods and systems for dynamic content provisioning
CN110738538A (zh) 识别相似物品的方法和装置
CN112288512B (zh) 信息处理方法、装置、电子设备及可读存储介质

Legal Events

Date Code Title Description
NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 18866024

Country of ref document: EP

Kind code of ref document: A1