WO2020125015A1 - 场景化商家召回 - Google Patents

场景化商家召回 Download PDF

Info

Publication number
WO2020125015A1
WO2020125015A1 PCT/CN2019/098210 CN2019098210W WO2020125015A1 WO 2020125015 A1 WO2020125015 A1 WO 2020125015A1 CN 2019098210 W CN2019098210 W CN 2019098210W WO 2020125015 A1 WO2020125015 A1 WO 2020125015A1
Authority
WO
WIPO (PCT)
Prior art keywords
merchant
word
merchants
recall
platform
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.)
Ceased
Application number
PCT/CN2019/098210
Other languages
English (en)
French (fr)
Inventor
阎晗
杨涵
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Sankuai Online Technology Co Ltd
Original Assignee
Beijing Sankuai Online Technology Co Ltd
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 Beijing Sankuai Online Technology Co Ltd filed Critical Beijing Sankuai Online Technology Co Ltd
Publication of WO2020125015A1 publication Critical patent/WO2020125015A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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

Definitions

  • This application relates to the field of data mining technology, specifically to the scenario-based business recall.
  • the scenario-based business recall is based on different scenarios to classify the recall of the merchants stationed on the Internet platform, which can achieve better management of the merchants.
  • the supervised learning method is generally used to obtain the relevance of items and each scene into a classification task, and the probability of belonging to each scene category is used as the correlation between the merchant and the scene; Or, through unsupervised learning methods, such as by directly characterizing item text and scene text as vectors, the similarity is obtained.
  • the supervised learning method requires manual labeling of a large amount of data, which consumes a lot of manpower and has a long time period; while the results generated by the unsupervised learning method cannot take into account accuracy and recall rate, the correlation of the recalled merchant list decays quickly from the head to the waist , Cannot produce results that can be used directly by industry.
  • the commonly used method for obtaining relevance of business is mainly applied to plain text recalls, such as news, documents, articles, etc. recalls, and many Internet platform recalls are merchants, and the text is distributed on different attributes, and the importance of each attribute on the item Different, and the text with different attributes contains a lot of redundant information and special stop words, which requires a lot of targeted optimization of the text.
  • a scenario-based merchant recall method including:
  • a scenario-based merchant recall device including:
  • the keyword set expansion module is used to obtain a given scene word, and expand the given scene word into a keyword set containing N words, where N is a natural number;
  • An initial seed merchant determination module used to determine an initial seed merchant from platform merchants based on each keyword in the keyword set;
  • a high-precision seed merchant determination module configured to filter the initial seed merchant according to the keyword set to obtain a high-precision seed merchant
  • the final recall merchant confirmation module is used to determine the final recall merchant of the given scene word according to the high-precision seed merchant.
  • an electronic device including:
  • a processor a memory, and an acquisition machine program stored on the memory and executable on the processor.
  • the processor executes the program, the aforementioned scenario-based merchant recall method is implemented.
  • a readable storage medium which, when instructions in the storage medium are executed by a processor of an electronic device, enables the electronic device to perform the aforementioned scenario-based merchant recall method.
  • a given scene word can be obtained, and the given scene word can be expanded into a keyword set containing N words, where N is a natural number; based on each key in the keyword set Word, determine the initial seed merchant from the platform merchants; filter the initial seed merchant based on the keyword set to obtain a high-precision seed merchant; determine the final scene word for the given scene based on the high-precision seed merchant Recalling merchants has achieved the beneficial effects of reducing time costs and labor costs while taking into account accuracy and recall rate and improving adaptability.
  • FIG. 1 shows a flowchart of steps of a scenario-based business recall method according to an embodiment of the present application
  • FIG. 2 shows a flowchart of steps of a scenario-based business recall method according to an embodiment of the present application
  • FIG. 3 shows a schematic structural diagram of a scenario-based merchant recall device according to an embodiment of the present application.
  • FIG. 4 shows a schematic diagram of an electronic device according to an embodiment of the present application.
  • FIG. 5 shows a block diagram of another electronic device for performing the method according to the present application.
  • FIG. 6 shows a storage unit for holding or carrying program code implementing the method according to the present application.
  • FIG. 1 a flowchart of steps of a scenario-based business recall method in an embodiment of the present application is shown.
  • Step 110 Obtain a given scene word, and expand the given scene word into a keyword set containing N words, where N is a natural number.
  • the given scene words described in the embodiments of the present application are used to indicate the current merchant recall scene.
  • the recall rate is low. Therefore, in the embodiment of the present application, in order to improve the recall rate, after the given scene words are obtained, the given scene words may be expanded into a keyword set containing N words.
  • N is a natural number.
  • a given scene word for the same scene may specifically include at least one word, and the specific value of N may be preset according to requirements, which is not limited in this embodiment of the present application.
  • the scene-specific words can be expanded into a keyword set containing N words by any available semantic similarity method.
  • the word-bag model can be used to expand a set of scene words to a keyword set containing N words.
  • the Bag of Words (BOW) model assumes that for a text, it ignores its word order, grammar, and syntax, and treats it as a collection of words, or a combination of words. The appearance of each word is independent and does not depend on whether other words appear, or when the author of this article chooses a vocabulary at any position, it is independently selected by the previous sentence.
  • the keyword set of a given scene word may also include the corresponding given scene word, of course, it may not include the corresponding given scene word, and the specific setting can be preset according to requirements. This embodiment of the present application is not limited.
  • the step 110 may further include:
  • Sub-step 111 Obtain a given scene word, and obtain N words with the highest correlation with the given scene word through word2vec to construct a keyword set of the given scene word.
  • Word2vec is a high-efficiency tool for characterizing words as real-valued vectors open sourced by Google in mid-2013. It uses deep learning ideas to simplify the processing of text content into vector operations in K-dimensional vector space through training , And the similarity in the vector space can be used to represent the semantic similarity of the text.
  • the word vectors output by Word2vec can be used to do many NLP (Natural Language Processing Natural Language Processing) related work, such as clustering, finding synonyms, part-of-speech analysis and so on. If you change your mind and use words as features, Word2vec can map the features to a K-dimensional vector space, which can seek a deeper feature representation for text data.
  • NLP Natural Language Processing Natural Language Processing
  • Word2vec uses the distributed vector representation of distributed representation.
  • Distributed representation was first proposed by Hinton in 1986. The basic idea is to map each word into a K-dimensional real number vector (K is generally a hyperparameter in the model) through training, and determine the distance between words (such as cosine similarity, Euclidean distance, etc.) Semantic similarity. It uses a three-layer neural network, input layer-hidden layer-output layer.
  • Huffman Huffman
  • Word2vec is so popular is its efficiency.
  • This three-layer neural network itself models the language model, but at the same time obtains a representation of the word in the vector space, and this side effect is the real goal of Word2vec.
  • the basic idea of Word2Vec is to express each word in natural language as a short vector with a unified meaning and a unified dimension.
  • the top N similar words with the highest similarity of the given scene words can be directly calculated using the formula for calculating the Top N words of the given arbitrary word in the official document of word2vec, which can be constructed to obtain the given Set of keywords for scene words.
  • Step 120 Based on each keyword in the keyword set, determine an initial seed merchant from the platform merchants.
  • Each word in the keyword set is a keyword.
  • the initial seed merchants can be determined from the platform merchants based on each keyword in the keyword set.
  • the platform merchants may include all merchants included in the platform involved in the merchant recall.
  • the initial seed merchant can be determined based on the matching degree of each keyword in the keyword set with the platform merchant; or, all keywords in the keyword set can be taken as a whole, and then based on all keywords and platform merchants The degree of matching, determine the initial seed merchant; etc.
  • the matching degree with the platform merchant may include the matching degree with various attributes of the platform merchant.
  • the attributes can include business name, recommended dishes, group purchase information, reviews, etc.
  • the merchants whose matching degree of each keyword is greater than or equal to the second preset matching degree can be selected from the platform merchants as the initial recall merchants corresponding to the corresponding keywords, or separately from the platform merchants
  • the top S merchants with the highest matching degree for each keyword are selected as the initial recall merchant corresponding to the corresponding keyword, and so on.
  • the initial recall merchants corresponding to each keyword need to be aggregated to obtain the initial seed merchant corresponding to the corresponding keyword set.
  • S is an integer greater than 1, and the second preset matching degree and the specific value of S can be preset according to requirements, which is not limited in this embodiment of the present application.
  • each keyword in the keyword set can be directly hit in the merchant review, and the initial count is filtered by the hit count Seed merchants. Specifically, if any of the merchant reviews of the platform merchant includes any keyword in the keyword set, confirm that the corresponding keyword hits the corresponding platform merchant once, and then the keywords in the keyword set can be matched to the same The hit counts of all the merchant reviews of the platform merchants are weighted and summed to obtain the total hit counts of the keyword set to the corresponding platform merchants.
  • the top R platform merchants can be selected as the initial seed merchants based on the order in which the total number of hits for each platform merchant from the keyword set is high to low.
  • the specific value of R can be preset according to requirements, and this embodiment of the present application is not limited.
  • the weight of the number of hits corresponding to each keyword may be the weight of the degree of relevance between the corresponding keyword and a given scene word.
  • Step 130 Filter the initial seed merchant according to the keyword set to obtain a high-precision seed merchant.
  • a given scene word will be expanded into a set of keywords.
  • the accuracy of the initial seed merchant obtained by each keyword in the keyword set cannot be fully guaranteed, and the algorithm from the initial seed merchant to the final recall result has a very strong propagation ability. If the accuracy of an initial seed merchant is not enough Then, in the subsequent steps, it will continue to be amplified, thereby affecting the accuracy of the merchant's recall. Therefore, in this embodiment of the present application, the accuracy of the initial seed merchant is very high, but there is no requirement for the recall rate of the initial seed merchant.
  • the filtering strategy for filtering the initial seed merchant may be preset according to requirements, and this embodiment of the present application is not limited.
  • the filtering strategy may be set as if the initial seed merchant’s business name contains at least one keyword in the corresponding keyword set, then the corresponding initial seed merchant is confirmed to be a high-precision seed merchant, and if the initial seed merchant’s business name If any keyword in the corresponding keyword set is not included, the corresponding initial seed merchant can be filtered out, then the initial seed merchant retained after filtering is a high-precision seed merchant.
  • Step 140 According to the high-precision seed merchant, determine the final recall merchant of the given scene word.
  • the accuracy of the identified high-precision seed merchants is high, but the number is limited. Therefore, after the high-precision seed merchants are determined, the scope of the recall needs to be further expanded, that is, to determine the final recall merchant for a given scene word.
  • the similarity between each platform merchant and each corresponding high-precision seed merchant can be determined first.
  • the higher the similarity between a platform merchant and a high-precision seed merchant the greater the possibility that the platform merchant and the high-precision seed merchant correspond to the same scenario. Therefore, in the embodiment of the present application, based on the similarity between the merchants of various platforms and the high-precision seed merchants, the final recall merchant of the corresponding given scene word can be determined.
  • the similarity between the platform merchant and the high-precision seed merchant can be confirmed according to any available method, which is not limited in this embodiment of the present application.
  • the similarity between platform merchants and high-precision seed merchants can be determined by comparing the matching degree between platform merchants and high-precision seed merchants; or the attributes of platform merchants and high-precision seed merchants can be characterized as In the form of vectors, the similarity between the vectors corresponding to the same attribute is compared to determine the similarity between the platform merchant and the high-precision seed merchant; and so on.
  • the attributes therein may include at least one of a business name, a review, a special dish, etc. The specific above-mentioned attributes may be set in advance according to requirements, and this embodiment of the present application is not limited.
  • the high-precision merchant determined based on the given scene word can default to the final recall merchant corresponding to the corresponding given scene word, and of course, the high precision determined based on the given scene word can also be set according to requirements The merchant defaults to the final recalled merchant, and this embodiment of the application is not limited.
  • step 140 it may further include:
  • Step 150 Determine a set of search words for each scene according to user search records.
  • the final recall merchant generated through the above steps is a recall merchant based on a given scene word, but in actual applications, the query terms for different users in different scenes are not exactly the same, then after determining the final recall merchant in each scene, In order to be able to return the corresponding query results according to the personalized query words of different users, it is necessary to further perform regular filtering on the recalled merchants covered by each specific query word according to the query words and business requirements specifically covered by the scene to generate specific queries. A collection of recalled businesses that are highly relevant to the word and business.
  • the words filtered by the user according to the user search record may be searched or queried for different scenarios, thereby obtaining a set of search words for each scenario.
  • a set of search words belonging to the corresponding scene can be given through the user search record. This set is a closed set.
  • the set will contain the given scene words of the corresponding scene, as long as the user enters the search word set Any at least one word of will trigger the search process for the corresponding scene.
  • Step 160 Determine a filtering strategy for the search term set based on the search term set and business requirements.
  • the specific content of the filtering strategy, the business requirement, and the correspondence relationship between the filtering strategy, the set of search terms and the business requirement can be preset according to the demand, and this embodiment of the present application is not limited.
  • the filtering strategy of the search term set may also be determined based only on the search term set or business requirements, which is not limited in this embodiment of the present application.
  • Step 170 Based on the filtering strategy, the final recalled merchants of each search term in the search term set are filtered to obtain a recalled merchant set that matches the search term and the business requirements.
  • the final recalled merchants of each search term in the search term set may be filtered based on the filtering strategy to obtain a recalled merchant set that matches the search term and the business requirements. Since the set of search words contains a given scene word of the corresponding scene, the final recall merchant of the given scene word can be filtered based on the filtering strategy to obtain information related to the given scene word and the business requirements Matching collection of recalled businesses.
  • the merchants recalled by traditional methods based on keyword hit filtering and the like can be compared with the merchants recalled by the above steps 110-140, and the merchants recalled by the traditional method can be extracted and compared with the merchants recalled by the above steps 110-140.
  • the merger produces the final recall result.
  • a given scene word can be obtained, and the given scene word can be expanded into a keyword set containing N words, where N is a natural number; based on each key in the keyword set Word, determine the initial seed merchant from the platform merchants; filter the initial seed merchant based on the keyword set to obtain a high-precision seed merchant; determine the final scene word for the given scene based on the high-precision seed merchant Recall business. This reduces the time cost and labor cost of the merchant's recall, while taking into account the accuracy and recall rate of the merchant's recall.
  • FIG. 2 a flowchart of steps of a scenario-based business recall method in an embodiment of the present application is shown.
  • Step 210 Train the LDA model through pre-selected high-quality business reviews and generate a word-topic matrix.
  • the LDA model is a document topic generation model, also known as a three-layer Bayesian probability model, which contains a three-layer structure of words, topics and documents. Then, training the LDA model through pre-selected high-quality business reviews can generate the corresponding word-topic matrix.
  • the high-quality business reviews are the business reviews of different businesses selected in advance according to their needs.
  • high-quality business reviews may include reviews of Meituan reviews that have been subjected to manual processes or automatic filtering before being stored, so the quality of high-quality merchant reviews is guaranteed.
  • the step 210 may further include:
  • Sub-step 211 pre-processing the pre-selected high-quality merchant reviews.
  • pre-selected high-quality merchant reviews may be pre-processed.
  • the pre-processing may include, but is not limited to, at least one of special character removal processing and word segmentation processing.
  • the special characters may include but not limited to special symbols, low frequency words, stop words, and so on.
  • the special symbols may include symbols that are less frequently used and difficult to input directly than traditional or commonly used symbols, such as: mathematical symbols, unit symbols, tabs, etc.
  • Low-frequency words refer to words that have fewer occurrences.
  • a frequency threshold may be preset. If a word appears in a comment less than the frequency threshold, it may be confirmed as a low-frequency word.
  • Stop words can be included in information retrieval. To save storage space and improve search efficiency, certain words or words are automatically filtered out before or after processing natural language data (or text). These words or words are called Stop Words (stop words).
  • the stop word removal process may be performed based on the existing stop word list, and/or stop words for different services may be preset. For example, for the restaurant business name "restaurant" is the stop word, for the group purchase field text "full XXX minus XXX" is the stop word, for example, "photo" in the photo business is the stop word, and so on.
  • word segmentation processing can be performed on high-quality business reviews by any available word segmentation processing method, which is not limited in the embodiment of the present application.
  • sub-step 212 the pre-processed high-quality business reviews are used to train the LDA model and generate a word-topic matrix.
  • the number of hidden topics K in the LDA model may also be preset.
  • the number of hidden topics can be understood as the number of topics in the word-topic matrix obtained after training.
  • Step 220 Represent the given scene word as a mixture of multiple original topics based on the LDA model.
  • a mixed component based on the original topic can be generated for any given document.
  • the original topic is also the topic obtained by the above-mentioned trained LDA model.
  • the given scene word can be input into the trained LDA model, and then the probability of the given scene word under each original theme is obtained, that is, the given scene word is represented as a mixed component of multiple original themes.
  • the probability of "anime” under each original theme can be obtained based on the LDA model, that is, the theme distribution of "anime” is as follows:
  • the left column represents the original topic number
  • the right column represents the probability of a given scene word under the corresponding native topic.
  • Step 230 Determine the relevance weight of the given scene word and each word in the word-topic matrix based on the conditional probability represented by the mixed component.
  • the probability of the given scene word under the original topic can be expressed as P(topic k
  • word represents any word in the word-topic matrix
  • K represents the total number of original topics.
  • topic k ) can be obtained when the word-topic matrix is obtained.
  • Step 240 Select N words with the highest relevance weight to the given scene word from each word of the word-topic matrix to construct a keyword set of the given scene word.
  • N words with the highest relevance weight to the given scene word may be selected from each word of the word-topic matrix to construct a keyword set of the given scene word.
  • the value of N can be preset according to requirements, and this embodiment of the present application is not limited.
  • Step 250 Re-determine the relevance weight of each word in the word-topic matrix and the given scene word based on a preset rearrangement algorithm.
  • Step 260 Update the keyword set according to the newly determined relevance weight.
  • the keyword set of a given scene word obtained based on the above-mentioned correlation weight may contain some keywords that are not actually relevant, then at this time, the keyword can be re-determined based on the preset rearrangement algorithm.
  • the rearrangement algorithm may be preset according to requirements, and this embodiment of the present application is not limited.
  • the relevance weight of a given scene word and each word in the word-topic matrix can be re-determined based on multiple attributes of the word meaning and occurrence scene of the word in the given scene word and the word-topic matrix, and so on.
  • the keyword set corresponding to the given scene word may be updated based on the re-determined relevance weight.
  • a high-precision keyword set can be selected by a small amount of manual selection.
  • Step 270 Determine the vector representation of the merchants of each platform and the vector representation of the keyword set.
  • the vector representation of the merchants of various platforms and the vector representation of the keyword set may be determined first.
  • the vector representation of the merchants of various platforms and the vector representation of the keyword set can be determined by any available method, and this embodiment of the present application is not limited.
  • the step 270 may further include:
  • sub-step 271 the comments of the merchants on the same platform are aggregated into a comment text, and all keywords in the keyword set are aggregated into a keyword text.
  • the initial seed merchant is determined.
  • the comments of the merchants on the same platform can be aggregated into a comment text
  • all keywords in the keyword set can be aggregated into a keyword text.
  • the reviews belonging to the same platform business that are aggregated into one review text can be pre-screened reviews of higher quality and belonging to the corresponding platform business, of course, it can also be all the obtained reviews of the corresponding platform business, which can be based on
  • the requirements are preset, and this embodiment of the present application is not limited.
  • Sub-step 272 a vector representation of the platform merchant is determined based on the review text, and a vector representation of the keyword set is determined based on the keyword text.
  • the vector representation of the platform merchant can be determined based on the review text, and the vector representation of the platform merchant can be determined based on the keyword text, respectively.
  • the sub-step 272 may further include: determining the vector representation of the platform merchant according to the review text through the LDA model, and using the LDA model according to the The keyword text determines the vector representation of the keyword set; or, through the word2vec model, the vector representation of the platform merchant is determined based on the review text, and the vector representation of the keyword set is determined based on the keyword text.
  • the vector representation obtained by the LDA model can also be called the theme vector.
  • the aforementioned trained LDA model can be used to infer from the comment text
  • the vector representation of the merchant on the corresponding platform is obtained, and then the keyword text is used as the long text.
  • the vector representation of the corresponding keyword set is also inferred through the LDA model.
  • the vector representation of the platform merchant may also be determined according to the review text through the word2vec model, and determined based on the keyword text through the word2vec model
  • a vector representation of the set of keywords. word2vec is also called word embeddings.
  • the Chinese name "word vector” is used to convert words in natural language into dense vectors (Dense Vector) that can be understood by computers.
  • the LDA model may be preferably used to determine the vector representation, and if the keyword text and/or comment text is short text, it may be preferably used The word2vec model determines the vector representation.
  • the specific setting may be preset according to requirements, and this embodiment of the present application is not limited.
  • Step 280 Obtain the similarity between the vector representation of each platform merchant and the vector representation of the keyword set.
  • the similarity between the vector representation of each platform merchant and the vector representation of the keyword set may be determined based on the vector representation.
  • the similarity may include but not limited to any one of cosine similarity, Euclidean distance similarity, cross-entropy similarity, and relative entropy distance similarity.
  • any corresponding method can be used to determine the corresponding similarity, which is not limited in the embodiments of the present application.
  • step 290 the top M platform merchants with the highest similarity are selected as the initial seed merchants for the given scene word.
  • the top M platform merchants with the highest similarity can be selected as the initial seed merchants for the corresponding given scene words.
  • M is an integer greater than 1, and the value of M can be preset according to requirements, which is not limited in this embodiment of the present application.
  • Step 2110 Filter the initial seed merchant according to the keyword set to obtain a high-precision seed merchant.
  • Step 2120 Perform multiple iteration operations according to the high-precision seed merchant, and use the sum of the precise recall merchants obtained by each iteration calculation as the final recall merchant of the given scene word.
  • the final recalled merchant may also include the initially acquired high-precision seed merchant, and of course may not include the initially acquired high-precision seed merchant. Specifically, it may be preset according to requirements, which is not limited in this embodiment of the present application.
  • the process of each iteration operation includes:
  • Step S1 according to each attribute of the platform merchant and the preset weight of the attribute, determine the final similarity between each high-precision seed merchant and each other merchant in the platform merchant;
  • each platform merchant can include multiple different attributes, such as merchant name, special dishes, evaluation, etc., and the importance of different attributes in determining the similarity is not exactly the same, so in this book
  • the preset weight of each attribute of the platform merchant may be set in advance, and then each of the high-precision seed merchant and the platform may be determined according to each attribute of the platform merchant and the preset weight of the attribute The final similarity of each other business in the business.
  • other merchants corresponding to a given scene word may include other platform merchants other than the high-precision seed merchant corresponding to the corresponding given scene word in the platform merchant.
  • the similarity of high-precision seed merchants and other merchants under different attributes can be determined separately, and then the similarity of different attributes can be weighted and summed based on the preset weight of each attribute, thereby obtaining high-precision seed merchants and corresponding The final similarity of other businesses.
  • the text data corresponding to each attribute may also be pre-processed.
  • the preprocessing at this time may include washing the text data with specific stop words, where the specific stop words may include stop words preset based on business needs, for example, “restaurant” in the business name of the gourmet business is a specific stop word Words, for example, "full XXX minus XXX" in the text of the group purchase field is a specific stop word, for example, "photo” in the photo business is a specific stop word, and so on.
  • Step S2 based on a preset similarity threshold, determine the number of votes the high-precision seed merchant voted for the other merchants; the weight of the votes is the final similarity between the high-precision seed merchant and the platform merchant;
  • the similarity threshold may be preset according to requirements, and the preset similarity may be related to the type of similarity, which is not limited in this embodiment of the present application. Specifically, if the final similarity between a certain other merchant and a high-precision seed merchant exceeds the preset similarity threshold, it is determined that the corresponding high-precision seed merchant will vote for the corresponding other merchant, and the weight of the number of votes is the corresponding The final similarity between the platform merchant and the corresponding high-precision seed merchant; if the final similarity between a certain other merchant and a high-precision seed merchant does not exceed the preset similarity threshold, it is determined that the corresponding high-precision seed merchant will not Corresponding other businesses vote.
  • the final similarities between other merchants B, C, and D and high-precision seed merchant A1 are 0.6, 0.8, and 0.3, respectively, and the final similarities between other merchants B, C, and D and high-precision seed merchant A2 are respectively It is 0.7, 0.4, 0.3, and the larger the similarity value, the more similar it is.
  • the preset similarity threshold is 0.5, it can be determined that the number of votes of high-precision seed merchant A1 to other merchants B is 0.6, the number of high-precision seed merchants A1 to other merchants C is 0.8, and the high-precision seed merchants A1 will not To vote for other merchants D, the high-precision seed merchant A2 votes 0.7 for other merchants B, and the high-precision seed merchant A2 will not vote for other merchants C and D. Then it can be obtained that the number of votes obtained by other merchants B at this time is 1.3, the number of votes obtained by other merchants C is 0.8, and the number of votes obtained by other merchants D is 0.
  • Step S3 selecting the top L other merchants with the highest number of votes as the initial recall merchant, L is a natural number;
  • the higher the number of votes the higher the similarity between the corresponding other merchants and the high-precision seed merchants. Therefore, in this embodiment of the present application, the top L other merchants with the highest number of votes can be selected as the initials of the corresponding given scene words. Recall business.
  • the value of L can be preset according to requirements, and this embodiment of the present application is not limited.
  • the initial recalled merchants can be selected as other merchants B and C in turn.
  • Step S4 filtering the initial recall merchant according to the keyword set to obtain an accurate recall merchant
  • the initial recall merchant can also be filtered according to the keyword set to obtain an accurate recall merchant.
  • the specific filtering process may be similar to the foregoing filtering process for the initial seed merchants, and will not be repeated here.
  • Step S5 Select the first number of merchants with the highest number of votes from the precise recall merchants as the new high-precision seed merchants, and update the high-precision seed merchants in this iteration until the precision of the recall of the precise recall merchants drops Or the number of the exact recalled merchants has converged.
  • the first number of merchants from the precise recalled merchants obtained through this iteration can be selected as the new high-precision seed merchants, and the new high-precision seed merchants can be used to update the High-precision seed merchants, and then enter the next iteration process with the updated high-precision seed merchants, until the precision of the recalled merchants obtained decreases or the number of precision recalled merchants converges.
  • the specific value of the first quantity may be preset according to requirements, and this embodiment of the present application is not limited.
  • the process of updating the high-precision seed merchants of this iteration may be adding the first selected high-precision seed merchants to the high-precision seed merchants used in this iteration process, or the latest selected The first number of high-precision seed merchants replace the high-precision seed merchants used in this iteration, and so on.
  • it can be preset according to requirements, and this embodiment of the present application is not limited.
  • step S4 it can be further determined whether the recall accuracy of the precise recalled merchant has decreased, or whether the number of precise recalled merchants has converged, if the recall precision of the precise recalled merchant has decreased, or the precision If the number of recalled merchants has converged, the iteration can be stopped without performing step S5; and if the recall precision of the precise recalled merchant has not decreased, and the number of precise recalled merchants has not converged, then S5 can be performed and the iteration continues.
  • the attribute includes at least one of a business name, a recommended dish, group purchase information, and a comment.
  • the step S2 may further include:
  • Sub-step S21 characterizing each attribute of the platform merchant as a vector to obtain an attribute vector of the platform merchant
  • each attribute of the platform merchant needs to be characterized as a vector to obtain the attribute vector of the platform merchant.
  • the attributes can be characterized as vectors by any available method, which is not limited in the embodiments of the present application. For example, through the aforementioned LDA model, word2vec model, or doc2vec model, etc.
  • Sub-step S22 based on the attribute vector and the preset weight of the attribute, determine the final similarity between each high-precision seed merchant and each other merchant in the platform merchant.
  • the final similarity between each high-precision seed merchant and each other merchant in the platform merchant may be determined based on the attribute vector and the preset weight of the attribute.
  • the preset weight of each attribute may be preset according to requirements, and this embodiment of the present application is not limited.
  • the sub-step S21 may further include:
  • Sub-step S211 performing data cleaning on the text data corresponding to the attributes based on the specific stop words of the attributes
  • different specific stop words may be set for different attributes. For example, set "restaurant” in the business name of the food business as a specific stop word, set "full XXX minus XXX” in the text of the group purchase field as a specific stop word, and set "photo” in the photo business as a specific stop word. and many more. Then, the text data corresponding to the attributes may be cleaned based on the specific stop words of the attributes.
  • Sub-step S212 using word2vec to characterize the merchant name of the platform merchant after data cleaning as a vector
  • the attributes in the embodiments of the present application may include the name of the business, and since the general name of the business name will not be too many, then this time can be applied to characterizing the business name of the platform business after data cleaning by word2vec as Vector to obtain the attribute vector corresponding to the business name.
  • the space is generally larger than the name of the business, and the expression is more inclined to the document form. Therefore, in the embodiment of this application, the data after the data can be cleaned by doc2vec
  • the reference information of the platform merchant is characterized as a vector, and then the attribute vector corresponding to each reference information is obtained.
  • each participle in the reference information of the platform merchant is separately characterized as a vector, and the vector average of each participle in each reference information is obtained as the vector of the reference information.
  • the sub-step S22 may further include:
  • the similarity of the high-precision seed merchant corresponding to the current given scene word in the platform merchant and other merchants under each attribute can be obtained respectively. For example, similarity of business name, similarity of special dishes, similarity of group purchase information, etc.
  • sub-step S222 the similarity of each attribute is weighted and summed according to the preset weight of the attribute to obtain the final similarity between the high-precision seed merchant and the other merchant.
  • the similarity of each attribute can be weighted and summed based on the preset weight of each attribute, and then the final similarity between the corresponding high-precision seed merchant and the corresponding other merchant can be obtained.
  • the similarity may include but is not limited to any one of cosine similarity, Euclidean distance similarity, cross-entropy similarity, and relative entropy distance similarity.
  • step S5 before the step S5, it may further include:
  • Step S6 Determine the value of the first quantity according to the number of exact recalled merchants obtained by this iteration operation and the iteration order of this iteration operation.
  • the first number corresponding to each iteration may be the same or not the same. If the first number corresponding to each iteration is not the same, you can set the first number according to the iteration process. Value to further improve the accuracy and recall rate of the final recalled merchant. Therefore, in the embodiment of the present application, the value of the first quantity may be determined according to the number of accurate recall merchants obtained by this iterative operation and the iteration order of this iterative operation. The specific relationship between the number of exact recalled merchants obtained by this iteration operation and the iteration order of this iteration operation and the first number corresponding to this iteration can be pre-set according to requirements, and this embodiment of the present application is not applied limited.
  • the value of the first number can be set to T*step*n_iteration
  • step is the number of exact recalled merchants determined by this iteration operation
  • n_iteration is the iteration order of this iteration operation
  • T is a preset parameter.
  • the value of T can be set according to requirements, which is not limited in this embodiment of the present application, for example, T can be directly set to 1, and so on.
  • a given scene word can be obtained and expanded to a keyword set containing N words; based on each key in the keyword set Word, determine the initial seed merchant from the platform merchants; filter the initial seed merchant based on the keyword set to obtain a high-precision seed merchant; determine the final scene word for the given scene based on the high-precision seed merchant Recall business. This reduces the time cost and labor cost of the merchant's recall, while taking into account the accuracy and recall rate of the merchant's recall.
  • an LDA model can be trained through pre-selected high-quality merchant reviews and a word-topic matrix is generated; based on the LDA model, the given scene word is expressed as multiple original topics Based on the conditional probability represented by the mixed component, determine the relevance weight of the given scene word and each word in the word-topic matrix; select from each word in the word-topic matrix and The N words with the highest relevance weight of the given scene word construct a keyword set of the given scene word. Furthermore, the relevance weights of each word in the word-topic matrix and the given scene word are re-determined based on a preset rearrangement algorithm; and the keyword set is updated according to the re-determined relevance weights.
  • preprocessing the pre-selected high-quality business reviews; using the pre-processed high-quality business reviews to train the LDA model and generate a word-topic matrix; wherein, the pre-processing includes special character removal and word segmentation processing;
  • the special characters include at least one of special symbols, low-frequency words, and stop words. This can further improve the accuracy and recall rate of merchant recalls.
  • the vector representation of each platform merchant and the vector representation of the keyword set can be determined; the vector representation of each platform merchant and the vector of the keyword set can be obtained Represents the similarity between; select the top M platform merchants with the highest similarity as the initial seed merchants for the given scene word. Aggregate the reviews of merchants on the same platform into one review text, and aggregate all keywords in the keyword set into one keyword text; determine the vector representation of the platform merchant based on the review text, and based on the key The word text determines the vector representation of the set of keywords. It can also further improve the accuracy and recall rate of merchant recalls.
  • the process of each iteration operation includes: according to each attribute of the platform merchant and the preset weight of the attribute, determine each of the high-precision seed merchants and each other merchant in the platform merchant Final similarity; based on a preset similarity threshold, determine the number of votes the high-precision seed merchant voted for the other merchants; the weight of the votes is the final similarity between the high-precision seed merchant and the platform merchant Select the top L other merchants with the highest number of votes as the initial recall merchants; filter the initial recall merchants according to the keyword set to obtain accurate recall merchants; select the highest number of votes from the precise recall merchants The first number of merchants serves as a new high-precision seed merchant, and updates the high-precision seed merchant of this iter
  • each attribute of the platform merchant as a vector to obtain an attribute vector of the platform merchant; based on the attribute vector and the preset weight of the attribute, each high-precision seed merchant and the platform are determined
  • the final similarity of each other business in the business Perform data cleaning on the text data corresponding to the attribute based on the specific stop words of each attribute; characterize the business name of the platform merchant after data cleaning as a vector through word2vec; and clean the data of the platform merchant after data cleaning through doc2vec
  • the reference information is respectively characterized as a vector; or, word2vec is used to characterize each participle in the platform merchant's reference information after data cleaning as a vector, and the vector average of each participle in each reference information is obtained as the A vector of the reference information; wherein the reference information includes at least one of recommended dishes, group purchase information, and reviews.
  • the value of the first number is determined according to the number of accurate recall merchants obtained by this iterative operation and the iteration order of this iterative operation. It can also further reduce the time cost and labor cost of the merchant's recall, while improving the precision and recall rate of the merchant's recall.
  • a scenario-based merchant recall device provided by an embodiment of the present application is described in detail.
  • FIG. 3 shows a schematic structural diagram of a scenario-based merchant recall device in an embodiment of the present application.
  • the keyword set expansion module 310 is used to obtain a given scene word, and expand the given scene word into a keyword set containing N words, where N is a natural number;
  • An initial seed merchant determination module 320 configured to determine an initial seed merchant from platform merchants based on each keyword in the keyword set;
  • the high-precision seed merchant determination module 330 is configured to filter the initial seed merchant according to the keyword set to obtain a high-precision seed merchant
  • the final recall merchant confirmation module 340 is used to determine the final recall merchant of the given scene word according to the high-precision seed merchant.
  • a given scene word can be acquired and expanded to a keyword set containing N words; based on each key in the keyword set Word, determine the initial seed merchant from the platform merchants; filter the initial seed merchant based on the keyword set to obtain a high-precision seed merchant; determine the final scene word for the given scene based on the high-precision seed merchant Recall business. This reduces the time cost and labor cost of the merchant's recall, while taking into account the accuracy and recall rate of the merchant's recall.
  • An embodiment of the present application of the present invention also discloses an electronic device, including:
  • the processor executes the program, the aforementioned scenario-based merchant recall method is implemented.
  • the electronic device 400 includes a memory 410 and a processor 420, and the memory 410 and the processor 420 are communicatively connected by a bus.
  • the memory 410 stores a computer program, and the computer program can run on the processor 420. Furthermore, the steps in the method described in any of the above embodiments of the present application are implemented.
  • An embodiment of the present application also discloses a readable storage medium.
  • the instructions in the storage medium are executed by the processor of the electronic device, the electronic device can execute the aforementioned scenario-based merchant recall method.
  • the scenario-based merchant recall device disclosed in the embodiments of the present application can implement the aforementioned scenario-based merchant recall method.
  • the specific implementation of each module of the device corresponds to the corresponding method step Therefore, the description is relatively simple, and the relevant part can be referred to the description of the method embodiment.
  • modules in the device in the embodiment can be adaptively changed and set in one or more devices different from the embodiment.
  • the modules or units or components in the embodiments may be combined into one module or unit or component, and in addition, they may be divided into a plurality of submodules or subunits or subcomponents. Except that at least some of such features and/or processes or units are mutually exclusive, all features disclosed in this specification (including the accompanying claims, abstract and drawings) and any methods so disclosed or All processes or units of equipment are combined. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract and drawings) may be replaced by alternative features serving the same, equivalent or similar purpose.
  • Each component embodiment of the present application may be implemented by hardware, or implemented by a software module running on one or more processors, or implemented by a combination thereof.
  • a microprocessor or a digital signal processor (DSP) may be used to implement some or all functions of some or all components in the scenario-based merchant recall device according to an embodiment of the present application.
  • the present application may also be implemented as a device or device program (for example, an acquisition machine program and an acquisition machine program product) for performing part or all of the methods described herein.
  • Such a program for implementing the present application may be stored on an acquisition machine-readable medium, or may have the form of one or more signals.
  • Such a signal can be downloaded from an Internet website, or provided on a carrier signal, or provided in any other form.
  • FIG. 5 shows an electronic device that can implement the method according to the present application, which traditionally includes a processor 520 and a computer program product in the form of a memory 510 or a computer-readable storage medium.
  • the electronic device may be a PC, a mobile terminal, a personal digital assistant, a tablet computer, or the like.
  • the electronic device is a computing processing device.
  • the memory 510 may be an electronic memory such as flash memory, EEPROM (Electrically Erasable Programmable Read Only Memory), EPROM, hard disk, or ROM.
  • the memory 510 has a storage space 5101 for program code for performing any method steps in the above method.
  • the storage space 5101 for program code may include program code 5102 for implementing various steps in the above method, respectively.
  • the computer program product or the computer-readable storage medium has stored thereon the program code of the computer program, and when the program code is executed by the processor 520, the data reading method as described in Embodiment 1 and Embodiment 2 of the present application is implemented.
  • These program codes can be read from or written into one or more computer program products.
  • These computer program products include program code carriers such as hard disks, compact disks (CDs), memory cards or floppy disks.
  • Such a computer program product or computer-readable storage medium is usually a portable or fixed storage unit as described with reference to FIG. 6.
  • the storage unit is used for holding or carrying program code for implementing the method according to the present application.
  • the storage unit may have storage sections, storage spaces, and the like arranged similarly to the memory 510 in the electronic device of FIG. 5.
  • the program code may be compressed in an appropriate form, for example.
  • the storage unit includes computer readable code 6102', that is, code that can be read by a processor such as 520, for example, which causes the electronic device to perform various steps in the method described above when the electronic device is running .
  • any reference signs between parentheses should not be constructed as limitations on the claims.
  • the word “comprising” does not exclude the presence of elements or steps not listed in the claims.
  • the word “a” or “one” before an element does not exclude the presence of multiple such elements.
  • the application can be realized by means of hardware including several different elements and by means of a suitably programmed computer. In the unit claims enumerating several devices, several of these devices may be embodied by the same hardware item.
  • the use of the words first, second, and third does not indicate any order. These words can be interpreted as names.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Databases & Information Systems (AREA)
  • Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • Accounting & Taxation (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Development Economics (AREA)
  • Economics (AREA)
  • Finance (AREA)
  • Marketing (AREA)
  • Strategic Management (AREA)
  • General Business, Economics & Management (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

本申请公开了一种场景化商家召回方法、装置、电子设备及可读存储介质。所述方法，包括：获取给定场景词，并将所述给定场景词扩展为包含N个词的关键词集合，N为自然数；基于所述关键词集合中的各个关键词，从平台商家中确定初始种子商家；根据所述关键词集合，对所述初始种子商家进行过滤，得到高精度种子商家；根据所述高精度种子商家，确定所述给定场景词的最终召回商家。

Description

场景化商家召回
本申请要求在2018年12月17日提交中国专利局、申请号为201811545314.4、发明名称为“场景化商家召回方法、装置、电子设备及可读存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及数据挖掘技术领域，具体涉及场景化商家召回。
背景技术
商家召回是O2O(Online To Offline，在线离线/线上到线下)等互联网模式中的重要问题。其中，场景化商家召回是基于不同场景对互联网平台上入驻的商家进行分类召回，可实现对商家更好的管理。在现有的场景化商家召回相关技术中，一般是采用有监督学习方法将物品与每个场景的相关度获取转化为分类任务，将属于每个场景类别的概率作为商家跟场景的相关度；或者，通过无监督学习方法，如通过直接将物品文本和场景文本表征为向量，进而获取相似度。
但是，采用有监督学习方法需要人工标注大量数据，耗费大量人力并且时间周期长；而采用无监督学习方法生成的结果无法兼顾精度和召回率，召回商家列表从头部到腰部相关度衰减很快，无法产出工业界能够直接使用的结果。而且业务常用的相关度获取方法主要是应用于纯文本召回，比如新闻、文档、文章等召回，而很多互联网平台的召回物品是商家，文本分布在不同的属性上，各属性对物品的重要性不同，并且不同属性文本中包含大量冗余信息和特殊停用词，需要对文本进行大量针对性优化。
发明内容
依据本申请第一方面，提供了一种场景化商家召回方法，包括：
获取给定场景词，并将所述给定场景词扩展为包含N个词的关键词集合，N为自然数；
基于所述关键词集合中的各个关键词，从平台商家中确定初始种子商家；
根据所述关键词集合，对所述初始种子商家进行过滤，得到高精度种子商家；
根据所述高精度种子商家，确定所述给定场景词的最终召回商家。
根据本申请的第二方面，提供了一种场景化商家召回装置，包括：
关键词集合扩展模块，用于获取给定场景词，并将所述给定场景词扩展为包含N个词的关键词集合，N为自然数；
初始种子商家确定模块，用于基于所述关键词集合中的各个关键词，从平台商家中确定初始种子商家；
高精度种子商家确定模块，用于根据所述关键词集合，对所述初始种子商家进行过滤，得到高精度种子商家；
最终召回商家确认模块，用于根据所述高精度种子商家，确定所述给定场景词的最终召回商家。
根据本申请的第三方面，提供了一种电子设备，包括：
处理器、存储器以及存储在所述存储器上并可在所述处理器上运行的获取机程序，所述处理器执行所述程序时实现前述的场景化商家召回方法。
根据本申请的第四方面，提供了一种可读存储介质，当所述存储介质中的指令由电子设备的处理器执行时，使得电子设备能够执行前述的场景化商家召回方法。
根据本申请的场景化商家召回方法，可以获取给定场景词，并将所述给定场景词扩展为包含N个词的关键词集合，N为自然数；基于所述关键词集合中的各个关键词，从平台商家中确定初始种子商家；根据所述关键词集合，对所述初始种子商家进行过滤，得到高精度种子商家；根据所述高精度种子商家，确定所述给定场景词的最终召回商家，取得了降低时间成本以及人力成本，同时兼顾精度和召回率并提高适应性的有益效果。
上述说明仅是本申请技术方案的概述，为了能够更清楚了解本申请的技术手段，而可依照说明书的内容予以实施，并且为了让本申请的上述和其它目的、特征和优点能够更明显易懂，以下特举本申请的具体实施方式。
附图说明
通过阅读下文优选实施方式的详细描述，各种其他的优点和益处对于本领域普通技术人员将变得清楚明了。附图仅用于示出优选实施方式的目的，而并不认为是对本申请的限制。而且在整个附图中，用相同的参考符号表示相同的部件。在附图中：
图1示出了根据本申请一个实施例的一种场景化商家召回方法的步骤流程图；
图2示出了根据本申请一个实施例的一种场景化商家召回方法的步骤流程图；
图3示出了根据本申请一个实施例的一种场景化商家召回装置的结构示意图；以及
图4示出了根据本申请一个实施例的电子设备的示意图；
图5示出了用于执行根据本申请的方法的另一电子设备的框图；以及，
图6示出了用于保持或者携带实现根据本申请的方法的程序代码的存储单元。
具体实施例
下面将参照附图更详细地描述本公开的示例性实施例。虽然附图中显示了本公开的示例性实施例，然而应当理解，可以以各种形式实现本公开而不应被这里阐述的实施例所限制。相反，提供这些实施例是为了能够更透彻地理解本公开，并且能够将本公开的范围完整的传达给本领域的技术人员。
实施例一
详细介绍本申请实施例提供的一种场景化商家召回方法。
参照图1，示出了本申请实施例中一种场景化商家召回方法的步骤流程图。
步骤110，获取给定场景词，并将所述给定场景词扩展为包含N个词的关键词集合，N为自然数。
本申请实施例中所述的给定场景词用于指示当前商家召回场景。在实际应用中，如果直接基于给定场景词进行商家召回，召回率较低。因此在本申请实施例中，为了提高召回率，在获取得到给定场景词之后，可以将给定场景词扩展为包含N个词的关键词集合。其中，N为自然数。一般而言，针对同一场景的给定场景词具体可以包括至少一个词语，而且N的具体取值可以根据需求进行预先设置，对此本申请实施例不加以限定。
而且，在本申请实施例中，可以通过任何可用的语义相似度方法将定场景词扩展为包含N个词的关键词集合。例如，可以通过词袋模型，将定场景词扩展为包含N个词的关键词集合。其中，在信息检索中，词袋(Bag of words，BOW)模型假定对于一个文本，忽略其词序和语法、句法，将其仅仅看做是一个词集合，或者说是词的一个组合，文本中每个词的出现都是独立的，不依赖于其他词是否出现，或者说当这篇文章的作者在任意一个位置选择一个词汇都不受前面句子的影响而独立选择的。
而且，在本申请实施例中，给定场景词的关键词集合中也可以包含相应的给定场景词在内，当然也可以不包含相应的给定场景词，具体可以根据需求进行预先设置，对此本申请实施例不加以限定。
可选地，在本申请实施例中，所述步骤110进一步可以包括：
子步骤111，获取给定场景词，并通过word2vec获取与所述给定场景词的相关度最高的N个词构建所述给定场景词的关键词集合。
Word2vec是Google在2013年年中开源的一款将词表征为实数值向量的高效工具,其利用深度学习的思想，可以通过训练，把对文本内容的处理简化为K维向量空间中的向量运算，而向量空间上的相似度可以用来表示文本语义上的相似度。Word2vec输出的词向量可以被用来做很多NLP(Natural Language Processing自然语言处理)相关的工作，比如聚类、找同义词、词性分析等等。如果换个思路，把词当做特征，那么Word2vec就可以把特征映射到K维向量空间，可以为文本数据寻求更加深层次的特征表示。
Word2vec使用的是Distributed representation(分布式表示)的词向量表示方式。Distributed representation最早由Hinton在1986年提出。其基本思想是通过训练将每个词映射成K维实数向量(K一般为模型中的超参数)，通过词之间的距离(比如余弦相似度、欧氏距离等)来判断它们之间的语义相似度。其采用一个三层的神经网络，输入层-隐层-输出层。有个核心的技术是根据词频用Huffman(哈夫曼)编码，使得所有词频相似的词隐藏层激活的内容基本一致，出现频率越高的词语，他们激活的隐藏层数目越少，这样有效的降低了计算的复杂度。而Word2vec大受欢迎的一个原因正是其高效性。这个三层神经网络本身是对语言模型进行建模，但也同时获得一种单词在向量空间上的表示，而这个副作用才是Word2vec的真正目标。Word2Vec的基本思想是把自然语言中的每一个词，表示成一个统一意义统一维度的短向量。
在本申请实施例中，可以直接使用word2vec官方文档中计算给定任意词的Top N个词的公式直接计算得到给定场景词的相似度最高的前N个相似词，进而可以构建得到给定场景词的关键词集合。
步骤120，基于所述关键词集合中的各个关键词，从平台商家中确定初始种子商家。
关键词集合中的每个词都是一个关键词，那么此时为了扩大召回商家的数量以及范围，可以分别基于关键词集合中的每个关键词，从平台商家中确 定初始种子商家。其中的平台商家可以包括进行商家召回所涉及的平台中所包含的全部商家。
具体的，可以基于关键词集合中的每个关键词与平台商家的匹配度，确定初始种子商家；或者，可以将关键词集合中的全部关键词作为一个整体，然后基于全部关键词与平台商家的匹配度，确定初始种子商家；等等。而且，与平台商家的匹配度可以包括与平台商家的各个属性的匹配度。其中的属性可以包括商家名、推荐菜、团购信息、评论，等等。在本申请实施例中，可以分别从平台商家中筛选出与每个关键词的匹配度大于等于第二预设匹配度的商家作为相应关键词对应的初始召回商家，或者是分别从平台商家中筛选出与每个关键词的匹配度最高的前S个商家作为相应关键词对应的初始召回商家，等等。而且如果分别获取得到各个关键词对应的初始召回商家，那么还需要将各个关键词对应的初始召回商家汇总，得到相应的关键词集合所对应的初始种子商家。其中，S为大于1的整数，第二预设匹配度以及S的具体取值可以根据需求进行预先设置，对此本申请实施例不加以限定。
或者，在本申请实施例中，根据相关场景在语料中的稀疏度，在确定初始种子商家时可以直接使用关键词集合中的各个关键词在商家评论中直接命中，并通过命中计数来筛选初始种子商家。具体的，如果在平台商家的任意一条商家评论中包含关键词集合中的任意一个关键词，则确认相应的关键词命中相应的平台商家一次，进而可以将关键词集合中的各个关键词对同一平台商家的全部商家评论的命中次数进行加权求和，得到关键词集合对相应的平台商家的总命中次数。进而可以基于关键词集合对各个平台商家的总命中次数从高到低的顺序，选取前R个平台商家作为初始种子商家。其中，R的具体取值可以根据需求进行预先设置，对此本申请实施例不加以限定。各个关键词对应的命中次数的权重可以为相应关键字与给定场景词的相关度权重。
步骤130，根据所述关键词集合，对所述初始种子商家进行过滤，得到高精度种子商家。
如前述步骤可知，为了提高召回率会将给定场景词扩展为一个关键词集合。但是并不能完全保证经关键词集合中的每个关键词得到的初始种子商家的精确度，而且初始种子商家到最终召回结果的算法具有极强的传播能力，如果一个初始种子商家的精确度不够，那么在后续步骤中其会被继续放大，从而影响商家召回的精确度，所以在本申请实施例中对初始种子商家的精度要求很高，但是对初始种子商家的召回率可以无任何要求。
因此，在本申请实施例中，为了能够在提高商家召回率的同时提高召回准确性，在获取得到初始召回商家之后，还需要根据相应的关键词集合，对所述初始种子商家进行过滤，得到高精度种子商家。具体的，根据关键词集合，对所述初始种子商家进行过滤的过滤策略，可以根据需求进行预先设置，对此本申请实施例不加以限定。
例如，可以设置过滤策略为如果初始种子商家的商家名中包含相应的关键词集合中的至少一个关键词，则确认相应的初始种子商家为高精度种子商家，而如果初始种子商家的商家名中不包含相应的关键词集合中的任意一个关键词，则可以将相应的初始种子商家过滤掉，那么经过滤后保留的初始种子商家则为高精度种子商家。
步骤140，根据所述高精度种子商家，确定所述给定场景词的最终召回商家。
如前述，确定的高精度种子商家的精度较高，但是数量有限，因此在确定了高精度种子商家之后，还需要进一步扩大召回范围，也即确定给定场景词的最终召回商家。
具体的,可以首先确定各个平台商家与相应的每个高精度种子商家的相似度。很明显，某一平台商家与高精度种子商家的相似度越高，则说明该平台商家与高精度种子商家对应同一场景的可能性就越大。因此，在本申请实施例中，可以基于各个平台商家与高精度种子商家之间的相似度，确定相应的给定场景词的最终召回商家。而且，具体的可以根据任何可用方法确认平台商家与高精度种子商家之间的相似度，对此本申请实施例不加以限定。
例如，可以通过比对平台商家与高精度种子商家各个属性之间的匹配程度，进而确定平台商家与高精度种子商家的相似度；或者是分别将平台商家与高精度种子商家的各个属性表征为向量形式，进而分别比较同一属性对应的各个向量之间的相似度，从而确定平台商家与高精度种子商家的相似度；等等。其中的属性可以包括商家名、评论、特色菜，等等中的至少一种，具体的上述属性可以根据需求进行预先设置，对此本申请实施例不加以限定。
而且，需要说明的是，其中基于给定场景词确定的高精度商家可以默认为相应的给定场景词对应的最终召回商家，当然也可以根据需求设置不将基于给定场景词确定的高精度商家默认为最终召回商家，对此本申请实施例不加以限定。
可选地，在本申请实施例中，在所述步骤140之后，还可以包括：
步骤150，根据用户搜索记录确定每个场景的搜索词集合。
经上述的步骤生成的最终召回商家是基于给定场景词的召回商家，但是在实际应用中，针对不同场景不同用户的查询词并不完全相同，那么在确定了各个场景的最终召回商家之后，为了能够根据不同用户的个性化查询词返回相应的查询结果，还需要进一步根据场景具体覆盖的查询词以及业务要求，对每个具体的查询词覆盖的召回商家进行规则化过滤，产生与具体查询词以及业务高度相关的召回商家集合。
那么首先则需要根据用户搜索记录确定每个场景的搜索词集合。例如，根据用户搜索记录筛选出的用户针对不同的场景可能会搜索或查询的词，从而得到各个场景的搜索词集合。对于每个场景，都可以通过用户搜索记录给定一个属于相应的各个场景的搜索词集合，这个集合是闭集合，通常集合里都会包含相应场景的给定场景词，只要用户输入搜索词集合里的任意至少一个词就会触发对相应场景的搜索流程。
步骤160，根据所述搜索词集合以及业务需求，确定所述搜索词集合的过滤策略。
在实际应用中，由于搜索目的、搜索发起者等不同，相应的业务需求也会有所不同。例如，可以分为对公搜索业务需求、个人搜索业务需求、高精度业务需求、高召回率业务需求，等等。
在本申请实施例中，其中的过滤策略、业务需求的具体内容，以及过滤策略与搜索词集合和业务需求的对应关系可以根据需求进行预先设置，对此本申请实施例不加以限定。当然，在本申请实施例中，也可以仅根据所述搜索词集合或者是业务需求，确定所述搜索词集合的过滤策略，对此本申请实施例不加以限定。
步骤170，基于所述过滤策略对所述搜索词集合中的各个搜索词的最终召回商家进行过滤，得到与所述搜索词以及所述业务需求匹配的召回商家集合。
在确定了过滤策略之后，则可以基于所述过滤策略对所述搜索词集合中的各个搜索词的最终召回商家进行过滤，得到与所述搜索词以及所述业务需求匹配的召回商家集合。由于所述搜索词集合中包含相应场景的给定场景词，因此可以基于所述过滤策略对所述给定场景词的最终召回商家进行过滤，得到与所述给定场景词以及所述业务需求匹配的召回商家集合。例如，用户搜索“老北京火锅”时，需要分别首先将“火锅”改写成“涮肉”、“铜锅涮肉”、“涮羊肉”，然后同时使用“火锅”、“涮肉”、“涮羊肉”、“铜锅涮肉”从老北京的最终召回商家里面过滤出“火锅”的结果。
另外，在本申请实施例中，为了避免经由上述的步骤110-140得到的最终召回商家相对于现有的场景化商家召回方法有遗漏，还可以在步骤140得到最终召回商家的基础上进行补召回。具体的可以将传统基于关键词命中过滤等方法召回的商家与由上述步骤110-140召回的商家进行比较，提取出经传统方法额外召回的商家，并与由上述步骤110-140召回的商家进行合并生成最终召回结果。
根据本申请的场景化商家召回方法，可以获取给定场景词，并将所述给定场景词扩展为包含N个词的关键词集合，N为自然数；基于所述关键词集合中的各个关键词，从平台商家中确定初始种子商家；根据所述关键词集合，对所述初始种子商家进行过滤，得到高精度种子商家；根据所述高精度种子商家，确定所述给定场景词的最终召回商家。由此降低了商家召回的时间成本以及人力成本，同时兼顾了商家召回的精度和召回率。
实施例二
详细介绍本申请实施例提供的一种场景化商家召回方法。
参照图2，示出了本申请实施例中一种场景化商家召回方法的步骤流程图。
步骤210，通过预选的高质量商家评论训练LDA模型并生成word-topic矩阵。
为了能够生成高精度且高召回的结果，首先需要从全方位去描述给定场景，需要将给定场景词扩展为带有相关度权重的关键词集合，所以我们选择了考虑全局上下文且发散性更强的LDA(Latent Dirichlet Allocation)模型。LDA模型是一种文档主题生成模型，也称为一个三层贝叶斯概率模型，包含词、主题和文档三层结构。那么，通过预选的高质量商家评论对LDA模型进行训练则可以生成相应的word-topic(主题词)矩阵。
其中的高质量商家评论是预先根据需求选定的不同商家的商家评论。例如，对于美团点评而言，高质量商家评论可以包括在存储之前已经进行过多道工序进行过人为筛选或者自动筛选的美团点评的评论，所以高质量商家评论的质量是有保证的。
可选地，在本申请实施例中，所述步骤210进一步可以包括：
子步骤211，对所述预选的高质量商家评论进行预处理。
在实际应用中，用户在进行评价时可能会输入无关紧要的内容，例如一些特殊符号、停用词等等，而这些内容都可能影响最终的训练结果。因此，在本申请实施例中，为了提高训练效果，可以先对预选的高质量商家评论进 行预处理。而且，其中的预处理可以包括但不限于去除特殊字符处理、分词处理中的至少一种。其中特殊字符可以包括但不限于特殊符号、低频词、停用词等等。
其中，特殊符号可以包括相对于传统或常用的符号外，使用频率较少字符且难以直接输入的符号，比如：数学符号，单位符号，制表符等。低频词是指出现次数较少的词，在本申请实施例中可以预先设置一出现频率阈值，如果评论中某一词的出现次数少于该出现频率阈值，则可以确认其为低频词。停用词可以包括在信息检索中，为节省存储空间和提高搜索效率，在处理自然语言数据(或文本)之前或之后会自动过滤掉某些字或词，这些字或词即被称为Stop Words(停用词)。在本申请实施例中，可以基于现有的停用词表执行去除停用词处理，和/或可以预先设置不同业务的停用词。例如，对于美食业务商家名中“餐厅”就是停用词，对于团购字段文本中“满XXX减XXX”就是停用词，比如写真业务中“写真”就是停用词，等等。
另外，在本申请实施例中，可以通过任何可用分词处理方法对高质量商家评论进行分词处理，对此本申请实施例不加以限定。
子步骤212，利用预处理后的高质量商家评论训练LDA模型并生成word-topic矩阵。
在对高质量商家评论进行预处理之后，则可以进一步，利用预处理后的高质量商家评论训练LDA模型并生成word-topic矩阵，矩阵中包括标识主题的词。其中，在本申请实施例中，还可以预先设置LDA模型中的隐含话题数K。其中的隐含话题数可以理解为训练后得到的word-topic矩阵中的topic(主题)数量。在本申请实施例中，基于多轮实验结果，优选地可以设置K为2500。
步骤220，基于所述LDA模型将所述给定场景词表示为多个原始主题的混合成分。
利用训练好的LDA模型可以对任意一篇给定文档生成一个基于原始主题的混合成分。其中的原始主题也即上述的训练后的LDA模型得到的topic。
此时可以将给定场景词输入训练后的LDA模型，进而得到给定场景词在各个原始主题下的概率，也即将给定场景词表示为多个原始主题的混合成分。
例如，假设给定场景词是“动漫”，那么基于LDA模型可以得到“动漫”在各个原始主题下的概率，也即“动漫”的主题分布如下：
1297:0.89444275499683
1320:0.044182015916109305
617:0.013558739479591326
1024:0.010889265702068057
1885:0.009441767594398784
182:0.005473920591330183
......
其中，左侧一列代表原生主题编号，右侧一列代表给定场景词在相应原生主题下的概率。
步骤230，基于所述混合成分所表征的条件概率，确定所述给定场景词与所述word-topic矩阵中各个词的相关度权重。
其中，上述给定场景词在原始主题下的概率可以表示为P(topic k|d)，其中topic k表示为第k个原始主题，d表示给定场景词。而且如果给定场景词具体包含多个词，那么d则可以表示给定场景词构成的文档。
进而可以基于上述的P(topic k|d)得到给定场景词与所述word-topic矩阵中各个词的相关度权重为：
Figure PCTCN2019098210-appb-000001
其中，word表示为word-topic矩阵中的任意一个词，K表示原始主题的总数。P(word|topic k)的取值可以在获取得到word-topic矩阵时同时获取得到。
步骤240，从所述word-topic矩阵的各个词中选取与所述给定场景词的相关度权重最高的N个词构建所述给定场景词的关键词集合。
在计算得到给定场景词与所述word-topic矩阵中各个词的相关度权重之后，很明显相关度权重越大则说明相应的词与给定场景词越接近，越可能对应同一场景。因此，在本申请实施例中，可以进一步从word-topic矩阵的各个词中选取与给定场景词的相关度权重最高的N个词构建所述给定场景词的关键词集合。其中N的取值可以根据需求进行预先设置，对此本申请实施例不加以限定。
例如，对于前述的给定场景词“动漫”，根据其与word-topic矩阵中每个词的相关度权重，可以得到其关键词集合为{可爱,卡通,漫画,动画,海贼王,榻榻米,趣味,玩偶,桌游,女仆,日式,居酒屋,二次元,玩具,公仔,任务,贴纸,彩绘,机器猫,卡哇伊,奇幻}，此时N的取值为21。
步骤250，基于预设的重排算法重新确定所述word-topic矩阵中各个 词与所述给定场景词的相关度权重。
步骤260，根据重新确定的相关度权重更新所述关键词集合。
在实际应用中，基于上述的相关度权重得到的给定场景词的关键词集合中可能会包含一些实际相关性不大的关键词，那么此时则可以基于预设的重排算法重新确定所述word-topic矩阵中各个词与所述给定场景词的相关度权重。其中的重排算法可以根据需求进行预先设置，对此本申请实施例不加以限定。
例如，可以基于给定场景词与word-topic矩阵中各个词的词义、出现场景等等多个属性重新确定给定场景词与word-topic矩阵中各个词的相关度权重，等等。
在重新确定了word-topic矩阵中各个词与所述给定场景词的相关度权重之后，则可以基于重新确定的相关度权重更新给定场景词对应的关键词集合。
而且，如果重排后得到的给定场景词的关键词集合仍达不到要求精度，则可以通过少量人工筛选出高精度的关键词集合。
步骤270，确定各个平台商家的向量表示，以及所述关键词集合的向量表示。
为了进一步基于关键词集合中的各个关键词从平台商家中筛选出初始种子商家，则需要确定各个平台商家与关键词集合的相似度。但是基于文本形式的平台商家和关键词集合很难准确地确定相似度。因此在本申请实施例中，可以首先确定各个平台商家的向量表示，以及所述关键词集合的向量表示。在本申请实施例中，可以通过任何可用方法确定各个平台商家的向量表示，以及所述关键词集合的向量表示，对此本申请实施例不加以限定。
可选地，在本申请实施例中，所述步骤270进一步可以包括:
子步骤271，将同一平台商家的评论聚合为一个评论文本，并将所述关键词集合中的全部关键词聚合为一个关键词文本。
在本申请实施例中，为了充分考虑同一商家平台的各个评论，以及关键词集合中的全部关键词，进而确定初始种子商家。在本申请实施例中，可以先将同一平台商家的评论聚合为一个评论文本，并将所述关键词集合中的全部关键词聚合为一个关键词文本。
其中，聚合为一个评论文本的属于同一平台商家的评论可以为经过预先筛选的较高质量的且属于相应平台商家的评论，当然也可以为获取到的相应平台商家的全部评论，具体的可以根据需求进行预先设置，对此本申请实施 例不加以限定。
子步骤272，基于所述评论文本确定所述平台商家的向量表示，并基于所述关键词文本确定所述关键词集合的向量表示。
在获取得到评论文本以及关键词文本之后，则可以分别基于评论文本确定所述平台商家的向量表示，以及基于关键词文本确定所述平台商家的向量表示。
可选地，在本申请实施例中，所述子步骤272进一步可以包括:通过所述LDA模型，根据所述评论文本确定所述平台商家的向量表示，并通过所述LDA模型，根据所述关键词文本确定所述关键词集合的向量表示；或者，通过word2vec模型，根据所述评论文本确定所述平台商家的向量表示，并根据所述关键词文本确定所述关键词集合的向量表示。其中，通过LDA模型得到的向量表示也可以称为主题向量。
由于LDA模型只对长文本有好的表征效果，而经聚合后得到的评论文本以及关键词文本可以作为长文本，因此在本申请实施例中，可以使用前述训练后的LDA模型根据评论文本推断出相应平台商家的向量表示，接着将关键词文本作为长文本同样通过LDA模型推断出相应的关键词集合的向量表示。
或者，在本申请实施例中，根据场景在语料中的稀疏度，也可以通过word2vec模型，根据所述评论文本确定所述平台商家的向量表示，并通过word2vec模型，根据所述关键词文本确定所述关键词集合的向量表示。word2vec也叫word embeddings，中文名“词向量”，作用就是将自然语言中的字词转为计算机可以理解的稠密向量(Dense Vector)。
在本申请实施例中，如果关键词文本和/或评论文本为长文本，则可以优选地利用LDA模型确定向量表示，而如果关键词文本和/或评论文本为短文本，则可以优选地利用word2vec模型确定向量表示。当然，具体的可以根据需求进预先设置，对此本申请实施例不加以限定。
步骤280，获取各个所述平台商家的向量表示与所述关键词集合的向量表示之间的相似度。
进一步的，则可以基于向量表示确定各个平台商家的向量表示与所述关键词集合的向量表示之间的相似度。其中的相似度可以包括但不限于余弦相似度、欧氏距离相似度、交叉熵相似度、相对熵距离相似度中的任意一种。
而且，根据相似度的具体种类，则可以采用相应的任何可用方法确定相应的相似度，对此本申请实施例不加以限定。
步骤290，选取相似度最高的前M个平台商家作为所述给定场景词的初 始种子商家。
很明显，平台商家与关键词集合的相似度越高，则说明相应的平台商家与关键词集合对应同一场景的可能性越高，因此，在确定了各个平台商家与关键词集合的相似度之后，则可以选取相似度最高的前M个平台商家作为相应给定场景词的初始种子商家。其中M为大于1的整数，M的取值可以根据需求进行预先设置，对此本申请实施例不加以限定。
步骤2110，根据所述关键词集合，对所述初始种子商家进行过滤，得到高精度种子商家。
步骤2120，根据所述高精度种子商家，进行多次迭代运算，并以每次迭代运算得到的精确召回商家的总和作为所述给定场景词的最终召回商家。
为了进一步扩大召回商家数量，以提高召回率，在本申请实施例中，在确定得到高精度种子商家之后，还可以基于高精度种子商家，进行多次迭代运算，并以每次迭代运算得到的精确召回商家的总和作为所述给定场景词的最终召回商家。同样地，最终召回商家中还可以包括初始获得的高精度种子商家，当然也可以不包括初始获得的高精度种子商家，具体的可以根据需求进行预先设置，对此本申请实施例不加以限定。
其中，每次迭代运算的过程包括：
步骤S1，根据所述平台商家的各个属性以及所述属性的预设权重，确定每个所述高精度种子商家与所述平台商家中每个其他商家的最终相似度；
在实际应用中，每个平台商家又可以包括多个不同的属性，例如商家名、特色菜、评价等等属性，而且不同的属性在确定相似度时的重要程度并不完全一致，因此在本申请实施例中，可以预先设置平台商家的各个属性的预设权重，进而可以根据所述平台商家的各个属性以及所述属性的预设权重，确定每个所述高精度种子商家与所述平台商家中每个其他商家的最终相似度。其中，某一给定场景词对应的其他商家可以包括平台商家中除去相应的给定场景词对应的高精度种子商家之外的其他平台商家。
具体的，可以分别确定高精度种子商家与其他商家在不同属性下的相似度，进而基于各个属性的预设权重，对不同属性的相似度进行加权求和，从而得到高精度种子商家与相应的其他商家的最终相似度。
而且，在本申请实施例中，为了提高相似度的准确性，同样可以先对各个属性对应的文本数据进行预处理。此时的预处理可以包括用特定停用词对文本数据进行清洗，其中的特定停用词可以包括基于业务需求预先设定的停用词，例如美食业务商家名中“餐厅”就是特定停用词，比如团购字段文本 中“满XXX减XXX”就是特定停用词，比如写真业务中“写真”就是特定停用词，等等。
步骤S2，基于预设的相似度阈值，确定所述高精度种子商家对所述其他商家所投的票数；所述票数的权重为所述高精度种子商家与所述平台商家的最终相似度；
其中的相似度阈值可以根据需求进行预先设置，而且相似度预设的设置可以与相似度的种类有关，对此本申请实施例不加以限定。具体的，如果某一其他商家与一高精度种子商家的最终相似度超过预设的相似度阈值，则确定相应的高精度种子商家会对相应的其他商家投票，而且所投票数的权重为相应平台商家与相应的高精度种子商家的最终相似度；而如果某一其他商家与一高精度种子商家的最终相似度未超过预设的相似度阈值，则确定相应的高精度种子商家不会对相应的其他商家投票。
例如，假设其他商家B、C和D与高精度种子商家A1之间的最终相似度分别为0.6、0.8、0.3，其他商家B、C和D与高精度种子商家A2之间的最终相似度分别为0.7、0.4、0.3，且相似度数值越大表示越相似。如果预设的相似度阈为0.5，那么可以确定高精度种子商家A1对其他商家B所投票数为0.6，高精度种子商家A1对其他商家C所投票数为0.8，高精度种子商家A1不会对其他商家D投票，高精度种子商家A2对其他商家B所投票数为0.7，高精度种子商家A2不会对其他商家C、D投票。那么可以得到此时的其他商家B得到的票数为1.3，其他商家C得到的票数为0.8，其他商家D得到的票数为0。
步骤S3，选取票数最高的前L个其他商家作为所述初始召回商家，L为自然数；
很明显，得到票数越高则说明相应的其他商家与高精度种子商家的相似度越高，因此在本申请实施例中可以选取票数最高的前L个其他商家作为相应的给定场景词的初始召回商家。其中L的取值可以根据需求进行预先设置，对此本申请实施例不加以限定。
如前述的其他商家B、C和D与高精度种子商家A1、A2，如果此时设置L的取值为2，那么则可以选取得到初始召回商家依次为其他商家B、C。
步骤S4，根据所述关键词集合，对所述初始召回商家进行过滤，得到精确召回商家；
为了提高召回精确率，同样地可以根据关键词集合，对所述初始召回商家进行过滤，得到精确召回商家。具体的过滤过程可以与前述的对初始种子 商家的过滤过程类似，在此不加以赘述。
步骤S5，从所述精确召回商家中选取票数最高的第一数量个商家作为新的高精度种子商家，并更新本次迭代的所述高精度种子商家，直至所述精确召回商家的召回精度下降或者所述精确召回商家的数量收敛。
为了进一步扩大召回范围，则可以从经本次迭代得到的精确召回商家中选取第一数量个商家作为新的高精度种子商家，并通过所述新的高精度种子商家更新本次迭代的所述高精度种子商家，然后以更新后的高精度种子商家进入下一次迭代过程，直至得到的精确召回商家的召回精度下降或者是精确召回商家的数量收敛。其中第一数量的具体取值可以根据需求进行预先设置，对此本申请实施例不加以限定。而在更新高精度种子商家之后，则可以基于更新后的高精度种子商家进入下一次迭代过程，也即依次执行上述的步骤S1-S5，以更新后的高精度种子商家，确定每个高精度种子商家与所述平台商家中每个其他商家的最终相似度，进而基于预设的相似度阈值，确定更新后的高精度种子商家对相应的其他商家所投的票数，其中票数的权重为相应的高精度种子商家与平台商家的最终相似度，重新得到初始召回商家并且经过滤得到新的精确召回商家，然后再次更新高精度种子商家，直至当前次迭代的精确召回商家的召回精度下降或者所述精确召回商家的数量收敛。
其中，更新本次迭代的所述高精度种子商家的过程可以为将最新选取的第一数量个高精度种子商家添加到本次迭代过程中所用的高精度种子商家中，或者是以最新选取的第一数量个高精度种子商家替换本次迭代过程中所用的高精度种子商家，等等。具体的可以根据需求进行预先设置，对此本申请实施例不加以限定。
而且，经步骤S4之后，还可以进一步判断所述精确召回商家的召回精度是否下降，或者是所述精确召回商家的数量是否收敛，如果所述精确召回商家的召回精度下降，或者是所述精确召回商家的数量收敛，则可以停止迭代，也无需执行步骤S5；而如果精确召回商家的召回精度未下降，且所述精确召回商家的数量未收敛，则可以执行S5并继续进行迭代。
可选地，在本申请实施例中，所述属性包括商家名、推荐菜、团购信息、评论中的至少一种。
可选地，在本申请实施例中，所述步骤S2进一步可以包括：
子步骤S21，将所述平台商家的各个属性表征为向量，得到所述平台商家的属性向量；
在本申请实施例中，为了计算高精度种子商家与每个其他商家的最终相 似度，而且是基于平台商家的各个属性确定其中的高精度种子商家与所述平台商家中每个其他商家的最终相似度，那么同样地则需要将平台商家的各个属性表征为向量，得到所述平台商家的属性向量。而且，在本申请实施例中，可以通过任何可用方法将属性表征为向量，对此本申请实施例不加以限定。例如，通过前述的LDA模型，word2vec模型，或者是doc2vec模型，等等。
子步骤S22，基于所述属性向量以及所述属性的预设权重，确定每个所述高精度种子商家与所述平台商家中每个其他商家的最终相似度。
在得到各个属性的属性向量，则可以进一步基于属性向量以及所述属性的预设权重，确定每个所述高精度种子商家与所述平台商家中每个其他商家的最终相似度。其中各个属性的预设权重可以根据需求进行预先设置，对此本申请实施例不加以限定。
可选地，在本申请实施例中，所述子步骤S21进一步可以包括：
子步骤S211，基于各个属性的特定停用词对所述属性对应的文本数据进行数据清洗；
在本申请实施例中，可以针对不同的属性设置不同的特定停用词。例如，将美食业务商家名中“餐厅”设置为特定停用词，将团购字段文本中“满XXX减XXX”设置为特定停用词，将写真业务中“写真”设置为特定停用词，等等。那么则可以基于各个属性的特定停用词对所述属性对应的文本数据进行数据清洗。
子步骤S212，通过word2vec将数据清洗后的所述平台商家的商家名表征为向量；
如前述，在本申请实施例中的属性可以包括商家名，而且由于商家名一般字数不会过多，那么此时则可以适用于通过word2vec将数据清洗后的所述平台商家的商家名表征为向量，进而得到商家名对应的属性向量。
子步骤S213，通过doc2vec将数据清洗后的所述平台商家的参考信息分别表征为向量；或者，通过word2vec将数据清洗后的所述平台商家的参考信息中的每个分词分别表征为向量，并获取每个参考信息中的每个分词的向量平均值作为所述参考信息的向量。
而对于推荐菜、团购信息、评论等参考信息而言，其篇幅一般会大于商家名，而且表述上更倾向于文档形式，因此在本申请实施例中，可以通过doc2vec将数据清洗后的所述平台商家的参考信息分别表征为向量，进而得到各个参考信息对应的属性向量。
或者，为了方便利用word2vec将数据清洗后的文本数据表征为向量， 还可以对平台商家的参考信息进行分词处理，进而将其处理为多个分词，那么则可以通过word2vec将数据清洗后的所述平台商家的参考信息中的每个分词分别表征为向量，并获取每个参考信息中的每个分词的向量平均值作为所述参考信息的向量。
例如，对于推荐菜而言，则可以对数据清洗后的推荐菜进行分词处理，进而通过word2vec将推荐菜的每个分词表征为向量，并获取推荐菜的全部分词的向量平均值，进而得到推荐菜所对应的文本数据的属性向量。而对于其他的参考信息也类似。
可选地，在本申请实施例中，所述子步骤S22进一步可以包括：
子步骤S221，基于所述属性向量，分别获取所述高精度种子商家与所述其他商家在每个属性下的相似度；
在本申请实施例中，可以基于平台商家的属性向量，分别获取平台商家中的与当前的给定场景词对应的高精度种子商家与其他商家在每个属性下的相似度。例如商家名相似度，特色菜相似度、团购信息相似度，等等。
子步骤S222，按照所述属性的预设权重，将各个属性的相似度加权求和，得到所述高精度种子商家与所述其他商家的最终相似度。
在获取得到各个属性下的相似度之后，则可以基于各个属性的预设权重，将各个属性的相似度进行加权求和，进而得到相应的高精度种子商家与相应的其他商家的最终相似度。
可选地，在本申请实施例中，所述相似度可以包括但不限于余弦相似度、欧氏距离相似度、交叉熵相似度、相对熵距离相似度中的任意一种。
可选地，在本申请实施例中，在所述步骤S5之前，还可以包括：
步骤S6，根据本次迭代运算得到的精确召回商家的数量，以及本次迭代运算的迭代次序，确定所述第一数量的取值。
在实际应用中，对于每一次迭代对应的第一数量可以相同，也可以不完全相同，那么如果每次迭代所对应的第一数量不完全相同，则可以根据迭代过程中设置第一数量的取值，从而进一步提高得到的最终召回商家的准确率以及召回率。因此，在本申请实施例中，可以根据本次迭代运算得到的精确召回商家的数量，以及本次迭代运算的迭代次序，确定所述第一数量的取值。具体的本次迭代运算得到的精确召回商家的数量以及本次迭代运算的迭代次序与本次迭代对应的第一数量之间的对应关系可以根据需求进行预先设置，对此本申请实施例不加以限定。
例如，可以设置第一数量的取值为T*step*n_iteration，step为经 本次迭代运算确定的精确召回商家的数量，n_iteration为本次迭代运算的迭代次序，T为预设参数。其中T的取值可以根据需求进行设置，对此本申请实施例不加以限定，例如可以直接设置T为1，等等。
第一方面，根据本申请的场景化商家召回方法，可以获取给定场景词，并将所述给定场景词扩展为包含N个词的关键词集合；基于所述关键词集合中的各个关键词，从平台商家中确定初始种子商家；根据所述关键词集合，对所述初始种子商家进行过滤，得到高精度种子商家；根据所述高精度种子商家，确定所述给定场景词的最终召回商家。由此降低了商家召回的时间成本以及人力成本，同时兼顾了商家召回的精度和召回率。
第二方面，根据本申请的场景化商家召回方法，可以通过预选的高质量商家评论训练LDA模型并生成word-topic矩阵；基于所述LDA模型将所述给定场景词表示为多个原始主题的混合成分；基于所述混合成分所表征的条件概率，确定所述给定场景词与所述word-topic矩阵中各个词的相关度权重；从所述word-topic矩阵的各个词中选取与所述给定场景词的相关度权重最高的N个词构建所述给定场景词的关键词集合。并且，基于预设的重排算法重新确定所述word-topic矩阵中各个词与所述给定场景词的相关度权重；根据重新确定的相关度权重更新所述关键词集合。以及，对所述预选的高质量商家评论进行预处理；利用预处理后的高质量商家评论训练LDA模型并生成word-topic矩阵；其中，所述预处理包括去除特殊字符处理、分词处理；所述特殊字符包括特殊符号、低频词、停用词中的至少一种。从而可以进一步提高商家召回的准确率以及召回率。
第三方面，根据本申请的场景化商家召回方法，可以确定各个平台商家的向量表示，以及所述关键词集合的向量表示；获取各个所述平台商家的向量表示与所述关键词集合的向量表示之间的相似度；选取相似度最高的前M个平台商家作为所述给定场景词的初始种子商家。将同一平台商家的评论聚合为一个评论文本，并将所述关键词集合中的全部关键词聚合为一个关键词文本；基于所述评论文本确定所述平台商家的向量表示，并基于所述关键词文本确定所述关键词集合的向量表示。同样可以进一步提高商家召回的准确率以及召回率。
第四方面，根据本申请的场景化商家召回方法，可以根据所述高精度种子商家，进行多次迭代运算，并以每次迭代运算得到的精确召回商家的总和作为所述给定场景词的最终召回商家；其中，每次迭代运算的过程包括：根据所述平台商家的各个属性以及所述属性的预设权重，确定每个所述高精度 种子商家与所述平台商家中每个其他商家的最终相似度；基于预设的相似度阈值，确定所述高精度种子商家对所述其他商家所投的票数；所述票数的权重为所述高精度种子商家与所述平台商家的最终相似度；选取票数最高的前L个其他商家作为所述初始召回商家；根据所述关键词集合，对所述初始召回商家进行过滤，得到精确召回商家；从所述精确召回商家中选取票数最高的第一数量个商家作为新的高精度种子商家，并更新本次迭代的所述高精度种子商家，直至所述精确召回商家的召回精度下降或者所述精确召回商家的数量收敛。并且，将所述平台商家的各个属性表征为向量，得到所述平台商家的属性向量；基于所述属性向量以及所述属性的预设权重，确定每个所述高精度种子商家与所述平台商家中每个其他商家的最终相似度。基于各个属性的特定停用词对所述属性对应的文本数据进行数据清洗；通过word2vec将数据清洗后的所述平台商家的商家名表征为向量；通过doc2vec将数据清洗后的所述平台商家的参考信息分别表征为向量；或者，通过word2vec将数据清洗后的所述平台商家的参考信息中的每个分词分别表征为向量，并获取每个参考信息中的每个分词的向量平均值作为所述参考信息的向量；其中，所述参考信息包括推荐菜、团购信息、评论中的至少一种。基于所述属性向量，分别获取所述高精度种子商家与所述其他商家在每个属性下的相似度；按照所述属性的预设权重，将各个属性的相似度加权求和，得到所述高精度种子商家与所述其他商家的最终相似度。根据本次迭代运算得到的精确召回商家的数量，以及本次迭代运算的迭代次序，确定所述第一数量的取值。同样可以进一步降低商家召回的时间成本以及人力成本，同时提高商家召回的精度和召回率。
对于方法实施例，为了简单描述，故将其都表述为一系列的动作组合，但是本领域技术人员应该知悉，本申请实施例并不受所描述的动作顺序的限制，因为依据本申请实施例，某些步骤可以采用其他顺序或者同时进行。其次，本领域技术人员也应该知悉，说明书中所描述的实施例均属于优选实施例，所涉及的动作并不一定是本申请实施例所必须的。
实施例三
详细介绍本申请实施例提供的一种场景化商家召回装置。
参照图3，示出了本申请实施例中一种场景化商家召回装置的结构示意图。
关键词集合扩展模块310，用于获取给定场景词，并将所述给定场景词扩展为包含N个词的关键词集合，N为自然数；
初始种子商家确定模块320，用于基于所述关键词集合中的各个关键词，从平台商家中确定初始种子商家；
高精度种子商家确定模块330，用于根据所述关键词集合，对所述初始种子商家进行过滤，得到高精度种子商家；
最终召回商家确认模块340，用于根据所述高精度种子商家，确定所述给定场景词的最终召回商家。
第一方面，根据本申请的场景化商家召回装置，可以获取给定场景词，并将所述给定场景词扩展为包含N个词的关键词集合；基于所述关键词集合中的各个关键词，从平台商家中确定初始种子商家；根据所述关键词集合，对所述初始种子商家进行过滤，得到高精度种子商家；根据所述高精度种子商家，确定所述给定场景词的最终召回商家。由此降低了商家召回的时间成本以及人力成本，同时兼顾了商家召回的精度和召回率。
本发明本申请实施例中还公开了一种电子设备，包括：
处理器、存储器以及存储在所述存储器上并可在所述处理器上运行的获取机程序，所述处理器执行所述程序时实现前述的场景化商家召回方法。如图4所示，电子设备400包括：存储器410和处理器420，存储器410与处理器420之间通过总线通信连接，存储器410中存储有计算机程序，该计算机程序可在处理器420上运行，进而实现本申请上述任一实施例所述的方法中的步骤。
本申请实施例中还公开了一种可读存储介质，当所述存储介质中的指令由电子设备的处理器执行时，使得电子设备能够执行前述的场景化商家召回方法。
本申请实施例公开的场景化商家召回装置能够实现前述场景化商家召回方法，对于装置实施例而言，由于其与方法实施例基本相似，装置的各个模块的具体实施方式与相应的方法步骤对应，所以描述的比较简单，相关之处参见方法实施例的部分说明即可。
在此提供的算法和显示不与任何特定获取机、虚拟系统或者其它设备固有相关。各种通用系统也可以与基于在此的示教一起使用。根据上面的描述，构造这类系统所要求的结构是显而易见的。此外，本申请也不针对任何特定编程语言。应当明白，可以利用各种编程语言实现在此描述的本申请的内容，并且上面对特定语言所做的描述是为了披露本申请的最佳实施方式。
在此处所提供的说明书中，说明了大量具体细节。然而，能够理解，本申请的实施例可以在没有这些具体细节的情况下实践。在一些实例中，并未 详细示出公知的方法、结构和技术，以便不模糊对本说明书的理解。
类似地，应当理解，为了精简本申请并帮助理解各个发明方面中的一个或多个，在上面对本申请的示例性实施例的描述中，本申请的各个特征有时被一起分组到单个实施例、图、或者对其的描述中。然而，并不应将该公开的方法解释成反映如下意图：即所要求保护的本申请要求比在每个权利要求中所明确记载的特征更多的特征。更确切地说，如下面的权利要求书所反映的那样，申请方面在于少于前面公开的单个实施例的所有特征。因此，遵循具体实施方式的权利要求书由此明确地并入该具体实施方式，其中每个权利要求本身都作为本申请的单独实施例。
本领域那些技术人员可以理解，可以对实施例中的设备中的模块进行自适应性地改变并且把它们设置在与该实施例不同的一个或多个设备中。可以把实施例中的模块或单元或组件组合成一个模块或单元或组件，以及此外可以把它们分成多个子模块或子单元或子组件。除了这样的特征和/或过程或者单元中的至少一些是相互排斥之外，可以采用任何组合对本说明书(包括伴随的权利要求、摘要和附图)中公开的所有特征以及如此公开的任何方法或者设备的所有过程或单元进行组合。除非另外明确陈述，本说明书(包括伴随的权利要求、摘要和附图)中公开的每个特征可以由提供相同、等同或相似目的的替代特征来代替。
此外，本领域的技术人员能够理解，尽管在此所述的一些实施例包括其它实施例中所包括的某些特征而不是其它特征，但是不同实施例的特征的组合意味着处于本申请的范围之内并且形成不同的实施例。例如，在下面的权利要求书中，所要求保护的实施例的任意之一都可以以任意的组合方式来使用。
本申请的各个部件实施例可以以硬件实现，或者以在一个或者多个处理器上运行的软件模块实现，或者以它们的组合实现。本领域的技术人员应当理解，可以在实践中使用微处理器或者数字信号处理器(DSP)来实现根据本申请实施例的场景化商家召回设备中的一些或者全部部件的一些或者全部功能。本申请还可以实现为用于执行这里所描述的方法的一部分或者全部的设备或者装置程序(例如，获取机程序和获取机程序产品)。这样的实现本申请的程序可以存储在获取机可读介质上，或者可以具有一个或者多个信号的形式。这样的信号可以从因特网网站上下载得到，或者在载体信号上提供，或者以任何其他形式提供。
例如，图5示出了可以实现根据本申请的方法的电子设备，该电子设备 传统上包括处理器520和以存储器510形式的计算机程序产品或者计算机可读存储介质。所述电子设备可以为PC机、移动终端、个人数字助理、平板电脑等。所述电子设备为计算处理设备。所述存储器510可以是诸如闪存、EEPROM(电可擦除可编程只读存储器)、EPROM、硬盘或者ROM之类的电子存储器。存储器510具有用于执行上述方法中的任何方法步骤的程序代码的存储空间5101。例如，用于程序代码的存储空间5101可以包括分别用于实现上面的方法中的各种步骤的程序代码5102。所述计算机程序产品或者计算机可读存储介质，其上存储有计算机程序的程序代码，该程序代码被处理器520执行时实现如本申请实施例一和实施例二所述的数据读取方法。这些程序代码可以从一个或者多个计算机程序产品中读出或者写入到这一个或者多个计算机程序产品中。这些计算机程序产品包括诸如硬盘，紧致盘(CD)、存储卡或者软盘之类的程序代码载体。这样的计算机程序产品或者计算机可读存储介质通常为如参考图6所述的便携式或者固定存储单元。该存储单元用于保持或者携带实现根据本申请的方法的程序代码，该存储单元可以具有与图5的电子设备中的存储器510类似布置的存储段、存储空间等。程序代码可以例如以适当形式进行压缩。通常，存储单元包括计算机可读代码6102’，即可以由例如诸如520之类的处理器读取的代码，这些代码当电子设备运行时，导致该电子设备执行上面所描述的方法中的各个步骤。
本文中所称的“一个实施例”、“实施例”或者“一个或者多个实施例”意味着，结合实施例描述的特定特征、结构或者特性包括在本公开的至少一个实施例中。此外，请注意，这里“在一个实施例中”的词语例子不一定全指同一个实施例。
在此处所提供的说明书中，说明了大量具体细节。然而，能够理解，本申请的实施例可以在没有这些具体细节的情况下被实践。在一些实例中，并未详细示出公知的方法、结构和技术，以便不模糊对本说明书的理解。
在权利要求中，不应将位于括号之间的任何参考符号构造成对权利要求的限制。单词“包含”不排除存在未列在权利要求中的元件或步骤。位于元件之前的单词“一”或“一个”不排除存在多个这样的元件。本申请可以借助于包括有若干不同元件的硬件以及借助于适当编程的计算机来实现。在列举了若干装置的单元权利要求中，这些装置中的若干个可以是通过同一个硬件项来具体体现。单词第一、第二、以及第三等的使用不表示任何顺序。可将这些单词解释为名称。
应该注意的是上述实施例对本申请进行说明而不是对本申请进行限制， 并且本领域技术人员在不脱离所附权利要求的范围的情况下可设计出替换实施例。在权利要求中，不应将位于括号之间的任何参考符号构造成对权利要求的限制。单词“包含”不排除存在未列在权利要求中的元件或步骤。位于元件之前的单词“一”或“一个”不排除存在多个这样的元件。本申请可以借助于包括有若干不同元件的硬件以及借助于适当编程的获取机来实现。在列举了若干装置的单元权利要求中，这些装置中的若干个可以是通过同一个硬件项来具体体现。单词第一、第二、以及第三等的使用不表示任何顺序。可将这些单词解释为名称。

Claims (18)

  1. 一种场景化商家召回方法，包括：
    获取给定场景词，并将所述给定场景词扩展为包含N个词的关键词集合，N为自然数；
    基于所述关键词集合中的各个关键词，从平台商家中确定初始种子商家；
    根据所述关键词集合，对所述初始种子商家进行过滤，得到高精度种子商家；
    根据所述高精度种子商家，确定所述给定场景词的最终召回商家。
  2. 根据权利要求1所述的方法，所述获取给定场景词，并将所述给定场景词扩展为包含N个词的关键词集合的步骤，包括：
    通过预选的高质量商家评论训练LDA模型并生成word-topic矩阵；
    基于所述LDA模型将所述给定场景词表示为多个原始主题的混合成分；
    基于所述混合成分所表征的条件概率，确定所述给定场景词与所述word-topic矩阵中各个词的相关度权重；
    从所述word-topic矩阵的各个词中选取与所述给定场景词的相关度权重最高的N个词构建所述给定场景词的关键词集合。
  3. 根据权利要求2所述的方法，在所述基于所述条件概率以及预设的概率阈值，确定所述给定场景词的关键词集合的步骤之后，还包括：
    基于预设的重排算法重新确定所述word-topic矩阵中各个词与所述给定场景词的相关度权重；
    根据重新确定的相关度权重更新所述关键词集合。
  4. 根据权利要求2所述的方法，所述通过预选的高质量商家评论训练LDA模型并生成word-topic矩阵的步骤，包括：
    对所述预选的高质量商家评论进行预处理；
    利用预处理后的高质量商家评论训练LDA模型并生成word-topic矩阵；
    其中，所述预处理包括去除特殊字符处理、分词处理；所述特殊字符包括特殊符号、低频词、停用词中的至少一种。
  5. 根据权利要求1所述的方法，所述基于所述关键词集合中的各个关键词，确定初始种子商家的步骤，包括：
    确定各个平台商家的向量表示，以及所述关键词集合的向量表示；
    获取各个所述平台商家的向量表示与所述关键词集合的向量表示之间的相似度；
    选取相似度最高的前M个平台商家作为所述给定场景词的初始种子商家，M为自然数。
  6. 根据权利要求5所述的方法，所述确定各个商家的向量表示，以及所述关键词集合的向量表示的步骤，包括：
    将同一平台商家的评论聚合为一个评论文本，并将所述关键词集合中的全部关键词聚合为一个关键词文本；
    基于所述评论文本确定所述平台商家的向量表示，并基于所述关键词文本确定所述关键词集合的向量表示。
  7. 根据权利要求6所述的方法，所述基于所述评论文本确定所述平台商家的向量表示，并基于所述关键词文本确定所述关键词集合的向量表示的步骤，包括：
    通过LDA模型，根据所述评论文本确定所述平台商家的向量表示和根据所述关键词文本确定所述关键词集合的向量表示；
    或者，通过word2vec模型，根据所述评论文本确定所述平台商家的向量表示和根据所述关键词文本确定所述关键词集合的向量表示。
  8. 根据权利要求1所述的方法，所述根据所述高精度种子商家，确定所述给定场景词的最终召回商家的步骤，包括：
    根据所述高精度种子商家，进行多次迭代运算，并以每次迭代运算得到的精确召回商家的总和作为所述给定场景词的最终召回商家；
    其中，每次迭代运算的过程包括：
    根据所述平台商家的各个属性以及所述属性的预设权重，确定每个所述高精度种子商家与所述平台商家中每个其他商家的最终相似度；
    基于预设的相似度阈值，确定所述高精度种子商家对所述其他商家所投的票数；所述票数的权重为所述高精度种子商家与所述平台商家的最终相似度；
    选取票数最高的前L个其他商家作为所述初始召回商家，L为自然数；
    根据所述关键词集合，对所述初始召回商家进行过滤，得到精确召回商家；
    从所述精确召回商家中选取票数最高的第一数量个商家作为新的高精 度种子商家，并更新本次迭代的所述高精度种子商家，直至所述精确召回商家的召回精度下降或者所述精确召回商家的数量收敛。
  9. 根据权利要求8所述的方法，所述根据所述平台商家的各个属性以及所述属性的预设权重，确定每个所述高精度种子商家与所述平台商家中每个其他商家的最终相似度的步骤，包括：
    将所述平台商家的各个属性表征为向量，得到所述平台商家的属性向量；
    基于所述属性向量以及所述属性的预设权重，确定每个所述高精度种子商家与所述平台商家中每个其他商家的最终相似度。
  10. 根据权利要求9所述的方法，所述将所述平台商家的各个属性表征为向量，得到所述平台商家的属性向量的步骤，包括：
    基于各个属性的特定停用词对所述属性对应的文本数据进行数据清洗；
    通过word2vec将数据清洗后的所述平台商家的商家名表征为向量；
    通过doc2vec将数据清洗后的所述平台商家的参考信息分别表征为向量，或者，通过word2vec将数据清洗后的所述平台商家的参考信息中的每个分词分别表征为向量，并获取每个参考信息中的每个分词的向量平均值作为所述参考信息的向量。
  11. 根据权利要求9所述的方法，所述基于所述属性向量以及预设的所述属性的预设权重，确定每个所述高精度种子商家与所述平台商家中每个其他商家的最终相似度的步骤，包括：
    基于所述属性向量，分别获取所述高精度种子商家与所述其他商家在每个属性下的相似度；
    按照所述属性的预设权重，将各个属性的相似度加权求和，得到所述高精度种子商家与所述其他商家的最终相似度。
  12. 根据权利要求8所述的方法，在所述从所述精确召回商家中选取票数最高的第一数量个商家作为新的高精度种子商家，并更新本次迭代的所述高精度种子商家的步骤之前，还包括：
    根据本次迭代运算得到的精确召回商家的数量，以及本次迭代运算的迭代次序，确定所述第一数量的取值。
  13. 根据权利要求1所述的方法，其特征在于，在所述根据所述高精度种子商家，确定所述给定场景词的最终召回商家的步骤之后，还包括：
    根据用户搜索记录确定每个场景的搜索词集合；
    根据所述搜索词集合以及业务需求，确定所述搜索词集合的过滤策略；
    基于所述过滤策略对所述搜索词集合中的各个搜索词的最终召回商家进行过滤，得到与所述搜索词以及所述业务需求匹配的召回商家集合。
  14. 根据权利要求1所述的方法，所述获取给定场景词，并将所述给定场景词扩展为包含N个词的关键词集合的步骤，包括：
    获取给定场景词，并通过word2vec获取与所述给定场景词的相关度最高的N个词构建所述给定场景词的关键词集合。
  15. 一种场景化商家召回装置，包括：
    关键词集合扩展模块，用于获取给定场景词，并将所述给定场景词扩展为包含N个词的关键词集合，N为自然数；
    初始种子商家确定模块，用于基于所述关键词集合中的各个关键词，从平台商家中确定初始种子商家；
    高精度种子商家确定模块，用于根据所述关键词集合，对所述初始种子商家进行过滤，得到高精度种子商家；
    最终召回商家确认模块，用于根据所述高精度种子商家，确定所述给定场景词的最终召回商家。
  16. 一种电子设备，包括：
    处理器、存储器以及存储在所述存储器上并可在所述处理器上运行的获取机程序，所述处理器执行所述获取机程序时实现如权利要求1-14中的任一项所述的场景化商家召回方法。
  17. 一种可读存储介质，当所述存储介质中的指令由电子设备的处理器执行时，使得电子设备能够执行如权利要求1-14中的任一项所述的场景化商家召回方法。
  18. 一种计算机程序，包括计算机可读代码，当所述计算机可读代码在电子设备上运行时，导致所述电子设备执行根据权利要求1-14中的任一项所述的场景化商家召回方法。
PCT/CN2019/098210 2018-12-17 2019-07-29 场景化商家召回 Ceased WO2020125015A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811545314.4A CN109815392A (zh) 2018-12-17 2018-12-17 场景化商家召回方法、装置、电子设备及可读存储介质
CN201811545314.4 2018-12-17

Publications (1)

Publication Number Publication Date
WO2020125015A1 true WO2020125015A1 (zh) 2020-06-25

Family

ID=66601718

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/098210 Ceased WO2020125015A1 (zh) 2018-12-17 2019-07-29 场景化商家召回

Country Status (2)

Country Link
CN (1) CN109815392A (zh)
WO (1) WO2020125015A1 (zh)

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109815392A (zh) * 2018-12-17 2019-05-28 北京三快在线科技有限公司 场景化商家召回方法、装置、电子设备及可读存储介质
CN110751504B (zh) * 2019-09-11 2025-05-27 汉海信息技术(上海)有限公司 商户聚类方法及装置、商户推荐方法及装置、电子设备
CN110851571B (zh) * 2019-11-14 2022-11-25 拉扎斯网络科技（上海）有限公司 数据处理方法、装置、电子设备及计算机可读存储介质
CN111178970B (zh) * 2019-12-30 2023-06-30 微梦创科网络科技（中国）有限公司 广告投放的方法及装置、电子设备和计算机可读存储介质
CN111159393B (zh) * 2019-12-30 2023-10-10 电子科技大学 一种基于lda和d2v进行摘要抽取的文本生成方法
CN111984698B (zh) * 2020-08-07 2021-03-19 北京芯盾时代科技有限公司 一种信息预测方法、装置及存储介质
CN114116977B (zh) * 2021-11-24 2025-08-01 Oppo广东移动通信有限公司 融合多特征的相似缺陷推荐的方法及相关装置
CN115952333B (zh) * 2023-01-18 2026-04-14 支付宝(杭州)数字服务技术有限公司 结合文本表征的召回方法和系统
CN118445264B (zh) * 2024-07-08 2024-09-03 长春医学高等专科学校(长春职工医科大学长春市医学情报所) 一种电子归档数据存储方法及系统

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7831582B1 (en) * 2005-08-23 2010-11-09 Amazon Technologies, Inc. Method and system for associating keywords with online content sources
CN102629257A (zh) * 2012-02-29 2012-08-08 南京大学 一种基于关键字的电子商务网站商品推荐方法
CN103577432A (zh) * 2012-07-26 2014-02-12 阿里巴巴集团控股有限公司 一种商品信息搜索方法和系统
CN104063523A (zh) * 2014-07-21 2014-09-24 焦点科技股份有限公司 一种电子商务搜索评分与排名的方法及系统
CN105069086A (zh) * 2015-07-31 2015-11-18 焦点科技股份有限公司 一种优化电子商务商品搜索的方法及系统
CN107862027A (zh) * 2017-10-31 2018-03-30 北京小度信息科技有限公司 检索意图识别方法、装置、电子设备及可读存储介质
CN108241699A (zh) * 2016-12-26 2018-07-03 百度在线网络技术（北京）有限公司 用于推送信息的方法和装置
CN109815392A (zh) * 2018-12-17 2019-05-28 北京三快在线科技有限公司 场景化商家召回方法、装置、电子设备及可读存储介质

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20090216734A1 (en) * 2008-02-21 2009-08-27 Microsoft Corporation Search based on document associations
KR101057075B1 (ko) * 2008-12-31 2011-08-16 주식회사 솔트룩스 정보 검색 방법 및 이를 수행할 수 있는 프로그램이 수록된컴퓨터로 읽을 수 있는 기록 매체
CN103365904B (zh) * 2012-04-05 2018-01-09 阿里巴巴集团控股有限公司 一种广告信息搜索方法和系统
CN105677769B (zh) * 2015-12-29 2018-01-05 广州神马移动信息科技有限公司 一种基于潜在狄利克雷分配（lda）模型的关键词推荐方法和系统
CN107491518B (zh) * 2017-08-15 2020-08-04 北京百度网讯科技有限公司 一种搜索召回方法和装置、服务器、存储介质

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7831582B1 (en) * 2005-08-23 2010-11-09 Amazon Technologies, Inc. Method and system for associating keywords with online content sources
CN102629257A (zh) * 2012-02-29 2012-08-08 南京大学 一种基于关键字的电子商务网站商品推荐方法
CN103577432A (zh) * 2012-07-26 2014-02-12 阿里巴巴集团控股有限公司 一种商品信息搜索方法和系统
CN104063523A (zh) * 2014-07-21 2014-09-24 焦点科技股份有限公司 一种电子商务搜索评分与排名的方法及系统
CN105069086A (zh) * 2015-07-31 2015-11-18 焦点科技股份有限公司 一种优化电子商务商品搜索的方法及系统
CN108241699A (zh) * 2016-12-26 2018-07-03 百度在线网络技术（北京）有限公司 用于推送信息的方法和装置
CN107862027A (zh) * 2017-10-31 2018-03-30 北京小度信息科技有限公司 检索意图识别方法、装置、电子设备及可读存储介质
CN109815392A (zh) * 2018-12-17 2019-05-28 北京三快在线科技有限公司 场景化商家召回方法、装置、电子设备及可读存储介质

Also Published As

Publication number Publication date
CN109815392A (zh) 2019-05-28

Similar Documents

Publication Publication Date Title
WO2020125015A1 (zh) 场景化商家召回
CN113297369B (zh) 基于知识图谱子图检索的智能问答系统
CN109815308B (zh) 意图识别模型的确定及检索意图识别方法、装置
CN106649818B (zh) 应用搜索意图的识别方法、装置、应用搜索方法和服务器
CN110162630B (zh) 一种文本去重的方法、装置及设备
JP6657124B2 (ja) 会話理解システムのためのセッションコンテキストモデリング
WO2018049960A1 (zh) 一种为文本信息匹配资源的方法及装置
CN104199833B (zh) 一种网络搜索词的聚类方法和聚类装置
CN115129864B (zh) 文本分类方法、装置、计算机设备和存储介质
CN110096567A (zh) 基于qa知识库推理的多轮对话回复选择方法、系统
EP3314461A1 (en) Learning entity and word embeddings for entity disambiguation
CN111522886B (zh) 一种信息推荐方法、终端及存储介质
CN109829045A (zh) 一种问答方法和装置
WO2013049529A1 (en) Method and apparatus for unsupervised learning of multi-resolution user profile from text analysis
CN109408802A (zh) 一种提升句向量语义的方法、系统及存储介质
CN113988057A (zh) 基于概念抽取的标题生成方法、装置、设备及介质
CN105740448B (zh) 面向话题的多微博时序文摘方法
CN107895303A (zh) 一种基于ocean模型的个性化推荐的方法
CN110728135B (zh) 文本主题标引方法、装置、电子设备及计算机存储介质
CN106776782B (zh) 基于人工智能的语义相似度获取方法及装置
CN102929962B (zh) 一种搜索引擎的评测方法
CN108062355B (zh) 基于伪反馈与tf-idf的查询词扩展方法
CN103226601A (zh) 一种图片搜索的方法和装置
WO2016210203A1 (en) Learning entity and word embeddings for entity disambiguation
CN115329083A (zh) 文档分类方法、装置、计算机设备和存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19899225

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19899225

Country of ref document: EP

Kind code of ref document: A1