Vertical search method, device and system for procurement scenarios

Abstract

The application discloses a kind of vertical search method, device and system of procurement scene, it is related to wisdom supply chain technical field.The specific embodiment of the method includes: in response to user input search text, at least one keyword is obtained from search text, the user intent corresponding to keyword in preset dictionary is determined as the first initial selection intent of user;User's historical procurement data is input into preset procurement periodicity rule model and procurement timeliness rule model, and the user intent indicated by the periodicity rule and timeliness rule matched by historical procurement data is determined as the second initial selection intent of user;According to first initial selection intent and second initial selection intent, the current intent of user is determined;Query is carried out using current intent and search text, and at least one recall result found in the query is sorted, and the search result corresponding to search text is returned to user.The embodiment can carry out efficient and high-quality search for procurement scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent supply chain, and particularly relates to a vertical search method, device and system for a procurement scene. BACKGROUND

[0002] Vertical search is a professional search method for a certain industry, and provides search services for specific groups of people, specific fields and specific needs. At present, the vertical search in the enterprise procurement scene generally uses a SQL (Structured Query Language) query method based on a database table and a social e-commerce search method. The former has poor search efficiency, accuracy and recall rate. The latter is mainly aimed at the search scene of individual users and cannot reflect the search rules of the enterprise procurement scene, and thus is not suitable for being used as a search engine for the procurement scene. SUMMARY

[0003] Therefore, the embodiments of the present application provide a vertical search method, device and system for a procurement scene, which can perform efficient and high-quality search for the procurement scene.

[0004] To achieve the above object, according to one aspect of the present application, a vertical search method for a procurement scene is provided.

[0005] The vertical search method for a procurement scene of the embodiments of the present application comprises the following steps: in response to a user inputting a search text, at least one keyword is obtained from the search text, and a user intention corresponding to the keyword in a preset dictionary is determined as a first initial intention of the user; wherein the dictionary contains a mapping relationship between a plurality of words and a plurality of user intentions; historical procurement data of the user is input into a preset procurement periodicity rule model and a procurement timeliness rule model, and a user intention indicated by a periodicity rule and a timeliness rule matched by the historical procurement data is determined as a second initial intention of the user; wherein the procurement periodicity rule model contains at least one periodicity rule, and the procurement timeliness rule model contains at least one timeliness rule; the current intention of the user is determined according to the first initial intention and the second initial intention; the current intention and the search text are used for query, and at least one recall result obtained by the query is sorted to obtain a search result corresponding to the search text, which is returned to the user.

[0006] Optionally, the periodic rule and the timeliness rule comprise a discriminant condition part and a user intention part; and the inputting the historical purchase data of the user into the preset purchase periodic rule model and the purchase timeliness rule model, and determining the user intention indicated by the periodic rule and the timeliness rule matched with the historical purchase data as the second initial intention of the user, comprises: inputting the historical purchase data of the user in a first historical time period into the purchase periodic rule model, and determining the user intention indicated by the periodic rule with the discriminant condition part consistent with the historical purchase data as the second initial intention; inputting the historical purchase data of the user in a second historical time period into the purchase timeliness rule model, and determining the user intention indicated by the timeliness rule with the discriminant condition part consistent with the historical purchase data as the second initial intention; wherein the starting time of the first historical time period is earlier than the starting time of the second historical time period.

[0007] Optionally, the method further comprises: inputting the segmented search text into a pre-trained intention classification model based on machine learning to obtain a third initial intention of the user; and the determining the current intention of the user according to the first initial intention and the second initial intention comprises: determining the current intention in combination with the first initial intention, the second initial intention and the third initial intention.

[0008] Optionally, the querying by using the current intention and the search text comprises: querying in a preset database by using the segmented search text, adjusting the query scores of the query results according to the item category indicated by the current intention, and determining a plurality of query results with the largest query scores as the recall results of the database; inputting the word vector features of the segmented search text into a pre-trained vector recall model based on machine learning, adjusting the query scores of the output results according to the item category indicated by the current intention, and determining a plurality of output results with the largest query scores as the recall results of the vector recall model.

[0009] Optionally, the sorting of the at least one recall result comprises: for any item category in the recall results, determining the sorting score of the any item category according to the feature data extracted from the behavior log of the user and whether the item category is the item category indicated by the current intention, and determining the arrangement order between the item categories in the recall results according to the sorting scores; and for the same item category in the recall results, calculating the sorting scores of the recall results belonging to the item category by using a pre-trained sorting model based on machine learning, and determining the arrangement order of the recall results in the item category according to the sorting scores.

[0010] Optionally, the obtaining the search result corresponding to the search text comprises: selecting a preset number of the recall results arranged in front as preselected results; and adjusting arrangement orders of the preselected results by using a preset reordering rule to obtain the search result.

[0011] Optionally, the obtaining the at least one keyword from the search text comprises: determining an association word corresponding to the input text from a preset association word library during the inputting of the search text by the user, and outputting the association word to the user; determining whether the search text is an error text after the search is initiated by the user; correcting the search text when the search text is the error text; determining at least one synonym corresponding to the search text in a preset synonym library as an expansion word of the search text, and performing intent recognition and query of the recall results according to the expansion word.

[0012] To achieve the above object, according to another aspect of the present application, a vertical search device for a procurement scenario is provided.

[0013] The vertical search device for a procurement scenario can comprise: a first intent recognition unit configured to obtain at least one keyword from the search text in response to the input of the search text by the user, and determine a user intent corresponding to the keyword in a preset dictionary as a first initial intent of the user; wherein the dictionary contains mapping relationships between a plurality of words and a plurality of user intents; a second intent recognition unit configured to input historical procurement data of the user into a preset procurement periodicity rule model and a procurement timeliness rule model, and determine a user intent indicated by a periodicity rule and a timeliness rule matched by the historical procurement data as a second initial intent of the user; wherein the procurement periodicity rule model contains at least one periodicity rule, and the procurement timeliness rule model contains at least one timeliness rule; and a search unit configured to determine a current intent of the user according to the first initial intent and the second initial intent, perform a query by using the current intent and the search text, and sort at least one recall result obtained by the query to obtain a search result corresponding to the search text, and return the search result to the user.

[0014] To achieve the above object, according to another aspect of the present application, a vertical search system for a procurement scenario is provided.

[0015] The vertical search system of the procurement scenario can include an application layer, a service layer, a data layer, a log collection layer, and a data calculation layer. The service layer deploys a plurality of microservices for supporting the application layer, including an intent recognition service, a recall service, and a ranking service. The log collection layer collects user behavior logs through the application layer. The data calculation layer performs calculations according to the behavior logs and stores the calculation results in the data layer. The calculation results include historical procurement data of the user. The intent recognition service obtains at least one keyword from a search text input by the user, determines a user intent corresponding to the keyword in a preset dictionary as a first preliminary intent of the user. The dictionary contains a mapping relationship between a plurality of words and a plurality of user intents. The intent recognition service inputs historical procurement data of the user in the data layer into a preset procurement periodicity rule model and a procurement timeliness rule model, and determines a user intent indicated by a periodicity rule and a timeliness rule matched by the historical procurement data as a second preliminary intent of the user. The procurement periodicity rule model contains at least one periodicity rule, and the procurement timeliness rule model contains at least one timeliness rule. The intent recognition service determines a current intent of the user according to the first preliminary intent and the second preliminary intent. The recall service queries the data layer using the current intent and the search text. The ranking service ranks at least one recall result obtained by the query to obtain a search result corresponding to the search text. The application layer returns the search result to the user.

[0016] Optionally, the behavior logs include search logs and order logs, the calculation results further include feature data calculated from the search logs, and the data layer includes a full-text retrieval database for performing the query and a cache database for storing the calculation results. The log collection layer includes a relational database for storing a part of the search logs and the order logs and a file system for storing another part of the search logs. The data calculation layer includes a stream computing engine for calculating the search logs and a batch computing engine for calculating the order logs. The stream computing engine stores the calculated feature data in the cache database, and the batch computing engine stores the calculated historical procurement data in the cache database.

[0017] Optionally, the system can further comprise: a first conversion module, a second conversion module and a message queue module; wherein the first conversion module converts the search log stored in the relational database into stream data, and sends the stream data to the stream computing engine through the message queue module; the second conversion module converts the search log stored in the file system into stream data, and sends the stream data to the stream computing engine through the message queue module.

[0018] Optionally, the vertical search system can further comprise: an algorithm model layer comprising an intent classification model supporting the intent recognition service, a vector recall model supporting the recall service and a ranking model supporting the ranking service; the intent classification model, the vector recall model and the ranking model are all pre-trained machine learning models.

[0019] Optionally, the microservice can further comprise: an association word service, a correction service, a synonym service, a word segmentation service and a reordering service; the full-text retrieval database comprises Elastic Search (ES), the cache database comprises Redis, the relational database comprises Mysql, the file system comprises Hadoop Distributed File System (HDFS), the stream computing engine comprises Flink, the batch computing engine comprises Spark, the first conversion module comprises Flume, the second conversion module comprises Beats, and the message queue module comprises Kafka.

[0020] To achieve the above object, according to another aspect of the present application, an electronic device is provided.

[0021] The electronic device of the present application comprises: one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the vertical search method for procurement scenarios provided by the present application.

[0022] To achieve the above object, according to still another aspect of the present application, a computer readable storage medium is provided.

[0023] The computer readable storage medium of the present application has a computer program stored thereon, when the program is executed by a processor, the vertical search method for procurement scenarios provided by the present application is implemented.

[0024] According to the technical solution of the present application, the embodiments in the above application have the following advantages or beneficial effects:

[0025] The pre-deployed intention recognition service accurately judges the real intention of the user from multiple aspects, thereby embodying the search rule of the procurement scene to improve the search efficiency and quality. Specifically, in a first aspect, after the user inputs the search text, the keywords are determined from the search text, and the user intention corresponding to the keywords in the preset dictionary is determined as the first initial intention of the user. In a second aspect, the historical procurement data of the user is input into a preset procurement periodicity rule model and a procurement timeliness rule model, the procurement periodicity rule model reflects the periodicity procurement rule embodied by the user in a long time, and the procurement timeliness rule model reflects the short-term timeliness procurement rule of the user, and the two models are combined to determine the second initial intention of the user. In a third aspect, the search text is input into a pre-trained intention classification model after being segmented, thereby obtaining the third initial intention of the user. Finally, the current intention (i.e., the real intention) of the user is determined by combining the first initial intention, the second initial intention and the third initial intention, and thereafter, the current intention of the user is used to guide the subsequent recall and sorting process, thereby realizing a high-availability vertical search engine suitable for the procurement scene. In addition, the application also provides a specific architecture of the vertical search system suitable for the procurement scene, including an application layer, a service layer, a data layer, a log collection layer and a data calculation layer and the like, and these layers realize data query and log collection and processing through interaction and cooperation, thereby forming a strong support for the above vertical search engine.

[0026] The further effects of the above non-conventional optional mode will be described in the following with reference to the specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0027] The accompanying drawings are used to better understand the application and do not constitute an improper limitation on the application. Among them:

[0028] Figure 1 is the main step schematic diagram of the vertical search method of the procurement scene in the embodiment of the application;

[0029] Figure 2 is the micro-service calling schematic diagram of the embodiment of the application;

[0030] Figure 3 is the component schematic diagram of the vertical search device of the procurement scene in the embodiment of the application;

[0031] Figure 4 is the architecture schematic diagram of the vertical search system of the procurement scene in the embodiment of the application;

[0032] Figure 5 is an exemplary system architecture diagram to which embodiments of the application can be applied;

[0033] Figure 6 is an electronic device structure schematic diagram used to realize the vertical search method of the procurement scene in the embodiment of the application. DETAILED DESCRIPTION

[0034] The exemplary embodiments of the present application will be described hereinafter with reference to the accompanying drawings, in which the various features of the present application are shown to provide a full understanding of the present application. It should be noted that the embodiments of the present application and the technical features in the embodiments can be combined with each other without conflict.

[0035] It should be noted that the embodiments of the present application and the technical features in the embodiments can be combined with each other without conflict.

[0036] Figure 1 is a schematic diagram of main steps of the vertical search method for the procurement scenario according to an embodiment of the present application.

[0037] As shown in Figure 1 , the vertical search method for the procurement scenario according to an embodiment of the present application can be specifically executed according to the following steps:

[0038] Step S101: In response to a user inputting a search text, at least one keyword is obtained from the search text, and a user intent corresponding to the keyword in a preset dictionary is determined as a first selected intent of the user.

[0039] The vertical search method provided by the embodiment of the present application can be applied to the enterprise procurement scenario. Unlike the personal user purchase scenario, the procurement scenario often has a long-time range periodicity rule and a short-time range similarity rule. For example, a user purchased a certain item on a certain day last year, and there is a certain probability that the user will purchase the same item or a related item near this date this year; a user purchased a certain item a week ago, and there is a certain probability that the user will purchase the same item or a related item currently. The above examples are procurement demand prediction rules. In actual application, the procurement demand prediction rules are generally determined by procurement experts according to the specific circumstances of the industry where the user is located and the procurement item.

[0040] In an actual scenario, when a user inputs a search text in a search box, the server can determine an association word corresponding to the input text from a preset association word library, and display the association word in a drop-down list of the search box, from which the user can select the search text to be input. It can be understood that the association word library contains a plurality of association words with high use frequency. During the process of the user inputting the search text, the server can match each word in the association word library with the input text as a prefix, and sequentially display the matched association words to the user according to use frequency and other factors. The above association word function can be realized by a preset association word service in the server.

[0041] After the user determines the complete search text through the input mode or the selection of the association word and initiates the search through the click of the search button or the like, the server can determine whether the search text is an incorrect text based on preset rules or a pre-trained machine learning model. If it is determined that the search text is an incorrect text, the search text is corrected, and the user can be prompted to select whether to use the corrected text or the original search text. If it is determined that the search text is correct, no processing is performed. The above correction function can be implemented through a preset correction service in the server.

[0042] Thereafter, the server can determine at least one synonym of the search text (which can be the corrected search text or the original search text) in the preset synonym library as an expansion word of the search text, and independently perform subsequent processing similar to the search text on any expansion word, i.e., independently perform the subsequent tokenization, intent recognition, and recall processes to be described below on any expansion word. The recall results of any expansion word and the recall results of the search text are aggregated to perform the sorting and reordering processes. The tokenization, intent recognition, recall, sorting, and reordering processes are only illustrated below using the search text as an example. The processing of the expansion word is similar to the search text. It should be noted that, compared with the search text, the recall results of the expansion word have a lower weight in the source dimension in the sorting and reordering processes. It can be understood that the above search text can be various characters, and the above association word and synonym can be word vocabularies of various characters, such as Chinese characters and words, and English words. The above synonym function can be implemented through a preset synonym service in the server.

[0043] Thereafter, the server uses a preset tokenization service to tokenize the search text, and performs intent recognition based on the tokenization results. The above intent recognition is used to determine the purchasing demand of the preset user, i.e., to determine the real intent (hereinafter referred to as the current intent) of the user based on the search text and the historical behavior of the user. The above intent can correspond to one or more categories of items. In actual applications, the current intent of the user can be comprehensively determined from two aspects. The first aspect is based on a dictionary. Specifically, the server uses a keyword extraction tool to obtain at least one keyword from the search text, and determines the user intent corresponding to the keyword in the preset dictionary as the first selected intent of the user. It can be understood that the above dictionary contains a mapping relationship between multiple words and multiple user intents, which can be obtained by statistical analysis of a large amount of historical data. For example, the search text is “high-quality mature Pu”, the server can determine the keyword “mature Pu” therefrom, and determine “tea” corresponding to “mature Pu” in the above dictionary as the first selected intent of the user.

[0044] The second aspect is based on rules. Specifically, the server inputs the historical purchase data of the user into a preset purchase periodicity rule model and a purchase timeliness rule model, and determines the user intent indicated by the periodicity rule and the timeliness rule matched with the historical purchase data as the second preliminary intent of the user (step S102 above). In a specific application, the historical purchase data can include purchase-related data of the user in a historical period, such as purchase items, purchase quantities, purchase times, etc., which can be extracted from the user's behavior log. The user's behavior log can include a search log and an order log. The search log can record the search text, exposed items, clicked items, etc. of a search process, and the order log can record the purchase-related data of the user. The above historical purchase data is extracted from the order log.

[0045] Generally, the above purchase periodicity rule model is used to embody the periodicity rule in the long time range as described above, and contains at least one periodicity rule; and the above purchase timeliness rule model is used to embody the similarity rule in the short time range as described above, and contains at least one timeliness rule. The above periodicity rule and timeliness rule can be determined by purchase experts after fully mining the field characteristics and scene characteristics of enterprise purchase searches. Specifically, the periodicity rule and timeliness rule include a judgment condition part and a user intent part. For example, a certain periodicity rule is: if a user purchased electrolytic capacitors in February a year ago, then if the user initiates a search in February or March this year (judgment condition part), the intent is capacitors and resistors (user intent part); and a certain timeliness rule is: if a user purchased women's summer tops within a month, then if the user initiates a search between May and August this year (judgment condition part), the intent is women's summer tops and skirts (user intent part). It can be understood that the timeliness rule can embody the seasonality and immediacy rule of the current search time in addition to the similarity rule in the short time range as described above.

[0046] In the second aspect, the server can input the historical purchase data of the user in a first historical time period into the purchase periodicity rule model, compare each historical purchase data with each periodicity rule of the purchase periodicity rule model, and determine the user intent indicated by the periodicity rule whose judgment condition part matches the historical purchase data as the second preliminary intent; and input the historical purchase data of the user in a second historical time period into the purchase timeliness rule model, compare each historical purchase data with each timeliness rule of the purchase timeliness rule model, and determine the user intent indicated by the timeliness rule whose judgment condition part matches the historical purchase data as the second preliminary intent. The first historical time period is a time period earlier than the current time, and the second historical time period is a time period later than the current time, i.e., the start time of the first time period is earlier than the start time of the second historical time period. For example, the first time period is one year ago to the current time, and the second time period is one month ago to the current time.

[0047] Step S103: determining the current intent of the user according to the first primary intent and the second primary intent; querying by using the current intent and the search text, and sorting the at least one recall result obtained in the querying, to obtain the search result corresponding to the search text and return the search result to the user.

[0048] In this step, the current intent of the user can be determined in combination with the first primary intent and the second primary intent, for example, the union or intersection of the first primary intent and the second primary intent is determined as the current intent of the user. In specific applications, the current intent of the user can correspond to multiple categories of items, and each category of item can have different weights, which can affect the processing results in the subsequent recall process and sorting process. In actual scenarios, the above weights can be adjusted according to the specific circumstances of the first primary intent and the second primary intent, for example, when the union of the first primary intent and the second primary intent is determined as the current intent of the user, the weight of the category of item corresponding to the intersection of the first primary intent and the second primary intent can be increased.

[0049] As a preferred scheme, the intent recognition can also be further performed by the third aspect. Specifically, the server can input the search text after tokenization into a pre-trained intent classification model based on machine learning, to obtain a third primary intent of the user. The above intent classification model can be a Distilled Bert model based on Bert (Bidirectional Encoder Representation from Transformers, a language model). Considering that the search scenario has a high requirement for the online processing speed of the model, the traditional Bert model has too many parameters, and although it has high accuracy, the model processing speed is relatively slow, therefore, the model distillation technology is adopted in the embodiment of the application, and the Distilled Bert model is distilled based on the Bert model, so that the online processing speed of the model is greatly improved while the accuracy of the model is ensured.

[0050] Thereafter, the server determines the current intent of the user in combination with the first primary intent, the second primary intent and the third primary intent, that is, the union or intersection of the first primary intent, the second primary intent and the third primary intent is determined as the current intent of the user. Through the above steps, the purchasing demand of the user can be comprehensively predicted from the three aspects of the word library, the rule and the machine learning model in combination with the search text and the historical procurement data, the prediction accuracy is improved by means of the periodicity rule and the timeliness rule of the procurement scenario, which is helpful to realize the efficient and high-quality procurement search service. The above intent recognition function can be realized by the pre-set intent recognition service in the server.

[0051] After determining the current intent of the user, the following recall process can be performed. Specifically, the server can recall results from multiple paths. In one path, the server uses the segmentation result of the search text to query in a preset database (e.g., Elastic Search ES), adjusts the query score of the query result according to the item category indicated by the current intent, for example, increases the query score of the corresponding query result by different degrees according to different weights of the item category indicated by the current intent, and finally determines multiple query results with the largest query scores as the recall results of the database. In the recall process based on the above database, the item category indicated by the current intent can also be determined as the query range, and then the database is queried in the query range, and the query result is taken as the recall result of the database; the search text can also be queried in the database first, and then the query result matching the item category indicated by the current intent is arranged in advance, and finally the query result in the front of the preset number is selected as the recall result of the database.

[0052] In another path, the server can input the word vector feature of the segmentation result of the search text into a pre-trained vector recall model based on machine learning, adjust the query score of the output result according to the item category indicated by the current intent, for example, increase the query score of the corresponding output result by different degrees according to different weights of the item category indicated by the current intent, and finally determine multiple output results with the largest query scores as the recall results of the vector recall model. In the recall process based on the above vector recall model, the item category indicated by the current intent can also be determined as the query range, and then the output result of the vector recall model is obtained, and the query range is used for screening, and the output result in the query range is taken as the recall result of the vector recall model; the output result of the vector recall model can also be obtained, and the output result matching the item category indicated by the current intent is arranged in advance, and finally the output result in the front of the preset number is selected as the recall result of the vector recall model.

[0053] In other paths, the current intent of the user and the search text can also be used to query in a preset specific database (e.g., a new product database, a long-tail item database), to obtain recall results similar to the aforementioned intent recognition process of the database. Finally, the recall results of all paths can be integrated, and a subsequent sorting process can be performed. It can be understood that if an expansion word is determined in the previous step, independent queries need to be performed for each expansion word, and finally the recall results based on the search text and the expansion word are selected, and in the selection process, the recall result of the expansion word has a lower weight in the source dimension. Through the above steps, accurate and comprehensive data recall can be achieved based on the search text and the current intent of the user, combined with multiple paths, which helps to improve the search quality. In actual scenarios, the above recall function can be realized by a preset recall service in the server.

[0054] Preferably, in the sorting process, for any item category in the recall result, the server determines the sorting score of any item category according to the feature data extracted from the behavior log of the user and whether the item category is the item category indicated by the current intention of the user, and determines the arrangement order between each item category in the recall result according to the sorting score. The above feature data can be user features and commodity features extracted from the search log of the user. In the actual scenario, the calculation method of the above sorting score is that if a certain item category has a high correlation with the above feature data, a higher sorting score is set for it; if a certain item category belongs to the item category indicated by the current intention of the user, a corresponding higher sorting score is set for it according to the weight of the category indicated by the current intention. It can be understood that in the above sorting method, the recall results of the same category are arranged in adjacent positions, and the recall results of different categories are arranged in the whole according to the sorting score, and the arrangement positions are not mixed.

[0055] In the same item category in the recall result, the server calculates the sorting score of each recall result belonging to the item category by using a pre-trained sorting model based on machine learning, and determines the arrangement order of each recall result in the item category according to the sorting score. Specifically, the search text and the description information (including the item name, the brand, etc.) of each recall result can be input into the sorting model to calculate the sorting score of each recall result, and the sorting score is positively correlated with the correlation between the recall result and the search text. Finally, a preset number of recall results arranged in the front are selected as the preselected results. In the actual scenario, the above sorting can be divided into two stages of rough sorting and fine sorting, and the execution principles of the two stages are similar, but different sorting logic and sorting models can be used. The rough sorting stage is used to filter out a large number of weakly related recall results at a faster operation speed, and the fine sorting stage is used to accurately locate the final preselected results. It can be understood that if the recall result contains the recall result based on the expansion word, these recall results have a lower weight in the source dimension when determining the sorting score. In practice, the above sorting process can be realized by a pre-set sorting service. It can be seen that through the inter-category sorting based on the current intention of the user and the feature data and the intra-category sorting based on the search text, the sorting accuracy can be maximized to ensure that the items strongly related to the user's demand can be displayed in front of the user.

[0056] Finally, the above pre-selected results can be re-ranked, which is used to reflect more important business rules. In practice, the re-ranking service can be implemented by pre-setting the re-ranking rules. Specifically, the server adjusts the arrangement order of each pre-selected result by using the pre-set re-ranking rules to obtain the final search result. The server can also combine the re-ranking rules with the user features or commodity features calculated based on the user behavior log to obtain the search result. Before obtaining the search result, the server can perform the deduplication processing. Finally, the server returns the search result to the user.

[0057] In particular, the above association word service, correction service, synonym service, word segmentation service, intent recognition service, recall service, ranking service and re-ranking service all belong to microservices. Each microservice runs in an independent process, and the microservices communicate with each other through a lightweight communication mechanism and are independently deployed to the production environment. Referring to Figure 2 , the microservices have certain calling relationships, and the execution process of the microservices is consistent with the search process of the embodiments of the present application.

[0058] In the technical scheme of the embodiments of the present application, the intelligent search engine is developed around the enterprise procurement scene. Natural language processing, machine learning, deep learning, big data analysis and other technologies are used to create a special word segmentation tool and a word segmentation library, establish a search word correction mechanism, optimize the association word rules, embed the user intent prediction algorithm, establish the intelligent ranking rules, mine the search data value, build a sustainable optimization search engine technology and operation system, improve the search experience of the procurement scene, and help each enterprise to deploy and adapt the professional search engine of the procurement scene and then integrate the engine into the general engine of the enterprise.

[0059] It should be noted that, for the foregoing method embodiments, in order to facilitate description, they are described as a series of action combinations, but those skilled in the art should know that the present application is not limited to the order of the actions described, and some steps can actually be performed in other orders or at the same time. In addition, those skilled in the art should know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the implementation of the present application.

[0060] In order to better implement the above scheme of the embodiments of the present application, the related device for implementing the above scheme is also provided.

[0061] Please refer to Figure 3 The vertical search device for the procurement scene 300 provided by the embodiments of the present application can include a first intent recognition unit 301, a second intent recognition unit 302 and a search unit 303.

[0062] The first intention recognition unit 301 can be configured to: in response to a user input search text, acquire at least one keyword from the search text, and determine a user intention corresponding to the keyword in a preset dictionary as a first initial intention of the user; the dictionary contains a mapping relationship between a plurality of words and a plurality of user intentions; the second intention recognition unit 302 can be configured to: input historical purchase data of the user into a preset purchase periodicity rule model and a purchase timeliness rule model, and determine a user intention indicated by a periodicity rule and a timeliness rule matched by the historical purchase data as a second initial intention of the user; the purchase periodicity rule model contains at least one periodicity rule, and the purchase timeliness rule model contains at least one timeliness rule; the search unit 303 can be configured to: determine a current intention of the user according to the first initial intention and the second initial intention, query by using the current intention and the search text, sort at least one recall result obtained by querying, and return a search result corresponding to the search text to the user.

[0063] In the embodiment of the application, the periodicity rule and the timeliness rule include a discriminant condition part and a user intention part; the second intention recognition unit 302 can be further configured to: input historical purchase data of the user in a first historical time period into the purchase periodicity rule model, and determine a user intention indicated by a periodicity rule with a discriminant condition part consistent with the historical purchase data as a second initial intention; input historical purchase data of the user in a second historical time period into the purchase timeliness rule model, and determine a user intention indicated by a timeliness rule with a discriminant condition part consistent with the historical purchase data as a second initial intention; the start time of the first historical time period is earlier than the start time of the second historical time period.

[0064] In a specific application, the device 300 can further include a third intention recognition unit configured to: input a segmented search text into a pre-trained intention classification model based on machine learning to obtain a third initial intention of the user; and the search unit 303 can be further configured to: determine a current intention of the user in combination with the first initial intention, the second initial intention and the third initial intention.

[0065] In actual application, the search unit 303 can be further configured to: query in a preset database by using a segmented result of the search text, adjust a query score of a query result according to an item category indicated by the current intention, and determine a plurality of query results with the largest query score as recall results of the database; input a word vector feature of the segmented result of the search text into a pre-trained vector recall model based on machine learning, adjust a query score of an output result according to an item category indicated by the current intention, and determine a plurality of output results with the largest query score as recall results of the vector recall model.

[0066] As a preferred solution, the searching unit 303 can be further configured to: for any item category in the recall result, determine a ranking score of the any item category according to the feature data extracted from the behavior log of the user and whether the item category is the item category indicated by the current intention, determine the arrangement order between the item categories in the recall result according to the ranking scores; and within the same item category in the recall result, calculate the ranking scores of the recall results belonging to the item category by using a pre-trained machine learning-based ranking model, and determine the arrangement order of the recall results within the item category according to the ranking scores.

[0067] Preferably, the searching unit 303 can be further configured to: select a preset number of the recall results arranged in the front as preselected results; and adjust the arrangement order of the preselected results by using a preset reordering rule to obtain the search results.

[0068] In addition, in the embodiment of the present application, the device 300 can further comprise a preprocessing unit configured to: during the process that the user inputs the search text, determine an association word corresponding to the input text from a preset association word library, and output the association word to the user; after the user initiates the search, determine whether the search text is an error text; when the search text is an error text, correct the search text; determine at least one synonym corresponding to the search text in a preset synonym library as an expansion word of the search text, and perform intention recognition and query of the recall result according to the expansion word.

[0069] According to the technical solution of the embodiment of the present application, the intention recognition service is pre-deployed to accurately determine the real intention of the user from multiple aspects, thereby embodying the search rule of the procurement scenario to improve the search efficiency and quality. Specifically, in the first aspect, after the user inputs the search text, the key word is determined from the search text, and the user intention corresponding to the key word in the preset dictionary is determined as the first initial intention of the user. In the second aspect, the historical procurement data of the user is input into a preset procurement periodicity rule model and a procurement timeliness rule model, the procurement periodicity rule model reflects the periodicity procurement rule embodied by the user in a long time, and the procurement timeliness rule model reflects the short-term timeliness procurement rule of the user, and the two models are combined to determine the second initial intention of the user. In the third aspect, the search text is input into a pre-trained intention classification model after being segmented, thereby obtaining the third initial intention of the user. Finally, the current intention (i.e., the real intention) of the user is determined by combining the first initial intention, the second initial intention and the third initial intention, and thereafter the current intention of the user is used to guide the subsequent recall and ranking process, thereby realizing a high-availability vertical search engine suitable for the procurement scenario.

[0070] Figure 4is a schematic diagram of an architecture of a vertical search system for a procurement scenario in an embodiment of the present application, as shown Figure 4 The vertical search system for the procurement scenario in the embodiment of the present application includes an application layer, a service layer, a data layer, a log collection layer, a data calculation layer, and an algorithm model layer.

[0071] The service layer deploys a plurality of microservices for supporting the application layer, and the above microservices include associated word services, error correction services, synonym services, word segmentation services, intent recognition services, recall services, sorting services, and re-sorting services. These microservices can realize their respective functions and can be specifically selected as needed. The service layer needs the support of the data layer, and the data layer can include a full-text search database (such as Elastic Search ES) for executing data queries and a cache database (such as Redis) for storing the calculation results based on user behavior logs. The recall service of the service layer needs to be queried in the full-text search database to obtain the recall results, and the intent recognition service, the sorting service, and the re-sorting service need to use the calculation results stored in the cache database.

[0072] The user behavior logs can include search logs and order logs, and the calculation results based on the behavior logs can include user historical procurement data calculated based on the order logs and feature data (such as user features and item features) calculated based on the search logs. From the search execution results, the search logs can be divided into successful logs in which the user generates a click or purchase behavior and failed logs in which the user does not generate a click behavior, and the two kinds of logs can be used to guide subsequent product line construction and related operation strategies. In actual applications, the above intent recognition service needs to use the user historical procurement data, and the above sorting service and re-sorting service need to use the above feature data.

[0073] When using the above various microservices, models of the algorithm model layer can need to be called. These models are generally pre-trained models based on machine learning, such as an intent classification model supporting the intent recognition service, a vector recall model supporting the recall service, and a sorting model supporting the sorting service.

[0074] The above log collection layer can collect user behavior logs through application-layer event tracking. Specifically, the log collection layer may include: a relational database (e.g., MySQL) for storing a portion of search logs and order logs, and a file system (e.g., HDFS) for storing another portion of search logs. The former stores log data in database tables, while the latter stores log data in files. The data computation layer performs calculations based on user behavior logs and stores the results in storage units within the data layer. The data computation layer may include: a stream processing engine (e.g., Flink) for calculating search logs and a batch processing engine (e.g., Spark) for calculating order logs; the stream processing engine stores the calculated feature data in a cache database, and the batch processing engine stores the calculated historical purchase data in a cache database. Figure 4 As can be seen, the batch calculation engine can directly obtain order logs from the relational database and then perform calculations. In addition, the batch calculation engine can also pre-calculate the second preliminary intent based on the user's historical purchase data offline and store it in the cache database, which is then called by the intent recognition service.

[0075] Preferably, the vertical search system for procurement scenarios in this embodiment of the invention may further include: a first conversion module (e.g., Flume), a second conversion module (e.g., Beats), and a message queue module (e.g., Kafka). The first conversion module converts database table-style search logs stored in a relational database into streaming data and sends this streaming data to the stream computing engine via the message queue module. The second conversion module converts file-format search logs stored in a file system into streaming data and sends this streaming data to the stream computing engine via the message queue module. The message queue module also has the function of managing context information (e.g., historical execution records) and traffic shaping.

[0076] As can be seen, through Figure 4 The system architecture shown in this embodiment of the invention enables the vertical search system for procurement scenarios to achieve data querying, log collection and processing through the interaction and collaboration of each layer, thus providing stable support for the vertical search engine for procurement scenarios.

[0077] The following explains the working principles of the various microservices mentioned above.

[0078] The keyword suggestion service is used to: determine the suggested words corresponding to the entered text from a pre-set keyword library during the user's search text input process, and output the suggested words to the user. When building the keyword library, it can be based on the user's search logs, or it can be built by combining analysis of item description information, such as extracting entity words and modifiers from item description information and combining them to generate new suggested words.

[0079] The error correction service is configured to determine whether the search text is an error text after the user initiates the search, and correct the search text when the search text is an error text.

[0080] The synonym service is configured to determine at least one synonym corresponding to the search text in a preset synonym library as an extended word of the search text, and perform intent recognition and query of the recall result according to the extended word.

[0081] The intent recognition service is configured to obtain at least one keyword from the search text input by the user, determine a user intent corresponding to the keyword in a preset dictionary as a first initial intent of the user, wherein the dictionary contains a mapping relationship between a plurality of words and a plurality of user intents, input historical procurement data of the user in a data layer into a preset procurement periodicity rule model and a procurement timeliness rule model, and determine a user intent indicated by a periodicity rule and a timeliness rule matched by the historical procurement data as a second initial intent of the user, wherein the procurement periodicity rule model contains at least one periodicity rule, and the procurement timeliness rule model contains at least one timeliness rule, and determine a current intent of the user according to the first initial intent and the second initial intent.

[0082] The recall service is configured to query in the data layer according to the current intent and the search text, the sorting service is configured to sort at least one recall result to obtain a search result corresponding to the search text, and the application layer is configured to return the search result to the user.

[0083] In the embodiment of the application, the periodicity rule and the timeliness rule include a judgment condition part and a user intent part, and the intent recognition service is further configured to input historical procurement data of the user in a first historical time period into the procurement periodicity rule model, determine a user intent indicated by a periodicity rule whose judgment condition part is consistent with the historical procurement data as the second initial intent, input historical procurement data of the user in a second historical time period into the procurement timeliness rule model, and determine a user intent indicated by a timeliness rule whose judgment condition part is consistent with the historical procurement data as the second initial intent, wherein the start time of the first historical time period is earlier than the start time of the second historical time period.

[0084] As a preferred scheme, the intent recognition service is further configured to input the search text after tokenization into a pre-trained intent classification model based on machine learning to obtain a third initial intent of the user, and determine the current intent of the user in combination with the first initial intent, the second initial intent and the third initial intent.

[0085] Preferably, the recall service is further configured to: query the preset database using the segmentation result of the search text, adjust the query scores of the query results according to the item category indicated by the current intent, and determine the query results with the highest query scores as the recall results of the database; input the word vector features of the segmentation result of the search text into a pre-trained vector recall model based on machine learning, adjust the query scores of the output results according to the item category indicated by the current intent, and determine the output results with the highest query scores as the recall results of the vector recall model.

[0086] In a specific application, the ranking service is further configured to: for any item category in the recall results, determine a ranking score of the item category according to the feature data extracted from the user's behavior log and whether the item category is the item category indicated by the current intent, determine the arrangement order between the item categories in the recall results according to the ranking scores, calculate the ranking scores of the recall results belonging to the same item category in the recall results by using a pre-trained ranking model based on machine learning, and determine the arrangement order of the recall results belonging to the same item category according to the ranking scores; and select a preset number of the recall results arranged in the front as preselected results.

[0087] In an actual application, the re-ranking service is configured to: adjust the arrangement order of the preselected results by using a preset re-ranking rule, and obtain the search results.

[0088] In the technical solution of the embodiments of the present application, a specific architecture of a vertical search system suitable for a procurement scenario is provided, including an application layer, a service layer, a data layer, a log collection layer, and a data calculation layer, and the like. These layers realize data query and log collection and processing through interaction and cooperation, and form a strong support for the vertical search engine of the procurement scenario.

[0089] Figure 5 An exemplary system architecture 500 of the vertical search method or the vertical search device of the procurement scenario to which the embodiments of the present application can be applied is shown.

[0090] As shown in Figure 5 The system architecture 500 can include terminal devices 501, 502, and 503, a network 504, and a server 505 (this architecture is only an example, and the components included in the specific architecture can be adjusted according to the specific application). The network 504 is used to provide a communication link medium between the terminal devices 501, 502, 503, and the server 505. The network 504 can include various connection types, such as wired, wireless communication links, or optical fiber cables, etc.

[0091] The user can use the terminal devices 501, 502, 503 to interact with the server 505 through the network 504 to receive or send messages, etc. Various client applications can be installed on the terminal devices 501, 502, 503, such as a vertical search application for a procurement scenario (only as an example).

[0092] The terminal devices 501, 502, 503 can be various electronic devices with display screens and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, desktop computers, etc.

[0093] The server 505 can be a server providing various services, such as a background server supporting a vertical search application for a procurement scenario operated by the user using the terminal devices 501, 502, 503 (only as an example). The background server can process a received search request and feed back a processing result (such as a search result--only as an example) to the terminal devices 501, 502, 503.

[0094] It should be noted that the vertical search method for a procurement scenario provided by the embodiments of the present application is generally executed by the server 505, and correspondingly, the vertical search apparatus for a procurement scenario is generally arranged in the server 505.

[0095] It should be understood that the number of terminal devices, networks and servers in Figure 5 is merely illustrative. Any number of terminal devices, networks and servers can be provided according to implementation needs.

[0096] The present application also provides an electronic device. The electronic device of the embodiments of the present application comprises one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the vertical search method for a procurement scenario provided by the present application.

[0097] Reference is made below to Figure 6 which shows a structural schematic diagram of a computer system 600 suitable for implementing the electronic device of the embodiments of the present application. Figure 6 The electronic device shown is merely an example and should not bring any limitation to the functions and use range of the embodiments of the present application.

[0098] As Figure 6As shown, the computer system 600 includes a central processing unit (CPU) 601 which can perform various appropriate actions and processes in accordance with programs stored in a read only memory (ROM) 602 or loaded into a random access memory (RAM) 603 from a storage section 608. In the RAM 603, various programs and data required for the operation of the computer system 600 are also stored. The CPU 601, the ROM 602, and the RAM 603 are connected to each other through a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0099] Connected to the I / O interface 605 are an input section 606 including a keyboard, a mouse, etc.; an output section 607 including a display device such as a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN card, a modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the I / O interface 605 as necessary. A removable recording medium 611 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc. is attached to the drive 610 as necessary, so that a computer program read therefrom is installed into the storage section 608 as necessary.

[0100] In particular, according to the embodiments disclosed herein, the processes described in the above main step diagrams can be implemented as a computer software program. For example, the embodiments of the present application include a computer program product comprising a computer program carried on a computer readable medium, the computer program containing program code for executing the method shown in the main step diagrams. In the above embodiments, the computer program can be downloaded and installed from a network by the communication section 609, and / or installed from the removable recording medium 611. When the computer program is executed by the central processing unit 601, the above functions defined in the system of the present application are performed.

[0101] It should be noted that the computer-readable medium shown in the present application can be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may, for example, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or component, or any combination of the above. More specific examples of computer-readable storage media can include, but are not limited to, an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device or component. In the present application, the computer-readable signal medium can include a data signal carried in a baseband or as a part of a carrier wave, which carries computer-readable program code. Such a propagated data signal can take various forms, including but not limited to an electromagnetic signal, an optical signal or any suitable combination of the above. The computer-readable signal medium can also be any computer-readable medium other than the computer-readable storage medium, which can send, propagate or transmit a program for use by or in conjunction with an instruction execution system, device or component. The program code contained on the computer-readable medium can be transmitted by any suitable medium, including but not limited to wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0102] The flowcharts and block diagrams in the drawings illustrate the possible implementation architectures, functions and operations of the systems, methods and computer program products according to various embodiments of the present application. In this regard, each block in the flowcharts or block diagrams can represent a module, a program segment or a part of code containing one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur in different order than that shown in the drawings. For example, two blocks that are shown in succession can actually be executed substantially in parallel, and sometimes in reverse order, depending on the involved functions. It should also be noted that each block in the block diagrams or flowcharts, and the combination of blocks in the block diagrams or flowcharts, can be implemented by a dedicated hardware-based system that performs the specified functions or operations, or can be implemented by a combination of dedicated hardware and computer instructions.

[0103] The units described in the embodiments of the present application can be implemented in the form of software, or can be implemented in the form of hardware. The described units can also be arranged in a processor, for example, a processor can be described as including a first intent recognition unit, a second intent recognition unit, and a search unit. Among them, the names of these units do not constitute a limitation on the units themselves in some cases, for example, the first intent recognition unit can also be described as "a unit that provides the search unit with a first preliminary intent".

[0104] As another aspect, the present application also provides a computer readable medium, which can be included in the device described in the above embodiments, or can exist independently without being assembled into the device. The computer readable medium carries one or more programs, and when the one or more programs are executed by the device, the steps performed by the device include: in response to a user input search text, obtaining at least one keyword from the search text, determining a user intent corresponding to the keyword in a preset dictionary as a first preliminary intent of the user; wherein the dictionary contains mapping relationships between multiple words and multiple user intents; inputting historical purchase data of the user into a preset purchase periodicity rule model and a purchase timeliness rule model, and determining a user intent indicated by a periodicity rule and a timeliness rule matched by the historical purchase data as a second preliminary intent of the user; wherein the purchase periodicity rule model contains at least one periodicity rule, and the purchase timeliness rule model contains at least one timeliness rule; determining a current intent of the user according to the first preliminary intent and the second preliminary intent; querying using the current intent and the search text, and sorting at least one recall result obtained by querying to obtain a search result corresponding to the search text, and returning the search result to the user.

[0105] In the technical scheme of the embodiment of the present application, the pre-deployed intention recognition service accurately judges the real intention of the user from multiple aspects, thereby embodying the search rule of the procurement scene to improve the search efficiency and quality. Specifically, in the first aspect, after the user inputs the search text, the key word is determined from the search text, and the user intention corresponding to the key word in the preset dictionary is determined as the first initial selected intention of the user. In the second aspect, the historical procurement data of the user is input into the preset procurement periodicity rule model and the procurement timeliness rule model, the procurement periodicity rule model reflects the periodicity procurement rule embodied by the user in a long time, and the procurement timeliness rule model reflects the short-term timeliness procurement rule of the user, and the two models are combined to determine the second initial selected intention of the user. In the third aspect, the search text is input into the pre-trained intention classification model after being segmented, thereby obtaining the third initial selected intention of the user. Finally, the current intention (i.e., the real intention) of the user is determined by combining the first initial selected intention, the second initial selected intention and the third initial selected intention, and thereafter the current intention of the user is used to guide the subsequent recall and sorting process, thereby realizing a high-availability vertical search engine suitable for the procurement scene. In addition, the present application also provides a specific architecture of the vertical search system suitable for the procurement scene, including an application layer, a service layer, a data layer, a log collection layer and a data calculation layer and the like, and these layers realize data query and log collection and processing through interaction and cooperation, thereby providing strong support for the above vertical search engine.

[0106] The above detailed description does not constitute a limitation on the protection scope of the present application. Those skilled in the art should understand that various modifications, combinations, sub-combinations and substitutions can occur depending on design requirements and other factors. Any modification, equivalent replacement and improvement within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A vertical search method for a procurement scenario, characterized in that, include: In response to user input of search text, at least one keyword is obtained from the search text, and the user intent corresponding to the keyword in a preset dictionary is determined as the user's first preliminary intent; wherein, the dictionary contains mapping relationships between multiple words and multiple user intents; The user's historical purchasing data is input into a preset purchasing periodicity rule model and a purchasing timeliness rule model. The user intent indicated by the periodicity rule and timeliness rule matched with the historical purchasing data is determined as the user's second preliminary intent. The purchasing periodicity rule model reflects the user's periodic purchasing patterns over a longer period and contains at least one periodic rule. The purchasing timeliness rule model reflects the user's short-term timeliness purchasing patterns and contains at least one timeliness rule. The longer period is a time period exceeding a set first threshold, and the short-term period is a time period not exceeding a set second threshold, wherein the first threshold is greater than the second threshold. The search text is segmented and then input into a pre-trained machine learning-based intent classification model to obtain the user's third preliminary intent. The user's current intent is determined based on the first initial intent, the second initial intent, and the third initial intent; a query is performed using the current intent and the search text, and at least one recall result is sorted to obtain the search result corresponding to the search text and return it to the user; The step of querying using the current intent and the search text includes: The search results are queried using the word segmentation results of the search text in the preset database. The query scores of the query results are adjusted according to the item category indicated by the current intent. The multiple query results with the highest query scores are determined as the recall results of the database. The word vector features of the word segmentation results of the search text are input into a pre-trained, machine learning-based vector recall model. The query score of the output results is adjusted according to the item category indicated by the current intent. The multiple output results with the highest query scores are determined as the recall results of the vector recall model. Using the current intent and the search text, a query is performed in a preset specific database to obtain the recall results of the specific database; The recall results are combined with those from the database, the vector recall model, and the specific database.

2. The method according to claim 1, characterized in that, The periodic rules and the timeliness rules include a discrimination condition part and a user intent part; and, the step of inputting the user's historical procurement data into a preset procurement periodic rule model and procurement timeliness rule model, and determining the user intent indicated by the periodic rules and timeliness rules matched by the historical procurement data as the user's second preliminary intent, includes: The user's historical procurement data for the first historical period is input into the procurement periodic rule model, and the user intent indicated by the periodic rule that matches the historical procurement data in the discrimination condition part is determined as the second preliminary selection intent; The user's historical procurement data in the second historical time period is input into the procurement timeliness rule model, and the user intent indicated by the timeliness rule that matches the historical procurement data in the judgment condition part is determined as the second preliminary intent; wherein, the start time of the first historical time period is earlier than the start time of the second historical time period.

3. The method according to claim 1, characterized in that, The sorting of at least one retrieved result includes: For any item category in the recall results, a ranking score is determined based on the feature data extracted from the user's behavior log and whether the item category is the item category indicated by the current intent. The order of the item categories in the recall results is determined according to the ranking score. Within the same item category in the recall results, a ranking score for each recall result belonging to that item category is calculated using a pre-trained, machine learning-based ranking model, and the order of each recall result within that item category is determined according to the ranking score.

4. The method according to claim 3, characterized in that, Obtaining the search results corresponding to the search text includes: Select the first preset number of the recall results as the pre-selected results; The search results are obtained by adjusting the order of the pre-selected results using preset reordering rules.

5. The method according to any one of claims 1-4, characterized in that, Obtaining at least one keyword from the search text includes: During the process of the user inputting the search text, the associated words corresponding to the input text are determined from the preset associated word library, and the associated words are output to the user; After the user initiates a search, determine whether the search text is incorrect; if the search text is incorrect, correct the search text. At least one synonym of the search text in the preset thesaurus is identified as an extended word of the search text, and intent recognition and recall result query are performed based on the extended word.

6. A vertical search device for a procurement scenario, characterized in that, include: The first intent recognition unit is configured to: in response to a user inputting search text, obtain at least one keyword from the search text, and determine the user intent corresponding to the keyword in a preset dictionary as the user's first preliminary intent; wherein the dictionary contains mapping relationships between multiple words and multiple user intents; The second intent recognition unit is configured to: input the user's historical procurement data into a preset procurement periodicity rule model and a procurement timeliness rule model; and determine the user intent indicated by the periodicity rule and timeliness rule matched with the historical procurement data as the user's second preliminary intent; wherein, the procurement periodicity rule model reflects the user's periodic procurement pattern over a longer period and contains at least one periodic rule, and the procurement timeliness rule model reflects the user's short-term timeliness procurement pattern and contains at least one timeliness rule; the longer period is a time period whose length exceeds a set first threshold, and the short-term period is a time period whose length does not exceed a set second threshold, wherein the first threshold is greater than the second threshold; The third intent recognition unit is used to: segment the search text into words and input it into a pre-trained machine learning-based intent classification model to obtain the user's third preliminary intent; The search unit is configured to: determine the user's current intent based on a first preliminary intent, a second preliminary intent, and a third preliminary intent; perform a query using the current intent and the search text, sort at least one of the retrieved results, and return the search results corresponding to the search text to the user. The search unit is further configured to: perform a query in a preset database using the word segmentation results of the search text; adjust the query score of the query results according to the item category indicated by the current intent; and determine the multiple query results with the highest query scores as the recall results of the database; input the word vector features of the word segmentation results of the search text into a pre-trained, machine learning-based vector recall model; adjust the query score of the output results according to the item category indicated by the current intent; and determine the multiple output results with the highest query scores as the recall results of the vector recall model; perform a query in a preset specific database using the current intent and the search text to obtain the recall results of the specific database; and combine the recall results of the database, the recall results of the vector recall model, and the recall results of the specific database.

7. A vertical search system for procurement scenarios, characterized in that, include: Application layer, service layer, data layer, log collection layer, and data computation layer; among them, The service layer is deployed to support multiple microservices of the application layer, including an intent recognition service, a recall service, and a ranking service. The log collection layer collects user behavior logs through the application layer; the data calculation layer performs calculations based on the behavior logs and stores the calculation results in the data layer; wherein, the calculation results include the user's historical purchasing data; The intent recognition service obtains at least one keyword from the search text input by the user, and determines the user intent corresponding to the keyword in the preset dictionary as the user's first preliminary intent; wherein, the dictionary contains mapping relationships between multiple words and multiple user intents; The intent recognition service inputs the user's historical procurement data from the data layer into a preset procurement periodicity rule model and a procurement timeliness rule model. The user intent indicated by the periodicity rules and timeliness rules matched with the historical procurement data is determined as the user's second preliminary intent. The procurement periodicity rule model reflects the user's periodic procurement patterns over a longer period and contains at least one periodic rule. The procurement timeliness rule model reflects the user's short-term timeliness procurement patterns and contains at least one timeliness rule. The longer period is a time period exceeding a set first threshold, and the short-term period is a time period not exceeding a set second threshold, wherein the first threshold is greater than the second threshold. The intent recognition service segments the search text into words and inputs it into a pre-trained, machine learning-based intent classification model to obtain the user's third preliminary intent. The intent recognition service determines the user's current intent based on a first preliminary intent, a second preliminary intent, and a third preliminary intent; the recall service uses the current intent and the search text to perform a query in the data layer; the sorting service sorts at least one of the retrieved recall results to obtain the search results corresponding to the search text; and the application layer returns the search results to the user. The recall service is further configured to: perform a query in a preset database using the word segmentation results of the search text; adjust the query score of the query results according to the item category indicated by the current intent; and determine the multiple query results with the highest query scores as the recall results of the database; input the word vector features of the word segmentation results of the search text into a pre-trained, machine learning-based vector recall model; adjust the query score of the output results according to the item category indicated by the current intent; and determine the multiple output results with the highest query scores as the recall results of the vector recall model; perform a query in a preset specific database using the current intent and the search text to obtain the recall results of the specific database; and combine the recall results of the database, the recall results of the vector recall model, and the recall results of the specific database.

8. The system according to claim 7, characterized in that, The behavior log includes: search logs and order logs, and the calculation result further includes: feature data calculated from the search logs; and, The data layer includes: a full-text search database for executing the query and a cache database for storing the calculation results; The log collection layer includes: a relational database for storing a portion of the search logs and the order logs, and a file system for storing another portion of the search logs; The data computation layer includes: a stream computing engine for performing calculations on the search logs and a batch computing engine for performing calculations on the order logs; the stream computing engine stores the calculated feature data in the cache database, and the batch computing engine stores the calculated historical purchase data in the cache database.

9. The vertical search system according to claim 8, characterized in that, The system further includes: a first conversion module, a second conversion module, and a message queue module; wherein, The first conversion module converts the search logs stored in the relational database into streaming data, and sends the streaming data to the streaming computing engine through the message queue module; The second conversion module converts the search logs stored in the file system into streaming data and sends the streaming data to the streaming computing engine through the message queue module.

10. The vertical search system according to claim 9, characterized in that, The vertical search system further includes: an algorithm model layer containing an intent classification model that supports the intent recognition service, a vector recall model that supports the recall service, and a ranking model that supports the ranking service; The intent classification model, the vector recall model, and the ranking model are all pre-trained machine learning models.

11. The vertical search system according to claim 10, characterized in that, The microservices further include: word association service, error correction service, synonym service, word segmentation service, and reordering service; The full-text search database includes Elasticsearch (ES), the caching database includes Redis, the relational database includes MySQL, the file system includes HDFS, the stream processing engine includes Flink, the batch processing engine includes Spark, the first transformation module includes Flume, the second transformation module includes Beats, and the message queue module includes Kafka.

12. An electronic device, characterized in that, include: One or more processors; Storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-5.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-5.

Images