Search result generating method and device

Abstract

The present application provides a search result generation method and apparatus, wherein the method includes: obtaining search information input by a target user; screening a collection of materials related to the search information from a preset full collection of materials as a candidate collection of materials; matching the search information with historically relevant information, and selecting the user corresponding to the matched historically relevant information as a candidate contact; determining a collection of associated materials corresponding to the search information based on the search information and a pre-established knowledge graph and a standard question-themed corpus; and determining search results corresponding to the search information based on the candidate collection of materials, candidate contacts, and associated collection of materials. The present application only requires inputting search information once to return search results determined based on the candidate collection of materials, candidate contacts, and associated collection of materials, resulting in more comprehensive search results that better meet user needs.

Description

Technical Field

[0001] The present application relates to the field of computer applications, and in particular to a method and device for generating search results. Background Art

[0002] Financial markets and other fields often have specialized systems, such as information management systems and transaction management systems. When users use unfamiliar scenarios and functions within specialized systems, they often encounter questions like, "How do I use a certain function?", "I'm having a problem in the system. How do I solve it?", or "I'm having a problem in the system. Who should I contact?" In these cases, users can search within the specialized systems to find solutions.

[0003] Currently, dedicated systems implement search and user help functions independently. For search, the search bar generally only supports full-text or fuzzy search. Search results are primarily calculated using a fixed matching algorithm based on the page title and HTML content, and only content containing the searched keywords is returned. For user help, dedicated systems typically display help documents in a flat format, requiring users to find a title matching their needs in the directory or search within the help document library based on their needs.

[0004] However, whether it is the search function or the user help function, for some users who are unable to accurately describe their help needs, such as colloquial expressions or typos, empty results will be returned in the search bar, resulting in users being unable to successfully find solution information; in addition, dedicated systems usually do not support the provision of candidate contacts and do not support the provision of related information.

[0005] In summary, there is an urgent need for a dedicated system that can provide an integrated search feature so that search results containing solution materials, related materials, and candidate contacts can be returned in one search. Summary of the Invention

[0006] In view of this, the present application provides a search result generation method and apparatus for returning search results containing solution materials, related materials, and candidate contacts in a single search. The technical solution is as follows:

[0007] A search result generating method, comprising:

[0008] Obtain the search information entered by the target user;

[0009] Filtering data sets related to the search information from the preset full data set as candidate data sets;

[0010] Match the search information with historical related information, and select the user corresponding to the matched historical related information as a candidate contact, wherein the historical related information includes historically published relevant materials and / or historical answer users, and the historical answer users are users who have historically answered related questions;

[0011] Determine the associated data set corresponding to the search information based on the search information and the pre-established knowledge graph and standard question topic corpus. The knowledge graph is generated based on system log data and existing solution data, and the standard question topic corpus is generated based on existing solution data.

[0012] Determine search results corresponding to the search information based on the candidate data set, candidate contacts, and associated data set.

[0013] Optionally, data sets related to the search information are screened from the preset full data set as candidate data sets, including:

[0014] Determine whether the search information is of the intended type;

[0015] If so, a solution set related to the search information is screened from a preset stock solution corpus, and a page set related to the search information is screened from a preset full page set. The screened solution set and page set are used as a candidate data set;

[0016] If not, a solution set related to the search information is screened from the existing solution corpus, and the screened solution set is used as the candidate data set.

[0017] Optionally, matching the search information with historical related information and selecting users corresponding to the matched historical related information as candidate contacts includes:

[0018] Matching the search information with historically published related information through a data matching model, and selecting users corresponding to the matched related information as candidate contacts, wherein the data matching model is trained using the first training search information as training data and using users corresponding to the training related information that are labeled to match the first training search information as sample labels;

[0019] And / or, the search information is matched with the historical answer users through a user matching model, and the matched historical answer users are used as candidate contacts, wherein the user matching model is trained using the second training search information as training data and the training answer users matched with the labeled second training search information as sample labels.

[0020] Optionally, based on the search information and a pre-established knowledge graph and standard topic corpus, a set of related materials corresponding to the search information is determined, including:

[0021] If the search information is valid input, the search information is matched with the standard question subject corpus to obtain the standard question corresponding to the search information, and the standard question corresponding to the search information is matched with the knowledge graph to obtain the associated data set corresponding to the search information;

[0022] If the search information is an invalid input, the historical operation sequence related to the search information is matched with the standard question topic corpus to obtain the standard question corresponding to the historical operation sequence, and the standard question corresponding to the historical operation sequence is matched with the knowledge graph, and the matched related data set is used as the related data set corresponding to the search information, where the historical operation sequence is determined by the system log data.

[0023] Optionally, determining search results corresponding to the search information based on the candidate data set, candidate contacts, and associated data set includes:

[0024] Match the candidate materials in the candidate material set with the search information based on semantic similarity, and obtain several candidate materials with high semantic similarity as the target material set;

[0025] The target data set, candidate contacts and associated data set are used as search results corresponding to the search information.

[0026] Optionally, semantic similarity matching is performed on candidate materials in the candidate material set and the search information to obtain several candidate materials with high semantic similarity, including:

[0027] The deep learning pre-trained language model is used to match the candidate materials in the candidate material set with the search information for semantic similarity, and obtain several candidate materials with high semantic similarity. The deep learning pre-trained language model is trained using the training search information and the corresponding training materials as training data, and the semantic similarity between the labeled training search information and the training materials as sample labels.

[0028] Optionally, also include:

[0029] Generate a search return reason corresponding to the search result, where the search return reason is used to explain the reason for returning the search result.

[0030] Optional methods for building a knowledge graph include:

[0031] Obtain system log data and existing solution data. System log data is used to record historical usage of users when solving problems.

[0032] Based on the cleaning of system log data, several historical operation sequences are obtained, and correlation and dependency relationships of the several historical operation sequences are extracted to obtain the extracted correlation and dependency relationships;

[0033] Extract label relationships and author relationships based on existing solution data to obtain extracted label relationships and author relationships;

[0034] Based on the correlation, dependency, label and author relationships, a knowledge graph is established with solutions, answerers and topic labels as the three nodes of the knowledge graph.

[0035] A search result generating device comprises: a search information acquisition module, a candidate data set determination module, a candidate contact determination module, a related data set determination module and a search result determination module;

[0036] Search information acquisition module, used to obtain the search information input by the target user;

[0037] A candidate data set determination module is used to select data sets related to the search information from the preset full data set as candidate data sets;

[0038] A candidate contact determination module is configured to match search information with historically relevant information and select users corresponding to the matched historically relevant information as candidate contacts, wherein the historically relevant information includes historically published relevant materials and / or historically answering users, wherein historically answering users refer to users who have historically answered related questions;

[0039] A related document set determination module is used to determine the related document set corresponding to the search information based on the search information and a pre-established knowledge graph and standard question topic corpus, wherein the knowledge graph is generated based on system log data and existing solution data, and the standard question topic corpus is generated based on existing solution data;

[0040] The search result determination module is used to determine the search results corresponding to the search information based on the candidate data set, candidate contacts and related data set.

[0041] Optionally, the candidate data set determination module includes: a purpose type determination submodule, a first candidate data set determination submodule, and a second candidate data set determination submodule;

[0042] The purpose type judgment submodule is used to judge whether the search information is of the purpose type;

[0043] A first candidate material set determination submodule is configured to, if the purpose type determination submodule determines that the search information is of the purpose type, filter a solution set related to the search information from a preset stock solution corpus, and filter a page set related to the search information from a preset full page set, and use the filtered solution set and page set as a candidate material set;

[0044] The first candidate material set determination submodule is used to screen solution sets related to the search information from the existing solution corpus if the purpose type determination submodule determines that the search information is not of the purpose type, and the screened solution sets are used as candidate material sets.

[0045] Through the above technical solution, it can be seen that the search result generation method provided by this application first obtains the search information input by the target user, and then filters the data set related to the search information from the preset full data set as the candidate data set, and then matches the search information with the historical related information, and uses the user corresponding to the matched historical related information as the candidate contact, and then determines the associated data set corresponding to the search information based on the search information and the pre-established knowledge graph and standard question subject corpus, and finally determines the search result corresponding to the search information based on the candidate data set, candidate contacts and associated data set. It can be seen that this application only needs to input the search information once to return the search results determined based on the candidate data set, candidate contacts and associated data set, and the search results are more comprehensive and more in line with user needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without any creative work.

[0047] Figure 1 A flowchart of a method for generating search results according to an embodiment of the present application;

[0048] Figure 2 A schematic diagram of the structure of a search result generating device provided in an embodiment of the present application;

[0049] Figure 3 This is a hardware structure block diagram of the search result generation device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0051] In view of the problems existing in the prior art, the inventors of this case conducted in-depth research and finally proposed a search result generation method. The search result generation method provided by this application will be introduced in detail through the following embodiments.

[0052] See also Figure 1 , which shows a flow chart of a method for generating search results provided in an embodiment of the present application. The method for generating search results may include:

[0053] Step S101: Acquire search information input by a target user.

[0054] Here, the search information is the information entered by the target user when seeking solutions to the difficulties (usage requirements) encountered when currently using the dedicated system, for example, "how to use a certain function", "how to solve a certain problem encountered in the system", "how to use user basic information", "how to generate a comprehensive transaction query report" and other search information.

[0055] In this step, when the target user encounters the above-mentioned usage requirements in the dedicated system, he or she may enter search information in the search bar of the dedicated system.

[0056] Step S102: Filter a data set related to the search information from the preset full data set as a candidate data set.

[0057] In this step, a corresponding full data set is preset for each dedicated system, such as a full page set, a set of existing solution data contained in the existing solution corpus, etc. When a target user enters search information, this step allows the user to filter data sets related to the search information from the preset full data set, so that the target user can use the filtered data sets to meet their current needs.

[0058] It should be noted that the stock solution data contained in the stock solution corpus refers to data related to the current solutions of the dedicated system. Here, the solutions are used to solve the user's usage needs.

[0059] Optionally, the existing solution data includes solution identification (such as solution ID), solution tag (used to mark the standard question scenario corresponding to the solution), solution author, solution scenario description, solution content, solution title, question-asking user and other data.

[0060] It should be noted that if the search information input by the target user is valid, the candidate material set screened out in this step includes several candidate materials. If the search information input by the target user is invalid, the candidate material set screened out in this step is empty.

[0061] Step S103: Match the search information with the historical related information, and select the user corresponding to the matched historical related information as a candidate contact.

[0062] It is understandable that there may be situations where the target user does not understand the candidate data set or the candidate data set is insufficient to meet the target user's usage needs. In order to further help the target user, the search information can also be matched with historical related information to obtain candidate contacts that can help the target user meet their usage needs.

[0063] In an optional embodiment, the historical related information provided in this step includes one or more of the following information: historically released related materials and historical answer users. Here, historical answer users refer to users who have historically answered related questions, and the related questions refer to questions related to the search information.

[0064] That is, this step can match the search information with the relevant information published historically to determine candidate contacts from the responsible users who have published relevant information historically, and / or match the search information with the historical answer users to determine candidate contacts from the historical answer users.

[0065] Step S104: Determine a set of related documents corresponding to the search information based on the search information and a pre-established knowledge graph and standard question topic corpus.

[0066] Specifically, this step can generate a knowledge graph based on system log data and existing solution data. This knowledge graph associates various pieces of information within the system log data and existing solution data. This step can also generate a standard question topic corpus based on the existing solution data. Optionally, a standard question topic corpus can be established based on structured data such as [solution identifier, solution title, solution content, solution tags]. Here, the solution title in the standard question topic corpus is referred to as a standard question.

[0067] Optionally, the process of generating a knowledge graph based on system log data and existing solution data in this step may include the following S1 to S4:

[0068] S1. Obtain system log data and existing solution data, wherein the system log data is used to record the historical usage of users when solving problems.

[0069] S2. Based on the system log data cleaning, several historical operation sequences are obtained, and correlation and dependency relationships are extracted from the several historical operation sequences to obtain extracted correlation and dependency relationships.

[0070] In this step, the system log data is in the form of [User][Operation Op][Time of Execution T]. In this embodiment, the system log data can be used to clean out several historical operation sequences in the form of [Operation A, Operation B, Operation C]. The historical operation sequence is a set of standard solution steps generated based on the system log data.

[0071] In this step, the relationship between solutions can be extracted through the operation sequence to obtain related (RELATED) and dependent (RELY) relationships.

[0072] S3. Extract label relationships and author relationships based on the existing solution data to obtain extracted label relationships and author relationships.

[0073] For example, optionally, the solution ID and solution label information in the existing solution data can be used to extract the label (LABEL) relationship, and the solution ID and solution author information in the existing solution data can be used to extract the author (AUTHOR) relationship.

[0074] S4. Based on the correlation, dependency, label and author relationships, a knowledge graph is established with solutions, answerers and topic labels as the three nodes of the knowledge graph.

[0075] Specifically, when building a knowledge graph, the solution, the answerer and the topic label are in a label (LABEL) relationship, the solutions are in a related (RELATED) relationship or a dependent (RELY) relationship, and the solution and the answerer are in an author (AUTHOR) relationship.

[0076] After establishing the knowledge graph and standard question topic corpus, the corresponding related data set can be determined based on the search information and the pre-established knowledge graph and standard question topic corpus. Here, the related data set serves as "potentially helpful information" to assist the target user in solving their current needs.

[0077] That is, this step can mine the implicit input information of the target user based on the knowledge graph, and then determine the associated solutions based on the implicit input information, so that the final search results are more comprehensive.

[0078] Step S105: Determine search results corresponding to the search information based on the candidate data set, candidate contacts, and associated data set.

[0079] Specifically, this step may directly use the candidate data set, candidate contact set, and associated data set as the search results corresponding to the search information, or may further process the candidate data set, candidate contact set, and associated data set, and use the processed candidate data set, candidate contact set, and associated data set as the search results corresponding to the search information.

[0080] The search result generation method provided by this application first obtains the search information input by the target user, then filters the data set related to the search information from the preset full data set as the candidate data set, then matches the search information with the historical related information, and uses the user corresponding to the matched historical related information as the candidate contact, and then determines the associated data set corresponding to the search information based on the search information and the pre-established knowledge graph and standard question subject corpus, and finally determines the search result corresponding to the search information based on the candidate data set, candidate contacts and associated data set. It can be seen that this application only needs to input the search information once to return the search results determined based on the candidate data set, candidate contacts and associated data set, and the search results are more comprehensive and more in line with the user's needs.

[0081] Considering that in some scenarios, it is sufficient to simply return the corresponding answer (also known as a solution) for the search information entered by the target user, while in other scenarios, it is necessary to return the corresponding answer and a page containing an operation shortcut control for the search information entered by the target user, so that the target user can quickly fill in the relevant information on the result page and quickly perform the operation. For example, when the target user needs to perform an approval operation on a dedicated system, they can enter the search information "I want to approve an order with the order number xx,..." in the search bar. In this case, the present embodiment needs to return the corresponding answer and a page containing an operation shortcut control based on the search information, so that the target user can quickly perform the approval process after filling in the relevant information in the operation shortcut control.

[0082] Based on this, the aforementioned step "Step S102, screening a data set related to the search information from the preset full data set as a candidate data set" may optionally include:

[0083] Step S1021: Determine whether the search information is of the target type.

[0084] Here, the purpose type refers to the problem type that can provide solutions and operation shortcut controls (ie, pages).

[0085] Step S1022: If yes, filter the solution set related to the search information from the preset stock solution corpus, and filter the page set related to the search information from the preset full page set. The filtered solution set and page set are used as candidate data sets.

[0086] When the search information is determined to be of the target type in the above step, this step can be used to filter a solution set related to the search information from the existing solution corpus as a candidate solution set, and a page set related to the search information can be filtered from the preset full page set as a candidate page set. The candidate solution set and the candidate page set are the candidate data set in this step.

[0087] Step S1023: If not, then filter the solution sets related to the search information from the existing solution corpus, and use the filtered solution sets as candidate data sets.

[0088] If the search information is not of the target type, it means that only solutions for the search information need to be returned. Based on this, this step only filters the solution set related to the search information from the existing solution corpus.

[0089] This embodiment filters different candidate data sets according to the different types of search information input by the user, so that this embodiment can be used in more scenarios. When the search information is the target type, the page containing the operation shortcut control can be filtered out to facilitate the target user to perform quick operations, thereby improving the target user's operation efficiency.

[0090] In one embodiment of the present application, the process of the aforementioned "step S103, matching the search information with the historical related information, and selecting the user corresponding to the matched historical related information as a candidate contact" is introduced.

[0091] Optionally, the process of “step S103, matching the search information with historical related information, and selecting the user corresponding to the matched historical related information as a candidate contact” may include at least one of the following steps S1031 and S1032:

[0092] Step S1031: Match the search information with the relevant information published in the past through a data matching model, and use the user corresponding to the matched relevant information as a candidate contact.

[0093] If the historical related information includes historically released related information, this step can match the search information with the historically released related information through a pre-trained information matching model, and use the user corresponding to the matched related information as a candidate contact.

[0094] Here, the data matching model is trained using the first training search information as training data and using users corresponding to the training-related data that are marked to match the first training search information as sample labels.

[0095] It is worth noting that this step matches the search information with each relevant document published in the history through the document matching model. Therefore, the "matched relevant documents" in this step refer to the top several relevant documents that are most matched.

[0096] Optionally, the above-mentioned data matching model can be specifically a q2a2u matching model (first training search information query-matching standard question-matching contact), which is obtained by modeling the joint probability space of editing a specific solution for a specific contact based on [user, operation sequence] data.

[0097] In an optional embodiment, before the data matching model matches the search information with the historically published relevant data, the User2Vec model may be first used to obtain the user representation vector (i.e., user embedding, i.e., the user's low-dimensional dense vector) of the target user based on the search information, and the user representation vector of the user corresponding to each historically published relevant data may be obtained based on each historically published relevant data. Then, the step of "matching the search information with the historically published relevant data through the data matching model and taking the users corresponding to the matched relevant data as candidate contacts" may specifically be performing cosine similarity calculation on the user representation vector of the target user and the user representation vector of the user corresponding to each historically published relevant data through the data matching model, and obtaining the top K1 users with the largest similarity calculation results as candidate contacts.

[0098] Here, the training data set of User2Vec is divided into user answer sequence S, that is, given a user User i A time series record of answers or solution edits. The training goal is to maximize the logarithmic maximum likelihood value ζ = ∑log p(Context(item)item) for all user solution sequences. (Context refers to the context item' of a specific solution item in the solution sequence S).

[0099] Step S1032: Match the search information with historical answer users through a user matching model, and use the matched historical answer users as candidate contacts.

[0100] If the historical related information includes historical answer users, this step can match the search information with the historical answer users through a pre-trained user matching model, and use the matched historical answer users as candidate contacts.

[0101] Here, the user matching model is trained using the second training search information as training data and the training answer users matched with the labeled second training search information as sample labels.

[0102] It is worth noting that this step matches the search information with each historical answer user through the user matching model, so the "matched historical answer users" in this step refers to the top several historical answer users that are most matched.

[0103] Optionally, the user matching model may specifically be a q2u matching model (second training search information query-matching contact), which is obtained by modeling a joint probability space of a specific contact's answer to a question based on [user, user corpus] data.

[0104] In an optional embodiment, before the user matching model matches the search information with the historical answer users, the user representation vector of the target user can be first obtained based on the search information through the User2Vec model, and the user representation vector of each historical answer user can be obtained based on the related questions answered by the historical answer users. Then, this step of "matching the search information with the historical answer users through the user matching model, and using the matched historical answer users as candidate contacts" can specifically be performing cosine similarity calculation on the user representation vector of the target user and the user representation vector of each historical answer user through the user matching model, and obtaining the top K2 users with the largest similarity calculation results as candidate contacts.

[0105] In summary, this embodiment performs a joint recall calculation based on the data matching model and / or the user matching model, so that the present application can obtain more matching candidate contacts.

[0106] As described in the background section, "some users are unable to accurately describe their help needs, such as through colloquial expressions or typos, resulting in empty results being returned in the search bar, preventing users from finding solutions." To ensure that users can still obtain relatively good search results even when they are unable to accurately describe their help needs, this embodiment provides two different processes for determining related resource sets in step S104, corresponding to valid and invalid search information. These two processes are described below.

[0107] Based on this, the process of "step S104, determining a set of related documents corresponding to the search information based on the search information and the pre-established knowledge graph and standard question topic corpus" may optionally include the following steps S1041 or S1042:

[0108] Step S1041: If the search information is a valid input, the search information is matched with the standard question subject corpus to obtain the standard question corresponding to the search information, and the standard question corresponding to the search information is matched with the knowledge graph to obtain the associated data set corresponding to the search information.

[0109] Optionally, if the search information is of the purpose type, this step may use the full log information knowledge graph (a part of the knowledge graph established in this embodiment) as the target pool to be matched, and use the user's implicit input information (including system usage logs, etc.) to match the associated page set related to the search information; at the same time, this step may use another part of the knowledge graph as the target pool to be matched, and use the user's implicit input information to match the associated solution set related to the search information.

[0110] Step S1042: If the search information is an invalid input, the historical operation sequence related to the search information is matched with the standard question subject corpus to obtain the standard question corresponding to the historical operation sequence, and the standard question corresponding to the historical operation sequence is matched with the knowledge graph, and the matched related data set is used as the related data set corresponding to the search information, wherein the historical operation sequence is determined by the system log data.

[0111] Optionally, this step can first use the User2Vec model to perform user representation vectorization on the historical operation sequence related to the search information. The processed user implicit operation sequence is then matched with the closest standard question from the standard question topic corpus, and the adjacent content of the standard question corresponding to the historical operation sequence in the knowledge graph is used as the associated data set corresponding to the search information.

[0112] Optionally, when displaying and outputting the associated document set corresponding to the search information, the most similar standard questions may also be displayed and output together in the form of abbreviations.

[0113] This embodiment provides a method for determining a set of associated solutions for both valid and invalid search information inputs. This way, regardless of whether the user can accurately describe their help needs, the final search results will not be empty, providing a better search experience for the user.

[0114] Step S102 above uses a full-text search engine to search for a candidate document set. This method is a coarse-grained screening approach that doesn't consider the semantic similarity between the search information and the candidate document set. This can result in some candidate documents in the candidate document set not matching the search information's semantics. If the candidate document set were presented directly in the search results, the target user would need to browse through the candidate document set one by one and select helpful documents, which is time-consuming.

[0115] To save the target user's time, this embodiment can be used to perform fine-grained screening on the coarse-grained screening results to filter out candidate profiles that do not match the semantics. Based on this, the aforementioned "Step S105, determining search results corresponding to the search information based on the candidate profile set, candidate contacts, and associated profile set" can optionally include:

[0116] Step S1051: Match the candidate documents in the candidate document set with the search information based on semantic similarity, and obtain several candidate documents with high semantic similarity as the target document set.

[0117] That is, in this step, the search information may be matched with each candidate document in the candidate document set in terms of semantic similarity, and then the top K3 candidate documents with the highest semantic similarity are used as the target document set.

[0118] Optionally, this step can be implemented through a deep learning pre-trained language model, that is, the candidate materials in the candidate material set are matched with the search information by semantic similarity through the deep learning pre-trained language model to obtain several candidate materials with high semantic similarity.

[0119] Here, the deep learning pre-trained language model is trained using the training search information and the corresponding training materials as training data, and the semantic similarity between the labeled training search information and the training materials as sample labels.

[0120] Optionally, the deep learning pre-trained language model can be composed of a pre-trained language model under a multi-head attention mechanism, and this embodiment can train different pre-trained language models for different subject domains. Here, the multi-head attention mechanism is the core processing mechanism in mainstream deep learning natural language processing models, such as the transformer model and the BERT model. The multi-head attention mechanism obtains a vector representation of the text containing contextual semantics through matrix operations of multiple attention head devices and matrix linear transformation operations. In this embodiment, the multi-head attention mechanism mainly plays its significant advantages in processing the contextual semantics of long texts, improves the accuracy of tasks such as global log information mining, and supports the establishment of knowledge graphs.

[0121] Step S1052: The target data set, candidate contacts, and associated data set are used as search results corresponding to the search information.

[0122] This embodiment can perform fine-grained screening on the candidate data set and filter out semantically inconsistent candidate data, thereby making the final search results more accurate and saving the target user's time.

[0123] In summary of the above embodiments, optionally, if the search information is a valid input and is of the purpose type, the search results in this embodiment include three parts, namely, a target solution set and a target page set, candidate contacts (help-seeking contacts), and an associated solution set and an associated page set (potentially helpful information); if the search information is a valid input and is not of the purpose type, the search results in this embodiment include three parts, namely, a target solution set, candidate contacts (help-seeking contacts), and an associated solution set (potentially helpful information); if the search result is an invalid input, the search result in this embodiment includes two parts, namely, candidate contacts and an associated data set.

[0124] In this embodiment, the specific process of obtaining the search results can be referred to the introduction in the above embodiment, which will not be repeated here.

[0125] In an optional embodiment, considering that the existing technology can only return search results containing keywords in the search information when the search information is valid input, but cannot return the reasons for generating the search results, and users often need a set of corresponding return reasons, this embodiment can also generate a set of search return reasons corresponding to the search results when the search results are obtained, so as to explain the reasons for returning the search results.

[0126] In summary, this embodiment integrates information of different dimensions, such as page content, page title, page function description, historical question and answer information, and persons in charge of related functions, and uses information of each dimension to generate solutions, so that users can understand the system functions and the scenarios they are in and solve problems in one stop. Moreover, this embodiment uses knowledge graph information to mine the global behavioral elements of the system, and matches the information of each dimension corresponding to the highly relevant information nodes in the graph as candidate answers, so that users can obtain valuable information based on implicit data such as their own behavior without valid input. In addition, the results of knowledge graph matching are interpretable, which better meets the user's understanding needs for solution attribution.

[0127] It provides an end-to-end solution from search information to corresponding function pages, and from purpose description to transaction instructions or corresponding function pages. At the same time, it solves user operation problems based on explicit input information such as full search content, and implicit input information such as knowledge graph mining and user continuous operation records, and provides a more complete solution to improve user operation efficiency.

[0128] The embodiment of the present application further provides a search result generating device. The search result generating device provided in the embodiment of the present application is described below. The search result generating device described below and the search result generating method described above can refer to each other.

[0129] See also Figure 2, shows a schematic diagram of the structure of the search result generating device provided by an embodiment of the present application, such as Figure 2 As shown, the search result generating apparatus may include: a search information acquisition module 201 , a candidate data set determination module 202 , a candidate contact determination module 203 , a related data set determination module 204 and a search result determination module 205 .

[0130] The search information acquisition module 201 is used to acquire the search information input by the target user.

[0131] The candidate data set determining module 202 is configured to select data sets related to the search information from the preset full data set as candidate data sets.

[0132] The candidate contact determination module 203 is used to match the search information with historical related information and select the user corresponding to the matched historical related information as a candidate contact, wherein the historical related information includes historically published related materials and / or historical answer users, and historical answer users refer to users who have historically answered related questions.

[0133] The associated document set determination module 204 is used to determine the associated document set corresponding to the search information based on the search information and a pre-established knowledge graph and standard question topic corpus, wherein the knowledge graph is generated based on system log data and existing solution data, and the standard question topic corpus is generated based on existing solution data.

[0134] The search result determination module 205 is configured to determine the search result corresponding to the search information based on the candidate data set, the candidate contacts, and the associated data set.

[0135] The search result generation device provided by the present application first obtains the search information input by the target user, and then filters the data set related to the search information from the preset full data set as the candidate data set, then matches the search information with the historical related information, and uses the user corresponding to the matched historical related information as the candidate contact, and then determines the associated data set corresponding to the search information based on the search information and the pre-established knowledge graph and standard question subject corpus, and finally determines the search result corresponding to the search information based on the candidate data set, candidate contacts and associated data set. It can be seen that the present application only needs to input the search information once to return the search results determined based on the candidate data set, candidate contacts and associated data set, and the search results are more comprehensive and more in line with the user's needs.

[0136] In a possible implementation, the candidate material set determining module 202 may include: a purpose type determination submodule, a first candidate material set determining submodule, and a second candidate material set determining submodule.

[0137] The purpose type determination submodule is used to determine whether the search information is of a purpose type.

[0138] The first candidate data set determination submodule is used to screen a solution set related to the search information from a preset stock solution corpus, and to screen a page set related to the search information from a preset full page set, if the purpose type judgment submodule determines that the search information is of the purpose type. The screened solution set and page set are used as candidate data sets.

[0139] The first candidate material set determination submodule is used to screen solution sets related to the search information from the existing solution corpus if the purpose type determination submodule determines that the search information is not of the purpose type, and the screened solution sets are used as candidate material sets.

[0140] In a possible implementation, the candidate contact determination module 203 may include: a first candidate contact matching submodule and / or a second candidate contact matching submodule.

[0141] Among them, the first candidate contact matching submodule is used to match the search information with the relevant information published in the history through the data matching model, and use the user corresponding to the matched relevant information as the candidate contact. Among them, the data matching model is trained using the first training search information as training data and the users corresponding to the training relevant information matched by the marked first training search information as sample labels.

[0142] The second candidate contact matching submodule is used to match the search information with the historical answer users through the user matching model, and use the matched historical answer users as candidate contacts, wherein the user matching model is trained using the second training search information as training data and the training answer users that are marked to match the second training search information as sample labels.

[0143] In a possible implementation, the associated data set determining module 204 may include: a first associated data set determining submodule or a second associated data set determining submodule.

[0144] Among them, the first related data set determination submodule is used to match the search information with the standard question topic corpus if the search information is valid input, to obtain the standard question corresponding to the search information, and to match the standard question corresponding to the search information with the knowledge graph to obtain the related data set corresponding to the search information.

[0145] The first associated data set determination submodule is used to match the historical operation sequence related to the search information with the standard question subject corpus if the search information is invalid input, obtain the standard question corresponding to the historical operation sequence, and match the standard question corresponding to the historical operation sequence with the knowledge graph, and use the matched associated data set as the associated data set corresponding to the search information, wherein the historical operation sequence is determined through system log data.

[0146] In a possible implementation, the search result determination module 205 may include: a target document set determination submodule and a search result determination submodule.

[0147] The target data set determination submodule is used to perform semantic similarity matching between candidate data in the candidate data set and the search information, and obtain several candidate data with high semantic similarity as the target data set.

[0148] The search result determination submodule is used to take the target data set, candidate contacts and related data set as search results corresponding to the search information.

[0149] In one possible implementation, the above-mentioned target data set determination submodule can be specifically used to match the candidate data in the candidate data set with the search information by semantic similarity through a deep learning pre-trained language model to obtain several candidate data with high semantic similarity, wherein the deep learning pre-trained language model is trained using the training search information and the corresponding training data as training data, and the semantic similarity between the labeled training search information and the training data as sample labels.

[0150] In a possible implementation, the search result generating apparatus provided in the embodiment of the present application may further include: a search return reason generating module.

[0151] The search return reason generating module is used to generate a search return reason corresponding to the search result, wherein the search return reason is used to explain the reason for returning the search result.

[0152] In a possible implementation, the process of establishing a knowledge graph by the associated data set determination module 204 may include: a data acquisition submodule, a first relationship extraction submodule, a second relationship extraction submodule, and a knowledge graph establishment submodule.

[0153] The data acquisition submodule is used to obtain system log data and existing solution data, wherein the system log data is used to record the historical usage of users when solving problems.

[0154] The first relationship extraction submodule is used to obtain a number of historical operation sequences based on system log data cleaning, and extract correlation and dependency relationships from the number of historical operation sequences to obtain extracted correlation and dependency relationships.

[0155] The second relationship extraction submodule is used to extract label relationships and author relationships based on the existing solution data to obtain extracted label relationships and author relationships.

[0156] The knowledge graph establishment submodule is used to establish a knowledge graph based on correlation, dependency, label and author relationships, with solutions, answerers and topic labels as the three nodes of the knowledge graph.

[0157] The embodiment of the present application also provides a search result generating device. Optionally, Figure 3 The hardware structure diagram of the search result generating device is shown. Figure 3 , the hardware structure of the search result generating device may include: at least one processor 301, at least one communication interface 302, at least one memory 303 and at least one communication bus 304;

[0158] In the embodiment of the present application, the number of the processor 301, the communication interface 302, the memory 303, and the communication bus 304 is at least one, and the processor 301, the communication interface 302, and the memory 303 communicate with each other through the communication bus 304;

[0159] The processor 301 may be a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present invention;

[0160] The memory 303 may include a high-speed RAM memory, and may also include a non-volatile memory, such as at least one disk memory;

[0161] The memory 303 stores a program, and the processor 301 can call the program stored in the memory 303, and the program is used to:

[0162] Obtain the search information entered by the target user;

[0163] Filtering data sets related to the search information from the preset full data set as candidate data sets;

[0164] Match the search information with historical related information, and select the user corresponding to the matched historical related information as a candidate contact, wherein the historical related information includes historically published relevant materials and / or historical answer users, and the historical answer users are users who have historically answered related questions;

[0165] Determine the associated data set corresponding to the search information based on the search information and the pre-established knowledge graph and standard question topic corpus. The knowledge graph is generated based on system log data and existing solution data, and the standard question topic corpus is generated based on existing solution data.

[0166] Determine search results corresponding to the search information based on the candidate data set, candidate contacts, and associated data set.

[0167] Optionally, the detailed functions and extended functions of the program may refer to the above description.

[0168] An embodiment of the present application further provides a readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the search result generating method as described above is implemented.

[0169] Optionally, the detailed functions and extended functions of the program may refer to the above description.

[0170] Finally, it should be noted that, in this document, relational terms such as and and the like are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus comprising the element.

[0171] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0172] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A search result generating method, characterized in that: include: Obtain the search information entered by the target user; Filtering a data set related to the search information from a preset full data set as a candidate data set; Matching the search information with historical related information, and selecting users corresponding to the matched historical related information as candidate contacts, wherein the historical related information includes historically published relevant materials and / or historical answer users, wherein the historical answer users are users who have historically answered related questions; Determining a set of related documents corresponding to the search information based on the search information and a pre-established knowledge graph and a standard question topic corpus, wherein the knowledge graph is generated based on system log data and existing solution data, and the standard question topic corpus is generated based on the existing solution data; Performing semantic similarity matching between candidate documents in the candidate document set and the search information using a deep learning pre-trained language model to obtain a plurality of candidate documents with high semantic similarity as a target document set, wherein the deep learning pre-trained language model is trained using the training search information and corresponding training documents as training data and using the semantic similarity between the labeled training search information and the training documents as sample labels; and using the target document set, the candidate contacts, and the associated document set as search results corresponding to the search information; The method for establishing the knowledge graph includes: Acquire the system log data and the existing solution data, wherein the system log data is used to record historical usage of the user when solving the problem; A plurality of historical operation sequences are obtained based on the cleaning of the system log data, and correlation and dependency relationships are extracted from the plurality of historical operation sequences to obtain the extracted correlation and dependency relationships; Extracting label relationships and author relationships based on the existing solution data to obtain the extracted label relationships and author relationships; Based on the correlation relationship, the dependency relationship, the label relationship and the author relationship, the knowledge graph is established with solutions, answerers and topic labels as the three nodes of the knowledge graph.

2. The search result generating method according to claim 1, characterized in that: The step of screening a data set related to the search information from the preset full data set as a candidate data set includes: Determining whether the search information is of a purpose type, where the purpose type is a problem type that can provide a solution and an operation shortcut control; If so, a solution set related to the search information is screened from a preset stock solution corpus, and a page set related to the search information is screened from a preset full page set, and the screened solution set and page set are used as the candidate data set; If not, a solution set related to the search information is screened from the existing solution corpus, and the screened solution set is used as the candidate material set.

3. The search result generating method according to claim 1, wherein: The step of matching the search information with the historical related information and selecting the user corresponding to the matched historical related information as a candidate contact includes: Matching the search information with the historically published relevant information using a data matching model, and using the user corresponding to the matched relevant information as the candidate contact, wherein the data matching model is trained using the first training search information as training data and using the users corresponding to the training relevant information that are marked as matching the first training search information as sample labels; And / or, the search information is matched with the historical answer user through a user matching model, and the matched historical answer user is used as the candidate contact, wherein the user matching model is trained using the second training search information as training data and the training answer users that are marked as matching the second training search information as sample labels.

4. The search result generating method according to claim 1, wherein: Determining a set of related documents corresponding to the search information based on the search information and a pre-established knowledge graph and a standard question topic corpus includes: If the search information is a valid input, the search information is matched with the standard question subject corpus to obtain the standard question corresponding to the search information, and the standard question corresponding to the search information is matched with the knowledge graph to obtain the associated data set corresponding to the search information; If the search information is an invalid input, the historical operation sequence related to the search information is matched with the standard question subject corpus to obtain the standard question corresponding to the historical operation sequence, and the standard question corresponding to the historical operation sequence is matched with the knowledge graph, and the matched associated data set is used as the associated data set corresponding to the search information, wherein the historical operation sequence is determined by the system log data.

5. The search result generating method according to claim 1, characterized in that: Also includes: Generate a search return reason corresponding to the search result, wherein the search return reason is used to explain the reason for returning the search result.

6. A search result generating device, characterized in that: include: Search information acquisition module, candidate data set determination module, candidate contact determination module, associated data set determination module and search result determination module; The search information acquisition module is used to acquire the search information input by the target user; The candidate data set determining module is used to select a data set related to the search information from a preset full data set as a candidate data set; The candidate contact determination module is configured to match the search information with historically relevant information and select users corresponding to the matched historically relevant information as candidate contacts, wherein the historically relevant information includes historically published relevant materials and / or historically answering users, wherein the historically answering users are users who have historically answered relevant questions; The associated document set determination module is configured to determine the associated document set corresponding to the search information based on the search information and a pre-established knowledge graph and a standard question topic corpus, wherein the knowledge graph is generated based on system log data and existing solution data, and the standard question topic corpus is generated based on the existing solution data; The search result determination module is configured to perform semantic similarity matching between candidate documents in the candidate document set and the search information using a deep learning pre-trained language model, thereby obtaining a plurality of candidate documents with high semantic similarity as a target document set, wherein the deep learning pre-trained language model is trained using the training search information and corresponding training documents as training data and using the semantic similarity between the labeled training search information and the training documents as sample labels; and using the target document set, the candidate contacts, and the associated document set as search results corresponding to the search information; The process of establishing the knowledge graph includes: Acquire the system log data and the existing solution data, wherein the system log data is used to record historical usage of the user when solving the problem; A plurality of historical operation sequences are obtained based on the cleaning of the system log data, and correlation and dependency relationships are extracted from the plurality of historical operation sequences to obtain the extracted correlation and dependency relationships; Extracting label relationships and author relationships based on the existing solution data to obtain the extracted label relationships and author relationships; Based on the correlation relationship, the dependency relationship, the label relationship and the author relationship, the knowledge graph is established with solutions, answerers and topic labels as the three nodes of the knowledge graph.

7. The search result generating device according to claim 6, characterized in that: The candidate data set determination module includes: a purpose type determination submodule, a first candidate data set determination submodule, and a second candidate data set determination submodule; The purpose type determination submodule is used to determine whether the search information is of a purpose type, where the purpose type is a problem type that can provide a solution and an operation shortcut control; The first candidate material set determination submodule is configured to, if the purpose type determination submodule determines that the search information is of a purpose type, filter a solution set related to the search information from a preset stock solution corpus, and filter a page set related to the search information from a preset full page set, and use the filtered solution set and page set as the candidate material set; The first candidate material set determination submodule is used to screen a solution set related to the search information from the existing solution corpus if the purpose type judgment submodule determines that the search information is not of the purpose type, and the screened solution set is used as the candidate material set.

Images