A search keyword recommendation method, device and equipment and storage medium

Abstract

The application provides a search keyword recommendation method and device, equipment and a storage medium. The method is applied to a network community, and includes: obtaining a search keyword input by a user; performing text matching on the search keyword input by the user and candidate keywords in a preset candidate keyword library to obtain target candidate keywords meeting a preset matching condition; determining target recommendation keywords matched with the target candidate keywords based on target matching keyword pairs of the candidate keywords in the preset candidate keyword library and recommendation keywords in a preset recommendation keyword library; wherein the recommendation keywords in the preset recommendation keyword library include search keywords in historical search records, the search keywords having a search frequency greater than a second preset search frequency in a second preset time period and an evaluation index of search recall greater than a preset evaluation index; and the target recommendation keywords are recommended to the user. Through the technical solution of the application, the user can be recommended search keywords more matched with the search intention, and the search efficiency and accuracy of the user are improved.

Description

Technical Field

[0001] One or more embodiments of this application relate to the field of Internet technology, and in particular to a method, apparatus, device and storage medium for recommending search keywords. Background Technology

[0002] With the continuous development of computer science and technology, more and more users are using online searches to query and obtain various information. Typically, when a user enters a search term into a search engine, other related search terms are recommended below the search box. Traditional search term recommendation mechanisms recommend terms to users based on the search popularity of various search terms within the site and the user's personal search history. This recommendation mechanism is more suitable for general online communities and not applicable to highly specialized online communities.

[0003] Taking vertical communities as an example, these are product communities with a clear content focus. Consequently, they contain a large number of domain-specific terms and words with special meanings. When users search without relevant domain knowledge, they often encounter problems such as inaccurate search keywords and unclear search intent, leading to users failing to find the information they need or finding information that does not meet their needs, thus negatively impacting the user experience. Summary of the Invention

[0004] This application provides a method for recommending search keywords to address the shortcomings of related technologies.

[0005] According to a first aspect of one or more embodiments of this application, a method for recommending search keywords is provided, the method being applied to an online community, the method comprising:

[0006] Get the search terms entered by the user;

[0007] The user-input search term is matched with the candidate terms in the pre-set candidate term library to obtain target candidate terms that meet the preset matching conditions; the candidate terms in the pre-set candidate term library include search terms in historical search records whose search frequency is greater than the first preset search frequency within a first preset time period.

[0008] Based on the target matching word pairs between the candidate words in the pre-selected word library and the recommended words in the pre-recommended word library, target recommended words that match the target candidate words are determined; wherein, the recommended words in the pre-recommended word library include search words in the historical search records whose search frequency in the second pre-premised time period is greater than the second pre-premised search frequency, and whose search recall evaluation index is greater than the pre-premised evaluation index; wherein, the first pre-premised time period is longer than the second pre-premised time period, and / or the first pre-premised search frequency is less than the second pre-premised search frequency;

[0009] The target keywords are recommended to the user.

[0010] According to a second aspect of one or more embodiments of this application, a keyword recommendation device is provided, the device being applied to an online community, the device comprising:

[0011] The acquisition unit is used to acquire the search terms entered by the user;

[0012] The text matching unit is used to perform text matching between the user-input search term and the candidate terms contained in the pre-set candidate term library to obtain target candidate terms that meet the preset matching conditions; the candidate terms contained in the pre-set candidate term library include search terms in historical search records whose search frequency is greater than the first preset search frequency within a first preset time period.

[0013] The determining unit is configured to determine a target recommended word that matches the target candidate word based on the target matching word pair between the candidate words contained in the pre-selected word library and the recommended words contained in the pre-recommended word library; wherein, the recommended words contained in the pre-recommended word library include search words in the historical search records whose search frequency in a pre-preset second time period is greater than a second pre-preset search frequency, and whose search recall evaluation index is greater than a pre-preset evaluation index; wherein, the first pre-preset time period is longer than the second pre-preset time period, and / or the first pre-preset search frequency is less than the second pre-preset search frequency;

[0014] The recommendation unit is used to recommend the target keywords to the user.

[0015] According to a third aspect of one or more embodiments of this application, an electronic device is provided, comprising:

[0016] processor;

[0017] Memory used to store processor-executable instructions;

[0018] The processor executes the executable instructions to implement the method described in the first aspect above.

[0019] According to a fourth aspect of one or more embodiments of this application, a computer-readable storage medium is provided that stores computer instructions thereon, which, when executed by a processor, implement the steps of the method as described in the embodiments of the first aspect above.

[0020] As can be seen from the above technical solutions, in one or more embodiments of this application, a pre-selected term library and a preset recommended term library are set, and the user-input search terms are matched and filtered twice based on these two term libraries. Since the recommended terms contained in the preset recommended term library are high-frequency and highly rated search terms in historical search records, recommending the matched target recommended terms to the user can make the recommended terms more consistent with the user's search intent, thereby allowing the user to quickly obtain search results related to their search intent and improving the efficiency of the user's search.

[0021] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0022] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0023] Figure 1 This is a flowchart of a method for recommending search keywords according to an exemplary embodiment of this application.

[0024] Figure 2 This is a flowchart of a multi-path recall matching method provided according to an exemplary embodiment of this application.

[0025] Figure 3 This is a flowchart illustrating a training method for a search keyword recommendation model according to an exemplary embodiment of this application.

[0026] Figure 4 This is a schematic diagram of the structure of an electronic device according to an exemplary embodiment of this application.

[0027] Figure 5 This is a block diagram illustrating a keyword recommendation device according to an exemplary embodiment of this application. Detailed Implementation

[0028] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of this application as detailed in the appended claims.

[0029] It should be noted that the steps of the corresponding methods in other embodiments are not necessarily performed in the order shown and described in this application. In some other embodiments, the methods may include more or fewer steps than those described in this application. Furthermore, a single step described in this application may be broken down into multiple steps in other embodiments; and multiple steps described in this application may be combined into a single step in other embodiments.

[0030] With the development of science and technology, more and more users are searching for and obtaining information in online communities. Different online communities encompass varying amounts of information. In some highly specialized online communities, there are numerous domain-specific terms and words with special meanings, such as financial communities and automotive communities. When users search for information in these specialized online communities, they are easily unable to find the desired information due to a lack of relevant knowledge, leading to inaccurate search keywords.

[0031] This application provides a method for recommending search keywords, which helps users clarify their search intent, conduct precise searches, and improve search efficiency by recommending relevant proper nouns and words with clear meanings in the field.

[0032] Figure 1 This is a flowchart illustrating a method for recommending search keywords, as provided in an exemplary embodiment. Figure 1 As shown, this method is applied to online communities and may include the following steps:

[0033] S101: Obtain the search terms entered by the user.

[0034] In this embodiment, the search terms entered by the user represent the user's search request. These search terms can be single words, phrases of multiple words, Chinese characters, pinyin, or different word orders with the same meaning. For example, search terms could be "yueliang," "Shanghai food distribution," or "food distribution in Shanghai," etc.

[0035] For example, if a user wants to learn about vehicle supercharging, the search terms they enter could be "vehicle supercharging", "supercharging", "supercharging technology", "vehicle charging", etc.

[0036] S102: Perform text matching between the user-input search term and the candidate terms contained in the pre-set candidate term library to obtain target candidate terms that meet the preset matching conditions; the candidate terms contained in the pre-set candidate term library include search terms in historical search records whose search frequency is greater than the first preset search frequency within a first preset time period.

[0037] Text matching aims to study the relationship between two texts and is commonly used in fields such as text similarity, information search, and question-answering matching. Traditional text matching techniques include algorithms such as BoW, TF-IDF, BM25, Jaccard, and SimHash.

[0038] In this embodiment, Minhash_Jaccard_Sim is used to perform approximate text matching between the user-input search term and the candidate terms contained in the pre-set candidate term library to obtain target candidate terms that meet the preset matching conditions. Referring to the embodiment in S101, the filtering condition for the target candidate terms is set as follows: sorted by matching degree from high to low, the top 5 candidate terms with the highest matching degree are selected as target candidate terms. This matching method can also be called text nearest neighbor matching. Therefore, when the user-input search term is "vehicle charging", the target candidate terms could be "supercharging", "charging pile", "supercharging station", "fast charging", and "supercharging price".

[0039] In this embodiment, the candidate terms in the pre-selected keyword library are search terms whose search frequency in historical search records is greater than a first preset search frequency within a first preset time period. The historical search records involved here refer to the historical search records of all users in a certain online community. It should be noted that user historical search records involve user privacy, and user consent is required to obtain user historical search records. The first preset time period and the first preset search frequency can be set according to business needs. For example, the first preset time period can be set to 365 days, and the first preset search frequency can be set to 5 times. That is, if a search term appears more than 5 times in the historical search records of the community within the most recent 365 days, then that search term can be used as a candidate term in the pre-selected keyword library. By setting the pre-selected keyword library using historical search records within different preset time periods, the candidate terms contained in the pre-selected keyword library can change over time, thereby ensuring the timeliness of the candidate terms.

[0040] S103: Based on the target matching word pairs of the candidate words contained in the pre-selected word library and the recommended words contained in the pre-recommended word library, determine the target recommended words that match the target candidate words; wherein, the recommended words contained in the pre-recommended word library include search words in the historical search records whose search frequency in the second pre-premised time period is greater than the second pre-premised search frequency, and whose search recall evaluation index is greater than the pre-premised evaluation index; wherein, the first pre-premised time period is longer than the second pre-premised time period, and / or the first pre-premised search frequency is less than the second pre-premised search frequency.

[0041] In this embodiment, the recommended terms in the preset recommended term library are those whose search frequency in historical search records is greater than the second preset search frequency within a second preset time period, and whose search recall evaluation index is greater than the preset evaluation index. Similar to the preset recommended term library, the historical search records involved here are the historical search records of all users in the same online community. The second preset time period, the second preset search frequency, and the preset evaluation index can be set according to business needs. By setting the preset recommended term library using historical search records within different preset time periods, the recommended terms contained in the preset recommended term library can change over time, thereby ensuring the timeliness of the recommended terms. In one embodiment, the search recall evaluation index may include the number of articles recalled and the click-through rate of the search recall results. When conducting search recall based on a certain search term, the more articles recalled or the higher the click-through rate of the search recall results, the closer the search term is to keywords with special and professional meaning in the field. Therefore, by filtering recommended terms from historical search records according to preset rules to form a preset recommended term library, it is possible to recommend relevant proper nouns and words with clear meanings in the field to users, so as to help users clarify their search intent and thus improve the efficiency and accuracy of users' searches.

[0042] Based on the embodiment in S101, the second preset time period can be set to the most recent 60 days. The search frequency, the number of articles retrieved by the search terms, and the click-through rate of the retrieved results for all search terms in the historical search records within these 60 days are sorted from high to low. Search terms with the highest search frequency (top 50%), the highest number of articles retrieved (top 30%), and the highest click-through rate of the retrieved results (top 30%) are selected as recommended terms, forming a preset recommended term library. Then, based on the target matching word pairs between the candidate terms in the preset candidate term library and the recommended terms in the preset recommended term library, a recommended term matching the target candidate term is queried in the preset recommended term library, and this recommended term is determined as the target recommended term. Based on the target matching word pairs, the target recommended term matching the target candidate term "supercharging" can be "supercharging station," "battery," or "fast charging port."

[0043] In one embodiment, the target matching word pairs between candidate words in the pre-set optional word library and recommended words in the pre-set recommended word library can be generated through multi-path recall matching. Multi-path recall matching can include: text matching, semantic vector matching, and collaborative filtering matching.

[0044] Figure 2 This is an exemplary flowchart of multi-way recall matching, which will be discussed below. Figure 2 The process of three-way recall matching will be introduced in turn.

[0045] (1) Text matching

[0046] In this embodiment, text matching mainly involves calculating the text similarity matrix between candidate words in the pre-set candidate word library and recommended words in the pre-set recommended word library. The Jaccard similarity algorithm can be used for this calculation. Then, based on the text similarity results between each candidate word and each recommended word, text matching word pairs with a text similarity greater than a first preset text similarity threshold are output. Each text matching word pair contains one candidate word and one recommended word. The first preset text similarity threshold can be set according to actual needs; for example, in this embodiment, it can be set to 0.5.

[0047] (2) Semantic vector matching

[0048] Semantic vector matching is a fundamental problem in natural language processing. It involves converting text words into vector representations and then calculating the similarity between two vectors to determine the degree of matching between two text words. Methods for encoding text words into corresponding word vectors include BERT (Bidirectional Encoder Representations from Transformers) and Word2Vec, while the similarity between two vectors can be calculated using methods such as cosine similarity, Euclidean distance, and Manhattan distance. This application does not impose any restrictions on these methods.

[0049] In this embodiment, BERT vector generation and cosine similarity calculation are used to calculate the semantic vector similarity between candidate words in the pre-selected word library and recommended words in the pre-selected word library. Semantic vector matching word pairs with a similarity greater than a pre-set threshold are output. Each semantic vector matching word pair contains one candidate word and one recommended word. The pre-set vector similarity threshold can be set according to actual needs; for example, it can be set to 0.5 in this embodiment.

[0050] (3) Collaborative filtering matching

[0051] First, text matching is performed on historical search records based on a pre-set candidate term library to obtain the target historical search records. Specifically, a one-to-one text similarity calculation is performed between the candidate terms in the pre-set candidate term library and the historical search terms in the historical search records. The candidate term with the highest text similarity is selected. Then, all historical search terms with a text similarity greater than a second preset text similarity threshold are uniformly mapped to the candidate term with the highest text similarity, while historical search terms with a text similarity less than the second preset text similarity threshold are retained in their original form without processing. Therefore, the obtained target historical search records contain the candidate term with the highest text similarity and historical search terms retained in their original form.

[0052] In this embodiment, the Jaccard similarity algorithm can be used to perform text matching on the pre-set candidate word library and historical search records, and the second preset text similarity threshold is set to 0.5. Of course, the second preset text similarity threshold can be set according to requirements. By uniformly mapping all search terms greater than the second preset text similarity threshold to the candidate words with the highest text similarity, noise reduction processing of the data is achieved, which helps to reduce the complexity of the subsequent matching process.

[0053] Then, based on the collaborative filtering algorithm, the collaborative similarity between historical search terms in the target's historical search records is calculated, resulting in multiple collaborative matching word pairs and the collaborative similarity of each pair. Collaborative filtering recommendation obtains the similarity between different search terms by comprehensively analyzing the same user groups across different search terms. The collaborative filtering used in this application is content-based collaborative filtering, which analyzes the commonalities between different content based on user interaction behavior to obtain the similarity between different content.

[0054] Finally, candidate matching word pairs are selected from multiple collaborative matching word pairs. Each collaborative matching word pair contains two search terms, which are compared with candidate terms in the pre-selected word library and recommended terms in the pre-recommended word library. If one search term in a collaborative matching word pair is a candidate term in the pre-selected word library and the other search term is a recommended term in the pre-recommended word library, then the collaborative matching word pair can be considered a candidate matching word pair.

[0055] Because the target's historical search records were collaboratively filtered, the selected terms were the most frequently used historical search terms by most users in the user community. This made the recommended terms selected based on these historical search terms more closely related to the proper nouns and words with clear meanings in the field.

[0056] In summary, text matching, semantic vector matching, and collaborative filtering all output multiple matching word pairs and their similarity scores. The similarity is calculated by combining the matching word pairs output by these three methods, and then selecting matching word pairs whose combined similarity is greater than a preset similarity threshold as target matching word pairs. For example, methods such as addition or weighted averaging can be used for the combination calculation; this application does not impose any restrictions on this approach.

[0057] Taking weighted average as an example, in this embodiment, the priority of the three-way recall matching is determined according to business needs, with higher priority terms having greater weights and lower priority terms having smaller weights. The similarity of the concatenated three-way recall matching is calculated based on the weight of each recall matching path, and this similarity is sorted from high to low. The top 10 pairs of candidate words with the highest similarity are selected to establish target matching word pairs with the recommended words. Specifically, when there is a lot of user interaction in the community, collaborative filtering matching has a larger weight; when there is less user interaction in the community, collaborative filtering matching has a smaller weight, while semantic vector matching and text matching have larger weights. When the accuracy of semantic vector representation of proper nouns in the community is low, it is necessary to consider each search term at the text level, which can correspondingly increase the weight of text matching and decrease the weight of semantic vector matching. Conversely, for general search terms with high accuracy in semantic vector representation, the weight of semantic vectors can be increased and the weight of text matching decreased.

[0058] By obtaining recommended words from candidate words under multi-path matching recall conditions through semantic matching, text matching, and collaborative filtering matching, the diversification of recommended words can be achieved. For example, the candidate word "supercharging" can be matched with three recommended words: "supercharging station", "battery", and "fast charging port", instead of just "supercharging station".

[0059] S104: Recommend the target keywords to the user.

[0060] In one embodiment, the target recommended terms obtained in S103 can be recommended to the user. In S101-S103, the user-input search terms are matched and filtered twice based on a pre-selected term library and a preset recommended term library. The resulting target recommended terms are proper nouns and words with clear meanings within the community's domain, or search terms very close to these two types of words. Therefore, when users use the target recommended terms to search, they can more quickly find content related to their search intent, avoiding the problem of inaccurate search results due to a lack of domain knowledge.

[0061] In another embodiment, since a network community contains a massive number of users, the profile characteristics of different users are not entirely the same. Therefore, the target recommendation terms can be further processed to recommend personalized search terms to different users. The target recommendation terms, the attribute features of the target recommendation terms, and the user profile features are input into the search keyword recommendation model, and personalized recommendation terms are output. The attribute features of the target recommendation terms may include the search frequency of the target recommendation terms in different time periods and the evaluation indicators of search recall, such as the number of articles recalled and the click-through rate of the search recall results.

[0062] In this embodiment, user profile features can include user community identity features, user community points features, and user community level features. Because different users have slightly different profile features, the recommendation weight of personalized recommendation words will vary for different users. Recommendation weight is a probability score; the higher the probability score of a recommendation word, the greater its recommendation weight, and the more likely it is to be clicked by the user. Personalized recommendation words can be sorted according to their recommendation weight, and then a certain number of personalized recommendation words can be selected and recommended to the user. For example, the top 5 personalized recommendation words with the highest recommendation weight can be selected and recommended to the user, thus achieving personalized search keyword recommendations for the user based on the target recommendation words.

[0063] In one embodiment, the search keyword recommendation model can generate probability scores for personalized recommendation words using the DropoutNet algorithm. DropoutNet is an embedding learning model applied to cold-start scenarios. In the recommendation field, a cold-start scenario refers to the situation where, when a new user or item joins the recommendation system, the system lacks sufficient historical interaction data, making it impossible to obtain accurate recommendation content or recommend the item to the appropriate user. One solution is to generalize the new user profile, moving towards broader concepts in terms of attributes or topics. In this application, the recommendation model can be trained based on the historical search records and user profile features of other users within the community. This allows the model to learn generalized attribute or topic feature weights, and then use the trained recommendation model to recommend personalized search keywords to cold-start users with the same generalized attributes or topic features. For example, if a cold-start user's identity is "beginner," their level is "primary," and other interaction information is unknown, then the model can predict and rank the target recommendation words for the current user based on the weight information learned from the "beginner" and "primary" levels, generating personalized recommendation words, and then recommending these personalized words to the cold-start user. This method of recommending search keywords enables relatively accurate search keyword recommendations even for users starting out.

[0064] Therefore, in response to the aforementioned search keyword recommendation model, this application also proposes a corresponding model training scheme, which will be discussed below. Figure 3 A detailed introduction will follow.

[0065] Figure 3 This is a schematic flowchart illustrating a training method for a search keyword recommendation model, provided as an exemplary embodiment. Figure 3 As shown:

[0066] First, a training sample set needs to be obtained, including a training positive sample set and a training negative sample set. The training positive sample set includes the sample user's historical search records, the attribute features of the historical search terms contained in the historical search records, the user profile features, and the target probability score of the training positive sample set. The training negative sample set includes predefined sample search terms, the attribute features of the predefined sample search terms, the user profile features, and the target probability score of the training negative sample set. Note that the predefined sample search terms do not include the historical search terms contained in the historical search records.

[0067] In one embodiment, users whose historical search records were included within a certain period prior to model training can be selected as the sample user set for model training. The profile features of the sample users in the sample user set are obtained, along with the attribute features of the historical search terms contained in the sample's historical search records. These attribute features may include the search frequency of the historical search terms during that period, the number of articles retrieved, the click-through rate of the retrieved results, etc. The profile features of the sample users include their community identity features, community points features, and community level features. For example, the sample users' historical search records within the 30 days prior to model training can be selected, and the sample users' historical search records within these 30 days, the attribute features of the historical search terms contained in the historical search records, and the sample users' profile features can be combined to form a positive training sample set.

[0068] Correspondingly, a training negative sample set can be obtained through predefined rules. The training negative sample set includes predefined sample search terms and their attribute features, as well as the profile features of sample users. In this embodiment, the historical search records of all users in the community within the 30 days prior to model training can be filtered to obtain a set of historical search terms with a search frequency greater than 100. Then, the set of IDs of all users with search records in the community within the 30 days prior to model training is correlated with this set of historical search terms using a Cartesian product, from which historical search terms not searched by sample users are filtered out. These historical search terms not searched by sample users are used as predefined sample search terms.

[0069] Because the training positive sample set is obtained based on the sample users' historical search records, it represents the sample users' search interests to a certain extent. The training negative sample set, on the other hand, consists of words that sample users have not searched for, representing, to some extent, users' lack of interest in the predefined sample search terms in the training negative sample set. Therefore, the target probability score for the training positive sample set can be set to 1, and the target probability score for the training negative sample set can be set to 0.

[0070] Then, the training sample set is input into the model to be trained, and the predicted probability score of each search term in the training sample set is output. The sample search terms include the historical search terms in the positive training sample set and the predefined sample search terms in the negative training sample set.

[0071] Finally, the model is iteratively trained based on the predicted probability score and the target probability score of each sample's search term.

[0072] In this embodiment, after generating the predicted probability score of the sample search term using the DropoutNet algorithm, the predicted probability score of the sample search term is compared with the target probability score, and the search keyword recommendation model is iteratively trained based on the difference between the two. A preset difference threshold can be set. If the difference between the predicted probability score and the target probability score of the sample search term is less than the preset difference threshold, the model is considered to have achieved the expected training effect, and the model training can be terminated. The model can then be used to achieve, for example... Figure 1 The recommended search keywords are shown.

[0073] Corresponding to the above-described method embodiment for recommending search keywords, this application also provides an embodiment of a device for recommending search keywords.

[0074] Figure 4 This is a schematic diagram illustrating the structure of an electronic device according to an exemplary embodiment of this application. (Reference) Figure 4 At the hardware level, the electronic device includes a processor 402, an internal bus 404, a network interface 406, memory 408, and non-volatile memory 410, and may also include other hardware required for business operations. The processor 402 reads the corresponding computer program from the non-volatile memory 410 into memory 408 and then runs it. Of course, in addition to software implementation, this application does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution subject of the following processing flow is not limited to individual logic units, but can also be hardware or logic devices.

[0075] Figure 5 This is a block diagram illustrating a keyword recommendation device according to an exemplary embodiment of this application. (Refer to...) Figure 5 The device includes an acquisition unit 502, a text matching unit 504, a determination unit 506, and a recommendation unit 508, wherein:

[0076] The acquisition unit 502 is configured to acquire the search terms entered by the user.

[0077] The text matching unit 504 is configured to perform text matching between the user-input search term and the candidate terms contained in the pre-set candidate term library to obtain target candidate terms that meet preset matching conditions; the candidate terms contained in the pre-set candidate term library include search terms in historical search records whose search frequency is greater than the first preset search frequency within a first preset time period.

[0078] The determining unit 506 is configured to determine a target recommended word that matches the target candidate word based on the target matching word pair between the candidate words contained in the pre-selected word library and the recommended words contained in the preset recommended word library; wherein, the recommended words contained in the preset recommended word library include search words in the historical search records whose search frequency in a second preset time period is greater than a second preset search frequency, and whose search recall evaluation index is greater than a preset evaluation index; wherein, the first preset time period is longer than the second preset time period, and / or the first preset search frequency is less than the second preset search frequency.

[0079] Recommendation unit 508 is configured to recommend the target recommendation words to the user.

[0080] Optionally, the evaluation metrics for the search recall include: the number of articles retrieved and / or the click-through rate of the search recall results.

[0081] Optionally, the device further includes:

[0082] The matching word pair generation unit 510 is configured to perform text matching on the historical search records according to the pre-selected word library to obtain target historical search records; perform collaborative filtering matching on the target historical search records to obtain at least one collaborative matching word pair and the collaborative similarity of the at least one collaborative matching word pair; filter candidate matching word pairs from the at least one collaborative matching word pair, wherein the filtering condition for the candidate matching word pair is: one search word in the candidate matching word pair is a candidate word contained in the pre-selected word library and the other search word is a recommended word contained in the preset recommendation word library; and determine the target matching word pair based on the candidate matching word pair and the similarity of the candidate matching word pair.

[0083] Optionally, the matching word pair generation unit 510 is specifically used to perform semantic vector matching between candidate words in the pre-selected word library and recommended words in the preset recommended word library to obtain the similarity between semantic vector matching word pairs; to perform text matching between candidate words in the pre-selected word library and recommended words in the preset recommended word library to obtain the similarity between text matching word pairs; and to determine the target matching word pair based on the similarity between semantic vector matching word pairs, the similarity between text matching word pairs, and the similarity between candidate matching word pairs.

[0084] Optionally, the device further includes:

[0085] The personalized recommendation unit 512 is configured to input the target recommendation word, the attribute features of the target recommendation word, and the user's profile features into the search keyword recommendation model to obtain personalized recommendation words; and to recommend the personalized recommendation words to the user.

[0086] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0087] The apparatus or unit described in the above embodiments can be implemented by a computer chip or physical entity, or by a product with a certain function. A typical implementation device is a computer, which can be a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email sending and receiving device, game console, tablet computer, wearable device, or any combination of these devices.

[0088] In a typical configuration, a computer includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0089] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0090] Computer-readable media, including both permanent and non-permanent, removable and non-removable media, can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage, quantum memory, graphene-based storage media or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0091] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method of recommending a search keyword, characterized by, The method is applied to online communities, and the method includes: Get the search terms entered by the user; The user-input search term is matched with the candidate terms in the pre-set candidate term library to obtain target candidate terms that meet the preset matching conditions; the candidate terms in the pre-set candidate term library include search terms in historical search records whose search frequency is greater than the first preset search frequency within a first preset time period. Based on the target matching word pairs of candidate words in the pre-selected word library and recommended words in the pre-recommended word library, target recommended words that match the target candidate words are determined; wherein, the target matching word pairs are determined based on matching word pairs filtered through multi-way recall matching and the similarity of the matching word pairs, and the matching word pairs include one candidate word and one recommended word; the recommended words in the pre-recommended word library include search words in the historical search records whose search frequency in the second pre-recommended time period is greater than the second pre-recommended search frequency, and whose search recall evaluation index is greater than the pre-recommended evaluation index; wherein, the first pre-recommended time period is longer than the second pre-recommended time period, and / or the first pre-recommended search frequency is less than the second pre-recommended search frequency; The target keywords are recommended to the user.

2. The method of claim 1, wherein, The evaluation metrics for search recall include: the number of articles recalled and / or the click-through rate of the search recall results.

3. The method of claim 1, wherein, Also includes: The target historical search record is obtained by performing text matching on the pre-selected word library. Collaborative filtering matching is performed on the target historical search records to obtain at least one collaborative matching word pair and the collaborative similarity of the at least one collaborative matching word pair; Candidate matching word pairs are selected from the at least one collaborative matching word pair, wherein the selection criteria for the candidate matching word pair are: one of the search terms in the candidate matching word pair is a candidate word contained in the pre-set optional word library, and the other search term is a recommended word contained in the pre-set recommended word library; The target matching word pair is determined based on the similarity between the candidate matching word pairs and the candidate matching word pairs.

4. The method of claim 3, wherein, Also includes: Semantic vector matching is performed between the candidate words in the pre-selected word library and the recommended words in the pre-recommended word library to obtain the similarity between the semantic vector matching word pairs and the semantic vector matching word pairs; The candidate words in the pre-selected word library and the recommended words in the pre-recommended word library are matched by text to obtain the similarity between the text matching word pairs and the text matching word pairs. Determining the target matching word pair based on the similarity between the candidate matching word pairs includes: The target matching word pair is determined based on the similarity between the semantic vector matching word pairs, the similarity between the text matching word pairs, and the similarity between the candidate matching word pairs.

5. The method of claim 1, wherein, The step of recommending the target keywords to the user includes: The target recommendation word, the attribute features of the target recommendation word, and the user's profile features are input into the search keyword recommendation model to obtain personalized recommendation words; The personalized recommendation words are then recommended to the user.

6. The method according to claim 5, characterized in that, The search keyword recommendation model is trained in the following way: A training sample set is obtained, comprising a positive training sample set and a negative training sample set. The positive training sample set includes the sample user's historical search records, the attribute features of the historical search terms contained in the historical search records, the user profile features, and the target probability score of the positive training sample set. The negative training sample set includes predefined search terms, the attribute features of the predefined search terms, the user profile features, and the target probability score of the negative training sample set. The predefined search terms do not include the historical search terms contained in the historical search records. The training sample set is input into the model to be trained, and the predicted probability score of each sample search term in the training sample set is output. The sample search terms include the historical search terms in the positive training sample set and the predefined sample search terms in the negative training sample set. The model is iteratively trained based on the predicted probability score and the target probability score of each sample search term.

7. A recommendation device for search keywords, characterized in that, The device is used in online communities, and the device includes: The acquisition unit is used to acquire the search terms entered by the user; The text matching unit is used to perform text matching between the user-input search term and the candidate terms contained in the pre-set candidate term library to obtain target candidate terms that meet the preset matching conditions; the candidate terms contained in the pre-set candidate term library include search terms in historical search records whose search frequency is greater than the first preset search frequency within a first preset time period. The determining unit is configured to determine a target recommended word that matches the target candidate word based on the target matching word pair between the candidate words in the pre-selected word library and the recommended words in the pre-recommended word library; wherein the target matching word pair is determined based on the matching word pairs filtered through multi-way recall matching and the similarity of the matching word pairs, and the matching word pair includes one candidate word and one recommended word; the recommended words in the pre-recommended word library include search words in the historical search records whose search frequency in a pre-premised second time period is greater than a second pre-premised search frequency, and whose search recall evaluation index is greater than a pre-premised evaluation index; wherein the first pre-premised time period is longer than the second pre-premised time period, and / or the first pre-premised search frequency is less than the second pre-premised search frequency; The recommendation unit is used to recommend the target keywords to the user.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor implements the method as described in any one of claims 1-6 by executing the executable instructions.

9. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, this instruction implements the steps of the method as described in any one of claims 1-6.

10. A vehicle comprising a memory, a processor, and computer instructions stored in the memory and executable on the processor, characterized in that, When the processor executes the instructions, it implements the steps of the method as described in any one of claims 1-6.

Images