A method for intelligent auxiliary search input

Through multiple recall and comprehensive analysis technology, combined with user behavior and content information, the problems of insufficient relevance, timeliness and personalization in existing search engine assisted input technologies are solved, and more efficient search result prompts are achieved, which improves users Experience and satisfaction.

CN114385925BActive Publication Date: 2024-11-15DATAGRAND TECH INC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210083151.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-01-25
Publication Date
2024-11-15
Estimated Expiration
2042-01-25

AI Technical Summary

Technical Problem

The existing search engine-assisted input technology has problems such as single text correlation calculation method, single data source, insufficient user search behavior data, no time impact and insufficient personalization, resulting in coverage, correlation, timeliness and The degree of personalization is insufficient, which affects user experience and efficiency.

Method used

Using multiple recall candidate sets, static analysis and dynamic analysis methods, combined with user query statements, historical search behavior, content information and timestamps, preorder matching, word embedding, and text semantics are performed through Trie tree, word2vec model and Elasticsearch technologies. Recall, and static and dynamic analysis are carried out, and finally the calculation quality scores are integrated to improve the coverage, relevance, timeliness and personalization of the prompt results.

Benefits of technology

It effectively improves the coverage, relevance, timeliness and personalization of prompt results, improves user experience and search efficiency, and improves user satisfaction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114385925B_ABST
    Figure CN114385925B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for intelligent auxiliary search input, which includes the following steps: S1. data pre-processing; S2. multi-way recall candidate set: multi-way recall of the search query #imgabs0# to be prompted, then merging the multi-way recall results, and filtering duplicates to obtain the final M candidate results; S3. static analysis and dynamic analysis: for the M candidate result sets obtained in step S2, the #imgabs1# result is recorded as #imgabs2#, and static analysis and dynamic analysis are performed respectively; S4. fusion sorting: according to the static analysis and dynamic analysis results in step S3, the overall quality score of #imgabs3# is fused and calculated, and then the M candidate result sets are sorted from large to small according to the quality score, and the top K results ranked in front are taken and returned as the final result of intelligent auxiliary input. The present invention can effectively improve the coverage, relevance, timeliness, personalization, etc. of the prompt results, greatly improve the user experience, and improve the efficiency and satisfaction of user search.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the fields of the Internet and big data, and in particular to a method for intelligent auxiliary search input. Background Art

[0002] With the advent of the big data era and the rise of mobile Internet, information technology has completely changed people's daily lifestyles, and various types of APP applications have covered all aspects of people's lives. Obtaining information is a rigid need for human cognition of the world, survival and development, but people are surrounded by massive amounts of information every day. How to obtain information better and faster has become an urgent problem that needs to be solved. Search engines undoubtedly play an irreplaceable role in this process. When our needs are very clear, enter the content you want to find in the search box, and the search engine can quickly and accurately return the results the user wants. Limited by the size of the mobile phone screen and the input method, the typing speed is reduced, the time required for input is increased, and typos are prone to occur, which greatly affects people's efficiency in obtaining information and reduces the user experience. Therefore, assisting user input has become a standard function of search engines.

[0003] The existing auxiliary user input technologies in the industry are mainly based on the auto-completion function and prefix tree matching provided by open source tools, but they have the following problems:

[0004] 1. The text relevance calculation method is single, resulting in insufficient text relevance of the returned prompt results, which is even worse when there are typos, pinyin, Martian text, etc.;

[0005] 2. The data source used is too single. In addition to the user's historical search logs, there are also content tags, titles, descriptions, texts and other information that need to be taken into account and can all be used as input data to increase the coverage and relevance of the results;

[0006] 3. Not making full use of user search behavior data, including the query statements entered and the clicks and selections after the prompt results are given, which directly reflect the user's real information needs;

[0007] 4. The impact of time is not taken into account. As time goes by, a large amount of new content is constantly generated, and the returned prompt results also need to consider the effectiveness of the results;

[0008] 5. The returned results are not personalized enough. People’s information demands vary, and the results need to be more personalized. Summary of the invention

[0009] In order to address the deficiencies in the above-mentioned prior art, the present invention provides a method for intelligent assisted search input, which can effectively improve the coverage, relevance, timeliness, personalization, etc. of the prompt results, greatly improve the user experience, and improve the efficiency and satisfaction of user search.

[0010] In order to achieve the above-mentioned object of the invention, the present invention provides the following technical solutions:

[0011] A method for intelligent auxiliary search input, the method comprising the following steps:

[0012] S1. Data pre-processing: take the most recent period Including search behavior and prompt click behavior Behavioral dataset , based on the behavioral dataset Search behavior Data, obtain the user query statement dataset after filtering , the user query statement dataset The length of , the behavior dataset Search behavior The number of , the prompt click behavior The number of , the user query statement dataset All are stored in the tree Trie, recorded as , the user query statement dataset All word vectors are generated using the word2vec model and stored in Faiss and recorded as , import all the content information to be searched into Elasticsearch and record it as , the content information includes field information, the field information includes title, tag, description and text; the timestamp of the current time is recorded as , for the search query to be prompted , the length of the string is recorded as ;

[0013] S2. Multi-way recall candidate set: search query to treat prompts Perform multi-way recall, which includes pre-order matching recall, word embedding recall and text semantic recall, and then merge the results of the pre-order matching recall, word embedding recall and text semantic recall, and perform duplicate filtering to obtain the final M candidate results;

[0014] S3. Static analysis and dynamic analysis: For the M candidate result sets obtained in step S2, The results are recorded as , respectively, static analysis and dynamic analysis are performed, the static analysis includes pre-order matching analysis and global matching analysis, the static analysis is used to analyze the text relevance of candidate results, the dynamic analysis includes behavior data analysis, user interest analysis, context analysis and timeliness analysis, the dynamic analysis is used to conduct a comprehensive analysis of candidate results combined with user historical behavior data;

[0015] S4. Fusion sorting: Based on the static analysis and dynamic analysis results in step S3, fusion calculation The overall quality score of the task is calculated, and then the M candidate result sets are sorted from large to small according to the quality score, and the top K results are taken and returned as the final result of the intelligent auxiliary input.

[0016] In step S2, the method for recalling the preceding order matches includes: According to the previous order Perform the longest match query, then sort the returned results from high to low according to the tree depth, and return the top N results; the word embedding recall method includes: Use word2vec to generate word vectors The most similar TopN results are returned in the query; the text semantic recall method includes: Call The suggest function returns the top N results.

[0017] In step S3, the method for preamble matching analysis includes the following steps: matching the M candidate result sets obtained in step S2 one by one from the front to the back to obtain the length of the longest matching string , express Length, express and The smallest one, and Respectively and the search query to be suggested In the The value at the position is calculated as follows: and Pre-order matching degree:

[0018] .

[0019] In step S3, the method for global matching analysis includes the following steps: and Preprocess them respectively to get new word sets, which are recorded as and , the preprocessing includes word segmentation, removal of stop words and filtering of duplicates, and then it is calculated as follows and The global matching degree of:

[0020] ,

[0021] in, Indicates that it appears at the same time and The total number of words in express The number of words in .

[0022] In step S3, the method for analyzing behavioral data includes the following steps: Medium Calculation Corresponding search behavior The quantity is and prompt click behavior The quantity is , and then calculate it as follows Corresponding user voting points:

[0023] ,

[0024] in, Is a constant with a value between 0 and 1.

[0025] In step S3, the method for analyzing user interests includes the following steps: obtaining the most recent P query statements of the current user, and obtaining the vectors of each statement processed by word2vec and summing them up, corresponding to the vector , search queries to be prompted The corresponding word2vec processed vector is recorded as vector , and then calculate the user pair as follows Level of interest:

[0026] .

[0027] In step S3, the context analysis method includes the following steps: obtaining the most recent P query statements of the current user, and preprocessing each statement to obtain a new word set, denoted as , treat the prompted search query Preprocessing obtains a word set, denoted as , the sentence preprocessing includes word segmentation, stop word removal, merging and duplicate filtering, and then it is calculated as follows Contextual relevance:

[0028] ,

[0029] in Indicates that it appears at the same time and The total number of words in express The number of words in .

[0030] In step S3, the timeliness analysis method includes the following steps: get The corresponding latest timestamp , then, calculated as follows The time factor of:

[0031] .

[0032] In step S4, the fusion calculation The overall quality score is calculated as follows:

[0033] ,

[0034] in , , , , , are all constants, with values ​​between 0 and 1, and , The value is positively correlated with the amount of historical search data and the regularity of historical searches. The more searches a user makes, the more regular the searches are. The larger the value; The value size is positively correlated with the concentration of search query intent. The more the search query intent is concentrated on some core words, The larger the value; The value is positively correlated with the amount of clicks on the auxiliary input results by users. The more clicks a user makes on the intelligent auxiliary input results, the more likely it is to be clicked. The larger the value; The value size is positively correlated with the amount of user behavior data and the accuracy of user interest mining. The larger the behavior data, the higher the confidence of the user interest measurement value. The larger the value; The value is positively correlated with the intensity of the user's current search. When the user has intensive search needs within 10 minutes of the current search, The larger the value; The value is positively correlated with the timeliness of search results. When the newer the search results are, the higher the ranking will be. The larger the value.

[0035] According to the above technical scheme, the present invention has the following advantages: the present invention takes the title, text and other information of the content to be searched and the user's historical search behavior as input, and combines the previous maximum hit similarity, global hit similarity, the number of historical search results, the number of historical associated selections of results, the user's personal historical search record analysis, the user's current search context, the effectiveness of the results and other aspects for comprehensive consideration for sorting, and ultimately can effectively improve the coverage, relevance, timeliness, personalization, etc. of the prompt results, greatly improve the user experience, and improve the efficiency and satisfaction of user searches. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a flowchart of a method for intelligent auxiliary search input of the present invention. DETAILED DESCRIPTION

[0037] The present invention is further described in detail below with reference to the accompanying drawings and specific examples, in order to provide a clearer understanding of its structural type and usage, but this shall not be used to limit the protection scope of the patent of the present invention.

[0038] like Figure 1 As shown, take the most recent period Including search behavior and prompt click behavior Behavioral dataset ,based on middle Behavioral data, obtain the user query statement data set after filtering ,in Length is , middle The quantity is , The quantity is .Will All are stored in the Trie tree, recorded as .Will All word vectors are generated using word2vec and stored in Faiss, denoted as Import all the content information to be searched, such as title, tag, description, text and other field information, into Elasticsearch and record it as The current time is timestamped as For search queries to be suggested , whose string length is , do the following:

[0039] 1. Multi-way recall candidate set:

[0040] Predecessor matching recall: For According to the previous order Perform the longest match query, then sort the returned results from high to low according to the tree depth, and return the top N results.

[0041] Word embedding recall: Use word2vec to generate word vectors The top N most similar results are returned.

[0042] Text Semantic Recall: Call The suggest function returns the top N results.

[0043] 2. Result merging

[0044] The recall results of the three methods in 1 are merged and repeated filtering is performed to obtain the final M candidate results.

[0045] 3. Static analysis and dynamic analysis

[0046] For the M candidate result sets after merging in 2, The results are , static analysis and dynamic analysis are performed respectively.

[0047] 3.1 Static Analysis

[0048] It mainly analyzes the text relevance of candidate results.

[0049] 1) Preamble matching analysis

[0050] First, match one by one from front to back to get the length of the longest matching string , express Length, express and, The smallest one, and Respectively and In the The value at the position.

[0051] pseudocode:

[0052]

[0053] Then, we calculate and Pre-order matching degree of:

[0054] ,

[0055] 2) Global matching analysis

[0056] First of all, and After preprocessing, including word segmentation, stop word removal, and duplicate filtering, a new word set is obtained, denoted as and , and then calculated as follows and The global matching degree of:

[0057] ,

[0058] in Indicates that it appears at the same time and The total number of words in express The number of words in .

[0059] 3.2 Dynamic Analysis,

[0060] Mainly, candidate results are comprehensively analyzed in combination with user historical behavior data, which is divided into four dimensions: behavior data, user interests, context, and timeliness.

[0061] 1) Behavioral data analysis

[0062] First, in Medium Calculation Corresponding The quantity is and The quantity is .

[0063] Then calculate it as follows Corresponding user voting points:

[0064] ,

[0065] in, It is a constant with a value between 0 and 1, which will be adjusted according to the actual situation.

[0066] 2) User interest analysis

[0067] First, obtain the current user’s most recent P query statements, and obtain the word2vec processed vectors of each statement and sum them up, corresponding to the vector , The corresponding word2vec processed vector is recorded as vector .

[0068] Then, the user pair is calculated as follows Level of interest:

[0069] .

[0070] 3) Contextual Analysis

[0071] First, obtain the most recent P query statements of the current user, preprocess each statement, including word segmentation, removal of stop words, merging, filtering, etc., to obtain a new word set, denoted as .right Preprocessing obtains a word set, denoted as .

[0072] Then, we calculate Contextual relevance:

[0073] ,

[0074] in Indicates that it appears at the same time and The total number of words in express The number of words in .

[0075] 4) Timeliness analysis

[0076] First, from get The corresponding latest timestamp Then, we calculate The time factor of:

[0077] .

[0078] 3.3 Fusion Sorting

[0079] After the above static analysis and dynamic analysis, the fusion calculation is performed in the following way Overall quality score:

[0080]

[0081] in , , , , , are all constants, with values ​​between 0 and 1, and , will be configured according to the actual situation. The value is closely related to the amount of historical search data and the regularity of historical searches. The more searches a user makes, the more regular the searches are. The larger the value, the higher the It mainly depends on whether the search query intent is concentrated on some core words. The more the search intent is concentrated on some core words, The larger the value, the higher the It is mainly related to the amount of clicks on the auxiliary input results by users. The more clicks users make on the intelligent auxiliary input results, The larger the value, the It has a positive relationship with the amount of user behavior data and the accuracy of user interest mining. The larger the behavior data, the higher the confidence of the user interest measurement value. The larger the value, the It is strongly related to the intensity of the user's current search. When the user has intensive search needs within 10 minutes of the current search, The larger the value, the It mainly affects the timeliness of search results. When you want newer search results to be ranked higher, The value can be adjusted to be larger.

[0082] According to the calculation results of 3.3, the M candidate result sets are then sorted from large to small according to the quality scores, and the top K results are taken and returned as the final result of the intelligent auxiliary input.

[0083] Undoubtedly, the present invention has other similar structural compositions and usage modes besides the above-mentioned embodiments. In summary, the present invention also includes other obvious changes and substitutions for those skilled in the art.

Claims

1. A method for intelligent auxiliary search input, characterized in that: The method comprises the following steps: S1. Data pre-processing: take the most recent period Including search behavior and prompt click behavior Behavioral dataset , based on the behavioral dataset Search behavior Data, obtain the user query statement dataset after filtering , the user query statement dataset The length of , the behavior dataset Search behavior The number of , the prompt click behavior The number of , the user query statement dataset All are stored in the tree Trie, recorded as , the user query statement dataset All word vectors are generated using the word2vec model and stored in Faiss and recorded as , import all the content information to be searched into Elasticsearch and record it as , the content information includes field information, the field information includes title, tag, description and text; the timestamp of the current time is recorded as , for the search query to be prompted , the length of the string is recorded as ; S2. Multi-way recall candidate set: search query to treat prompts Perform multi-way recall, which includes pre-order matching recall, word embedding recall and text semantic recall, and then merge the results of the pre-order matching recall, word embedding recall and text semantic recall, and perform duplicate filtering to obtain the final M candidate results; S3. Static analysis and dynamic analysis: For the M candidate result sets obtained in step S2, The results are recorded as , respectively, static analysis and dynamic analysis are performed, the static analysis includes pre-order matching analysis and global matching analysis, the static analysis is used to analyze the text relevance of candidate results, the dynamic analysis includes behavior data analysis, user interest analysis, context analysis and timeliness analysis, the dynamic analysis is used to conduct a comprehensive analysis of candidate results combined with user historical behavior data; S4. Fusion sorting: Based on the static analysis and dynamic analysis results in step S3, fusion calculation The overall quality score of the task is calculated, and then the M candidate result sets are sorted from large to small according to the quality score, and the top K results are taken and returned as the final result of the intelligent auxiliary input.

2. The method of intelligent auxiliary search input according to claim 1, characterized in that: In step S2, the method for recalling the preceding order matches includes: According to the previous order Perform the longest match query, then sort the returned results from high to low according to the tree depth, and return the top N results; the word embedding recall method includes: Use word2vec to generate word vectors The most similar TopN results are returned in the query; the text semantic recall method includes: Call The suggest function returns the top N results.

3. The method of intelligent auxiliary search input according to claim 1, characterized in that: In step S3, the method for preamble matching analysis includes the following steps: matching the M candidate result sets obtained in step S2 one by one from the front to the back to obtain the length of the longest matching string , express Length, express and The smallest one, and Respectively and the search query to be suggested In the The value at the position is calculated as follows: and Pre-order matching degree: 。 4. The method of intelligent auxiliary search input according to claim 1, characterized in that: In step S3, the method for global matching analysis includes the following steps: and Preprocess them respectively to get new word sets, which are recorded as and , the preprocessing includes word segmentation, removal of stop words and filtering of duplicates, and then it is calculated as follows and The global matching degree of: , in, Indicates that it appears at the same time and The total number of words in express The number of words in .

5. The method of intelligent auxiliary search input according to claim 1, characterized in that: In step S3, the method for analyzing behavioral data includes the following steps: Medium Calculation Corresponding search behavior The quantity is and prompt click behavior The quantity is , and then calculate it as follows Corresponding user voting points: , in, Is a constant with a value between 0 and 1.

6. The method of intelligent auxiliary search input according to claim 1, characterized in that: In step S3, the method for analyzing user interests includes the following steps: obtaining the most recent P query statements of the current user, and obtaining the vectors of each statement processed by word2vec and summing them up, corresponding to the vector , search queries to be prompted The corresponding word2vec processed vector is recorded as vector , and then calculate the user pair as follows Level of interest: 。 7. The method of intelligent auxiliary search input according to claim 1, characterized in that: In step S3, the context analysis method includes the following steps: obtaining the most recent P query statements of the current user, and preprocessing each statement to obtain a new word set, recorded as , treat the prompted search query Preprocessing obtains a word set, denoted as , the sentence preprocessing includes word segmentation, stop word removal, merging and duplicate filtering, and then it is calculated as follows Contextual relevance: , in Indicates that it appears at the same time and The total number of words in express The number of words in .

8. The method of intelligent auxiliary search input according to claim 1, characterized in that: In step S3, the timeliness analysis method includes the following steps: get The corresponding latest timestamp , then, calculated as follows The time factor of: 。 9. The method of intelligent auxiliary search input according to claim 1, characterized in that: In step S4, the fusion calculation The overall quality score is calculated as follows: , in , , , , , are all constants, with values ​​between 0 and 1, and , The value is positively correlated with the amount of historical search data and the regularity of historical searches. The more searches a user makes, the more regular the searches are. The larger the value; The value size is positively correlated with the concentration of search query intent. The more the search query intent is concentrated on some core words, The larger the value; The value is positively correlated with the amount of clicks on the auxiliary input results by users. The more clicks a user makes on the intelligent auxiliary input results, the more likely it is to be clicked. The larger the value; The value size is positively correlated with the amount of user behavior data and the accuracy of user interest mining. The larger the behavior data, the higher the confidence of the user interest measurement value. The larger the value; The value is positively correlated with the intensity of the user's current search. When the user has intensive search needs within 10 minutes of the current search, The larger the value; The value is positively correlated with the timeliness of search results. When the newer the search results are, the higher the ranking will be. The larger the value.