KR20170042080A - Method and apparatus for providing personalized search results - Google Patents

Abstract

According to an embodiment of the present invention, a method for providing a personalized search result comprises the steps of: extracting a keyword from a document which is opened by a user; calculating a weighted value by considering a frequency of the keyword in the document; generating user model data on the user by mapping the weighted value to the keyword; and receiving a query from a terminal of the user, and aligning a search result on the query by using the user model data.

Description

TECHNICAL FIELD The present invention relates to a personalized search result providing method and apparatus,

The present invention relates to a method and apparatus for providing personalized search results. More particularly, the present invention relates to a method of digitizing a user's response to a search result and providing a user with a document that is more interesting, and an apparatus for performing the method.

We often use search to get only the desired information among various information. At this time, the user performs a search by inputting a query word for searching for desired information. However, since most of the query words input by the user are only one or two words, the quality of the search often deteriorates. That is, when the user inputs an optimal query for finding the desired information, it is a prime number.

To solve this problem, search engines sort their search results by their own algorithms, and provide services such as similarity search queries or suggestions. However, most of these services are provided in correspondence with query terms, and even if they are different users, they provide the same search results, the same similar query, and the same query when the same query is input.

The information to be obtained through the query may differ depending on the user even if the query is the same. An extreme example of a homophonic word or a variant, for example, if you enter the same query term "night", someone might be searching for information about the night, which means when the sun was darkened, May be to retrieve information about the chestnut, which means the fruit of the chestnut.

A problem to be solved by the present invention is to provide a method and apparatus for providing personalized search results.

It is another object of the present invention to provide a method and apparatus for providing an optimal combination of algorithms to provide personalized search results.

The technical problems of the present invention are not limited to the above-mentioned technical problems, and other technical problems which are not mentioned can be clearly understood by those skilled in the art from the following description.

According to an aspect of the present invention, there is provided a personalized search result providing method including: extracting a keyword of a document from a document viewed by a user; Calculating a weight, mapping the weight to the keyword, generating user model data for the user, receiving a query word from the terminal of the user, and using the user model data, The search results may be sorted.

In one embodiment, the step of calculating the weights may include a step of correcting the weights in consideration of the distribution of the keywords and the weights.

In another embodiment, calculating the weighting may include correcting the weighting, taking into account information about the time the user has viewed the document.

In yet another embodiment, the step of sorting the search results for the query term includes extracting a search result keyword from each document belonging to the search result, comparing the search result keyword with the keyword, Calculating the importance of each document based on the weights, and sorting the search results for the query terms using the importance of each document.

In another embodiment, the step of sorting the search results for the query term may include the step of feeding back the user's response to the sorted search results and updating the user model data.

In yet another embodiment, generating the user model data for the user may comprise clustering a plurality of users based on user model data to construct a cluster.

According to another aspect of the present invention, there is provided a personalized search result providing apparatus including a keyword extracting unit for extracting a keyword of a document from a document viewed by a user, A user model data generation unit for generating a user model data for the user by mapping the weights to the keyword and a query input unit for receiving a query word from the user terminal, And a search result personalization unit for sorting search results for the query term using the search result personalization unit.

In one embodiment, the weight calculation unit may include a weight correction unit that corrects the weight by considering the distribution of the keyword and the weight.

In another embodiment, the search result personalization unit may further include a search result keyword extracting unit for extracting a search result keyword from each document belonging to the search result, a comparison unit for comparing the search result keyword with the keyword, A document importance calculation unit for calculating the importance of each document, and a search result sorting unit for sorting search results for the query term using the importance of each document.

In another embodiment, the search result personalization unit may include a user model data updating unit that updates the user model data by feeding back the response of the user to the sorted search result.

According to another aspect of the present invention, there is provided a personalized search result providing apparatus including a network interface, at least one processor, a memory for loading a computer program executed by the processor, Model data, and algorithm combination data. Here, the computer program may further include a weight calculation operation for calculating a weight, in consideration of a keyword extraction operation for extracting a keyword of the document and a frequency of the keyword in the document, A user model data generation operation for generating the user model data for the user by mapping the weights to the keyword and a plurality of users by clustering a plurality of users based on the user model data, An algorithm combination learning operation for mechanically learning a response of a user belonging to the cluster to an algorithm, a weight correction algorithm and an importance computation algorithm to generate the algorithm combination data of the cluster, And a search result personalization operation for receiving a query word and using the user model data to sort search results for the query term.

According to the present invention as described above, the user's reaction to the search result is numerically quantified and collected as user model data, so that when the user performs the next search, Can be offered first.

In addition, a plurality of algorithms that can be used in this process can be used to provide individualized search results by setting the algorithms for each user through machine learning.

The effects of the present invention are not limited to the effects mentioned above, and other effects not mentioned can be clearly understood to those of ordinary skill in the art from the following description.

1 is a conceptual diagram for explaining an existing search result providing method.
2 to 3 are conceptual diagrams illustrating a method of providing personalized search results according to an embodiment of the present invention.
4 is a flowchart of a personalized search result providing method according to an exemplary embodiment of the present invention.
5 is a conceptual diagram illustrating a step of extracting a keyword from a user's interest document and calculating a weight according to an embodiment of the present invention.
6 to 7 are conceptual diagrams illustrating a step of calculating a weight based on time information related to the browsing of a document according to an embodiment of the present invention.
8 to 9 are conceptual diagrams illustrating a step of personalizing search results using user model data according to an embodiment of the present invention.
Figure 10 shows an algorithm that can be used in each step of the method for providing personalized search results.
11 shows a result of testing the importance of a specific document according to each algorithm combination.
12 is a flowchart illustrating a method of providing a personalized search result by machine learning of an algorithm combination according to an embodiment of the present invention.
FIG. 13 shows a result of testing the response of the user according to each algorithm combination.
14 is a hardware configuration diagram of a personalized search result providing apparatus according to an embodiment of the present invention.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS The advantages and features of the present invention and the manner of achieving them will become apparent with reference to the embodiments described in detail below with reference to the accompanying drawings. The present invention may, however, be embodied in many different forms and should not be construed as being limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art. Is provided to fully convey the scope of the invention to those skilled in the art, and the invention is only defined by the scope of the claims. Like reference numerals refer to like elements throughout the specification.

Unless defined otherwise, all terms (including technical and scientific terms) used herein may be used in a sense commonly understood by one of ordinary skill in the art to which this invention belongs. Also, commonly used predefined terms are not ideally or excessively interpreted unless explicitly defined otherwise. The terminology used herein is for the purpose of illustrating embodiments and is not intended to be limiting of the present invention. In the present specification, the singular form includes plural forms unless otherwise specified in the specification.

It is noted that the terms "comprises" and / or "comprising" used in the specification are intended to be inclusive in a manner similar to the components, steps, operations, and / Or additions.

Hereinafter, the search and search results used in the present invention are targeted to documents created mainly in terms of keywords and texts. That is, it does not target such areas as image search or video search. Here, the document prepared in the text-oriented manner may be a general web document, and in some cases, it may be a document such as a mail in a groupware, a publication such as a bulletin board, or a document written in a document program such as Hangul, word or PowerPoint.

Hereinafter, the present invention will be described in more detail with reference to the accompanying drawings.

1 is a conceptual diagram for explaining an existing search result providing method.

Referring to FIG. 1, when different user A 101a and user B 101b perform a search with the same query term (a query step), the search engine 111 searches the corresponding document 113, And searching result (② searching step). The search engine 111 performs sorting of the inquired search result 115 by its own algorithm ((3) ranking step). In this case, the sorting algorithm can use information such as the frequency of occurrence of the keyword in the search target document, for example. Or information such as when the document to be searched for was created or modified could be used in the sorting algorithm. The sorted search results 115 are provided to the user A 101a and the user B 101b (step 4).

However, since the user A 101a and the user B 101b input the same query, the search result 115 to be searched at that time must be the same. That is, since the search result 115 reflecting the taste or the tendency of the user is not provided, satisfaction with the search result 115 depends on the user.

For example, the user A 101a entered "night" because he wanted to search for a dark night, user B 101b entered "night" because he wanted to search for chestnut berries, ), If the information about the chestnut fruit is full, the user A (101a) must check each one of the search results 115 by several pages in order to search for the desired information, or perform a search again after correcting the query term something to do.

If it is possible to provide different personalized search results 115 in consideration of the taste or the tendency of the user A 101a and the user B 101b at this time, It will be less.

2 to 3 are conceptual diagrams illustrating a method of providing personalized search results according to an embodiment of the present invention.

Referring to FIG. 2, the search result 115 is personalized and sorted according to each user by using user model data (personalized ranking step). Here, the user model data 117 is data generated by reflecting the taste or the tendency of each user, and is data obtained by, for example, data on the types and characteristics of documents that the user may like. That is, data that models a user's taste or disposition.

Of course, the user model data 117 may be data received directly from a user through a user preference menu or the like, but may be data generated by monitoring a user's reaction to a search result. That is, the search results are sorted based on documents that are similar to the documents that the user has browsed in the next search based on which documents the user has focused on in the search results.

Even if a query word "night" is input in the same manner, the user A 101a is harmed and provides mainly dark dark night related documents 119a. In addition, the user B 101b provides documents related to chestnut berries night (119b). That is, the user model data 117 may be generated for each user and used to provide personalized search results 119a and 119b.

Referring to FIG. 3, a process of generating user model data 117 can be seen. That is, the personalized search results 119a and 119b are used to revise the user model data 117 by using which documents each user mainly browses (step (5) feedback). As the user performs the search, documents similar to the documents viewed by the user can be provided as the search results, so that the quality of the search results can be enhanced.

1 to 3, a conventional search result providing method and a personalized search result providing method according to an embodiment of the present invention have been conceptually described. Hereinafter, a personalized search result providing method according to an embodiment of the present invention will be described in detail with reference to FIG. 4 through FIG.

4 is a flowchart of a personalized search result providing method according to an exemplary embodiment of the present invention.

Referring to FIG. 4, keywords are extracted from the documents of interest of each user (S1100). Here, the user's interest document refers to the documents that the user has responded to in the search result as described above. That is, the information that a specific user desires to obtain through a search is inferred from the documents read by the specific user.

In a search result obtained by inputting such a query, a user who browses these documents mainly is to digitize information on documents of interest in order to give priority to documents similar to the documents in the next search. In order to quantify the user's interest documents, various information about the document of interest can be used. However, we focus on the words and keywords in the document to characterize the document.

After extracting the keyword from the document of interest, each keyword weight can be calculated (S1200). When calculating the keyword weight, the weight can be calculated on the basis of how often the keyword appears in the interest document and how often it appears in other documents. Furthermore, the weight can be further corrected according to the distribution of the keyword and the weight.

After extracting the keyword from the user's interest document and calculating the weight of each keyword, the keyword weight is summed and stored as user model data for each user (S1300). In step S1400, even if the user performs a search by inputting another query, the user can personalize the retrieved documents based on the user's response to the previously retrieved documents.

Since the documents read for each user are different from each other, even if the search is performed by inputting the same query words, the user model data are different from each other and the search results are also provided differently. Through this, the user can reduce the trouble of searching the search results one by one in order to find the information desired by the user.

For example, if a user A (101a) who entered "night" to search for dark nights due to sunshine enters a query term "travel" after that, he / she may prioritize the nighttime travel destination If the user B (101b) who entered "night" to search for chestnut berries, then enters the query "travel," he or she can search for a destination where night production or night picking experience programs can be performed You will be able to provide it first.

5 is a conceptual diagram illustrating a step of extracting a keyword from a user's interest document and calculating a weight according to an embodiment of the present invention.

First, keywords are extracted from the documents that the user has viewed. Here, the keyword can be extracted from the title, author, and text of the document. If the document is used in groupware, information related to the author's department or position system can be extracted as keywords. The extracted keywords are weighted based on the frequency in the document and the frequency in other documents.

The calculated weight can be used as the user model data as it is, but the weight can be corrected if necessary. Various statistical methods can be applied to the weight correction. The weight correction operation may be performed in several steps. That is, weight correction may be performed several times by various correction algorithms. After the weights are calculated and corrected for each document, the keyword weights of the respective documents are summed up to generate and store the user model data of the corresponding user.

Referring to FIG. 5, a process of extracting a keyword from a specific document 121 among the documents viewed by a user and calculating a weight value can be seen. You can extract keywords such as "business", "report data", "Hong Gil Dong", "A business division" by extracting keywords from the document created by "Hong Gil Dong" However, here, keywords are extracted from the information related to the title and author due to limitations on the paper, and keywords can be extracted from various text information related to the text of the document or other documents.

The extracted keyword 123 primarily calculates the weight 125 using the frequency in the document and the frequency in the other document. Referring to FIG. 5, it can be seen that "business" has a weight of 0.12, "report data" has a weight of 1.32, "Hong Kil Dong" has a weight of 2.12, and "A business unit" has a weight of 4.12. When the primary task of calculating the weight is completed, the weight can be corrected appropriately.

The step of correcting the weight can be performed several times as needed. For example, if the deviation of the weight is too severe, it can be corrected. In the example of FIG. 5, it can be seen that the respective weights are corrected to 0.02, 0.23, 0.11 and 0.12 through the correction.

When the corrected weights 127 in the document are added together, the user model data is completed. In the example of FIG. 5, keywords are extracted from other documents viewed by the user in addition to the specific document 121 viewed by the user, and the weights are calculated. It is possible to add the weights of keywords "business", "deadline", "application", "Kim Ji-young" extracted from other documents in addition to "business", "report data", " The user model data 129 can be generated by summing the keywords of the various documents and the keyword weights. Here, a concrete method of calculating the weights will be described in detail later with reference to FIG. 10 to FIG.

6 to 7 are conceptual diagrams illustrating a step of calculating a weight based on time information related to the browsing of a document according to an embodiment of the present invention.

When calculating the weights by monitoring the user's reaction to the search results, calculating the weights based on whether or not the documents are simply browsed may not reflect the weight among the documents. In general, the user must check the documents provided as a result of the search to see if there is desired information. In other words, when a user browses a document provided as a search result, it is necessary to consider the time at which each document is viewed. Therefore, it is necessary to calculate the weight by considering the document in which the user has the desired information among the search results.

In addition, it is necessary for the user to input the query phrase and reflect the severity of the documents retrieved from the retrieved results and the documents retrieved from the retrieved results by inputting the query phrase recently. That is, it is necessary that the documents that have been searched and read recently have a larger weight than when the weight is calculated. To do this, the user needs to monitor the time the document is viewed in the search results. In other words, there is a need for a process of correcting the weights based on the cumulative time of the user viewing the document and the latest viewing date and time.

Referring to FIG. 6, the user can input the query term "settlement" in the search window 131 to see the search result 133 that performed the search. The search results retrieved with the query "settlement" entered by the user in the groupware may include documents such as "development definition .pptx", "screen design .pptx", "settlement documents .xls" The user refers to the title or summary (not shown) of each document in the search result to determine which document has the desired information. Then, the user browses some documents in the search result, . At this time, when the user starts to browse the document and monitors the time at which the reading is finished, the time consumed by the user to browse the document and the latest browsing date and time of the document can be obtained.

In the example of Fig. 6, it takes 10 seconds to browse the " development definition .pptx ", 8 seconds to view "screen design .pptx ", 6 minutes to view" It can be seen that the time consumed in seconds is consumed. Perhaps the information the user wanted to retrieve from the search results is likely to be in the "settlement documents .xls". If the user's viewing time for each document is monitored and stored separately in the log data 135, not only the order of reading after each document but also the cumulative viewing time can be obtained. In other words, you can see which documents were most recently viewed and how much time was spent browsing them.

Referring to FIG. 7, it can be seen that the weight is calculated on the basis of the time consumed in browsing the document obtained in FIG. 6 and the time when the document was recently viewed. In FIG. 5, the keyword weighting included in each document is calculated. In FIG. 7, the time weighting for each document is calculated.

If we do not take into account the time weight of each document, we will simply generate user model data by simply summing the keyword weights of the documents that the user has viewed. 7, the keyword weights of the document 1 141a and the document 2 141b are calculated, and then the sum of the keyword weights is added to the user model data 151 without considering the time weight. In this case, "business", which is a common keyword, has a weight of 0.02 + 0.04 = 0.06, and the other keywords do not have a common keyword. "Report data" is 0.23, User model data 151 having a weighting value of 0.11, 0.12, 0.12, a deadline of 0.36, a reference of 0.17, and a weight of 0.24 were generated.

However, as described in FIG. 6, it is necessary to calibrate a document that is older than the scanned document so as to have a lower weight as the document consumes less time for reading. In contrast to the keyword weighting in FIG. 5, this is called time weighting.

The weights 143a and 143b based on the browsing time have higher weights as the browsing time is longer. The document 1 141a has a time weight of 0.5 (143a) because of a large viewing time, and the viewing time of the document 2 (141b) is small and has a time weight of 0.2 (143b). At this time, the time weight according to the viewing time can be converged to a specific value as the viewing time increases, if necessary. In the example of FIG. 7, the time weight according to the viewing time is set to converge to 1.

In addition to the weights 143a and 143b based on the browse time, it is possible to have a higher weight for the document that has been recently viewed. The weights 145a and 145b based on the recent browsing are as follows. In the case of the document 1 141a, the oldest document has a weight of 0.1. In the case of the document 2 141b, the most recently viewed document has a weight of 0.6 Can be seen. At this time, if necessary, the most recently viewed document can have a specific value. In the example of FIG. 7, the most recently viewed document is set to have a value of 1. The longer the viewing date and time, the smaller the time weight and eventually converges to zero.

It can be seen that the user model data 157 considering the time weight is different from the user model data 151 generated without consideration of the time weight. That is, when the time weight is taken into consideration, the keyword weight of each document can be corrected by multiplying the time weight based on the browse time first and the time weight based on the second most recent browse. In the case of the document 1 141a, the keyword weight 147a is corrected by 0.05 (= 0.5 * 0.1) in consideration of the time weight. In the case of the document 2 141b, * 0.6) of the new keyword weight 147b. Now, by summing up these two, new user model data 157 reflecting the time weight can be obtained, and if the search results are sorted using the new user model data 157, a document suited to the user's interest can be provided first.

8 to 9 are conceptual diagrams illustrating a step of personalizing search results using user model data according to an embodiment of the present invention.

5 to 7, the process of calculating the weight by considering the keyword weight and the time weight in order to generate the user model data has been described. The user model data thus generated can be used to sort search results. That is, even if the same keyword is input, the search results provided to the user are different because of different user model data. For example, the user A 101a may be harmed to give priority to a document related to a dark night, and the user B 101b may preferentially provide documents related to a chestnut fruit night.

Referring to FIG. 8, an algorithm for determining the order for sorting each document of the retrieved result is introduced. There are three algorithms that can be used at this time. One is a matching algorithm, the other is a unique matching algorithm, and the last is a language model algorithm.

The matching algorithm extracts keywords from the retrieved documents and calculates the importance of the documents using the frequency of the keywords in the retrieved documents and the weights of the keywords in the user model data.

In the example of FIG. 8, when the matching algorithm is applied to the search result document 1 163, the weight of the user model data 161 of the "business" appearing twice is 0.1, and the user model data of the "report data" The weight in the user model data 161 of the " A business unit ", which appeared twice, and the weight in the user model data 161 of the "meeting " (2 * 0.1) + (1 * 0.2) + (2 * 0.4) + (1 * 0) = 1.2 The significance of 1.2 can be obtained by the following equation.

The unique matching algorithm is similar to the matching algorithm, but differs in that it does not consider the frequency of keywords in the retrieved documents. That is, if the keyword appears in the retrieved document, it is a method of calculating the importance of the document without making a difference whether it appears once or several times.

In the example of FIG. 8, when a unique matching algorithm is applied to the search result document 163, the keyword "business", "report data", "business A", "conference" The significance of 0.7 can be obtained by adding the weights 0.1, 0.2, 0.4, and 0 in the user model data 161 to the formula 0.1 + 0.2 + 0.4 + 0 = 0.7.

Compared to the two previous algorithms, the language model algorithm takes a slightly different approach. Weights of each keyword are summed up to obtain a weighted sum (w _total ), and a log of the weighted sum is calculated to calculate the importance. In this process, 1 is added to each keyword weight to obtain logarithm. Using a language model algorithm, unlike a matching algorithm or a unique matching algorithm, a keyword having a weight of 0 in the user model data 161 may affect the importance.

In the example of FIG. 8, when the language model algorithm is applied to the search result document 163, 2 * log ((0.1 + 1) /1.5) + 1 * log ( (0.4 + 1) /1.5) + 1 * log ((0 + 1) / 1.5) = 4.8.

As shown in FIG. 8, even if importance is calculated using the same user model data 161 for the same search result document 163, it can be seen that the importance of the document can be changed according to the applied algorithm. Let's take a closer look at each algorithm in conjunction with Figure 9.

In FIG. 9, the process of calculating the importance of the document by the respective algorithms with respect to the search result document 2 165 shown in FIG. 8 can be seen.

The process of calculating the importance of the document for the search result document 1 (163) and the search result document 2 (165) will be described as follows.

keyword matching Unique matching Language model business 2 * 0.1 = 0.2 0.1 2 * log ((0.1 + 1) /1.5) = 1.467 Reporting 1 * 0.2 = 0.2 0.2 1 * log ((0.2 + 1) / 1.5) = 0.8 A Division 2 * 0.4 = 0.8 0.4 2 * log ((0.4 + 1) /1.5) = 1.867 conference 1 * 0 = 0 0 1 * log ((0 + 1) / 1.5) = 0.667 Sum 1.2 0.7 4.8

keyword matching Unique matching Language model Hong Gil Dong 2 * 0.3 = 0.6 0.3 2 * log ((0.3 + 1) /1.5) = 1.733 A Division 1 * 0.4 = 0.4 0.4 1 * log ((0.4 + 1) /1.5) = 0.933 Kim Ji Young 1 * 0.5 = 0.5 0.5 1 * log ((0.5 + 1) / 1.5) = 1 Sum 1.5 1.2 3.67

Table 1 summarizes the importance calculation for the search result document 1 (163), and Table 2 lists the importance calculation for the search result document 2 (165). Referring to Table 1 and Table 2, when the importance is calculated by the matching algorithm or the unique matching algorithm, the importance 1.5 or 1.2 of the search result document 2 (165) is compared with the importance 1.2 or 0.7 of the search result document 1 (163) The search result document 2 165 can be treated as a more important document. However, if the importance level is calculated by the language model algorithm, the importance degree 3.67 of the search result document 2 165 is lower than the importance degree 4.8 of the search result document 1 163, and the search result document 1 163 is more important It can be treated as a document.

As in the example of FIG. 8, the sort order of the search results may be reversed according to the algorithm applied to calculate the importance of the document. The step of calculating the importance of the document of Fig. 8 corresponds to the step of personalizing the search result (S1400) in the flowchart of Fig. In addition to the step of personalizing the search result of FIG. 8 (S1400), the sorting order of the search results may vary depending on the algorithm applied in the remaining steps (S1100, S1200, S1300). This will be described with reference to FIGS. 10 to 13. FIG.

FIGS. 10 to 13 are conceptual diagrams for explaining a process of learning a combination of algorithms used in personalizing search results according to an embodiment of the present invention.

Figure 10 shows an algorithm that can be used in each step of the method for providing personalized search results. First, algorithms such as n-gram, context, stemming, part-of-speech tagging, and NER (named entity recognition) may be used in step S1100 of extracting keywords from the document of interest. For each algorithm, the n-gram algorithm includes uni-gram, bi-gram, and tri-gram, extracting uni-gram by one letter, bi-gram by two letters, and tri- gram by three letters Method. The n-gram algorithm is applied to the example of "Go to school".

Original She goes to school. uni-gram Iron, water, autumn, school, bridge bi-gram Withdraw, go to school, school tri-gram The withdrawal goes to school.

Next, the context algorithm is a method of extracting a name, a work name, a department name, and a date from a text. Table 4 below shows the application of the context algorithm to the sample sentence "This is the data released by accounting department Kim Chul Su."

Original This accounting data is published by Chul Soo Kim, accounting department. keyword Name: Chul Soo Kim, Department: Accounting Department, Title: Accounting

Next, the part-of-speech tagging algorithm is a method of tagging parts of speech. Apply the part-of-speech tagging algorithm to the example sentence "Go to school."

Original She goes to school. keyword Withdrawal / noun, verb / investigation, school / noun, in / survey, verb / verb

Next, the stemming algorithm is a method of extracting one common word by a specific rule for English text. A simple example of this is shown in Table 6 below.

Original engineering, engineer, engine keyword engine

Next, the NER (Named Entity Recognition) algorithm is a method of extracting the date, name, and place name in a manner similar to the context algorithm by object name recognition. Applying the NER algorithm to the example sentence "Go to school" is shown in Table 7 below.

Original She goes to school. keyword Name: Chul Su, Place: School

Other keywords may be extracted according to the applied keyword extraction algorithm, and the extracted keywords may be classified according to parts of speech or personality. Keyword extraction algorithms can be affected by the nature of the document. Although not shown in FIG. 10, for example, an algorithm capable of extracting affirmative / negative keywords in a search that analyzes and provides a review of a movie or a music will be useful, and an algorithm for extracting keywords When searching gossip news, the algorithms for extracting keywords will be different.

In step S1200 of calculating the weight of the keyword in FIG. 5, it is mentioned that the weight of each keyword can be calculated using the frequency of the keyword. The TF-IDF algorithm can be applied in a general way. In this paper, we propose a TF-IDF algorithm that can be applied to a large number of words in a document, The more the word, the lower the weight is reflected. In other words, the more words that appear only in the document, the higher the weight is. The TF-IDF is a concrete expression that can be obtained by dividing the frequency of the keyword in the document by the frequency in other documents. However, here the frequency in other documents can be calculated by taking a log of inverses.

Next, the TF-IDF normalized algorithm is a method of calculating the weight of a keyword by performing normalization on the document length unlike the general TF-IDF algorithm. That is, the longer the length of the document, the higher the frequency with which the keyword appears. Therefore, the weight of the keyword is calculated in consideration of this. In other words, even if the keyword has the same frequency, the keyword is evaluated with a higher weight for a document having a shorter length. This is an algorithm used to weaken the weight of a word appearing in a document having a long document length because it can always be high.

Next, the BM25 algorithm is an algorithm that evaluates weights differently by adjusting some parameters similar to the TF-IDF normalized method. So far, the algorithm for calculating the weight of a keyword can be expressed by a simple formula as shown in Table 8 below.

TF-IDF

TF-IDF (norm)

BM25

After the weights of the keywords are calculated, the weights may be further corrected as necessary. The singular value decomposition (SVD) algorithm decomposes U, which represents a term component, sigma, which represents a singular value, and V (transpose matrix), which represents a document component, from a matrix of terms and documents, It is a method used for reduction.

Next, the gaussian mixture model (GMM) algorithm is a mixed model of the Gaussian distribution model. The closer to the center of the distribution of values, the higher the weighting is. When the characteristics of a general Gaussian distribution appear in various forms, they can be formed as a mixed model.

Finally, the algorithms that may be used in step S1400 of generating user model data and personalizing the search results using the same may include matching algorithms, unique matching algorithms, language model algorithms, etc., as illustrated in FIGS. 8 through 9 . The matching algorithm is a method of considering the number of keywords in the search result document in calculating the sum of the keyword weights when the keyword of the user model data is in the search result document. And the language model algorithm divides the keyword weight by the weighted sum of all the keywords.

Algorithms that may be used at each step of the method of providing personalized search results may be various algorithms other than the algorithm shown in FIG. Depending on the document, the algorithm that is appropriate for applying according to the keyword may vary depending on the user.

11 shows a result of testing the importance of a specific document according to each algorithm combination.

Referring to FIG. 11, the combinations of the respective algorithms were tested on the same user model data and the same search result document. In the case of using TF-IDF, TF-IDF (norm), BM25, and applying the weight correction algorithm to calculate the weight, POSTagging algorithm (not shown) I tested it when it does not apply. It can be confirmed that the importance of the search result document is calculated in various ways according to the algorithm applied to the same user model data and the same search result document.

When the relevance of the search result document is calculated to be highest (TF-IDF (norm) / GMM / unique matching algorithm), 0.55 is used. When the importance of the search result document is calculated to be lowest TF-IDF / language model algorithm) or (BM25 / language model algorithm) or (TF-IDF (norm) / GMM / matching algorithm) combination. It can be seen that the significance of the search result document from 0.3 to 0.55 can be calculated variously according to the combination of the algorithms.

When the retrieved results are sorted and provided, the importance of each document can be obtained, and the order of the documents exposed to the user can be sorted and provided using the obtained importance. In this case, factors affecting the importance of the document may affect the combination of the algorithm used in the personalized search result providing method as well as the user model data generated based on the documents already viewed by the user. If you can personalize the combination of algorithms for a particular user, you will be able to provide users with more reliable search results. However, since there are many difficulties in finding a suitable combination of algorithms for a specific user, some methods for simply performing this algorithm will be described with reference to FIG.

12 is a flowchart illustrating a method of providing a personalized search result by machine learning of an algorithm combination according to an embodiment of the present invention.

Referring to FIG. 12, after performing step S1300 of generating user model data, a step S1310 of clustering users based on the generated user model data may be performed. In other words, it is necessary to group users based on user model data because there is not enough sample or experiment results to learn the algorithm combination by user. Considering the keywords included in the user model data and the weight of each keyword, clustering among users having similar keyword weights can learn algorithm combinations more efficiently.

After clustering the user model data, the user's reaction is machine-learned by using algorithm for extracting keywords for each cluster, algorithm for calculating weight, algorithm for correcting weight, and algorithm for calculating document importance. (S1320). That is, when the time consumed for browsing a specific document in the search result is used, it is possible to infer which document the user has searched for through a search. Therefore, The percentage of documents included can be used to learn which algorithm combination is more useful for the cluster. In the machine learning process, the statistical Naive Bayes algorithm, the SVM (Support Vector Machine) algorithm and the kNN (k-Nearest Neighbor) algorithm can be applied.

FIG. 13 shows a result of testing the response of the user according to each algorithm combination.

In the case of providing a search result by a combination of several algorithms for a specific user model data cluster, it is determined whether or not the user gives priority to the information to be obtained through search based on the user's reaction to the search result provided to the user Respectively.

In the specific user model data cluster, the combination of POSTagging as a keyword extraction algorithm, BM25 as a primary weight calculation algorithm, SVD as a secondary weight correction algorithm, and a matching algorithm (not shown) ) Was 0.79, 0.49 for recall and 0.60 for F-measure.

If the algorithm combination is learned for the user cluster based on each user model data as in the example of FIG. 13, not only the user model data becomes more refined by feedback of the document read by the user, but also the user- The algorithm combination to be applied to the cluster to which the model data belongs may be more elaborate. This makes it possible to provide personalized search results more efficiently.

14 is a hardware configuration diagram of a personalized search result providing apparatus according to an embodiment of the present invention.

14, the personalized search result providing apparatus 10 may include one or more processors 510, a memory 520, a storage 560, and a network interface 570. As shown in FIG. The processor 510, the memory 520, the storage 560, and the interface 570 transmit and receive data via the system bus 550.

The processor 510 executes a computer program loaded into the memory 520 and the memory 520 loads the computer program from the storage 560. [ The computer program may include a keyword extraction operation 521, a weight calculation operation 523, a user model data generation operation 525, an algorithm combination learning operation 527, and a search result personalization operation 529.

The keyword extracting operation 521 can extract a keyword by applying an algorithm such as Context or POSTagging to a document viewed by the user among the document data 569 of the storage 560. Also, in order to calculate the importance of each document retrieved by inputting a query word among the document data 569 of the storage 560 using the user model data 561, keywords can be extracted for each retrieved document.

The weight calculation operation 523 can calculate the weight of each keyword by applying algorithms such as TF-IDF and BM25 to the keywords extracted from the keyword extraction operation 521. [ In addition, the weight calculation operation 523 may perform correction by applying algorithms such as GMM and SVD to the calculated weights as necessary or applying time weighting based on the time the user browses the document, can do.

The user model data generation operation 525 may generate weight model information 561 by summing weight information corresponding to the keywords and keywords calculated in the weight calculation operation 523 on the basis of a document viewed by the user. The user model data 561 generated at this time is stored in the user model data 561 of the storage 560 via the system bus 550.

The algorithm combination learning operation 527 classifies users into a plurality of clusters based on the user model data 561 generated in the user model data creation operation 525 and classifies the clusters into a plurality of clusters based on the keyword extraction algorithm, , Weights correction algorithm, and document importance calculation algorithm are input, and the user's reaction to the search result is monitored, so that a combination of algorithms suitable for the cluster can be learned. The thus learned user cluster-specific algorithm combination is stored in the algorithm combination data 563 of the storage 560 via the system bus 550.

The search result personalization operation 529 receives a query term from the user via the network interface 570 and searches the document data 569 of the storage 570 for the retrieved documents, And algorithm combination data (563). The search results are sorted according to the calculated importance, and personalized search results are provided to each user.

Each component in FIG. 14 may refer to software or hardware such as an FPGA (Field Programmable Gate Array) or an ASIC (Application-Specific Integrated Circuit). However, the components are not limited to software or hardware, and may be configured to be addressable storage media, and configured to execute one or more processors. The functions provided in the components may be implemented by a more detailed component, or may be implemented by a single component that performs a specific function by combining a plurality of components.

While the present invention has been described in connection with what is presently considered to be practical exemplary embodiments, it is to be understood that the invention is not limited to the disclosed embodiments, but, on the contrary, You will understand. It is therefore to be understood that the above-described embodiments are illustrative in all aspects and not restrictive.

Claims (17)

Extracting a keyword of the document on a document viewed by a user;
Calculating a weight in consideration of a frequency of the keyword in the document;
Mapping the weight to the keyword to generate user model data for the user; And
Receiving a query term from the terminal of the user and using the user model data to sort the search results for the query term;
A method for providing personalized search results. The method according to claim 1,
The step of calculating the weight includes:
And correcting the weighting factor in consideration of the distribution of the keyword and the weighting factor.
A method for providing personalized search results. The method according to claim 1,
The step of calculating the weight includes:
And correcting the weight by considering information on a time when the user browses the document.
A method for providing personalized search results. The method of claim 3,
The step of correcting the weight includes:
And correcting the weight significantly as the user is viewing the document for a longer period of time.
A method for providing personalized search results. The method of claim 3,
The step of correcting the weight includes:
And correcting the weight significantly as the time the user has browsed the document,
A method for providing personalized search results. The method according to claim 1,
Wherein the step of sorting the search results for the query term comprises:
Extracting a search result keyword from each document belonging to the search result;
Comparing the search result keyword with the keyword, and calculating importance of each document based on a weight mapped to the keyword; And
And arranging search results for the query terms using the importance of each document.
A method for providing personalized search results. The method according to claim 1,
Wherein the step of sorting the search results for the query term comprises:
And updating the user model data by feeding back the response of the user to the sorted search result.
A method for providing personalized search results. The method according to claim 1,
Wherein generating user model data for the user comprises:
Clustering a plurality of users based on user model data to construct a cluster,
A method for providing personalized search results. 9. The method of claim 8,
Wherein configuring the cluster comprises:
Determining a keyword extraction algorithm of the cluster by mechanically learning a response of a user belonging to the cluster to a keyword extraction algorithm;
A method for providing personalized search results. 9. The method of claim 8,
Wherein configuring the cluster comprises:
Learning the response of a user belonging to the cluster to a weight calculation algorithm to determine a weight calculation algorithm of the cluster.
A method for providing personalized search results. 9. The method of claim 8,
Wherein configuring the cluster comprises:
Learning a response of a user belonging to the cluster to a weight correction algorithm, and determining a weight correction algorithm of the cluster;
A method for providing personalized search results. 9. The method of claim 8,
Wherein configuring the cluster comprises:
Learning a response of a user belonging to the cluster to an importance computation algorithm to determine an importance computation algorithm of the cluster;
A method for providing personalized search results. A keyword extracting unit for extracting a keyword of the document from a document viewed by a user;
A weight computing unit for computing a weight by considering the frequency of the keyword in the document;
A user model data generation unit for mapping the weights to the keyword to generate user model data for the user; And
And a search result personalization unit that receives a query term from the terminal of the user and sorts the search result for the query term using the user model data,
A personalized search result providing device. 14. The method of claim 13,
The weight computing unit may include:
And a weight correction unit for correcting the weight by considering the distribution of the keyword and the weight,
A personalized search result providing device. 14. The method of claim 13,
The search result personalization unit,
A search result keyword extracting unit for extracting a search result keyword from each document belonging to the search result;
A document importance calculation unit for comparing the search result keyword with the keyword and calculating importance of each document based on a weight mapped to the keyword; And
And a search result sorting unit for sorting search results for the query terms by using the importance of each of the documents.
A personalized search result providing device. 14. The method of claim 13,
The search result personalization unit,
And a user model data updating unit for updating the user model data by feeding back the response of the user to the sorted search result.
A personalized search result providing device. Network interface;
One or more processors;
A memory for loading a computer program executed by the processor; And
A storage for storing document data, user model data, algorithm combination data,
The computer program comprising:
A keyword extraction operation for extracting a keyword of the document from a document viewed by a user among the document data;
A weight calculation operation for calculating a weight in consideration of a frequency of the keyword in the document;
A user model data generation operation for mapping the weights to the keyword to generate the user model data for the user;
A plurality of users are clustered on the basis of user model data to constitute a cluster, and a reaction of a user belonging to the cluster with respect to a keyword extraction algorithm, a weight calculation algorithm, a weight correction algorithm and an importance calculation algorithm are machine- An algorithm combination learning operation for generating algorithm combination data; And
A search result personalization operation that receives a query term from the terminal of the user and uses the user model data to sort the search results for the query term;
A personalized search result providing device.

Images