JP5551665B2 - Information extraction apparatus, information extraction method, and information extraction program - Google Patents

Description

  The present invention is input / selected as a narrowing-down condition for obtaining a product list among the conditions in which a user browses a product detail page with respect to a product list presented on, for example, an EC (Electronic Commerce) site. The present invention relates to a technique for extracting hidden conditions other than conditions as browsing conditions.

  In recent years, EC sites, online shopping malls, cyber malls, and the like have spread on the Web, and users who purchase products using the Internet are increasing. As a method for such a user to search for a product for purchase purpose, a method is used in which a product list is obtained by narrowing down a product group by keyword search or category search, and a product suitable for the purchase purpose is browsed from the product list. ing. At this time, the user uses the condition of the product suitable for the purchase purpose as the viewing condition.

  Marketers, on the other hand, estimate the browsing conditions actually intended by users browsing the site in order to understand customer needs, and use the input words such as keywords and categories used for narrowing down and the group of products viewed. .

  However, since the user does not necessarily input all browsing conditions as input words, estimation of browsing conditions using input words cannot estimate all browsing conditions actually intended by the user. Moreover, in the estimation of browsing conditions using the browsed product group, even if the user intends only a part of the conditions common to the browsed products as the viewing conditions, it cannot be distinguished, and all are browsed As a result of condition estimation.

  Therefore, Non-Patent Document 1 proposes a method for estimating a user's viewing condition using other information. The method of Non-Patent Document 1 uses a search log that is a user's browsing behavior history, expands an input word to a related word, and estimates a viewing condition based on the expanded word. In this search log, an input word input as a search query by a user, a list of Web pages presented as a search result based on the input word, and a Web page viewed from the list are recorded as information. Here, the browsing condition of the user is estimated with the condition that is not an input word among the browsing conditions of the Web page browsed from the list as a hidden condition.

  As a method for extracting the hidden condition, there is a method for estimating the hidden condition by extracting a word from a sentence describing a Web page. At this time, the hidden condition degree indicating the index of the hidden condition is calculated so as to be given higher to words that appear more frequently in the browsed web page group than in the web page group that has not been browsed. Here, a word having a high degree of the hidden condition is extracted as a specific word, that is, a hidden condition.

Takenobu Tokunaga "Language and Computation 5 Information Retrieval and Language Processing" The University of Tokyo Press, November 25, 1999 First edition issued p. 154-159

  However, in the conventional method, a high degree of hidden condition is not given to an important word in marketing, and there is a possibility that a problem may occur in extraction of browsing conditions.

  That is, in marketing, a word representing a viewing condition that can explain a viewed product is important. In order to be able to explain a product, it is necessary that the viewing conditions indicate an attribute value (a value that gives a specific value to the attribute), not an attribute (an abstract one such as a color or price). For example, in the case of the product attribute “color”, the product cannot be described by “colored product”, but the product can be described by “black product” in which the product attribute “color” is embodied. Therefore, in order to estimate viewing conditions in marketing, it is required to estimate attribute values as viewing conditions.

  However, since words that appear more frequently in the viewed product group than the product group that the user did not browse are not necessarily attribute values, there is a possibility that a high hidden condition degree may not be given to the attribute value in the conventional method. For example, in the case of fashion sites, words that are not important as hidden conditions among terms used in fashion sites such as “color”, “fashion”, and “net” appear in common in the viewed product group. There is a risk of giving a high degree of hidden condition to the attribute, not the attribute value.

  The present invention has been made to solve the problems of the prior art as described above, and does not give a high degree of hiding condition to insignificant words among terms used in a website, and high hiding for important words. The problem to be solved is to provide a degree of condition and improve the accuracy of the extracted browsing conditions.

  Therefore, the present invention extracts a common word common to each product in the product list, and pays attention to a dependency structure between the specific word and the common word. That is, a specific word having a high appearance frequency of a dependency structure whose dependency is a dependency on a common word and a dependency on a common word is extracted as a hidden condition.

  One aspect of the present invention is that browsing conditions other than a condition that is input and selected as a narrowing-down condition for obtaining a product list among viewing conditions in which a user browses products from a product list presented from a website are used as browsing conditions. An information extraction device for extracting, a common word extraction means for extracting a word common to each product listed in the product list as a common word according to the appearance order of the word, and a user browsing in the product list Gives a higher score to a word for a product viewed than a product that has not been browsed, and accumulates character string information for all products registered on the website, and a unique word extraction means for extracting as a unique word according to the ranking of the given score The frequency of occurrence of the dependency structure of the dependency language of the dependency language is obtained by referring to the database, and the dependency language of the dependency language of the dependency language of the dependency language is acquired. Comprising Re calculates condition of the hidden conditional calculation means for extracting a condition hidden 該特 Yugo according to order of the hidden conditional degree calculated, the.

  According to another aspect of the present invention, a hidden condition other than a condition input / selected as a narrowing condition for obtaining a product list among viewing conditions for a user viewing a product from a product list presented from a website is viewed. Information extraction method of the apparatus for extracting as a common word extraction step of extracting a word common to each product listed in the product list as a common word according to the appearance order of the word, A unique word extraction step for giving a higher score to a product viewed than a product that is not browsed, and extracting as a specific word according to the ranking of the given score, and a character string for all products registered on the website Refer to the database that stores the information, acquire the appearance frequency of the dependency structure where the unique word is the dependency and the common word is the dependency source, and the acquired appearance frequency is Having a hidden conditional calculation step of calculating a hidden condition of the 該特 Yugo, extracted condition hidden 該特 Yugo according to order of the hidden conditional degree calculated by household.

  In addition, this invention is good also as an aspect of the program which makes a computer function as said apparatus. This program can be provided through a network or a recording medium.

  According to the present invention, a high hiding condition level is given to important words without giving a high hiding condition level to unimportant words among terms used on the website, and the accuracy of the browsing conditions to be extracted is improved. Can do.

The block diagram of the information extraction apparatus which concerns on embodiment of this invention. The process flowchart. The processing flowchart of the same common word extraction part. The processing flowchart of the unique word extraction unit. The processing flowchart of the same hidden condition degree calculation part. Explanatory drawing of a dependency structure. Frequency S (a _i, b _j) explanatory diagram showing an example of. Explanatory view showing another example of frequency _{S (a i, b j)} . The image figure of the goods list which concerns on an Example. The figure which shows the state which browsed the product detail page. The figure which shows the same input word set, goods list group, and a specific score. The figure which shows the example of extraction of the same hidden condition.

≪Configuration example≫
A configuration example of an information extraction apparatus according to an embodiment of the present invention will be described based on FIG. This information extraction apparatus 1 extracts browsing conditions of a user who browses an EC site in order to grasp customer needs. Here, conditions other than input words such as search keywords and categories are extracted as browsing conditions from the narrowing-down conditions of the product list browsed by the user.

  Specifically, the information extraction apparatus 1 is configured by a server group that manages and operates an EC site, and the hardware resources of a normal computer such as a CPU. A storage device such as a memory (RAM) or a hard disk drive device is provided. As a result of the cooperation of the hardware resource and the software resource (OS. Application, etc.), the information extraction apparatus 1 has a common word extraction unit 2. 2. Specific word extraction unit It functions as the hidden condition degree calculation unit 4.

  The information extraction device 1 includes a product information input device 5. 5. Hidden condition output device All product information database 7 is connected. The database 7 stores all products on the EC site and character string information for each product as product information data. Note that the buffers B1 to B9 in FIG. 1 are constructed in the storage device, and the database 7 may be built in the information extraction device 1.

  The processing contents will be described with reference to FIG. 2. Various browsing information of the user at the EC site is written into the buffers B1 to B3 of the information extraction device 1 through the product information input device 5 as input data (S01). That is, the product list word group viewed by the user as a processing target is written in the product information group input buffer B1 of the common word extraction unit 2. Further, the total word group for all products listed in the product list is written in the total word appearance number group buffer B2. Further, the input words used by the user for the product list narrowing condition, that is, the words obtained by dividing the character string such as the search keyword and category selection by morphological analysis are written in the input word group buffer B3.

Tables 1 to 3 show examples of data written to the buffers B1 to B3. Here, it is assumed that the user has narrowed down the product list using the search keyword “eye shadow”. In this product list, it is assumed that the product detail page is clicked to browse the product detail page, while the product 2 is not clicked and the product detail page is not browsed. The comments on the product detail page for this product 1.2 are as follows.
・ Product 1 Product name “Popular xxx Mystic Eyes” Comment “Eye shadow set with four beautifully harmonious colors of xxx”
・ Product 2 Product name “Yyyy Secret Collection Popular Comment” “Yyyyy's playful eye shadow”

  Table 1 shows an example of a product list word group written in the product information group input buffer B1. This product list word group is created based on words extracted from sentences (for example, product descriptions, comments, etc.) on the product detail page displayed after clicking the product for all products listed in the product list. Here, for each word number of each extracted word, the product number, the presence / absence of product browsing, and the number of appearances of the word are written in each row as product list word information. This product number indicates a number uniquely assigned to a product in which the word number appears on the product detail page. The presence / absence of product browsing is indicated as “browsing” when the user is browsing the product detail page of the product, and is indicated as “non-browsing” when the user is not browsing.

  Table 2 shows an example of the total word appearance number group written in the total word appearance number group buffer B2. Here, the total number of words appearing on the product detail page is written for each product number. Table 3 shows input words written to the input word group buffer B3.

  Table 4 shows an example of product information data stored in the database 7. The product information data includes character string information on a product detail page for all products prepared in advance by hand, for example, all products registered in an EC site (in this case, a fashion site) that is a hidden condition extraction target. This character string information is composed of a pair of a product number of a product and a character string group of a product detail page.

  Then, the common word extraction unit 2 extracts, as a common word, a word having a high appearance frequency on the product detail page of the product listed in the product list based on the product list word group written in the product information group input buffer B1 (S02). ). Moreover, the specific word extraction part 3 extracts a specific word based on the product list word group written in the product information group input buffer B1. Here, in the product list, a higher specific score is given to a word that appears on the product detail page of the product viewed than the product detail page of the product that the user has not browsed, and a word higher in the specific score is extracted as a specific word ( S03). This unique score is calculated by the ratio between the word appearance probability on the product detail page of the product viewed by the user and the word appearance probability on the product detail page of the product not viewed.

  Further, the hidden condition calculation unit 4 acquires the appearance frequency of the dependency structure where the unique word is the dependency destination and the common word is the dependency source from the stored data of the database 7, and the unique word is used by using the acquired appearance frequency. The degree of the hidden condition is calculated (S04). A unique word having a higher rank of the calculated hidden condition degree is output as a hidden condition to the hidden condition output device 6 (S05). At this time, the specific words corresponding to the input words written in the input word group buffer B3 are excluded.

  The extracted hidden conditions are output to the EC site management server or the like via the hidden condition output device 6 and used for marketing of the EC site. Hereinafter, the specific processing content of each part 2-4 is demonstrated.

≪Processing contents of common word extraction unit 2≫
The specific processing contents of the common word extraction unit 2 will be described with reference to FIG. Here, the detailed processing content of S02 is shown, and the common word extraction unit 2 starts processing when the product list word group is written from the product information input device 5 to the product information group input buffer B1. It is assumed that the data in Table 1 is written in the product information group input buffer B1.

  S11. S12: When the process is started, the common word extraction unit 2 extracts one product list word information from the product information group input buffer B1 (S11). The number of products including the words in the product list information extracted here is counted and written in the common word group buffer B4 (S12). Table 5 shows an example of data written in the common word group buffer B4, and the number of appearing products is written for each word of the product list information.

  S13. S14: The common word extraction unit 2 checks whether or not there is product list word information that has not been processed in S12 in the product information group input buffer B1 (S13). If there is unprocessed product list word information as a result of the confirmation, the process returns to S12 and processing such as counting the number of products is resumed. On the other hand, if there is no unprocessed product list word information, the process proceeds to S14.

That is, the common word extraction unit 2 rearranges the words in the common word group buffer B4 in descending order of the number of appearing products, and extracts the top N _e using a predetermined constant N _e (S14). The extracted top N _e words are written as common words in the common word group input buffer B5 of the hidden condition degree calculation unit 4, and the process is terminated. Table 6 shows an example of data written in the common word group input buffer B5, in which N _e common word groups are written.

≪Processing contents of specific word extraction unit 3≫
Specific processing contents of the specific word extraction unit 3 will be described with reference to FIG. Here, the detailed processing content of S03 is shown, and the specific word extraction unit 3 starts the process when the common word group is written from the common word extraction unit 2 to the common word group input buffer B5. At this time, the data in Table 1 is written in the product information group input buffer B1, the data in Table 2 is written in the total word appearance number group buffer B2, and the common word group in Table 6 is stored in the common word group input buffer B5. It shall be written.

  S21. S22: When the process is started, the specific word extraction unit 3 extracts one product list word information word from the product information group input buffer B1 (S21). Here, in Table 1, words in an arbitrary column are taken out. A characteristic score of the extracted word is calculated, and a pair of the extracted word and the calculated specific score is written in the specific word group buffer B6 (S22).

Table 7 shows an example of data written in the specific word group buffer B6, and a specific score is written for each word number of the word. This unique score is calculated so that the word of the product indicated as “browse” is higher in score than the product indicated as “non-browse” in Table 1. The calculation method is as follows: the occurrence probability P _{c, t of} the product number of the product number (viewed product) indicated as “browsing” in Table 1, the product of the product number indicated as “non-viewed” (non-viewed product) If the word appearance probability P _{s, t in} , the unique score Sp _{c, t} can be calculated by the equation (1).

However, in the case of “appearance probability P _{s, t} = 0”, the unique score Sp _{c, t} is a predetermined “constant T _click (arbitrary constant)”. The appearance probability P _{c, t} can be calculated by the number of appearances of words in all browsed products (total number of appearances of browsed products) / number of products of browsed products (total number of browses). Similarly, the appearance probability P _{s, t} can be calculated by the number of appearances of words in all non-viewed products (total number of non-viewed products) / number of non-viewed products (total number of non-viewed products).

In addition, the appearance probability P _{c, t} is calculated by the total number of appearances of browsing products / the total number of words of browsing products, and the appearance probability P _{s, t} is calculated by the total number of appearances of non-viewing products / the total number of words of non-viewing products. Also good. As the total word number, the total word appearance number in the total word appearance number group buffer B2 (Table 2) may be used.

  S23. S24: The specific word extraction unit 3 checks whether or not product list word information that has not been processed in S22 exists in the product information group input buffer B1 (S23). If there is unprocessed product list word information as a result of the confirmation, the process returns to S22 and processing such as counting the number of products is resumed. On the other hand, if there is no unprocessed product list word information, the process proceeds to S24.

That is, specific word extraction section 3 rearranges word in specific word group buffer B6 in descending order of specific scores, higher N _o number is extracted by using constants N _o stipulated in advance (S24). The extracted upper N _o number of words as specific word, to end the writing and process the unique word group input buffer B7 hidden condition calculation unit 4. Note that Table 8 shows an example of data written in the specific language group input buffer B7, are written specific score for each word number to N _o number of unique word group.

≪Processing contents of hidden condition degree calculation unit 4≫
The processing contents of the hidden condition degree calculation unit 4 will be described with reference to FIG. Here, S04. The details of S05 are shown, and the hidden condition degree calculation unit 4 starts processing when a specific word group is written from the specific word extraction unit 3 to the specific word group input buffer B7. At this time, each buffer B5. In B7, Table 6. It is assumed that the data in Table 8 is written and the product information data in Table 4 is written in the database 7.

  S31 to S34: The hidden condition degree calculation unit 4 extracts one specific word from the specific word group input buffer B7 (S31), and extracts one common word from the common word group input buffer B5 (S32). One piece of product information data is extracted (S33), and then the process proceeds to S34.

  In S34, first, a character string is extracted from the product information data, and the appearance frequency of the dependency word in the character string is calculated. At this time, the dependency word is a character string composed of “common words of specific words” as shown in FIG. Here, for example, the appearance frequency (number of appearances) for “xxx eye shadow” in the extracted character string is counted for each product.

  Next, the same dependency word and the appearance frequency of the dependency word are referred to from the dependency word group buffer B8, and if the same dependency word exists, the current appearance frequency is added to the reference appearance frequency to determine the dependency word. Write to group buffer B8. On the other hand, if the same dependency word does not exist, the current dependency word and the appearance frequency are written in the dependency word group buffer B8. Table 9 shows an example of data written in the dependency word group buffer B8, in which dependency words and appearance frequencies are written for each product number.

  S35 to S37: The hidden condition degree calculation unit 4 performs the confirmation process of S34 to S37. That is, it is checked whether or not there is a character string that has not been processed in S35 in the product information data extracted in S34 (S35). As a result of the check, if there is an unprocessed character string, the process of S34 is resumed. If there is no unprocessed character string, the process proceeds to S36.

  In S36, it is confirmed whether or not there is product information data in the database 7 that has not been processed in S33 and thereafter. As a result of the confirmation, if there is unprocessed product information data, the process returns to S33 to continue the process, while if there is no unprocessed product information data, the process proceeds to S37.

  In S37, it is confirmed whether or not there is a common word that has not been processed in S32 and subsequent steps in the common word group input buffer B5. As a result of the check, if there is an unprocessed common word, the process returns to S32 to continue the process, whereas if there is no unprocessed common word, the process proceeds to S38.

  S38: The hidden condition degree calculation unit 4 calculates the hidden condition degree of the specific word extracted in S31 from the dependency word group buffer B8, and pairs the calculated hidden condition degree with the specific word to cover the hidden condition degree group buffer. Write to B9.

That is, for all the dependency words included in the dependency word group buffer B8, the specific word (a ₁ ) extracted in S31 is the dependency destination, and each of the processed common words (b in S32 to S37) ₁ , b ₂ ,...) Using the frequency S (a _i , b _j ) in which the dependency structure (dependency word) that becomes the dependency source appears, the hidden condition degree C of the specific word (a ₁ ) _ai is calculated by equation (2).

An example of the frequency S (a _i , b _j ) will be described based on FIG. Here, it is assumed that the unique word (a ₁ ) is “xxxx” and each common word (b ₁ , b ₂ ,...) Is “popular”, “eyeshadow”, and “set”. Therefore, there are three types of dependency structures: “xxxx-popularity”, “xxxx-eye shadow”, and “xxxx-set”.

At this time, the appearance frequency of each dependency structure is written in the dependency word group buffer B8 for each item (product). Here, based on the data in the dependency word group buffer B8, the number of appearance products of each dependency structure is used as the frequency S (a _i , b _j ).

According to the example of FIG. 7, “xxxx-popular” appears in item 1.2, “xxxx-eyeshadow” appears in items 1 to 3, and “xxxx-set” appears in item 3. . Therefore, the frequency S (a _i , b _j ) is (“xxxx-popularity” = 2 “xxxx-eyeshadow” = 3 “xxxx-set” = 1).

Another example of the frequency S (a _i , b _j ) will be described based on FIG. Here, the frequency S (a _i , b _j ) obtained by normalizing the number of appearances of all the dependency structure items by the number of items is used.

  According to the example of FIG. 8, “xxxx-popularity” appears twice in item 1, once in item 2, and 0 in item 3 for a total of 3 times. In addition, “xxxx-eye shadow” appears once for each of items 1 to 3 in total. Furthermore, “xxxx-set” appears once in total, 0 times for item 1, 0 times for item 2, and 1 time for item 3. The total number of appearances is divided by the number of items “3” and normalized (averaged).

Therefore, the frequency S (a _i , b _j ) is (“xxxx-popularity” = 3/3 = 1 “xxxx-eyeshadow” = 3/3 = 1 “xxxx-set” = 1/3 = 0.33 )

  Table 10 shows an example of data written in the hidden condition degree group buffer B9. Here, the degree of hidden condition is written for each unique word.

  S39. S40: The hidden condition degree calculation unit 4 checks whether or not there is a unique word that has not been processed in S31 and subsequent processing in the unique word group input buffer B7 (S39). As a result of the confirmation, if there is an unprocessed unique word, the process returns to S31 to continue the process, but if not, the process proceeds to S40.

That is, the hidden condition degree calculation unit 4 rearranges the specific word groups in the hidden condition degree group buffer B9 in descending order of the hidden condition degree, and is a predetermined constant N that is not a word included in the input word group buffer B3. _The top N _h pieces are extracted as hiding conditions using _h , the extracted hiding conditions are output to the hiding condition output device 6 and input to the device 6 (S40), and the process is terminated. Table 11 shows hidden conditions input to the hidden condition output device 6.

As described above, according to the information extraction apparatus 1, since the degree of the hidden condition is calculated using the dependency structure frequency S (a _i , b _j ), the word indicating the attribute (for example, “color”) is not set as the hidden condition. It is possible to set an attribute value (for example, “black”) as a hidden condition. Therefore, for example, terms used in fashion sites such as “color”, “fashion”, “net”, etc., but not important words are not given a high hiding condition level, and only important words are given a high hiding condition level. It is output as a hidden condition, and the accuracy of the output viewing condition can be improved. As a result, the marketer can more accurately estimate the viewing conditions of the browsing user of the EC site, and can contribute to grasping customer needs.

<Example>
Hereinafter, an embodiment using the information extraction apparatus 1 will be described with reference to FIGS. Here, the user inputs the keyword of the input word “T-shirt zzz”, and the product list of FIG. 9 is presented. At this time, the user's hiding condition is that “the product of the navy shirt is viewed from the product list (for the purpose of extracting navy)”. In addition, 1-12 in FIG. 9 shall show the goods of goods list publication.

  It is assumed that the user browses a product detail page by clicking on a product of interest with respect to this product list. In FIG. 10, the state which clicked the goods 1 as an example and browsed the goods detail page is shown. Here, similarly, the product detail page is also viewed for the product 7.8.12, and viewing / non-viewing of the product detail page is “total number of browsing = 4. Total number of non-viewing = 8”. A product list word group to be written in the product information group input buffer is created based on the product information extracted from each product detail page. Further, the product information data of the database 7 is created based on the product information on the detailed page of each product registered in the EC site regardless of the user's browsing.

  In the product information group input buffer B1, as shown in FIG. 11, presence / absence of browsing for each product number (product 1-5) for each word extracted from the product detail page of the product listed in the product list ( (Browsing or non-browsing)-The number of browsing times is written in a pair.

In calculating the specific score of each word, if “P _c = total number of browsing products / total number of browsing. P _s = total number of non-viewing products / total number of non-viewing”,
・ Navy P _c = 3/4 P _s = 0/8
One size _{P c = 3/4 P s} = 1/8
・ Material P _c = 3/4 P _s = 0/8
Nirvana P _c = 0/4 P _s = 2/8
Is calculated.

And, according to the processing of S21~S24, words included many in the product details pages viewed than the product detail page that does not view the user (big word of the ratio of P _s of P _c), that is, the "navy" As shown in FIG. 12, “material” is written in the specific word group input buffer B7 as a specific word. Here, it is assumed that “T-shirt”, “zzz”, and “Made in China” included in the product detail pages of all the products are written in the common language group input buffer B5 in advance through the processing of S11 to S14.

At this time, the large-scale product information (product information data group) of the dependency structure (common word of unique words), that is, the number of appearances in the database 7 is “navy * (common language)” = 2000 times. * (Common language) ”= 50 times. Here, it is assumed that a constant N _h = 1000 times is set. Then, “Navy” is output as a hidden condition through S31 to 40, whereas “Material” is not a hidden condition and is not output.

  Therefore, according to the information extraction apparatus 1, the specific words that match the user's hiding conditions are output, the specific words that do not match the user's hiding conditions are not output, and the user's viewing conditions can be estimated correctly.

≪Programs≫
The present invention can also be configured as an information extraction program that causes a computer to function as some or all of the units 2 to 4 of the information extraction apparatus 1. According to this program, S01 to S05. S11-S14. S21-S24. It is possible to cause the computer to execute part or all of S31 to S40.

  The program can be provided through a network such as a website or e-mail. The program is stored in a recording medium such as a CD-ROM, DVD-ROM, CD-R, CD-RW, DVD-R, DVD-RW, MO, HDD, BD-ROM, BD-R, or BD-RE. It is also possible to record, save and distribute. This recording medium is read using a recording medium driving device, and the program code itself realizes the processing of the above embodiment, so that the recording medium also constitutes the present invention.

DESCRIPTION OF SYMBOLS 1 ... Information extraction apparatus 2 ... Common word extraction part (common word extraction means)
3 ... Specific word extraction unit (specific word extraction means)
4 ... Hidden condition degree calculation unit (hidden condition degree calculation means)
5 ... Product information input device 6 ... Hidden condition output device 7 ... All product information database (database)
B1 ... Product information group input buffer B2 ... Total word appearance number group buffer B3 ... Input word group buffer B4 ... Common word group buffer B5 ... Common word group input buffer B6 ... Special word group buffer B7 ... Special word group input buffer B8 ... Related Language group buffer B9 ... Hidden condition degree group buffer

Claims (7)

An information extraction device that extracts hidden conditions other than the conditions input / selected as a narrowing-down condition for obtaining a product list, among browsing conditions for a user browsing a product from a product list presented from a website. And
A common word extraction means for extracting a word common to each product listed in the product list as a common word according to the appearance order of the word;
In the product list, a specific word extraction unit that gives a higher score to a product viewed than a product that the user has not browsed and extracts as a specific word according to the rank of the given score,
The appearance frequency of the dependency structure where the unique word is the dependency destination and the common word is the dependency source is obtained by referring to a database storing character string information for all products registered on the Web site, and the acquired appearance frequency A hidden condition degree calculating means for calculating the hidden condition degree of the specific word using the and extracting the specific word as a hidden condition according to the rank of the calculated hidden condition degree;
An information extraction device comprising: The information extraction apparatus according to claim 1, wherein the hidden condition degree calculation means uses the appearance frequency as the number of products in which the dependency structure appears in the character string information. 2. The information according to claim 1, wherein the hidden condition degree calculation means uses the numerical value obtained by normalizing the number of appearances of the dependency structure in the character string information by the number of products in which specific words appear as the appearance frequency. Extraction device. Information extraction method for an apparatus that extracts hidden conditions other than the conditions input / selected as a narrowing-down condition for obtaining a product list, among browsing conditions for a user browsing products from a product list presented from a Web site Because
A common word extraction step of extracting a word common to each product listed in the product list as a common word according to the appearance order of the word;
In the product list, a specific word extraction step of giving a higher score to the product viewed than the product not viewed by the user and extracting as a specific word according to the ranking of the given score,
The appearance frequency of the dependency structure where the unique word is the dependency destination and the common word is the dependency source is obtained by referring to a database storing character string information for all products registered on the Web site, and the acquired appearance frequency A hidden condition degree calculating step of calculating a hidden condition degree of the specific word using the and extracting the specific word as a hidden condition according to the rank of the calculated hidden condition degree;
An information extraction method characterized by comprising: The information extraction method according to claim 4, wherein the hidden condition degree calculating step uses the number of products in which the dependency structure appears in the character string information as the appearance frequency. 5. The information according to claim 4, wherein the hidden condition degree calculating step uses the numerical value obtained by normalizing the number of appearances of the dependency structure in the character string information by the number of products in which specific words appear as the appearance frequency. Extraction method.   The information extraction program which makes a computer function as each means of the information extraction apparatus of any one of Claims 1-3.

Images