JP5819239B2 - Important word / phrase extraction apparatus, method, and program - Google Patents

Description

  The present invention relates to an important phrase extracting apparatus, method, and program for extracting an important phrase from a document.

  When a user searches and browses a document with a certain request, it is assumed that a portion including information closely related to the user's request in the document to be searched or browsed is an important portion to be extracted. The unit of the important part can be a word (important word) indicating information closely related to the user's request, a set of sentences such as sentences or paragraphs including the important word, and the like. Hereinafter, this important part is referred to as an important phrase. Here, the user's request does not refer to specific things such as “cheap and delicious shops” and “scenery places”, but the user places importance on searching or browsing documents. Information. Conventionally, an important phrase related to such a user request is extracted from a document and presented to the user.

  For example, for the purpose of easily grasping the user's browsing history, the user's browsing history is collected, and a plurality of browsing histories to be processed are classified into clusters based on the similarity between the browsing histories. A summary generation device has been proposed in which keywords are extracted for each cluster, the keywords and information related to the browsing history for each cluster are generated as a summary, and presented to the user (see, for example, Patent Document 1).

  In addition, as a keyword extraction method for extracting words having high importance for the user from the history of documents viewed by the individual user, a word having a high frequency of appearance is extracted from the document viewed by the user, and co-occurrence with the word is performed. There has been proposed a method of extracting a word having a large bias as a keyword having high importance for the user (see, for example, Non-Patent Document 1).

JP 2011-100350 A

Yutaka Matsuo, Hayato Fukuda, Mitsuru Ishizuka, “Supporting Browsing by Extracting Keywords from User's Individual Browsing History”, Journal of the Japanese Society for Artificial Intelligence, Volume 18, 2003

  However, in the related art, when a word that is regarded as an important word appears multiple times in a document, a plurality of parts including the important word are extracted as important words. For example, let us consider a case where the word “cheap” is an important word when extracting an important word / phrase in sentence units. If the document being processed for search or browsing contains sentences such as "It was cheap", "There is a menu for children and it's cheaper", and "Special price even cheaper with coupons!" Since the word “cheap” is used for each sentence, all these sentences are extracted as important phrases. There are many types of important words, and as the number of important words increases, the number of important words extracted increases, and there is a possibility that most parts in the document are extracted as important words.

  However, in actuality, a statement such as “Special price even cheaper with a coupon!” Or “Low price of 800 yen for lunch” is more demanding that the user wants “Cheap” than a sentence that says “It was cheap”. It seems to contain a lot of information about. Whether such a sentence includes many user requests does not depend on the number of characters or words in the sentence. For example, the sentence "There is a menu for children and it's cheap" does not mean fewer letters and words than the sentences "Special price even cheaper with a coupon!" . However, the sentence “There is a menu for children and it's cheap” contains information about the requirements of “cheap” and “with children”, but it searches for documents with only the request of “cheap” Or it cannot be said that it is a sentence with much information for the browsing user.

  In addition, the method of extracting key words using word co-occurrence increases the co-occurrence with similar words when frequent words in the browsing history are synonyms. There is a possibility that the bias of the origin does not increase and is not extracted. For example, if the frequently used words in the browsing history include “price”, “price”, and “fee”, even if there is an important word such as “cheap” in the document to be processed, Since co-occurrence with frequent words is high, there is a high possibility that they are not extracted as important words.

  The present invention has been made in view of the above-mentioned facts, and it also supports the case where information related to a user's request is expressed in various words, and extracts important phrases that are deeply related to the user's request. An object of the present invention is to provide an important word / phrase extraction device, method, and program.

To achieve the above object, keyword extraction device of the present invention includes a receiving means for receiving each request word set representing each of the plurality of documents to be processed containing the word, and the user of the plurality of requests, the receiving Document dividing means for dividing the document received by the means into a plurality of divided word sets according to the length of the important words to be extracted, each of the words included in the divided word set divided by the document dividing means, Based on the similarity between words with each of the words included in the requested word set accepted by the accepting means, the maximum value of the similarity between word sets between the divided word set and each of the requested word sets , or weighted score, and the score calculating means for calculating as a score of the divided words each set, the calculated score is higher split word set than a predetermined threshold by said score calculating means, the document Extraction means for extracting a keyword included, is configured to include a.

According to the important word / phrase extraction apparatus of the present invention, the accepting unit accepts each of a processing target document including a plurality of words and a requested word set representing each of a plurality of user requests. Then, the document dividing unit divides the document received by the receiving unit into a plurality of divided word sets according to the length of the important phrase to be extracted, and the score calculating unit uses the divided word set divided by the document dividing unit. The maximum similarity between word sets between the divided word set and each of the requested word sets based on the similarity between words between each of the words included in the request word and each of the words included in the requested word set received by the receiving unit A value or a weighted score is calculated as a score for each divided word set. Then, the extraction unit extracts a divided word set whose score calculated by the score calculation unit is higher than a predetermined threshold as an important phrase included in the document.

  As described above, when the score based on the similarity between words between the requested word set representing the user request and the divided word set obtained by dividing the document is used, information regarding the user request is expressed in various words. It is possible to extract important words / phrases that are also closely related to the user's request.

Further, it is possible to request the User chromatography The even when there are multiple, and extracts a turn keyword from the portion that most represents a user request from the document.

Also, keyword extraction method of the present invention includes a receiving unit receives each request word set representing each of the plurality of word document to be processed containing, and user of the plurality of requests, document divider means, the accepting The document received by the means is divided into a plurality of divided word sets according to the length of the important phrase to be extracted, and the score calculating means uses the word set similarity between the divided word set and each of the requested word sets. Or the weighted score is calculated as a score for each divided word set, and the extraction means includes a divided word set whose score calculated by the score calculating means is higher than a predetermined threshold in the document. This is a method of extracting as an important phrase.

  Moreover, the important phrase extraction program of this invention is a program for functioning a computer as each means which comprises said important phrase extraction apparatus.

  As described above, according to the important phrase extracting device, method, and program of the present invention, the score based on the similarity between words between the requested word set representing the user's request and the divided word set obtained by dividing the document is used. Thus, it is possible to extract an important phrase that is compatible with a case where information related to the user's request is expressed in various words and has a deep relationship with the user's request.

It is the schematic which shows the structure of the keyword extraction apparatus which concerns on this Embodiment. It is explanatory drawing for demonstrating an example of the production | generation method of a request word set. It is explanatory drawing for demonstrating the calculation method of the score for every division | segmentation word set. It is a flowchart which shows the content of the important phrase extraction process routine in the important phrase extraction apparatus which concerns on this Embodiment.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

<Configuration of key word extraction device>

  The important phrase extracting device according to the present embodiment includes a CPU (Central Processing Unit), a RAM (Random Access Memory), and a ROM (Read Only Memory) storing a program for executing a later-described important phrase extracting process routine. It is comprised with the computer provided with. As shown in FIG. 1, this computer functionally includes a document reception unit 12, a user request reception unit 14, a setting reception unit 16, a morpheme analysis unit 20, a document division unit 22, a score calculation unit 24, And an important phrase extraction unit 18 including the extraction unit 26, an inter-word similarity determination database (DB) 28, and an output unit 30.

  The document receiving unit 12 receives a document (text data) to be processed, which is input by an input device such as a known keyboard, mouse, or storage device. The document to be processed can be a document of a web page that the user is currently browsing using a personal computer or a document at a link destination of a link attached to the document being browsed.

  The user request receiving unit 14 receives a user request input by the input device. The user request is a word set including one or more words representing a request that the user has when searching or browsing a document. The part of speech of each word included in the word set representing the user request is not limited. For example, when a user searches or browses a document with a request that “I want to save you a great deal”, the word “price” may be used as a user request, or two or more “price is cheap” A word set including words may be a user request. For example, the user request may be expressed by a word set including a synonym such as “price / price” or a related word such as “coupon sale”.

  The input method of the user request received by the user request receiving unit 14 is not particularly limited. The user request input from the input device by the user is received in advance, and the user request received in advance is applied to any document. Alternatively, a user request may be accepted for each document.

  Further, the user request accepting unit 14 may accept text data input in the form of a natural sentence by a user, and automatically extract a user request from the text data and accept it. For example, when text data in the form of a natural sentence “I want to save you a great deal” input by the input device is accepted, this text data is analyzed and a word that indicates a user request such as “Budget” Can be converted to a set.

  Further, the user request receiving unit 14 presents a plurality of user requests in a format such as a selection list to the user in order to save the user input, and receives the user request selected by the user. Also good. For example, as a word set representing a user request, words having similar meanings such as word set 1: “price price coupon sale”, word set 2: “child family parent-child baby baby food”, word set 3:. A plurality of sets are prepared and presented to the user as a selection list. When the user has a request for “I want to save you cheaply”, the user selects the word set 1 from the selection list, and the user request reception unit 14 selects the selected word set 1: “price price”. Coupon sale "can be accepted. Note that the number of words included in the word set representing each user request may be different.

  Further, the number of user requests is not limited to one and may be plural. In that case, a plurality of user requests, each represented by a word set as described above, are accepted. For example, if the user has a request that “I want to save you at a reasonable price” and a request that “I want something that is suitable for children”, “Price Price Coupon Sale” and “Children Family Parent and Child Baby Baby Food” Are accepted as two user requests. In addition, when the natural sentence is analyzed and converted into the user request as described above, for example, the text data “I want to save you money and are suitable for children” is analyzed. Can be converted into two word sets. Further, when the selection list is presented as described above, a plurality of selections from the selection list are permitted. For example, in the above example, when the user selects the word set 1 and the word set 2 from the selection list, the user request receiving unit 14 can receive the two word sets.

  Here, an example of a method for generating a word set representing a user request will be described. As shown in FIG. 2, a query log that is a word or a combination of words input by a user when searching for a document is used. Extract only queries that contain multiple words from the query log. For example, “Minato Mirai Lunch” and “Minato Mirai Lunch Coupon”. A plurality of words in which words having similar meanings and deep relationships are gathered by clustering the words included in such a query using the K-means method with the similarity between the words as a distance. Create a set.

  The word set generated in this way can be used when the natural sentence is analyzed and converted into a word set representing a user request or presented as a selection list. Note that the method of generating a word set representing a user request is not limited to the above example.

  The setting reception unit 16 receives information indicating various settings input by the input device. The various settings include settings relating to a unit of an important word to be extracted (for example, a word set having an arbitrary length such as a word, a sentence, and a paragraph) and a threshold value th used in a score calculation unit 24 described later.

  The important word / phrase extraction unit 18 divides the document to be processed in accordance with the unit of the important word / phrase to extract, and obtains a word set representing the user's request from the divided word sets. Hereinafter, each part of the important phrase extracting unit 18 will be described in detail.

  The morphological analysis unit 20 performs morphological analysis on the processing target document (text data) received by the document reception unit 12 by a conventionally known method.

As shown in FIG. 3, the document dividing unit 22 divides a document to be processed into a plurality of word sets l ₁ based on the analysis result of the morphological analysis unit 20 and the unit of the important word phrase extracted by the setting reception unit 16. , divided into ··· _{l h.} A divided word set (hereinafter referred to as "split word set") and the l _y w _{y1, ···,} are included t number of words w _yt. When the unit to be extracted is a word, each is regarded as a word set including one word.

The score calculation unit 24 acquires a word set (hereinafter referred to as “request word set”) a ₁ ,..., A _i representing the user request received by the user request reception unit 14. _Assume that a certain requested word set a _x includes k words of n _x1 ,..., N _xk . Score calculation unit 24 first obtains the divided word set _{l y} the requested word set _a similarity between the word _{n x1} included in _{x s (l y, n x1} ). The similarity s (l _y , n _x1 ) is, for example, the word n _x1 among the words w _y1 ,..., W _yt included in the divided word set l _y as shown in the following equation (1). And the average of the top j similarities in descending order. However, j is an arbitrary integer from 1 to _t, w _{y1, ···,} w the j number of words in the descending order of the degree of similarity among the _{w yt 'y1, ···, w} ' yj Distant, word _Assume that the similarity between w ′ _y1 and the word n _x1 is sim (w ′ _y1 , n _x1 ).

Then, according to the following expression (2), the similarity s between each divided word set l _y and each word n _x1 ,..., N _xk included in the requested word set a _x is averaged to obtain a divided word set l _y. And a similarity S (l _y , a _x ) between the request word set a _x and the requested word set a _x are calculated.

  Here, for the similarity between words, a concept base is used as the inter-word similarity determination DB 28. The concept base is a database of concept vectors in which the singular value decomposition is performed on the co-occurrence matrix that records the co-occurrence frequency of words in the corpus, and the words are expressed as vectors with reduced dimensions (Reference: Bessho) Katsuto, Kurosekura, Kataoka Ryoji, “Concept vector generation method based on co-occurrence of words and semantic attributes”, Japanese Society for Artificial Intelligence, 2006). The method of obtaining the similarity between words is not limited to a method using a concept base, and a method using a thesaurus or the like may be used.

Score calculation unit 24, if the user request is one, the calculated similarity S and split word set _l score _y score _{(l y).} Also, if the user request is two or more, required word set a ₁ representing each user _request, ..., and calculates each similarity between a _i and the divided word set l _y, the following equation (3) as shown in, the most value divided those large word set _l score _y score _{(l y).} Further, as shown in the following equation (4), a score obtained by adding a weight ω to each similarity between the divided word set l _y and each requested word set a ₁ ,..., A _i may be used as the score. Good. The weight ω is a weight that is inclined from the one having a large similarity between the divided word set l _y and each requested word set a ₁ ,..., A _i , or according to the strength of the user request. Or weights.

For example, if the user's request is “I want to know a shop that is cheap and has a beautiful scenery suitable for families,” the demand for money “cheap”, the facilities and atmosphere of the store “family oriented”, etc. A request and a request regarding a view of “beautiful scenery” are included. In the prior art, since the importance of words closely related to each request increases, the number of extracted important words increases as the number of user requests increases, and it may not make sense to extract. In this embodiment, as described above, the request word set a ₁ representing each user _request, ..., by calculating the score using the similarity between a _i and the divided word set l _y, most It is possible to extract as important words in order from a part including a lot of information and a part most representative of the user's request. For example, when there is a phrase including information about “cheap” and a phrase including information about “cheap” and “for families”, the latter is more likely to be extracted, and the user You can quickly find the part that contains more information about your request and get to the information.

Furthermore, the score calculation unit 24, a plurality of sentences and phrases in the divided word set l _y, if it contains such a word, obtains a respective score was subdivided into the unit, the product of the product and weighting the scores or the like may be used as the score of the divided word set l _y. The score calculation unit 24 calculates this score for all divided word sets.

  The extraction unit 26 extracts a divided word set whose score calculated by the score calculation unit 24 exceeds an arbitrary threshold th as an important phrase. The threshold th can be freely determined depending on how much important phrases need to be extracted. For example, it may be decided to extract the top k ranks of the score as important words or phrases, and the average value of the scores may be set to th so that words having a score higher than the overall average are extracted as important words. As the threshold th is increased, only the divided word set that best represents the information related to the user request in the document is extracted. As the threshold value is decreased, the divided word set having various information related to the user request is extracted. Many are extracted. An important word / phrase may be extracted by setting the same threshold th for any document, or may be extracted by adjusting the content, length, etc. of each document. Further, in addition to the threshold th, there may be a restriction that limits the number of words and characters to be extracted, and extracts important words / phrases within a range that falls within that range.

  The output unit 30 outputs an important word / phrase so that the important word / phrase extracted by the extraction unit 26 is highlighted and displayed in the document being browsed by the user.

<Operation of key word extraction device>

  Next, the operation of the keyword extraction device 10 according to the present embodiment will be described.

  When a user searches or browses a document with a personal computer or the like, when the user request and information indicating various settings are input by the user, the important phrase extracting device 10 executes the important phrase extracting process routine shown in FIG. Executed.

In step 100, the document receiving unit 12 receives a document (text data) to be processed, and the user request receiving unit 14 receives a request word set a _x (x = 1,..., I) representing a user request, Information indicating various settings including the unit of the important phrase extracted by the setting receiving unit 16 is received.

Next, in step 102, the morpheme analysis unit 20 performs morpheme analysis on the processing target document received in step 100. Next, in step 104, the document dividing unit 22 divides the document to be processed into a plurality of divisions based on the analysis result of the morphological analysis unit 20 in step 102 and the unit of the extracted important word accepted in step 100. Divide into word sets l _y (y = 1,..., H).

Next, in step 106, the score calculation unit 24, split word set _{l y} the requested word set _{a x} a similarity _{S (l} y, _{a x)} on the basis of each split word set _l score _y score ( l _y ).

  Next, in step 108, the extraction unit 26 extracts and outputs the divided word set whose score calculated in step 106 exceeds the threshold th as an important phrase, and ends the important phrase extraction processing routine.

  As described above, according to the important word / phrase extraction device according to the present embodiment, a space between words between a space representing a user's request (requested word set) and a portion obtained by dividing the document into predetermined units (divided word set). By using a score based on the distance (similarity) in the semantic space, it is possible to prevent a word having the same meaning from being extracted as in the prior art using the bias of co-occurrence and to be processed In this document, it is possible to cope with a case where information relating to a user request is expressed in various words. Also, by extracting the important words by comparing the score and the threshold, it is possible to prevent most of the document from being extracted as important words and to extract the important words that are closely related to the user's request. Can do.

  In addition, when there are a plurality of user requests, the maximum value of the similarity between the divided word set and the request word set representing each of the user requests or the weighted similarity is used as a score, thereby including the most information on the user request. It is possible to extract as important words in order from the part and the part that most represents the user's request, and it is possible to prevent most of the document from being extracted as important words.

  Further, in the conventional method, a user's browsing history is required, but in the present embodiment, information related to each user's request can be extracted even if there is no browsing history.

  The important phrases extracted in this way can be used, for example, when browsing or searching a document in a document on the web. When presenting a document to the user, it is possible to quickly find the information that the user wants by highlighting the extracted important phrases and presenting it to the user. Can be searched. In addition, it presents the extracted important words as a snippet of the search result of the document, or combines the important words extracted from multiple documents with regard to user requests from blogs and review articles written by various people. It is also possible to collect and present information. In particular, when searching for information using a mobile terminal such as a smartphone when going out, etc., it is troublesome to obtain information by browsing all the information of various documents. By extracting and presenting to the user, the user's trouble can be saved. Furthermore, when browsing not only documents on the web but also a large amount of accumulated document data, electronic books, etc., information related to the user's request can be efficiently obtained.

  Note that the present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  Moreover, although the above-mentioned important phrase extracting device has a computer system inside, the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used. .

  In the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium. The present invention can also be realized by installing a program on a known computer via a medium or a communication line.

<Example>

  Examples are shown below. Consider a case where the user has two requirements of “cheap and profitable” and “shop atmosphere”. The following was used as a request word set representing a user request.

a ₁ = {“set”, “coupon”, “cheap”, “discount”, “price”, “arrangement”, “fee”, “tour”, “cheap”, “price”, “cheap”, “free” “,“ Cheap ”,“ Package ”,“ Profit ”,“ Plan ”,“ Cost ”,“ High ”,“ Campaign ”,“ Free Plan ”}, k ₁ = 20
a ₂ = {“cute”, “sense”, “fashion”, “fashionable”, “flower cart”, “fashion”, “design”, “elegant”, “elegant”, “casual”, “chic”, “clothes "," Coordinate "," Sexy "," Cute "," Accessory "," Fashion "," Brand "," Cute "," Fashion "}, k ₂ = 20

Moreover, the unit of the important phrase to be extracted is a sentence unit, and as the word included in the divided word set, the word whose part of speech is a noun or adjective is used among the words included in the sentence which is one unit of the divided word set. Here, j in the formula (1) is set to j = 2, and since the score value is small, a logarithmic value is used. The score calculation method will be described using the following sentence (divided word set l ₁ ).

l ₁ = “On the day of lunch and pasta…”
The words of the above-mentioned conditions to be included in the divided word set l _y that is,
w ₁ = {“Lunch”, “Pasta”, “Day”}
And t = 3. The similarity between this divided word set l ₁ and the following request word set a ₁ is obtained.

First, the similarity between the word n ₁₁ = “set” included in the requested word set a ₁ and the divided word set l ₁ is obtained. At this time,
sim (“Lunch”, “Set”) = 0.0667
sim ("pasta", "set") = 0.0315
sim (“day”, “set”) = 0.0570
So that
w ′ ₁ = {“Lunch”, “Sun”}
It becomes. Using this w ′ ₁ , the similarity between the word n ₁₁ = “set” included in the requested word set a ₁ and the divided word set l _{1 according} to the following formula (5) in which j = 2 is applied to the formula (1) Find the degree.

This is calculated for all words included in the requested word set a ₁ and averaged. That is, the similarity S (l ₁ , a ₁ ) between the divided word set l ₁ and the requested word set a ₁ is calculated by the following formula (6) in which k = k ₁ = 20 is applied to the formula (2).

Similarity S (l ₁ , a ₂ ) is similarly calculated for request word set a ₂ representing another user request, and the logarithm of the largest similarity is set as score score (l ₁ ) of divided word set l _1. did. In the following output example, the threshold th is determined by the following equation (7).

  Here, thlv is a parameter for adjusting how many important phrases are extracted. In the following output example, a sentence with a circle mark at the beginning of the sentence is a sentence extracted when thlv = 0.85 and the threshold value th is set higher. A sentence with a triangle mark at the beginning of the sentence is a sentence extracted in addition to the sentence marked with a circle when thlv = 0.6 and the threshold value th is set low. It can be seen that a sentence that represents the user's request is selected as the sentence with a circle, and a sentence that is related to the user's request is selected as a sentence with a triangle. The number in parentheses at the end of the sentence represents the score. For sentences whose part of speech does not include noun or adjective words, the score cannot be calculated under the conditions of this example, so “null” is added to the score and it is excluded from extraction.

(Example output)
・ On the day of pasta at lunch ... (-85.4484637354695)
・ ABC Cafe (-88.0978786776667)
○ This is a reasonable and easy Italian lunch shop (-34.9677383949862)
○ It looks like a stylish cafe (-33.9715381578441)
○ Lunches are always lined up and everyone sits in a chair outside the store and waits, so they want to enter early or late lunch after 12-13 o'clock (-43.8880971908371)
○ The shop was bright and lively (-49.225867499909)
△ Here at lunchtime, pasta is served with bread, salad and drink (-60.503565658478)
△ Besides, it is free of charge (-63.4664379716312)
・ That ’s right, old and young .... (-98.540852414816)
・ Bread and salad (-184.423315416741)
・ Salad dressing is also delicious (-76.8044815622487)
・ I got it (null)
・ Drinks are orange juice (-94.2192436418884)
・ Pasta is oil base of tuna and fresh tomato (-66.83119207037)
・ <It is large> (-184.613623137593)
・ I was satisfied because I was hungry (-70.390793748566)
・ The taste is normally delicious (-72.6345729746803)
○ It is worth mentioning that this volume and contents are priced at 950 yen (-38.5748480842723)
・ Moreover, it seems to change by about 5 seasons (-141.449011600035)
△ Moreover, there is a stamp card for lunch (-60.3557579207)
△ Collect 10 and get 1 free lunch! I want to go! ! (-58.9474098383104)
△ At night, I feel a little more relaxed (-59.2116103942359)
△ May be good for a date (-61.5268202682332)

  The extracted sentence (divided word set) may differ depending on the number of words used as the requested word set representing the user's request or by adding a new word. Moreover, as a value used as the similarity, a concept vector is used this time, but the extracted sentence may be different by changing this.

DESCRIPTION OF SYMBOLS 10 Important phrase extraction apparatus 12 Document reception part 14 User request reception part 16 Setting reception part 18 Important phrase extraction part 20 Morphological analysis part 22 Document division part 24 Score calculation part 26 Extraction part 30 Output part

Claims (3)

Receiving means for receiving each of a processing target document including a plurality of words and a request word set representing each of a plurality of user requests;
A document dividing unit that divides the document received by the receiving unit into a plurality of divided word sets according to the length of an important word to be extracted;
Based on the word similarity between each of the words included in the divided word set divided by the document dividing unit and each of the words included in the requested word set received by the receiving unit, the divided word set and the Score calculating means for calculating a maximum value of similarity between word sets with each of the requested word sets , or a weighted score as a score for each of the divided word sets;
Extraction means for extracting a segmented word set whose score calculated by the score calculation means is higher than a predetermined threshold as an important phrase included in the document;
An important word / phrase extraction device. Accepting means accepts each request word set representing each of the plurality of documents to be processed containing the word, and the user of the plurality of requests,
The document dividing means divides the document received by the receiving means into a plurality of divided word sets according to the length of the important word to be extracted,
The score calculation means is based on the inter-word similarity between each of the words included in the divided word set divided by the document dividing means and each of the words included in the requested word set received by the receiving means. Calculating a maximum value of similarity between word sets between each divided word set and each of the requested word sets , or a weighted score, as a score for each divided word set;
An important phrase extraction method, wherein the extraction means extracts a divided word set whose score calculated by the score calculation means is higher than a predetermined threshold as an important phrase included in the document. Computer, keyword extraction program for causing to function as each means constituting the keyword extraction device according to claim 1 Symbol placement.

Images