JPH08329106A - Method for generating dictionary, method for constructing word set, method for constructing document set, and conception supporting system using these methods - Google Patents

Abstract

PURPOSE: To obtain a dictionary more appropriate to each given text by extracting a specific character stirring from plural character strings appearing in a text and generating a dictionary. CONSTITUTION: This dictionary generating method generates a dictionary by extracting a character string R satisfying a condition including the character string R and excluding a character string S appearing at the same appearance frequency as that of the character string R and longer than the character string R from a text stored in a data base consisting of one or more texts. In the case of executing the method by a dictionary generating means 2, each text in the data base is each document stored in an information storing means 1. A keyword based upon a structured word set can be automatically added to the dictionary and relation between documents can be presented based upon an automatically structured document set.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to automatic classification of document information and visualization of an information space of document information, and more particularly, to a dictionary generation method for automatically generating a dictionary from each word included in a document, and a word set. Structured word set structuring method and document set structuring method automatically structuring the document set, and the information space of the document set by visualizing words and documents by associating these methods The present invention relates to a bottom-up type idea support system that provides an environment that looks freely.

[0002]

2. Description of the Related Art An idea support system is a system that provides an environment in which information such as words, images, and sounds can be handled in order to support human intellectual creative activities. As a conventional idea support system that handles words and documents, there is a system that collects ideas given by a user as keywords and arranges them in space to encourage the user to further think. In addition, there is a device that associates a thesaurus prepared in advance and prompts the user to come up with an idea. However, in each case, the word level of the keyword is used to prompt the user to come up with an idea, and the processing is not performed in correspondence with the text database.

As a concrete example of a conventional idea support system,
For example, Journal of Artificial Intelligence, Volume 9, No. 1 (1994 1
Moon, published by Japan Society for Artificial Intelligence), P. 139-P. 1
47, "Thinking support system by arranging text objects in space". This system is to write ideas and text fragments that come to the mind of users as virtual cards in a space.

Next, the operation will be described. the system,
First, input ideas or document fragments that come to the mind of the user and arrange them in space as a virtual card called a text object, specifically on the screen. At this time, what is displayed in the space is not the text of the document but the name given to each text object by the user. Also, the user subjectively attaches some keywords to the text object. Next, the system reconstructs the space in consideration of the relation between the text objects and presents the reconstructed space to the user. The system determines the degree of association between any two text objects by using the degree of keyword sharing. That is, the system increases the relevance accordingly if any two text objects share the same keyword.

The user can see the reconstructed space and find the meaning of the axis of the space or the cluster. The discovery will stimulate new ideas for system users.

[0006]

Since the conventional idea support system is configured as described above, it has the following problems. (1) The user has to add keywords to all texts, which is inconvenient. (2) Since the keywords are determined by the subjectivity of the user, the keywords are not controlled and the keywords are likely to be unrelated to each other. Further, if there are many text objects, it becomes difficult to match the keywords. (3) Since there is no text object search function, it is difficult for users to grasp the mutual relationship. (4) The text object is composed only of the memo created by the user. Therefore, the subject matter of the idea is limited.

The present invention has been made in order to solve the above-mentioned problems, and it is possible to automatically attach a keyword with a structured word set to a document.
It is an object of the present invention to obtain an idea support system that can present the relationships between documents based on automatically structured document sets. Another object of the present invention is to obtain a dictionary generation method, a word set structuring method, and a document set structuring method suitable for being applied to such an idea support system.

[0008]

According to a first aspect of the present invention, there is provided a dictionary generating method, wherein a frequency of appearance of a character string in one or more texts is calculated, and the appearance frequency is selected from the character strings appearing in the text. Extract a character string that meets the condition that it does not occur less frequently than the character string, and that there is no character string that contains the character string and is longer than the character string, and adds the extracted character string as an item to the dictionary. Is generated.

A word set structuring method according to a second aspect of the present invention obtains the latest low-order word and the latest high-order word of a word, associates the word with the latest low-order word and the latest high-order word, and determines a predetermined one of these relationships. The word network is generated after excluding associations that meet the conditions.

In the word set structuring method according to the third aspect of the present invention, in a set of texts having a headline and a body, the word KW1 included in the headline of the text DOC1 is included in the body of the text DOC2 and the word KW1 of the text DOC2 is included. If the word KW2 in the heading is included in the body of the text DOC3, the word KW1 is included in the body of the text DOC3, and the word KW3 is included in the heading of the text DOC3, the three words KW1, KW2, K
A basic chain of word KW1 → word KW2 → word KW3 is generated by W3, and basic chains having a common part among the generated basic chains are combined to generate a word network.

In the word set structuring method according to the invention of claim 4, the word network in the invention of claim 2 is the first word network, and the word network in the invention of claim 3 is the second word network. Then, the first word network and the second word network are integrated using the designated word as a key.

According to a fifth aspect of the present invention, there is provided a document set structuring method for converting each text into a word vector, each of which includes a plurality of texts. A plurality of clusters set to are set, and a word representative of each cluster is obtained.

In the document set structuring method according to the sixth aspect of the present invention, when setting a plurality of clusters, the centroids of all word vectors are obtained, and the vector farthest from the obtained centroids is set as the first cluster. For each word vector, find the distance to the nearest cluster center for each word vector, and add the word vector with the maximum distance as the cluster center, classify each word vector into the cluster with the closest cluster center, and include the words in each cluster. The operation with the average of the vectors as the new cluster center is continued until the cluster center does not change.

According to a seventh aspect of the present invention, there is provided a method of structuring a document set, which extracts a large number of predetermined values from the elements of each word vector at the center of a cluster, and extracts words corresponding to each extracted word vector. It is defined as a word representing a cluster.

The idea generation support system according to the invention described in claim 8 is a dictionary generation means for generating a dictionary by the method according to claim 1 for the text stored in the information storage means, and a dictionary generated by the dictionary generation means. Is used as a word set, the word set structuring means for structuring the word set by the method according to claim 4, the information selecting means for selecting a predetermined text from the text set in the information storing means, and the text set in the information storing means. And a text set structuring means for structuring the text set selected by the information selecting means by the method according to claim 7, the processing result of the word set structuring means, and the respective processing results of the document set structuring means. And an information space display means for controlling the information space.

In the idea generation support system according to the ninth aspect of the present invention, when the search keyword is input, the information selecting means selects each text based on the input keyword, and the document group structuring means selects the information. It is configured such that the document set is structured for each text selected by the means.

In the idea generation support system according to the tenth aspect of the present invention, when the information space display means requests a related display, the processing result of the word set structuring means and the text set in the information storage means are displayed. On each display screen of the processing result of the document set structuring means and the processing result of the document set structuring means for the text set selected by the information selecting means, the element corresponding to the request for the related display is highlighted. It has become.

[0018]

In the step of extracting the character string in the invention described in claim 1, expressions (eg, "Hana no Expo") that are not found in ordinary dictionaries and expressions depending on the type of text (eg,
When a newspaper article is targeted, it is possible to automatically extract a meaningful character string from a given text, such as extracting “Gorbachev”, which means President Gorbachev.

In the step of generating a word network in the invention according to claim 2, it is possible to automatically extract upper / lower words of an arbitrary word and provide an appropriate search word that can be a keyword for information search. To do.

The step of generating a word network in the invention according to claim 3 makes it possible to automatically extract a related word of an arbitrary word and provide an appropriate search word that can be a keyword for information search.

The step of integrating with the word network in the invention according to claim 4 is to carry out arbitrary words and their upper /
It is possible to automatically connect a subordinate word and a related word, and to provide a more appropriate search word that can be a keyword for information search.

The step of setting a plurality of clusters according to the fifth aspect of the invention makes it possible to automatically classify the text set according to the content. Also, the step of finding words representative of clusters makes it possible to provide suitable words representative of each group after classification.

The step of adding the word vector having the maximum distance as the cluster center in the invention described in claim 6 enables each cluster to be effectively arranged in a discrete manner. Further, the step of continuing the operation with the average of the word vectors included in the cluster as the new cluster center until the cluster center does not change makes it possible to optimize the configuration of the cluster including each word vector.

In the step of obtaining a word representing a cluster in the invention described in claim 7, a word having a higher degree of importance in the text corresponding to each word vector forming the cluster is selected as a word representing the cluster. To enable that.

The information space display means according to the invention of claim 8 presents the user with the relevance of each word in the text and presents the classified texts to the user.

The information space display means in the invention according to claim 9 presents to the user each text classified based on the keyword input by the user.

The information space displaying means in the invention according to the tenth aspect makes it easy to display the relevance of the word designated by the user to other words and the relevance of the word designated by the user to each text. To do.

[0028]

【Example】

Example 1. FIG. 1 is a block diagram showing the configuration of an idea generation support system according to an embodiment of the present invention. In the figure, 1
Is an information accumulating means for accumulating document information, 2 is a dictionary generating means for generating a dictionary from the document information accumulated in the information accumulating means 1, and 3 is a word for performing a process for structuring a word set by each word in the document information. A set structuring means, 4 is an information selecting means for selecting the document information of the information accumulating means 1 according to a user's instruction, 5 is a selection information accumulating means for accumulating the document information selected from the information accumulating means 1, and 6 is A document set structuring unit that performs a process of structuring the document set of each document in the information storage unit 1 and the document set of each document in the selection information storage unit 5,
Reference numeral 7 denotes a structured word set accumulating means 8 for accumulating the processing result of the word set structuring means 3 and a structured document set accumulating means 9 and a structured selection document set accumulating means 10 for accumulating the processing result of the document set structuring means 6. And an information space display means for presenting the information space for the idea support to the user.

When the idea generation support system is realized by a computer, the information storage means 1, the selection information storage means 5,
Structured word set storage means 8 and structured document set storage means 9
The structured selection document set accumulating unit 10 is realized by the storage unit in the computer system. The dictionary generating means 2, the word set structuring means 3 and the document set structuring means 6 can be realized by software. The information selecting means 4 can be realized by an input unit and software in a computer system. The information space display means 7 can be realized by a display unit and software in a computer system.

FIG. 2 is a flow chart showing a dictionary generating method according to the first embodiment of the present invention. This dictionary generation method can be applied to the dictionary generation means 2 in the system shown in FIG. Of course, this dictionary generation method can be applied to other systems.

Next, the operation will be described. This dictionary generation method is a character string S that includes a character string R and appears at the same frequency as the appearance frequency of the character string R in a text in a database including one or a plurality of texts, and is longer than the character string R. This is a method of generating a dictionary by extracting a character string R that satisfies the condition that the character string S does not exist. When the dictionary generation means 2 in the system shown in FIG. 1 executes this method, each text in the database is each document stored in the information storage means 1. In that case, the flowchart shown in FIG. 2 shows the processing executed by the dictionary generation means 2.

A temporary file T-FI is used to generate the dictionary.
Prepare areas for LE, dictionary DICT, temporary dictionary T-DICT, and variable I. First, the temporary file T-FILE, the dictionary DICT, and the temporary dictionary T-DICT are set to the "empty" state. Also, the length L of the text is determined (step ST130). The text length is the number of characters of the text being handled or a length less than that. Then, the value L is substituted for the variable I (step ST131). The temporary file T-FILE temporarily stores each partial character string that appears in each text and the number of appearances thereof during execution of the process of searching for a character string that satisfies the above conditions. The temporary dictionary T-DICT temporarily stores each partial character string that satisfies the above condition for the value I of each length and the number of appearances thereof while executing the process of searching for a character string that satisfies the above condition.

Thereafter, the processes of steps ST132 to ST136 are repeated until the value of the variable I becomes "1". Therefore, first, it is checked whether or not the value of the variable I is "1", and if it is not "1", a process of searching for a character string satisfying the above conditions is performed (step ST132). In the process of searching for a character string satisfying the above conditions, first, the number of appearances of all partial character strings having a length of I, which is a portion in each text, is counted, and each partial character string is displayed together with the number of appearances in a temporary file T-FILE. To store (step ST13
3).

Next, all the character strings set in the dictionary DICT are compared with each partial character string in the temporary file T-FILE. Then, among the partial character strings in the temporary file T-FILE, those that are not the partial character strings of the character string in the dictionary DICT, and the partial character strings of the character string in the dictionary DICT, Those that exceed the frequency of appearance of character strings in the dictionary DICT are recorded in the temporary dictionary T-DICT (step ST13).
4). In the first process, that is, when I = L, the dictionary DICT is “empty”, so the temporary file T-FILE
All substrings inside are recorded in the temporary dictionary T-DICT.
Then, all the character strings in the temporary dictionary T-DICT are added to the dictionary DICT (step ST135).

Next, the temporary file T-FILE and the temporary dictionary
The T-DICT is set to the “empty” state, the value of the variable I is decremented by 1, and the process returns to step ST132 (step ST13).
6). Then, the processes of steps ST132 to ST136 are repeated until I = 1. When I = 1, the process ends. At this time, the dictionary DICT is an automatically generated dictionary composed of character strings satisfying the above conditions.

In the above processing, the dictionary DICT is first
That is, although it is set to "empty" in step ST130, a general dictionary may be set. In that case, in addition to the contents of the general dictionary, the character string automatically extracted as described above is set in the dictionary. Also, first, set the dictionary created for a certain text as the dictionary DICT,
You may extract a new string from another text and add it to the dictionary DICT. Then, by setting the dictionary thus created as the dictionary DICT in step ST130 and repeating the above processing, the dictionary can be generated step by step.

Example 2. 3 and 4 are flowcharts showing the process of creating the first word network in the word set structuring method according to the second embodiment of the present invention. The processing by this word set structuring method can be applied in the word set structuring means 3 in the system shown in FIG. Of course, this process can be applied to other systems.

According to this method, it is possible to obtain the first word network that reflects the relationship between the most recent lower word and the most recent upper word of a word A. The nearest lower word of a word A is a word B including the word A. However, the word C including the word B is not the closest to the word A. The most recent superordinate word of a certain word A is each word Dn forming the word A.
However, the Dnn forming the word Dn is not the closest to the word A. For example, “nuclear safety” is the most recent subordinate word of the word “safety”, but “nuclear safety white paper” is not the most recent subordinate word of the word “safety”. Although "16 bits" is the most recent broader word of the word "16-bit personal computer", "bit" is not the most recent broader word of the word "16-bit personal computer".

Next, the operation will be described. When the word set structuring means 3 in the system shown in FIG. 1 executes this method, the document stored in the information storage means 1 or the dictionary generated by the dictionary generation means 2 is targeted, and the latest lower word and the most recent word are used. The high-order word is extracted.

First, referring to the flowchart of FIG. 3, the word WO in the process of generating the first word network.
The process of extracting the immediately lower word of RD will be described. Word WORD, list WORDS, area WO to extract the most recent lower word
Areas of RD2, list WORDS-KAI, list WORDS-KAI-SAKUJO and list 1-LEVEL-KAI are prepared. When the word set structuring means 3 in the system shown in FIG. 1 executes this method, the word WORD is a word specified by the user or each word in the list, and the word from which the latest subordinate word is extracted. The list of is a dictionary created by the dictionary creating means 2, for example.

First, the word list is substituted into the list WORDS (step ST51). Next, all words including the word WORD as a character string are selected from the list WORDS, and a list WORDS-KAI including the selected words is created (step ST52). For example, when the word WORD is “safe”, the list WORDS-KAI becomes [“Safety Center”, “Nuclear Safety”, “Nuclear Safety White Paper”, ...].

Then, a process of deleting lower-order words other than the latest lower-order word from the list WORDS-KAI is performed. Therefore, first, the list WORDS-KAI-SAKUJO is set to the "empty" state (step ST53). Next, set one word in the list WORDS-KAI to the area WORD2. Then, it is checked whether the word set in the area WORD2 includes the word existing in the list WORDS-KAI. If included, list the word WO
Add to RDS-KAI-SAKUJO (step ST54). For example, the “nuclear safety white paper” contains the string “nuclear safety” in the list WORDS-KAI-SAKUJO, so the list WORDS-
Added to KAI-SAKUJO. And this process, list
Execute for all words in WORDS-KAI.

Then, from list WORDS-KAI to list WORD
Excluding words in S-KAI-SAKUJO, then list WORDS-KAI
Each word in the list is added to the list 1-LEVEL-KAI (step ST55). As described above, the most recent subordinate word is a list
Set to 1-LEVEL-KAI. For example, the list 1-LEVEL-KAI about the word "safety" becomes ["safety center", "nuclear safety", ...].

Further, from the list 1-LEVEL-KAI, a process of marking a word which may not be said to be the most recent subordinate word is performed. Here, the rule for finding a false upper / lower relationship is applied. The following are examples of rules. That is, among the words belonging to list 1-LEVEL-KAI, if the character string of the difference from the word set in word WORD is one alphabet, one katakana or one hiragana, A high-order / low-order relationship is marked (step ST56). This rule is a heuristic rule that excludes, for example, those that do not clearly have a superior / subordinate relationship in meaning, such as "Aji" of fish and "Asia" of neighboring countries.

Even when the difference from the word set in the word WORD exceeds one character, the difference character string is
Gong (“yu”, “u”, etc.) and consonant (“tsu”, “tsu”)
There is a false superior / inferior relation when it begins with a character that indicates a syllable (“n”, “n”) or a long sound. For example, "chart"
Since the character string of the difference between "quartet" and "quartet" is "t", there is a false upper / lower relationship. Since the character string of the difference between "combi" and "corned beef" is "oo", there is a false upper / lower relationship. You may further apply such a rule.

Next, with reference to the flow chart of FIG. 4, the process for obtaining the most recent upper word of the word WORD will be described. Word WORD, list WORDS to extract the most recent upper word
, Area WORD2, list WORDS-JOUI, list WORDS-JOUI-
Prepare areas for SAKUJO and Listing 1-LEVEL-JOUI.
When the word set structuring means 3 in the system shown in FIG. 1 executes this method, the word WORD is a word specified by the user or each word in the list, and the word from which the most recent superordinate word is extracted. The list of is a dictionary created by the dictionary creating means 2, for example.

First, the word list is substituted into the list WORDS (step ST51). Next, the words included in the word WORD as a character string are selected from the list WORDS, and a list WORDS-JOUI composed of the selected words is created (step ST
58). For example, when the word WORD is "16-bit personal computer", the list WORDS-JOUI becomes ["16-bit", "bit", "personal computer"].

Then, the process of deleting the upper words other than the most recent upper word from the list WORDS-JOUI is performed. First, the list WORD
Put S-JOUI-SAKUJO in the "empty" state (step ST5
9). Next, put one word in the list WORDS-JOUI into the region WO
Set to RD2. Then, it is checked whether or not the character string of the word set in the area WORD2 is included in the words existing in the list WORDS-JOUI. If it is included, the word in the area WORD2 is added to the list WORDS-KAI-SAKUJO (step ST60). For example, "bit" is the list WORDS-JO
As it is included in "16 bits" existing in the UI, list
Added to WORDS-KAI-SAKUJO. And this process
Execute for every word in the list WORDS-JOUI.

Next, from list WORDS-JOUI to list WORD
S-JOUI-SAKUJO Excluding words in list WORDS-JO
Each word in the UI is added to the list 1-LEVEL-JOUI (step ST61). As described above, the most recent upper word is set in the list 1-LEVEL-JOUI. For example, the list 1-LEVEL-JOUI for the word "16-bit personal computer" is ["16
Bit "," PC "].

Further, from the list 1-LEVEL-JOUI, a process of marking those that may not be said to be the most recent superordinate word is performed. Here, the rule for finding a false upper / lower relationship is applied. The following are examples of rules. That is, among the words belonging to list 1-LEVEL-JOUI, if the character string of the part that differs from the word set in word WORD is one alphabet, one katakana or one hiragana, The upper / lower relation of is marked (step ST62). This rule is a heuristic rule that excludes, for example, sickness “cancer” and country name “Uganda” that do not have a significant upper / lower relationship in meaning.

As described above, it is possible to obtain the upper / lower relation for each word. The relationship between a word and its nearest lower-ranked word is called the nearest lower-ranked relationship, and the relationship between a word and its nearest higher-ranked word is called the nearest higher-ranked relationship. The one including each link is expressed as a first word network. In the first word network, the relation marked with a false upper / lower relation can be distinguished in link type as necessary.

When the first word network is applied to the system shown in FIG. 1, the word set structuring means 3 creates the first word network and stores it in the structured word set storage means 8. . The information space display means 7 can also visually display the first word network in response to the user's request. In that case, the upper / lower relationship for each word is displayed as the first word network.

FIG. 5 is an explanatory diagram showing an example of the first word network. It is possible to increase the recall rate of information retrieval in the idea generation support system by using the first word network. For example, if a user wants to retrieve a document regarding "kidney transplantation" from a certain database, if the user refers to the first word network created based on the database, the character string "kidney transplantation" and the character string "kidney" can be referred to. It is understood that the keywords are "transplant" and the character string "kidney transplant". If the first word network is not provided, only the character string “kidney transplant” is used as a keyword, and the recall of the search may be deteriorated. Here, the recall ratio of the search is (the number of documents that the user desired in the retrieved documents) ÷ (the number of documents that should be provided to the database user).

Further, the accuracy of information retrieval in the idea generation support system can be improved by utilizing the first word network. For example, if a user wants to retrieve a document concerning a microorganism "virus" from a database, if the user refers to the first word network created based on the database, the character string "virus" is used as a keyword, and It turns out that the character string "computer virus" should be an exclusive keyword. If the first word network is not provided, computer-related documents may be extracted by using only the character string “virus” as a keyword. That is, the accuracy of the search may be deteriorated. Here, the accuracy of the search is
(Number of documents that the user desired in the retrieved documents) ÷ (Number of retrieved documents).

Next, the second word network in the word set structuring method will be described. The processing by this word set structuring method can be applied in the word set structuring means 3 in the system shown in FIG. of course,
This process can be applied to other systems. The second word network is configured as follows.

The keyword KWj is added to the heading of the document DOCi.
Is included is represented as KWj (DOCi). Although one word may include a plurality of meanings, the keyword KWj here means a meaning used in the context of the document DOCi.

The fact that the keyword KWk included in the headline of the document DOCm is included in the body of the document DOCi and the keyword KWj is included in the headline is expressed as KWk (DOCm) => KWj (DOCi). This relationship is called a co-occurrence relationship. Here, as shown in FIG. 6, when there is a relationship of KW1 (DOC1) => KW2 (DOC2) KW2 (DOC2) => KW3 (DOC3) KW1 (DOC1) => KW3 (DOC3), KW1, KW2, I think that KW3 has an associative relationship. If it is considered that KW1, KW2, and KW3 are in an associative relationship only by the occurrence of the co-occurrence relationship between the first item and the second item, it is not possible to exclude the case where these two terms happen to cause the co-occurrence relationship. Therefore, the third item is also considered. The co-occurrence relation of the above 3 terms is expressed as
OC3). This notation is the basic chain of association.

When there is an intersection between some basic chains, these basic chains are combined to form a network. This network is the second word network. For example, KW1 (DOC1) → KW2 (DOC2) → KW3 (D
OC3) KW2 (DOC2) → KW3 (DOC3) → KW4 (D
OC4) KW1 (DOC1) → KW2 (DOC2) → KW5 (D
The three basic chains of OC5) are combined as follows.

[0059]

[Table 1]

Further, when DOCi and DOCj are the same type of document, KWk (DOCi) and KWk (D
And OCj). For example, KW1 (DOC1) → KW2 (DOC2) → KW3 (D
OC3) KW4 (DOC4) → KW1 (DOC6) → KW5 (D
OC5) and DOC1 and DOC6 are the same type of document,

[Table 2] To generate a second word network. The same type of document means a document having the same theme, and is, for example, each document to which the user gives the same index, or each document which is classified into the same cluster by the automatic classification described later.

In summary, as shown in FIG. 6, in a set of documents having a headline and a body, the word KW1 included in the headline of the document DOC1 is included in the body of the document DOC2 and the word KW2 in the headline of the document DOC2. Document DO
The word KW1 contained in the body of C3 and the word KW1 is document DOC3
, And the word KW3 is included in the heading of the document DOC3, the three words KW1, KW
2, word KW1 → word KW1 → word KW2 → word KW
A second word network is generated by constructing three basic chains and further combining each basic chain.

For example, as newspaper article data, Freon (DOCa), carbon dioxide (DOCb), greenhouse effect (DO)
Cc), warming (DOCd), environmental diplomacy (DOCe), CFCs (DOCa) → carbon dioxide (DOCb) → greenhouse effect (DOCc) carbon dioxide (DOCb) → greenhouse effect (DOCc) → warming (DOCd) Greenhouse effect (DOCc) → warming (DOCd) → environmental diplomacy (DOCe) carbon dioxide (DOCb) → freon (DOCa) → global warming (DOCd) If there are basic chains,

[Table 3] A second word network such as In addition,
Here, the document ID is omitted.

As the newspaper article data, Freon (DOC)
a), carbon dioxide (DOCb), greenhouse effect (DOCc),
There are global warming (DOCd), refrigerator (DOCf), ozone layer (DOCg), and freon (DOCh). Carbon dioxide (DOCb) → greenhouse effect (DOCc) → warming (DOCd) freon (DOCa) → greenhouse effect (DOCc) ) → Global warming (DOCd) Refrigerator (DOCf) → Freon (DOCh) → Ozone layer (DOCg) There are basic chains, and document DOCa and document D are available.
If OCh is the same type of document,

[Table 4] A second word network such as

When the second word network is applied to the system shown in FIG. 1, the word set structuring means 3 creates the second word network and stores it in the structured word set storage means 8. . The information space display means 7 visually displays the second word network in response to the user's request. That is, the association relation for each word is displayed as the second word network. The displayed second word network can be applied to the knowledge extraction support function in the idea generation support system. The relationship of concepts (words) that the user is interested in is obtained by searching a path between the items (words) of interest in the presented second word network. For example, "refrigerator" and "warming"
For the relationship with the above, refer to the second word network presented as in (1) above, and read, "A refrigerator that uses CFCs as a refrigerant and that CFCs diffuse into the atmosphere The effect warms the earth.
The greenhouse effect is also caused by carbon dioxide. Can be recalled. Furthermore, the information space display means 7 can also display a combination of the first word network and the second word network.

Next, referring to the flowchart of FIG.
A specific process for creating the word network will be described. First, the headline portion of all documents is divided into words (step ST110). Of the obtained words, those that are different from each other are extracted. Then, each extracted word is set as KEYWORDi (i = 1 to p) (step S
T111). It is assumed here that p different words are extracted. In the example shown in FIG. 6, p = 3. That is, KEYWORD1 = KW1, KEYWORD2 = K
W2, KEYWORD3 = KW3.

Then, for each of the KEYWORDi, the document including it in the index section is searched. The found document is set as IDMij (j = 1 to q) (step ST112). Here, it is assumed that q documents are found. In the example shown in FIG. 6, for i = 1, q =
1, q = 1 for i = 2 and q = 1 for i = 3. That is, IDM11 = DOC1, IDM21 = D
OC2 and IDM31 = DOC3.

Further, for each of the KEYWORDi, the document including it in the text is searched. Let the found document be IDHik (k = 1 to r) (step S
T113). Here, it is assumed that r documents have been found. In the example shown in FIG. 6, r = 2 for i = 1,
For i = 2, r = 1. That is, IDH11 =
DOC2, IDH12 = DOC3, IDH21 = DOC
It is 3.

Further, for each of KEYWORDi, j satisfying IDHib = IDMja is obtained (step ST114). Here, a and b are indefinite values.
In the example shown in FIG. 6, IDH11 = DOC2, IDM21
= DOC2, and IDH12 = DOC3, IDM31
= DOC3, for i = 1, j = 2,3
Is. IDH21 = DOC3, IDM31 = DOC3
Therefore, for i = 2, j = 3.

Then, for the obtained i, j, ID
A k that satisfies Hid = IDMkc and IDHje = IDMkc is obtained (step ST115). Where d,
c and e are indefinite values. In the example shown in FIG. 6, i = 1, j
= 2, there exists k = 3 that satisfies both expressions, and i = 1,
When j = 3 and when i = 2 and j = 3, there is no k satisfying both expressions. That is, i = 1, j = 2, k
= 3, both equations are satisfied.

Then, KEYWORDi (*) → KEY
WORDj (IDMja) → KEYWORDk (IDM
Let kc) be a basic chain (basic fragment) of an associative relationship (conceptual chain). In the example shown in FIG. 6, i = 1, j = 2, k =
Therefore, after all, KEYWORD1 (*) → KEYWORD2 (IDM2
a) → KEYWORD3 (IDM3c) That is, KW1 (DOC1) → KW2 (DOC3) →
The basic chain of KW3 (DOC3) is obtained. In addition,
(*) Indicates a document compatible with KEYWORDi.

Further, if there is a partial chain match among the basic chains obtained in step ST116, they are combined (step ST11).
7). This process is executed for all basic chains.

Although the first word network and the second word network are obtained by the processing shown in FIGS. 3, 4 and 7, they are integrated using the designated word as a key to structure the word set. Is completed. In the system shown in FIG. 1, the word set structuring unit 3 integrates the first word network and the second word network centering on the word designated by the user. FIG. 8 is an explanatory diagram showing an example of the processing result by the word set structuring method. In the system shown in FIG. 1, the word set structuring means 3
Performs such structuring, and the information space display means 7 is shown in FIG.
Visually display as shown in.

In the example shown in FIG. 8, when the word "greenhouse effect" is designated, the links of the first word network indicated by dotted lines, the most recent upper / lower words, and the solid line are shown. It is an example in which links of the second word network and each word having an associative relationship are shown. For the "greenhouse effect", the user selects the related word set in the second word network such as "CFC", "carbon dioxide", "warming", and the first word such as "greenhouse", "effect". It can be seen that there are related word sets in the word network.
Thus, when a certain word is designated, the word set related to the word can be easily recognized.

In this embodiment, each word has been described as having only one meaning, so polysemous words are treated as if different words exist for each meaning.

Example 3. FIG. 9 is a flowchart showing a document set structuring method according to the third embodiment of the present invention. This document set structuring method can be applied to the document set structuring means 6 in the system shown in FIG.
Of course, this document set structuring method can be applied to other systems. This method is a method in which a plurality of documents are divided into several groups and a word representative of the content of each group is obtained.

Here, the weight V of the word j included in the document i
_ij , that is, the importance of the word j in the document i is defined as follows. V _ij = F _ij × log (N / N _j ) × log (L _j ) ... (2) N is the number of documents in the document set, N _j is the number of documents containing the word j, and L _j is the word The number of characters constituting j, F _ij is the document i
Is the frequency with which the word j appears. Expression (2) represents the following.

It is considered that the word is more important as the number of times it appears in a document is larger, and more important from the viewpoint of classification as the number of documents in which the word appears is smaller.
Therefore, in such a case, the weight V _{ij of the} word j becomes large. When the word j appears in all the documents in the document set, N _j = N and log (N / N _j ) = 0. Therefore, the weight V _{ij of the} word j is also 0. That is, the words appearing in all the documents cannot be distinguished from each other by the words, and therefore the importance is 0 from the viewpoint of document classification. Also, lo
g (L _j ) is a factor indicating that the larger the number of characters that make up a word, the more important it is. For example, "
Words represented by one letter such as "mountain" and "river" are very common words, and their importance is considered to be low. The more letters that make up the word, the more concepts are included. The compound noun is often a compound noun, which is important from the viewpoint of document classification, and therefore has a large weight V _ij .

Next, the operation will be described with reference to the flowchart of FIG. First, each document is given by equation (2)
In accordance with the above, the weight of each word contained in it is converted into a vector having elements (step ST149). Therefore, each vector corresponding to each document is obtained. Next, the number of clusters to be obtained is set in the variable K (step ST150). When the document set structuring means 6 in the system shown in FIG. 1 executes this method, the user inputs the number of clusters to be obtained. Then, the centroids of all the vectors are obtained, and the vector farthest from that is set as the cluster center (step ST151).

Then, steps ST152 and ST153
The above process is repeated until K clusters are obtained. In step ST152, the distance to the closest cluster center is obtained for each vector, and the vector corresponding to the maximum value of those distances is further set as the cluster center. When K cluster centers are obtained, that is,
When K clusters are obtained, the process proceeds to cluster classification processing (step ST153).

In the cluster classification process, first, each vector is classified into the closest cluster center cluster (step ST154). Then, in each cluster, the average of the vectors included in the cluster is set as a new cluster center (step ST155). Step ST1
When there is a cluster center changed by the process of 55, the process of step ST154 is executed again (step ST156). When the cluster center does not change, the process of finding a representative word in each cluster is executed. That is, in each cluster, L elements having large values are extracted from the elements of the cluster-centered vector. Then, each word corresponding to the extracted L elements is set as a representative word representing the content of the cluster (step ST157).

As described above, each document is classified into K clusters, and L representative words are set in each cluster. That is, the document set is structured. In the process of step ST157, when the value of the element of the cluster-centered vector is smaller than a certain threshold, it is considered that the element does not represent the feature of the cluster, and therefore it is excluded from the representative words. May be.

Examples of clusters and typical words obtained by the above method will be shown below. (A) An example in which each article including the word “TV” is set as a document set for economic-related newspaper article data and the document set is automatically classified. Cluster number Representative word 1 Index 2 service, JR West, Sanyo Shinkansen 3 NHK, satellite broadcasting, development ………………………………… 6 CM, commercial, consumer 7 Oversized garbage, fee, fee 8 Energy, 2010 Fiscal year, energy conservation 9 France, public notice, tobacco ………………… 11 Mall main, foreigner, government 12 Automobile, export, semiconductor ……………

(B) An example in which each article including the word "official discount rate" is set as a document set for economic-related newspaper article data and the document set is automatically classified. Cluster number Typical words 1 Fixed amount savings, High interest rate, Interest rate 2 Finance minister, Crude oil price, Marc 3 Monetary tightening, Borrowing, Manufacturing 4 Long-term prime rate, Lending interest rate, Reduction 5 Money supply, Growth rate, Lending 6 Shotoku Taishi, High-value banknotes, Ministry of Finance 7 rough estimate request, crude oil price, LDP 8 government bond, SAMA, investment …………………………………………………………………………………………………

(C) 1 Personal asset management, deposit and savings ..... 2 Abolition of interest rate “cap method” ..... 3 Banks of Naganoshin banks, long plastics ........ 4 Metropolitan bank, this time drastically reduced .... 5 Long plastics 0.3% In addition, although the case where the cluster center is determined from the beginning has been described in this embodiment, when adding a document after the document set has already been structured, step ST
The process may be started from 154. In that way, documents can be added in sequence and structured for all documents.

Example 4. 10 and 11 are flowcharts showing the operation of the idea generation support system according to the fourth embodiment of the present invention. The configuration of this system is as shown in FIG.

Next, the operation will be described. The information storage unit 1 stores document information including a plurality of documents (texts) input from an input unit (not shown) (step ST).
400). The dictionary generating means 2 executes the method described in the first embodiment, and generates a dictionary using such a method (step ST410).

The word set structuring means 3 executes the method described in the first embodiment, and the structure of each word in the dictionary generated by the dictionary generating means 2 using such a method. Conversion processing is performed (step ST412). The processing result is, for example, the first word network and the second word network as described above, and the word set structuring unit 3 stores them in the structured word set storage unit 8 (step ST414). . Further, the document set structuring unit 6 performs a document structuring process for each document in the information storage unit 1 (step ST420). The structuring process is performed as described in the third embodiment.
The document set structuring unit 6 stores the processing result in the structured document set storage unit 9 (step ST422).

When the user inputs the information indicating the selected document, the information selecting means 4 performs a process of selecting each selected document from the information storing means 1 (step ST430).
When selecting, the user can use keywords. The information selection unit 4 can search the document set in the information storage unit 1 by the logical sum or logical product of the input keywords and select the corresponding documents. Alternatively, each document can be selected by vector search with a vector of words. Each selected document is stored in the selection information storage means 5 (step ST432).
The document set structuring unit 6 performs a document set structuring process for each document in the selection information storage unit 5 (step S
T434), the processing result is stored as a structured selection document set storage unit 1.
It is stored in 0 (step ST436).

The information space displaying means 7 reads out the contents of the structured word set accumulating means 8, the structured document set accumulating means 9 and the structured selected document set accumulating means 10 and uses them as a key with the word specified by the user as a key. Are displayed in association with each other (step ST438). For example, the displays shown in the upper and middle rows of FIG. 12 are displayed.

The user can make various processing requests to the system via the input means (not shown) (step ST440). When the user requests a new “keyword search”, the system re-executes the processing from step ST430 onward based on the input keyword.
That is, do a document search for a new keyword,
Structures each document retrieved and displays it.

When the user requests "display cluster contents", the information space display means 7 outputs the document IDs and headings belonging to the cluster designated by the user as the structured document set storage means 9 and the structured selection. It is read from the document set accumulating means 10 and displayed (step ST442).
When the user requests "display body" together with the document ID, the information space display means 7 reads the relevant document from the information storage means 1 and displays the contents (step ST4).
44).

When the user requests the "related display", the information space display means 7 highlights the element related to the element designated by the user (step ST446). FIG. 12 shows an example of a related display when the user inputs the word word-il as an element. The information space display means 7 obtains the relation between the word-il and the word from the word network stored in the structured word set storage means 8. And, for example, word-il
The most recent upper word and the most recent lower word of are highlighted. From the presented screen, the user has the latest words, word-a, word-b, and word-c, as words related to word-il, and conceptually word-
Words higher than il are word-d, word
It is easy to see that there is -e. Also, as words having meanings close to those words, word-f,
word-g, word-h, word-j, word
It is easily understood that there are -k, word-p, word-m, word-n, and the like.

From the relationship and the display contents of the structured document, the user can understand that the cluster having word-il as a representative word is the cluster of cluster number i. Therefore, "Display cluster contents" for cluster number i
Request, the document IDs and headings belonging to that cluster are displayed. Therefore, from the display contents as shown in FIG.
It is easily understood that the documents oc-a2, ..., Doc-ai, ..., Doc-an are included. In addition, w
It can also be seen that the one including ord-il in the heading is doc-ai. Therefore, if necessary, the user can refer to the contents by requesting "text display". In this way, the user uses the specified word
The relationship between the d-il and each document can be easily known.

When the user makes a request for "end", the processing of the system ends. The information storage means 1
Has been described as if it were one device to which information is set from the outside, but it may be composed of a plurality of storage means distributed over a computer network. Further, the contents of the selection information storage means 5 and the structured document set storage means 9 may be saved when the system ends.
In that case, the system can read them at the time of restart and restart the process from step ST438.

Next, a specific example of using the idea generation support system will be described. For example, it has already been explained that the word set structuring means 3 can generate the following second word network.

[Table 5]

This second word network is visually displayed by the information space display means 7. The user can obtain the relation of the concept of interest by searching the path between items in the second word network.
For example, regarding the relationship between “refrigerator” and “warming”, “a refrigerator uses CFCs as a refrigerant, and the greenhouse effect caused by diffusion of the CFCs into the atmosphere causes the earth to warm. The greenhouse effect is also caused by carbon dioxide. "

The user can access the information of interest from the structured document set as described above. (A) An example in which each article including the word “TV” is set as a document set for economic-related newspaper article data and the document set is automatically classified. Cluster number Representative word 1 Index 2 service, JR West, Sanyo Shinkansen 3 NHK, satellite broadcasting, development ………………………………… 6 CM, commercial, consumer 7 Oversized trash, fee, charge 8 Energy, 2010 Fiscal year, energy conservation 9 France, public notice, tobacco ………………… 11 Mall main, foreigner, government 12 Automobile, export, semiconductor ……………

In the structured document set, as a result of the user requesting "keyword search" with the keyword "TV", the information selecting means 4 extracts each relevant document from the information storing means 1 (step ST430). , ST43
2), the document set structuring means 6 performs the structuring process and the information space display means 7 displays it.

The user who has obtained information based on the vague keyword "TV" sees the automatic classification result and the representative words of the clusters, and obtains the following information about "TV". be able to. (1) From the viewpoint of the broadcasting business such as NHK and satellite broadcasting from the cluster of cluster number 3. (2) From the viewpoint of commercial television broadcasting such as commercials and commercials from the cluster with cluster number 6. (3) From the viewpoint of disposal of the television receiver such as oversized waste from the cluster of cluster number 7. (4) Viewpoint of a television receiver as a commodity such as an automobile, an export, and a semiconductor from the cluster of cluster number 12.

It is easy to forget the negative point of disposing of television receivers such as oversized trash, but as described above, by targeting a large amount of ideas,
It can also provide perspectives that the user may have forgotten. In other words, this system is an effective idea generation support system that allows the user to search and access information of interest while viewing the processing result of the system.

Further, the user can access the information of interest from the structured document set as described above in (b) and (c). (B) is an example in which each article including the word "official discount rate" is set as a document set for economic-related newspaper article data and the document set is automatically classified. (C) is the heading of the document in the cluster of cluster number 4 in the classification of (b). As can be expected from the typical words of the cluster, it can be seen that documents related to long-term interest rates and lending rates are gathered.

As described above, according to this embodiment, the structured document set obtained by structuring the document set accumulated in the information accumulating means 1 and the selection information selected by the information selecting means 4 are structured. Since it can be associated with the structured selection document set, the user can search information while grasping the entire information space, and the user can take the initiative to search the information space. It is possible. Further, since the user can freely look around the information space, the user may be able to find something different from the initial intention.

[0103]

As described above, according to the first aspect of the present invention, the dictionary generation method uses the character string appearing in the text, the appearance frequency of which is not lower than that character string, and
Since it is configured to generate a dictionary by extracting a character string that satisfies the condition that there is no character string that contains that character string and is longer than that character string, a normal dictionary is created according to each given text. There is an effect that a character string that does not exist can be automatically extracted and a more appropriate dictionary can be provided for each given text.

According to the second aspect of the present invention, the word set structuring method relates the words to the latest low order word and the latest high order word, and excludes the relationships that meet a predetermined condition from those relationships. Since it is configured to generate a word network, it is possible to automatically obtain the upper / lower words of a certain word and provide an environment for finding an appropriate search word in information retrieval, and improve the recall and accuracy of the search. There is an effect that you can get what you can.

According to the third aspect of the present invention, a word set structuring method is used to generate a basic chain by three words based on a co-occurrence relation between texts in a set of documents having a headline and a body. , It is configured to combine the basic chains that have a common part among the generated basic chains to generate the word network, so that it is possible to provide an environment that can easily understand the word set related to a certain word. effective. When this method is applied to the idea generation support system, there is an effect that the user can easily understand the relation of the concept of interest.

According to the fourth aspect of the present invention, the word set structuring method is configured to integrate the first word network and the second word network with the designated word as a key. On the other hand, there is an effect that it is possible to provide an environment in which a word set related to the word can be easily understood.

According to the fifth aspect of the invention, the document set structuring method includes a plurality of clusters, each of which includes a plurality of texts and which are discretely set by using the distance between word vectors corresponding to the respective texts. Since it is configured to obtain a word representative of each cluster, there is an effect that a text set is automatically classified according to the content and a word representative of each classification is obtained. When this method is applied to the idea generation support system, there is an effect that it is possible to provide an environment in which the user can find an attribute that the user did not notice about a theme.

According to the sixth aspect of the present invention, in the document set structuring method, the vector farthest from the center of gravity of all word vectors is set as the center of the first cluster, and the distance between each word vector and the closest cluster center is set. Then, the word vector with the maximum distance is added as the cluster center, each word vector is classified into the cluster with the closest cluster center, and the average of the word vectors included in each cluster is used as the new cluster center. Since it is configured to continue until does not change, there is an effect that each cluster is effectively arranged discretely in the document set space.

According to the seventh aspect of the invention, the document set structuring method extracts a predetermined number of large values from the elements of each word vector in the cluster, and extracts words corresponding to each extracted word vector. Is defined as a word representing the cluster, so that a word with higher importance in the document corresponding to each word vector forming the cluster can be selected as the word representing the cluster. effective.

According to the invention described in claim 8, since the idea generation support system is configured to display the processing result of the word set structuring means and the processing result of each of the document set structuring means, it is possible to correlate them. There is an effect that an information space consisting of a set of words and a structured set of texts can be visualized, and a user can extract a clue of an associative relationship by a word network.

According to the invention described in claim 9, since the idea generation support system is configured to select each text based on the keyword of the search and to structure the document set for each selected text, the user There is an effect that what can be confirmed by information retrieval that a person thinks from the clue of knowledge can be obtained.

According to the invention of claim 10,
The idea generation support system, when there is a request for a related display, on each display screen for a structured word set, a structured text set and a structured selection text set,
Since it is configured to highlight the element corresponding to the request for the related display, the user can easily understand the relationship between the word specified by the user and other words and the relationship between the word specified by the user and each text. There is an effect that you can get what you can understand.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of an idea generation support system according to an embodiment of the present invention.

FIG. 2 is a flowchart showing a dictionary generation method according to the first embodiment of the present invention.

FIG. 3 is a flowchart showing a process of creating a first word network in the word set structuring method according to the second embodiment of the present invention.

FIG. 4 is a flowchart showing a process of creating a first word network in the word set structuring method according to the second embodiment of the present invention.

FIG. 5 is an explanatory diagram showing an example of a first word network.

FIG. 6 is an explanatory diagram for explaining a co-occurrence relationship.

FIG. 7 is a flowchart showing a process of creating a second word network in the word set structuring method according to the second embodiment of the present invention.

FIG. 8 is an explanatory diagram showing an example of a processing result by the word set structuring method.

FIG. 9 is a flowchart showing a document set structuring method according to a third embodiment of the present invention.

FIG. 10 is a flowchart showing an operation of the idea generation support system according to the fourth example of the present invention.

FIG. 11 is a flowchart showing the operation of the idea generation support system according to the fourth example of the present invention.

FIG. 12 is an explanatory diagram showing an example of display by the information space display means.

[Explanation of symbols]

1 information storage means, 2 dictionary generation means, 3 word set structuring means, 4 information selecting means, 6 document set structuring means, 7 information space display means.

Front page continued (72) Inventor Terumasa Yasui 8-1-1 Tsukaguchihonmachi, Amagasaki City Mitsubishi Electric Corporation Central Research Laboratory (72) Inventor Shinichiro Tsutaka 8-1-1 Tsukaguchihonmachi, Amagasaki Mitsubishi Electric Corporation Company Central Research Institute

Claims (10)

[Claims] 1. A step of obtaining an appearance frequency of a character string in one or a plurality of texts, wherein the appearance frequency is not lower than the character string among the character strings appearing in the text, and the character A dictionary generation method comprising: a step of extracting a character string that satisfies a condition that a character string including a string does not exist longer than the character string; and a step of adding the extracted character string as an item to generate a dictionary. 2. A step of obtaining a latest lower-rank word of a word and a latest higher-rank word of the word, and associating the word with the latest lower-rank word and the latest higher-rank word, and a relationship satisfying a preset condition among those relationships. A method of structuring a word set, the method comprising: 3. The first included in the heading of the first text
Is included in the body of the second text, the second word in the heading of the second text is included in the body of the third text, and the first word is the third text. The first word, if the third word is included in the heading of the third text, the first word,
A word set structure including a step of generating a basic chain connecting the second word and the third word, and a step of combining basic chains having a common part among the generated basic chains to generate a word network. Method. 4. A step of obtaining a latest lower-order word of a word and a latest higher-order word of the word from a plurality of texts, and associating the word with the latest lower-order word and the latest higher-order word, and presetting those relationships. Generating a first word network excluding associations that meet the specified conditions, the first word included in the heading of the first text being included in the body of the second text, and the first word being included in the body of the second text. The second word in the headline is contained in the body of the third text, the first word is contained in the body of the third text, and the third word is contained in the body of the third text. First word, second word, if included in heading
Generating a basic chain connecting the first word and the third word, combining the basic chains having a common part among the generated basic chains to generate a second word network, and the specified word And a step of integrating the first word network and the second word network with the key as a key. 5. A step of converting each text into a word vector, and setting a plurality of clusters each including a plurality of texts of the respective texts and discretely set using a distance between the word vectors. Steps to
A method of structuring a document set, comprising the step of obtaining a word representing each of the clusters. 6. The step of setting a plurality of clusters comprises:
The step of obtaining the center of gravity of all word vectors, the step of setting the vector farthest from the obtained center of gravity as the center of the first cluster, and the distance to the closest cluster center for each of the word vectors are obtained, and the maximum distance And adding each word vector as a cluster center, and classifying each of the word vectors into the cluster with the closest cluster center and using the average of the word vectors included in each cluster as a new cluster center, the cluster center does not change. 6. The document set structuring method according to claim 5, further comprising the step of continuing up to. 7. The step of obtaining a word representative of a cluster includes a step of extracting a large number of predetermined values from the elements of each word vector at the center of the cluster and a step of extracting a word corresponding to each extracted word vector. 7. The document set structuring method according to claim 6, further comprising the step of defining words as representative of clusters. 8. An information storage unit for storing document information, a dictionary generation unit for generating a dictionary by the method according to claim 1 for the text stored in the information storage unit, and the dictionary generation unit. A word set structuring means for structuring the word set by the method according to claim 4 using a dictionary as a word set, an information selecting means for selecting a predetermined text from a text set in the information storing means, and the information storing means. 8. A document set structuring means for structuring the text set in the text set and the text set selected by the information selecting means by the method according to claim 7, the processing result of the word set structuring means, and the document set structuring means. Idea support system including an information space display means for displaying the processing result of. 9. When a search keyword is input, the information selection means selects each text based on the keyword, and the document set structuring means selects each text based on the keyword for each text selected by the information selection means. The idea generation support system according to claim 8, wherein the document set is structured. 10. The information space display means, when there is a request for a related display, the processing result of the word set structuring means, the processing result of the document set structuring means for the text set in the information storage means, and the information selecting means. 10. The idea generation support system according to claim 9, wherein an element corresponding to the request for the related display is highlighted on each display screen of the processing result of the document set structuring unit for the text set selected by.