JP4592629B2 - Document search support method and apparatus, program, and computer-readable recording medium - Google Patents

Description

  The present invention relates to a document search support method and apparatus, a program, and a computer-readable recording medium. In particular, in order to efficiently perform a document search, a document that matches a search condition requested by a user is used. The present invention relates to a document search support method and apparatus, a program, and a computer-readable recording medium.

  In a normal document search, a technique for presenting a document that directly matches a user's search request is used. However, when the number of target documents is enormous, the number of documents cannot be sufficiently narrowed down by simply matching with the conditions, and the user cannot efficiently obtain the documents.

  For this reason, there is a search method in which a plurality of conditions such as document creation time and document genre information are specified in addition to keywords (see, for example, Patent Document 1).

In addition, additional information such as genre information of documents such as “economy” and “society” and type information of words and phrases appearing in the document, such as “International Atomic Energy Agency = Organization” and “United Nations = Organization”, is stored in advance. A technology that supports search by assigning to a document, narrowing down the document by a keyword given by the user, automatically organizing the document using the additional information, and presenting representative information called a label There is.
JP 2005-339139 A

  However, the above existing technique is based on the premise that genre information, phrase type information, etc. are assigned to each document, and a special dictionary that performs manual assignment work or records the type of phrase. Need to prepare.

  The present invention has been made in view of the above points, and provides a document search support method and apparatus for organizing and presenting search results based on search conditions designated by a user without requiring special preparation in advance. An object is to provide a program and a computer-readable recording medium.

  FIG. 1 is a diagram for explaining the principle of the present invention.

The present invention (Claim 1) is a document search support apparatus having a document storage means storing a document to be searched, a document search means, a center of gravity selection means, a vector generation means, a vector storage means, a cluster generation means, and a cluster storage means. A document search support method for searching for a document that matches a search condition requested by a user,
A document search step (step 1) in which the document search means acquires from the document storage means a document that matches the search condition input by the user;
For each document acquired by the centroid selection means in the document search step, a distance between a position where a word A included in each input search condition and a word B included in the same search condition other than the word A appear in the document The center-of-gravity selection step of assigning a score to the word A appearing in the document and specifying the position in the document of the word having the smallest score as the center- of- gravity position for all words included in the search condition (Step 2)
The vector generation means totals the words appearing within a certain range centered on the center of gravity position and the number of appearances of the words from each document acquired from the document storage means, acquires a vector for each document, and stores the vector A vector generation step (step 3) to be stored in the means;
A cluster generation unit calculates a distance between vectors of each document stored in the vector storage unit, integrates documents having close distances to form a cluster, and stores the cluster in the cluster storage unit (step 4). ) And do.

The present invention (Claim 2) is a document search support apparatus further comprising a main keyword acquisition unit.
The main keyword acquisition means refers to the vector of the vector storage means based on the cluster configuration document stored in the cluster storage means , finds the sum of the weights for each word, and determines the top N words from the words with the highest sum A main keyword acquisition step (step 5) for selecting as a main keyword is further performed.

The present invention (Claim 3) is a document search support apparatus further comprising representative information generating means.
A representative information generation step (step 6) in which the appearance position of the main keyword is specified for each of the clustered documents stored in the cluster storage means, and a certain range centered on the appearance position is acquired as an important part of the document. Is further performed.

Further, the present invention (Claim 4), in the gravity center selection step,
When the centroid selection unit can divide the search condition input by the user into a plurality of words, the centroid selection unit divides the search condition into words W _{1 to} W _n, and when the word W ₁ appears multiple times in the document, the word W _A weight determined by the distance to W ₂ closest to the first appearance position of 1, a weight determined by the distance from W ₃ closest to the first appearance position of word W ₁ ,. weights determined by the distance between the Wn located closest to the _first appearance position, the sum recorded in the storage means as the first score of the W _1, Similarly, the second word W _1, third, ... of The score is also calculated for the appearance position, the appearance position of W ₁ having the smallest score is determined as the centroid of the word W _{1 in} the document, and similarly, the centroid positions of the other words W _{2 to} W _n are determined. Thus, the centroid positions of the words W _{1 to} W _n constituting the search condition are determined for each document. ,
If the search condition cannot be divided into a plurality of words, the appearance position closest to the beginning of the sentence in the document of the search condition is determined as the center of gravity.

  FIG. 2 is a principle configuration diagram of the present invention.

The present invention (Claim 5) is a document search support apparatus for searching for a document that matches a search condition requested by a user,
A document search unit 306 for acquiring a document that matches the search condition input by the user from the document storage unit 301 storing the document to be searched;
For each document acquired by the document search means 306, based on the distance between the position where a word A included in each input search condition and the word B included in the same search condition other than the word A appear in the document , A centroid selection unit 307 that assigns a score to the word A appearing in the document and performs a process of specifying the position in the document of the word having the smallest score as the centroid position for all words included in the search condition ;
From each document acquired from the document storage unit 301, the words appearing within a certain range centered on the center of gravity position and the number of appearances of the words are totaled, and a vector for each document is acquired and stored in the vector storage unit 309. Vector generating means 308 to perform,
A cluster generation unit 310 that calculates a distance between vectors of each document stored in the vector storage unit, integrates documents having close distances to form a cluster, and stores the cluster in the cluster storage unit 311;

The present invention (Claim 6) refers to the vector in the vector storage unit 309 based on the cluster configuration document stored in the cluster storage unit 311 to obtain the sum of weights for each word, and the sum There is further provided a main keyword acquisition means 312 for selecting the top N words from the large words as the main keywords.
Main keyword acquisition means for referring to and selecting a vector of the vector storage means based on a cluster configuration document stored in the cluster storage means;
The document search support apparatus according to claim 5, further comprising:

  According to the present invention (Claim 7), the appearance position of the main keyword is specified for each of the cluster-structured documents stored in the cluster storage means, and a certain range centered on the appearance position is designated as an important part of the document. As a representative information generation means.

Further, the present invention (Claim 8) is provided in the center-of-gravity selection means 307.
When the search condition input by the user can be divided into a plurality of words, the search condition is divided into words W _{1 to} W _n, and when the word W ₁ appears multiple times in the document, the first word W ₁ weights determined by the distance between the W ₂ located closest to the appearance position, the weights determined by the distance between W ₃ located closest to the first occurrence position of the word W _1, ..., from the appearance position of the word W ₁ weights determined by the distance between a certain W _n closest, the sum recorded in the storage means as the first score of the W _1, Similarly, the second word W _1, third, also ... occurrence position of The score is calculated, the appearance position of W ₁ having the smallest score is determined as the centroid of the word W _{1 in} the document, and similarly, the centroid positions of the other words W _{2 to} W _n are determined. Determining the centroid positions of the words W _{1 to} Wn constituting the search condition for each document;
In addition, when the search condition cannot be divided into a plurality of words, the search condition includes means for determining the appearance position closest to the beginning of the sentence as the center of gravity in the document of the search condition.

The present invention (Claim 9) is a document search support program for causing a computer to function as each means constituting the document search support apparatus according to any one of Claims 5 to 8.

The present invention (claim 10) is a recording medium computer-readable, characterized in that storing the document search support program according to claim 9.

  According to the present invention, the distance between words constituting a search condition is used, and a certain range centered on the appearance position is acquired as an important part of the document, thereby improving the accuracy of subsequent document organization processing. It is.

  In addition, as a result of integration of documents having a small distance between vectors, that is, containing many of the same morphemes into the same cluster, a plurality of clusters of similar documents can be obtained.

  Further, by determining a main keyword that is a word representing each cluster from the clusters, it is possible to select a word that has a strong influence in the process of forming each cluster.

  As described above, when displaying search results, similar documents are clustered using words that appear in the vicinity of the search condition in each document, and information representing each cluster also uses the peripheral information of the search condition. Therefore, the search results can be organized and presented in accordance with the user's search intention without giving knowledge of a special dictionary or the like in advance.

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

  FIG. 3 is a configuration diagram of the document search support apparatus according to the embodiment of the present invention.

  The document search support apparatus shown in FIG. 1 includes a document recording unit 301, a document analysis unit 302, a word recording unit 303, a search request statement generation unit 304, a search request statement recording unit 305, a document search unit 306, a centroid selection unit 307, a vector. A generation unit 308, a vector recording unit 309, a cluster generation unit 310, a cluster recording unit 311, a main keyword acquisition unit 312, a main keyword recording unit 313, a representative information generation unit 314, and a search result display unit 315 are configured.

  The document recording unit 301 stores a document to be searched in the present invention. FIG. 4 shows an example of document data stored in the document recording unit 301. The document recording unit 301 records a document ID uniquely assigned to each document and the text of each document.

  The document analysis unit 302 performs in advance an analysis process that is a preparation for searching for a document that matches a condition from the documents recorded in the document recording unit 301. After obtaining a pair of the document ID and the text of the document from the document recording unit 301 and analyzing the text to acquire a word in the text, the appearance position information of each word in the text and the word along with the document ID Records in the recording unit 303. For the analysis processing, for example, a morpheme analysis technique that is frequently used in the field of natural language processing is used, and the morpheme is divided into “morphemes” that are the minimum constituent units constituting the sentence. In the morpheme analysis, the part of speech information is given to each divided morpheme, but the document analysis unit 302 of the present embodiment acquires only the morpheme having the part of speech information of “noun” among all the morphemes. . In the present embodiment, hereinafter, only nouns acquired from the text will be referred to as words and will be handled in subsequent processing, but are not necessarily limited to nouns.

  FIG. 5 is a flowchart of processing performed by the document analysis unit according to the embodiment of the present invention.

  Step 101) The document analysis unit 302 acquires the document ID and text for one document from the document recording unit 301.

  Step 102) A morphological analysis process is performed on the acquired text.

  Step 103) A word (here, a noun) is acquired by performing morphological analysis.

  Step 104) The word group acquired from one document is recorded in the word recording unit 303 together with the appearance position information in the document and the document ID.

  Step 105) The processing is repeated until the analysis processing of all the documents in the document recording unit 301 is completed.

  As a result of this processing, word data as shown in FIG. “START” and “END” in FIG. 6 represent the start position and end position in the text in bytes. The appearance position information is given while the document analysis unit 302 performs morpheme analysis processing and then sequentially tracks the morpheme from the beginning of the sentence. For example, in the document with the document ID “0001”, since the sentence head is “Government is 1 day”, the word “Government” appears in the 0th byte, and one Japanese character has an information amount of 2 bytes. Considering this, it can be seen that the 0th to 3rd bytes are “government” appearance positions. Since it is understood that the word “consumption tax” that appears second in the document ID “0001” appears from the 12th byte to the 17th byte, “12” is similarly set to “END” as the value of “START”. “17” is stored in the word recording unit 303 as the value of. The same applies to the subsequent words and documents after ID “0002”.

  If the data as shown in FIG. 6 is created in advance and exists in an external storage device or the like instead of an essential component of the search support device of the present invention, the document analysis unit 301 is included in the configuration of the search support device. Not necessary.

  The search request statement generation unit 304 generates a search request statement to be sent to the document search unit 306 of the present invention from the search conditions input by the user. For example, when a user inputs two search conditions such as “Prime Minister” and “Diplomacy Issue”, a search request sentence “q = (Cabinet or Prime Minister or Minister) and (Diplomacy or issue)” is generated. And recorded in the search request sentence recording unit 305.

  Hereinafter, the processing of the search request sentence generation unit 304 will be described in detail.

  FIG. 7 is a flowchart of search request statement generation processing in an embodiment of the present invention.

  Step 201) First, the search request sentence generation unit 304 acquires the first condition “Prime Minister”.

  Step 202) This is sent to the document analysis unit 302 to perform morphological analysis processing and word (herein, noun) acquisition.

  Step 203) The acquired three words “Cabinet”, “Prime Minister” and “Minister” are temporarily stored in a buffer (not shown).

  Step 204) It is determined whether or not all the search conditions have been analyzed. If the analysis has not been performed yet, the process returns to Step 201 to perform the processing of the “diplomatic problem” that is the second condition. Two words “problem” are held in a buffer (not shown) in the same manner as in step 203. If all the search conditions have been analyzed, the process of step 205 is performed.

  Step 205) The above search request sentence is generated. Note that the above search request sentence is a search for “a document that includes any one or more words of the cabinet, prime minister, or minister and includes any one or more words of diplomacy or problems”. It is a sentence that indicates what is required. If the search condition entered by the user is only one word (for example, “Prime Minister only”), the search request will be “q = (Cabinet or Prime Minister or Minister)”, and multiple morphological analysis processing will be performed. In the case of search conditions that are not divided into words (for example, “Cabinet” and “diplomatic”), no division is performed, and the search request sentence is “q = (Cabinet) and (diplomatic)”.

  The search request text generated by the search request text generation unit 304 in the above procedure is stored in the search request text recording unit 305.

  The document search unit 306 acquires a search request sentence from the search request sentence recording unit 305, queries the word recording unit 303 to search for a document that meets the conditions, and sends a list of document IDs to the centroid selection unit 307.

  The center-of-gravity selection unit 307 selects, for each searched document, a location (center of gravity position) where words in the search request sentence appear densely at positions close to each other in the text. The barycentric position here is the position information obtained for each word of the search condition entered by the user. In the above example, the barycentric positions of “Prime Minister” and “Diplomatic Issues” are obtained. Become. In addition, the gravity center selection unit 307 does not include words such as “diplomatic” and “problem” that constitute the search request sentence in the actual document, such as “diplomatic problem” and “diplomatic problem”. Even in such a case, if they appear at positions close to each other with a short word such as “no” or “related”, it is determined that the position is the center of gravity. For this reason, a set of a plurality of words appearing at different positions may be referred to as one center of gravity position.

  FIG. 8 is a diagram for explaining a method for determining the position of the center of gravity according to the embodiment of the present invention.

  The figure shows the text of one document. In the text, the word “problem” appears 701, 704, 705 three times, and the word “diplomatic” appears 702, 703 twice. Further, it is assumed that “q = (Cabinet or Prime Minister or Minister) and (Diplomatic or Problem)” is given as a search request sentence. As described above, the position of the center of gravity is determined for each search condition entered by the user.Here, the center of gravity position of the “Prime Minister” has already been determined. This will be described with reference to FIG.

  FIG. 9 is a flowchart of the center-of-gravity position determination process according to the embodiment of the present invention.

  Step 301) The center-of-gravity selection unit 307 acquires “diplomatic” that is one word constituting the search condition.

  Step 302) Among the words constituting the “diplomatic problem” that is the same search condition, one word other than the word acquired in step 301, that is, “problem” is acquired.

  Step 303) For each of the words “diplomatic” appearing in the text, the distance to “problem” is checked, and the distance to the nearest “problem” is added in a buffer (not shown). For example, if “Diplomacy” in 702 of FIG. 8 and “Problem” that is closest in the text are “Problem” in 701 and the distance between these two words is 30 bytes, “Diplomacy” in 702 30 is added as a score and stored in a buffer (not shown). Subsequently, with respect to “Diplomacy” of 703, since the closest “Problem” is adjacent to 704, the distance “0” is added as the score of “Diplomacy” of 703 and stored in a buffer (not shown).

  Step 304) After adding the shortest distances to all words other than “diplomatic”, the process proceeds to step 305.

  Step 305) Among the “diplomacy” appearing twice, the one having the smallest score stored in the buffer (not shown) is selected as the barycentric position of “diplomacy”. That is, a location where the search conditions input by the user appear close together in the text is selected as the position of the center of gravity. In the example of FIG. 8 above, as a result of comparing “30” and “0”, which are the scores at the time of step 304 Yes of “Diplomacy” of 702 and 703, “Diplomacy” of 703 having the smallest score is selected. The center of gravity is determined. When there are a plurality of the same scores, all of them are determined as the barycentric position.

  Step 306) For “problem”, the processes of steps 301 to 305 are repeated in the same manner, and the center positions of “diplomatic” and “problem” are determined. Since “Diplomacy” in 703 and “Problem” in 704 are adjacent to each other, the “Problem” score in 704 is smaller than the “Problem” score in 701 and 705, and the center of gravity of “Problem” is 704 It is determined.

  Up to this point, the description has been made with reference to FIG. 8, but in actuality, the center-of-gravity selection unit 307 collates the search request sentence with the information in the word recording unit 303 and records “START” and “END” recorded as shown in FIG. The shortest distance is searched using the position information “”.

  The center-of-gravity position determination method in the above-described center-of-gravity selection unit 307 is not limited to the “diplomatic problem” but the word “problem” in a meaning different from the user's search intention, such as “unemployed person problem” and “problem remark”. When there is a document being used, the correct character string “diplomatic problem” is found to improve the accuracy of the subsequent document organization process. Also, instead of simply matching the search condition with the text to determine the position of the center of gravity, it uses the distance between the words that make up the search condition, so other characters like “diplomatic issues” Even when pinched, it can be detected correctly as the center of gravity.

  For example, when a search condition that is not divided into a plurality of words even if morphological analysis processing is performed, such as “Cabinet”, the distance from other words cannot be measured as described above. In this case, priority is given to the “Cabinet” that appears at the position closest to the beginning of the sentence, and the center of gravity is set.

  When the centroid selection unit 307 finishes determining the centroid position in one document, the centroid selection unit 307 sends the document ID and the centroid position to the vector generation unit 308. For example, in the document with the document ID “9876”, the starting point of the center of gravity of each document is “0” for “Cabinet”, “4” for “Prime Minister”, and “8” for “Minister”. Similarly, when “Diplomacy” is “100” and “Problem” is “104”, the vector generation unit 308 has ID “9876” and the center of gravity position 0. , 4, 8, 100, 104 are transmitted.

  The vector generation unit 308 collates the acquired document ID and centroid position with the contents of the word recording unit 303, and the word appearing within a certain range before and after the centroid position in the document (for example, 50 bytes before and after). To get. In the above-described example of “ID“ 9876 ””, the ranges of 50 bytes before and after the center of gravity of 0, 4, 8, 100, and 104 are −50 to 50, −46 to 54, 42 to 58, and 50, respectively. ~ 150, 54 ~ 154. All words included in any of these ranges are acquired from the document with ID “9876”. If an overlap or a negative value is removed from these ranges, “0 to 58 or 50 to 154” is the range to be acquired. Subsequently, in the document with ID “9876”, the total number of occurrences of the word in the above range is recorded in the vector recording unit 309 together with the document ID as a word weight.

  The vector generation method is not limited to the above, and other existing techniques may be used.

  When the above is performed for all the documents sent from the document search unit 306 to the gravity center selection unit 307, a vector for each document as shown in FIG. 10 is obtained. Each vector is a list of words and word weights (the number of appearances in the present embodiment). Through the processing described so far, when a user inputs a search condition and executes a search, the center of gravity position is determined, and a vector is generated and stored for each document that matches the condition.

  In the present embodiment, the vector generation unit 308 uses the number of appearances of each word as the word weight. For example, the weight of each word is set by TF-IDF (vector space method) often used in the natural language processing field. You may calculate.

  The cluster generation unit 310 performs clustering processing for acquiring all document vectors from the vector storage unit 309 and dividing the vectors into groups of similar vectors.

  There are a plurality of existing techniques for processing methods for performing clustering. Although an example using a clustering method called the longest distance method is described here, other clustering methods may be used.

  Clustering by the longest distance method is performed according to the following procedure.

  First, each document to be clustered is regarded as a cluster, and the distance between the clusters is calculated. The distance between the clusters is the distance between the farthest documents among the documents constituting the cluster, and the Euclidean distance between the vectors acquired from the vector recording unit 309 is regarded as the distance between the documents. However, since all the clusters are each composed of one document at the time of this processing, obtaining the distance between clusters is equivalent to obtaining the distance between documents.

  Subsequently, the two clusters having the closest distance are integrated to form a new cluster, and the distance between the new cluster and another existing cluster is calculated again by the method described above and updated.

  Thereafter, this integration process is repeated to grow the clusters. If the distance between the two clusters to be integrated (the distance between the two nearest clusters) exceeds a predetermined constant value, the integration is performed. And finish the clustering process.

  As a result of the above processing, as a result of integration of documents having a small distance between vectors, that is, documents containing many of the same morpheme into the same cluster, a plurality of clusters of similar documents are obtained.

  After performing the clustering process, the cluster generation unit 310 records the processing result in the cluster recording unit 311 as shown in FIG. A unique cluster ID such as C001, C002,... Is assigned to the generated cluster, and document IDs of documents belonging to the respective clusters are managed. Clusters that have not been integrated with any other clusters in the process of cluster generation are excluded and recorded.

  The main keyword acquisition unit 312 acquires a document ID list of documents in each cluster from the cluster recording unit 311 one cluster at a time, extracts a main keyword representing each cluster from the documents in the cluster, Records in the recording unit 313.

  The main keyword is a word representing each cluster, and it is necessary to select a word representing the content of the document in the cluster. In the present embodiment, in the clustering process performed by the cluster generation unit 310, the following method is used to select words that have a strong influence in the process of forming each cluster.

  The acquisition of the main keyword in the main keyword acquisition unit 312 is performed for each cluster. The main keyword acquisition process for one cluster will be described with reference to FIG.

  FIG. 12 is a flowchart of main keyword acquisition processing according to an embodiment of the present invention.

  Step 401) The main keyword acquisition unit 312 acquires a document ID for one cluster from the cluster recording unit 311. In this step, for example, four IDs “9875”, “9877”, “9898”, and “9900”, which are document IDs of documents belonging to the cluster C123, are acquired.

  Step 402) For the first document “9875”, the vector recording unit 309 acquires a corresponding vector.

  Step 403) The weights of each word are added and totaled. In the example shown in FIG. 10, the vector of the document ID “9875” is “UN = 3, Iraq = 4, today = 1, resolution = 2, assertion = 1,...”. A buffer (not shown) in the keyword acquisition unit 312 adds and holds the same words. However, since the buffer (not shown) is empty at the time of processing the first first document, the weight of each word is held as it is.

  Step 404) The above-described processing of Step 402 and Step 403 is repeated until all the document IDs for one cluster acquired in Step 401 are completed, and the sum of the weights is obtained for each word constituting the document vector for one cluster. .

  Step 405) The top N words are selected from the words having the largest sum of weights, and are recorded in the main keyword recording unit 313 together with the cluster ID C123.

  FIG. 13 shows an example of data recorded in the main keyword recording unit 313 at the time after steps 401 to 405 in FIG. 12 are repeated for each cluster. Main keywords such as “Iraq”, “Self Defense Force” and “Dispatch” are selected from the cluster C123. In the example, N = 3, and three main keywords are acquired from each cluster. In the clustering process, words with large weights have a strong effect on cluster formation, so adding the weight of each word across all the documents in the cluster increases the weight of the word itself and causes many words to appear. Is selected.

  The representative information generation unit 314 acquires the main keywords of each cluster from the main keyword recording unit 313, generates cluster representative information for each cluster, and sends it to the search result display unit 315.

  Since the representative information is information presented to the user for the purpose of grasping the outline of the group of documents in the cluster, the main keyword itself may be used as the representative information. In this embodiment, a part of each document in the cluster is used. A method of extracting and presenting (hereinafter referred to as an important part) will be described.

  The important part is extracted by a method similar to the process in which the center-of-gravity selection unit 307 determines the position of the center of gravity of each document, and the vector generation unit 308 acquires words in a certain range centered on the center of gravity. That is, three main keywords for one cluster are acquired from the main keyword recording unit 313, and document IDs in the cluster are acquired from the cluster recording unit 311, and the position where the main keyword appears in each document is determined. The content of the word recording unit 303 is collated, and the centroid position of each main keyword is determined by the above-described centroid position determination procedure. However, since not all main keywords appear in all documents, when the main keyword is not included in the document, the position of the center of gravity of the main keyword is not determined. Subsequently, a character string in a certain range before and after the center of gravity corresponding to each main keyword is acquired from the document recording unit 301 and used as an important part of the document.

  If the above processing is performed for each document in the cluster and further performed for other clusters, representative information of each cluster can be obtained. The acquired representative information is sent to the search result display unit 315 and displayed on the screen of the search result display unit 315.

  Like the display example of the screen of the search result display unit 315 shown in FIG. 14, excerpts of important parts of each document in the cluster are displayed for each cluster as representative information. The underline represents the main keyword. When the user selects any important part, the text of the corresponding document is displayed.

  In the present embodiment, the document analysis unit 302 analyzes the document data stored in the document recording unit 301 and allows the user to perform a search. The document group obtained by the search method may be used as the input of the present invention.

  Further, the representative information is generated from each cluster generated by the cluster generation unit 310 in the subsequent processing. However, the generation of the representative information is not essential to the present invention, and the generated cluster is used for other purposes. May be.

  In addition, the representative information generated by the representative information generation unit 314 is displayed on the search result display unit 315. However, instead of displaying the representative information on the screen, the representative information is stored in a storage device. It may be used for purposes.

  It is possible to construct the function of the document search support apparatus shown in FIG. 3 as a program, install it on a computer used as the document search support apparatus, execute it, or distribute it via a network.

  Further, the constructed program can be stored in a portable storage medium such as a hard disk, a flexible disk, or a CD-ROM, and installed in a computer or distributed when the present invention is implemented. .

  The present invention is not limited to the above-described embodiment, and various modifications and applications can be made within the scope of the claims.

  The present invention is applicable to an apparatus for clustering a document group for each similar document, an apparatus for determining a main word of the cluster, an apparatus for determining an important portion in a document, and the like for supporting document search. .

It is a figure for demonstrating the principle of this invention. It is a principle block diagram of this invention. It is a block diagram of the document search assistance apparatus in one embodiment of this invention. It is an example of the document data in the document recording part in one embodiment of this invention. It is a flowchart of the process which the document analysis part in one embodiment of this invention performs. It is an example of the word data in the word recording part in one embodiment of this invention. It is a flowchart of the search request sentence production | generation process in one embodiment of this invention. It is a figure for demonstrating the determination method of the gravity center position in one embodiment of this invention. It is a flowchart of the determination process of the gravity center position in one embodiment of this invention. It is an example of the vector recorded on the vector recording part in one embodiment of this invention. It is an example of the data recorded on the cluster recording part in one embodiment of this invention. It is a flowchart of the main keyword acquisition process in one embodiment of this invention. It is an example of the data recorded on the main keyword recording part in one embodiment of this invention. It is an example of a display of the screen of the search result display part in one embodiment of the present invention.

Explanation of symbols

301 Document storage unit, document recording unit 302 Document analysis unit 303 Word recording unit 304 Search request sentence generation unit 305 Search request sentence recording unit 306 Sentence search unit, sentence search unit 307 Center of gravity selection unit, center of gravity selection unit 308 Vector generation unit, vector Generation unit 309 Vector storage unit, vector storage unit 310 Cluster generation unit, cluster generation unit 311 Cluster storage unit, cluster recording unit 312 Main keyword acquisition unit, main keyword acquisition unit 313 Main keyword recording unit 314 Representative information generation unit, representative information generation Part 315 Search result display part

Claims (10)

Retrieval conditions requested by the user in the document retrieval support apparatus having document storage means, document retrieval means, centroid selection means, vector generation means, vector storage means, cluster generation means, and cluster storage means for storing documents to be searched A document search support method for searching for documents that match
A document search step in which the document search means acquires from the document storage means a document that matches the search condition input by the user;
For each document acquired by the centroid selection means in the document search step, a position where a word A included in each input search condition and a word B included in the same search condition other than the word A appear in the document Based on the distance, a score is assigned to the word A appearing in the document, and the center of gravity for performing processing for specifying the position in the document of the word with the smallest score as the center of gravity position for all words included in the search condition A selection step;
The vector generation unit obtains a vector for each document by totaling words appearing within a certain range centered on the center of gravity position and the number of appearances of the word from each document acquired from the document storage unit. A vector generation step of storing in the vector storage means;
A cluster generation step in which the cluster generation unit calculates a distance between vectors of each document stored in the vector storage unit, forms a cluster by integrating documents having a short distance, and stores the cluster in the cluster storage unit When,
A document search support method characterized in that: In a document search support device further having a main keyword acquisition means,
Based on the cluster configuration document stored in the cluster storage unit, the main keyword acquisition unit refers to the vector in the vector storage unit to obtain a sum of weights for each word, and the word having a large sum The main keyword acquisition step of selecting the top N words as the main keyword from
The document search support method according to claim 1, further performed. In the document search support apparatus further comprising representative information generation means,
A representative information generating step of identifying an appearance position of the main keyword for each of the clustered documents stored in the cluster storage means, and acquiring a certain range centered on the appearance position as an important part of the document; ,
The document search support method according to claim 2 further performed. In the center of gravity selection step,
When the centroid selection unit can divide the search condition input from the user into a plurality of words, the centroid selection unit divides the search condition into words W _{1 to} W _n, and when the word W ₁ appears multiple times in the document, weights determined by the distance between W ₃ located closest to the first occurrence position of the weight determined by the distance, the words W ₁ and W ₂ located closest to the first occurrence position of the word W _1, ..., weights determined by the distance between the Wn located closest to the occurrence position of the word W _1, the sum recorded in the storage means as the first score of the W _1, Similarly, the second said word W _1, third ,... Is also calculated, and the appearance position of W ₁ having the smallest score is determined as the centroid of the word W _{1 in} the document. Similarly, the centroids of the other words W _{2 to} W _n . By determining the position, the barycentric positions of the words W _{1 to} W _n constituting the search condition are determined. Decide for each document,
If the search condition cannot be divided into a plurality of words, the appearance position closest to the beginning of the sentence in the document of the search condition is determined as the center of gravity.
The document search support method according to claim 1. A document search support device for searching for a document that matches a search condition requested by a user,
A document search means for acquiring from the document storage means a document that matches the search condition input by the user;
For each document acquired by the document search means, based on the distance between the position where the word A included in each input search condition and the word B included in the same search condition other than the word A appear in the document , A centroid selection unit that assigns a score to the word A that appears in the document, and performs a process of specifying the position in the document of the word having the smallest score as the centroid position for all words included in the search condition ;
From each document acquired from the document storage means, the words appearing within a certain range centered on the center of gravity position and the number of appearances of the words are totaled to obtain a vector for each document, and the vector storage means Vector generation means for storing;
Calculating a distance between vectors of each document stored in the vector storage means, integrating documents having a short distance to form a cluster, and storing the cluster generation means in the cluster storage means;
A document search support apparatus comprising: Based on the cluster configuration document stored in the cluster storage means, the vector of the vector storage means is referred to find the sum of the weights for each word, and the top N words from the words with the highest sum are used as the main keywords. The main keyword acquisition means to select
The document search support apparatus according to claim 5, further comprising: Representative information generating means for specifying the appearance position of the main keyword for each of the cluster-structured documents stored in the cluster storage means and acquiring a certain range centered on the appearance position as an important part of the document; ,
The document search support apparatus according to claim 6, further comprising: The center of gravity selection means includes
Wherein when dividable search conditions input by the user into a plurality of words, divided into words W ₁ to _W-n, when the word W ₁ appears more than once in the document, the first word W ₁ weights, ..., occurrence position of a word W ₁ that is determined by the distance between W ₃ located closest to the weight, the first occurrence position of a word W ₁ that is determined by the distance between the W ₂ located closest to the occurrence position of the nearest determined by the distance between the W _n in the position weight, the total value is recorded in the storage means as the first score of the W _1, Similarly, the second said word W _1, third, ... occurrence position from also calculates a score for the occurrence position of the score is the smallest W _1, determines that the center of gravity of the word W ₁ in the document in the same manner, to determine the position of the center of gravity of the other word W ₂ to _W-n that Then, the centroid position of the words W _{1 to} Wn constituting the search condition is determined for each document,
8. A document search support apparatus according to claim 5, further comprising means for determining, when the search condition cannot be divided into a plurality of words, the position of the search condition closest to the beginning of the sentence as a centroid. .   A document search support program for causing a computer to function as each means constituting the document search support apparatus according to claim 5.   A computer-readable recording medium storing the document search support program according to claim 9.

Images