JP4205753B2 - Document search system - Google Patents

Description

  The present invention relates to an electronic document search system that registers an electronic document group as an index and searches for a document including a search term of a designated character string.

  There is a full text search as a document search method. However, this method basically requires scanning the contents of all registered documents at the time of retrieval, so that a large amount of retrieval processing time is required for a large number of documents. In order to solve the problem, attempts have been made to increase the speed by devising the index structure and the retrieval processing method. As an index structure, a method of associating only a document ID with an index word (so that only the presence / absence of an index word in each registered document can be known from the index) was initially mainstream. Search index (excess hit) is unavoidable because it is broken down into words and index matching is performed, and post-processing by full-text scanning is required to remove noise, and there is a limit to speeding up. In order to further increase the speed, a method in which the index has additional information such as the appearance position of the index word in each document has been recently proposed.

  For example, in Patent Document 1, a specific number of character chains in a registered document are registered together with an appearance position in an index, and a document including a search word is recorded in the index related to the character chain in the search word when searching. Use to specify.

In Patent Document 2, each character in a registered document is compressed together with the appearance position and registered in the index. At the time of search, a document including the search word is recorded as the appearance position recorded in the index related to the character chain in the search word. Use to identify.
However, in these methods, if the length of the index word is short, the inspection time increases. If the length of the index word is long, the search word having a short index word cannot be searched. If the search word is long, the search time increases. There are problems such as.
JP-A-6-52222 JP-A-8-101848

  The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a document search system that improves search efficiency by devising a character string dividing method and a search evaluation order.

According to the first aspect of the present invention, document dividing means for dividing a registered document into index words based on each character string serving as an index heading, a document frequency representing the number of documents including the index word as an index word, and an index word The index that holds the identifier of the document to be included, the frequency in the document that indicates the number of occurrences of the index word in each document, and the list of occurrence positions in each document of the index word, and the search word in the search condition are recorded in the index Search word dividing means for dividing into index words, and there is no corresponding document to be created corresponding to the case where the search terms are analyzed and the index words or the search word dividing means does not output any index words from the search words A search condition analysis means for generating a search condition tree in which an empty document set representing one or both of them is combined with an operator, and information relating to an index word is extracted from the index based on the search condition tree, and a search result combining process Run A search condition evaluation unit that obtains a search result, and when the search term dividing unit divides the search term into two or more index terms, the search condition analyzing unit includes the two or more search term analyzing units. A candidate decision condition tree obtained by combining index terms with a product set operator that takes the product set of a plurality of search results, and a minimum number of search terms covering the search terms from the two or more index terms are selected between the appearance positions. A detailed decision condition tree synthesized by a position operator that specifies a distance is created, and the search condition evaluation means first obtains a candidate for the corresponding document using the candidate decision condition tree, and then uses the detailed decision condition tree. A candidate document is determined from candidate documents in consideration of the positional relationship, and the candidate determination condition tree includes an index word used for the detailed determination condition tree and the index word used before and after the other index words. Index terms used for detailed decision condition trees Using a remote document frequency is small index word, set union the search condition set union operator taking a set sum of a plurality of search results in the tree if it is nested, is a child node to the set union operator Only when the number of operator child nodes is less than or equal to a threshold value, the latter child node is used as the former child node, and flattening is performed to remove the latter from the former element .

The invention of claim 2 includes a document dividing means for dividing a registered document into index words based on each character string serving as an index headline, a document frequency representing the number of documents including the index word, and the index word using the index word as a headline. The index that holds the document identifier, the frequency in the document that indicates the number of occurrences of the index word in each document, and the list of occurrence positions in each document of the index word, and the search word in the search condition are recorded in the index. Search term dividing means that divides into index words, and there is no corresponding document to be created corresponding to the case where the search terms are analyzed and the index word or the search word dividing means does not output any index words from the search terms A search condition analysis means for generating a search condition tree by combining an empty document set representing one or both of them with an operator, and extracting search result information from the index based on the search condition tree. Run A document search system having search condition evaluation means for obtaining a search result, wherein when the search word dividing means divides a search word into two or more index words, the search condition analyzing means includes the two or more indexes. A candidate decision condition tree obtained by combining words with a product set operator that takes a product set of a plurality of search results, and covers the search words from the two or more index words, and the total document frequency for each index word is minimized. A detailed decision condition tree synthesized by a position operator that selects such an index word and designates a distance between appearance positions, and the search condition evaluation means first uses the candidate decision condition tree as a candidate for the corresponding document. Next, the relevant document is determined from candidate documents in consideration of the positional relationship in the detailed determination condition tree, and the candidate determination condition tree includes an index word used in the detailed determination condition tree and , Before and after other index terms If serial detailed document frequency than index terms used in the determination condition trees using fewer index word, set union operator taking a set sum of a plurality of search results in the search condition trees are nested, the set Only when the number of child nodes of the set sum operator that is a child node is less than or equal to the threshold value, the latter child node is made the former child node and the latter is removed from the former element. It is characterized by that.

The invention according to claim 3 is the invention according to claim 1 or 2 , wherein the set product operator is a candidate decision condition tree obtained by combining the candidate decision condition trees of child nodes with the set product operator, and the search condition The evaluation means is characterized in that, in the evaluation of the set product operator, first, a candidate for the corresponding document is obtained using the candidate determination condition tree, and then a search result is determined using a child node of the set product operator. Is.

The invention of claim 4 is characterized in that, in the invention of claim 3 , an index word node which is a child node of the set product operator is also added to the condition tree for candidate determination of the set product operator.

According to the present invention, since an index word consisting of one character is always extracted, the search processing for a search word consisting of one character can be speeded up.

According to the present invention, candidates are determined using only the index word selected based on the document frequency of the index word, so that the search process can be speeded up.

According to the present invention, when the set sum operator is nested in the search condition, the set sum operator is flattened in consideration of the number of child nodes, so that the search process can be speeded up.

According to the present invention, when the set product operator and the set sum operator are nested in the search condition, the sum standard form conversion is performed in consideration of the number of child nodes, so that the search process can be speeded up.

According to the present invention, when a set condition operator is included in the search condition, the search is performed using the child node candidate determination condition tree, so that the search process can be speeded up.

According to the present invention, when a set difference operator is included in the search condition, the search is performed using the child node candidate determination condition tree, so that the search process can be speeded up.

According to the present invention, when a search condition includes a set sum operator, the search is performed using the child node candidate determination condition tree, so that the search process can be speeded up.

  FIG. 1 is a block diagram showing an outline of a document search system according to the present invention. The present invention includes a document dividing unit 1 that divides a registered document into character strings (hereinafter referred to as index words) serving as index headings, and the number of documents (hereinafter referred to as document frequency) and documents including the index words as index words. The identifier of the document including the index word (hereinafter, document identifier or document ID), the number of occurrences of the index word in each document (hereinafter, frequency in the document), and the appearance position of the index word in each document And a search word dividing means 3 for dividing the search word in the search condition into the index words recorded in the index, and the search word dividing means 3 for dividing the search word in the search condition into index words recorded in the index. Analyzing the search conditions, the index word, or the empty word set indicating that there is no corresponding document created when the search word dividing means does not output any index word from the search word, or one or both of them Synthesized by operator Document having search condition analysis means 4 for generating a condition tree, and search condition evaluation means 5 for appropriately retrieving information on index terms from the index based on the search condition tree and executing search result synthesis processing to obtain search results Search system.

  In the present invention, information necessary for an index for high-speed search is recorded in a document group to be searched by a registration process. As a prior application filed by the same person as the present application, there is an electronic document search system and storage medium disclosed in Japanese Patent Application No. 10-25694. In the system of claim 2 of the above application, when the search word is less than n characters, the search at the end of the registered document may not be correctly searched. On the other hand, in the first embodiment of the present application, the document dividing unit divides a registered document into index words that are n character chains for an integer n of 1 or more. At this time, if n> 1, if n ′ is an integer less than n, in addition to the n-character chain index word, the index word that is the n′-character chain including the last character of the registered document is also divided. As a result. Now, document 1 = “Oh”, document 2 = “Aiueo”, document 3 = “Ae”, and document 4 = “say” are registered. If n = 2, information such as the table shown in FIG. 2 is recorded in the index. Here, the appearance information for each document is the appearance in one document in the range surrounded by {,}, the first field is the document ID, the second field is the document frequency, the third field (( )) Is a sequence of appearance positions.

  The first embodiment is different from the system of claim 2 of the above-mentioned prior application in that an index word having a length of 1 (“A”, “I”, “U”, “E”, “O”) is registered. If n = 3, for example, from document 2, in addition to “Aoi” “Iue” “Ue”, “Eo” “ “O” is also extracted.

  When the search word has n + 1 characters or more, the search word dividing unit divides the search word into index words that are two or more n-character chains covering the search word, and the search condition analysis unit is configured to search between the appearance positions of the index words. Combining with the position operator that specifies the distance. Now, it is assumed that search is performed for a document in which the index word A and the index word B are at a distance x characters by #distance [x] (A, B). For example, if n = 2 and the search word is “Yes”, the search word dividing means divides the search word into two index words “ai” and “say”, and the search condition analyzing means is #distance [1]. A search condition tree corresponding to (Ai, Say) is created. The search condition evaluation means reads the transposed list regarding “ai” and “say”, and these two index words appear at the same time, and this is the search result.

  When the search word is n characters, the search word dividing unit uses the search word as an index, and the search condition analyzing unit creates a search condition including the index word. For example, if n = 2 and the search word is “No”, the search word dividing unit extracts the index word “No” from the search word, and the search condition analysis unit creates a search condition tree corresponding to “No”. . In this case, the search result is document 3.

  If n> 1 and the search word is less than n characters, the search word dividing means outputs all index words existing in the index that matches the search word from the first character, and the search condition analysis means These index words are synthesized by a union operator that takes a union of a plurality of search results. For example, if the search word is “e”, the search word dividing unit outputs “e” and “eo”, and the search condition analysis unit creates a search condition tree of #or. Here, #or (A, B) specifies that a union of a document set including the index word A and a document set including the index word B is searched. In this case, the documents 2 and 3 are search results. In the system of claim 2 of the prior application, only document 2 is searched, and document 3 is not searched.

  In the system of this embodiment, when n> 1, a search word consisting of one character is searched as the union operation result of a plurality of index words. Therefore, there is a problem that a search for a search word consisting of one character is slow. Therefore, in the second embodiment of the present application, index words having different lengths of 1 to N characters are extracted from the registered document to create an index. The index when the four documents used in the description of the first embodiment are registered in the system of the second embodiment with N = 2 is shown in the table of FIG. 2 is different from the table of FIG. 2 in that the appearance of one character is recorded.

  In this system, the processing when the search word is N + 1 characters or more is the same as the system of the first embodiment. On the other hand, when the search word is 1 character or more and N or less, the search word dividing means uses the search word as an index word, and the search condition analysis means creates a search condition including the index word. If the search word is “e”, the search word dividing unit outputs “e”, and the search condition analyzing unit creates a search condition tree of “e”. It can be seen that the documents 2 and 3 become the search results without performing the union operation as in the first embodiment.

  In the system of the second embodiment, the processing of a single character search term is speeded up, but the index is increased accordingly. That is, it is not desirable to set the length of the registered document division to 1 or more as in the second embodiment. Therefore, in the third embodiment, when n is an integer of 2 or more, index words having different lengths of n characters or more and N characters or less are extracted from the registered document to create an index. In this case, the search processing is divided into three cases as in the first embodiment. In the following description, it is assumed that n = 2 and N = 3. If the search term is “Aiue”, the length of the search term is N or more, so the search condition tree is #distance [1] (that is, Iue). If the search term is “Yes”, the search condition tree is “No” because it is n or more and N or less. Similarly, when the search term is “Ai”, the search condition tree is “Ai”. If the search term is “A”, it is less than n, so the search condition tree is #or (Ah, Ah, Ah,…, Ai, Aia,…, Ann,…, Ann) (Here, Hiragana It is assumed that a document consisting only of

  In the system of the third embodiment, in processing a search word with less than n characters, the search word dividing means outputs all index words existing in the index whose head part matches the search word, so The number of index words to be synthesized by the operator becomes very large. Therefore, in the fourth embodiment, the search word dividing means outputs an index word of n characters or less existing in an index whose head part matches the search word. Since n character chains existing in a registered document are always registered at the time of registration, it is sufficient to synthesize character chains of indexes of n or less with a union operator at the time of retrieval. By outputting only index words of n characters or less in this way, the index words to be synthesized by the union operator are reduced, and the search can be speeded up. For example, if n = 2 and N = 3 and the search term is “A”, the search condition tree is #or (Ah, Ah, Ai,..., A).

Japanese has multiple character types such as Katakana, Hiragana, and Kanji, and has the following characteristics.
・ Many words consist of only the same character type.
-The length of words consisting of only the same character type varies depending on the character type.

  Therefore, it is not efficient to ignore the character type when dividing the registered document / search word as in the systems of the first to third embodiments. On the other hand, in the system of the fifth embodiment, efficient registration / search processing is realized by considering the character type. Specifically, individual methods including the processing methods according to the first, second, and third embodiments can be designated for each character type. For example, if the division method for a certain character type is based on the system of the first embodiment, n for the character type, N for the character type in the second embodiment, and the character type in the third embodiment. N and N can be specified.

In the following description, it is assumed that there are three types of character types: Katakana, Kanji, and others. At this time, for example, an index can be created as follows.
For katakana, n = 2 and N = 3 in the process according to the third embodiment.
-For kanji, N = 2 in the process according to the second embodiment.
For others, n = 1 in the process according to the first embodiment.

  For example, if the registered document is “Create a search system”, “Search” “Search” “Search” “Sys” “System” “Ste” “Stem” “Tem” “M” “O” “Saku” “Ru” It is divided like “M” is cut out because it is a character chain of less than n located at the end of the continuous part of katakana “system”.

  The search process differs depending on whether or not the search word is composed of only the same character type. In the case where only the same character type is used, the search process is a search process corresponding to the character type division method. For example, if the search term is “search”, the processing of the second embodiment is performed, so the search condition tree is “search”. On the other hand, if it is composed of different character types, a search condition tree (this is called a sub-search condition tree) is created by the above method for consecutive parts of the same character type, and these are combined with a position operator. The final search condition tree. For example, assume that the search term is “search system”. In this case, first, a sub-search condition tree “search” for the continuous part “search” of kanji and a sub-search condition tree #distance [1] (system, stem) for the continuous part “system” of katakana are created. Next, #distance [2] (search, #distance [1] (system, system)) that combines two sub-search condition trees at the distance (two characters) between “search” and “system” is the final search. It becomes a conditional tree.

  However, there is a problem that the search is not efficient when the search word is composed of different character types, and the length of the portion other than the end of the continuous portion of the same character type is shorter than the minimum cutout length n of the character type. For example, if the search term is “M”, the sub search condition tree #or (M, Mua,...) For “M” and the sub search condition tree for “ The search condition tree is #distance [1] (#or (mu, mua,...)). However, among the index words expanded by the union operator in the sub search condition tree for "mu", since "m" is a character chain that is continuous with characters other than "o", " The distance between and cannot be 1. Therefore, even if the length of the character other than the end of the continuous portion of the same character type is shorter than the minimum cut-out length n of the character type, in the search process, for the continuous portion, the index that is the continuous portion itself Use words. That is, when the search term is “mu”, the sub search condition tree for “m” may be “mu”, and the final search condition tree is #distance [1] (mu,). In this case, the search condition tree becomes simple and the search can be speeded up.

  In the system of the fifth embodiment, the search word is composed of different character types, and the search is not efficient when the length of the last part of the continuous portion of the same character type is shorter than the minimum cutout length n of the character type. (This applies only when the cut-out method for the end of the continuous part of the same character type is n> 1 and the third embodiment in the first embodiment). For example, if the search term is “search”, the sub search condition tree “search” for “search” and the sub search condition tree #or (si, shear,...) The search condition tree is #distance [1] (#or (si, shear,...). However, if the union operator is included in the position operator in this way, the search processing becomes complicated, and the search time is increased. Will increase.

  On the other hand, in the system of the sixth embodiment, for the character type to which the case of n> 1 and the third embodiment is applied in the first embodiment, the document dividing unit determines the character type in the registered document. An index word that is an n-character chain with a continuous part, an index word that is an n'-character chain that includes the last character of the continuous part of the character type when n 'is an integer less than n, and a continuous part of the character type The index word is divided into n 'character chains including the first character. For example, if the registered document is “Create a search system”, “Search” “Search” “Search” “Shi” “Sys” “System” “Ste” “Stem” “Tem” “M” “Create” “Create” ”And“ Ru ”. The difference from the processing according to the fifth embodiment is that “shi”, which is a single character chain including the first character of the katakana continuous part, is cut out independently.

  The search process differs depending on whether or not the search word is composed of only the same character type, but if it is composed of only the same character type, the search process may be performed in exactly the same way as in the fifth embodiment. On the other hand, in the case of being composed of different character types, the processing in the case where the length of the last part of the continuous portion of the same character type is shorter than the minimum cut-out length n of the character type is different from the fifth embodiment. At this time, since the n ′ character chain including the first character of the continuous part of the same character type is extracted as an index word at the time of registration, the index word that is the continuous part itself is also applied to the continuous part at the time of search. Use it. That is, when the search word is “search”, the sub-search condition tree for “shi” may be “shi”, and the final search condition tree is #distance [1] (search, shi).

  In the system of the fifth embodiment, a two-character chain composed of different character types is not recorded in the index and is not used for searching. However, some combinations of character types are often specified as search terms. For example, a combination of kanji and hiragana constitutes a use word like “movement” and is often used as a search word. However, in the system of the fifth embodiment, “movement” is processed as a search condition tree of #distance [1] (motion, ki), and therefore it takes time to search.

  Therefore, in the system according to the seventh embodiment, a two-character chain composed of different designated character types is used as an index. The document dividing means divides each consecutive portion of the same character type in the registered document into index words corresponding to n or N for the character type, and also extracts a two-character chain consisting of designated different character types as an index word. For example, in addition to designating the processing method for each character type as in the fifth embodiment, a combination of kanji and hiragana is also used for the index. In this case, if the registered document is “Create a search system”, “Search” “Search” “Search” “Sys” “System” “Ste” “Stem” “Tem” “M” “O” “Saku” “ In addition to “Ru”, “Make”, which is a combination of kanji and hiragana, is also extracted.

  The search word dividing unit also divides the search word in the same manner as the document dividing unit. The search condition analysis means creates a search condition tree in the same manner as before when the search character analysis unit does not include a two-character chain consisting of different character types for which the search word is designated. When two character chains consisting of different character types are included, it is regarded as a target for creating a sub-search condition tree that connects the linked side portions connected by the position operator with the position operator, and the continuous portion of the same character type is divided. Two-character chains consisting of different character types are also extracted as index words. Then, the search condition analysis means assembles the extracted index terms into a sub search condition tree with a position operator.

  For example, when a combination of kanji and katakana is specified, if the search term is “search system”, the entire “search system” is regarded as the sub search condition tree creation target, and the continuous part of “kanji” “search” and katakana Extract "search", "system" and "stem" from the continuous part "system", and also extract "search" consisting of kanji and katakana. From this, a sub-search condition tree of #distance [1] (search, #distance [1] (search, #distance [1] (system, stem))) is created (in this case, one sub-search condition from the search term) Since the tree creation target is not created, this sub search condition tree becomes the final search condition tree).

  If the search term is “Create search system”, “Search system” and “Create” are largely divided into “mu” and “saku”, which are the sub search condition tree creation targets. This is because no combination is specified. The previous sub search condition tree is created from the “search system”, and the sub search condition tree “create” is created from “create”. Therefore, the final search condition tree is #distance [6] (#distance [1] (search, #distance [1] (search, #distance [1] (system, stem))), created).

  In the system of the seventh embodiment, when the search word includes a two-character type chain consisting of different character types designated, and the length of the continuous portion of the preceding character type is one character, the search processing is wasted. . For example, it is assumed that the division method for kanji and hiragana is n = 2 in the first embodiment, and kanji and hiragana are designated. At this time, if the search word is “movement”, the kanji continuous part “movement” is #or (movement, movement,...), And the hiragana continuous part “giga” is “kiga”. The final search condition tree that is created and combined with the 'motion' generated from the continuous "motion" of kanji and hiragana is #distance [1] (# distance [0] (# or (motion, motion , ...), movement), and giga). However, since #distance [0] (#or (motion, motion,...), Motion) is equivalent to 'motion', it is useless to create the above-described search condition tree.

  Therefore, in the system of the eighth embodiment, when the search word includes a two-character chain consisting of different character types designated, and the length of the continuous portion of the preceding character type is one character, the search word dividing means The index word corresponding to the character on the front side is not generated. That is, “motion” and “giga” are extracted from the search term “motion is”, and the final search condition tree is also #distance [1] (motion, kiga). As a result, the search process is simplified and the search process is also speeded up.

  In the system of the seventh embodiment, when the search term includes a two-character chain consisting of different character types designated, and the length of the continuous portion of the subsequent character type is one character, the search processing is wasted. . For example, it is assumed that the division method for kanji and hiragana is n = 2 in the first embodiment, and kanji and hiragana are designated. At this time, if the search word is “action”, “action” is created from the continuous part “motion” of kanji, and #or (ga, there, etc.) is created from the continuous part “ga” of hiragana. , And a combination of 'songs' generated from consecutive kaku and hiragana "sakuga", the final search condition tree #distance [1] (operation, #distance [1] (song, #or ( , There, ...))) is created. However, since #distance [1] (work is #or (has, is,...)) Is equivalent to “work is”, it is useless to create the search condition tree described above.

  Therefore, in the system of the ninth embodiment, when the search word includes a two-character chain consisting of different character types designated, and the length of the continuous portion of the subsequent character type is one character, the search word dividing means An index word corresponding to one character on the subsequent side is not generated. That is, “motion” and “sakuga” are extracted from the search term “motion is”, and the final search condition tree is also #distance [1] (motion is a product). As a result, the search process is simplified and the search process is also speeded up.

In the system of the seventh embodiment, there is a problem that the search is not efficient when the search word is composed of the same character type and the length is shorter than the minimum cut-out length n of the character type (this is the case). This applies only when the cutout method for the end of the continuous part of the same character type is the case of n> 1 and the third embodiment in the first embodiment). For example, if the division method for kanji and hiragana is n = 2, hiragana / kanji is specified in the first embodiment, and the search word is “a”, the search condition tree for “a” is #or (a , Oh, ..., ann, aa ...).
However, in the document registration, if “A” is followed by a different character type, “A” is extracted, so all documents including index words that are a series of hiragana and kanji are index corresponding to “A”. Is recorded. Therefore, it is useless for the search word dividing means to generate hiragana and kanji.

  In the system of the tenth embodiment, the search word dividing means is configured such that, when the search words are composed of the same character type and the length is shorter than the minimum cutout length n of the character type, An index word composed only of the character type existing in the index whose head part matches is output. For example, the search condition tree for “A” is #or (Ah, Ah,... An). As a result, the useless index word as described above is not used in the search process, so that the search is speeded up.

  In the conventional methods, when the search word is divided into two or more index words, the search is performed using only the search condition synthesized by the position operator. In this method, there is a possibility that useless position matching processing occurs. For example, Document 1 = “Aiueo”, Document 2 = “Aiuei”, Document 3 = “Aiueo”, Document 4 = “Aiuee”, Document 5 = “Aiueo”, and the system according to the first embodiment. When an index is created with n = 2 by the method, the table shown in FIG. 4 is obtained.

  When the search word “Ai” is processed by the method described in the first embodiment, “Ai”, “Good”, and “U” are obtained as index words, and #distance [2] (#distance [1] ( A search condition tree is created. However, if “ai” and “say” are at a distance of 2, there is always “good” between them, so the search condition tree #distance [2] (ai, say) is sufficient. . The search condition processing including the position operator first identifies the document ID in which all index words appear, and then checks whether the deviation of the appearance position between the index words in the document is as specified by the position operator. By doing so, it can be realized by the procedure of determining whether the document really meets the search condition. In the above search terms, two index words, “ai” and “say”, may be used when inspecting the deviation of the appearance position, but “good” should also be used when specifying the document ID. Increased efficiency. This is because “good” is not registered in the index, and it can be found that there is no document that corresponds to “ai” simply by checking whether “good” appears (this technique is based on the above-mentioned prior application patent application). (Proposed in claims 8 and 9 of Hei 10-256974). Hereinafter, the document ID specifying process is referred to as candidate determination, and the appearance position shift inspection process is referred to as detailed determination. The search condition tree used for the former is a candidate determination condition tree, and the search condition tree used for the latter is a detailed determination condition. Called a tree. Then, for this search term, the candidate determination condition tree: #and (ai, i.e.) and the detailed determination condition tree: #distance [2] (ai i.e.) are used. Here, #and uses the set product of the search results for each operand as the search result, and is called a set product operator.

  When the above method is applied to the search term “Iueo”, the candidate decision condition tree: #and (say, ue, eo), the detailed decision condition tree: #distance [2] (say, eo) Become. However, in this case, “up” always appears in the document in which “say” appears, so adding “up” to the candidate decision condition tree has no effect of narrowing down the candidate document, and the index Since the number of words increases, the processing increases, and there is a problem that the search becomes slow.

  In the eleventh and twelfth embodiments, only index words that can be expected to narrow down candidate documents are added to the candidate decision condition tree to speed up the search. That is, instead of simply adding an index word extracted from a search term, the candidate decision condition tree includes an index word used in the detail decision condition tree and detailed judgments before and after the other index words. Use index terms that have less document frequency than the index terms used in the conditional tree. For example, in the search word “Aii”, the document frequency of “Good” is 0 with respect to the document frequency 5 of “Ai” and “Ai” used for the detailed determination, so “Good” is used. On the other hand, in the search term “Iueo”, the document frequency of “U” is not as small as 5 compared to the document frequency of “U” used for detailed determination, so “U” is not used. In the eleventh embodiment and the twelfth embodiment, the index word used for the detailed determination is selected as the index word used in the eleventh embodiment while the twelfth embodiment selects the index word. The difference is that the document frequency with the smallest total is selected.

  There is Japanese Patent Application No. 10-020840 as another prior application filed by the same person as the present application. In claim 8 of the prior application, a set sum operation that takes a set sum of a plurality of search results in a search condition tree. If the child is nested, the latter child node is made the former child node, and the latter is removed from the former element. For example, if the search condition tree is #or (#or (Tokyo, Edo), Osaka), it is #or (Tokyo, Edo, Osaka). Here, #or is a set sum operator.

  However, when the set-sum operator has a large number of child nodes of the set-sum operator, which is a child node, processing cost is required for the flattening operation. Therefore, in the thirteenth embodiment of the present application, when a set sum operator that takes a set sum of a plurality of search results is nested in the search condition tree, a set sum operator that is a child node is included in the set sum operator. Only when the number of child nodes is less than or equal to the threshold value, the latter child node is used as the former child node, and the latter is removed from the former element. As a result, it is possible to avoid an increase in processing cost when the set sum operator has a large number of child nodes in the set sum operator.

  If there is a set-sum operator in the child node of the set-product operator that takes the set product of multiple search results in the search condition tree, the sum-standard form is equivalent to the set-product operator coming to the child node of the set-sum operator Can be converted. For example, #and (#or (Tokyo, Edo), Osaka) becomes #or (#and (Tokyo, Osaka), #and (Edo, Osaka)). If the transformation is performed in this way, the set of documents to be subjected to the set sum operation becomes smaller, so that the search can be made efficient.

  However, if the number of child nodes of the set sum operator in the child nodes of the set product operator is large, the number of child nodes of the set sum operator after conversion may become enormous. For example, if it is #and (#or (Tokyo, Tokyo, Tokyo, Tokyo, Tokyo, Edo, Edo, Ed, EDO, edo), #or (Osaka, Osaka, Osaka, OOSAKA, Oosaka)) The number of child nodes is as large as 10 × 5 = 50, and the processing cost for the conversion work increases.

  Therefore, in the fourteenth embodiment of the present application, when there is a set sum operator in the child node of the set product operator that takes the set product of a plurality of search results in the search condition tree, the set product is set in the child node of the set sum operator. Only when the number of child nodes of the set sum operator after conversion is less than or equal to the threshold value is converted into the equivalent form to which the operator comes. As a result, it is possible to avoid an increase in processing cost when the number of child nodes of the set sum operator after conversion increases.

  Consider a search condition in which search terms divided into a plurality of index terms are combined by a set product operator. For example, when an index is created by the processing (n = 2) by the system of the first embodiment, #and (#distance [2] (#distance [1] ] (Pre, phosphorus), data), #distance [2] (#distance [1] (sys, ste), tem)) is created. In this search condition tree, first, a document corresponding to “printer” is determined by #distance [2] (#distance [1] (pre, phosphorus), data), and the document is #distance [2] (# The distance [1] (sys, ste), tem) is used to determine whether it corresponds to “system”, and if so, it is evaluated by adding it to the search result.

  At this time, the search condition tree is and (#distance [2] (pre, data), #distance [2] (sys, tem)), and claims 8 and 9 of the Japanese Patent Application No. 10-256974. Alternatively, in the case of processing by the system of claims 11 and 12, for example, for #distance [2] (pre, enter), #and (pre, phosphorus, enter), #distance [2] (sys, #And (sys, stem, tem) is set as a candidate decision condition tree. Therefore, first, a candidate document including “printer” is determined by #and (pre, phosphorus, and data), and it is determined whether or not the document includes “system” by #and (sys, stem, and tem). If it is satisfied, it is determined whether the position condition for including “printer” is satisfied by #distance [2] (pre, data), and if it is satisfied, “system” is determined by #distance [2] (sys, tem). The document can be determined by the procedure of determining whether the position condition for inclusion is satisfied. In this procedure, the number of times of determining whether the position condition is satisfied is reduced, so that the search process can be speeded up.

  Therefore, in the fifteenth embodiment of the present application, a candidate determination condition tree is a combination of child node candidate determination condition trees using a set product operator. For example, the candidate decision condition tree for the previous search condition is #and (pre, phosphorus, data, cis, stem, tem). In this method, since the time and labor for determining candidate documents is reduced, the search process can be further speeded up.

  In the method of the fifteenth embodiment, for a search condition having an index word node as a child node such as #and (printer, apparatus), the node is not included in the candidate determination condition tree, and the candidate determination condition The tree is #and (pre, phosphorus, data). Therefore, there is a possibility that candidate documents are not sufficiently narrowed down and search processing is delayed. Therefore, in still another example of the present invention, a child node also adds an index word node as a condition tree for determining a candidate for a set product operator. For example, the candidate decision condition tree for the previous search condition is #and (pre, phosphorus, data, device). In this method, candidate documents are narrowed down, so that the search process can be speeded up.

  Consider the case of a search condition that is synthesized by a set difference operator (hereinafter referred to as # and-not) that takes a set difference between two search results. For example, when an index is created by the method of the first embodiment (n = 2), # and-not (#distance [2] (#distance [ 1] (pre, phosphorus), data), #distance [2] (#distance [1] sys, ste), tem)) is created. In this search condition tree, first, a document corresponding to “printer” is determined by #distance [2] (#distance [1] (pre, phosphorus), data), and the document is #distance [2] (#distance [1] (Sys, Ste), Tem), it is judged whether it corresponds to “system”, and if it does not, it is evaluated by the procedure of adding to the search result.

  At this time, the retrieval condition tree is # and-not (#distance [2] (pre, data), #distance [2] cis, tem)), and claims 8 and 256 of the above Japanese Patent Application No. 10-256974. 9 or claims 11 and 12, for example, for #distance [2] (pre, enter), #and (pre, phosphorus, enter), #distance [2] cis, tem) #And (Sys, Stem, Tem) is set as a candidate decision condition tree. Therefore, first, a document including “system” is determined using #and (pre, phosphorus, data) and #distance [2] (pre, data), and the document is #and (sys, stem, tem). To determine whether the document is a candidate, and if it is not satisfied, #distance [2] sys, tem) can be used to determine whether the document satisfies the position condition for including “system”. . This procedure is the seventeenth embodiment of the present application, and the number of times to determine whether the position condition is satisfied is reduced, so that the search process can be speeded up.

  Consider a search condition in which search terms divided into a plurality of index terms are combined by a set sum operator. For example, when an index is created by the method of the first embodiment (n = 2), # or # distance [2] (pre, data), #distance from the search condition #or (printer, system) [2] A search condition tree called “sys, tem)) is created.

  At this time, in the method of claim 8 or claim 9 or claim 11 or claim 12 of the Japanese Patent Application No. 10-25694, for example, for #distance [2] (pre, enter), # For and (pre, phosphorus, data) and #distance [2] (sys, tem), #and (cis, stem, tem) is set as a candidate determination condition. This search condition tree can be evaluated as follows. First, a document corresponding to “printer” is determined using #and (pre, phosphorus, data) and #distance [2] (pre, data). Next, a document corresponding to “system” is determined using #and (sys, stem, tem) and #distance [2] (sys, tem). Finally, the set sum of the two is obtained.

  However, when determining a document that corresponds to “system”, it has already been determined that a document that corresponds to “printer” is included in the search results, so it is necessary to accurately determine whether it corresponds to “system”. Absent. Therefore, in the second process in the above description, a candidate document including “system” is determined by #and (sys, stem, tem), and #distance [2] if the document does not correspond to “printer”. (Sys, Tem) “If the position condition for including the system is satisfied, it is added and if it is satisfied, it is added to the result set. It is unnecessary and can proceed to determination of the next candidate document. This process is implemented in the eighteenth embodiment of the present application, and in this method, the number of times of determining the position condition for the child node can be reduced, so that the search process is speeded up.

It is a block diagram which shows the outline | summary of the document search system by this invention. It is a table | surface for demonstrating 1st Embodiment. It is a table | surface for demonstrating 2nd Embodiment. It is a table | surface for demonstrating 10th Embodiment.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Document division | segmentation means, 2 ... Index, 3 ... Search word division | segmentation means, 4 ... Search condition analysis means, 5 ... Search condition evaluation means.

Claims (4)

Document dividing means for dividing a registered document into index words based on each character string serving as an index heading, document frequency representing the number of documents including the index word with the index word as a headline, an identifier of a document including the index word, and an index word An index that holds the frequency of occurrence in each document and the list of occurrence positions of each index word in each document, and the search word in the search condition is divided into index words recorded in the index. A search word dividing unit that analyzes the search condition and an empty word set indicating that there is no corresponding document to be created corresponding to an index word or a case where the search word dividing unit does not output any index word from the search word or A search condition analysis unit that generates a search condition tree obtained by combining one or both of them with an operator, and retrieves information on the index word from the index based on the search condition tree and executes a search result combining process to obtain a search result. obtain A search condition evaluation unit, wherein when the search word dividing unit divides the search word into two or more index words, the search condition analysis unit converts the two or more index words into a plurality of index words. A position for specifying a distance between occurrence positions by selecting a candidate decision condition tree synthesized by a product set operator that takes a product set of search results and a minimum number that covers the search word from the two or more index words A detailed decision condition tree synthesized by an operator, and the search condition evaluation means first obtains a candidate for the corresponding document using the candidate decision condition tree, and then considers the positional relationship in the detailed decision condition tree. The candidate document is determined from among the candidate documents, and the detailed decision condition tree in the candidate decision condition tree is the index word used in the detailed decision condition tree and the other of the other index words. Less frequent than index terms used in Using the index words have,
When a set sum operator that takes a set sum of a plurality of search results is nested in the search condition tree, the number of child nodes of the set sum operator that is a child node of the set sum operator is less than or equal to a threshold value A document retrieval system characterized by flattening only the latter child node as the former child node and removing the latter from the former element . Document dividing means for dividing a registered document into index words based on each character string serving as an index heading, document frequency indicating the number of documents including the index word, using the index word as a heading, identifier of the document including the index word, index word Divides the search term in the search condition into an index word recorded in the index and an index that holds the frequency in the document that represents the number of occurrences in each document and the list of occurrence positions in each document. Search term dividing means, an index word or an empty document set indicating that there is no corresponding document to be created corresponding to a case where the search term dividing means does not output any index word from the search term by analyzing the search condition Search condition analysis means for generating a search condition tree obtained by combining one or both of the above with an operator, information on index terms is extracted from the index based on the search condition tree, and search result combining processing is executed to obtain a search result Inspection A document search system having condition evaluation means, wherein when the search word dividing means divides a search word into two or more index words, the search condition analysis means searches the two or more index words for a plurality of searches. A candidate decision condition tree synthesized by a product set operator that takes the product set of the results, and an index word that covers the search term from the two or more index terms and minimizes the total document frequency for each index term A detailed decision condition tree synthesized by a position operator that selects and designates the distance between the appearance positions is created, and the search condition evaluation means first obtains a candidate for the corresponding document from the candidate decision condition tree, and then A relevant document is determined from candidate documents in consideration of the positional relationship in the detailed determination condition tree, and the candidate determination condition tree includes an index word used for the detailed determination condition tree and other indexes. The detailed judgment condition before and after the word Use the document less frequently index terms than the index words used to,
When a set sum operator that takes a set sum of a plurality of search results is nested in the search condition tree, the number of child nodes of the set sum operator that is a child node of the set sum operator is less than or equal to a threshold value A document retrieval system characterized by flattening only the latter child node as the former child node and removing the latter from the former element . 3. The document retrieval system according to claim 1, wherein the set product operator is a candidate decision condition tree obtained by combining the candidate decision condition trees of child nodes with the set product operator, and the search condition evaluation means includes: In the evaluation of the set product operator , a document search system characterized in that first a candidate for a corresponding document is obtained using the candidate decision condition tree, and then a search result is determined using a child node of the set product operator . . 4. The document search system according to claim 3 , wherein an index word node that is a child node of the set product operator is also added to the condition tree for determining a candidate for the set product operator .

Images