JP5426710B2 - Search support device, search support method and program - Google Patents

Description

Embodiments described herein relate generally to a search support apparatus , a search support method, and a program .

  Document search is a technique for searching a document including a search keyword designated by a user from a set of documents to be searched. Here, the document includes not only the digitized document itself but also various contents having text data. Conventionally, various search support methods have been proposed in order to reduce a user's operation burden in document search.

  For example, a method of presenting search keyword candidates to a user based on a history of past search expressions is known. In this method, when a search keyword “diffusion weighted image”, for example, is input by the user's input operation, “delayed phase”, “fat”, “high signal”, “axial position” are selected as subsequent search keyword candidates. Suggest words that frequently occur in the history of past search formulas. According to this method, it is possible to easily generate a search expression including a plurality of search keywords and reduce the operation burden on the user. However, this method requires a lot of history in order to propose appropriate search keyword candidates, and if the history is biased or insufficient, the quality of the proposal decreases, and the target document There is a concern that it cannot be searched.

  There is also known a method of proposing search keyword candidates to a user using a co-occurrence dictionary that defines a combination of two words having a co-occurrence relationship. In this method, when a word is input as a search keyword by the user's input operation, a word registered in the co-occurrence dictionary as a word having a high co-occurrence probability for the input search keyword as a subsequent search keyword candidate Propose. According to this method, it is possible to easily generate a search expression including a plurality of search keywords and reduce the operation burden on the user. However, with this method, it is necessary to prepare a co-occurrence dictionary in which co-occurrence relationships between words are defined in advance, and if the prepared co-occurrence dictionary does not match the document set to be searched, There is a concern that the quality will deteriorate and the target document cannot be retrieved.

  As described above, in the conventional technology that facilitates the generation of a search expression including a plurality of search keywords, the quality of the proposal may be deteriorated and the target document may not be searched, and improvement is required.

Japanese Patent No. 2850952 JP 2006-48286 A

The problem to be solved by the present invention is to propose search keyword candidates suitable for a set of documents to be searched without requiring the user to perform complicated operations in advance, and appropriately support document search. A search support apparatus , a search support method, and a program are provided.

The search support device of the embodiment includes an extraction unit, a calculation unit, a first detection unit, a first generation unit, a second generation unit, a first recognition unit, a co-occurrence propagation unit, and a second detection unit. And a third generation unit, a second recognition unit, a presentation unit, and a search unit. The extraction unit extracts keyword candidates from the document set to be searched. The calculation unit calculates a co-occurrence probability, which is a probability that one keyword candidate appears in the same document in the document set together with another keyword candidate, for the extracted combination of two keyword candidates. The first detection unit detects a co-occurrence keyword set that is a combination of two keyword candidates in which the co-occurrence probability satisfies the first condition. The first generation unit generates a co-occurrence dictionary that is a set of dictionary elements having one keyword candidate of the co-occurrence keyword set as a headword and the other keyword candidate as a co-occurrence word. The second generation unit generates a character string complement rule that is a rule for complementing the input character string to obtain keyword candidates included in the co-occurrence keyword set. A 1st recognition part recognizes the keyword candidate obtained by complementing an input character string according to the said character string complementation rule as an input keyword. The co-occurrence propagation unit refers to the co-occurrence dictionary, acquires a co-occurrence word of a dictionary element having the input keyword as an entry word, and acquires a co-occurrence word of a dictionary element having the acquired co-occurrence word as an entry word Repeat the process. The second detection unit detects a zero co-occurrence keyword set that is a combination of two keyword candidates having the co-occurrence probability of zero. The third generation unit generates a zero co-occurrence dictionary which is a set of dictionary elements in which one keyword candidate of the zero co-occurrence keyword set is a headword and the other keyword candidate is a zero co-occurrence word. The second recognizing unit refers to the zero co-occurrence dictionary, and includes the zero co-occurrence keyword set in a word string connecting the input keyword and the co-occurrence word acquired by the processing by the co-occurrence propagation unit. If there is a word string that simultaneously includes two keyword candidates that constitute a word string, the word string is excluded, and among the remaining word strings, a word string that satisfies the second condition is recognized as a proposed word string. The presenting unit presents the proposed word string. When the presented proposed word string is selected, the search unit generates a search formula based on the proposed word string and searches the document set.

FIG. 1 is a block diagram illustrating a functional configuration of the search support apparatus according to the first embodiment. FIG. 2 is a diagram illustrating an example of a document set to be searched. FIG. 3 is a diagram illustrating an example of keyword candidates extracted from a search target document set. FIG. 4 is a diagram illustrating an example of the relationship between the extracted keyword candidates and the appearance frequency. FIG. 5 is a diagram illustrating an example of the appearance frequency of a combination of two keyword candidates. FIG. 6 is a diagram illustrating an example of the co-occurrence probability between two keyword candidates. FIG. 7 is a diagram illustrating an example of a co-occurrence network. FIG. 8 is a flowchart illustrating an example of processing by the co-occurrence keyword set detection unit. FIG. 9 is a diagram illustrating an example of the data structure of dictionary elements that constitute the co-occurrence dictionary. FIG. 10 is a diagram illustrating an example of a co-occurrence dictionary. FIG. 11 is a diagram illustrating an example of a Patricia tree. FIG. 12 is a diagram illustrating an example of a data structure of a word string. FIG. 13 is a flowchart illustrating an example of processing by the input keyword recognition unit, the co-occurrence propagation unit, and the suggested word string recognition unit. FIG. 14 is a block diagram illustrating a functional configuration of the search support apparatus according to the second embodiment. FIG. 15 is a diagram illustrating a combination of two keyword candidates having a co-occurrence probability of zero. FIG. 16 is a diagram illustrating an example of a data structure of dictionary elements constituting the 0 co-occurrence dictionary. FIG. 17 is a diagram illustrating an example of a 0 co-occurrence dictionary. FIG. 18 is a flowchart illustrating an example of processing by the input keyword recognition unit, the co-occurrence propagation unit, and the suggested word string recognition unit.

  Hereinafter, a search support device and a search support method of an embodiment will be described in detail with reference to the drawings.

(First embodiment)
FIG. 1 is a block diagram illustrating a functional configuration of the search support apparatus 100 according to the first embodiment. As shown in FIG. 1, the search support apparatus 100 according to the present embodiment includes a keyword candidate extraction unit (extraction unit) 101, an index generation unit 102, a co-occurrence probability calculation unit (calculation unit) 103, and a co-occurrence keyword set. A detection unit (first detection unit) 104, a co-occurrence dictionary generation unit (first generation unit) 105, a Patricia tree generation unit (second generation unit) 106, an input reception unit 107, and an input keyword recognition unit (first 1 recognition unit) 108, co-occurrence propagation unit 109, proposed word string recognition unit (second recognition unit) 110, display unit (presentation unit) 111, and search unit 112.

  The keyword candidate extraction unit 101 extracts keyword candidates that are search keyword candidates from the search target document set 300. The method for extracting keyword candidates from the search target document set 300 is not particularly limited, and various methods can be used. For example, referring to a vocabulary dictionary in which many words that can be used as search keywords are registered, the words registered in the vocabulary dictionary are extracted as keyword candidates from each document included in the search target document set 300. it can.

  Here, the search target document set 300 can be arbitrarily designated by the user using the search support apparatus 100 of the present embodiment. That is, in the search support apparatus 100 according to the present embodiment, when the user specifies the document set 300 to be searched, the keyword candidate extraction unit 101 executes a process of extracting keyword candidates from the document set 300 specified by the user. . Hereinafter, as a specific example of the document set 300 to be searched, a book information set as illustrated in FIG. 2 is illustrated. In the example of FIG. 2, five pieces of book information are shown as documents 1 to 5, and documents 1 to 5 are “book name”, “author name”, “translator name”, and “publisher”, respectively. , “ISBN”, and “summary”, corresponding book information is described in, for example, text data. The keyword candidate extraction unit 101 extracts keyword candidates from text data described for each item of each document. Note that the search target document set 300 is generally a set of many documents, but only five documents are illustrated here for the sake of simplicity.

  FIG. 3 shows an example of keyword candidates extracted by the keyword candidate extraction unit 101 from the document set 300 illustrated in FIG. FIG. 4 shows an example of the relationship between the extracted keyword candidates and the appearance frequency. The appearance frequency of the keyword candidates represents the number of documents in which the same keyword candidate appears among the documents in the document set 300. For example, in the document set 300 illustrated in FIG. 2, “Hanako”, which is one of the keyword candidates, is extracted from each of the documents 1 to 5, so the frequency of appearance of the keyword candidate “Hanako” is 5.

  When the keyword candidate extraction unit 101 extracts keyword candidates from the document set 300, the keyword candidate extraction unit 101 outputs the extracted keyword candidates to the index generation unit 102 together with document pointer information indicating a document in which the keyword candidates appear. Further, the keyword candidate extraction unit 101 outputs a combination of the extracted two keyword candidates to the co-occurrence probability calculation unit 103. Further, it is desirable that the keyword candidate extraction unit 101 retains information as shown in FIG. 4 indicating the relationship between the extracted keyword candidates and the appearance frequency using a storage unit (not shown).

  The index generation unit 102 generates an index 301 for the search target document set 300 based on the keyword candidates and the document pointer information input from the keyword candidate extraction unit 101. The index 301 holds document pointer information indicating a document in which a search keyword appears when a certain search keyword is input.

  The co-occurrence probability calculation unit 103 is a probability that one keyword candidate appears in the same document in the document set 300 together with another keyword candidate for each combination of two keyword candidates input from the keyword candidate extraction unit 101. Calculate the probability of occurrence.

  FIG. 5 is a diagram illustrating an example of the appearance frequency of a combination of two keyword candidates. FIG. 6 is a diagram illustrating an example of the co-occurrence probability between two keyword candidates calculated from the appearance frequency of FIG. 5 and the total number of documents in the document set 300. FIG. 7 is a diagram illustrating an example of a co-occurrence network generated based on the co-occurrence probability of FIG. The appearance frequency of the combination of two keyword candidates represents the number of documents in which two keyword candidates appear simultaneously in the documents in the document set 300. For example, in the document set 300 illustrated in FIG. 2, the keyword candidates “Hanako” and “Fuchu” appear simultaneously in Document 1, Document 2, and Document 5, and therefore a combination of “Hanako” and “Fuchu”. The appearance frequency is 3.

  The co-occurrence probability calculation unit 103 first obtains the appearance frequency of each combination for each combination of two keyword candidates input from the keyword candidate extraction unit 101. Then, the co-occurrence probability calculation unit 103 calculates the co-occurrence probability between the two keyword candidates by dividing the appearance frequency of the obtained combination by the total number of documents included in the document set 300. Then, the co-occurrence probability calculation unit 103 generates a co-occurrence network as shown in FIG. 7 that holds the calculated co-occurrence probability, and outputs the generated co-occurrence network to the co-occurrence keyword set detection unit 104. As shown in FIG. 7, the co-occurrence network is such that keyword candidates are nodes, nodes in a co-occurrence relationship are linked by links, and the co-occurrence probabilities are associated with the links.

  The co-occurrence keyword set detection unit 104 detects a co-occurrence keyword set that is a combination of two keyword candidates whose co-occurrence probability satisfies the first condition based on the co-occurrence network input from the co-occurrence probability calculation unit 103, Output to co-occurrence dictionary generation unit 105 and Patricia tree generation unit 106. Here, as the first condition, for example, a condition that the co-occurrence probability is larger than the threshold value α with respect to the threshold value α set so that the number of co-occurrence keyword sets finally obtained is equal to or larger than the threshold value β. Can be used. The first condition is not limited to this example, and various predetermined conditions can be used.

  FIG. 8 is a flowchart illustrating an example of processing performed by the co-occurrence keyword set detection unit 104 when a co-occurrence keyword set having a co-occurrence probability larger than the threshold value α is detected.

  The co-occurrence keyword set detection unit 104 first sets a threshold value α to a predetermined initial value, and extracts a combination of keyword candidates whose co-occurrence probability calculated by the co-occurrence probability calculation unit 103 is larger than the threshold value α ( Step S101). Specifically, the co-occurrence keyword set detection unit 104 takes, for example, one node of the co-occurrence network input from the co-occurrence probability calculation unit 103, and sets the co-occurrence probability larger than the threshold α for the node. Nodes connected by corresponding links are specified, and a combination of two keyword candidates indicated by these two nodes is extracted. Then, the co-occurrence keyword set detection unit 104 repeats the above processing for all nodes of the co-occurrence network, and extracts all keyword candidate combinations having a co-occurrence probability larger than the threshold value α.

  Next, the co-occurrence keyword set detection unit 104 counts the number of keyword candidate combinations extracted in step S101, and determines whether or not the number of extracted keyword candidate combinations is equal to or greater than a threshold value β (step S102). If the number of combinations of extracted keyword candidates is less than the threshold β as a result of this determination (step S102: No), the co-occurrence keyword set detection unit 104 decreases the threshold α by a predetermined amount (step S103), and step S101. Return to and repeat the above process. On the other hand, if the number of extracted keyword candidate combinations is greater than or equal to the threshold β (step S102: Yes), the extracted keyword candidate combinations are set as co-occurrence keyword sets to the co-occurrence dictionary generation unit 105 and the Patricia tree generation unit 106. In response, the data is output (step S104).

  The co-occurrence dictionary generation unit 105 generates the co-occurrence dictionary 302 based on the co-occurrence keyword set input from the co-occurrence keyword set detection unit 104. As shown in FIG. 9, the co-occurrence dictionary 302 is a set of dictionary elements having a data structure in which three elements of (headword, co-occurrence word, co-occurrence probability) are set, and a headword is designated. When searching, a dictionary element having the headword can be searched to obtain a co-occurrence word and a co-occurrence probability. The co-occurrence dictionary generation unit 105 sets one keyword candidate of each co-occurrence keyword set as a headword and the other keyword candidate as a co-occurrence word for all the co-occurrence keyword sets input from the co-occurrence keyword set detection unit 104. A dictionary element describing the co-occurrence probability between these keyword candidates is generated, and a co-occurrence dictionary 302 which is a set of these dictionary elements is generated. An example of the co-occurrence dictionary 302 generated by the co-occurrence dictionary generation unit 105 is shown in FIG. Since the entire co-occurrence dictionary 302 becomes large, only a part is shown in FIG. 10 for convenience.

  Based on the co-occurrence keyword set input from the co-occurrence keyword set detection unit 104, the Patricia tree generation unit 106 complements the input character string to obtain a keyword candidate included in the co-occurrence keyword set (characters Column completion rule). An example of the Patricia tree 303 generated by the Patricia tree generation unit 106 is shown in FIG. In addition, since the whole Patricia tree 303 becomes large, in FIG. 11, only a part is shown for convenience.

  As shown in FIG. 11, the Patricia tree 303 is represented as a tree structure in which each node has a partial character string. The nodes are sequentially traced from the root to the terminal side, and the partial character strings of the nodes are sequentially combined, and the character string obtained when the leaf that is the terminal node is reached is obtained as the output character string. The Patricia tree generation unit 106 divides the keyword candidates included in the co-occurrence keyword set input from the co-occurrence keyword set detection unit 104 into partial character strings, and merges the common partial character strings to thereby obtain the co-occurrence keyword set. A Patricia tree 303 is generated in which the keyword candidates included in the are output character strings. In the range of the Patricia tree 303 shown in FIG. 11, “orion”, “orinpic”, “orinpos”, “toyo”, “tokyo”, “haruko”, and “hanako” are output characters, respectively. It becomes a column.

  The input receiving unit 107 receives an operation input from the user. For example, when the user performs an operation of inputting a character string using an input device such as a keyboard, the input receiving unit 107 receives the input of the character string by the user and inputs the input character string (input character string). Output to the keyword recognition unit 108. Here, the input character string includes a partial character string that does not constitute one word. Here, even an input of only one character is treated as an input character string. The input receiving unit 107 outputs a new input character string of a change word to the input keyword recognizing unit 108 every time the input character string changes. In addition, when the user performs a predetermined operation for confirming the input character string, such as pressing a line feed key, the input receiving unit 107 outputs the confirmed input character string as a search keyword to the search unit 112. To do. Further, when an operation for selecting any one of the suggested word strings displayed on the display unit 111 described later is performed by the user, the input receiving unit 107 receives the operation and selects the proposed keyword Each keyword candidate included in the column is output to the search unit 112 as a search keyword.

  The input keyword recognition unit 108 recognizes a keyword candidate obtained by complementing the input character string received by the input reception unit 107 as an input keyword according to the Patricia tree 303 generated by the Patricia tree generation unit 106, and recognizes the input. The keyword is output to the co-occurrence propagation unit 109. That is, the input keyword recognition unit 108 receives the input character string that the user has input so far from the input reception unit 107 and searches the Patricia tree 303 to estimate a character string that may be input subsequently. Thus, a character string matching the keyword candidate included in the co-occurrence keyword set with the input character string as the head is obtained. At this time, the obtained character strings are all of the keyword candidates included in the co-occurrence keyword set that match the input character string in the forward direction. If all the character strings obtained here are input keywords, there is a concern that the processing load on the subsequent stage will increase significantly. Therefore, the input keyword recognition unit 108 will recognize a character string recognized as an input keyword among the obtained character strings. It is desirable to narrow down. As a narrowing-down method, for example, as described above, when information as shown in FIG. 4 showing the relationship between the keyword candidate extracted by the keyword candidate extraction unit 101 and the appearance frequency is held, this information is referred to. A method of recognizing the upper M0 items (M0 is a predetermined natural number) as the input keyword in descending order of appearance frequency can be used.

  In addition, the input keyword recognition unit 108 performs the next node corresponding to the newly input character with respect to the newly input character from the user, with the node of the Patricia tree 303 obtained by the previous input as a base point. To obtain a keyword candidate that matches forward with the input character string obtained by adding the newly input character. Then, the top M0 items are recognized as input keywords and output to the co-occurrence propagation unit 109. The input keyword recognition unit 108 repeats the above process every time a new character is input by the user.

  The co-occurrence propagation unit 109 refers to the co-occurrence dictionary 302 generated by the co-occurrence dictionary generation unit 105, acquires a co-occurrence word of a dictionary element having the input keyword input from the input keyword recognition unit 108 as a headword, Furthermore, the process of acquiring a co-occurrence word of a dictionary element having the acquired co-occurrence word as a headword is repeated. Specifically, the co-occurrence propagation unit 109 searches the co-occurrence dictionary 302 for each of the M0 input keywords input from the input keyword recognition unit 108 to obtain a co-occurrence word for the input keyword. Among the words, the top M1 co-occurrence words (M1 is a predetermined natural number) in the descending order of the co-occurrence probability with the input keyword are temporarily stored as one-stage co-occurrence words. The co-occurrence propagation unit 109 searches the co-occurrence dictionary 302 for each of the M1 first-stage co-occurrence words to obtain a co-occurrence word for the first-stage co-occurrence word. In the descending order of the co-occurrence probability between the first-stage co-occurrence words, the top M2 co-occurrence words (M2 is a predetermined natural number) are temporarily stored as the second-stage co-occurrence words. At this time, the co-occurrence propagation unit 109 adds the co-occurrence probabilities between the first-stage co-occurrence word and the second-stage co-occurrence word to the co-occurrence probabilities between the input keyword and the first-stage co-occurrence word. The integrated co-occurrence probability that is a value is obtained and temporarily stored. Further, if the calculated co-occurrence probability is greater than the threshold value γ, the co-occurrence propagation unit 109 may change from the second-stage co-occurrence word to the third-stage co-occurrence word, the third-stage co-occurrence word, and the like. The process of sequentially obtaining the keyword candidates that the co-occurrence relationship propagates from the input keyword is repeated, and the integrated co-occurrence probability that is a value obtained by sequentially integrating the co-occurrence probabilities between the keyword candidates having the co-occurrence relationship is obtained. Then, the co-occurrence propagation unit 109 stops the process when the obtained integrated co-occurrence probability is equal to or less than the threshold value γ.

  That is, the co-occurrence propagation unit 109 sets a dictionary element co-occurrence word having an input keyword as an entry word as a 1-stage co-occurrence word and L-1 stage co-occurrence word (L is a natural number of 2 or more) as an entry word. When the dictionary element co-occurrence word is an L-stage co-occurrence word, the process of obtaining the L-stage co-occurrence word is repeated while increasing L by 1, and the co-occurrence between the input keyword and the 1-stage co-occurrence word When the cumulative co-occurrence probability, which is a value obtained by sequentially integrating the co-occurrence probabilities between the L-1 stage co-occurrence word and the L-stage co-occurrence word with respect to the probability, becomes equal to or less than the threshold value γ, the processing is stopped. Then, the co-occurrence propagation unit 109 connects the input keyword to the L-stage co-occurrence word obtained at the stage where the processing is stopped, or the word string from the input keyword to the L-1 stage co-occurrence word. The sequence is output to the proposed word sequence recognition unit 110 together with the cumulative co-occurrence probability of each word sequence. In order to avoid repeated propagation between keywords with high co-occurrence probabilities, word strings with the same keyword candidates overlapping are excluded from processing targets. The co-occurrence propagation unit 109 performs the above process every time a new input keyword is input from the input keyword recognition unit 108.

  FIG. 12 is a diagram illustrating an example of a data structure of a word string output from the co-occurrence propagation unit 109 to the proposed word string recognition unit 110. From the co-occurrence propagation unit 109 to the proposed word string recognition unit 110, for example, as shown in FIG. 12, a word string connecting input keywords, one-stage co-occurrence words,. Data combining the cumulative co-occurrence probabilities of the word string is output. When a word string is formed by connecting the input keyword to the L-1 stage co-occurrence word, a word string connecting the input keyword,..., L-1 stage co-occurrence word and the integration of the word string are shared. Data combining the occurrence probabilities is output from the co-occurrence propagation unit 109 to the proposed word string recognition unit 110.

  In the above example, it has been described that the co-occurrence propagation unit 109 stops the process of acquiring the L-stage co-occurrence word when the cumulative co-occurrence probability is equal to or less than the threshold γ, but the co-occurrence propagation unit 109 stops the process. The condition to be performed is not limited to this example. For example, when the number of obtained word strings reaches the upper limit, the process of the co-occurrence propagation unit 109 is stopped or the co-occurrence word connected to the one-stage co-occurrence word A condition may be used in which the processing of the co-occurrence propagation unit 109 is stopped when the number reaches the upper limit. Further, in the above example, when the L-stage co-occurrence word is acquired from the first-stage co-occurrence word, the upper M1 cases and the upper ML cases are narrowed down in descending order of the co-occurrence probability of the obtained co-occurrence words. However, the co-occurrence probability that is a condition for acquiring the co-occurrence word is determined in advance, and the ones having the predetermined co-occurrence probability or higher are sequentially acquired from the first-stage co-occurrence word to the L-stage co-occurrence word. May be.

  The proposed word string recognition unit 110 recognizes a word string satisfying the second condition among the word strings input from the co-occurrence propagation unit 109 as a proposed word string, and outputs the recognized proposed word string to the display unit 111. . Here, as the second condition, for example, a condition of top N cases (N is a predetermined natural number) in descending order of the cumulative co-occurrence probability can be used. The second condition is not limited to this example, and various predetermined conditions can be used.

  FIG. 13 is a flowchart illustrating an example of processing by the input keyword recognition unit 108, the co-occurrence propagation unit 109, and the suggested word string recognition unit 110.

  When the input keyword recognition unit 108 inputs the input character string from the input reception unit 107, the input keyword recognition unit 108 acquires keyword candidates obtained by complementing the input character string based on the Patricia tree 303, and among the acquired keyword candidates, the appearance frequency The top M0 keyword candidates are recognized as input keywords in descending order (step S201). Then, the input keyword recognition unit 108 outputs the recognized input keyword to the co-occurrence propagation unit 109.

  Next, when the co-occurrence propagation unit 109 inputs M0 input keywords from the input keyword recognition unit 108, the co-occurrence propagation unit 109 refers to the co-occurrence dictionary 302 and obtains co-occurrence words of the M0 input keywords, respectively. Among the words, the top M1 keyword candidates are acquired as one-stage co-occurrence words in descending order of the co-occurrence probability with the input keyword (step S202). Further, the co-occurrence propagation unit 109 refers to the co-occurrence dictionary 302 to obtain co-occurrence words of L-1 level co-occurrence words (L is a natural number of 2 or more), and among these co-occurrence words, L- The top ML keyword candidates are acquired as the L-stage co-occurrence word in descending order of the co-occurrence probability with the first-stage co-occurrence word (step S203). The co-occurrence propagation unit 109 sequentially calculates the co-occurrence probabilities between the L-1 stage co-occurrence word and the L-stage co-occurrence word with respect to the co-occurrence probability between the input keyword and the 1-stage co-occurrence word. It is determined whether or not the integrated co-occurrence probability that is the integrated value is equal to or less than the threshold value γ (step S204). If the cumulative co-occurrence probability is greater than the threshold γ as a result of this determination (step S204: No), the co-occurrence propagation unit 109 increments (+1) the value of L (step S205) and returns to step S203. Repeat the process of acquiring the co-occurrence words. On the other hand, if the cumulative co-occurrence probability is less than or equal to the threshold γ (step S204: Yes), the co-occurrence propagation unit 109 adds the word string connecting the input keyword and the acquired co-occurrence word to the cumulative co-occurrence of each word string. It outputs to the proposal word sequence recognition part 110 with a probability.

  Next, when a word string is input from the co-occurrence propagation unit 109, the proposed word string recognition unit 110 recognizes the top N word strings as the suggested word string in descending order of the cumulative co-occurrence probability. (Step S206). Then, the suggested word string recognition unit 110 outputs the recognized suggested word string to the display unit 111 (step S207).

  The display unit 111 displays the suggested word string input from the suggested word string recognition unit 110 and presents it to the user. The suggested word string displayed by the display unit 111 can be selected by the user using an input device such as a mouse or a keyboard. When the user selects one of the suggested word strings displayed on the display unit 111, the operation is accepted by the input accepting unit 107, and each keyword candidate included in the selected proposed word string is used as a search keyword. Is output. The display unit 111 displays the search result when the search result is input from the search unit 112 described later.

  The search unit 112 receives each keyword candidate included in the selected suggested word string when any of the suggested word strings displayed on the display unit 111 is selected by the user, and searches all of these keyword candidates. A search expression including keywords is generated, and the document set 300 is searched. Specifically, the search unit 112 refers to the index 301 generated by the index generation unit 102 and sequentially searches for document pointer information including the search keyword included in the search expression, and all the search keywords included in the search expression. Document pointer information for a document containing Next, the search unit 112 obtains a document indicated by the obtained document pointer information from the document set 300 to be searched, and outputs necessary information to the display unit 111 as a search result.

  The search support apparatus 100 according to the present embodiment includes, for example, a control device such as a CPU, an internal storage device such as a ROM and a RAM, an input device such as a keyboard and a mouse, a display device such as a liquid crystal panel, a hard disk, and a CD. A hardware configuration using an ordinary computer having an external storage device such as a DVD or a flash memory can be employed. In this case, the functional configuration of the search support apparatus 100 according to the present embodiment can be realized by the computer executing the program. The index 301, the co-occurrence dictionary 302, and the Patricia tree 303 can be stored in an external storage device such as a hard disk or a flash memory, for example.

  Next, a specific example of processing performed by the search support apparatus 100 according to this embodiment will be described by taking as an example a case where a search is performed using the document set 300 illustrated in FIG. 2 as a search target. In the following, it is assumed that the co-occurrence dictionary 302 illustrated in FIG. 10 and the Patricia tree 303 illustrated in FIG. 11 are generated from the document set 300 illustrated in FIG.

First, when a user inputs a character string “O” using an input device such as a keyboard, the input receiving unit 107 receives the input of the character string, and inputs the input character string “O” to the input keyword recognizing unit 108. Output. The input keyword recognizing unit 108 receives the input character string “o” from the input receiving unit 107, searches the Patricia tree 303 illustrated in FIG. 11, and performs processing for recognizing the input keyword. First, in order to obtain a node corresponding to the input character string “o”, the input keyword recognition unit 108 sequentially searches the Patricia tree 303 illustrated in FIG. 11 from the root, and sets “ori” as a node corresponding to “o”. Get the node. Next, the input keyword recognition unit 108 identifies a leaf group traced from the “ori” node, obtains a character string indicated by the leaf as a character string subsequent to “ori”, and complements the input character string “o”. Search for keyword candidates. In this case, “Olympic”, “Orion” and “Olympos” are obtained. The input keyword recognition unit 108 recognizes the top M0 items as the input keyword in descending order of appearance frequency among the obtained keyword candidates. Here, it is assumed that M0 = 3. In this case, the input keyword recognition unit 108 recognizes “Olympic”, “Orion”, and “Olympos” as input keywords, and outputs these input keywords to the co-occurrence propagation unit 109.

The co-occurrence propagation unit 109 receives the input keywords “Olympic”, “Orion”, and “Olympos” from the input keyword recognition unit 108, and refers to the co-occurrence dictionary 302 illustrated in FIG. Find the dictionary element to do. Here, the following dictionary elements are obtained.
(Orion, Hanako, 0.75)
(Orion, Komukai, 0.5)
(Orion, Isogo, 0.5)
(Orion, Jiro, 0.5)
(Orion, World Cup, 0.5)
(Orion, Publishing, 0.4)
(Orion, Kamakura, 0.33)
(Orion, Sugita, 0.33)
(Orion, Toyo, 0.25)
(Orion, Fuchu, 0.25)
(Olympic, Oriental, 0.67)
(Olympic, Horikawa, 0.5)
(Olympic Games, Shibaura, 0.5)
(Olympic, Kawasaki, 0.5)
(Olympic, Haruko, 0.5)
(Olympic, Jiro, 0.5)
(Olympic, World, 0.5)
(Olympic, land, 0.5)
(Olympic, Taro, 0.5)
(Olympic, publication, 0.4)
(Olympic, Hanako, 0.4)
(Olympic, Fuchu, 0.25)
(Olympos, Kamakura, 0.5)
(Olympos, Tokyo, 0.5)
(Olympos, Fuchu, 0.33)
(Olympus, Publishing, 0.2)
(Olympus, Hanako, 0.2)

  Next, the co-occurrence propagation unit 109 sets the M1 co-occurrence words in the descending order of the co-occurrence probability with the headword (input keyword) from the co-occurrence words of the obtained dictionary elements. Get as a utterance. Here, it is assumed that M1 = 3. For co-occurrence words with the same co-occurrence probability with headwords, for example, the co-occurrence words to be acquired are based on the order of appearance frequency of headwords, the order of appearance frequency of co-occurrence words, etc. Can be determined. In this case, “Hanako” having a co-occurrence probability of 0.75 with the input keyword “Orion”, “Toyo” having a co-occurrence probability of 0.67 with the input keyword “Olympic”, and the input keyword “Orion”. “Komukai” having a co-occurrence probability of 0.5 is acquired as a one-stage co-occurrence word.

Further, the co-occurrence propagation unit 109 refers to the co-occurrence dictionary 302 again in order to obtain a two-stage co-occurrence word having a co-occurrence relationship with the first-stage co-occurrence word. Find the element. Here, the following dictionary elements are obtained.
(Hanako, Orion, 0.75)
(Hanako, Fuchu, 0.6)
(Hanako, Kamakura, 0.4)
(Hanako, Sugita, 0.4)
(Hanako, Olympics, 0.4)
(Toyo, Olympics, 0.67)
(Toyo, Publishing, 0.6)
(Toyo, Komukai, 0.33)
(Toyo, Taro, 0.33)
(Toyo, Horikawa, 0.33)
(Toyo, Shibaura, 0.33)
(Komukai, Sugita, 0.5)
(Komukai, Tokyo, 0.5)
(Komukai, Orion, 0.5)

  Next, the co-occurrence propagation unit 109 selects a co-occurrence word having a cumulative co-occurrence probability larger than the threshold γ from the obtained dictionary element co-occurrence words, and connects it to the one-stage co-occurrence word. Get as a word. Here, it is assumed that the threshold γ = 0.37. In this case, for the first-stage co-occurrence word “Hanako”, the co-occurrence probability with the input keyword “Orion” is 0.75, so that the cumulative co-occurrence probability is greater than 0.37. It is necessary to acquire a two-stage co-occurrence word having a co-occurrence probability of 0.49 or more, and “Orion” and “Fuchu” satisfying this condition are acquired as two-stage co-occurrence words. However, since “Orion” overlaps with the input keyword, it is excluded from processing. Also, for the first-stage co-occurrence word “Toyo”, the co-occurrence probability with the input keyword “Olympic” is 0.67, so that the cumulative co-occurrence probability is larger than 0.37. It is necessary to acquire a two-stage co-occurrence word having an occurrence probability of 0.55 or more, and “Olympic” and “Publishing” that satisfy this condition are acquired as two-stage co-occurrence words. However, since “Olympic” overlaps with the input keyword, it is excluded from processing. In addition, for the one-stage co-occurrence word “Komukai”, the co-occurrence probability with the input keyword “Orion” is 0.5, so that the cumulative co-occurrence probability is greater than 0.37. It is necessary to acquire a two-stage co-occurrence word having a co-occurrence probability of 0.8 or more. However, since there is no co-occurrence word satisfying this condition, a two-stage co-occurrence word cannot be obtained.

  As a result of the above processing, (Orion, Hanako, Fuchu, 0.45) and (Olympic, Toyo, Publishing, 0.4) are used as word strings connecting the input keyword, the first-stage co-occurrence word, and the second-stage co-occurrence word. ) Is obtained.

Thereafter, the co-occurrence propagation unit 109 refers to the co-occurrence dictionary 302 again to obtain a three-stage co-occurrence word having a co-occurrence relationship with the second-stage co-occurrence word, and a dictionary having the second-stage co-occurrence word as an entry word. Find the element. Here, the following dictionary elements are obtained.
(Fuchu, Kamakura, 0.67)
(Fuchu, Publishing, 0.6)
(Fuchu, Hanako, 0.6)
(Publishing, Toyo, 0.6)
(Publishing, Fuchu, 0.6)
(Publishing, Orion, 0.4)

  Next, the co-occurrence propagation unit 109 connects the co-occurrence words whose cumulative co-occurrence probability is greater than the threshold γ = 0.37 to the two-stage co-occurrence words from the obtained dictionary element co-occurrence words. The process of acquiring as a three-stage co-occurrence word is performed. In this case, for the two-stage co-occurrence word “Fuchu”, since the cumulative co-occurrence probability so far is 0.45, the co-occurrence probability is 0.82 for the cumulative co-occurrence probability to be greater than 0.37. Although it is necessary to acquire the above three-stage co-occurrence words, there is no co-occurrence word that satisfies this condition, and therefore, a three-stage co-occurrence word cannot be obtained. Also, for the two-stage co-occurrence word “publishing”, since the cumulative co-occurrence probability so far is 0.4, the co-occurrence probability is 0.99 or more for the cumulative co-occurrence probability to be larger than 0.37. However, since there is no co-occurrence word that satisfies this condition, a three-stage co-occurrence word cannot be obtained. For this reason, the co-occurrence propagation unit 109 stops the process of propagating the co-occurrence words, and is a word string (Orion, Hanako, Fuchu, 0.45) and (Olympics, Toyo, Publication, 0.4) is output to the proposed word string recognition unit 110.

  The proposed word string recognition unit 110 receives the word strings (Orion, Hanako, Fuchu, 0.45) and (Olympic, Toyo, Publishing, 0.4) from the co-occurrence propagation unit 109, and among these word strings, The top N items are recognized as a proposed word string in descending order of the cumulative co-occurrence probability. Here, it is assumed that N = 2. For word strings having the same cumulative co-occurrence probability, for example, the order of appearance frequency of input keywords, order of appearance frequency of first-stage co-occurrence words, order of appearance frequency of second-stage co-occurrence words, etc. As a result, N proposal word strings can be determined. In this case, the proposed word string recognition unit 110 recognizes (Orion, Hanako, Fuchu) and (Olympic, Toyo, Publishing) as a proposed word string, and outputs these suggested word strings to the display unit 111.

  Upon receiving the suggested word strings (Orion, Hanako, Fuchu) and (Olympic, Toyo, Publishing) from the suggested word string recognition unit 110, the display unit 111 displays these suggested word strings and presents them to the user. When the user performs an operation of selecting, for example, (Orion, Hanako, Fuchu) from the suggested word string displayed on the display unit 111 using an input device such as a mouse, the selection operation is performed by the input receiving unit 107. The keyword candidates “Orion”, “Hanako”, and “Fuchu” included in the suggested word string are output to the search unit 112, respectively. In the state where the suggested word string is displayed on the display unit 111, the user continues to input the character string without selecting any of the suggested word strings. For example, when the user inputs the character string “Orin”, The input keyword recognition unit 108 recognizes “Olympic” and “Olympos” as input keywords, and the same processing is performed thereafter.

  Upon receiving “Orion”, “Hanako”, and “Fuchu” from the input receiving unit 107, the search unit 112 generates a search expression including all of these keyword candidates as search keywords, and performs a search on the document set 300. Do. As a result, [Document 2] is obtained from the document set 300 illustrated in FIG. 2, and information on [Document 2] is displayed on the display unit 111 as a search result.

  As described above in detail with specific examples, according to the search support apparatus 100 of the present embodiment, the user simply specifies the document set 300 to be searched, and the keywords included in the document set 300 are included. Among the candidates, a co-occurrence keyword set that is a combination of keyword candidates having a large co-occurrence probability is detected, a co-occurrence dictionary 302 is generated based on the co-occurrence keyword set, and the input character string is complemented to generate the co-occurrence keyword. A Patricia tree 303, which is a character string complement rule for obtaining keyword candidates constituting a set as input keywords, is generated. Then, when the user inputs a character string, a suggested word string connecting the input keyword corresponding to the input character string and the co-occurrence word in which the co-occurrence relationship propagates from the input keyword is presented to the user, and the proposed proposal When the user selects a word string, a document including all keyword candidates included in the proposed word string is searched from the search target document set 300. Therefore, according to the search support apparatus 100 of the present embodiment, the search keyword candidate suitable for the document set 300 to be searched is proposed and the document search is appropriately performed without requiring the user to perform complicated operations in advance. Can help.

  Further, according to the search support apparatus 100 of the present embodiment, the co-occurrence propagation unit 109 refers to the co-occurrence dictionary 302 and sets a co-occurrence word of a dictionary element having an input keyword as a headword as a one-stage co-occurrence word, When a co-occurrence word of a dictionary element having an L-1 stage co-occurrence word as an entry word is an L-stage co-occurrence word, the process of acquiring the L-stage co-occurrence word while repeating L is incremented by 1 Since the process is stopped when the occurrence probability is equal to or less than the threshold value γ, the integrated co-occurrence probabilities of the suggested word string presented to the user can be made to be close to the threshold value γ. This leads to effectively proposing a plurality of keyword candidates included in the document set 300 as search keywords without making unnecessary proposals to the user. That is, in many cases, the word sequence in which the cumulative co-occurrence probability is excessive is generally an enumeration of highly related words. However, in the search support apparatus 100 of this embodiment, the cumulative co-occurrence probability is close to the threshold value γ. Since the suggested word string of values is presented to the user, it is possible to make a proposal that is considered preferable for the user without proposing a word string generally related to the word.

(Second Embodiment)
Next, a second embodiment will be described. In the second embodiment, a zero co-occurrence keyword set, which is a combination of two keyword candidates whose co-occurrence probability is zero, is detected in advance, and a zero co-occurrence dictionary is generated to connect the input keyword and each co-occurrence. If there is a word string that includes two keyword candidates that simultaneously constitute a zero co-occurrence keyword set in the word string, the word string is excluded from the target of the proposed word string. That is, when sequentially acquiring keyword candidates that follow the input keyword by propagating the co-occurrence relationship from the input keyword, the input keyword, the first-stage co-occurrence word, the first-stage co-occurrence word, and the second-stage co-occurrence word A co-occurrence relationship is guaranteed between consecutive keyword candidates, such as between, but for example, there is always a co-occurrence relationship between distant keyword candidates, such as between an input keyword and a two-stage co-occurrence word. Is not limited. Therefore, for example, there is no co-occurrence relationship between the input keyword and the two-stage co-occurrence word (that is, a document in which the input keyword and the two-stage co-occurrence word appear simultaneously does not exist in the search target document set). In such a case, when such a word string is presented to the user as a suggested word string, and the user selects the proposed word string and performs a search on the document set, no search result is obtained (that is, the number of hits becomes zero). ) Problems arise. Therefore, in the present embodiment, a search result is always obtained by searching a document set based on a proposed word string by excluding a word string that simultaneously includes two keyword candidates whose co-occurrence probability is zero from the target of the proposed word string. Is to be obtained. Hereinafter, only the characteristic part of the present embodiment will be described, and the same components as those in the first embodiment will be denoted by the same reference numerals, and redundant description will be omitted.

  FIG. 14 is a block diagram illustrating a functional configuration of the search support apparatus 200 according to the second embodiment. As shown in FIG. 14, the search support apparatus 200 of the present embodiment includes a zero co-occurrence keyword set detection unit (second detection unit) 201, a zero co-occurrence dictionary, in addition to the configuration of the search support apparatus 100 of the first embodiment. A generation unit (third generation unit) 202 is provided. Further, the search support apparatus 200 of the present embodiment replaces the co-occurrence propagation unit 109 and the proposed word string recognition unit 110 included in the search support apparatus 100 of the first embodiment with a co-occurrence propagation unit 209 and a proposed word string recognition unit. 210.

  The zero co-occurrence keyword set detection unit 201 inputs a co-occurrence network as illustrated in FIG. 7 from the co-occurrence probability calculation unit 103, and based on this co-occurrence network, two keyword candidates whose co-occurrence probability is zero Are detected and output to the zero-co-occurrence dictionary generation unit 202. Specifically, the zero co-occurrence keyword set detection unit 201 selects, for example, keyword candidates whose appearance frequency exceeds a predetermined value from the keyword candidates extracted by the keyword candidate extraction unit 101, and selects each of these selected keyword candidates. For a keyword candidate with a co-occurrence probability of zero, a zero co-occurrence keyword set is obtained. Here, the keyword candidates whose appearance frequency exceeds a predetermined value are targeted because it is assumed that keyword candidates whose appearance frequency is extremely low are not keyword candidates that constitute a co-occurrence keyword set in the first place.

  FIG. 15 is a diagram illustrating a combination of two keyword candidates having a co-occurrence probability of 0. Of the co-occurrence probabilities between the two keyword candidates illustrated in FIG. 6, those having a co-occurrence probability of zero are extracted. It is a thing. In the example of FIG. 15, each keyword candidate having an appearance frequency of 2 or more represents a combination with another keyword candidate having a co-occurrence probability of zero.

  The zero co-occurrence dictionary generation unit 202 generates a zero co-occurrence dictionary 304 based on the zero co-occurrence keyword set input from the zero co-occurrence keyword set detection unit 201. As shown in FIG. 16, the zero co-occurrence dictionary 304 is a set of dictionary elements having a data structure in which two elements of (headword, zero co-occurrence word) are set, and when a headword is designated. In addition, a zero co-occurrence word can be obtained by searching a dictionary element having the entry word. The zero co-occurrence dictionary generation unit 202, for all zero co-occurrence keyword sets input from the zero co-occurrence keyword set detection unit 201, finds one keyword candidate narrowed down by appearance frequency as a headword and the other keyword candidate as zero co-occurrence. A dictionary element is generated as a word, and a zero co-occurrence dictionary 304 which is a set of these dictionary elements is generated. An example of the zero co-occurrence dictionary 304 generated by the zero co-occurrence dictionary generation unit 202 is shown in FIG. Since the entire zero co-occurrence dictionary 304 is large, only a part is shown in FIG. 17 for convenience.

  The co-occurrence propagation unit 209 searches the co-occurrence dictionary 302 for each of the M0 input keywords input from the input keyword recognition unit 108 to obtain a co-occurrence word for the input keyword. The top M1 co-occurrence words (M1 is a predetermined natural number) in the descending order of the co-occurrence probability with the input keyword are temporarily stored as one-stage co-occurrence words. The co-occurrence propagation unit 109 searches the co-occurrence dictionary 302 for each of the M1 first-stage co-occurrence words to obtain a co-occurrence word for the first-stage co-occurrence word. A co-occurrence word having a co-occurrence probability between the first-stage co-occurrence word and the threshold value δ or more is temporarily stored as a second-stage co-occurrence word. Further, the co-occurrence propagation unit 209 starts from the second-stage co-occurrence word to the third-stage co-occurrence word until the number of co-occurrence words connected to the first-stage co-occurrence word reaches a predetermined value A. The process of sequentially obtaining co-occurrence words connected to the first-stage co-occurrence word, such as four-stage co-occurrence words, is repeated, and when the number of co-occurrence words connected to the first-stage co-occurrence word reaches a predetermined value A, the process is stopped.

  That is, the co-occurrence propagation unit 209 sets the co-occurrence word of the dictionary element having the input keyword as an entry word as a 1-stage co-occurrence word, and uses the L-1 stage co-occurrence word (L is a natural number of 2 or more) as an entry word. When the dictionary element co-occurrence word is an L-stage co-occurrence word, the process of acquiring the L-stage co-occurrence word is repeated while increasing L by 1. When L reaches a predetermined value A, the L-stage co-occurrence word is obtained from the input keyword. The integrated co-occurrence probability up to the word is obtained and the process is stopped. Then, the co-occurrence propagation unit 209 converts the word string obtained by connecting the input keyword and the L-stage co-occurrence word obtained at the stage where the processing is stopped together with the proposed word string recognition unit 210 together with the cumulative co-occurrence probability of each word string. Output for. As in the first embodiment, in order to avoid repeated propagation between keywords having a high co-occurrence probability, word strings with the same keyword candidate overlapping are excluded from processing targets. The co-occurrence propagation unit 209 performs the above process every time a new input keyword is input from the input keyword recognition unit 108.

  In the above example, it has been described that the process of the co-occurrence propagation unit 209 is stopped when the number of co-occurrence words connected to the first-stage co-occurrence word reaches a predetermined value A, but the co-occurrence propagation unit 209 performs the process. The condition for stopping is not limited to this example. For example, as in the first embodiment, when the integrated co-occurrence probability is equal to or less than the threshold γ, the co-occurrence propagation unit 209 stops or is obtained. A condition that the processing of the co-occurrence propagation unit 209 is stopped when the number of word strings reaches the upper limit may be used.

  Similar to the proposed word string recognition unit 110 of the first embodiment, the proposed word string recognition unit 210 recognizes a word string satisfying the second condition among the word strings input from the co-occurrence propagation unit 209 as a proposed word string. The recognized suggested word string is output to the display unit 111. However, the proposed word string recognition unit 210 first refers to the zero co-occurrence dictionary 304, and includes a word string that simultaneously includes two keyword candidates constituting a zero co-occurrence keyword set in the word string input from the co-occurrence propagation unit 210. If such a word string exists, the word string is excluded and the word string satisfying the second condition is displayed as the suggested word string among the remaining word strings. Output to the unit 111. As the second condition, as in the first embodiment, for example, the condition of the top N cases (N is a predetermined natural number) in descending order of the cumulative co-occurrence probability can be used.

  FIG. 18 is a flowchart illustrating an example of processing performed by the input keyword recognition unit 108, the co-occurrence propagation unit 209, and the suggested word string recognition unit 210 of the search support apparatus 200 according to the present embodiment.

  When the input keyword recognition unit 108 inputs the input character string from the input reception unit 107, the input keyword recognition unit 108 acquires keyword candidates obtained by complementing the input character string based on the Patricia tree 303, and among the acquired keyword candidates, the appearance frequency The top M0 keyword candidates are recognized as input keywords in descending order (step S301). Then, the input keyword recognition unit 108 outputs the recognized input keyword to the co-occurrence propagation unit 209.

  Next, when the co-occurrence propagation unit 209 inputs M0 input keywords from the input keyword recognition unit 108, the co-occurrence propagation unit 209 refers to the co-occurrence dictionary 302 and obtains co-occurrence words of the M0 input keywords, respectively. Among the words, the top M1 keyword candidates are acquired as one-stage co-occurrence words in descending order of co-occurrence probability (step S302). Further, the co-occurrence propagation unit 209 refers to the co-occurrence dictionary 302 to obtain co-occurrence words of L-1 level co-occurrence words (L is a natural number of 2 or more), and among these co-occurrence words, L- Keyword candidates whose co-occurrence probability with the first-stage co-occurrence word is equal to or greater than the threshold δ are acquired as L-stage co-occurrence words (step S303). Then, the co-occurrence propagation unit 209 determines whether or not the value of L has reached a predetermined value A (step S304). As a result of this determination, if the value of L is less than the predetermined value A (step S304: No), the co-occurrence propagation unit 209 increments (+1) the value of L (step S305), and returns to step S303. Repeat the process of acquiring the co-occurrence words. On the other hand, if the value of L has reached the predetermined value A (step S304: Yes), the co-occurrence propagation unit 209 uses the word string connecting the input keyword and the acquired co-occurrence word as the integrated value of each word string. Along with the occurrence probability, it is output to the proposed word string recognition unit 210.

  Next, when a word string is input from the co-occurrence propagation unit 209, the proposed word string recognition unit 210 refers to the zero co-occurrence dictionary 304 and forms two co-occurrence keyword sets in the input word string. It is determined whether there is a word string that includes keyword candidates at the same time (step S306). Then, when there is a word string that simultaneously includes two keyword candidates that constitute the zero co-occurrence keyword set (step S306: Yes), the proposed word string recognition unit 210 deletes the word string (step S307). On the other hand, if there is no word string that simultaneously includes two keyword candidates that constitute the zero co-occurrence keyword set in the input word string (step S306: No), the process of step S307 is skipped. Next, the proposed word string recognition unit 210 recognizes the top N word strings as the proposed word string in descending order of the cumulative co-occurrence probability among the remaining word strings excluding the word string deleted in step S307 (step S308). ). Then, the suggested word string recognition unit 210 outputs the recognized suggested word string to the display unit 111 (step S309).

  Similar to the search support apparatus 100 of the first embodiment, the search support apparatus 200 of the present embodiment has a hardware configuration such as a control device such as a CPU, an internal storage device such as a ROM or a RAM, a keyboard or a mouse, and the like. A hardware configuration using a normal computer that includes an input device, a display device such as a liquid crystal panel, and an external storage device such as a hard disk, CD, DVD, or flash memory can be employed. In this case, the functional configuration of the search support apparatus 200 according to the present embodiment can be realized by the computer executing the program. The index 301, the co-occurrence dictionary 302, the Patricia tree 303, and the zero co-occurrence dictionary 304 can be stored in an external storage device such as a hard disk or a flash memory, for example.

  Next, a specific example of processing performed by the search support apparatus 200 according to the present embodiment will be described with reference to an example in which a search is performed using the document set 300 illustrated in FIG. 2 as a search target. In the following, it is assumed that the co-occurrence dictionary 302 illustrated in FIG. 10, the Patricia tree 303 illustrated in FIG. 11, and the zero co-occurrence dictionary 304 illustrated in FIG. 17 are generated from the document set 300 illustrated in FIG. .

  First, when a user inputs a character string “O” using an input device such as a keyboard, the input receiving unit 107 receives the input of the character string, and inputs the input character string “O” to the input keyword recognizing unit 108. Output. The input keyword recognizing unit 108 receives the input character string “o” from the input receiving unit 107, searches the Patricia tree 303 illustrated in FIG. 11, and performs processing for recognizing the input keyword. First, in order to obtain a node corresponding to the input character string “o”, the input keyword recognition unit 108 sequentially searches the Patricia tree 303 illustrated in FIG. 11 from the root, and sets “ori” as a node corresponding to “o”. Get the node. Next, the input keyword recognition unit 108 identifies a leaf group traced from the “ori” node, obtains a character string indicated by the leaf as a character string subsequent to “ori”, and complements the input character string “o”. Search for keyword candidates. In this case, “Olympic”, “Orion” and “Olympos” are obtained. The input keyword recognition unit 108 recognizes the top M0 items as the input keyword in descending order of appearance frequency among the obtained keyword candidates. Here, it is assumed that M0 = 3. In this case, the input keyword recognition unit 108 recognizes “Olympic”, “Orion”, and “Olympos” as input keywords, and outputs these input keywords to the co-occurrence propagation unit 209.

The co-occurrence propagation unit 209 receives the input keywords “Olympic”, “Orion”, and “Olympos” from the input keyword recognition unit 108, and refers to the co-occurrence dictionary 302 illustrated in FIG. Find the dictionary element to do. Here, the following dictionary elements are obtained.
(Orion, Hanako, 0.75)
(Orion, Komukai, 0.5)
(Orion, Isogo, 0.5)
(Orion, Jiro, 0.5)
(Orion, World Cup, 0.5)
(Orion, Publishing, 0.4)
(Orion, Kamakura, 0.33)
(Orion, Sugita, 0.33)
(Orion, Toyo, 0.25)
(Orion, Fuchu, 0.25)
(Olympic, Oriental, 0.67)
(Olympic, Horikawa, 0.5)
(Olympic Games, Shibaura, 0.5)
(Olympic, Kawasaki, 0.5)
(Olympic, Haruko, 0.5)
(Olympic, Jiro, 0.5)
(Olympic, World, 0.5)
(Olympic, land, 0.5)
(Olympic, Taro, 0.5)
(Olympic, publication, 0.4)
(Olympic, Hanako, 0.4)
(Olympic, Fuchu, 0.25)
(Olympos, Kamakura, 0.5)
(Olympos, Tokyo, 0.5)
(Olympos, Fuchu, 0.33)
(Olympus, Publishing, 0.2)
(Olympus, Hanako, 0.2)

  Next, the co-occurrence propagating unit 209 sets the upper M1 co-occurrence words from the co-occurrence words of the obtained dictionary elements in the descending order of the co-occurrence probability with the headword (input keyword). Get as a utterance. Here, it is assumed that M1 = 3. For co-occurrence words with the same co-occurrence probability with headwords, for example, the co-occurrence words to be acquired are based on the order of appearance frequency of headwords, the order of appearance frequency of co-occurrence words, etc. Can be determined. In this case, “Hanako” having a co-occurrence probability of 0.75 with the input keyword “Orion”, “Toyo” having a co-occurrence probability of 0.67 with the input keyword “Olympic”, and the input keyword “Orion”. “Komukai” having a co-occurrence probability of 0.5 is acquired as a one-stage co-occurrence word.

Further, the co-occurrence propagation unit 209 refers to the co-occurrence dictionary 302 again in order to obtain a two-stage co-occurrence word having a co-occurrence relationship with the first-stage co-occurrence word. Find the element. Here, the following dictionary elements are obtained.
(Hanako, Orion, 0.75)
(Hanako, Fuchu, 0.6)
(Hanako, Kamakura, 0.4)
(Hanako, Sugita, 0.4)
(Hanako, Olympics, 0.4)
(Toyo, Olympics, 0.67)
(Toyo, Publishing, 0.6)
(Toyo, Taro, 0.33)
(Toyo, Horikawa, 0.33)
(Toyo, Shibaura, 0.33)
(Komukai, Sugita, 0.5)
(Komukai, Tokyo, 0.5)
(Komukai, Orion, 0.5)

  Next, the co-occurrence propagation unit 209 selects two keyword candidates whose co-occurrence probability with the headword (one-stage co-occurrence word) is equal to or greater than the threshold δ from the co-occurrence words of the obtained dictionary elements. Acquired as a co-occurrence word. Here, it is assumed that the threshold δ = 0.4. In this case, “Orion”, “Fuchu”, “Kamakura”, and “Sugita” with a co-occurrence probability of 0.4 or more are used as the second-stage co-occurrence words “Hanako”. To be acquired. However, since “Orion” overlaps with the input keyword, it is excluded from processing. For the first-stage co-occurrence word “Toyo”, “Olympic” and “Publishing” having a co-occurrence probability of 0.4 or more are acquired as two-stage co-occurrence words, respectively. However, since “Olympic” overlaps with the input keyword, it is excluded from processing. For the first-stage co-occurrence word “Komukai”, “Sugita”, “Tokyo”, and “Orion” having co-occurrence probabilities of 0.4 or more are acquired as second-stage co-occurrence words. However, since “Orion” overlaps with the input keyword, it is excluded from processing.

  As a result of the above processing, as a word string connecting the input keyword, the first-stage co-occurrence word, and the second-stage co-occurrence word, (Orion, Hanako, Fuchu), (Orion, Hanako, Kamakura), (Orion, Hanako, Sugita) ), (Olympic, Toyo, Publishing), (Orion, Komukai, Sugita), (Orion, Komukai, Tokyo).

  Next, the co-occurrence propagation unit 209 checks whether or not the number of co-occurrence words connected to the first-stage co-occurrence word has reached a predetermined value A. Here, it is assumed that the predetermined value A = 1 is set for simplicity. In this case, the co-occurrence propagation unit 209 stops the process when the two-stage co-occurrence word is acquired, obtains an integrated co-occurrence probability for each of the word strings obtained in the process so far, and (Orion, Hanako) , Fuchu, 0.45), (Orion, Hanako, Kamakura, 0.45), (Orion, Hanako, Sugita, 0.3), (Olympic, Toyo, Publishing, 0.4), (Orion, Komukai, Sugita, 0.25), (Orion, Komukai, Tokyo, 0.25) is output to the proposed word string recognition unit 210.

  The proposed word string recognition unit 210 is a word string from the co-occurrence propagation unit 209 (Orion, Hanako, Fuchu, 0.45), (Orion, Hanako, Kamakura, 0.45), (Orion, Hanako, Sugita, 0 .3), (Olympic, Toyo, Publishing, 0.4), (Orion, Komukai, Sugita, 0.25), (Orion, Komukai, Tokyo, 0.25) Referring to the dictionary 304, if a word string that includes two keyword candidates that simultaneously constitute a zero co-occurrence keyword set exists in these word strings, the word string is deleted. The zero co-occurrence dictionary 304 illustrated in FIG. 17 includes dictionary elements having “Orion” as a headword and “Tokyo” as a zero co-occurrence word. Among the above word strings (Orion, Komukai, Tokyo, (0.25) includes “Orion” and “Tokyo” at the same time, so (Orion, Komukai, Tokyo, 0.25) is deleted.

  Next, the proposed word string recognition unit 210 performs the remaining word strings (Orion, Hanako, Fuchu, 0.45), (Orion, Hanako, Kamakura, 0.45), (Orion, Hanako, Sugita, 0.3), (Olympic, Toyo, Publishing, 0.4), (Orion, Komukai, Sugita, 0.25), the top N items are recognized as the suggested word sequence in descending order of the cumulative co-occurrence probability. Here, it is assumed that N = 4. For word strings having the same cumulative co-occurrence probability, for example, the order of appearance frequency of input keywords, order of appearance frequency of first-stage co-occurrence words, order of appearance frequency of second-stage co-occurrence words, etc. As a result, N proposal word strings can be determined. In this case, the proposed word string recognition unit 210 recognizes (Orion, Hanako, Fuchu), (Orion, Hanako, Kamakura), (Orion, Hanako, Sugita), and (Olympic, Toyo, Publishing) as the proposed word strings. These suggested word strings are output to the display unit 111.

  Each of the proposed word strings output to the display unit 111 by the proposed word string recognition unit 210 is a word string from which a search result can be reliably obtained (that is, the number of hits does not become zero). For example, if (Orion, Hanako, Fuchu) is selected, [Document 2] is obtained in the document set 300 illustrated in FIG. 2, and if (Orion, Hanako, Kamakura) is selected, it is illustrated in FIG. If [Document 2] is obtained in the document set 300 and (Orion, Hanako, Sugita) is selected, [Document 3] is obtained in the document set 300 illustrated in FIG. If “Toyo Publishing” is selected, [Document 1] is obtained in the document set 300 illustrated in FIG.

  When the display unit 111 receives the proposed word sequence (Orion, Hanako, Fuchu), (Orion, Hanako, Kamakura), (Orion, Hanako, Sugita), (Olympic, Toyo, Publishing) from the proposed word sequence recognition unit 210, These suggested word strings are displayed and presented to the user. When the user performs an operation of selecting, for example, (Olympic, Toyo, Publishing) from the suggested word string displayed on the display unit 111 using an input device such as a mouse, the selection operation is performed by the input receiving unit 107. The keyword candidates “Olympic”, “Toyo”, and “Publishing” included in the suggested word string are output to the search unit 112, respectively. In the state where the suggested word string is displayed on the display unit 111, the user continues to input the character string without selecting any of the suggested word strings. For example, when the character string “Orio” is input, The input keyword recognizing unit 108 recognizes “Orion” as an input keyword, and the same processing is performed thereafter.

  When the search unit 112 receives “Olympic”, “Toyo”, and “Fuchu” from the input receiving unit 107, the search unit 112 generates a search expression including all of these keyword candidates as search keywords, and searches the document set 300. Do. As a result, [Document 1] is obtained from the document set 300 illustrated in FIG. 2, and information on [Document 1] is displayed on the display unit 111 as a search result.

  As described above in detail with specific examples, according to the search support apparatus 200 of the present embodiment, the user selects the document set 300 to be searched in the same manner as the search support apparatus 100 of the first embodiment. By simply specifying, a co-occurrence keyword set that is a combination of keyword candidates having a high co-occurrence probability is detected from the keyword candidates included in the document set 300, and a co-occurrence dictionary 302 based on the co-occurrence keyword set is generated. At the same time, a Patricia tree 303, which is a character string complement rule for complementing the input character string and obtaining keyword candidates constituting the co-occurrence keyword set as the input keyword, is generated. Then, when the user inputs a character string, a suggested word string connecting the input keyword corresponding to the input character string and the co-occurrence word in which the co-occurrence relationship propagates from the input keyword is presented to the user, and the proposed proposal When the user selects a word string, a document including all keyword candidates included in the proposed word string is searched from the search target document set 300. Therefore, according to the search support apparatus 200 of the present embodiment, the search keyword candidates suitable for the search target document set 300 are proposed and the document search is appropriately performed without requiring the user to perform complicated operations in advance. Can help.

  Further, according to the search support apparatus 200 of the present embodiment, a zero co-occurrence keyword set that is a combination of two keyword candidates whose co-occurrence relationship is zero is detected, and a zero co-occurrence dictionary 304 is generated. Then, the proposed word string recognition unit 210 refers to the zero co-occurrence dictionary 304, and displays a word string excluding a word string that simultaneously includes two keyword candidates constituting the zero co-occurrence keyword set as a suggested word string. Output to. Therefore, according to the search support apparatus 200 of the present embodiment, the search is appropriately supported so that only the suggested word string that can reliably obtain the search result is presented to the user and the user does not perform a useless search. can do.

  Each functional configuration in the search devices 100 and 200 according to the first and second embodiments described above, for example, when a computer is used as the hardware configuration of the search support devices 100 and 200, executes a predetermined program on this computer. Can be realized. A program executed by a computer used as the search support apparatus 100 or 200 is a file in an installable format or an executable format, and is a CD-ROM (Compact Disk Read Only Memory), a flexible disk (FD), a CD-R (Compact It is recorded on a computer-readable recording medium such as a disk recordable (DVD) or a DVD (Digital Versatile Disc) and provided as a computer program product.

  Further, a program executed by a computer used as the search support apparatus 100 or 200 may be stored on another computer connected to a network such as the Internet and provided by being downloaded via the network. . Further, a program executed by a computer used as the search support device 100 or 200 may be provided or distributed via a network such as the Internet. In addition, a program executed by a computer used as the search support device 100 or 200 may be provided by being incorporated in advance in a ROM or the like inside the computer.

  The programs executed on the computers used as the search support devices 100 and 200 are the main components of the search support devices 100 and 200 (keyword candidate extraction unit 101, co-occurrence probability calculation unit 103, co-occurrence keyword set detection unit 104, zero Co-occurrence keyword set detection unit 201, co-occurrence dictionary generation unit 105, Patricia tree generation unit 106, zero co-occurrence dictionary generation unit 202, input keyword recognition unit 108, co-occurrence propagation units 109 and 209, suggested word string recognition unit 110, 210, a display unit 111, and a search unit 112). As actual hardware, for example, a CPU (processor) reads out and executes a program from a storage medium, whereby each of the above components Are loaded on the main storage, and the above components are generated on the main storage. There. The main components of the search support apparatuses 100 and 200 according to the embodiment are partially or wholly used by using dedicated hardware such as ASIC (Application Specific Integrated Circuit) or FPGA (Field-Programmable Gate Array). It can also be realized.

  According to the search support apparatus of at least one embodiment described above, the search keyword suitable for the set of documents to be searched without requiring the user to perform complicated operations in advance by including the main components described above. Candidates can be proposed and document search can be supported appropriately.

  As mentioned above, although some embodiment of this invention was described, these embodiment is shown as an example and is not intending limiting the range of invention. These novel embodiments can be implemented in various other forms, and various omissions, replacements, and changes can be made without departing from the scope of the invention. These embodiments and modifications thereof are included in the scope and gist of the invention, and are included in the invention described in the claims and the equivalents thereof.

DESCRIPTION OF SYMBOLS 100 Search support apparatus 101 Keyword candidate extraction part 103 Co-occurrence probability calculation part 104 Co-occurrence keyword set detection part 105 Co-occurrence dictionary generation part 106 Patricia tree generation part 108 Input keyword recognition part 109 Co-occurrence propagation part 110 Proposed word sequence recognition part 111 Display unit 112 Search unit 200 Search support device 201 Zero co-occurrence keyword set detection unit 202 Zero co-occurrence dictionary generation unit 209 Co-occurrence propagation unit 210 Proposed word string recognition unit

Claims (7)

An extraction unit for extracting keyword candidates from a set of documents to be searched;
For a combination of two extracted keyword candidates, a calculation unit that calculates a co-occurrence probability that is a probability that one keyword candidate appears in the same document in the document set together with another keyword candidate;
A first detection unit that detects a co-occurrence keyword set in which the co-occurrence probability satisfies a first condition and is a combination of two keyword candidates;
A first generation unit that generates a co-occurrence dictionary that is a set of dictionary elements having one keyword candidate of the co-occurrence keyword set as a headword and the other keyword candidate as a co-occurrence word;
A second generation unit that generates a character string completion rule that is a rule for complementing an input character string and obtaining keyword candidates included in the co-occurrence keyword set;
A first recognition unit that recognizes a keyword candidate obtained by complementing an input character string according to the character string completion rule as an input keyword;
Co-occurrence is repeated by referring to the co-occurrence dictionary, acquiring a co-occurrence word of a dictionary element having the input keyword as an entry word, and acquiring a co-occurrence word of the dictionary element having the acquired co-occurrence word as an entry word A propagation section;
A second detection unit that detects a zero-co-occurrence keyword set that is a combination of two keyword candidates for which the co-occurrence probability is zero;
A third generation unit for generating a zero co-occurrence dictionary that is a set of dictionary elements in which one keyword candidate of the zero co-occurrence keyword set is a headword and the other keyword candidate is a zero co-occurrence word;
Two keyword candidates that constitute the zero co-occurrence keyword set in a word string that refers to the zero co-occurrence dictionary and connects the input keyword and the co-occurrence word acquired by the processing by the co-occurrence propagation unit A second recognition unit that recognizes a word string that satisfies the second condition among the remaining word strings as a proposed word string,
A presentation unit for presenting the suggested word string;
A search support apparatus, comprising: a search unit that generates a search formula based on the proposed word string and searches the document set when the proposed word string is selected.   The co-occurrence propagation unit is a dictionary in which a co-occurrence word of a dictionary element having the input keyword as an entry word is a first-stage co-occurrence word and an L-1 stage co-occurrence word (L is a natural number of 2 or more) is an entry word. When the co-occurrence word of an element is an L-stage co-occurrence word, the process of acquiring the L-stage co-occurrence word is repeated while increasing L by 1, and between the input keyword and the first-stage co-occurrence word When the cumulative co-occurrence probability, which is a value obtained by sequentially integrating the co-occurrence probabilities between the L-1 stage co-occurrence word and the L-stage co-occurrence word with respect to the co-occurrence probability, becomes equal to or less than a first threshold value, The search support apparatus according to claim 1, wherein the search support apparatus is stopped.   The second recognizing unit is higher in the descending order of the cumulative co-occurrence probability in a word string connecting the input keyword and the co-occurrence word acquired while the co-occurrence propagation unit continues the process. 3. The search support device according to claim 2, wherein N (N is a predetermined natural number) word string is recognized as the proposed word string.   The co-occurrence propagation unit is a dictionary in which a co-occurrence word of a dictionary element having the input keyword as an entry word is a first-stage co-occurrence word and an L-1 stage co-occurrence word (L is a natural number of 2 or more) is an entry word. When the co-occurrence word of an element is an L-stage co-occurrence word, the process of acquiring the L-stage co-occurrence word is repeated while increasing L by 1, and the process is stopped when L reaches a predetermined value. The search support apparatus according to claim 1, wherein:   The first detection unit obtains the number of combinations of two keyword candidates whose co-occurrence probability is greater than a second threshold, and if the obtained number is less than a third threshold, the second threshold is reduced by a predetermined amount. , Repeating the process of obtaining the number of combinations of two keyword candidates having a greater co-occurrence probability than the second threshold value reduced by a predetermined amount, and the two keyword candidates when the obtained number is equal to or greater than the third threshold value The search support apparatus according to claim 1, wherein the combination is detected as the co-occurrence keyword set. A search support method executed in a search support device,
An extraction unit of the search support device extracting keyword candidates from a set of documents to be searched;
The calculation unit of the search support device calculates a co-occurrence probability that is a probability that one keyword candidate appears in the same document in the document set together with another keyword candidate for the combination of the extracted two keyword candidates. When,
A step of detecting a co-occurrence keyword set in which the first detection unit of the search support device is a combination of two keyword candidates in which the co-occurrence probability satisfies a first condition;
A step of generating a co-occurrence dictionary which is a set of dictionary elements having one keyword candidate of the co-occurrence keyword set as a headword and the other keyword candidate as a co-occurrence word;
A step of generating a character string completion rule that is a rule for obtaining a keyword candidate included in the co-occurrence keyword set by complementing an input character string by the second generation unit of the search support device;
A step of recognizing a keyword candidate obtained by complementing an input character string according to the character string complementing rule as an input keyword by the first recognition unit of the search support device;
The co-occurrence propagation unit of the search support device refers to the co-occurrence dictionary, acquires a co-occurrence word of a dictionary element having the input keyword as an entry word, and obtains a dictionary element having the acquired co-occurrence word as an entry word. Repeating the process of obtaining co-occurrence words;
A second detection unit of the search support device detecting a zero co-occurrence keyword set that is a combination of two keyword candidates having a co-occurrence probability of zero;
The third generation unit of the search support device generates a zero co-occurrence dictionary that is a set of dictionary elements in which one keyword candidate of the zero co-occurrence keyword set is a headword and the other keyword candidate is a zero co-occurrence word. Steps,
The second recognition unit of the search support device, in the reference to the zero cooccurrence dictionary, word sequence obtained by connecting the occurrence word acquired by treatment with the input keyword and the co-occurrence propagating unit, the zero When there is a word string that includes two keyword candidates that simultaneously constitute a co-occurrence keyword set, the word string is excluded, and among the remaining word strings, a word string that satisfies the second condition is recognized as a proposed word string. Steps,
The presenting unit of the search support device presenting the suggested word string;
The search unit of the search support device includes a step of generating a search formula based on the proposed word string and searching the document set when the presented suggested word string is selected. Search support method.   On the computer,
  A function of an extraction unit that extracts keyword candidates from a set of documents to be searched;
  A function of a calculation unit that calculates a co-occurrence probability that is a probability that one keyword candidate appears in the same document in the document set together with another keyword candidate for a combination of two extracted keyword candidates;
  A function of a first detection unit that detects a co-occurrence keyword set in which the co-occurrence probability satisfies a first condition and is a combination of two keyword candidates;
  A function of a first generation unit that generates a co-occurrence dictionary that is a set of dictionary elements in which one keyword candidate of the co-occurrence keyword set is a headword and the other keyword candidate is a co-occurrence word;
  A function of a second generation unit that generates a character string completion rule that is a rule for complementing an input character string and obtaining keyword candidates included in the co-occurrence keyword set;
  A function of a first recognition unit that recognizes a keyword candidate obtained by complementing an input character string according to the character string completion rule as an input keyword;
  Co-occurrence is repeated by referring to the co-occurrence dictionary, acquiring a co-occurrence word of a dictionary element having the input keyword as an entry word, and acquiring a co-occurrence word of the dictionary element having the acquired co-occurrence word as an entry word The function of the propagation part;
  A function of a second detection unit for detecting a zero co-occurrence keyword set that is a combination of two keyword candidates with the co-occurrence probability being zero;
  A function of a third generation unit that generates a zero co-occurrence dictionary that is a set of dictionary elements in which one keyword candidate of the zero co-occurrence keyword set is a headword and the other keyword candidate is a zero co-occurrence word;
  Two keyword candidates that constitute the zero co-occurrence keyword set in a word string that refers to the zero co-occurrence dictionary and connects the input keyword and the co-occurrence word acquired by the processing by the co-occurrence propagation unit If there is a word string that simultaneously contains the word string, the function of the second recognition unit that excludes the word string and recognizes a word string satisfying the second condition as a proposed word string among the remaining word strings;
  A function of a presentation unit for presenting the suggested word string;
  A program for realizing a function of a search unit that generates a search formula based on the proposed word string and performs a search on the document set when the proposed word string is presented.

Images