JP5866084B2 - Search device - Google Patents

Description

  The present invention relates to a search apparatus capable of performing an ambiguous search in a large number of registered address data and facility name data using not only an official name but also an abbreviated name or a name of a familiar memory as a search key.

When searching for an address or a facility name, the user does not necessarily store an accurate name, and may search for a common name, abbreviation, a misrecognized name, or the like as a search key.
In addition, in a terminal or device that does not have a keyboard as an input device, such as a car navigation system or a smartphone, a search may be performed based on a result of voice recognition of a voice signal from a microphone or a result of character recognition of input from a touch panel. . In input using these input devices, there are input errors due to user operation errors such as recognition errors and keystroke errors.
In any of the above cases, not only the official name but also a technique for fuzzy searching for a character string or a name with similar pronunciation is required.

  As an example of a technique that enables the fuzzy search as described above, for example, there is a search device disclosed in Patent Document 1. The search device disclosed in Patent Literature 1 searches for similar word candidates from the input keywords using the matching degree of the partial character strings, and further extracts similar words whose edit distance is close to the input keyword from these similar word candidates Then, an ambiguous full-text search is performed by adding it to the search keyword. For example, when “acetaldehyde” is input as a search keyword, similar word candidates including “aceto”, “alde”, “hide”, and the like that are partial character strings are searched. When similar word candidates such as “acetoaldade” and “acetoaldol” are obtained as a result of this search, the edit distance between the input keyword “acetaldehyde” and each similar word candidate is calculated. Of these, the edit distance is short. A full-text search using the similar term “acetoaldade” also reduced search leaks.

JP 2005-11078 A

  However, since the technique described in Patent Document 1 uses similar word candidates for prior narrowing down, there is a problem that it is difficult to cope with a case where the input from the user is a continuous writing including a plurality of words. . In addition, it is necessary to divide words by subjecting the search target to morphological analysis in advance.

  In addition, when searching for similar word candidates, it is necessary to use an index larger than trigrams in order to improve the precision of prior narrowing down, and the dictionary data size becomes enormous, so it is suitable for use in embedded devices such as car navigation systems and mobile phones. There was also a problem that was not suitable. On the other hand, considering the memory capacity, if the index used for pre-restriction is limited to the bigram or less, effective pre-restriction becomes difficult and it takes time to calculate the edit distance.

  In addition, since error bias is not considered at the pre-restriction stage, differences are concentrated in front of the character string or conversely in the backward direction. As a result, there were cases in which it was not possible to sufficiently narrow down before calculating the edit distance. For example, it is assumed that the user erroneously inputs “TSUNAHIMA” when attempting to input “TSUNAHIMA”. At this time, in the prior art, “KUMASHIMA” (TS → K), “TSUNAHARA” (IM → AR), and the like are included as a result of prior narrowing down. Character strings with differences are hardened in one place and unnatural.

  The present invention has been made to solve such a problem, and an object of the present invention is to obtain a search device that can obtain an ambiguous search result at a high speed with a reduced memory.

  The search device according to the present invention includes a syllable bigram dictionary in which a reading of data to be searched is divided and stored in units of two syllables, a name search dictionary in which official names are stored, and a part of characters in a given reading character string Based on the similar reading candidates created by the similar reading candidate creation unit that searches for similar reading strings from the syllable bigram dictionary and creates similar reading candidates by combining the searched character strings. A name search unit for searching for an official name from the name search dictionary.

  The search device according to the present invention searches for a reading character string similar to a part of the character string of the reading character string from the syllable bigram dictionary, creates a similar reading candidate by combining the searched character strings, and searches for a formal name. It is what I did. As a result, the ambiguous search result can be obtained at high speed with less memory.

It is a block diagram which shows the search device of Embodiment 1 of this invention. It is a flowchart which shows the search process in the search device of Embodiment 1 of this invention. It is a block diagram which shows the detail of the similar reading candidate preparation part in the search device of Embodiment 1 of this invention. It is explanatory drawing of the syllable bigram dictionary in the search device of Embodiment 1 of this invention. It is explanatory drawing which shows the concept of the similar reading candidate tree structure data in the search device of Embodiment 1 of this invention. It is explanatory drawing of the similar reading candidate tree structure data in the search device of Embodiment 1 of this invention. It is a flowchart which shows the similar reading candidate creation process in the search device of Embodiment 1 of this invention. It is a flowchart which shows the syllable bigram reading character string search process in the search device of Embodiment 1 of this invention. It is explanatory drawing which shows the relationship with the dictionary name search part for name searches in the search device of Embodiment 1 of this invention. It is explanatory drawing of the name list in the search device of Embodiment 1 of this invention. It is a block diagram which shows the search device of Embodiment 2 of this invention. It is a block diagram which shows the detail of the name prefetching similar reading candidate preparation part in the search device of Embodiment 2 of this invention. It is a flowchart which shows the name prefetching similar reading candidate creation process in the search device of Embodiment 2 of this invention. It is a block diagram which shows the search device of Embodiment 3 of this invention. It is a block diagram which shows the detail of the division | segmentation limited similar reading candidate preparation part in the search device of Embodiment 3 of this invention. It is a flowchart which shows the division | segmentation limited similar reading candidate creation process in the search device of Embodiment 3 of this invention. It is explanatory drawing which shows the example of candidate rejection by the input reading division | segmentation determination process in the search device of Embodiment 3 of this invention.

Hereinafter, in order to explain the present invention in more detail, modes for carrying out the present invention will be described with reference to the accompanying drawings.
Embodiment 1 FIG.
FIG. 1 is a block diagram showing a search device according to Embodiment 1 of the present invention. The facility name search in the car navigation system will be described below as an example. However, the present invention is not limited to the facility name search in the car navigation system, and search processing performed on an embedded device such as an address search or an electronic manual search. It can be applied in general.

  The search apparatus shown in FIG. 1 includes an input unit 101, a similar reading candidate creation unit 103, a syllable bigram dictionary 104, a name search unit 106, and a name search dictionary 107. The input unit 101 accepts an input operation by a user such as a software keyboard or voice recognition, and converts it into read character string data 102. The similar reading candidate creation unit 103 creates a similar reading candidate list 105 with reference to the syllable bigram dictionary 104 stored by dividing the reading of the search target data into units of two syllables. The name search unit 106 refers to the name search dictionary 107 and outputs name data including each reading character string in the similar reading candidate list 105 as search result data 108.

  The search device of the present invention is configured using a computer, and the input unit 101, the similar reading candidate creation unit 103, and the name search unit 106 are configured by the CPU executing software corresponding to each function. . Alternatively, any of the processing units may be configured with dedicated hardware. The syllable bigram dictionary 104 and the name search dictionary 107 are stored on a storage device such as a hard disk device.

Hereinafter, the processing contents of the first embodiment will be described with reference to FIGS. 1 to 10 as appropriate. FIG. 2 shows a basic processing flow of search processing.
Step ST201 is a similar reading candidate creation process. The similar reading candidate creation unit 103 refers to the syllable bigram dictionary 104 and creates a similar reading candidate by combining the syllable bigrams of the search results. Details of the similar reading candidate creation process will be described later.
Step ST202 is a name search process. The name search unit 106 refers to the name search dictionary 107 and outputs data included in a part of the reading including the similar reading candidate list 105 as the search result data 108. Details of the name search process will be described later.

  As described above, dividing the step ST201 for generating similar reading candidates and the step ST202 for searching for names has the following advantages. In other words, the ambiguous search process that increases the index data capacity and the calculation amount is a process based on the syllable bigram in step ST201, thereby reducing the number of target data and suppressing an increase in the capacity and calculation amount. In addition, the name search process in step ST202 in which the number of search objects is extremely large can be processed with emphasis on speed performance and memory performance by performing a simple forward match search process without performing an ambiguous search.

Next, details of the similar reading candidate creation process (step ST201) will be described.
FIG. 3 shows a detailed configuration diagram of the similar reading candidate creation unit 103. The similar reading candidate tree structure data generation unit 301 obtains information on syllable bigrams similar to each partial character string of the input reading character string by the syllable bigram dictionary search unit 302, and combines these information with similar reading candidate tree structure data 303. Is generated. Details of this data will be described later. The similar reading candidate selection unit 304 includes a syllable reading editing distance determination unit 305, a reading editing distance accumulated value determination unit 306, a reading editing distance accumulated value calculation unit 307, a new candidate number determination unit 308, and a similar reading candidate score order selection unit 309. Then, each means of the reading appearance probability cumulative value calculation unit 310 is called, and the similar reading candidate tree structure data 303 generated by the similar reading candidate tree structure data generation unit 301 is selected according to each predetermined condition. The operation and selection conditions of each processing unit will be described later. The similar reading candidate extraction unit 311 generates and selects the similar reading candidate tree structure data 303 for all character positions of the input reading character string, and then selects the similar reading candidate character string from the similar reading candidate tree structure data 303. (Similar reading candidate list 105) is extracted.

  The syllable bigram dictionary 104 divides the reading character string assigned to the search target data into syllable units in advance, and based on the reading character strings of all two syllables appearing in the syllable-based reading character string. create. A syllable is a type of segmental unit that separates consecutive speech sounds. It is a unit composed of a vowel as a center, with the vowel alone or with one or more consonants before and after the vowel. In Japanese, it is a unit corresponding to one hiragana character. The syllable bigram composed of the concatenation of two syllables has the advantage that the number of the syllable bigrams is very small compared to the character trigram used in the apparatus described in Patent Document 1, and the dictionary data size can be reduced.

  FIG. 4 shows an example of a syllable bigram dictionary. The dictionary data includes a data storage portion (data storage unit 401) and a character index 402 for searching the dictionary at high speed. The data storage unit 401 stores the syllable reading bigram ID (104a), the first syllable reading (104b), the second syllable reading (104c), and the appearance probability (104d) in each record. These are read character strings with syllable breaks added to the search target data, for example, from the read character string “MI TSU BI SHI DE N KI” of the search target data, “MI TSU”, “TSU BI”, “BI”. The syllable bigrams “SHI”, “SHI DE”, “DEN”, and “NKI” are extracted and stored in each record of the data storage unit 401. At that time, the appearance frequency in the original data of each syllable bigram is calculated in advance separately, and the appearance probability (104d) is calculated and stored based on this.

  The character index 402 stores index information for accessing the data storage unit 401 at a high speed. Specifically, it has transposed index information (104f) for acquiring a syllable bigram including each character in the input reading character string at high speed. In the example of FIG. 4, the character bigram (104e) “AK” is shown to appear in the record with the syllable reading bigram ID of 2 and the record with the ID of 100.

  Next, the outline of the similar reading candidate tree structure data 303 will be described with reference to FIG. The similar reading candidate tree structure data 303 is an intermediate data structure generated to enumerate various possibilities. The syllable bigram candidate [n] shown in the figure indicates information obtained by searching the syllable bigram dictionary 104. Each candidate holds the reading of the first syllable, the reading of the second syllable, the character position in the input reading character string, etc. doing). The syllable bigram candidates [1] to [3] in the figure are candidates similar to the partial character string starting at the character position 0 of the input reading character string. Syllable bigrams [5] and [6] are candidates similar to a partial character string starting at character position 1. Syllable bigrams [7] to [10] are candidates similar to a partial character string starting at character position 3. Here, the syllable bigram candidates [5] and [6] are candidates connected to the syllable bigram candidate [3], and represent similar reading candidates “TO SU NA” and “TO SU MA”, respectively. The syllable bigram candidates [7] and [8] are candidates that are connected to the syllable bigram candidate [1], and represent similar reading candidates “TSU NA SHI” and “TSU NA KI”, respectively. As described above, the similar reading candidate tree structure data 303 represents reading character string candidates similar to the input reading character string by a combination of syllable bigrams.

  FIG. 6 shows a specific example of the similar reading candidate tree structure data 303. This is an embodiment of the tree structure conceptually shown in FIG. One row in the table of FIG. 6 corresponds to one of the syllable bigram candidates shown in FIG. The data of each candidate includes candidate ID (303a), first syllable reading (303b), second syllable reading (303c), first syllable start position (303d), second syllable start position (303e), and second syllable end. It consists of position (303f), first syllable editing distance (303g), second syllable editing distance (303h), appearance probability (303i), and parent node ID (303j). How each information is set will be described later. The tree-structured link information is expressed as a parent node ID of 303j.

  FIG. 7 shows a detailed flow of the similar reading candidate creation process (step ST201). Hereinafter, a case where a reading character string “TSUNAHIMA” illustrated in FIG. 5 is input will be described as an example. Steps ST701 to ST713 are loop processes performed for the number of input reading characters, and the processes from step ST702 to step ST712 are repeated while increasing the variable cur by one. Here, the variable “cur” represents the character position in the input reading character string, and when the reading character string shown in FIG.

  Step ST702 is a syllable bigram reading character string search process. Here, the syllable bigram dictionary search unit 302 searches the syllable bigram dictionary 104 for a syllable bigram having a reading similar to the partial character string starting at the character position cur of the input reading character string.

  FIG. 8 shows a detailed flow of the syllable bigram reading character string search process (step ST702). Step ST801 is a character unit search process. Here, the syllable bigram including each character of the partial character string starting at the character position cur is searched using the character index 402 shown in FIG. When cur = 0, data with syllable reading bigram IDs 200, 201, 202, and 203 is retrieved. When cur = 1, data with syllable reading bigram IDs 204 and 205 is retrieved. When cur = 3, data with syllable reading bigram IDs of 100, 101, 102, and 103 is retrieved. This search result is provisionally registered in the similar reading candidate tree structure data 303. When cur = 0, candidate IDs 1, 2, 3, and 4 are candidates. When cur = 1, candidate IDs 5 and 6 are candidates. When cur = 3, candidate IDs 7 and 7. 8, 9, and 10 candidates are provisionally registered. First, the first syllable reading (303b), the second syllable reading (303c), and the appearance probability (303i), which are information obtained from the syllable bigram dictionary 104, are stored in each candidate. Subsequently, based on the character position cur and the reading length of each syllable, the first syllable start position (303d), the second syllable start position (303e), and the second syllable end position (303f) are calculated and stored.

  Step ST802 is edit distance calculation processing. Here, the edit distance between the syllable bigram data searched in step ST801 and the partial character string starting at the character position cur of the input character string is calculated. Since the edit distance calculation is a very expensive calculation process, ranking is performed in consideration of the number of hit characters, the appearance probability of characters, and the like at the time of retrieval in units of characters in step ST801. The edit distance is calculated for a predetermined number of higher ranking results, and the ranking is re-performed based on the edit distance calculation result. Here, the edit distance calculation is performed for the partial character string of the first syllable reading and the input reading, and the partial character string of the second syllable reading and the input reading, respectively. The calculation result is stored in the first syllable editing distance (303g) and the second syllable editing distance (303h) of the data temporarily registered in the similar reading candidate tree structure data 303.

  Step ST703 is a syllable-specific reading edit distance determination process. Here, the syllable-by-syllable reading edit distance determination unit 305 leaves candidates whose edit distance for each syllable calculated in step ST802 is a predetermined value or less, and rejects other candidates. Here, a candidate whose edit distance for each syllable is 1 or less is left, and the syllable bigram candidate [4] in FIG. 5 is rejected. In order to indicate that the candidate is rejected, the parent node ID (303h) is changed to -1.

  In the double loop processing from step ST704 to step ST710, it is determined whether all syllable bigram candidates obtained as a result of step ST703 and all existing syllable bigram candidates can be connected, and according to the reading edit distance accumulated value. After selecting candidates, a new syllable bigram candidate is generated. When cur = 0, there is no previous candidate, so the connection determination in step ST706 is always true and the process proceeds to step ST707. When cur> 0, connection determination between each syllable bigram candidate newly obtained up to step ST703 and an existing syllable bigram candidate is performed. Connection determination is performed based on whether the first syllable reading of the new syllable bigram and the second syllable reading of the existing syllable bigram match. For example, when cur = 1, the first syllable reading of the syllable bigram candidate [5] [6] and the second syllable reading of the syllable bigram candidate [3] match, so that the syllable bigram candidate [5] [6] 3 is stored in the parent node ID. Similarly, when cur = 3, 1 is stored in the parent node ID of the syllable bigram candidate [7] [8], and 2 is stored in the parent node ID of the syllable bigram candidate [9] [10].

  Step ST707 is candidate selection processing corresponding to the reading editing distance accumulated value. Here, the reading / editing distance cumulative value determination unit 306 selects a syllable bigram candidate according to the reading / editing distance cumulative value calculated by the reading / editing distance cumulative value calculation unit 307.

  The reading editing distance accumulated value is calculated based on the first syllable editing distance (303g) and the second syllable editing distance (303h) of the similar reading candidate tree structure data 303. The cumulative edit distance of a syllable bigram candidate adds the second syllable edit distance (303h) to the syllable bigram candidate at character position 0 while following the parent node ID, and finally adds the first syllable edit distance (303g). It is obtained by doing. For example, in the case of the syllable bigram candidate [8] shown in FIGS. 5 and 6, the cumulative reading edit distance value is 1 + 0 + 0 = 1. Further, in the case of the syllable bigram candidate [10], the cumulative reading edit distance value is 1 + 1 + 0 = 2.

  The threshold value Th1 for determining whether or not a candidate is accepted is determined by the length L of the input reading character string. If the input character string is long, it is highly possible that an error is included, so a large value is set. For example, L / 4 is set as a threshold value, and similar reading candidates whose reading editing distance accumulated value is larger than this are rejected. When cur = 0, all candidates of syllable bigram candidates [1] [2] [3] have a reading edit cumulative value of 1 or less, and all candidates are adopted without being rejected. In the case of cur = 1, the cumulative value of the edit distance is 2 in the syllable bigram candidate [6], and is rejected. In the case of cur = 3, the syllable bigram [10] is similarly rejected.

  Step ST708 is a similar reading candidate tree structure update process. The candidates provisionally registered in step ST801 are formally registered. For example, when cur = 0, the rows with candidate IDs 1, 2, and 3 in FIG. 6 are formally registered. Since the candidate ID of 4 is already rejected in step ST703, it is not officially registered. Although it is also described in FIG. 6 to facilitate the correspondence with FIG. 5, it may be overwritten by processing after cur = 1 in order to suppress the memory usage. Similarly, when cur = 1, a candidate with a candidate ID of 5 is officially registered, and a candidate with a candidate ID of 6 is rejected in step ST707 and is not officially registered. Similarly, when cur = 3, candidates with candidate IDs 7, 8, and 9 are formally registered.

  When the double loop process from step ST704 to step ST710 ends, the process proceeds to step ST711. In step ST711, the new candidate number determination unit 308 performs selection processing determination based on the number of new syllable bigram candidates added at the current character position cur.

  When the number of new candidates is greater than the predetermined threshold value Th2, the process proceeds to step ST712, and the similar reading candidate score order selection unit 309 arranges the syllable bigram candidates in order of scores, leaves the higher score candidates, and rejects the lower score candidates. . Here, the threshold Th2 is appropriately set according to external conditions that the corresponding function should satisfy, such as memory usage and speed performance.

  The score calculation is performed by the reading appearance probability cumulative value calculation unit 310. Specifically, the score of a syllable bigram candidate is obtained by sequentially adding the appearance probability (303i) up to the syllable bigram candidate at character position 0 while following the parent node ID. Also, the edit distance accumulated value calculated by the reading edit distance accumulated value calculation unit 307 may be used as a score. The method for calculating the edit distance accumulated value is the same as that in step ST707.

  When the loop processing from step ST701 to step ST713 ends, the process proceeds to step ST714. Although not shown in FIG. 7, the connection determination in step ST706 may be performed immediately after step ST701. That is, if none of the existing syllable bigram candidates is the second character syllable end position (303f) at the current character position cur, even if the processes of steps ST702 and ST703 are performed, all of them are rejected at step ST706. Such a prior connection test is performed at the beginning of the loop processing, and if there is no connectable candidate, the processing after step ST702 can be skipped to prevent useless processing.

  In step ST714, the similar reading candidate extraction unit 311 extracts an actual similar reading candidate character string from the similar reading candidate tree structure data 303. The similar reading candidate character string is a character that follows the parent node ID for the syllable bigram candidate whose second syllable end position (303f) reaches the end of the input reading character string at the end of the tree structure data. This is obtained by sequentially connecting the second syllable readings (303c) up to the syllable bigram candidate at position 0, and finally reversing the character string obtained by connecting the first syllable readings (303b).

  The above is the details of the similar reading candidate creation process in step 201 of FIG. Next, the name search process in step ST202 will be described.

  In the name search process, the name search unit 106 searches the name data including the similar reading candidate list 105 with reference to the name search dictionary 107 and outputs it. Here, for example, a search method disclosed in JP 2010-205119 A is used. Since the details of the search method are described in the literature, the outline of the search process will be described below.

  FIG. 9 shows a configuration example of the name search dictionary 107. The name search unit 106 searches for a node corresponding to the given similar reading candidate character string based on the double sequence index data 901. Subsequently, based on the minimum / maximum child node index 902, the node that becomes the minimum character string and the node that becomes the maximum character string in the dictionary order among the child nodes of the node searched above are searched. Further, all the names from the name corresponding to the minimum node searched above to the name corresponding to the maximum node are extracted from the name list 903 as search result data 108.

  FIG. 10 shows an example of the name list 903. The name list 903 includes at least a name ID (903a) that uniquely identifies each name, a reading character string (903b) of the name, and type information (903c). In FIG. 10, two lines having the same name ID of 3 are shown in two places, but this is expanded in advance so that the reading character string (903b) can be searched even from characters in the middle. This is because it is indexed.

  In the above description, the search method using the double array index described in Japanese Patent Application Laid-Open No. 2010-205119 is shown as an example. However, the name search process in step ST202 is performed by each similar reading candidate included in the similar reading candidate list 105. Any method can be applied as long as the name data including the reading character string can be searched from the character string at high speed. For example, a database for an embedded device may be used, or a structure embedded in the tree structure index data for high-speed retrieval of information included in the name list 903 shown in FIG.

As described above, in the first embodiment, the syllable bigram dictionary 104 in which the reading of the search target data is divided and stored in units of two syllables, and the reading character string similar to a part of the character string input by the user. Is prepared from the syllable bigram dictionary 104 and combined to create similar reading candidates, so that there is no need for a prior word break and multiple users such as Japanese, Korean, Chinese, etc. Flexible and high-speed fuzzy search is possible in a language in which words can be continuously input.
Further, by providing a similar reading candidate selection unit 304 for creating a similar reading candidate while appropriately selecting candidates according to a predetermined selection criterion, the number of input characters increases, so that the combination does not increase rapidly. It is possible to create a similar reading character string candidate at high speed in memory.

  As described above, according to the search device of the first embodiment, the syllable bigram dictionary storing the reading of the search target data divided into units of two syllables, the name search dictionary storing the official names, are given. A similar reading candidate creation unit that searches a similar reading candidate character string from a syllable bigram dictionary and creates a similar reading candidate by combining the searched character strings. Since it has a name search unit that searches the official name from the name search dictionary based on the similar reading candidates that have been created, even if the search key is a common name, abbreviation, or misrecognized name As a search device capable of fuzzy searching and outputting the official name, it is possible to realize memory-saving and high-speed processing.

  In addition, according to the search device of the first embodiment, the similar reading candidate creation unit includes the similar reading candidate selection unit that selects similar reading candidates created by combining the searched character strings. A vague search result can be obtained.

  Further, according to the search device of the first embodiment, the similar reading candidate selection unit edits the reading of each syllable searched from the syllable bigram dictionary and the corresponding partial character string of the given reading character string. A syllable-by-syllable reading edit distance determination unit that selects similar reading candidates whose distance is equal to or smaller than a predetermined value and rejects other similar reading candidates is provided, so that an ambiguous search result can be obtained at high speed with less memory.

  Further, according to the search device of the first embodiment, the similar reading candidate selection unit edits the reading of each syllable searched from the syllable bigram dictionary and the corresponding partial character string of the given reading character string. A reading editing distance cumulative value calculation unit that calculates the cumulative value of distances, and a similar reading candidate whose reading editing distance cumulative value calculated by the reading editing distance cumulative value calculation unit is equal to or less than a predetermined value, and other similar reading candidates Is provided with a reading edit distance cumulative value determination unit that rejects the ambiguity, so that an ambiguous search result can be obtained at high speed with a reduced memory.

  Further, according to the search device of the first embodiment, the similar reading candidate selection unit determines the predetermined value by the new candidate number determination unit that determines whether or not the number of similar reading candidates exceeds a predetermined value, and the new candidate number determination unit. If it is determined that the number of similar reading candidates has been exceeded, the score of the similar reading candidate is calculated, and the similar reading candidate score order selection unit that leaves only the similar reading candidate having a predetermined value higher in the score is provided. Can do.

  Further, according to the search device of the first embodiment, the similar reading candidate score order selection unit calculates the score of the similar reading candidate based on the cumulative value of the reading editing distance calculated by the reading editing distance cumulative value calculation unit. As a result, ambiguous search results can be obtained at high speed with less memory.

  Further, according to the search device of the first embodiment, the similar reading candidate selection unit has the reading appearance cumulative value calculation unit that calculates the cumulative value of the appearance probability of each syllable searched from the syllable bigram dictionary, Since the reading candidate score order selection unit calculates the score of the similar reading candidate based on the calculation result of the reading appearance probability cumulative value calculation unit, the ambiguous search result can be obtained at high speed with less memory.

Embodiment 2. FIG.
FIG. 11 is a configuration diagram illustrating the search device according to the second embodiment. In the second embodiment, a name pre-read similar reading candidate creation unit 1101 is provided instead of the similar reading candidate creation unit 103 in the first embodiment. The name pre-read similar reading candidate creation unit 1101 creates the similar reading candidate list 105a with reference to the syllable bigram dictionary 104 that stores the search target data divided into two syllable units. At that time, the name search dictionary 107 is also referred to as part of the candidate selection process. Since other configurations are the same as those of the first embodiment shown in FIG. 1, the same reference numerals are given to corresponding portions, and descriptions thereof are omitted.

  FIG. 12 shows a detailed configuration diagram of the name prefetching similar reading candidate creation unit 1101. The name presence / absence determination unit 1201 refers to the name search dictionary 107 to determine whether a similar reading candidate exists in the name dictionary. Since the configuration other than this is the same as the configuration of the first embodiment shown in FIG. 3, the same reference numerals are given to corresponding portions, and the description thereof is omitted.

FIG. 13 is a flowchart illustrating a name prefetching similar reading candidate creation process according to the second embodiment. Since the processing other than the name search presence / absence determination processing in step ST1301 is the same as the processing shown in FIG. 7 of the first embodiment, description of processing other than step ST1301 is omitted. In step ST1301, first, a similar reading candidate character string is extracted by the similar reading candidate extraction unit 311 for each syllable bigram candidate of the processing result of step ST707. Next, the name presence / absence determination unit 1201 refers to the name search dictionary 107 to determine whether or not this similar reading candidate character string exists in the name search dictionary 107. If it does not exist, the search fails in the subsequent name search process, so the candidate is rejected at this point.
Note that the final name data itself is not necessary in the name presence / absence determination process in step ST1301, and therefore a faster search process is performed by using only the double array index data 901 (see FIG. 9).

  As described above, according to the search device of the second embodiment, the similar reading candidate selection unit selects a similar reading candidate that can be part of the name by searching the name search dictionary for the similar reading candidate. In addition, a name presence / absence determination unit that rejects other similar reading candidates is provided, so that only those that can be obtained as a result of an actual name search among similar reading candidates can be selected. Processing is possible.

Embodiment 3 FIG.
FIG. 14 is a block diagram showing the search device of the third embodiment. In the third embodiment, instead of the input unit 101 and the similar reading candidate creation unit 103 in the first embodiment, a reading conversion input unit 1401 and a delimited limited similar reading candidate creation unit 1403 are provided. In the reading conversion input unit 1401, the user inputs a reading character string and converts it into a notation character string, such as Japanese Kana-Kanji conversion, Chinese Pinyin Kanji conversion, and Korean character Hangul conversion. The result is acquired as delimited read character string data 1402 in which a delimiter is given to the read character string so as to correspond to each character of the notation. A delimited limited similar reading candidate creation unit 1403 creates a similar reading candidate list 105b by using the delimited reading character string data 1402 and referring to the syllable bigram dictionary 104 as in the first embodiment. Since other configurations are the same as those of the first embodiment shown in FIG. 1, the corresponding parts are denoted by the same reference numerals and description thereof is omitted.

  FIG. 15 is a configuration diagram illustrating details of the delimiter limited similar reading candidate creation unit 1403. The delimitation limited similar reading candidate tree structure data generation unit 1501 acquires information on syllable bigrams similar to each partial character string in the delimited reading character string data 1402 by the syllable bigram dictionary search unit 302, and combines these information to obtain similar readings. Candidate tree structure data 303 is generated. Since the configuration other than this is the same as the configuration of the first embodiment shown in FIG. 3, the same reference numerals are given to corresponding portions, and the description thereof is omitted.

  FIG. 16 is a flowchart showing a segmentation limited similar reading candidate creation process according to the third embodiment. Since the processing other than the input reading delimitation determination processing in step ST1601 is the same as the processing shown in FIG. 7 of the first embodiment, the description of processing other than step ST1601 is omitted. In step ST1601, it is determined whether or not to register the syllable bigram search result in the similar reading candidate tree structure data 303 based on the input reading break information. Specifically, it is determined whether or not the current character position cur and the reading length of the first syllable of the search result match the input reading character string delimiter. If they do not match, the candidate is rejected. .

  An example in which candidates are rejected in step ST1601 is shown in FIG. In this example, the delimited input reading character string is “TSU / NA / HI / MA”, whereas the syllable bigram candidate [3] corresponds to the delimited input input “TSU” for two syllables. It is rejected because the syllable break does not match.

  As described above, according to the search device of the third embodiment, the syllable bigram dictionary in which the reading of the search target data is divided and stored in units of two syllables, the name search dictionary that stores official names, and the syllable units The syllable bigram dictionary searches for a reading string that is similar to a part of the separated reading string separated by, and creates a similar reading candidate by combining the searched strings. On the other hand, a delimiter limited similar reading candidate creation unit that performs a match determination between a syllable delimiter in a delimiter reading character string and a search result of the syllable bigram dictionary, rejects a syllable bigram search result that does not match the delimiter, and creates a similar reading candidate Since it has a name search unit that searches for a formal name from the dictionary for name search based on the similar reading candidate created by the delimiter limited similar reading candidate creation unit, it matches the input Not to reject the unwanted syllables bigram candidates, it is possible to further saving memory and fast fuzzy search.

  In the present invention, within the scope of the invention, any combination of the embodiments, or any modification of any component in each embodiment, or omission of any component in each embodiment is possible. .

  As described above, the search device according to the present invention searches the syllable bigram dictionary for a reading character string similar to a part of the character string of the reading character string, and creates a similar reading candidate by combining the searched character strings. The official name is searched, and can be applied to search processing performed on an embedded device such as a car navigation facility name search, an address search, and an electronic manual search.

  101 Input unit, 102 Reading character string data, 103 Similar reading candidate creation unit, 104 Syllable bigram dictionary, 105, 105a, 105b Similar reading candidate list, 106 Name search unit, 107 Name search dictionary, 108 Search result data, 301 Similar Reading candidate tree structure data generation unit, 302 syllable bigram dictionary search unit, 303 similar reading candidate tree structure data, 304 similar reading candidate selection unit, 305 syllable-specific reading editing distance determination unit, 306 reading editing distance accumulated value determination unit, 307 reading Edit distance cumulative value calculation unit, 308 New candidate number determination unit, 309 Similar reading candidate score order selection unit, 310 Reading appearance probability cumulative value calculation unit, 311 Similar reading candidate extraction unit, 1101 Name prefetched similar reading candidate creation unit, 1201 Name Presence / absence determination unit, 1401 Reading conversion input unit, 1402 Sticky reading character string data, 1403 delimiter limited similar reading candidate creation unit, 1501 delimiter limited similar reading candidate tree structure data generation unit.

Claims (9)

A syllable bigram dictionary that stores the search target data divided into two syllable units;
A dictionary for name search that stores the official name,
A similar reading candidate creation unit that searches the syllable bigram dictionary for a reading character string similar to a part of a given reading character string, and creates a similar reading candidate by combining the searched character strings;
A search device comprising: a name search unit that searches for a formal name from the name search dictionary based on the similar reading candidate created by the similar reading candidate creation unit.   The search device according to claim 1, wherein the similar reading candidate creation unit includes a similar reading candidate selection unit that selects similar reading candidates created by combining the searched character strings.   The similar reading candidate selection unit selects similar reading candidates whose edit distance between a reading of each syllable searched from the syllable bigram dictionary and a partial character string of the given reading character string corresponding thereto is a predetermined value or less. The search device according to claim 2, further comprising a syllable-by-syllable reading edit distance determination unit that selects and rejects other similar reading candidates.   The similar reading candidate selection unit calculates a cumulative value of an editing distance between a reading of each syllable searched from the syllable bigram dictionary and a corresponding partial character string of the given reading character string. Cumulative value calculation unit and reading editing distance cumulative value determination that selects similar reading candidates whose reading editing distance cumulative value calculated by the reading editing distance cumulative value calculation unit is a predetermined value or less and rejects other similar reading candidates The search device according to claim 2, further comprising: a section.   When the similar reading candidate selection unit determines that the number of similar reading candidates exceeds a predetermined value, and the new candidate number determination unit determines that the number of similar reading candidates exceeds a predetermined value, the similar reading candidate selection unit The search device according to claim 2, further comprising a similar reading candidate score order selection unit that calculates a score and leaves only the similar reading candidates of the predetermined value higher in the score.   The said similar reading candidate score order selection part calculates the score of the said similar reading candidate based on the accumulated value of the reading editing distance calculated by the said reading editing distance cumulative value calculation part. Search device.   The similar reading candidate selection unit includes a reading appearance cumulative value calculation unit that calculates a cumulative value of appearance probabilities of each syllable searched from the syllable bigram dictionary, and the similar reading candidate score order selection unit includes the reading The search device according to claim 5, wherein the score of the similar reading candidate is calculated based on the calculation result of the appearance probability cumulative value calculation unit.   The similar reading candidate selection unit selects a similar reading candidate that can be a part of a name by searching the name search dictionary with respect to the similar reading candidate, and determines whether there is a name that rejects other similar reading candidates The search device according to claim 2, further comprising a section. A syllable bigram dictionary that stores the search target data divided into two syllable units;
A dictionary for name search that stores the official name,
A reading character string similar to a partial character string of a delimited reading character string divided in syllable units is searched from the syllable bigram dictionary, a similar reading candidate is created by combining the searched character strings, and the separation A delimiter that creates a similar reading candidate by rejecting a syllable bigram search result that does not match the syllable bigram dictionary search result for the diacritic character string A limited similar reading candidate creation section;
A search device comprising: a name search unit that searches a formal name from the name search dictionary based on the similar reading candidates created by the delimitation limited similar reading candidate creation unit.

Images