JPH08161325A - Kana (japanese syllabary)/kanji (chinese character) conversion device and method therefor - Google Patents

Abstract

PURPOSE: To check modification between clauses which are written with a space and to learn on modification when a next candidate is selected. CONSTITUTION: When the flag of modification learning is turned on (step S600), a modified word exists in a modifying word clause (steps S610 and 620). When the next candidate is selected in the modifying word clause or a modified word clause (steps S630 and 540), modification information is generated and is registered in a dictionary (steps S650 and 660). Only when the part of speech on the next candidate is the same as the part of speech on the candidate which is judged that modification is satisfied, information is registered. Consequently, the judgement of modification can widely be executed without making more a modification dictionary 98 uselessly large in volume.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a kana-kanji conversion device and a kana-kanji conversion method, and more specifically, it uses kana-kanji characters used to select kanji candidates for words constituting a bunsetsu, using information on the dependency between words. The present invention relates to a conversion device and a kana-kanji conversion method.

[0002]

2. Description of the Related Art Conventionally, various kana-kanji conversion devices for converting a kana character string input from a keyboard into a desired kana-kanji mixed sentence have been proposed as an input device for Japanese sentences or an editing device for Japanese sentences. Has been done. In such a kana-kanji conversion device, it is not necessary for the user to specify the delimiter positions of words and phrases one by one, and it is desired that the converted character string has the notation desired by the user. Since there are many homophones and homonomys in Japanese, in order to get the desired mixed kana-kanji sentence without error, it is probably necessary to finally analyze the meaning of the sentence. In order to analyze the meaning, a knowledge base of at least tens of thousands of words that are organically related is necessary, which is extremely difficult to realize.

Therefore, in the conventional kana-to-kanji conversion device, devising the process of segmenting and writing bunsetsu, and the learning process in selecting homonyms and oppositions, an attempt is made to obtain the desired result without the analysis of meaning. As for the processing of phrase segmentation,
For the 2-bunsetsu longest matching method in which the 2nd bunsetsu that can form the longest bunsetsu is obtained as the first candidate, and the words and combinations of words that can be candidates for the words that make up the bunsetsu. There is a minimum cost method in which a cost is added and a clause whose score satisfies a predetermined condition is set as a first candidate. In the learning process, the most recently selected homophonic word from the homonym is used as the next candidate for learning the homonym object, and the length of the phrase containing a word is used immediately before. It is known to learn bunsetsu length, which gives top priority to the length specified by the person.

Furthermore, recently, attention has been paid to a specific relationship between words (for example, "hot" and "tea" in "hot tea" or "hot" and "summer" in "hot summer", and this relationship By preparing a dictionary that stores, when one word (for example, “tea”) is specified, the word related to this word (for example, “hot” among candidates for “hot”) is set as the first word.
Those which are selected as candidates have also been proposed (for example, "Kana-Kanji conversion device" in JP-A-3-105664, "Kana-Kanji conversion device" in JP-A-4-278661). The particular relationship between these words is called "dependency" or "co-occurrence."

[0005]

However, the syntax rules of the language actually used are extremely complicated, and it is extremely difficult to prepare all the dependency relations in the dictionary. In addition, it is not necessary for all the Japanese sentences created by a specific individual to be modified, and if you create a dictionary that assumes the existence of all modified relationships, the modification will be established between any clauses. In some cases, it may not be possible to improve the accuracy of Kana-Kanji conversion.

Furthermore, there are various sentence patterns in Japanese. For example, there are relationships such as "red" + "flower" and "flower" + "red", or "vegetable" + "eat" and "vegetable". +
There is a causative or passive relationship such as "eat" and "vegetable" + "eatable", and if you try to create a dictionary that shows whether a dependency relationship is established including these relationships, the dictionary will be huge. There was a problem that it became a thing.

The kana-kanji conversion device and the kana-kanji conversion method of the present invention are intended to clarify such a problem and to obtain a desired kana-kanji mixed sentence by utilizing the dependency relation. It was

[0008]

The kana-kanji conversion device according to claim 1 is a kana-kanji conversion device for inputting a kana character string and referring to a grammar dictionary to generate a kana-kanji mixed character string candidate. And a word segmenting unit for segmenting the input kana character string into segment segments by referring to the grammar dictionary, and a candidate word displaying unit that indicates, in each segment segment, the words constituting the segment in a converted form. When,
A dependency information dictionary that stores information on the dependency between predetermined phrases, and a single phrase segmented as the starting point is used as a starting point, and the dependency information dictionary is referenced to verify the existence of a dependency with another phrase. When a word different from the word displayed by the candidate word display means is selected in at least one of the clauses whose dependency relationship is verified by the dependency test means,
The gist of the present invention is to include a dependency learning unit that learns the selected word as dependency information.

In this kana-kanji conversion device, the segmentation writing unit refers to the grammar dictionary to segment the input Kana character string into segments, and in each segmented segment, the candidate word display unit includes the words constituting the segment. In converted form. In this state, the dependency checking means refers to the dependency information dictionary that stores the information of the dependency between the predetermined phrases, starting from one punctuation segment, and the presence of the dependency with other phrases. To test. When a word different from the word displayed by the candidate word display means is selected in at least one of the clauses for which the establishment of the dependency relationship is verified by this dependency test, the dependency learning means makes this selection. The learned word is learned as dependency information. Therefore, it is possible to learn new dependency relationships.

In the kana-kanji conversion device according to claim 2, when the dependency learning means of claim 1 selects a word different from the word displayed by the candidate word display means, the word has a dependency relationship. The learning is carried out only when the judgment means judges whether or not the grammatical property is the same as the word when the establishment is tested, and the grammatical property is judged by the judging means. The gist is that it consists of learning execution means.

Therefore, in this kana-kanji conversion device, the learning of the new dependency has the same grammatical property between the word displayed by the candidate word display means and the word selected in place of this word. This is done only when, and there is no learning of bunsetsu relationships by words that have nothing to do with the originally tested dependencies.

In the kana-kanji conversion device according to claim 3, the determination means according to claim 2 stores, for a word that can form a clause having a dependency relation, a dictionary storing the category of the word and the category of both words. It is a gist to have means for judging that the grammatical properties are the same when the two are the same. Therefore, in this kana-kanji conversion device, it is possible to easily determine whether or not the grammatical properties are the same depending on the category assigned to the word. Information of this category can be held in the grammar dictionary.

The kana-kanji conversion device according to claim 4 is the device according to claim 1.
Dependency verification means of the dependency relation, the synonym + noun, the noun + noun, the noun + word, the adnominal form + noun, the noun +
The gist is that it is a means for testing at least one of the wording. Therefore, it is possible to test these dependencies.

In the kana-kanji conversion device according to claim 5, the dependency verification means according to claim 4 performs the dependency verification using a causative or passive relationship when analyzing the dependency pattern of the noun + phrase. A second means for learning the dependency relationship by returning the causative or passive relationship to the original form when the dependency means is verified by the first means. Use as a summary. Therefore, it is not necessary to separately store the dependency information or learn the dependency regarding the causative or passive relationship.

The kana-kanji conversion method according to claim 6 is a kana-kanji conversion method for inputting a kana character string and referring to a grammar dictionary to generate a kana-kanji mixed character string candidate. A string in which a character string is punctuated by referring to the grammar dictionary, the words constituting the bunsetsu are shown in a converted form in each of the punctuated bunsetsu, and information on the relationship between predetermined bunsetsus is stored. By referring to the dependency information dictionary, the existence of a dependency with another phrase is verified from one of the separated phrases as a starting point, and at least one of the phrases for which the establishment of the dependency relationship is verified by the dependency test. In one clause, when a word different from the displayed word is selected, the selected word is
The point is to learn as dependency information.

[0016]

Preferred embodiments of the present invention will be described below in order to further clarify the structure and operation of the present invention described above. FIG. 1 is a block diagram showing a control logic for kana-kanji conversion, and FIG. 2 is a block diagram showing hardware on which the kana-kanji conversion control logic actually operates. As shown in FIG. 2, this device is provided with the following parts centered around a well-known CPU 21 and mutually connected by a bus 31.
Each part mutually connected by the CPU 21 and the bus 31 will be briefly described.

ROM 22: Mask memory for storing kana-kanji conversion program, RAM 23: Readable and writable memory constituting main memory, keyboard interface 25: interface for controlling key input from keyboard 24, CRTC 27: display in color A CRT controller for controlling signal output to a possible CRT 26, a printer interface 29: an interface for controlling data output to the printer 28, a hard disk controller (HDC) 30; an interface for controlling the hard disk 32. The hard disk 32 stores various programs loaded and executed in the RAM 23, a kana-kanji conversion processing program provided in the form of a device driver, various conversion dictionaries referred to by the kana-kanji conversion processing program, and the like.

With the hardware thus configured,
Text is input, Kana-Kanji conversion, editing, display, printing, etc. That is, the character string input from the keyboard 24 is subjected to predetermined processing by the CPU 21, and the RAM 2
3 is stored in a predetermined area, and a CRT is sent via the CRTC 27.
It is displayed on the screen of 26.

Next, the function executed by the hardware thus constructed will be described with reference to FIG. The configuration and operation of each unit shown in FIG. 1 will be briefly described. The processing performed here is executed by the central processing unit (CPU 21) based on the data input from the keyboard 24. The CPU 21 performs all processing. Regarding kana-kanji conversion, when the keyboard 24 is operated, a predetermined interrupt process is activated, the input key image is converted into a corresponding kana-character string, and further this is converted into a kana-kanji mixed character string. Will start. Of course, if the computer can process in parallel,
Kana-Kanji conversion may be performed by one application (input method), and the conversion result may be delivered to the required application.
In this case, the input method collectively accepts the input from the keyboard 24.

The key image from the keyboard 24 is accepted by the character input unit 40, and converted into a corresponding kana character string here. When inputting romaji, it refers to a predetermined conversion table and converts into a kana character string. Every time one kana character is obtained, the character input unit 40 sends the kana character to the conversion control unit 42. This conversion control unit 42
Is a central part of kana-kanji conversion, controls various kana-kanji conversions described later, and sends the result to the converted character string output unit 44. Converted character string output unit 44
Actually sends a signal to the CRTC 27,
Display the converted string.

The conversion control unit 42 delivers the received kana characters to the character string input unit 50. The character string input unit 50 stores the kana character string in the character storage unit 52. Based on this character string, the independent word candidate creation unit 54 and the adjunct word candidate creation unit 64
And create word data candidates. The independent word candidate creation unit 54 uses the independent word dictionary 58 stored in advance in the hard disk 32, and under the management of the independent word analysis position management unit 56, performs processing for extracting independent word candidates from the obtained kana character string. To do. On the other hand, the adjunct word candidate creating unit 64 also performs the process of extracting an adjunct word candidate from the obtained kana character string under the control of the adjunct word analysis position managing unit 66 using the adjunct word dictionary 68. The process of extracting independent word candidates and adjunct word candidates while moving the analysis position will be described later.

Here, the independent word dictionary 58 is learned by learning.
Change the priority of homonyms and affixes. The learning process is performed by the dependency learning unit 70, the independent word learning unit 72, the auxiliary word learning unit 74, the affix learning unit 76, and the character conversion learning unit 78. When the user selects a word other than the word corresponding to the dependency under the condition that the dependency is satisfied, the dependency learning unit 70 gives priority to the combination selected by the user in the combination of the same words. This is to learn the relationship of dependency. The independent word learning unit 72 learns the word selected last in the independent word group in which homonyms exist as the highest priority candidate. The auxiliary word learning unit 74 learns in which word form an auxiliary word such as "please" is converted to "please" or "please". Further, the affix learning unit 76 is for learning conversion formats such as prefixes and suffixes (for example, "Go" and "Go"). The character conversion learning unit 78 learns the input character string as it is as hiragana or katakana and outputs the confirmed hiragana or katakana as a candidate in the subsequent conversion process. .

The independent word candidate creating unit 54 and the adjunct word candidate creating unit 64 obtain the created word candidates, and the word data creating unit 80 creates data for each word candidate. That is, as a result of the connection verification unit 82 referring to the connection verification table 84, the connection between the obtained independent words and adjunct words, independent words and independent words, and further the connection between the clauses consisting of “independent words + adjunct words”,
And the cost calculation unit 86 obtains the result of the cost calculation performed by the cost calculation unit 86 and outputs the result as data for each word.
This word data is once stored in the word data storage unit 100, receives the adjustment output from the dependency candidate adjustment unit 90, and is used for the phrase segmentation writing process.

The dependency candidate adjusting unit 90 receives word candidates from the independent word candidate creating unit 54 and the adjunct word candidate creating unit 64 via the word data creating unit 80, the word data storage unit 100, and the phrase segmentation writing unit 102. , To test the dependency. The modification check is performed by referring to the modification dictionary 98 prepared in advance on the hard disk 32. Note that the dependency dictionary stores only one dependency information even if the dependency relationship is reversed to reduce the capacity, and therefore the transposition information adjusting unit 99 is accompanied by grammatical analysis. Thus, the information in the dependency dictionary 98 is expanded to adjust the dependency candidates. For example, only the modification information of the modification word "Hana ga" + the modification word "beautiful" is stored in the modification dictionary 98,
He also tries to test the dependency word "beautiful" + the dependent word "flower."

The range in which the dependency test is performed is managed by the dependency range management unit 96. In addition, there are some admissible conditions in the verification of the dependency relationship, and these are determined by the causative / passive analysis unit 92, the particle admission analysis unit 94, and the like. The first bunsetsu segmentation candidate is determined from the bunsetsu candidates adjusted by the above dependency test, and stored in the bunsetsu data storage unit 106. The stored candidates are
The converted character string output unit 108 outputs the converted character string to the conversion control unit 42. The conversion control unit 42 displays this character string as a candidate character string, and since an undesired character string may be a candidate, the conversion control unit 42 receives instructions from the user and performs processing such as display and selection of the next candidate. To do. The results of these instructions and selections are input to the clause data storage unit 106 and the learning units 70 to 78 described above, and are used for partially fixing the clauses and rewriting priorities by learning. Although not shown, when the character string is confirmed by the user,
All the data temporarily stored in each unit are erased and prepared for the next conversion.

The process from the input of kana characters to the output of the conversion word character string has been outlined above, and the details of each process will be described below. First, a general phrase segmentation process will be described, and then a dependency process, which is an essential part of the present invention, will be described. FIG. 3 is a flow chart showing an outline of the processing of phrase segmentation by the minimum cost method. As shown in the figure, first, initialization processing such as erasing of temporarily stored data and initialization of the analysis position in the first digit (step S200) is performed, and then processing for obtaining the analysis position is performed ( Step S210). The analysis position is a position where the kana character string input so far is advanced one by one from the beginning. For example, if a kana character string “Kuru ma wa ko wo ko bu ko” is input as in the example sentence shown in FIG. 4, the first analysis position is the position of “KU” in the first digit. At this analysis position, a process of searching the independent word dictionary 58 and the auxiliary word dictionary 68 stored in the hard disk 32 is performed (step S220).

After the dictionary is searched, the word obtained is subjected to a process of checking the combination with the previous word (step S230), and when only the words which cannot be combined with each other are obtained, Further search the dictionary.
For example, in the example shown in FIG. 4, "ha" of "koohakobu"
The particle "ha" retrieved from the adjunct dictionary 68 for
Is determined to be impossible to combine with the immediately preceding case particle “o”, and is treated as invalid data by the connection test by the word data creation unit 80 and the connection verification unit 82. In FIG. 4, a word "x" is added to a word judged to be invalid by such a combination check. The connection between words is stored in advance in the connection verification table 84. This connection verification table 84 is a table that gives information about the possibility of combining the parts of speech of words. In the embodiment, 400 is used. It is given as a matrix of about × 400. When the dictionary search and the combination check at one analysis position are completed, the analysis positions are sequentially advanced and the processing is repeated.

For words that are likely to be combined, the cost is calculated next, and the minimum total cost of the words is calculated (step S240). This processing is performed by the cost calculation unit 86, and in the example shown in FIG.
"Car" can be divided into, for example, "ku" + "ru" + "ma", "car" + "ma" and "car". When applying words to these, independence word = 2, attached word It is assumed that the cost is = 0, and if "bitterness" (independent word) + "stream" (independent language), the total cost of "stream" is 4. At this time, the cost of "pause" is 4 because the minimum total cost is obtained, and "pain" + "flow" + "pause"
This is because the cost 4 in the case of “come” + “between” is adopted instead of the cost 6 of. Since "de" and "ha" are subordinate words, the word "car" with the lowest cost among the words before it =
The cost of 2 is its own cost. In FIG. 4, the cost of each word is shown in the lower right.

After the above cost calculation, the cost of each word is checked, and the process of invalidating the cost is performed (step S250). What is an inappropriate cost?
It is a combination of words that costs more than other combinations of words. That is, selecting a combination of words such as "ku" + "dome" is more expensive than other words "come" and "repeat" that can be obtained up to that position. Judgment is made and this is excluded from the phrase candidates. Words that are not adopted due to the idea of this minimum cost are shown as “●” in the upper right of the words in FIG. Note that, in FIG. 4, “◯” indicates that the word remains as a word that may form a phrase candidate as a result of the above-described combination check and cost check.

Next, a process of linking the word candidates thus given the cost is performed (step S260). That is, with respect to the words for which the combination is valid, the connection relation is related by setting a pointer. In the example of FIG. 4, "come", "roll", "car", "up", "de", "ha"
Since the minimum total cost was calculated as a word such as "in" which is not invalid, "up" is "" for "come" and "iteru".
The relationship is such that "car" is linked to "de" and "de" is linked to "car". The processes of the combination check, the cost calculation, and the linking are repeated while all the words are searched at one analysis position. Also, when the dictionary search at the analysis position is completed,
Further advance the analysis position, examine the establishment of a new word, and repeat the join check and cost calculation in the same way.

When the analysis position reaches the position of the last kana character that has already been input and the analysis has been completed for all words (step S265), the process of searching the path with the minimum cost is premised on the above process. (Step S27
0). This processing is performed by the phrase segmentation writing unit 102, and is a processing of searching for a combination of valid words that minimizes the total cost given to the words. In the case of "Walking until you come", as shown by the path of the solid line J in Fig. 4 (B), "By car" + "Box" +
The total cost is 18 for the "carry" segment.
Selected as the lowest cost. It should be noted that although the cost is not the minimum, other phrase segmentation candidates are also searched. For example, as shown by a path indicated by a broken line B in FIG. 4B, the segmentation is “by car” + “child” + “carry” (cost = 20). After creating the segmentation candidate in this way (step S280), this time, the process of creating the candidate inside each phrase is performed (step S290). That is,
Within a bunsetsu syllabary, candidates such as "box" and "box" for "hako" are prepared. These bunsetsu candidates and word candidates are used when the user instructs to change the way of bunsetsu division or to display the next candidate.

Next, the dependency verification process executed after the phrase segmentation process is performed and the kanji candidates are created for each phrase will be described. FIG. 5 is an explanatory view showing an example sentence in which kana-kanji conversion is performed, and FIG. 6 is a flowchart showing a dependency verification processing routine. Here, as an example sentence, as shown in FIG.
"I changed my car. ] Is taken up.
As a result of the segmentation process using the minimum cost method, FIG.
The phrase shown in (B) is extracted, and each phrase is shown in FIG. 5 (C).
It is assumed that the word shown in is displayed as the first candidate. Note that each word candidate is the first candidate learned so far. Regarding the dependency verification processing routine in this case,
It will be described with reference to FIG.

When the dependency verification process routine of FIG. 6 is started, first, a process of setting the phrase immediately before the end of the verification range (usually, all the separated phrases) as a dependent word is performed (step S400). ). In the embodiment, the dependency is searched for the corresponding dependent word by giving priority to the dependent word, so that the clause immediately before the end of the test range is set as the clause of the dependent word so as to have at least one dependent word. . Further, at this time, a process of initializing each variable of the process is also performed (for example, n is set to a value 1). Next, a process of setting a phrase n items behind the dependent word as a word is performed (step S410), and a process of searching the dependent dictionary 98 is performed (step S420). The dependent dictionary is recorded in the form of “received word root” + “dependent word”.

An example of the dictionary is shown in FIG. FIG. 7 (A) shows
It is a diagram schematically showing the contents of the dictionary in the case where there is a dependency of "quick and fast". This dictionary consists of the headwords "Kikiten" and the words "difficult" and "quick".
If the word is registered and the dependency is “to change the car +”, as shown in FIG. 7B, the reading of the entry word “kaeduma” and the words “replace” and “car” are used. The word is registered. In addition, attached word information for attached word allowance analysis described later is attached to the end. The actual dictionary is additionally provided with an index for searching, information indicating the word length, and the like.

As a result of searching the dependency dictionary from the word set as the dependent word and the word set as the dependent word, is there a headline (eg, "kaerumuma") created from the dependent word and the dependent word in the dictionary? It is determined whether or not (step S4
30) If the dependency is not found in the dictionary, then it is determined whether the end of the verification range has been verified (step S44).
0), if the end is not verified, the variable n is incremented by the value 1 (step S450), and step S4
The process is repeated from 10. Dependency is dependent dictionary 98
If it is found (step S430), it is determined whether the received word is already dependent (step S460). As shown in FIG. 8, when the dependency Q1 + R2 is once established, when the dependency is determined next for the preceding clause P, the processing that is different from the dependency Q1 + R2 already found is given priority. This is because.

If the dependency has not already been established for the received word side (that is, the word Q in the dependency determination of the word P and the word Q), the process proceeds to step S470 and thereafter.
First, the words that have a dependency are found with priority on the dependency. In order to search for a word that is dependent on priority depending on the dependent word, it is known from the determination in step S430 that there is at least one possible dependent word. The order of the related words is decided first. The state of this search is shown in FIG. When a plurality of words are found for the reading on the side of the dependent word, the first candidate X1 is fixed, and for the received word Y, according to the order of learning already arranged in the independent word dictionary 58, Y1 → Y2 →
It is checked whether the dependency is established in the order of Y3 → Y4 ... (Search by reference numeral A1 in FIG. 9). If a word satisfying the dependency is not found by this search, the next dependency word X2 is selected and the test is performed in the same manner (search for symbol A2).

When a combination of a dependent word and a dependent word that satisfies the modification read out from the modification dictionary 98 is found by such a search, the adjunct word is allowed to be analyzed (step S480). This processing will be described.

The admissibility analysis of particles is to determine whether or not the admissibility relation defined by the dependency type is satisfied, and has the following types for each dependency type. [I] Consecutive modifier noun + particle + particle in case of adjective Case particle “ga” “kara” “de” “to” “ni” “e” “yori” “o” “no” particle “ha” In the case of the noun + noun + noun + noun + noun (particle abbreviation type), the optional particles "ga", "ha", and adverbs [II] adnominal modifier noun + particle + noun particle In the case of "no" body language + body language (parallel) "ya""to" in the case of conjunctions + nouns In the case of conjunctions + nouns

That is, it is determined that the relationship is 2
If the relation between two words belongs to any of the above or any of the above, and if the adjunct word (mostly a particle or particle-like expression) that exists between the two words in the dependency relationship falls under any of the above, The dependency dictionary 98 is set with a particle that allows a word having a dependency relationship, and this is verified. For example, the dependency between "tact" and "taku" is the setting of the particle allowance (no ・ ga)
If it is accompanied by, it belongs to the above case (noun + particle + noun), so "no" and "ga" can exist between both words (quickly, witty → ○) , Other case particles such as "kara" and "de" are unacceptable.

With respect to each of the relationships (1) to (4), it is judged that the relationships other than those listed therein are allowed. The examples of judging that this is permissible are listed below, but these may include cases in which the dependency does not hold as an actual expression. However, dependency is a broad concept as an actual human language activity, and too strict dependency arrangements are often not practical. Further, too strict dependency arrangement only causes an unnecessarily large increase in the dependency dictionary 98, and also reduces the speed of dependency verification. Therefore, in this embodiment,
Regarding the admissibility of ancillary words, the relationship in which the dependency is generated is separated from that, and if there is a clear allowance / non-allowance, it is stored in the dependency dictionary along with the words that have the dependency relationship, and otherwise. With regard to

[III] Permissible expressions-in the case of continuous modified forms: noun + case particle expression + case particle expression in a noun "each""as""for""in""by", etc. Noun + particle + particle particle in adjectives “koso,” “sae,” “shika,” “but,” “mo,” etc. ・ Noun + auxiliary particle + auxiliary particle in adjectives “cut”, “about”, “gatsu”, “only” Etc. ・ Noun + Adverbial expression + Adverbial expression in adjectives "because", "Nara", etc. ・ Adjective + Particles + Particles in adjective "Noha" etc. ・ Connective particles "Node", "Kara""Kara","te", etc. ・ Connective particle-like expressions "kara ni""wa" and "ho"
"Ue", etc. ・ Expression in which verbs and verbs are arranged in parallel "ka""shi""tari"
"At the same time", etc.

[IV] Permissible expression-adjective modification form-Noun + particle expression + particle-like expression in nouns "in", "relating", "based on", etc. ・ Phrase + particle expression + particle in noun Expressions such as “for”, “to”, “to accompany”, “to”, “toki”, etc.

In accordance with the above rules, the admissibility of the adjunct word between the two words for which the dependency relationship is found is determined. In the case of "changing a car" and "changing", which are given as examples, the dependency is recognized. Therefore, this is determined (step S480), and when the dependency is established, the word for which the establishment is recognized is set as the first candidate of the dependent word and the dependent word in the independent word forming the clause. Perform (step S490). That is, the registration order for learning homonyms in the independent word dictionary 58 is changed. Further, the range from the dependent word to the dependent word thus found is registered as the dependent-established range, and processing for managing this is performed (step S500), and it is determined whether or not the dependent relationship is searched for in the entire range ( It proceeds to step S510).
If the dependency is denied due to an adjunct existing between the two words, the first candidate is not changed, and if there is other dependency information for the related word and the received word, the same change is performed. The test is performed (not shown), and if not, the process shifts to the judgment as to whether or not the whole range has been completed.

If the dependency test has not been completed for the entire range of the dependency test, that is, the entire dependency analysis range determined by the processing shown in FIG. 6, the variable n is initialized to the value 1 ( In step S520, the related word is further moved to the preceding clause (step S530), and the above-described processing (steps S410 to S500) is repeated until the modification verification of the entire range is completed.

If it is determined in step S460 that the dependent word is already dependent, it is determined whether the dependent word using the dependent word that is already dependent is established. Yes (step S540). That is,
In the example shown in FIG. 9A, when the dependency is established between the word Q1 and the word R2, when the dependency between the word P and the word Q is determined, The word Q1 is fixed and the dependency is determined. Therefore, in this case, even if the dependency P1 + Q2 exists, it is not adopted. However, the dependency P2 + with the received word as Q1
If Q1 is found, it is a dependency.
Therefore, as shown in FIG. 9B, P2 + Q1 + R2
The dependency is established. In the example sentence shown in FIG. 5, for “I changed the car”, which was defined as the verification range of the dependency, a dependency of “car” + “replaced” was found, and “car” “replaced” After learning as the first candidate, this dependency is not adopted because there is a dependency of "I am" + "replaced". On the other hand, if there is a dependency of "I" + "car", this is adopted.

Therefore, in this case, the first candidate of the related word (word P2 in the example) is determined (step S550). After that, similarly, the range of dependency is managed (step S50).
0) and whether the dependency check for the entire range is completed (step S510). When the dependency verification is completed for the entire range determined as the dependency verification range, the process goes to "END" and the present routine is terminated.

Next, regarding the modification learning routine,
This will be described with reference to FIG. Regarding the example sentence shown in FIG.
In the first three phrases, the words "I am", "car", and "replaced" are displayed as first candidates, and it is determined that "car" and "replaced" are established as dependencies. And When the dependency learning processing routine is started in this state, it is first determined whether or not the dependency learning flag is turned on (step S600). Whether or not the dependency learning is performed can be set, and this is determined by the flag. If the dependency learning is off, nothing is done and the process goes to "END" to end this routine. There are various kinds of learning in kana-kanji conversion. In the embodiment, independent word learning (a selected one of a plurality of independent word candidates is changed to the next kana-kanji conversion time) before the modification learning. Learning as the first candidate of the), learning of adjuncts (prefix "go""go" and classifier "times""kai"
Learning such as which one is prioritized) and bunsetsu learning (learning which one of the bunsetsu such as the length of the bunsetsu is prioritized) is performed. For these learnings, it is possible to set whether or not all learning is performed, and the flag is set to ON / OFF.
It is stored in the off state.

If the dependency learning flag is set to ON, the phrase in the next selected state, that is, the phrase currently being focused on for dependency learning during kana-kanji conversion is the dependency phrase. It is determined whether or not there is (step S610). Since it is determined and stored by the processing routine shown in FIG. 6 whether or not it is a dependent phrase, it is determined based on this information. If it is not a related word bunsetsu, nothing is performed and the process ends in "END", and if it is determined to be a related word bunsetsu, a process of searching for a phrase of the received word is performed (step S620). It is easy to find the dependent word corresponding to the dependent word from the dependent information. If the word on which the cursor under conversion is placed is "car", this word is determined to be a related word, and "substituted" is found as a received word.

Next, it is judged whether or not the next candidate is selected in the related word phrase (step S630). If the next candidate is not selected in the related word phrase, the next candidate is selected in the received word phrase. It is determined whether or not is selected (step S640). In these processes, in the phrase of the dependent word or the dependent word, for example, “car” is selected instead of “car” displayed as the first candidate because the dependent is established (the next candidate in the dependent word is selected). Selection) corresponds to the case where “changed” is selected instead of “changed” (selection of the next candidate in the received word).

In these cases, processing for creating dependency information is then performed (step S650). The creation of dependency information will be described later with reference to FIG. After creating the dependency information, the dependency dictionary 9 is used by using this information.
8 is registered (step S660), and then "E"
ND "is reached and this routine ends. If the next candidate is not selected in the phrase of the dependent word or the dependent word, there is no particular need to learn the dependent word, so nothing is done and the process ends and the routine ends. .

When the dependency learning processing routine of FIG. 10 explained above is executed, the dependency learning flag is ON, and the following dependency is detected in any of the dependency phrase and the dependency phrase for which the dependency is recognized. When a candidate is selected, dependency information is created and registered in the dependency dictionary 98.

Next, details of the modification information creation processing routine (step S650 in FIG. 10) will be described using the flowchart in FIG. In the modification information creation process, first, the modification word for which the next candidate is selected and the part-of-speech type of the modification word are checked (step S700), and it is determined whether or not the modification part is part-of-speech (step S7).
05). This is because the dependency was established when the first candidate was displayed, but it may be out of the dependency depending on the selection of the next candidate. For example, if the dependency "Kawano" + "Flow" is registered in the dictionary and its success is verified, "Kawano" (proprietary) instead of "Kawano" (noun + particle "no") When a noun, a family name) is selected, the part of speech is different, and usually the part of speech is not dependent on the corresponding word. If it is determined that the part of speech to be subjected to the dependency is no longer present, nothing is done and the process goes to "END" to end this processing routine.

When it is determined that the selected next candidate is the part of speech for which the modification is performed, the relationship between the modification word for which the modification is established next and the clause of the modification word may have a modification relationship. It is determined which of the following five types prepared as above belongs to. That is, "hymn" + "definition"
(Step S715), the "verb" +
Judgment as to whether it is a "defective" (step S720),
Judgment as to whether or not it is "hymn" + "hymn" (step S7)
25), "unified form" + "syllable" determination (step S730), and "union" + "dynamic" determination (step S735). If all of these determinations are "NO", it is determined that the combination is not a combination for performing dependency learning, the process ends in "END", and the present routine ends.

On the other hand, the related word and the phrase of the receiving word are "syntax"
If it is determined that the relation is ++ "defective" (step S715), then processing for analyzing the details of the dependency pattern of "predictive" + "defective" is performed (step S740). . The details of this analysis processing are shown in FIG.
It is a flowchart of FIG. There are various patterns in the dependency of "hymn word" + "defective word", so this should be analyzed in detail. This analysis processing will be described with reference to FIG.

When the pattern analysis processing routine of "syntax" + "synonym" is started, it is first judged whether or not there is a particle between "synomony" + "synonym" (step S80).
0) If it is determined that it does not exist, it is determined that an optional particle (for example, the case particle “ga”) is omitted, and this can be registered as a dependency in the dependency dictionary 98. It is determined (step S805). On the other hand, when it is determined that a particle is present between the “syntax” and the “defective” (step S800), the particle is “no” or “ya”.
It is determined whether or not (step S810). Normally, there is no possibility that there is a particle "no" or "ya" between the "composition" and the "verb" that make up the dependency, so it is judged that the dependency is established and the first word of each phrase Although the candidate was displayed, it is considered that the next candidate was selected because the judgment that the dependency was established was incorrect. Therefore, in this case, nothing is done and the process goes to "END" to end this routine.

On the other hand, when it is determined that there is a particle other than "no" or "ya" between "hymn" and "hymn" (step S810), "hymn" + "hymn" is used. It is determined whether or not there is the relationship (step S815). Causative and is a clause of the nominal "person (or thing) is",
And "What do" or "to the person (or thing)",
Since it has the property that it is composed of the phrase of the idiom, "make it XX," it is judged from the ending type of the idiom. For example, there can be a sentence such as “a parent bird feeds its chicks with food”. In this case, it is found that the causative dependency in the form of “hymn” + “defective” is “feed” + “feed”. "And" Hina "+" Feed ". When it is determined that the character is a causative character (step S815), it is further determined whether or not the particle located between the "syntax" and the "verb" is "wo" or "ni" (step S820). ). "Word of mouth"
＋ The particle located between "defectives" is "wo" or "ni"
If so, it is judged to be the role of the causative agent, and “hymn” +
In the form of "particles" + "defectives" (step S825),
It is possible to newly register as a dependency, step S805.
Move to

If it is determined in step S815 that the player is not a player, then it is determined whether or not the player is passive (passive) (step S830). The passive, "a person (or thing) is" or "to the person (or thing)" +
It is a syntax consisting of "verb passive type (○○ reru, uru)", such as "to him" + "rejected" (active type is "he" + "reject") or "door" + "open". It is a syntax such as "is" (active type is "open door" + "open"). If the clauses that create dependency information are in the “hymn” + “definitive” relationship and are judged to be passive, then “hymn”
It is determined whether or not the particle which exists between + and the "verb" is "ga" (step S835). This judgment is "YE
If it is "S", the particle between "hymn" and "definitive" is "wo"
(Step S840), the process proceeds to step S805, and the dependency can be registered.

"Passive" + "defective" even in the passive case
If the particle between is not "ga" (step S83
5) Next, it is determined whether or not the particle between "syntax" and "defective" is "ni" (step S845). If the passive particle is determined to be "ni", then "hymn" +
Change the particle between "defectives" to "ga" (step S8
50), the process proceeds to step S805, and the dependency can be registered.

If it is neither passive nor passive, and the interpositional particle is neither "ga" nor "ni", then it is a normal "body word" +
It is a dependency of a "verb" or a causative / passive that does not take a specific particle as a causative / passive (for example, "place" + "de").
+ "Verb causative or passive type"), the process proceeds to step S805, and the dependency can be registered, and then this routine ends. Through the above processing, the dependency pattern of “syntax” + “defective” is analyzed, and if a specific condition is satisfied, registration is possible.

Returning to FIG. 11, the description of the dependency information creation processing routine will be continued. In step S740, after the pattern analysis of the dependency of “syntax” + “defective” is performed, it is determined whether or not the combination of clauses to be analyzed can be registered (step S745). When it is determined by the pattern analysis shown in FIG. 12 that a new dependency can be registered, it is determined that registration is possible and the process proceeds to step S780 to generate dependency information. That is, for example, when it is possible to register as passive,
In the example of "feeding" + "feeding", the reading is "eating", the related word "bait", the receiving word "eat", the adjunct word "ga".
Create the information. In addition, for example, when it is possible to register as passive, if the "door" + "open", the information for registration is created as "active door" + "open". In this example, the reading "Aketobira", the adjective "Open", the related word "Door", the adjunct "Wo"
Information is created. Further, if "to him" + "will be refused", active information "he will" + "will be refused" will be used to create information for registration. In this example, information such as the reading “kotoware”, the received word “decline”, the related word “he”, and the adjunct word “ga” is created.

Next, the relationship between the dependent word and the clause of the dependent word, which is determined to have the dependent relationship, can be selected from among the five types prepared as the possible dependent relationship. A process performed when it is determined that the case belongs to any one of the cases will be described. When it is determined that “synonym” + “synonym” (step S720), it is determined whether or not the related word is a related form (step S750). (S755), the process proceeds to step S780, and dependency information is created. Even if it is determined that it is “synonym” + “synonym”, if it is determined that the related word is not the plural form, “END” is not performed.
To exit this routine.

If it is determined that "syntax" + "syntax" (step S725), "syntax" + "syntax".
It is determined whether or not there is a particle "no", "to", or "ya" in between (step S760). When these particles are put between "syntax" and "syntax", it can be regarded as showing a parallel relationship (A and B, A or B) or a limited relationship (A and B). Similarly, dependency information regarding the relationship is created (step S780). On the other hand, when these particles are not included between the "syntax" and the "syntax", nothing is done and the process goes to "END" to end this routine.

Similarly, when it is judged that the expression is "union form" + "noun" (step S730), "union word" +
When it is determined that the word is a "word of speech" (step S73)
5) As the relationship of dependency is established, dependency information is created as a relation modification type (step S780). “Union form” + “composition” is a case where the related word is a noun conjunction form and a compound word is found as a word, for example,
Dependency of "union form" + "hymn""Beautiful" + "Flower" when the next candidate "Dark" is selected "Dark" + "Flower" and dependency "Running" + "Vehicle" For example, “run” + “vehicle” when the candidate “vehicle” is selected. In addition, “adjunct” + “give” is, for example, “aru” + “woman” or “that” + “person”, and “aru” when the next candidate “aru” is selected. +
For example, "woman". In these cases, dependency information is created as the adnominal modification type. In other words, as the dependency information, the reading "Ukuhana", the receiving word "beautiful", the reading word "flower", the reading "Hashishayo", the receiving word "run", and the moving word "vehicle" are created. Be done. As a result, in the case of "union form" + "noun", "vehicle" +
In the form of "run" + particle "ga", that is, "body language" +
Dependency information is created and stored in the same format as the modification information in the case of a "verb". The dependency dictionary 98 is unified so that the dependency of the synonym and the idiom is the idiom and the idiom.

When the processing shown in FIGS. 10 to 12 described above is performed after the input kana character string has been segmented and written, it is determined that the dependency is established, and the first
When the second and subsequent candidates of the word are selected while the candidate word is selected and displayed, new dependency information is created for the selected next candidate (step S
780), which is registered in the dependency dictionary 98. Therefore, it is not necessary to enlarge the dependency dictionary to 98,
In addition, it is possible to increase the possibility that a desired character string can be obtained by Kana-Kanji conversion by storing a dependency pattern according to the habit and usage of a person who uses this Kana-Kanji conversion device. In addition, causatives and passives can be applied to any syntax because they are learned by returning to the original form (active form in this example).

In the present embodiment, the dependency test is
It starts from a position close to the end of the sentence within the range of the dependency verification range and gives priority to the dependency word. This configuration was extremely effective in making the determination of word candidates by modification more desirable. This is because in Japanese, the predicate side at the end of a sentence often carries the meaning of the sentence.
It is considered that the subject's actions often change rather than the same subject (generally described by the predicate on the sentence end side) and the same subject (generally the sentence description on the sentence side) changes.

When it is determined that the dependency is established, the range from the received word to the dependency is managed as the dependency establishment range, so that the dependency range does not intersect. Moreover, it is not judged that two or more dialects receive one related word. Moreover, since the establishment of the dependency is judged across the adjacent clauses (when n ≧ 2), the dependency test can be correctly performed even when the modifier such as an adverb enters between the dependencies. Therefore, when a plurality of dependencies are established, FIG. 13 (A)
As shown in FIG. 13, a combination in which independent dependencies are separately established, or as shown in FIG.
A combination that receives the above-mentioned dependency words or a combination that allows another dependency to be established so as to straddle one dependency, as shown in FIG. 13C, is permitted.

Next, a second embodiment of the present invention will be described.
In the second embodiment, first, the processing of FIG. 6 in the first embodiment,
That is, in the dependency verification processing routine, when the dependency dictionary 98 is searched (step S420) to determine whether or not there is a dependency (step S430), a specific dependency relationship is set in the dependency dictionary 98. Instead of storing it, as shown in FIG. 14, the information of the meaning classification part-of-speech is stored in the normal independent word dictionary 58 or the dependency dictionary 98, and the dependency is determined using this. . For example, in the first embodiment, “to section manager” + “say”, “to manager” + “say”, “to president” + “say”, etc. are stored in the dependency dictionary 98 as individual dependency information. Aside
The success of the dependency was verified individually. On the contrary,
In the second example, a new part-of-speech called the meaning class-of-speech is added to the independent word dictionary such as “section manager”, “department manager”, “president” in addition to the normal part-of-speech information. The meaning part of speech "position" indicating "/ position" is added. In addition, the dependency dictionary 98 stores information such as “composition (position)” + “say” + particle “ni” instead of individual dependency information. In this example, the reading is
4 ”, the receiving word is“ say ”, and the related word is“ status ”.

After segmenting the bunsetsu, the grammatical information of each bunsetsu is examined to read the meaning part-of-speech of the independent word forming the bunsetsu. Next, when performing the dependency test (Fig. 6),
The dependency dictionary is searched using this meaning part-of-speech (step S430). In the dependency dictionary 98, a dependency corresponding to “status” + “say” is found, so in this case, the dependency is established, and this is displayed as the first candidate of the word forming the phrase. In addition, when the next candidate of the word is selected in this state, whether the next candidate is a word to which the same semantic classification as the first candidate can be given is determined in the dependency information creation process (step S700 in FIG. 11). to decide. For example, if the second candidate has the same normal part of speech, it is determined that the same semantic classification as that of the first candidate can be given. If the first candidate is the "manager", then the "chief" or "head" from the next candidates
When is selected, the candidate parts of speech are the same as "common nouns", and the meaning parts of speech of "general chief" and "family head" are newly set as "position".

According to this embodiment, it is not necessary to individually store the dependency relation, and the dependency dictionary can be made extremely small. In addition, when a new dependency relation is added, it is only necessary to add the meaning part of speech of the word, so there is no need to modify the dependency dictionary. Therefore, the learning of dependency when the next candidate of the word candidates of the phrase once displayed is selected only by rewriting the meaning part of speech in the independent word dictionary 58. It should be noted that when the dependency is tested, only the dependency using the semantic part-of-speech may be the target, but both the individual dependency described in the first embodiment and the dependency using the semantic part-of-speech are tested. It doesn't matter.

The embodiments of the present invention have been described above.
The present invention is not limited to such an embodiment. For example, instead of the minimum cost method, a configuration using another bunsetsu segmentation method such as the two-bunsetsu longest matching method, and the dependent word priority or the dependent word in the modification check can be used. The present invention can be implemented in various modes within the scope not departing from the gist of the present invention, such as a configuration in which priority can be switched, a configuration in which the relationship of dependency is limited to one or more such as “hymn” + “verb”. Of course.

[0071]

As described above, in the kana-kanji conversion device according to claim 1 and the kana-kanji conversion method according to claim 6 of the present invention, the input kana character string is segmented into phrases by referring to the grammar dictionary, In each punctuation phrase, the words that make up the punctuation phrase are shown in a converted form. In this state,
The existence of a dependency with another bunsetsu is verified by referring to a dependency information dictionary in which the information of the dependency of a predetermined bunsetsu is stored, with one punctuation written as a starting point. When a word different from the word displayed by the candidate word display means is selected in at least one of the clauses for which the establishment of the dependency relationship has been verified by this dependency test,
The selected word is learned as dependency information.
Therefore, there is an excellent effect that it is possible to learn a new dependency relationship.

In the kana-kanji conversion device according to the second aspect, the learning of the new dependency has the same grammatical property as the word displayed by the candidate word display means and the word selected in place of this word. Since it is carried out only when, the effect of not learning the relation of the bunsetsu, which is not related to the tested dependency, as the dependency.

The kana-kanji conversion device according to claim 3 stores the category of a word that can form a phrase having a dependency relation, and when the category of the word is the same, the grammatical property is the same. Since it has a means to judge that they are the same,
It is possible to easily judge whether or not the grammatical properties are the same depending on the category given to the word. Information of this category can be held in the grammar dictionary.

Since the kana-kanji conversion device according to claim 4 verifies the dependency relation for at least one of the synonym + synonym, the synonym + synonym, the synonym + synamism, the adnominal form + synomonic, and the adjunct + synomical. , It is possible to reliably test the basic dependency relationship.

In the kana-kanji conversion device according to the fifth aspect, when the dependency pattern of the phrase + noun is analyzed, the dependency check using the causative or passive relation is performed, and the success of the dependency is verified. Since the causative or passive relationship is returned to the original form and the dependency relationship is learned, there is an excellent effect that it is not necessary to separately store the dependency information about the causative or passive relationship or to learn the dependency.

[Brief description of drawings]

FIG. 1 is a functional block diagram showing an implementation form of a kana-kanji conversion function in a kana-kanji conversion device according to an embodiment of the present invention.

FIG. 2 is a block diagram showing hardware for realizing a kana-kanji conversion device as an example.

FIG. 3 is a flowchart showing a phrase segmentation writing process executed in a segmentation segment writing unit 102.

FIG. 4 is an explanatory diagram showing an example of phrase segmentation according to the minimum cost method in the embodiment.

FIG. 5 is an explanatory diagram showing an example sentence that is a target of the dependency check in the embodiment.

FIG. 6 is a flowchart showing a dependency verification processing routine in the embodiment.

FIG. 7 is an explanatory diagram showing an example of a dependency dictionary according to the embodiment.

FIG. 8 is an explanatory diagram showing how dependency is established in a plurality of clauses.

FIG. 9 is an explanatory diagram showing a priority order in a modification check.

FIG. 10 is a flowchart showing a dependency learning processing routine.

FIG. 11 is a flowchart showing a modification information creation processing routine.

FIG. 12 is a flowchart showing a dependency pattern analysis processing routine of “syntax” + “defective”.

FIG. 13 is an explanatory diagram showing a relation of clauses for which dependency is established.

FIG. 14 is an explanatory diagram showing an example of a dictionary including meaning parts of speech according to the second embodiment.

[Explanation of symbols]

 21 ... CPU 22 ... ROM 23 ... RAM 24 ... Keyboard 25 ... Keyboard interface 26 ... CRT 27 ... CRTC 28 ... Printer 29 ... Printer interface 30 ... Hard disk controller (HDC) 31 ... Bus 32 ... Hard disk 40 ... Character input section 42 ... Conversion Control unit 44 ... Converted character string output unit 50 ... Character string input unit 52 ... Character storage unit 54 ... Independent word candidate creation unit 56 ... Independent word analysis position management unit 58 ... Independent word dictionary 64 ... Adjunct word creation unit 66 ... Adjunct analysis position management unit 68 ... Adjunct dictionary 70 ... Dependency learning unit 70 ... Learning unit 72 ... Independent word learning unit 74 ... Auxiliary word learning unit 76 ... Affix learning unit 78 ... Character conversion learning unit 80 ... Word data creation unit 82 ... Connection verification unit 84 ... Connection verification table 86 ... Cost calculation unit 90 ... Dependency candidate adjustment 92 ... Passive analysis unit 94 ... Particle admissible analysis unit 96 ... Dependency range management unit 98 ... Dependency dictionary 99 ... Dependency transposition information adjustment unit 100 ... Word data storage unit 102 ... Phrase segmentation writing unit 104 ... Dependency transposition information adjustment unit 106 ... clause data storage section 108 ... converted character string output section

Claims (6)

[Claims] 1. A kana-kanji conversion device for inputting a kana character string and referring to a grammar dictionary to generate a kana-kanji mixed character string candidate, wherein the inputted kana character string is referred to the grammar dictionary. And a punctuation means for puncturing each bunsetsu, a candidate word display means for showing the words constituting each bunsetsu in a converted form in each punctuated bunsetsu, and a dependency storing information on the dependency of a predetermined bunsetsu Starting from the information dictionary and one of the separated phrases, referring to the dependency information dictionary, the existence of a dependency with another phrase is detected.
When a word different from the word displayed by the candidate word display means is selected in at least one of the dependency verification means for testing and the clause for which the establishment of the modification relationship is verified by the modification test A kana-kanji conversion device comprising a dependency learning means for learning the selected word as dependency information. 2. The kana-kanji conversion device according to claim 1, wherein the dependency learning means selects a word different from the word displayed by the candidate word display means, and the dependency relationship of the word is changed. The judgment means for judging whether or not the grammatical property is the same as the word when the establishment of is verified, and the learning is performed only when the deciding means judges that the grammatical property is the same. A kana-kanji conversion device consisting of learning and performing means. 3. The kana-kanji conversion device according to claim 2, wherein the determination unit stores a category of words that can form a clause having a dependency relationship, and the both words. Kana-to-Kanji conversion device provided with means for judging that the grammatical properties are the same when the categories are the same. 4. The modification check means is a means for testing the relationship of modification for at least one of a synonym + synonym, a synonym + synonym, a synonym + hymn, an adnominal form + hymn, and a conjunction + hymn. The kana-kanji conversion device according to claim 1. 5. The kana-kanji conversion device according to claim 4, wherein the dependency check means performs a dependency check using a causative or passive relationship at the time of analyzing a dependency pattern of a noun + a noun. The kana including the first means for performing, and the learning means, when the first means determines that the dependency is established, restores the causative or passive relationship to the original form and learns the dependency. Kanji conversion device. 6. A kana-kanji conversion method for inputting a kana character string and referring to a grammar dictionary to generate a kana-kanji mixed character string candidate, wherein the inputted kana character string is referred to the grammar dictionary. Then, the words constituting the bunsetsu are shown in a converted form in each of the punctured bunsetsus, and the syllabus is referred to by referring to the modification information dictionary that stores the modification information of the predetermined syllabaries. The presence of a dependency with another bunsetsu is verified from one of the bunsetsus that have been set as the starting point, and at least one of the bunsetsus for which the establishment of the dependency relationship is verified by the verification of the bunsetsu, is displayed as described above. A kana-kanji conversion method in which when a word different from a word is selected, the selected word is learned as dependency information.