JPH0469766A - Kana to kanji converter - Google Patents

Abstract

PURPOSE:To improve the conversion speed by fetching a work from a divide-writing part at every character of an inputted character-string, discriminating the propriety of a connection to an immediately previous word in the input character-string with regard to the maximum likelihood word and executing a back-track with regard to the immediately previous word in the input character-string in the case it is not appropriate and selecting the maximum likelihood word. CONSTITUTION:A character is fetched by one character each by a character fetching means 101 from an inputted character-string, and stored in a read buffer 105. Subsequently, in a maximum likelihood evaluating and selecting means 103, while sorting a word which can be connected to the immediately previous maximum likelihood candidate word in order of frequency with regard to each word of a dictionary buffer 106, it is stored in a candidate word buffer 107, and the most superordinate word of sorting order is selected and stored in a maximum likelihood work buffer 108. Thereafter, these operations are executed with respect to the word corresponding to the next sentence clause, and when the word of the next sentence clause is found out, the present maximum likelihood candidate is written in a defined word buffer 109. In such a way, to output data for conversion to a KANJI (Chinese character) and KANA (Japanese syllabary) mixed sentence, an access is executed only once, therefore, the processing efficiency is improved.

Description

[Detailed description of the invention] [Summary 1] Regarding a kana-kanji conversion device that converts a human-written character string of kana characters into a sentence containing kanji and kana, the data length of the words handled is shortened and the buffer capacity is reduced even when pack tracking is performed. The purpose of this dictionary is to provide a kanji conversion device that can improve the conversion speed at the same time, and is composed of a data section containing pronunciations and parts of speech used only for retrieval, and an output data section containing pronunciations, kanji, etc. A word is extracted from the parting data section of the dictionary for each character of the input string, and candidate words are extracted from the retrieved words by maximum likelihood selection. Determine whether connection is possible or not, and if yes, perform the operation on the subsequent character string, and if not, select the most likely word by 8-trunking the previous word in the human character string using the above separation data. Then, the selected maximum likelihood word is converted into a sentence containing kanji and kana using the output data of the dictionary and is output.

[Industrial Application Field 1] The present invention relates to a kana-kanji conversion device for converting a human-written character string of kana characters into a sentence containing kanji and kana.

In systems such as word processors and workstations that perform Japanese language processing, there is a mechanism that automatically divides character strings spanning multiple phrases into phrases or words at the reception desk using kana manual input (solid input) and converts them into sentences containing kanji and kana. It is provided.

Conventional devices that perform such kana-kanji conversion use the ``yomiji J
The words corresponding to are stored. In this case, the word data includes data for outputting kanji, so when selecting the most suitable word candidate (maximum likelihood candidate), it is important to note that the word once selected is inappropriate. When it is determined by checking the connections, the process returns to binary and converts the more appropriate clause (character string) into two kana-kanji (this is called bank I racking).

It is desired that such kana-kanji conversion including eight-track ranking be realized at high speed and with as little memory capacity as possible.

[Prior Art] FIG. 6 is a block diagram of a conventional example, and FIG. 7 is a conversion processing flow diagram of the conventional example.

In FIG. 6, 61 is the keyboard 1", 62 is the file, 63 is the input 3h'J buffer, 64 is the reading buffer for dictionary access, 65 is the dictionary, 66 is the dictionary data buffer, 67 is the candidate word buffer, and 68 is the most Likelihood candidate buffer, 69 confirmed word buffer.

70 is a display device such as a CRT.

As shown in Figure 2, the dictionary 65 mentioned above has "
``Reading length'' (length of reading character string) 2 ``Reading'' (reading content), ``data length'', ``output data'' (kanji data), ``part of speech'', and ``frequency'' are set for each word. If there are multiple words stored with the same reading (homophones),
Corresponding to each word, "data length J + r output data (
Kanji data).

"Part of speech" and "frequency" are stored sequentially.

The operation of the configuration shown in FIG. 6 will be explained in terms of sections.

(2) When a character string (reading character) to be converted is input from the keyboard F61 or file 62 in FIG. 6, the data of the character string is stored in the input reading buffer 63.

■Reading buffer 6 for accessing the dictionary one character at a time
Take it out at 4.

(2) Search the dictionary 65 and read the searched word into the dictionary data buffer 66 each time the reading of one character is taken out to the reading buffer 64 for dictionary access.

(2) Write words that can be connected to the most likely candidate word of the immediately preceding clause into the candidate word buffer 67 while sorting them in order of frequency. In addition, for the maximum likelihood candidates in this case, all words in the same clause are sorted, giving priority to those with the longest reading length and giving priority to those with the highest frequency in the 5th order.

■The word with the highest rank as a result of the sorting is written in the 1st most likely candidate field 68.

When the maximum likelihood word of the 0th order clause is found, it is written to the confirmed word buffer 69.

However, if the input reading data for ■ is completed by repeating the process of extracting the fifth clause, ■, ■, ■, the word following the current maximum likelihood candidate is selected from the candidate word buffer Write to the buffer 68 (■ above). At this time, if there is no candidate word for the same clause, the processing clause is returned to the previous clause, the last word in the two-candidate word buffer is returned to the maximum likelihood candidate buffer 68, and the most likely candidate word in the candidate word buffer 67 is returned. The word following the most likely candidate word is written to the most likely candidate buffer 68, and the word is continued to be searched for (■, ■, ■). Such processing is generally called back 1 to racking.

■Once all input readings have been converted, the confirmed word buffer 6
The data for output is extracted from 9 and output as a mixed sentence of kanji and kana.

Next, the flow of the conversion process in the above conventional example will be explained using FIG. 7.

The input pronunciation (kana input) is extracted and if there is no manual reading data, it is converted to a sentence containing kanji and kana (for sentences for which the conversion process has been completed before then), and if there is, the word that matches the two pronunciations is added to the dictionary. (65 in FIG. 6) is read into memory (70 to 72 in FIG. 7). In this case, dictionary data including output data corresponding to each reading is read from the dictionary (65 in FIG. 6).

As a result, if there is a loaded word, the most likely word (''
2 (already mentioned in relation to Figure 6), and when there is a maximum likelihood word, the first step is to determine the connectability of the parts of speech (whether or not the combination of the parts of speech of the previous word and the subsequent word is possible). 73-76)
. If a connection is possible as a result, preprocessing is performed to start processing from the reading of the clause following the most likely word (the "yomi" after the most likely word is used when extracting the next input reading), and ■
The process moves to step 70 (step 77).

Further, if the possibility of connection based on one part of speech is checked and it is found that the connection is not possible, the process moves to step 74 according to route 2, and the most likely word of the next rank is extracted.

If there is no word in the dictionary that matches the input pronunciation, or if there is no maximum likelihood word, it is determined whether there is a previous clause, and if there is, it performs preprocessing to backtrunk and returns to processing the previous clause. 78.79).

If it is found in step 78 that there is no preceding phrase, it is determined that conversion is impossible and all input readings are converted to full-width hiragana and the process ends (step 80).

Finally, if there is no input reading (end of clause), the kanji code included in the selected most likely word is used to convert it into a sentence containing kanji and kana, and the process ends.

As mentioned above, in conventional Japanese conversion, bank trunking is repeated to find the longest pattern that can form a clause, and the conversion to a sentence containing kanji and kana is completed.

[Problem to be solved by the invention] According to the conventional method described above, there is a first-order problem.

■Since the data length of one word is large, when a word that matches the pronunciation is extracted, the candidate word buffer (67 in Figure 6)
The amount of data stored in the memory increases, putting pressure on memory capacity. For this reason, there are memory constraints and a sufficiently large "
When candidate words cannot be stored in the "candidate word buffer", a situation occurs in which all candidate words cannot be stored in the buffer. At this time, if backdragging occurs, it becomes necessary to reread the data that overflows from the memory, which increases the number of file accesses and reduces speed.

(2) If there are many words stored in the dictionary file, the number of words to be read as candidate data will increase, so the longer the data length of one word, the slower the processing speed will be.

■If there are restrictions on the size of the dictionary file, it will be necessary to reduce the number of words, and if the number of words is reduced, cases will likely occur where conversion will not be possible.

SUMMARY OF THE INVENTION An object of the present invention is to provide a Kana-Kanji conversion device that can shorten the data length of words to be handled even when bank tracking is performed, thereby reducing buffer capacity and improving conversion speed.

[Means to solve the problem] FIG. 1 is a diagram showing the basic configuration of the present invention.

In FIG. 1, 10 represents a processing device 311 including a CPU and memory, a dictionary, 12 an input section, and 13 an output section.

The processing device 10 includes a single character extraction means 101, a data retrieval means 102 for dividing data. Maximum likelihood word evaluation selection means 103
．． Each means of the output data search means 104 and a plurality of buffers 105 to 109 are provided. '', and an output data section 112 in which data and frequencies of kanji and kana are stored.

The present invention separately stores two parts in a dictionary: a parting data section containing parts of speech corresponding to the readings, and an output data section containing kana and kanji information corresponding to the readings, and selects the most likely candidate word. At the time of writing, the parting data is used, and after the word is determined, the output data section is used to perform kanji conversion.

[Operation] First, a character string to be converted is input to the processing device 10 from an input unit such as a keyboard or a file. Characters are extracted one by one by character extraction means 101 from the input character string and stored in reading buffer 105 . Every time one character is retrieved, the separation data search means 1.02 searches the separation data section of the dictionary 11 using the contents of the reading buffer 105 as a search key. The separation data (often there is a large number of pieces) obtained by the search is stored in the dictionary buffer 106.

Next, for each word in the dictionary buffer 106, the maximum likelihood word evaluation selection means 103 sorts words that can be connected to the previous maximum likelihood candidate word (determined by part of speech) in order of frequency (with priority given to words with long pronunciations). Candidate word buffer 1
After reading the words of the same clause in the first stage, select the word with the highest rank in the sort order and store it in the maximum likelihood word buffer] and 08. After the operation, each means IO1-103 performs an operation on the word corresponding to the tertiary clause, and when the word of the tertiary clause is found, the current maximum likelihood candidate is written into the confirmed word buffer 109.

This operation is repeated to write the confirmed word to the confirmed word buffer. If there is no word in the dictionary that connects to the first clause, backtracking is performed and the same process is performed going back to the previous word.

When processing is completed for all input readings, the output data 1.12 of the dictionary 11 is read out from the contents of the confirmed word buffer 109, and output data of sentences containing kanji and kana is extracted and outputted to the output unit 13.

In the operation described in F, the data handled by each means 102, 103 and each buffer 1.05 to 109 is short-length separation writing data, so even if hack tracking occurs, the amount of data to be processed is small. Since the output data for conversion into a kanji-kana mixed sentence is accessed only once, processing efficiency can be improved.

2 [Example] Fig. 2 is a processing flow diagram of an embodiment of the present invention, Fig. 3 is a configuration example of dictionary data according to the present invention, Fig. 4 is a table showing whether or not words can be connected according to parts of speech, Fig. 5 The figure shows an example of operation according to the present invention.

First, the structure of dictionary data according to the present invention will be explained using FIG.

A in FIG. 3 shows the data structure of the 1-minute writing data section (same as the dividing writing data section in FIG. 1), and each "
For "Yomi", the first is the total data length (the length of all data corresponding to the same reading; displayed in DL), the first is the reading length (the length of the reading data is displayed as Y1.), and the second part of speech is "Yomi". and frequency are stored, and if there are other words of the same "yomi" after this, their part of speech and frequency are stored in order.

Figure 3B shows the data structure of the output data section (same as the output data section in Figure 1), and for each "reading", first the data length (DL), reading length (Yil and The thing is,
It is stored in the same way as in Figure 3A, and then the number of words in the same part of speech (indicated by IIc) and the output data length (indicated by KL) are stored, and then the kanji data and their frequencies are set &J. There is. In this case as well, if there are multiple kanji for the same reading, the data length is 1 and the reading length is 5 for each funeral word.
The reading output data length, kanji data, and frequency are stored.

C9 and C9 in FIG. 3 show specific examples of data corresponding to some of the 1 readings 1 stored in the dictionary, C1 is an example of the data section for separation, and D corresponds to each data in C9. This is an example of an output data section.

FIG. 4 is a part-of-speech connection table, and this table is used when selecting a word following a previously determined word when selecting a maximum likelihood candidate.

In the figure, column A represents the part of speech of the preceding word,
B represents the part of speech of the word following A; 1. If the intersection of the two is ○, the connection is possible; if the intersection is ○, the connection is not possible. The classification of parts of speech used in this table can be subdivided as necessary, and some examples are shown.

Next, the conversion process flow shown in FIG. 2 will be explained.

First, data entered manually from a keyboard, a file, etc. is held in an input buffer (not shown), and then one character is manually read and retrieved (FIG. 2, 20). Here, it is determined whether or not there is a human power 8fe L'j (21), and if there is not, it is assumed that the selection of the maximum likelihood word has been completed, and the conversion process to a sentence containing kanji and kana (32 to 38), which will be described later, is carried out. Transition.

If there is manual reading (character data), all the words that match the pronunciation (minute writing data) are read from the dictionary's writing data section (C in Figure 3). If there is no clause, the process of step 29 described later is performed, and in some cases, the maximum likelihood evaluation method (prioritizes long clauses, and then Sort each candidate according to the criteria (giving priority to the highest one) (24).

In other words, give priority to long data, then add +=j to the written data from the most frequent parts, and connect the part of speech with the part of speech of the immediately preceding word (the most likely word selected in one process of ni). It is determined whether connection is possible or not based on the table (see 4th M), and the separated data that can be connected are sorted in order of frequency.

When the sorting is completed, the word with the highest priority is extracted as the most likely word from among the words of the most likely candidates that have been sorted (25). Next, determine whether there is a maximum likelihood word or not.
6), if there is not, determine whether there is a previous clause (29).
), if there is, preprocessing for hacking and racking (
Return to the processing of the immediately preceding clause) (30). in this case,
Return to step 25 using route ■, and for the previous clause,
The next ranked word among the candidate words is extracted, and if there is a word, it is determined whether or not one part of speech connection is possible (27).

If it is found that the words can be connected as a result, the most likely words that can be connected are stored in the confirmed single edict Hanwha (109 in Figure 1), and preprocessing is performed to start processing from the reading of the clause following the most likely word. (ibid. 28). Next, return to step 20 according to the path marked ■ in the figure, retrieve the subsequent readings in the input huffer, and
Perform the following similar processing.

If it is found in step 27 above that one-part-of-speech connection is not possible, return to step 25 via route ■.
Pick up other candidate words.

Note that the processing in steps 29 and 30 is performed only when the word (separation data) corresponding to the above input pronunciation is not found in the dictionary (
Step 23) is also executed.

If there is no previous clause in step 29,
Convert all input readings to full-width hiragana (31),
(If all hiragana are displayed like this, it means that it is completely unconvertible).

At step 2+1, if it is found that there is no input reading (end state), the process moves to step 32. here,
In the subsequent flow, "a data" represents the separation data (A, C in FIG. 3), and "b data" represents the output data (BD in FIG. 3).

In step 32, the b data of the dictionary is read from the conversion result of the a data (the first one if a plurality of words are included). It is determined whether or not there is output data in this b data, and if there is not, the a data is converted to full-width hiragana (37), and it is determined whether there is any remaining a data in this sentence (38).

At this time, if there is a sentence, the process returns to step 32 via the route (■), and if there is not, the conversion of all the clauses into sentences containing kanji and kana is completed, and the converted sentences containing kanji and kana are output.

In step 33, if b data has kanji, b
It is necessary to determine whether the data has the same part of speech (34), and since there are multiple kanji with the same part of speech, the a data is replaced with b data according to the frequency (35), and the processing target of the converted character string is The process is changed to the next phrase, the presence or absence of the remaining a data is determined (38), and the same processing as above is performed. ” In step 34, if the b data does not have the same part of speech,
a Convert the data to hiragana.

Next, the present invention shown in FIG. 5 will be explained in detail.

As shown in (1), for example, reader No.j says, ``Kyo wa Haretzu.''

(2) Processing of the first phrase "Ki", "Kyo", "Kyo", "Kyoha", "Kyohaha"...Read the 1-minute writing data and arrange them in descending order of the highest evaluation. The order is as shown in FIG. 5 (2) (maximum likelihood candidate). Here, "Kyoha" is the first candidate because long phrases are given top priority.

(3) Processing of the second clause The first data of the first clause [Kiyo wa (sect) I searched for a candidate with the reading following j, but since it was not in the dictionary, I hack trunked and converted the second clause of the first clause. Data “Kyo (Today)”
Search for pronunciations that follow (``ha'', ``haha'', ``hahare'', ``hahare''...) and create a candidate group as shown in Figure 5 (3).

(4) Processing of the third clause The first data of the second clause is searched for by the pronunciation that follows "haha (mother)", but since it is not in the dictionary, eight-track tracking is performed and the second data of the second clause is searched for. The reading that follows “ha (ha)” (
``Ha''``Hare''``Harezu'') Figure 5 (4)
Create one candidate group as shown in .

(5) Processing of the 4th clause The first data of the 3rd clause is the reading and 2 following the ``ha (ba)''.
I searched for a candidate for the 3rd data ``ha (ha)'', but it was not in the dictionary, so I searched for the pronunciation following the 3rd data ``hare (hare)'' in the 3rd clause by tracking. Noyo・
Create a group of sea urchin candidates.

(6) I searched for a candidate using the reading following the first data "de (de)" and the second data "de (de)" in the fourth clause, but it was not in the dictionary, so I hacked and racked it and found the third clause. Since all reading character strings are converted with the third data "Dezu", output data is created (not shown).

(7) Creating output data From the output data section of the dictionary, select "Kyouj""Ha""Hare"
It reads each data of "desu", creates a converted character string, and outputs it (not shown).

[Advantageous Effects of the Invention 1] According to the present invention, the conversion is completed by repeating backdragging to find the pattern that can form the longest phrase.When backtracking in Japanese conversion, it is possible to convert human-written character strings into phrases or words. Processing can be made more efficient because the amount of data to be handled can be reduced when writing separately.

In addition, although there are many homonyms in Japanese, the number of parts of speech is limited to a certain extent, and there are many homophones and different parts of speech, so by extracting the pronunciation and part of speech, the amount of data to be handled during partitioning can be reduced. I can do it.

By reducing the data length and number of data to be handled, the memory can be used with sufficient margin and the number of file accesses can be reduced.

FIG. 6 is a block diagram of a conventional example, and FIG. 7 is a conversion processing flow diagram of a conventional example.

In Figure 1 ]0: Processing device 11; Dictionary 111: Separating data section 112: Output data section 12: Input section 13: Output section 10 ■: Character extraction means 102・Separation data search means 103: Maximum likelihood word evaluation selection means 104; Output data search means

[Brief explanation of drawings]

Fig. 1 is a diagram showing the basic configuration of the present invention, Fig. 2 is a processing flow diagram of an embodiment of the present invention, Fig. 3 is an example of the structure of dictionary data according to the present invention, and Fig. 4 shows whether or not words can be connected according to parts of speech. Figure 5 is a table representing the patent applicant for the present invention: Bfn Co., Ltd. (one other person), sub-agent patent attorney Kazuo Hosaka (3) Separation data used to process the second clause Separation data

Claims (1)

[Scope of Claims] A kana-kanji conversion device that converts an input character string of kana characters into a sentence containing kanji and kana, which includes: a data section for separation that includes readings and parts of speech used only for separation search; and output data that includes readings, kanji, etc. For each character of the input character string, a word is extracted from the parting data section of the dictionary, candidate words are extracted from the retrieved words by maximum likelihood selection, and the input characters for the maximum likelihood word are extracted. It is determined whether the connection with the immediately preceding word in the string is possible, and if it is possible, the operation is performed on the subsequent character string, and if it is not possible, the immediately preceding word in the input string is backtracked using the above separation data and the maximum likelihood is calculated. A kana-kanji conversion device characterized by selecting a word, converting the selected maximum likelihood word into a sentence containing kanji and kana using the output data of the dictionary, and outputting the converted sentence.