JPS61128366A - 'kana'/'kanji' converter - Google Patents

Abstract

PURPOSE:To compress the occupying capacity of a memory by constituting the titled converter of n connection matrix list bodies obtained by extracting lines or rows having different arrangement of elements from a connection matrix list and registering the extracted lines or rows in each digit and n connection matrix list indexes indicating the positions of the elements in the bodies. CONSTITUTION:Each line in the connection matrix list divided into four rows is regarded as one record, the record is divided into an upper digit and a lower digit and an upper digit connection matrix list body registering only an upper digit different code having different arrangement of a bit string of the upper digits and a lower digit connection matrix body registering only a lower digit different bit string having the different arrangement of the lower digit bit string are registered in each digit in the order of the connection matrix list. In addition, a lower digit connection matrix list index and an upper digit connection matrix list index are formed individually for the lower digit connection matrix list body and an upper digit connection matrix list body. Thus, the occupying capacity of the memory can be reduced sharply and retrieval can be executed easily.

Description

TECHNICAL FIELD The present invention relates to a kana-kanji conversion device, and more particularly to a kana-kanji conversion device suitable for compressing a connection matrix table showing connectivity between words.

Conventional technology As an input method for a kana-kanji conversion device.

■Word unit method, ■Kanji section designation method, ■Phrase unit method,
■There are solid writing methods, etc., but ■other than the word-by-word method,■
Kanji section designation method, ■Bunsetsu unit method.

■In the solid writing method, etc., it is necessary to perform grammatical analysis on the input sentence &ζ. During this grammar analysis, it is necessary to determine the possibility of connection between words, and for this purpose, connection testers that display connection information between words are widely used.

A connectivity tester is usually represented in the form of a matrix and is called a connectivity matrix table.

FIG. 7 is a diagram showing a conventional general connection matrix table.

As shown in Figure 7, the item ``The word before r'' is filled with ``The word before r''.
The part of speech of the "later word" item l; is assigned the part of speech of the "later word", and the connectivity between the "r previous word" and the "later word" is the "r previous word". The connectivity of the "later word" at the part-of-speech level is determined. In Fig. 7, the connection value '0' indicates that connection is impossible, and the connection values '1', '2', and 3' indicate that connection is possible, and the larger the connection value, the higher the probability of connection. For example, after the verb l, the verb 1
It is impossible for these to be connected, and after 1 noun 1,
Verb 1, 2, 3. Nouns 1, 2. Particles 1 and 2 can connect, but particle 1 with connection value ``3'' has the highest probability of connecting.
The first rank is the particle 2 with the connective value "2', and the third rank is the verbs 1, 2, 3 and nouns 1, 2 with the connective value "l'.

Conventionally, the connection matrix table is about 256 rows x 128 columns, as shown in the paper published in rNHK Technical Research, Volume 25, No. 5, Aizawa, Ehara, "Kana-Kanji Conversion on Computers JPP 261-298". are used, but
If you simply memorize this in a table format, even if you express the connection between words with 1 bit (0, 1), it will be 256 x 12
8 = 32768 bits - 4096 bytes, which requires approximately 4 bytes of storage capacity. This connection matrix table is usually 1,
Since it is placed on the main memory, the main memory occupancy is a problem. Furthermore, the part-of-speech classification of words is further subdivided into 340 lines x 2
56 column), the possibility of connection is not 0.1, but as shown in Figure 7 above, the connection weight representing the strength of connection, or the connection probability (
For example, if it is expressed as 2 bits 0, 1, 2.3), then 340X256X2=174080!2176X8
= 2176 bytes, which requires a storage capacity of 22 bytes.

Traditionally, to compress this connectivity matrix table.

■Depending on the type of word, the range of words that can be connected is limited, and the O parts that do not connect are clustered, so categorize the words appropriately (e.g., nominal, particles, auxiliary verbs, etc.) and create a connection matrix table. How to divide and reduce capacity.

Method 1 Using a non-zero element table that collects only elements that are not +30 A general method for handling matrices with many 0 elements, such as dividing a matrix into several blocks and not storing blocks whose only element is 0, is adopted. It was getting worse.

However, method (2) requires a large number of tables, making it complicated to handle, and does not provide a very large compression effect.Furthermore, it is troublesome to handle exceptional words that are difficult to divide. In method (2), the procedure for reconfiguring the table into a one-element table is troublesome. “Objective” The object of the present invention is to solve the problems of the prior art as described above, to significantly reduce the amount of memory occupied, and to provide a kana-kanji conversion device equipped with a connection matrix table that can be easily searched. There is a particular thing.

Configuration In order to achieve the above object, the present invention includes a word dictionary,
It has a connection matrix table that shows connection information between words, and for character strings input as one phonetic character, kana-kanji conversion is performed using the word dictionary and a connection matrix table in which elements are expressed in n binary digits. In a kana-kanji conversion device that performs processing, when the connection matrix table is divided into multiple blocks in units of rows or columns, when elements are viewed by digit from each row or column formed in each block, n connection weights in which only rows or columns in which the elements are arranged differently are extracted and registered for each digit.

The present invention is characterized in that it is composed of a sequence table main body and n connection matrix table indexes indicating to which element of the connection matrix table each element of the n connection matrix table bodies corresponds.

Hereinafter, the configuration of the present invention will be explained in detail using an example.

FIG. 2 is a block diagram of a kana-kanji conversion device according to an embodiment of the present invention.6 In FIG. 5 is a dictionary, 5 is a connectivity test section, 6 is a connection matrix table, 7 is an evaluation section, 8 is a backtrack control section, and 9 is an output section.

FIG. 3 is a diagram showing an example of specific contents of the word dictionary 4 in FIG. 2.

As shown in Figure 3, the word dictionary IF4 includes "yomi", "
"Orthography", "part of speech", and "rank" necessary for homophone selection are listed.

A Japanese sentence is inputted from the input unit 1 using one phonetic character (one hiragana, one bakana, and the Roman alphabet), and an analysis target character string to be searched in a dictionary is created by an analysis target character string creation unit 2. The dictionary search unit 3 searches the word dictionary 4 from the beginning of the created character string to be analyzed, and all conversion candidates corresponding to the "yomi" are extracted.

The connectability test section 5 tests whether or not the conversion candidates extracted by the dictionary search section 3 can be connected to the immediately preceding converted word (conversion result) based on the connection matrix table 6, and determines connectable conversions. Test whether there are any candidates.

The evaluation unit 7 evaluates the connectable conversion candidates using an evaluation formula whose parameters are rank, reading length, connection weight, etc., and outputs the conversion candidate with the highest evaluation value as the conversion result from the output unit 9. Output.

As a result of the dictionary search, if there is no corresponding conversion candidate or if there is no conversion candidate that can be connected to the immediately previous converted word (conversion result), the backtrack control unit 8 The analysis may be incorrect, so
To redo the previous analysis without immediately processing unregistered words.

FIG. 1 is a diagram for explaining the process of compressing a connection matrix table according to an embodiment of the present invention.

Figure 1 (,) shows the connection matrix table before compression, with 340 rows x
There are 256 columns, each element is 2 bits (0, 'l, 2.

It has information on 4 stages of 3).

FIG. 1(b) is a diagram showing four tables (blocks) of 340 rows and 64 columns, which are obtained by dividing the connection matrix table of FIG. 1(a) into four equal parts vertically (by column). The four tables obtained by the division each have a row length of 64 columns, and each element has 2 bits of information.

Note that the symbols ■, ■, and ■ in Figure 1 are marked for convenience in order to identify the four tables obtained by division.

FIG. 1(e) is a diagram showing a connection matrix table index and a connection matrix table body according to an embodiment of the present invention.

First, consider each line of each of the four machines (1) to (2) in Figure 1(b) as one record. It has 4 levels of information, ゜2.3, and has 2 binary digits "oo".

Since it is expressed as '01', '10', '11',
If we divide the record into upper and lower digits, the upper and lower digits each have the same arrangement of bit strings. For example, in the "binary one-finger bit string by digit" column in Figure 1(d), ■ indicates the lower digit and ■ indicates the upper digit; The lower digits of both records are 00101101... and are the same. Therefore.

Only those in which the bit strings of the upper digits are arranged differently (this is called a bit string with different upper digits) and the bit string with a different arrangement of lower digits (this is called a bit string with different lower digits) are separated by digit.

Register in the order of tables 1 to 2 in FIG. 1(b). The upper digit connection matrix table itself is the one in which only the bit strings with different upper digits are registered,
A table in which only bit strings with different lower digits are registered is called a lower digit connection matrix table body. The lower digit connection matrix table body and the upper digit connection matrix table body are lower digit connection matrix table indexes that index the respective record positions.

It has a separate lower digit connection matrix table index.

In this case, the actual measurement results showed that 543 bit strings differed in the lower digits, and 182 bit strings differed in the upper digits. Therefore, when compressed using the method shown in Figure 1, the 11th
In 1(a), 340 rows x 256 columns x 2 bits = 2176
In contrast, in Figure 1 (C), the lower order connection matrix table body 264 columns x number of bit strings with different lower digits (543 rows) x l bits = 4344 bytes the upper order connection matrix table body 264 columns x The number of bit strings differs in the upper digits (182 rows) xi bits = 1456 bytes, and the lower digit connection matrix table index is 2 bytes.

If the upper digit connection matrix table index is expressed in 1 byte, the lower digit connection matrix table index: 340 rows x 4 pieces x 2 bytes = 2720
Byte upper digit connection matrix table index: 340 rows x 4 pieces x 1 byte = 1
36.0 bytes, and the entire connection matrix table is 4344 bytes + 145
6 bytes + 2720 bytes + 1360 bytes = 9880
It is approximately 9.9 bytes, and can be compressed to approximately 1/2.

FIG. 4 is a diagram showing the processing flow of the connectability verification unit when searching the main body of the lower digit and upper digit connection matrix table using the lower digit and upper digit connection matrix table index shown in FIG. 1(C). , 0, - Set the row address in the virtual connection matrix table (regular connection matrix table) before compression from the code indicating the "part of speech" of the previous word (obtained from the word dictionary 4) ( 401), then set the column address in the virtual connection matrix table before compression from the code indicating the "part of speech" of the next word (40
2).

From this row and two column addresses, the row addresses of the lower digit connection matrix table index and the upper digit connection matrix table index according to this embodiment, and the lower digit connection matrix table main body.

Find the column address of each upper digit connection matrix table body (
403).

Now, in the regular connection matrix table, let i be the row address indicating the position of the previous word, and j be the column address indicating the position of the subsequent word. In this case, if the first place of the quotient of j/64 is n, the following words belong to the (n, +1) table in FIG. 1(b). Therefore, the corresponding row address PL+P2 of the lower digit connection matrix table index and the lower digit connection matrix table index is PL or p 2 = i + n X 340
It can be obtained from (1).

On the other hand, the column address jT indicating the digit of the next word is the corresponding lower digit connection matrix table body, and the column address of the lower digit connection matrix table body is (11+'12).

(Il* or CIz=j-nX64' (
2).

When the row addresses P1+P2 of the lower digit connection matrix table index and the lower digit connection matrix table index corresponding to the previous word are obtained, the lower digit connection matrix table main body,
Since each connection number with the upper digit connection matrix table body can be recognized (404), the rows of the lower digit connection matrix table body and the upper digit connection matrix table body corresponding to the recognized connection numbers are searched, respectively, and the above One bit of information is obtained from each intersection with column address ql+'12 (405). After obtaining the two pieces of 1-bit information, binary information is generated from them, and this is used as a connection value (406).

In this way, in this example, the original table (
can be reconstructed into a regular connection matrix table).

FIG. 5 is a diagram for explaining the second embodiment. In this embodiment, when the connection matrix table is divided into four parts and viewed by lower digits and upper digits, the difference is that only the bit strings are registered in the lower digit connection matrix table body and the upper digit connection matrix table body, respectively.
Same as Figure 1.

The difference from Figure 1 is that the lower digit connection matrix table body is divided into pages of 256 records, and the lower digit connection matrix table index code is set to 10 bits (main page selection bit 2).
The key point is that the lower digit connection matrix table index is compressed by expressing it with 8 bits of address within 10 bits page).

However, in this embodiment, one page is made up of 256 records, so if the total number of rows in the lower digit connection matrix table is 543 as shown in FIG. All information is stored in , and after that it becomes empty.

According to this method, the lower digit connection matrix table index is 10 bits x 340 rows x 4 = 1700 bytes, and the upper digit connection matrix table index is 1360 bytes.

Combined with 4344 bytes of the lower digit connection matrix table body and 1456 bytes of the upper digit connection matrix table body.

The entire connection matrix table can be reduced to approximately 8.9 bytes,
The memory capacity can be further reduced by approximately IKby1 compared to the first embodiment.

FIG. 6 is a diagram showing a connection matrix table according to the third embodiment.

The same bit string exists between the lower digit different bit string of the lower digit connection matrix table shown in FIG. 1 and the upper digit bit string of the upper digit connection matrix table shown in FIG. Therefore, in this embodiment, the same bit strings are registered all at once in the main body of the common connection matrix table to avoid duplication.

In other words, the upper digits or bit strings are stored first in the front part of the common connection matrix table body, and the remaining lower digits are stored in the rear part after excluding the same bit strings between the lower digits or bit strings and the upper digits bit strings. A bit string was stored. In addition, contrary to the above storage method, the lower digit bit string is first stored in the front part of the common connection matrix table body, and the same bit string of the lower digit bit string and the upper digit bit string is stored in the rear part. It is also possible to store the excluded remaining high-order digits or bit strings.

According to actual measurement results, there were 30 identical bit strings, and the total number of rows in the main body of the connection matrix table was 695 rows. Therefore, the memory capacity of the common connection matrix table body is 64 columns x 695
Row x 1 bit = 5560 bytes. When the lower digit connection matrix table index is compressed to 1700 bytes as in the second embodiment, and the code of the upper digit connection matrix table index is expressed in 1 byte (1360 bytes), the entire connection matrix table according to this embodiment The memory capacity is 8620 bytes, making it possible to further reduce the memory capacity compared to the second embodiment.

In addition, although each of the above embodiments was an example in which a regular connection matrix table was divided into four, the present invention is not limited to four divisions.
The number of divisions is free.

In addition, although we have explained the example in which the elements of the connection matrix table are two binary digits from 0 to 3, when the elements are multi-valued with two or more binary digits, pay attention to the bit string of each digit in the same way. By configuring the connection matrix table using the same method as in each of the above embodiments, the memory capacity required for the connection matrix table can be significantly reduced. Further, although the explanation has been given using an example of dividing on a column-by-column basis, the same effect can be obtained when dividing on a row-by-row basis. In addition, in systems where memory occupancy is a problem, the lower order and upper digit connection matrix tables can be stored as external files.
It is also possible to search by indexing the lower digit and upper digit connection matrix table in the internal memory. Of course, the lower digits, the upper digit connection matrix table body, and the lower digits.

It is also possible to use both the upper digit connection matrix table indexes as external files. Further, although each of the above embodiments was applied to a kana-kanji conversion device using a solid writing input method, the present invention can also be applied to a kana-kanji conversion device using a kanji part specification method or a bunsetsu unit method. Needless to say.

Effects As explained above, according to the kana-kanji conversion device of the present invention, it is possible to significantly reduce the amount of memory occupied and to realize a connection matrix table that can be easily searched.

[Brief explanation of drawings]

Fig. 1 is a diagram for explaining a connection matrix table according to an embodiment of the present invention, Fig. 2 is a block diagram of a kana-kanji conversion device to which Fig. 1 is applied, and Fig. 3 is a diagram of a word dictionary in Fig. 2. A diagram showing an example, FIG. 4 is a diagram showing the processing flow of the connection possibility search unit in FIG. 2, FIG. 5 is a diagram for explaining the connection matrix table according to the second embodiment of the present invention, and FIG. 7 is a diagram showing a connection matrix table according to the third embodiment of the present invention, and FIG. 7 is a diagram showing a conventional general connection matrix table. Input section, 2: Analysis target character string creation section, 3: Dictionary search section, 4: Word dictionary, 5: Connectivity test section, 6: Connection matrix table, 7: Evaluation section, 8: Backtrack control section, 9 :output·
Department. Figure 1 Figure 1 (d) Figure 2 Figure 3 Figure 4 ■ Figure 6 Figure 7

Claims (3)

[Claims] (1) It has a word dictionary and a connection matrix table that shows connection information between words, and for character strings input in phonetic characters, the word dictionary and a connection matrix table in which elements are expressed in n binary digits. In a kana-kanji conversion device that performs kana-kanji conversion processing using n connection matrix table bodies in which only rows or columns whose elements are arranged differently when viewed by digit are extracted and registered for each digit, and each element of the n connection matrix table bodies is 1. A kana-kanji conversion device comprising n connection matrix table indexes indicating to which element it corresponds. (2) Part or all of the n connection matrix table bodies are divided into pages, and the connection matrix table index corresponding to the connection matrix table bodies divided into pages is the connection matrix table index corresponding to the divided connection matrix table bodies. 2. The kana-kanji conversion device according to claim 1, further comprising a page selection bit indicating a page number of the main body of the sequence table. (3) The kana-kanji conversion device according to claim 1, wherein the n connection matrix table bodies are integrated while avoiding duplication of arrangement of elements.