JPS58166435A - Japanese syllabary to chinese character conversion system using probability matrix - Google Patents

Abstract

PURPOSE:To express information of word-to-word concatenation in details and to perform high-level KANA(Japanese syllabary)-KANJI(Chinese character) conversion, by providing a probability information holding means for indicating the probability of concatenation between words and extracting a specific word on the basis of the probability of the concatenation between words obtained from said means. CONSTITUTION:A possible word extracting device 6 extracts words, one by one, from a possible word stack 5 and sets them in a possible word extraction buffer 7. Consequently, a possibility retrieving device 8 accesses a concatenation possibility matrix 9 on the basis of the part of speech of a word set in a last-word buffer 2 and that in said buffer 7 to set the concatenation possibility in a possibility buffer 10. An evaluated-value arithmetic device 11, on the other hand, sets the evaluated value of the word in an evaluated-value buffer 12 on the basis of data set in the possible-word extraction buffer 7. A multiplication part 13 multiplies the value set in the possibility buffer 10 by that in the evaluated- value buffer 12 and sets the product in a priority buffer 14.

Description

DETAILED DESCRIPTION OF THE INVENTION (1) Technical Field of the Invention The present invention relates to an 11h kana-kanji conversion method that automatically creates kana-mixed sentences in which kanji and kana characters are missing.

Especially the connection information between words! This invention relates to a kana-kanji conversion method that realizes high-performance kana-kanji conversion.

c) Conventional technology and problems, for example! As shown in FIG. 11, when connecting "nai" to the word company, it is necessary to convert it into "inside the company." In this case, each word connection information is used as shown in FIG. 1(b).

Nouns and nouns can be connected, but nouns and verbs cannot be connected...Conversion is performed based on the m5ts information of the group.

In this case, ``nai'' can be converted into the auxiliary verb ``nai,'' which expresses negation, or the noun ``uchi,'' so with this single candy threshold, is it converted to ``no company'' or ``inside the company''? It is unclear whether this will be done. In this case, the next KU and next KM degree information is used. In other words, in Japanese sentences, the use of St for ``uchi'' is louder than the frequency of use for ``nai'', so there is an undesired conversion to ``company nai''.

Let me clarify this point further.

The single-interval connection information used in conventional kanaji conversion devices expresses what types of words can be connected to a certain type of word. is a possibility bit matrix. According to this representation method, both the connection of words that do not occur often and the connection of words that occur frequently are expressed by 1 bit, so they are represented as "1". For example, in most cases, two nouns and the negative auxiliary verb ``nai'' do not connect, but they do connect with a special noun. As an example, r company J (noun) and ``nai'' do not connect, but ``howa J'' (noun) and ``nai J'' do.

In this way, an interlayer exists when a non-determined relationship is expressed using possibility bits.

Now, enter the rounded kana kanji conversion number #Kr force ishanai J to create the text ``J within the company''.

Kana-kanji conversion device is +! I 1 Figure 1o) Using the word association information 11 (connectivity matrix) shown in K.

Perform one ls vote as shown in Figure lI2.

To convert kana-kanji in this case, first, at level 1, search the dictionary for words that have a partial pronunciation of the left edge of the input ('
company,” “shellfish,” etc.) Then, assuming that the word "company" that is considered to be the best among these words is correct, the subsequent input string "nai" is analyzed at level 2.

The dictionary also says "nai" for this "nai".

"inside"... etc. are searched. The most likely word will then be selected from these words,
Which one is most likely is the longest match method that prioritizes long readings, or the method that takes #1 degree and reading length into account (e.g., maximum likelihood evaluation method, S degree x 32 x shi/guess). (evaluation; patent application filed) etc. are used. However, no matter which method is used to determine, in this case, when evaluating ``nai'' and ``in'', the trajectory is the same, so the priority of the selection is determined by the S degree, but ``nai'' Considering the degree of S that appears in the Nippon Ichi text of ``nai'', ``nai'' is larger [j is larger, so ``nai'' is given priority. Then, Figure 111I (b)
The state of connection between parts of speech is checked using the word relation information 11 shown in Figure 1, but since it is expressed in a matrix that nouns and auxiliary verbs can be connected, 0. As a result, the unnatural conversion result is ``there is no company.''

(3) Purpose of the Invention One aspect of the present invention is to improve the problem that information on word sequences that often appear in the rounding of single click sequences that do not appear often as described above does not have a single effect, and to develop a kana formula using a nine-rate matrix. It provides a kanji conversion method.

(4) Structure of the invention In order to achieve this objective, the kana-kanji conversion method using a probability matrix of the present invention has a single file,
In a kana-kanji conversion device that automatically converts Japanese input in katakana waru or kana to a sentence containing kanji and kana using the connection information of a single string, this is a sign that shows the connection between kanji and kana. It is interesting to note that a means for holding & holding is provided, and a specific word is extracted from the means for holding probability information based on the jealousy of the probabilistic information holding means.

151 'Jk Embodiment of the Invention Before explaining the present invention in detail based on one embodiment, the principle of the present invention will be briefly explained based on FIG.

In the present invention, the connection information between single wheels is displayed as a probability matrix as shown in FIG. 5(a), and the friend connection confirmation shown on this matrix is used to determine the priority order of word selection. In this i-table matrix, the conjunctions of each part of speech are displayed so that the sum of all rows becomes 1.
For example, the rate at which the verb is connected to the d-case particle indicates that it is the sum total of all parts of speech connected after the case particle. This probability can be statistically obtained from Japanese sentences.

Therefore when you type "Kaishanai".

First, the probability that ``company'' is converted, 9, and ``uchi'' is connected is 8 if we use ``uchi'' as the suffix, and the probability that the auxiliary verb is connected is mu.

Therefore, separately, using the longest match method or the Hiroshi maximum likelihood 1'F value method, the evaluation value F (within 1. within b) of "within" and "no" is calculated.
The method calculates F (+1. h no), multiplies it by the previous priority number, and outputs the larger one as the selection result.

This committee, II' (l not 1. h not) >F (I
Of these, 1. For example, F (l not 1. h not) - 1000, F (l not 1. h not) - 5.
000), the priority is calculated as s'o xF (l in l r h), -7, XF (+ not 1. h not), so there is a possibility of reversal.

An embodiment of the present invention will be described based on FIG.

Figure 4 (a) shows the configuration up to creating the priority number,
The same (b) shows a state in which a priority order candidate sequence is obtained in the priority order buffer based on the number of priorities.

In the figure, 1 is the input buffer and 2 is the previous word buffer.

5 is a dictionary, 4 is a search device, 5 is a candidate stack, 6 is a candidate extraction device, 7 is a candidate extraction buffer, 8 is a probability search device,
9 is a connection probability matrix, 10 is a one rate bathopha, 11
1 is an evaluation value calculation device, 12 is an evaluation value buffer, 13 is a multiplication unit, 14 is a priority number buffer, 15 is a Fi sorting device, 1
6 is a priority order buffer, and P is a priority number calculation device.

It's a hippo sofa! Kana input for requesting i-exchange is set. The previous word buffer 2 is used to set the word that was converted first among kana characters, and in the case of ``Kaishanai'', ``E company'' is set.

1111F43 is a file in which each Maki word is stored. This dictionary 3 includes kana reading, kanji, and frequency.

Parts of speech, etc. are also stored together.

The search device 4 searches the dictionary 5 based on the kana input set in the input buffer 1, and sets the output of the search results in the candidate stack 5.

The Iiura mound removal TIL 6 extracts the &sequence vermilion from the candidate stack 5 and sets it in the candidate extraction buffer 7.

The calculator search device 18 accesses the connection probability matrix 9 based on the part of speech of the rate candy set in the m rate candy buffer 2 and the part of speech of the main article set in the candidate extraction buffer 7, and searches the connection between the parts of speech. The probability is read out, and the read calculator is set in the probability buffer 10. The nine-connected frequency matrix 9 is constructed as shown in FIG. 5(a).

The #fiiIl noodle calculation device 11 calculates the evaluation value of the word set in the candidate extraction buffer 7.

The calculation method is well-known IIIL! The same #f value method may be used, or the maximum likelihood evaluation method may be used. The 1III value is then output to the evaluation value buffer 12.

The multiplier 13 multiplies the nine evaluation values set in the evaluation value buffer 12 by the probabilities set in the one-rate pathophone 10 to obtain the priority number of each word. 11<Set on sofa 14.

No) Based on the priority number set in the sofa 14, the device 15 sets the candidate words in order of priority (excluding the priority number 00) in the sofa 16 in descending order of priority, and outputs the candidate words in the output priority order. It is something that creates

The priority number calculation device dP calculates the priority number for each word, and uses the probability search device ms, *continuous probability matrix 9. A rate bathofa 1G,! 'F value calculation device 111
．． It is equipped with an evaluation value balancer 122, a multiplier 15, and the like.

next! The operation of v7&position shown in FIGS. 44(a) and 44(b) will be explained.

Here, ``1 kaisha yanai'' is input, and ``kaisha'' is converted to ``company'' and this is set in the previous word buffer 2. A case will be explained in which ``nai'' is converted.

■ The search device 14 searches for the character string “Nai” in the input buffer 1.
The dictionary 3 is searched for the word ``-'', and the resulting output is entered in the candidate stack 5.

(2) The candidate extraction device 6 extracts words one by one from the candidate stack 5 and sets them in the candidate extraction buffer 7.

■ From this K, the probability search result 118 is the previous word buffer 2.
The part of speech (in this case, noun) of the word set in , and the part of speech of the word set in the candidate extraction buffer 7 (see
In the case of (a), the noun connection probability matrix 9 is activated, the connection probability is read out, and this is set in the -table buffer 10.

■ On the other hand, the evaluation value calculation device 11 is set in the candidate extraction buffer 7 and calculates the evaluation value of the word based on the 9 data.
2) is stored in the evaluation value buffer 12.

(2) The multiplier 15 multiplies the multiplication table written in the G4 buffer 10 by 32 set in the #P value buffer 12 to obtain 1.6 and sets this in the priority number buffer 14.

(2) Such calculation of the priority number is performed for all candidate annual salaries output to the candidate stack 5. As a result, as shown in FIG. 4(b), based on the priority number set in the priority number buffer 14, the notebook device 1t15#'i bonded party Iit number of 00 is excluded, and the larger number of such devices is
They are sequentially arranged in the priority order buffer 16 to obtain priority order candidates 1j. From this K, as shown in FIG. 4(b), "inner" can be extracted as the highest priority conversion output candidate.

Through this process, it is possible to reverse the priority of ``uch'' and ``1'', which could not be realized with a fixed e-value, from the previous 4, and in this way, practical kana-kanji conversion can be achieved. It can be performed.

(6) Effects of the Invention According to the present invention, since the connection information of Fengjieguan is configured as a probability matrix, b! i1 It is possible to change the priority at the beginning of the year, which tends to be fixed, by changing the horizontal order of the nine words that have already appeared, and it can contribute to improving the accuracy rate of kana-kanji conversion and improving the conversion speed.

[Brief explanation of the drawing]

FIGS. 1 and 42 are explanatory diagrams of a conventional conversion method using single-to-one connection information, and FIGS. S and 1A are configuration diagrams of an embodiment of the present invention. In the figure, 1 is the input buffer, and 2 is the previous word buffer. 5 is a dictionary, 4 is a search device, 5 is a candidate stack, 6 is a candidate multiplication device, 7 is a candidate extraction buffer, 8 is a probability search value -1
9 is a tangent & probability matrix, 10 is a calculator pa sofa, 11
12 is an evaluation value calculation device, 12 is an evaluation value buffer, 15 is a butterfly multiplication unit, 14 is a priority number buffer, 15 is equipped with a notebook, 16
is a priority buffer, a PFi priority number calculation device. Patent Applicant ``D Shika Co., Ltd. Patent Attorney Yamatani #1''

Claims (1)

[Claims] (1) In a kana-Shiji conversion device that automatically converts a single file into a ``Shiji-Kana'' character inputted in Katakana or Kana and uses the connection information between words, the word A probability information holding means is provided which indicates the probability of a connection between the two, and the probability information holding means is characterized in that the probability information holding means extracts *@ of 111# based on the connection probability of 9 single 1 leap. Nine kana kanji conversion method using probability matrix.