JP2633824B2 - Kana-Kanji conversion device - Google Patents

Description

Description: FIELD OF THE INVENTION The present invention relates to a kana-kanji conversion device for converting a kana input to a kana-kana mixed sentence, and more particularly to a kana-kanji conversion device that can convert even if the input is not divided into phrases. The present invention relates to a kana-kanji conversion device suitable for so-called solid input.

[Background of the Invention]

Conventionally, the method of kana-kanji conversion of solid input is as follows.
Hitachi Review, May 63, 1981, Vol. 5, No. 5, "HITAC L-320 / 30
H, 50H document processing function ”(hereinafter referred to as reference (1)), a method combining longest match and backtrack,
1978 19th Annual Convention of the Information Processing Society of Japan 5E-4
"Kana-Kanji conversion system for solid text", hereinafter referred to as "Sentence line (2)", the two-phrase longest matching method. About the phrase structure analysis algorithm used and its efficiency "(hereinafter referred to as reference (3)), the minimum number of phrases method and the like are well known.

In the method of Document (1), an independent word is cut out from the left end of the input by the longest match, and then a phrase conversion process of assigning a grammatically connectable auxiliary word to the independent word is repeated until the right end is reached. If not reached, backtracking is performed and another phrase conversion is attempted and the process proceeds to the destination.

In the method of Document (2), all possible conversion candidates are brute-force extracted over two phrases from the left end of the input, and 2
This is a method in which the sum of the lengths of the clauses is selected to be the longest, and this is used as the cut of the first clause, and this point is used as a starting point and the same is repeated to determine the breaks of each clause and convert.

The method of Document (3) is a method in which a phrase is cut out from the left end in a brute force manner, and a combination that minimizes the number of phrases is selected from the cutout to obtain a conversion result.

In solid input, since there is room for selection of where to make a break between phrases, there are generally more polysemy candidates for conversion than in the case of phrase division writing. For example, there are many interpretations of the input, "On mathematics analysis,""On mathematics transfer,""Onmathematics,""Mathematics on archeological sites," and so on. It is possible. Therefore, reference (1),
There is a problem that it is difficult to obtain high conversion accuracy even by using the methods described in (2) and (3).

[Object of the invention]

SUMMARY OF THE INVENTION It is an object of the present invention to provide a means for solid writing input that efficiently extracts and holds a plurality of conversion candidates having a high likelihood of conversion certainty, and easily and quickly selects and determines a correct conversion result from the conversion results. It is in. It is a second object of the present invention to provide a learning mechanism for solid writing input for improving the accuracy of conversion.

[Summary of the Invention]

In order to achieve the above object, in the conversion, a conversion candidate is cut out from the left end of the input character string to the right, a likelihood is obtained, and processing is performed while holding the likelihood within a certain range. When the processing is completed up to the right end of the input character string, a candidate string having the highest likelihood is selected from the conversion candidates and displayed as a conversion result. At the same time, other candidate groups held in order from the left of the candidate column are displayed, so that selection and correction are possible.
If a candidate different from the displayed conversion result is selected and determined, the likelihood is re-evaluated for the remaining undetermined portion, and if necessary, the display of the conversion result is changed. The selected compound words and independent words with affixes are retained, and when the same character string is subsequently input, the likelihood is increased to improve the conversion accuracy.

In this way, by selecting and holding a candidate having a high possibility, it becomes possible to display and select another candidate at high speed even when a converted narration occurs, and the amount of storage device required for that is small. I'm done. If the conversion error is corrected by selecting another candidate, not only that part but also the following candidates are re-evaluated. For example, if the position of the break between candidates changes, Subsequent candidate columns starting from the new delimiter position are prepared, so that the number of subsequent correction steps can be reduced. Furthermore, the same error is not repeated for the conversion error part since learning is performed thereafter. In particular, since words with prefixes and suffixes and compound words are learned in long units, the effect of improving conversion accuracy is great.

(Example of the invention)

 Hereinafter, an embodiment of the present invention will be described.

FIG. 1 is a diagram showing the configuration of a kana-kanji conversion device of the present invention, which comprises an input / output unit 1, a control unit 2, and a storage unit 3. The input / output unit 1 includes a keyboard for inputting kana input, conversion and selection instructions, and a display unit for displaying conversion results and selection candidates. The control unit 2 controls execution of kana-kanji conversion. The storage unit 3 temporarily stores an input kana character string and data used for conversion.

FIG. 2 shows a method of executing conversion using the present kana-kanji conversion device. The user inputs a kana character string from the input / output unit 1 and instructs conversion (10). The control unit 2 performs a conversion process (20) of creating a candidate having a high likelihood of conversion certainty and storing the candidate in the storage unit 3 with respect to the kana input. Next, the control unit 2 creates an alternative candidate to be selected corresponding to the candidate row having the highest likelihood and the first candidate from the candidates in the storage unit 3 (30), and causes the input / output unit 1 to display the alternative candidate. (40). When the user selects a target candidate from the selection candidates (50), the control unit 2 determines the selected part and learns the selected candidate (6).
0). Next, the control unit 2 performs a display candidate creation process (30) again on the undetermined portion, and creates and displays a conversion result and a replacement candidate for the subsequent portion. The above display (40), selection (50),
The determination processing (60) and the display candidate creation processing (30) are repeated until all the conversion results are determined.

Hereinafter, the operation of the control unit 2 described above will be described in more detail.

FIG. 3 is a flowchart showing the operation of the conversion process (20). FIG. 4 shows the storage unit 3 according to the conversion processing (20).
It is a figure showing an example of data created above. The conversion process (20) in FIG. 3 will be specifically described using the example shown in FIG.

Now, it is assumed that the input character string is "in the world". 3 and 4, n is a pointer indicating the position of the character counted from the head of the input character string. In step 201 of FIG. 3, n is first set to 1 and the n is set at the head of the input character string. At step 202, it is checked whether there is a segment end where the position of n ends. It is assumed that the beginning of the character string is specially regarded as a clause end. Step 20
Go to step 3 to perform phrase conversion. In the phrase conversion, a conversion candidate capable of performing one phrase conversion is extracted with the position of n as the head of the phrase. Here, a phrase refers to a form in which one independent word is preceded by an optional prefix, an independent word, an optional suffix, and an optional appendix. Compound words are considered clauses. In the example of FIG.
"Number", "suck", "mathematical society", "mathematics world", "mathematics floor", "number", "suck" are extracted as conversion candidates from the top. The kana-kanji conversion for each phrase is
The method shown in NHK Technical Research Vol. 25, No. 5, May 1973, "Kana-Kanji Conversion by Computer" is well known. Next, in step 204, the likelihood of the conversion candidate is determined, and it is determined whether the conversion candidate should be stored in the storage unit 3 or discarded (called pruning). The likelihood of the certainty is basically, as described in the literatures (1), (2), and (3), that the input character string is a sequence of longer phrases, or in other words, fewer phrases. Is determined so that the likelihood is higher when the image is decomposed into columns. However, in the same phrase, adnominals such as "this" and "that" and formal nouns such as "koto" and "mono" are often attached to other phrases, so they are independent like nouns and verbs. When counting the number of phrases without considering it as one phrase, the value is set to a value smaller than 1. As described above, the number of phrases is obtained by multiplying the weight in consideration of the part of speech and the appearance frequency, and the smaller the number is, the higher the likelihood is. Specifically, nouns, verbs,
Weight 1 for adjectives and adjective verbs, 0.1 for formal nouns, auxiliary verbs, adverbs, etc., 0.5 for prefixes and suffixes as semi-independent words
Weight.

Pruning calculates the likelihood from the beginning of the character string to the end of the phrase that is currently the subject of the determination, and defines it as the likelihood at the character position at the end of the phrase. Is found, it is compared with that value, and if that value exceeds a certain allowable value, pruning is executed. In this embodiment, the allowable value is the minimum value of the number of weighted phrases at the same character position + 1. However, it goes without saying that the present invention is not limited by how to determine the likelihood of the certainty and the size of the allowable value of pruning. A feature of the present invention is that by extracting all phrase candidates within a certain range of likelihood of certainty and storing them in the storage unit 3, later selection and correction can be easily performed at high speed.

In FIG. 4, all the clauses extracted from the head are stored in the storage unit 3 (step 205). When storing the data, the data can be formed into a network as shown in FIG. 4 so that the occupied storage capacity can be reduced. For example, “number” and “suck” can share the right end of a phrase, thereby unifying subsequent phrases. A specific method of expressing such data is well known as list processing.

At step 206 in FIG. 3, 1 is added to n to advance the pointer to the next input character position. At step 207, it is determined whether or not an input is still present at the character position of the pointer n, and if there is, the process returns to step 202.

If n is 2, there is no right end of the phrase at this character position, so go to step 206 and immediately move the pointer n to the next character position. If n is 3, the end of the phrase "number" or "suck" is possible, so the next phrase is extracted from this position in step 203, and "forehead", "easy", "forehead", "Framework",
A phrase such as "forehead" can be obtained. N is 4 in step 204
"Mathematical" has already been obtained as a candidate at the position of, and the number of weighted phrases is one. In the phrase candidates of “amount” and “easy”, the number of weighted phrases at the position where n is 4 is 2 by adding 1 of “number” and 1 of “amount” or “easy”. From the pruning conditions, “forehead” and “easy” fall within the allowable range, and are registered in the network at step 205.

As described above, steps 207 to 207 are replaced by n.
Is repeated until 14 is obtained, whereby the complete network shown in FIG. 4 is obtained. However, here, "kai, world,
“Floor, family, chemical, up, place, state” is a suffix.

In step 208, the pointer P indicating the head position of the undetermined portion of the conversion result is initialized to 1, indicating that there is no determined portion yet.

Thus, the operation of the conversion process (20) in the control unit 2 is completed, and then the operation proceeds to the operation of the display candidate creation process (30).

FIG. 5 is a flowchart showing the operation of the display candidate creation process (30).

In step 301, a candidate string having a large likelihood from among the candidate strings from the point P to the end of the input character string is set in the display buffer as a conversion result. If there are a plurality of candidate strings having the same likelihood, a learning table which stores compound words and independent words with affixes from the already determined character strings is referred to in the storage unit 3 and is referred to in this table. Is selected as the conversion result. If the above is not determined uniquely, the candidate string having the largest sum of the independent word lengths is selected. Another way to select the conversion result is
There is also a method using the frequency of words, etc., and another selection method can be created by arbitrarily combining the factors described above.

In step 302, another candidate group starting from point P is set in the display buffer so that the user can select it.

In the case of the example shown in FIG. 4, when P is 1, “on mathematical analysis” is selected as a conversion result, and as a candidate group for selection,
"Mathematical", "number" and "suck" are extracted and set in the display buffer. Further, the conversion result and the corresponding portion of the input character string are highlighted to clearly indicate the character string portion currently being selected. (Display (40)).

FIG. 7 shows screen information displayed on the input / output unit 1 immediately after the input (10) is completed, and FIG. 8 shows screen information displayed after the display candidate creation process (30) is completed. In this screen, the bottom is an input line, the top is a conversion result output line, and the bottom right is a selection candidate display area.

The user looks at the conversion result on the screen and makes a selection (50).
The selection is performed by using a keyboard to select an object from the selection candidate group. If the conversion result highlighted at the top of the screen is correct, it can be selected by pressing a specific key. is there.

When the selection (50) is performed by the user, the control unit 2 next executes the operation of the confirmation process (60).

FIG. 6 is a flowchart showing the operation of the determination process (60).

In step 601, a mark is added to the network data in the storage unit 3 corresponding to the candidate selected and instructed by the user, the length of the selected candidate is added to P, and the start position of the subsequent undecided part is determined. Show.

In step 602, the selected candidate is read and stored in the learning table on the storage unit 3.

The control unit 2 re-evaluates the likelihood of the network of the candidate selected in the determination process 60 by the display candidate creation process 30 and the candidate after the point P, and the other conversion results obtained from the point P and other conversion results. Set the candidate group in the display buffer. FIG. 9 shows the screen display information after the user selects "math". The highlighting has moved to the "Analysis" section after "Mathematics". The candidate groups are “analysis”, “meeting”, “world”,
The “floor”, “family”, “ka”, and “ka” are displayed.

When the user selects “meeting” instead of “analysis” from among the candidates, as shown in FIG. 10, the conversion result after the “meeting” is displayed on “seat” obtained from the network of FIG. It will be changed automatically, and the selection candidates will be “Seat”, “Seki”,
"Cough" is displayed. Since the adjacency between candidates is maintained in this way, when a segment break is changed by selection (50), the subsequent display can be corrected at the same time, and the man-hours for subsequent corrections can be reduced. Can be reduced.

The learning table is also updated in the confirmation process (60) when the user selects "kai", but in this case, "math" has already been registered and "kai" is a suffix. From this, "Math" is added to the previously registered "Mathematics", and re-registration is performed in the form of "Mathematical Society". FIG. 11 shows the contents of the table after "Mathematical Society" has been registered.

The control unit 2 repeatedly executes the display candidate creation processing (30) and the determination processing (60) until there is no undetermined part. When all the conversion results have been determined, the process ends.

When the above-described series of processing is completed, the network data in the storage unit 3 is deleted, but the data in the learning table is retained. If the same character string is input thereafter, the data is preferentially adopted as the conversion result. Therefore, the conversion accuracy can be increased depending on the use. In the past, the learning function associated with conversion was used, but even if prefixes and suffixes were learned by themselves, comrades with the same sound meaning often appeared in the same sentence. Dropping was often a problem. Also, when a homonym appears, such as "news reporter is on a train", it was learned in units of reading "kisiya", so it was not possible to distinguish between "reporter" and "train" Was. According to the present invention, since the prefix and the suffix are learned together with the independence to which they are attached, and the compound word is learned in a long unit without being decomposed into the independent word, there is an effect that the accuracy can be further improved by learning.

〔The invention's effect〕

According to the present invention, since a candidate having a high likelihood of conversion certainty is held, another candidate can be immediately selected even if there is an error in the conversion result, and the correction can be speeded up. . By performing the processing for creating and holding the candidates in parallel during the input operation of the user, it is possible to further reduce the conversion waiting time of the user. In addition, the conversion accuracy can be improved by a compound word and an affixed self-sufficient learning function.

[Brief description of the drawings]

FIG. 1 is a block diagram of a kana-kanji conversion device according to the present invention, and FIG.
FIG. 3 is a diagram showing a method of executing the conversion, FIG. 3 is a flowchart showing an operation of the conversion process, FIG. 4 is a diagram showing an example of data, FIG. 5 is a display candidate creation process, and FIG. FIGS. 7 to 10 are diagrams showing screen display information, and FIG. 11 is a diagram showing an example of the contents of a learning table. 1 ... input / output unit, 2 ... control unit, 3 ... storage unit.

 ──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-57-14971 (JP, A) JP-A-58-115528 (JP, A) JP-A-56-17468 (JP, A) 129633 (JP, A) JP-A-58-19963 (JP, A) JP-A-59-121529 (JP, A)

Claims (4)

(57) [Claims] 1. A character string input means for inputting a kana character string which includes a plurality of phrases and is not separated in units of phrases, a storage means for storing the input kana character string, and a kanji character A conversion instructing unit for converting kana-kanji into a kana-mixed sentence; and a plurality of phrases cut out from the kana character string by the kana-kanji conversion instruction, and at least one conversion candidate starting with the first character of each phrase. A conversion means for extracting kana-kanji conversion by extracting the candidate kana-kanji character, and a display means for displaying a candidate character string composed of a conversion candidate having the highest probability for each phrase among the extracted conversion candidates. A phrase break indicating means for reselecting a conversion candidate constituting the candidate character string and indicating a break in a phrase of the input kana character string; and the kana sentence stored in the storage means. In the column, control means for causing the display means to display a conversion candidate obtained by re-evaluating the likelihood of conversion for all undetermined parts subsequent to the break of the indicated phrase, and Kana-kanji conversion device. 2. The kana-kanji conversion apparatus according to claim 1, wherein said storage means stores, together with said kana character string, a conversion candidate having a likelihood of certainty higher than a predetermined value among said conversion candidates. 3. The kana-kanji conversion device according to claim 2, wherein the likelihood is determined by the number of weighted components determined for each part of speech category of the components in the phrase. 4. The storage device according to claim 2, wherein said storage means stores said conversion candidates in the form of a network.
The kana-kanji conversion device described in section.