JPS59123926A - Voice input japanese processing system - Google Patents

Abstract

PURPOSE:To obtain a correct KANJI (Chinese character) and KANA (Japanese syllabary) mixed writing by preparing KANA strings as the 1st candidata closest to an input voice and another approximate candidate, and performing a grammatical check, maximum likelihood evaluation, etc., upon a KANA string made into morphemic form and outputting the KANA and KANJI mixed writing. CONSTITUTION:When a Japanese sentence to be printed out on a printer 26 is voiced in front of a microphone 10, a voice recognition part 14 recognizes it, syllable by syllable, and passes it to a KANA-KANJI conversion part 16 by KANA. Word dictionary consultation and decision on whether grammatical connection relation, therefore, words in morphemic form are proper or not are performed here to perform convesion to the KANA and KANJI mixed writing. Then, a document editing part 18 perform editing operations such as a line change, page change, sentence replacement, etc., and the edited KANA and KANJI mixed writing is displayed on a display device 20 through a voice synthesis/output control part 24 and further printed out on a printer 26 by an indication from a console panel 12, etc. Further, the input voice is recognized to prepare an approximate candidate in addition to the candidate closest to the right answer, and it is also checked to output the KANA and KANJI mixed writing close to the right answer.

Description

DETAILED DESCRIPTION OF THE INVENTION Technical Field of the Invention The present invention relates to a voice input Japanese processing method in which Japanese is input by voice and is converted into a sentence mixed with kana and kanji by a word processor.

Prior Art and Problems Traditionally, word processors with keyboard input have been the main way to create documents. Of course, for keyboard input, the operator needs to be proficient in keyboard operations. In response, a method has been developed that incorporates voice, the most natural means of information transmission for humans, into a man-machine interface, but this method displays the voice recognition results in kana on the display, leading to misunderstandings. ), have the speaker confirm it, make corrections by selecting the next candidate if necessary, convert the kana string obtained in this way to kana-kanji in the same way as a normal word processor, and at that time also check and make corrections by the operator. Go through the process of doing this. However, this method requires operations such as checking and correcting kana characters, and is not as easy to use as a keyboard-input word processor.

Purpose of the Invention The present invention is a method for processing voice input Japanese that does not require the speaker to confirm the kana sequence, although the process of converting the voice input into a kana string and converting the kana string into a sentence mixed with kana and kanji. This is what we are trying to achieve.

Structure of the Invention The present invention is an audio input Japanese processing method that recognizes input speech, converts it into a kana string, and converts the kana string into a kanji-kana mixed sentence. In addition to the kana sequence, if there are any kana whose approximation is not much different from the first candidate kana and the first candidate kana sequence has a pronunciation that differs from the kana-specific one, the second candidate kana is a collection of those kana. Prepare columns, perform morphemization on these kana strings using a word dictionary index, perform one or more of grammar checking, maximum likelihood evaluation, and linguistic evaluation on the variously morphemized kana strings, and perform these evaluations. The present invention is characterized in that the output data is a highly morphological Kanji-Kana mixed sentence, which will be explained in detail below with reference to the drawings.

Embodiment of the Invention FIG. 1 shows an outline of a voice input word processor, and 10
1 is a microphone, 12 is an operation console, 14 is a voice recognition unit, 1
6 is a kana-kanji converter, 18 is a document editor, and 20 is a CRT.
22 is a speaker, 24 is a voice synthesis/output control unit, and 26 is a printer. When a Japanese sentence to be printed out by the printer 26 is uttered into the microphone 10, the speech recognition section 14 recognizes it for each syllable (that is, kana) and passes it to the kana-kanji conversion section 16 as a kana string. At this point, a word dictionary index, grammatical connection relationships, and validity determination of the word are performed, and the sentence is converted into a sentence containing kana and kanji. The document editing section 18 performs line breaks, page breaks, text replacement, and other editing.The edited kana-kanji text is sent to the display 20 via the speech synthesis/output control section 24.
, and may be printed out from the printer 26 based on instructions from the operation console 12 or the like. In the conventional method, the syllables, that is, kana, recognized by the speech recognition unit 14 are displayed on the display 20.
is displayed, asking the speaker for confirmation and correction instructions, but since this is not necessarily necessary in the method of the present invention, what is displayed is the output of the text editing section 18 after the kana-kanji conversion. The basics of a speech input Japanese processing device are a speech recognition section and a kana-kanji conversion section, and the processing device can be configured by combining these parts. Both the speech recognition section and the kana-kanji conversion section are at a practical stage if they are completely or slightly narrowed down.

Note that some speech recognition units perform word recognition in addition to monosyllable recognition, but in this example, the former is used.

Kana-Kanji conversion compares the input kana character string with a word dictionary, generally performs part-of-speech decomposition using the longest match method, and assigns kanji or kana (generally hiragana) to the word corresponding to the kana character string. However, in the present invention, this is expanded as follows.

That is, in the present invention, the input voice is recognized and the answer closest to the correct answer (
In addition to the first candidate kana string (which has a short distance to the reference pattern), other candidates whose distances are less than or equal to the dark value by comparing the distance between the candidates with the dark value (knowing the degree of approximation information) are also picked up. That is, since recognition errors are inherent in voice recognition, a plurality of candidates are used to deal with these recognition errors, but as the number of candidates increases, a large number of combinations are created, which becomes extremely troublesome. For example, even in a two-word sentence, if each word has a first candidate and a second candidate, 2×2=4 different sentences will be created. In order to avoid this complexity, the degree of approximation (distance) is taken into consideration, and only candidates with a high degree of approximation are selected, and candidates with a low degree of approximation are discarded. Another option to adopt is to use the corresponding lowercase letter (for example, when pronouncing the ``tsu'' in ``Harashin'', first specify the lowercase letter by operating the console, etc.) for a kana string that is uttered without specifying a lowercase letter. '', but the kana sequence pronounced as ``tsu'' without doing this is pronounced as ``
"Tsu" for "Harashin". (lower case)
If the corresponding kana is in a kana string, the lowercase letter is used as another candidate (because it is a specific kana such as yu, yo, etc.), and the other character whose pronunciation and pronunciation are different or cannot be distinguished (for example, particles are In the document, they are wa, o, and eh, so we have wa, wa, and eh as alternative candidates for these, and since zutozu and jitoji are indistinguishable in pronunciation, one is an alternative candidate for the other. )and so on. For these other candidates, we assign analysis routes and priorities each time, and each route is analyzed individually by comparing character strings divided into morphemes (equivalent to dictionary headwords) or more with a word dictionary. The validity of the morpheme sequence is linguistically checked by dividing the morphemes into morphemes.For example, in the case of ``The present invention is'' and the morphemes that appear in one sentence, it is possible to connect the morphemes immediately before and after the morphemes, but in Japanese it is not possible. After pruning the root, assigning the corresponding kanji and kana, and changing the priority order of the root through other linguistic processing. (for example, the priority order of the root is evaluated and changed in terms of the presence or absence of unused words, the number of morphemes, etc.), and finally a string containing kana-kanji mixture of the root with the first priority is output. Furthermore, if necessary, the Kana/Kanji mixed string is converted back to a Kana string by converting it into a Kana string, and its validity is verified by comparing it with the root input source (which should originally match). Furthermore, if the desired sentence containing kana, kanji, and kanji cannot be obtained in one processing, reprocessing is performed using a reconversion command.

FIG. 2 is an explanatory diagram showing an overview of the above processing in a block diagram. The input such as a kana string from the speech recognition unit 14 is first analyzed as a command, and it is checked whether the command attached to the input is a document creation command, and if yes, it is sent to conversion route A, and if it is a document check, it is sent to reconversion route B. It will be done. conversion rule)
In A, input processing is first performed, that is, data is read, and then multi-route setting is performed by evaluating the degree of approximation of candidates.
Kana-Kanji conversion and linguistic checks for multiple routes, change of route ranking by linguistic checks, and verification of the best route by Kanji-Kana conversion are performed one after another, and then output. In the re-conversion route B, the route priority order is directly changed by linguistic checking, and the best route is verified by Kanji-kana conversion, and then output processing is performed.

FIG. 3 is a flowchart showing an outline of the processing of phonetic kana-kanji conversion. Read the input character string, that is, the voice-recognized kana string, cut out the processing units using punctuation marks, etc., and extract them from the recognition information.
Also configure multi) routes. Next, morpheme extraction,
It performs word dictionary access, grammatical connection check, and maximum evaluation check, and performs these for all routes, evaluates the routes, and outputs the appropriate one. Step (e
).

tf), when there is no valid morpheme, backtracking is performed. Steps marked * are additional parts of the invention. Next, a specific example of the process will be explained with reference to FIG.

In Figure 4, it is assumed that ``Calculate the total amount'' was inputted by voice, and the first candidate of the voice recognition result was ``Sougaouo, Kenigo nsuru,'' and the second candidate was ``Ofukakuho, Heihakuus.'' It is said that Note that each character of the second candidate corresponds to each character of the second candidate, and has the second highest degree of approximation. Since the voice input is syllable by syllable, it is input by saying "so", "tsu", "ga", etc., and the comma "," is input by saying the word "comma". Since the word input is limited to punctuation marks and other specific words (control words, etc.), the recognition rate is high and a second candidate is not necessary.

Next, extract the conversion unit and import the first kana sequence separated by commas. Next, the similarity information between the candidates is evaluated. In this evaluation, in this example, the degree of approximation of the second candidate to the first candidates G, T, G, O is low (the distances Ll, L2.L3°f, 4 of the first candidates G, T, G, O) are low. 2 candidates,
), force, distance Jl.

I12. 7!3. eg difference + L+, J2 L2
.

Next, we start setting the route and assigning priority, in this example,
``Sougawo'' is ranked 1st, and ``Sogakuo'' is ranked 2nd. The kana string with rank 1 leads to the prisoner's route, and the kana string with rank 2 leads to the route of time. These A.

Route B performs kana-kanji conversion and checking.

The processing contents for routes A and B are the same: first, a word dictionary is looked up, and corresponding words are selected in descending order of accuracy using the longest match method. Since this is the longest match, it is also assumed that the dictionary headword (morpheme) and the kana string have the highest probability of character matching from the beginning to the end as far as possible. In this example, the morphemes selected in this way were ``Uo'', ``Claw bud'', ``So'', and ``Total'' for the A route, as shown in the diagram. Perform a sentence 0 law check on these. The word dictionary has a part of speech for each morpheme, and there is also connection information indicating which part of speech this morpheme can be connected to, so it is checked whether the morphemes can be continuous with each other. Next, a maximum likelihood evaluation is performed, and a high rating is given to items with a large length (length and number of characters) and items with a large usage tendency.

Next, we will conduct a censored evaluation. Kana strings are converted into morphemes based on the longest match, but the longest match is not necessarily the correct answer.
If morphemes are erroneously converted, subsequent morphemes will not be connected. When we enter a dead-end branch, we discontinue the morphemization in that branch, change the previous morphemization to another one, two, etc., and redo the analysis.
Efforts will continue to arrive at the correct answer, but this evaluation provides a criterion for deciding where to cut off. The branches are evaluated based on depth, reading length, etc., and those whose evaluation values exceed the set values are discontinued. The next linguistic evaluation was mentioned above, but I would like to add that it is possible to make use of the characteristic that in Japanese, one character in hiragana does not have more than four consecutive morphemes, and one character in kanji does not have more than 21 consecutive morphemes. evaluate. For example, one character in hiragana is one morpheme, and multiple characters can be connected, but when connected, it becomes one morpheme, such as ``neta'' (shi).
It is judged that four or five independent morphemes that become morphemes and do not form a single morpheme even if they come together cannot continue. In addition, the words 日本, 本, and 国 are all words with one morpheme, as they are each found in dictionary headwords, and there is also the word ``Japan,'' which is a combination of these words, but this creates one morpheme with ``Japan.'' It is determined that there are no consecutive three or four independent morphemes that do not form one morpheme. For example, ``Subekudoryokusuru'' is changed to ``Subekudoryokusuru.''
When converted into morphemes, the kanji morphemes ``ku'', ``degree'', ``
It is strange because there are three ``power'' in a row, and they cannot be combined to form a single morpheme.This morpheme is judged to be incorrect.

Next, determine the highest priority route, in this example, is route A better?
Decide whether route B is better. Check whether it contains unused words and whether it has a large number of morphemes. A
If the root "Sugata" is a word that has never been used in this system, there will be unused words, and if the "Total amount" of route B is a word that has been used, there will be no unused words. , the latter is rated higher. In addition, the number of morphemes is 3 for root A, and the total number of morphemes for route B is 2, so the smaller the number of morphemes, the better. From these results, the kana string of route A had a higher degree of approximation to the input speech, but the analysis results within the chain line frame showed that the kana string of route B had a higher evaluation, and the priority was higher than that of route B. "Total amount" will be changed to ■, and "Sugata Uo" for route A will be changed to ■. Next, priority ■ "Total amount" is "
Since this is equivalent to the input data, it is determined that no error was introduced during the analysis, and the output data is ``Total amount''. Note that ``wo'' was determined to be a particle when the ``total amount'' was determined, and ``o'' was changed to ``wo''.

For example, if you convert ``Guiyo Ryoujiki Bubble'' into kana-kanji, it becomes ``substitute consul arrival bubble.'' In kanji-kana conversion, the morphological analysis is ``substitute (dairou)'' 3 consul (ryoji) arrival (jaku) bubble (bubble), so the conversion result is ``guiro ryoji jaku bubble''. In such cases, checking using Kanji-kana conversion is effective.

When performing dictionary indexing and morphemization, if the input kana string can be divided into parts of speech such as nouns, particles, and verbs, indexing becomes easier, the recognition rate increases, and the time required for recognition decreases. However, requiring the speaker to input information by part of speech impairs the simplicity and naturalness of voice input.

Therefore, it is effective to detect the phrases that the speaker makes naturally and use them as classification information. Specifically, the time of the input voice is monitored, and if the voice interruption time exceeds a set value, it is cut off. In the example above, the speaker probably uttered ``sou ga kuo'' in succession and did not say ``sou ga'', ``ri'', and ``o'', so the phrase up to ``sou ga kuo'' is treated as a series. Since the cut code is added after O), there is a high probability that you can easily find the ``total amount'' by finding the longest match.

In this example, there is a comma, so a punctuation code is not needed, but a punctuation code is effective in a sequence that has a large number of kana characters up to the punctuation mark.

As described in detail, the present invention outputs a correct kanji-kana mixed sentence by setting multiple routes and checking the suitability of each route, so that incorrectly recognized voice input kana strings can be translated into kanji with high accuracy. It can convert kanji to kanji, and there is no need for the speaker to check the kana sequence, and the kanji-kana mixed sentence can be immediately displayed on the display and asked for confirmation. Furthermore, when setting multi-routes, routes with a low degree of near-i are removed from the candidates, and routes that are easily misunderstood or cannot be distinguished are actively added to the candidates, so that an appropriate number of multi-routes can be secured. Furthermore, in the aptitude check for each route, we added unused words and the number of morphemes to the check items, so we can expect an improvement in the probability of selecting the correct route.

[Brief explanation of the drawing]

FIG. 1 is an explanatory diagram showing an outline of a voice input Japanese word processor, and FIGS. 2 to 4 are flow diagrams explaining the processing method of the present invention. 5 In the drawing, 10 is a microphone, 12 is an operation console, 14 is a voice recognition section, 16 is a kana-kanji conversion section, and 20 is a display. Applicant Fujitsu Limited Representative Patent Attorney Minoru Aoyagi6

Claims (1)

[Claims] In a speech input Japanese processing method that recognizes input speech, converts it into a kana string, and converts the kana string into a kanji-kana mixed sentence, a first candidate closest to the input speech is selected. In addition to the kana sequence, if there are any kana whose approximation is not much different from the first candidate kana and the first candidate kana sequence has a pronunciation that differs from the kana-specific one, the second candidate kana is a collection of those kana. Prepare columns, perform morphemization on these kana strings using a word dictionary index, perform one or more of grammar checking, maximum likelihood evaluation, and linguistic evaluation on the variously morphemized kana strings, and perform these evaluations. A voice input Japanese processing method characterized in that output data is a highly morphized kanji-kana mixed sentence.