WO2024042963A1

WO2024042963A1 - Error correcting translation device, error correcting translation method, program and storage medium for same

Info

Publication number: WO2024042963A1
Application number: PCT/JP2023/027192
Authority: WO
Inventors: 史昭菅谷; 将夫内山; 英一郎隅田; 照二小林
Original assignee: マインドワード株式会社; 国立研究開発法人情報通信研究機構
Priority date: 2022-08-26
Filing date: 2023-07-25
Publication date: 2024-02-29
Also published as: JP2024031617A

Abstract

This error correcting translation device comprises: a first terminal into which text in a first language is input; a text divider which divides the text into chunks to generate divided text; a translation machine which individually translates each chunk included in the divided text to generate translated chunks in a second language; and an error correction control unit which causes each translated chunk from the translation machine to be displayed individually on a second terminal. The second terminal accepts input of an identifier specifying a mistranslated chunk, and the error correction control unit causes the mistranslated chunk to be displayed on the first terminal in a different manner to other translated chunks, on the basis of the identifier input into the second terminal, and causes the first terminal to accept input of paraphrased text of the text in the first language in response to the display of the mistranslated chunk.

Description

Error correction translation device, error correction translation method, program, and storage medium thereof

The present invention relates to an error correction translation device, an error correction translation method, a program, and a storage medium thereof.

Conventionally, there is a known sentence division method for dividing input sequence data for simultaneous interpretation using machine translation.

For example, Patent Document 1 discloses a data segmentation device that can perform sequence data segmentation processing (for example, text segmentation processing) in real time.

JP2020-24277A

However, in the prior art, it has been difficult for the speaker and the listener to have a common understanding of the scope of errors that occur during interpretation. As a result, error correction conversations became complicated and redundant.

The present invention has been made in view of the above points, and is an error correction method that can reliably feed back errors in simultaneous interpretation to the source language side, makes error correction efficient, and improves the quality of conversation. The purpose of the present invention is to provide a translation device, an error-correcting translation method, a program, and a storage medium thereof.

An error correction translation device according to an embodiment of the present invention includes:
a first terminal into which text in a first language is input;
a text divider that divides the first language text into chunks to generate divided text;
a translator that individually translates each chunk included in the divided text to generate translated chunks in a second language;
a second terminal;
and an error correction control unit that causes each of the translation chunks from the translator to be individually displayed on the second terminal.

The second terminal receives an input of an identifier specifying a mistranslated chunk among the translated chunks,
The error correction control unit causes the mistranslation chunk to be displayed on the first terminal in a manner different from other translation chunks based on the identifier input to the second terminal, and the error correction control unit causes the mistranslation chunk to be displayed in a manner different from other translation chunks. In response, the first terminal is configured to accept input of paraphrased text of the text in the first language.

An error correction translation method according to another embodiment of the present invention includes:
inputting text in a first language into a first terminal;
a text dividing step of dividing the first language text into chunks to generate divided text;
a translation step of individually translating each chunk included in the divided text to generate translated chunks in a second language;
an error correction control step of individually displaying each of the translation chunks from the translator on the second terminal;
receiving, at the second terminal, an input of an identifier specifying a mistranslated chunk among the translated chunks;
a mistranslation chunk displaying step of displaying the mistranslation chunk on the first terminal in a manner different from other translated chunks based on the identifier input to the second terminal;
after the mistranslation chunk displaying step, receiving input of a paraphrase text of the text in the first language to the first terminal;
has.

An error correction translation program according to yet another embodiment of the present invention includes:
computer to translate text data,
a first terminal means into which text in a first language is input;
text dividing means for dividing the first language text into chunks to generate divided text;
Translation means for individually translating each chunk included in the divided text to generate translated chunks in a second language;
second terminal means;
functioning as error correction control means for individually displaying each of the translation chunks from the translation means on the second terminal means;
The second terminal means receives an input of an identifier specifying a mistranslated chunk among the translated chunks,
The error correction control means causes the first terminal means to display the mistranslated chunk in a manner different from other translation chunks based on the identifier input to the second terminal means, The method is characterized in that, in response to the display, the first terminal means accepts input of paraphrased text of the text in the first language.

1 is a block diagram showing the configuration of an error correction translation device 10 according to a first embodiment of the present invention. FIG. 3 is a flowchart showing a flow of error correction control executed by an error correction control unit 25. FIG. FIG. 3 is a diagram schematically showing a display screen of the first terminal 11 on which spoken text is displayed. FIG. 3 is a diagram schematically showing a display screen of the second terminal 21 on which translation chunks are separated and displayed in individual frames. FIG. 6 is a diagram schematically showing the display screen of the second terminal 21 when the designation of mistranslated chunks is received. FIG. 3 is a diagram schematically showing a display screen of the first terminal 11 on which mistranslated chunks are highlighted. FIG. 7 is a diagram schematically showing another example of the display screen of the first terminal 11 on which mistranslated chunks are highlighted. It is a figure which shows typically the display screen of the 2nd terminal 21 on which the paraphrase text was translated and displayed. 3D is a diagram schematically showing a modification example of the display of the first terminal 11 shown in FIG. 3D. FIG. 2 is a block diagram showing an example of the configuration of a text divider 13. FIG. It is a figure showing the result of morphological analysis of Japanese text using MeCab. It is a figure which shows the morphological analysis result of the English text using NLTK. 3 is a flowchart showing a chunk boundary determination processing procedure executed by the threshold value determination manager 34. FIG. FIG. 2 is a block diagram showing the configuration of an error correction translation device 30 according to a second embodiment of the present invention. FIG. 3 is a block diagram showing the configuration of an error correction translation device 50 according to a third embodiment of the present invention.

In the following, preferred embodiments of the present invention will be described, but these may be modified and combined as appropriate. Further, in the following description and the accompanying drawings, substantially the same or equivalent parts are designated by the same reference numerals.

[First embodiment]
FIG. 1 is a block diagram showing the configuration of an error correction translation device 10 according to a first embodiment of the present invention.

As shown in FIG. 1, the error correction translation device 10 is configured as a translation device that translates a conversation of a Japanese (first language) speaker to an English (second language) speaker. However, two-way simultaneous translation of Japanese (first language) speaker's conversation to English (second language) speaker's and English speaker's conversation to Japanese speaker's It may be configured as a translation device for an interpreter.

The first language (original language) and the second language (translation language) are not limited to Japanese and English, and can be applied to other languages as well.

The error correction translation device 10 includes a first terminal (terminal #1 in the figure) 11 which is a tablet terminal (input/output device), a text divider 14, and a Japanese-English/English-Japanese translator (hereinafter simply referred to as a translator). ) 15, a second terminal (terminal #2 in the figure) 21 which is a tablet terminal (input/output device), and an error correction control section 25.

The error correction control unit 25 also includes a first terminal controller 12 , a divided text correction manager (hereinafter simply referred to as a correction manager) 13 , and a second terminal controller 22 .

The first terminal 11 is equipped with a screen keyboard (KB) 11A. Kana characters input from the screen keyboard (hereinafter simply referred to as the keyboard) 11A by the Japanese speaker 1 are converted into Japanese text data (sequence data) by a processor (not shown) in the first terminal 11. is output from the first terminal 11 as . At this time, kana-kanji conversion, which is a known technique, may be used for the kana characters input from the keyboard 11A, if necessary.

The first terminal 11 also has a display that receives and displays data from the first terminal controller 12. Furthermore, the display has a function as a touch panel that receives operations from the Japanese speaker 1.

Similarly, the second terminal 21 is equipped with a keyboard (KB) 21A. Characters input from the keyboard 21A by the English speaker 2 are output from the second terminal 21 as English text data (sequence data) by a processor (not shown) in the second terminal 21.

The second terminal 21 also has a display that receives and displays data from the second terminal controller 22. Furthermore, the display has a function as a touch panel that receives operations from the English speaker 2.

Below, a case of simultaneous interpretation in which a conversation between a Japanese speaker and an English speaker is translated will be specifically explained with reference to the diagram. FIG. 2 is a flowchart showing the flow of the error correction translation method executed by the error correction control unit 25. Further, FIGS. 3A to 3G are diagrams schematically showing display screens of the first terminal 11 and the second terminal 21.

In the following, a case will be described in which a text consisting of one sentence is input to the error correction translation device 10 and translated.

The Japanese speaker 1 inputs the text "I'm going to Tokyo tomorrow" (hereinafter also referred to as uttered text) using the keyboard (KB) 11A. The first terminal 11 sends the input spoken text to the first terminal controller 12 .

The first terminal controller 12 generates the following data signal DS1 from the input text received from the first terminal 11 (step S1), and sends it to the correction manager 13. The first terminal controller 12 also sends the data signal DS1 to the first terminal 11, and as shown in FIG. 3A, the first terminal 11 receives the spoken text data [ 'I'm going to Tokyo tomorrow'] is displayed in the format of the text widget 11W (step S2).

Note that an example will be described in which the JSON (JavaScript Object Notation) format is used as the data format of the data signal.

Data signal DS1: {'id':10, 'jap': ['I'm going to Tokyo tomorrow']}
Here, 'id' is the ID of the utterance, and 'id' is an integer greater than or equal to zero such that id = {0, 1, . . .}. Therefore, "'id':10" means the 11th utterance.

The correction manager 13 of the error correction control unit 25 sends the data signal DS1 to the text divider 14 (step S2). The text divider 14 divides the spoken text data included in the data signal DS1 into chunks (partial sentences). The operation of the text divider 13 will be described in detail below with reference to FIG.

FIG. 4 is a block diagram showing an example of the configuration of the text divider 13. A text ST (indicated as <ST> in the figure) input to the input end 13I of the text divider 13 is divided into words by the word divider 31.

The word divider 31 performs, for example, morphological analysis to divide the text ST into words w ₁ to w _m . Each word obtained by the division is sequentially held as register values wr _{1 to wr m in registers r 1} _to _r _m of a word register 32, which is a FIFO (First In First Out) register. That is, the word w _i (i=1, 2, ·, m) is held in the register r _j (m is an integer of 1 or more).

Specifically, MeCab, which is a publicly known open source morphological analysis engine, was used. FIG. 5A shows the results of morphological analysis of Japanese text using MeCab.

When the Japanese text ``I'm going to Tokyo tomorrow'' is input to the word divider 31, it is divided into five words by morphological analysis, for example, as shown in FIG. 5A. Note that "nouns" and the like in the diagram indicate parts of speech in Japanese grammar. Furthermore, EOS (End Of Sentence) is a symbol indicating the end of the divided text.

That is, in the case of this embodiment, m=5, and the word sequences divided by the word divider 31 are 'Asu', 'Tokyo', 'ni', 'Go', and 'Masu'. .

Note that as the English word divider 31, a well-known morphological analysis technique such as NLTK (Natural Language Toolkit) can be used. FIG. 5B shows the results of morphological analysis of an English text using NLTK.

In this case, when the English text 'I'm going to Tokyo tomorrow' is input to the word divider 31, for example, as shown in FIG. 5B, 'I', 'm', 'going', 'to' , 'Tokyo', and 'tomorrow'. Note that PRP in the diagram indicates the part of speech in English grammar. Also, EOS is a symbol indicating the end of the divided text.

Next, with reference to the flowchart of FIG. 6, the chunk boundary determination processing procedure executed by the threshold value determination manager 34 will be described.

As shown in FIG. 4, threshold values θi (i=1, 2, . . . , m) are stored in advance in the threshold value register 35 in correspondence with each of the registers r ₁ to r _m of the word register 32.

When the register values wr ₁ to wr _m are held in the word register 32, the probability that the end of the word w _m (i=1, 2, ..., m) is the division position is determined by deep learning by the deep learning device 33. calculated. Let this probability be P(wr _i ).

Note that the probability P(wr _i ) (i=1, 2, ·, m) that the text is divided at the end of each word is given so that the sum of the probabilities that the word series is divided is 1.

First, the initial value i is set to 1 (step S21), and subsequent steps S22 to S25 are executed.

In step S22, the comparator 36 determines whether P(wr _i )>θi. If it is determined that P(wr _i )>θi, the end of the word w _i is determined to be the division position, that is, the boundary (chunk boundary) where the spoken text is divided into chunks (step S25). . If it is determined that P(wr _i )>θi is not satisfied, the process advances to the next step S23.

In step S23, it is determined whether the current word is the last word of the text (i≧m). If it is determined that the current word is the last word of the text, the process proceeds to step S23, where the end of the word is determined to be the division position.

If it is determined in step S23 that the current word is not the last word of the text, the process proceeds to step S24, where i is incremented by 1, and the above steps are repeated for the next word.

If the chunk boundary determination process has been completed up to the last word, this process ends (END).

The divided text output device 37 adds a division mark (separator) to the series of words w ₁ to w _m based on the chunk boundary determined by the threshold value judgment manager 34 and outputs the result.

For example, when adding a division mark represented by '<SEP>' between word w ₄ and word w ₅ , w ₁ , w ₂ , w ₃ , w ₄ , '<SEP>', w ₅ , . _.

That is, the text divider 14 divides the input text ST ("I'm going to Tokyo tomorrow") using a division engine, and creates the divided text STT as '<STT>Tomorrow<SEP>I'm going to Tokyo<SEP>'. is generated and output (FIG. 2, step S3).

Further, the divided text STT is sent to the translation machine 15 via a divided text correction manager (hereinafter simply referred to as a correction manager) 13. In the translator 15, the '<SEP>' mark is used as a delimiter in translation processing.

Specifically, the text divider 14 divides the uttered text data ['I'm going to Tokyo tomorrow]' in the data signal DS1 into two chunks 'Tomorrow' and 'I'm going to Tokyo', and generates the following data. A signal DS2 is generated and sent to the segmented text correction manager 13.

Note that in this specification, the term "chunk" may be any group of words that can be translated. Therefore, not limited to the above case, for example, the text 'I'm going to Tokyo tomorrow' may be divided into three chunks: 'Tomorrow', 'To Tokyo', and 'I'm going', or 'Tomorrow'. , 'Tokyo', 'ni', and 'go'.

Data signal DS2: {'id':10, 'jap': ['tomorrow', 'I'm going to Tokyo']}
The correction manager 13 also has a function of transmitting the data signal DS2 to the first terminal controller 12 in response to a request signal from the first terminal controller 12.

The divided text correction manager 13 sends the data signal DS2, which is the output of the text divider 14, to the translator 15, and translates the two chunks ``Asu'' and ``I'm going to Tokyo'' included in the data signal DS2, respectively.

The translator 15 translates each of the two chunks into English. More specifically, the translator 15 outputs the translation result for the input according to the function translate (input, 'dir'). Here, 'dir' defines the direction of translation. Here, 'dir' = 'je' to specify translation from Japanese to English. In other words, Japanese-English translation uses translate(input, 'je') as a function, while English-Japanese translation uses translate(input, 'ej') as a function.

The translator 15 translates the 'jap' element of the data signal DS2 according to the following procedure, and generates the data signal DS2A.
output＝translate('jap': ['tomorrow', 'I'm going to Tokyo'], 'je')
= translate('jap':['tomorrow'], 'je') + translate('jap': ['I'm going to Tokyo'], 'je')
='eng': ['ASU'] + 'eng: ['I go to Tokyo']
='eng': ['ASU', 'I go to Tokyo']
Data signal DS2A: 'eng': ['ASU', 'I go to Tokyo']

The correction manager 13 incorporates the data signal DS2A into the data signal DS2, generates the following data signal DS3 (step S4), and sends it to the second terminal controller 22 (FIG. 1).

Data signal DS3: {'id':10, 'eng': ['ASU', 'I go to Tokyo']}
That is, the data signal DS3 includes translated chunks (referred to as translated chunks) 'ASU' and 'I go to Tokyo' of the two chunks included in the data signal DS2.

The second terminal controller 22 sends the data signal DS3 to the second terminal 21, and the second terminal 21 receives the two translation chunks 'ASU' and 'ASU' included in the data signal DS3, as shown in FIG. 3B. 'I go to Tokyo' is separated and displayed in a separate frame.

More specifically, each translation chunk is displayed as an image display component called a text widget 21W surrounded by a frame line. The text widget 21W displays the text of the translation result, and the second terminal 21 receives a designation (mistranslation designation) of mistranslation or meaningless chunks from the user (English speaker 2) (step S5).

When the English speaker 2 finds any mistranslation or meaningless chunk in any of the translated chunks, the English speaker 2 can remove the mistranslation or meaningless chunk (hereinafter referred to as mistranslation chunk) by touching the text widget (frame) 21W. specify. The specified text widget 21W is changed from the standard display to the specified display.

Note that in this specification, the term "mistranslated chunk" includes a translated chunk that is mistranslated or has no meaning due to an error in the context judgment of the translator.

In step S5, if the English speaker 2 determines that there is no mistranslated chunk among the translated chunks, for example, the English speaker 2 can simply touch the SEND button, etc., and the designation (identifier) of the mistranslated chunk can be sent to the data. Not included in the signal.

In this case, the error correction control unit 25 determines that there is no mistranslation chunk specification, and control shifts to RETURN, and shifts to processing of the next spoken text (step S1) (step S6, No).

In this case, in step S1, the second terminal 21 is controlled to receive the spoken text (English text) in response to the previous spoken text (Japanese text), but the first terminal 11 Control may be performed to receive the next new spoken text (Japanese text) from.

On the other hand, if the English speaker 2 determines that there is a mistranslated chunk among the translated chunks, the designation (identifier) of the mistranslated chunk is included in the data signal by the following operation by the English speaker 2 (step S6, Yes).

Specifically, FIG. 3C shows the display on the second terminal 21 when the designation of mistranslated chunks by the English speaker 2 is accepted. In other words, this shows the display when the translation chunk 'ASU' is mistranslated and a mistranslation designation is accepted.

In other words, the text widget 21W for 'ASU' is highlighted (designated display), and the text widget 21W for 'I go to Tokyo' is not highlighted (standard display). Note that when the text widget 21W is repeatedly touched, the display is alternately changed to no highlighting, highlighting, and no highlighting.

Alternatively, change the text color or frame line color, etc., such as displaying the text or frame line of the text widget 21W for 'ASU' in red (specified display) and 'I go to Tokyo' in black (standard display), Alternatively, it is sufficient if the mistranslation designation can be visually recognized by blinking characters or the like. Alternatively, the system may be configured to notify the user by audio or the like so that the text widget designated as an incorrect translation can be confirmed.

In step S5, when the English speaker 2 presses the SEND button in FIG. (FIG. 1) (step S7).

Data signal DS4: {‘id’:10, ‘eng’:[‘ASU’,’I go to Tokyo’], ‘error’:[0]}

Here, chunk ID = 0 is assigned to the specified displayed (touched) chunk 'ASU', and the data signal DS4 has an identifier 'error':[ 0] has been added.

That is, an identifier (also referred to as a mistranslation identifier) is generated according to a signal (identification signal) specifying a mistranslation chunk inputted to the second terminal 21, and the error correction control unit 25 corrects the mistranslation based on the identifier. The chunk is displayed on the first terminal 11 in a manner different from other translated chunks (ie correctly translated chunks).

Note that the chunk ID [j] is assigned, for example, j=0, 1, 2, etc. in the chunk order of the uttered text (original language, Japanese text). Furthermore, the chunk ID [k] in the identifier 'error':[0] is assigned the ID of the translated chunk corresponding to the source language (Japanese) chunk ID [j].

Also, when each source language chunk of the spoken text is expressed as chk[j] (j=1, 2, ·j as chk1, chk2,·, the translation function is translate(chk1, chk2, · = translate(chk1) + It is expressed as translate(chk2) + ....

That is, the original language chunk chk[j] is translated one by one in chunk order. Therefore, translation chunks can be output without making the listener wait. Note that at this time, one chunk at a time, or two or more consecutive chunks may be translated one by one in chunk order. That is, at least one chunk may be translated sequentially.

The correction manager 13 receives the data signal DS4, extracts the element (translation chunk) of 'eng': ['ASU','I go to Tokyo'], and generates the following data signal DS4A. The data signal DS4A is the content of the English text displayed on the display screen (touch screen) of the second terminal 21.

Data signal DS4A: ‘eng’: [‘ASU’,’I go to Tokyo’]

Next, in order to convert the text back into Japanese using the translator 15, translation is performed using the following procedure to generate the following data signal DS5. The translator 15 has a back translation function that back translates the translation result (English) of the speaker's (Japanese speaker) text and generates a back translation chunk (Japanese).

That is, a back-translation chunk (Japanese) obtained by back-translating the translation result (English) of the speaker's text (Japanese) is returned to the speaker's first terminal 11.
translate('eng': ['ASU','I go to Tokyo'], 'ej')
= translate('eng': ['ASU'], 'ej) + translate('eng': ['I go to Tokyo']'ej')
='jap':['ASU'] + 'jap':['I'm going to Tokyo']
='jap': ['ASU', 'I'm going to Tokyo']
Data signal DS5: {'id':10, 'jap': [''ASU','I'm going to Tokyo'], 'error': [0]}

The first terminal controller 12 refers to the identifier 'error':[0] indicating an error (mistranslation) and its chunk ID=[0] in the data signal DS5, and refers to the back translation result 'jap': ['ASU Detects that 'ASU' in ','I'm going to Tokyo'] is an error.

The first terminal 11 displays the mistranslation chunk widget 11W ('ASU') with a highlight, as shown in FIG. 3D, based on the detection result of the first terminal controller 12 (step S8 ). Note that, as described above, the font color or frame color of the mistranslated chunk 'ASU' may be changed to a different color from the standard display.

Also, at this time, as shown in FIG. 3D, the English chunk translated by the translator 15 is back-translated by the translator 15 and displayed on the first terminal 11. That is, the first terminal 11 displays 'I'm going to Tokyo', which is the result (back translation chunk) of the translation (English) of the speaker's text (Japanese) back translated into Japanese. Therefore, the speaker can easily determine whether or not the translation has been correct.

Note that, as shown in FIG. 3E, the English chunk (translation chunk) translated by the translator 15 may be displayed as is on the first terminal 11.

Therefore, the Japanese speaker 1 can confirm that there is a mistranslation or meaningless translation (collectively referred to as "mistranslation") by highlighting or other specified display. It can be seen that the English speaker 2 has designated the chunk 'ASU' corresponding to 'tomorrow' as a chunk that is a mistranslation or a meaningless translation (mistranslation chunk).

The Japanese speaker 1 paraphrases the Japanese chunk corresponding to the chunk specified by the English speaker 2, and inputs the paraphrased text into the first terminal 11 (step S9). The paraphrased text is translated through the flow from step S1 described above, and translated chunks of the paraphrased text are displayed on the second terminal 21 (step S5).

Note that in this flow, all chunks included in the divided text may be translated, but only the paraphrased chunk may be translated, and the previous translated chunk may be used for the other translated chunks. It has advantages in terms of reducing translation capacity and improving translation speed.

In other words, Japanese speaker 1 can translate the Japanese part corresponding to the chunk specified by English speaker 2 into English by saying it in a different way, using a simpler expression, etc. can be conveyed to speaker 2.

For example, the Japanese speaker 1 can convey the accurate translated text to the English speaker 2 by rephrasing 'Asu' as 'Ashita'.

FIG. 3F is a diagram showing the second terminal 21 on which the paraphrased text is translated and displayed. This time, the translation is accurate, so by touching the SEND button, the translation of this spoken text, ``I'm going to Tokyo tomorrow'', is completed.

As explained above, since translation errors are reliably fed back to the source language side, error correction is efficient, accurate translation can be performed, and conversation can be carried out reliably.

Note that FIG. 3G shows a display that is a modified example of the display of the first terminal 11 shown in FIG. 3D. That is, the data signal DS2 is sent from the correction manager 13 to the first terminal 11 along with the data signal DS5 via the first terminal controller 12, and the first terminal 11 is sent the data signal DS2 together with the error 'ASU'. The utterance content 'tomorrow' of speaker 1 may be added and displayed.

In this case, along with the meaningless translation chunk ('ASU'), the corresponding utterance chunk ('Asu') is displayed on the first terminal 11, so 'ASU', which is the utterance content of the Japanese speaker 1, is displayed on the first terminal 11. ' is converted to 'ASU', and it is clearly and concisely displayed that the English speaker 2 has judged the sentence to be meaningless. Therefore, accurate translation helps in smooth conversation.

In other words, by correcting errors in chunks, errors can be localized and corrected locally, making error correction conversations more efficient and improving the quality of conversations during simultaneous interpretation. be able to.

[Second embodiment]
FIG. 7 is a block diagram showing the configuration of an error correction translation device 30 according to the second embodiment of the present invention.

In the error correction translation device 30 according to the present embodiment, a speech recognition section 11B and a speech recognition section 21B are provided in the first terminal 11 and the second terminal 21, respectively, in place of the keyboard 11A and the keyboard 21A. , is different from the error correction translation device 10 according to the first embodiment.

More specifically, in the error correction translation device 30, the first terminal 11 includes a speech recognition unit 11B that performs speech recognition on the utterance of the Japanese speaker 1 and outputs it as text data. The second terminal 21 also includes a speech recognition unit 21B that performs speech recognition on the utterances of the English speaker 2 and outputs the speech as text data.

When the Japanese speaker 1 speaks, the voice recognition unit 11B of the first terminal 11 performs voice recognition of the utterance and sends text data that is the voice recognition result to the first terminal controller 12.

Similarly, the second terminal 21 receives voice input from the English speaker 2, and sends the voice recognition result to the second terminal controller 22.

The speech recognition unit 11B and the speech recognition unit 21B have a speech recognition engine that can encode speech and convert it into data. For example, a method of encoding audio such as WAV or MP3, which is a known technology as a speech recognition method, can be used.

Note that in the error correction translation device 30 of this embodiment as well, mistranslation chunks are specified in the same way as in the first embodiment. That is, an identifier is generated that specifies that there is a mistranslation or a meaningless translation, and based on the identifier, the mistranslation chunk is displayed on the speaker's terminal in a manner different from other translation chunks.

In this embodiment, if a meaningless or unclear translation chunk is generated due to an error in speech recognition by the speech recognition unit 11B or the speech recognition unit 21B, the translation chunk is also processed as a mistranslation chunk. Ru.

Note that either one of the first terminal 11 and the second terminal 21 may be configured so that text data is input through voice recognition, and the other terminal is configured such that text data is input through a keyboard. Further, the first terminal 11 and the second terminal 21 may have a function of inputting text data by keyboard input and voice recognition.

[Third embodiment]
FIG. 8 is a block diagram showing the configuration of an error correction translation device 50 according to the third embodiment of the present invention.

The error correction translation device 50 according to the present embodiment has the translation function from Japanese (first language) to English (second language) of the error correction translation device 10 according to the first embodiment. A translation function is provided from Japanese (second language) to Japanese (first language). That is, the error correction translation device 50 is configured as a translation device capable of bidirectional simultaneous translation.

Note that the first language and the second language are not limited to Japanese and English, and can be applied to other languages as well.

Specifically, the text divider 54 divides the spoken text data included in the data signal DS1 in the Japanese-English translation as well as the data signal ES1 in the English-Japanese translation into chunks.

Furthermore, the translator 55 is configured as a Japanese-English/English-Japanese translator. Further, the correction manager 53 is configured to manage data exchange of data signals DS1 to DS5 during Japanese-English translation and data signals ES1 to ES5 during English-Japanese translation.

Further, the correction manager 53 is provided with an error correction completion determination unit 53A that determines whether error correction of one utterance text is completed. The error correction completion determining unit 53A determines whether error correction of the one utterance text is completed based on the presence or absence of an identifier that specifies that there is a mistranslation or a meaningless translation.

For example, in the case of bidirectional translation, when correcting errors in the uttered text from the speaker (Japanese speaker 1) to the receiver (English speaker 2), the identifier input from the second terminal 21 (first It is determined whether the error correction of the utterance text is completed based on the utterance identifier), and in the error correction of the utterance text from the speaker (English speaker 2) to the receiver (Japanese speaker 1), Based on the identifier (second identifier) input from the second terminal 11, it is determined whether or not the error correction of the uttered text has been completed. When a predetermined elapsed time has elapsed since each of the translation chunks was displayed on the terminal 11 of the terminal 11, it may be determined that the error correction of the uttered text of the one has been completed. This is because even if a mistranslated chunk is included in the translated chunk, the correct meaning may be understood. In this case, the error correction completion determination unit 53A has a timer that measures the elapsed time.

Alternatively, if the mistranslation chunk is not specified and input is simply made using the SEND button (or RETURN key or ENTER key), the error correction completion determination unit 53A determines that error correction of the utterance text has been completed. You may.

In other words, as explained in relation to step S6 etc. in FIG. 2, if the mistranslation identifier is present in the transmission/reception signal of the correction manager 53, it means that the error correction for the inputted utterance text has not been completed yet. , the transmitted/received signal is determined to be a transfer data signal for error correction. In this case, the correction manager 53 functions to prohibit input of the next new utterance text and prioritize completion of error correction.

Here, when translating from English (second speaker) to Japanese (first speaker), data signals ES1 to ES5 during English-Japanese translation correspond to data signals DS1 to DS5 during Japanese-English translation, respectively. This data signal is configured as an equivalent signal to the data signals DS1 to DS5 except that the language of the speaker and the direction of translation are different. Specifically, similar to the case described above, the following data signals ES1 to ES5 can be exemplified.
・Data signal ES1: {'id':11, 'eng': ['I'm heading to Boston tomorrow']}
・Data signal ES2: {'id':11,'eng': ['I'm heading to','Boston','tomorrow']}
・Data signal ES3: {'id':11,'jap': ['I'm going','To Boston','Tomorrow']}
・Data signal ES4: {'id':11,'jap': ['I'm going','To Boston','Tomorrow'],'error':[1]}
・Data signal ES5: {'id':11,'eng': ['I'm heading to','BOSTON','tomorrow'],'error': [1]}

In this case, the recipient (Japanese speaker 1) specifies 'To Boston' as meaningless and adds the identifier 'error':[1] (second identifier) containing chunk ID[1]. The resulting data signal ES4 is sent to the speaker (English speaker 2).

The second terminal 21 of the speaker (English speaker 2) receives an identifier in the data signal ES5 as a designated indication that 'BOSTON' is a meaningless translation (mistranslation), unlike other translation chunks. It is displayed in a different manner (for example, in uppercase letters), and the English speaker 2 can confirm that 'Boston' in the speech chunk is meaningless to the listener (Japanese speaker 1).

Note that the translator 55 has a context determination function. In this example, the translator 55 determines that 'Boston' means 'Boston Library' from the context of the conversation between Japanese speaker 1 and English speaker 2, and translates it. This is a mistranslation of "Boston" when it should have been translated.

Therefore, in this case, the English speaker 2 can convey the accurate translated text to the Japanese speaker 1 by rephrasing 'BOSTON' to 'the Boston Library'.

In the case of this embodiment, bidirectional simultaneous translation is possible between speakers of the first language and speakers of the second language. Therefore, both parties can confirm whether the meaning is unclear or there is a mistranslation by the designated display, and conversation can be smoothly carried out through accurate translation.

According to the present invention, by managing units of utterances with meaningful chunk IDs as translations, and by linking utterance information in the source language and translation information in the target language to the chunk IDs, errors can be avoided on the source language side. can provide reliable feedback.

In addition, by using the chunk-based error correction of the present invention, errors can be localized and corrected locally, making error correction conversations more efficient and improving the quality of conversations in simultaneous interpretation. can be improved.

Note that in the above example, a case was explained in which one sentence was input as text to the error correction translation device, but it is also configured such that text consisting of multiple sentences is input and translation is performed on the input text. may have been done.

In this case, for example, the first terminal controller 12 and the second terminal controller 22 receive signals indicating input confirmation such as 'SEND' and 'Enter' transmitted from the first terminal 11 and the second terminal 21. The end of the text can be determined based on .

Furthermore, although the case where data is exchanged by sending and receiving data such as data signals DS1 to DS5 is explained using the JSON format as an example, XML (Extensible Markup Language) or the like may also be used.

Furthermore, the error correction translation device described above may be configured as a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), or a server. Furthermore, it may be implemented in computers, information terminals, mobile phones, and the like.

Although the above-described error correction translation device and error correction translation method are configured as hardware, they may be configured as computer software (program). Alternatively, the program may be configured as a computer-readable storage medium in which the program is stored.

For example, any one of the first and second terminal controllers, text divider, translator, speech recognition unit, and error correction control unit may be configured as software. Alternatively, these may be configured as one software.

Although the embodiments of the present invention have been described in detail above, various modifications can be made without departing from the spirit of the present invention.

10, 30, 50: error correction translation device, 11: first terminal, 11A, 21A: keyboard, 11B, 21B: speech recognition unit, 11W, 21W: text widget, 12: first terminal controller, 13, 53: Divided text correction manager, 53A: Error correction completion determination unit, 14, 54: Text divider, 15, 55: Translator, 21: Second terminal, 22: Second terminal controller, 25: Error Correction control unit, 53A

Claims

a first terminal into which text in a first language is input;
a text divider that divides the first language text into chunks to generate divided text;
a translator that individually translates each chunk included in the divided text to generate translated chunks in a second language;
a second terminal;
an error correction control unit that causes each of the translation chunks from the translator to be individually displayed on the second terminal;
The second terminal receives an input of an identifier specifying a mistranslated chunk among the translated chunks,
The error correction control unit causes the mistranslation chunk to be displayed on the first terminal in a manner different from other translation chunks based on the identifier input to the second terminal, and the error correction control unit causes the mistranslation chunk to be displayed in a manner different from other translation chunks. An error correction translation device characterized in that, in response, the first terminal receives input of paraphrased text of the text in the first language.
The error correction translation according to claim 1, wherein the error correction control unit makes the paraphrase text the text in the first language, and displays each of the second language translation chunks of the paraphrase text on the second terminal. Device.
the translator back-translates the translated chunks in the second language to generate back-translated chunks;
The error correction translation device according to claim 1, wherein the error correction control unit causes the back translation chunk to be displayed on the first terminal.
The second terminal has a display with a touch panel function,
the translated chunk and the mistranslated chunk are displayed on the display as a text widget;
The error correction translation device according to claim 1, wherein the second terminal receives the input of the identifier by touching the display.
The error correction translation device according to claim 1, wherein the text divider performs morphological analysis to divide the text in the first language into the chunks.
The error correction translation device according to claim 5, wherein the text divider determines boundaries of the chunks by deep learning.
The error correction translation device according to claim 1, wherein the first terminal includes a speech recognition unit that recognizes spoken speech and generates text in the first language.
The error correction translation device according to claim 7, wherein the mistranslation chunk includes a translation error due to a speech recognition error by the speech recognition unit.
a first terminal and a second terminal;
Divide the text in a first language input from the first terminal and the text in a second language input from the second terminal into chunks to generate a first divided text and a second divided text, respectively. a text divider,
a translation machine that individually translates each of the chunks included in the first divided text and the second divided text to generate translated chunks in the second language and translated chunks in the first language, respectively;
individually displaying each of the translated chunks in the first language from the translator on the second terminal; and displaying each of the translated chunks in the second language from the translator on the first terminal. an error correction control unit that displays the information individually;
The second terminal receives an input of a first identifier specifying a mistranslated chunk among the translated chunks in the first language,
The first terminal receives an input of a second identifier specifying a mistranslated chunk among the translated chunks in the second language,
The error correction control unit includes:
When the first identifier is input to the second terminal, the mistranslated chunk in the first language is sent to the first terminal in a manner different from other translated chunks based on the first identifier. display,
When the second identifier is input to the first terminal, the mistranslated chunk in the second language is transmitted to the second terminal in a manner different from other translated chunks based on the second identifier. display,
in response to displaying the mistranslated chunk in the first language, causing the first terminal to accept input of paraphrased text of the text in the first language;
An error correction translation device that causes the second terminal to accept input of paraphrased text of the text in the second language in response to displaying the mistranslated chunk in the second language.
The error correction control unit is configured to perform an error correction based on the first identifier and the second identifier, or based on the elapsed time since the mistranslation chunk was displayed on the second terminal or the first terminal. 10. The error correction translation device according to claim 9, further comprising an error correction completion determination unit that determines whether error correction of one utterance text inputted from the first terminal or the second terminal is completed.
11. The error correction translation device according to claim 10, wherein the error correction completion determining unit prohibits input of a next new utterance text when determining that error correction of the first utterance text is not completed.
inputting text in a first language into a first terminal;
a text dividing step of dividing the first language text into chunks to generate divided text;
a translation step of individually translating each chunk included in the divided text to generate translated chunks in a second language;
an error correction control step of individually displaying each of the translation chunks from the translator on the second terminal;
receiving, at the second terminal, an input of an identifier specifying a mistranslated chunk among the translated chunks;
a mistranslation chunk displaying step of displaying the mistranslation chunk on the first terminal in a manner different from other translated chunks based on the identifier input to the second terminal;
after the mistranslation chunk displaying step, receiving input of a paraphrase text of the text in the first language to the first terminal;
An error correction translation method having the following.
computer to translate text data,
a first terminal means into which text in a first language is input;
text dividing means for dividing the first language text into chunks to generate divided text;
Translation means for individually translating each chunk included in the divided text to generate translated chunks in a second language;
second terminal means;
functioning as error correction control means for individually displaying each of the translation chunks from the translation means on the second terminal means;
The second terminal means receives an input of an identifier specifying a mistranslated chunk among the translated chunks,
The error correction control means causes the first terminal means to display the mistranslation chunk in a manner different from other translation chunks based on the identifier input to the second terminal means, and displays the mistranslation chunk on the first terminal means in a manner different from other translation chunks. An error correction translation program characterized in that, in response to a display, the first terminal means accepts input of paraphrased text of the text in the first language.
A computer-readable storage medium storing the error correction translation program according to claim 13.