JPH09223146A - Translating method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce the user's trouble for word selection and postediting by controlling the learning dictionary contents of a dictionary means to be made effective for subsequent translation. SOLUTION: An input part inputs a sentence to be translated and various commands. A control part 5 is the kernel of control of the whole system and controls the operation of the whole system. A word selection part 2 selects a translated word in editing mode. Then a learning control part 3 controls the recording, update, etc., of learning dictionary information in a learning dictionary part 7b of a dictionary part 7 by controlling the learning of plural selection contents wherein index words and translated words reciprocal, selection contents before last selection for one index word, etc., according to the contents of the translated word to control the learning dictionary contents of the learning dictionary part 7b so that effective learning information can be given for translation by a translation part 6. Further, the translation part 6 translates a source sentence which is inputted from an input part 1 and supplied through a control part 5 by referring to various information that the dictionary part 7 has.

Description

Detailed Description of the Invention

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a machine translation method for translating a sentence in a first language into a sentence in a second language.

2. Description of the Related Art Machine translation is a technique for automatically converting a sentence in one language into a sentence in another language using a computer. With the remarkable development of electronic technology and semiconductor technology including computer technology, anyone can easily use the computer, and also the machine translation technology such as the improvement of natural language processing technology and the improvement of calculation function can be achieved. With the further development of the environment surrounding the field, it has become possible to provide translated texts of high quality, and in recent years, machine translation systems have attracted attention. By the way, what is required for machine translation is to generate a translated sentence which is not strange as a language by reflecting the contents of the original sentence correctly and quickly. However, there are various types of original documents, and there are various expressions to provide a translation dictionary with a large amount of knowledge that is sufficient to translate all original sentences of all users in a satisfactory manner. I haven't arrived. At present, as a means for coping with this, various technical term dictionaries and user dictionaries are made available, and a translation learning function is provided only. The technical term dictionary is used for translating the contents of the original document, and for reflecting the translated words that depend on the subject field of the current document in the translated text. This is the dictionary to use. The user dictionary is for defining expressions and phrases unique to the user as headwords and designating translations thereof. In addition, the translated word learning function allows the user to select the translated word again by learning the translated word desired by the user for other general words according to the document and reflecting the learning content in the subsequent translations. Function so that the user's preferences can be naturally reflected in the translated text.
This is an important function. In the translated word learning function, when different translated words are selected in different places for the same headword, basically the last selection result is learned. However, in reality, when contradictory selections such as selecting different translations are made in different places like this, the translation of the latest selection is reflected in the learning result, and it affects the subsequent translation. It is not always preferable to do this. In other words, there are many cases where you want to change the translated word only depending on how the word is used in each sentence.In such a case, if you reflect the latest selection in the next translation, that part will be changed. The user will have to make a reselection, which is difficult to use. In addition, it may not be reflected in the subsequent translation (no need for learning), and it may be a valid selection only for the one sentence. Therefore, in the machine translation system, it is necessary for the user to consciously specify whether learning is necessary or unnecessary, or to avoid using the translation learning function for words that do not require learning. there were.

As described above, in the machine translation system, it is difficult to provide a large amount of knowledge as a translation dictionary so that the translated texts satisfying the users can be provided to all the users. Therefore, at present, various technical term dictionaries and user dictionaries are made available or a translation word learning function is provided as a means for coping with this. Then, by making the technical term dictionary and the user dictionary available, it is possible to perform translation corresponding to the field or according to the user's preference. Also, in order to reflect the user's translation usage tendency and to avoid translations that the user does not like as much as possible, there is a translation learning function, but with the conventional translation learning function, in general, it is time-based. Is designed to learn new selection results and reflect them in subsequent translations. Therefore, if the user selects a translation without being aware of it, learning may be performed that is valid only for one sentence and does not need to be reflected in the subsequent sentence, or if learning is necessary, the command is There was a problem that one side that was inconvenient was left, such as having to use properly. There is also a problem that it is not possible to learn different translated words for the same heading Satoru depending on the co-occurrence of words, and the later learning cancels the previous learning. Therefore, an object of the present invention is to cope with the above-mentioned problems, by making use of conflicting learning contents, it is possible to reflect the translated word with the highest expected value in the subsequent translation, and as a result, the user can select / post-edit the translated word. It is to provide a machine translation method capable of reducing the trouble of.

In order to achieve the above object, the present invention is as follows. That is, the present invention
Knowledge of dictionary means including an input step for inputting characters and a learning dictionary holding learning information, and holding knowledge information for converting the inputted natural language sentence in the first language into a second language A translation step of converting a word string of a first language into a word string of the second language using information, an output step of outputting a translation result in the translation step, and a translation word selection step for selecting a translation word. , A plurality of selection contents in which the relation between the headword and the translation word is contradictory based on the contents of the translation word selection in this translation word selection step, and 1
And a learning control step of controlling the contents of the learning dictionary of the dictionary means so that the contents selected before the last selection for one headword can be learned and the contents are effective in the subsequent translation. Then, when the natural language sentence in the first language is input in the input step, the knowledge of the dictionary means holding the knowledge information for converting the input natural language sentence in the first language into the second language in the translation step. The information is used to convert a word string in the first language into a word string in the second language. Then, in the output step, the translation result in the translation step is output. When there is a translation word to be corrected by looking at the translation result, the translation word selection step displays a list of candidates for the selection, and the translation word selection operation instruction can be performed to select the translation word. Then, in the learning control step, it is possible to learn a plurality of selection contents in which the relationship between the headword and the target word is contradictory, and the selection contents before the last selection for one headword, based on the contents of the target word selection in the target word selection step. At the same time, the learning dictionary contents of the dictionary means are controlled so as to be effective in the subsequent translation. By these, and further, in the translation step, in converting the word string of the first language into the word string of the second language, the contents of the learning dictionary are used to obtain the most expected value for a word in a sentence. By reflecting the high learning content, it is possible to make use of the contradictory learning content and reflect the translated word with the highest expected value in the subsequent translation, and as a result, the user's labor for selecting the translated word and post-editing can be reduced. become able to. The system configuration includes an input unit for inputting characters, a dictionary unit for holding knowledge information for converting an input natural language sentence in the first language into a second language, and the dictionary unit. First using knowledge information
A translation means for converting a word string of a language into a word string of the second language, an output means for outputting a translation result output by the translation means, a translation word selection means for selecting a translation word, and this translation word selection Based on the contents of the translation word selection made through the means, it becomes possible to learn a plurality of selection contents in which the relationship between the headword and the translation word is contradictory, and the selection contents before the last selection for one headword, and Learning control means for controlling the contents of the learning dictionary of the dictionary means to be effective in translation is provided. In this system, when a sentence and a word string in the first language, which is a natural language, are input from the input unit, the system sends the sentence and the word string to the control unit. The translation means performs translation processing using the knowledge in the dictionary means, and outputs the translated text to the output means via the control means. When the user selects a translated word with respect to the translation result via the translated word selecting means, the content is once sent to the control means. The control means sends a translation word replacement display command to the output means according to the selected content, and the translation word is replaced.
On the other hand, the selected content is also sent to the learning control means. The learning control means records past translation word selection contents in the learning dictionary. Each time a different selection is made for the same headword, the selected content is recorded, and the type of translation word for the relevant headword that occurs within the range of the sentence that has been translated / post-edited by that point The frequency of each is also recorded. Based on this frequency, a default translation word is selected from the learning translation word group in the subsequent translation processing and recorded in the learning dictionary. In the translation process in the translation means, the learning dictionary is referenced when determining the translated word. If the newly translated sentence matches any of the translated word selection condition groups recorded in the learning dictionary, the translated word selected under the condition is determined as the translated word of the sentence.
If the headword matches but the co-occurrence words do not match, the default translation is determined as the translation. By operating in this way, even if contradictory translations are selected for one headword, one is not invalidated, and by using the co-occurrence as a condition, all learning knowledge Can be used for translation. And, based on the frequency of the translated word that was used in the past, by estimating the translated word with the highest expected value in the later translation and storing it as the default translated word,
The post-editing work of the user can be reduced.

DETAILED DESCRIPTION OF THE INVENTION The present invention will now be described with reference to the drawings.
A concrete example of Ming will be described. FIG. 1 shows an example of the present invention.
Is a block diagram showing the configuration of a machine translation system
You. In the figure, 1 is an input unit, 2 is a translation selection unit, and 3 is a learning unit.
Learning control section, 4 output section, 5 control section, 6 translation section, 7
It is a dictionary section. The input unit 1 is used to input sentences to be translated and
Control unit for inputting seed commands
For 5, enter the sentence to be translated and enter various commands.
It is for doing power. As the input unit 1,
Usually, a keyboard and a mouse are used. Translation target sentence
In addition to using the keyboard, OCR (optical
Type character reader), floppy disk, magnetic tape, magnet
Consider reading from a magnetic disk, magneto-optical disk, etc.
Can be The output unit 4 outputs the translation result output from the translation unit 6.
Output or output from the input unit 1 or the translated word selection unit 2 to the control unit 5
For displaying responses from the control unit 5 to various commands
Things. The output unit 4 is a display means, for example, CR
T display device, liquid crystal display device, plasma display, etc.
Such as a display for image display is usually used.
However, the translation result is output by a printing machine (printer).
Printing means such as, or directly, floppy disk, magnetic
When outputting to a recording medium such as a magnetic tape or magnetic disk
And transmission means (communication means) for transmitting to other media
You can also. The control unit 5 controls the entire system.
Pivotal, for controlling the movement of the entire system
It is. The translation word selection unit 2 selects translation words in the edit mode.
Input section 1 in the edit mode.
Of the user who is given via the control unit 5 by the operation from
The translation word is selected according to the instruction. Learning control unit 3
Is the content of the translation selection made via the translation selection unit 2.
Based on multiple choices with conflicting headword and target terms
Content and before the last choice for one headword
Learning in the dictionary unit 7 by controlling learning of selection contents
Controls recording and updating of the learning dictionary information in the dictionary section 7b.
At the same time, provide effective learning information for translation in the translation unit 6.
In the learning dictionary of the learning dictionary unit 7b so that it can be given
It controls the content. The translation unit 6 is input from the input unit 1.
And the original text (text in the first language) given via the control unit 5
Referring to various information (knowledge) held by the dictionary unit 7
This is the part that performs the process of translating, and has the highest expected value.
The learning content selection means 6a for determining a translated word is provided.
The translation unit 6 uses the knowledge of the dictionary unit 7 in the translation process.
Use to perform translation. The dictionary unit 7 is necessary for translation processing.
It holds various dictionary information and rules, and the vocabulary section 7
a, learning dictionary unit 7b, morphological analysis rule 7c, syntax / meaning
Analysis rule 7d, conversion rule 7e, production rule 7f, morphomorphism
It consists of 7g of rules. Of these, the vocabulary section 7a is a word
This is a dictionary part that collects vocabulary information, and the learning dictionary part 7b is
Store the learning content learned by the user's translation selection
It is a dictionary. In addition, the morphological analysis rule 7c defines the input document
A part that holds the analysis rule information for dividing into
The syntax / semantic analysis rule 7d parses the input document,
This is the part that holds the rule information for semantic analysis.
The conversion rule 7e is a rule for conversion from the first language to the second language.
This is the part that holds the rule information. In addition, the production rule 7f is
A part that holds the rule information for generating two languages
Therefore, the morpheme generation rule 7g generates a morpheme of the second language.
This is the part that holds the rule information for. Next, the above configuration
The operation of the present system will be described. Processing target from input unit 1
When the original text of is input, this original text is sent to the control unit 5.
You. Then, the control unit 5 sends the original text data to the translation unit 6.
It will be sent and translated. Accept original text data
When the translation instruction is given by the user, the translation unit 6
Using the knowledge and rules of the writing department 7, for this input source text,
Second language sentence or second language containing first language character string
Translates the language into sentences. There is a translation process
When determining a translation for an entry word, the learning dictionary unit 7
b, and using the learning content selection means 6a, the most expected
Determine a translation with a high value. Translation processing for input source text
When the translation is completed, the translation unit 6 passes the translation result to the control unit 5.
Then, the control unit 5 sends this translation result to the output unit 4 and outputs it.
And complete the general processing. See the translation result
If the user determines that there is a need for a positive
Is performed from the input section 1, but the editing mode
Will enter. That is, the editing mode is
The user operates the input unit 1 with the translation result of
Then, the control unit 5 is instructed to activate the translated word selection function for a certain word.
It shifts by giving an order. And this translation selection function
When the activation command is given, the control unit 5 causes the translation word selection unit 2
A signal to start selecting a translated word is sent to. The translation word selection unit 2
When the signal to start word selection is received, the user is prompted to select a translated word.
Control the signal to display the screen for output on the output unit 1.
Send to Part 5. On the other hand, the control unit 5 gives a command to the translation unit 6.
For the word at the cursor position on the screen,
Get it from the writing department. The translation unit 6 obtains the acquired translation word candidates
Is passed to the control unit 5, and the control unit 5 sends it to the translated word selection unit 2.
Since it is passed, the translation word selection unit 2 makes a candidate list, and from this list
Display the screen on which you can select the target word and pass it to the control unit 5.
In addition, the control unit 5 displays this on the screen of the output unit 4.
To make it work. As a result, the output section 1 is guided to select.
A list of translation candidates to be displayed and operation buttons are displayed on the screen.
By looking at this screen, the user can select the translation
When performed from the input unit 1, the contents are given to the control unit 5,
The control unit 5 sends a display command according to the selected content to the output unit 4.
It is given and displayed, and the command (selection according to the selection
Instruction) to the translation word selection unit 2. The translation selection unit 2 is the finger
The instruction is further sent to the learning control unit 3 and also via the control unit 5.
And send it to the translation unit 6. By receiving this command,
A replacement command is issued, and on the other hand, the learning control unit 3
The contents of the learning dictionary 7b in the dictionary 7 and the current selection
Update learning contents of the relevant part of the learning dictionary unit 7b from contents etc.
Will be reflected in the subsequent translation process. Transliteration
For the translation results, edit the above-mentioned
By performing the processing described above by
You can obtain the translation result by specifying the desired translation word of
The learning information is learned by the dictionary unit 7. The above is roughly
This is an explanation of the kana processing operation.
Depending on how the word is used, a specific translation is desirable,
If you want to learn a specific translation, that learning
It is designed so as not to affect the translation after
Since it is characterized by these points, the details will be described in detail below.
You. <Interactive translation processing in control unit 3> Dialogue in control unit 3
The flow of a typical translation process is shown in FIG. The editing unit 5 is an input unit
If there is any key input from 1, each key
Perform processing according to the command. The end command from the input section 1
When input (step S201), all processing is terminated
You. If a translation instruction command is received from the input unit 1,
S202), the process proceeds to step S207, and the target sentence
And sends a translation processing start command to the translation unit 6. Translation department 6
For the flow of translation processing in
explain. When the translation process is completed, the process proceeds to step S208.
Then, a command for displaying the translated sentence is sent to the output unit 4. input
If a character key is input from copy 1 (step S
203), the process proceeds to step S209, and the original text character string is set.
And sends a display command to the output unit 4. From the input section 1
If there is an input (step S204), step
The process proceeds to S210, and editing processing is performed. Translated from input section 1
When the selection command is input (step S20)
5) proceeds to step S206, and the translation word selection unit 2
Send the command to activate the target word selection, and the target word selection function is activated.
Be done. Other keys / commands input from the input unit 1
In the case of, it progresses to step S211 and responds to the input.
Perform processing. <Processing in translated word selection unit 2>
FIG. 3 shows the flow of processing according to the above. From the control unit 5, a translation word selection machine
When the command to activate Noh is sent, the translated word selection unit 2 controls
Display unit 4 through unit 5 to display unit 4 together with a list of candidate words
To send a screen display command prompting selection (step S30
1). FIG. 4 is an example of the screen display. This example
Selecting a translation for the entry word "number" in English-Japanese translation
This is a screen for prompting. As shown in the figure, this screen displays "n
various translations for "number" (in the example, "number",
"Number", "Numerical", "Number", "No.") and "Apply button"
“Cancel button” and “Cancel button” are displayed.
Cursor, specify one of the translated words and click "Apply button".
Button or mouse button with the "Cancel button" specified
Operation, you can select or instruct it.
Wear. The black square on the left side of the translated word is the current
Currently, it is the translated word generated in the translated text. Then you
If the cancel button is pressed by the
(S302), the process is ended as it is. Press the apply button
When it is done (step S303), step S30
7 and sends various data related to selection to the learning control unit 3.
Attached. What you will learn is marked on the screen with a black square
It is a translated word. Upon receiving this, the learning control unit 3
The learning dictionary unit 7b is updated (step S308). this
Details of the update process will be described later using another figure.
You. If the user clicks on a translation
(Step S304), Step S305
The square next to the translated word is displayed in black and
Items that were displayed in black with are displayed in white.
To send a signal to the display unit 4. What the user wants to select
Translation you want to learn by clicking on a word and turning it into black
It means that the word is shown. The normal processing flow is
After displaying the word candidates on the display (step S30
1), the desired translation in the translation candidates displayed by the user
Mouse operation by pointing a word, for example, with the mouse cursor
Click the button (step S304)
The relevant word is marked on the display.
Display control is performed as described above (step S305). That
Then, the user presses the apply button (step S
303) The content of learning is determined. Other than the above depending on the user
If the command key of is input (step S30
6), go to step S307, and make a warning sound.
A signal is sent to the output unit 4 through the control unit 5. <Structure of learning contents in learning dictionary unit 7b> Learning dictionary unit 7b
The configuration of the learning content stored in will be described with reference to FIG.
The learning content is translated without learning for each one headword.
Words, learning default translations, and co-occurrence learning information
You. Co-occurrence is used together in an expression / sentence
The relationship between multiple words
Co-occurrence learning information is information obtained by learning information about the relationship between words and sentences
It is. "Translation without learning" refers to a certain headword.
Then, the translated words are output as translated words when no translated words are selected.
The translated word that was used to represent the headword
It means the translated word initially set as. Totally learned
A word that is not in the initial state is the default when it is referenced.
The translated word of the representative is adopted as the translated word during translation.
Be done. The co-occurrence learning information column is an item for distinguishing each translated word.
If the numbers are "[1]", "[2]", ... "[n]"
It is attached and described separately for each translated word, and is registered.
Item numbers "[1]", "[2]", ... "[n]"
Further, for example, "[1]" has a branch number of "[1,
1] ”,“ [1,2] ”, ...“ [1, m] ”,
"[2]" is "[2,1]", "[2,2]", ...
It is subdivided by adding it as "[2, m]". Soshi
For example, branch number “[1,1]”, ... “[2
1] ”,“ [3,1] ”, ... [n, 1]
The selected source text or the corresponding translation has been generated
In this state, the user selects "Translation"
Describe the co-occurrence condition and its frequency in the original text that did not work
Is done. When these co-occurrence learning information is updated,
For the target sentence of the update data, use another diagram later.
Will be described. In this system, for each part of speech of a headword,
The pattern to be stored as a condition is defined separately. So
And the source text does not match the pattern
If “X” is the maximum value of “[m]”, then “[1,
"X?", "[2, X]" and other "[? , X] ”(however,
? Is 1, 2, 3, 4, ...)
Only the frequency is remembered. The memory location for that frequency is
This is for the counting variable set corresponding to each translated word. Ma
In addition, each translated word has a "frequency meter" as a counting variable,
Is the frequency from branch number "[?, 1]" to "[?, X]"
The total value of is stored. Illustrated as an English-Japanese translation system
In this example, the common words defined for each part of speech
The following is a list of patterns stored as trigger conditions. Record
Enter the corresponding word in “????” in the description column.
You. If there are multiple items, separate them with commas
I do. (1) Noun (a) When a noun has a noun or an adjective for prefix modification Description method: (prefix "WORD") WORD: a noun or adjective for prefix modification (b) A suffix other than (a) When there is a modified prepositional phrase Description method: (Postfixed "PRE_NOUN") PRE: Preposition NOUN: Head noun of a noun phrase followed by a preposition (2) Adjective (a) Limited usage (when a noun is adnominal modified ) Writing method: (Limited "NOUN") NOUN: Noun to be modified (b) When there is a subject in narrative usage Writing method: (Narrative "NOUN") NOUN: Head noun of subject noun phrase (3) Verb (a) ) When there is an object (subject when passive) Description method: (object "NOUN") NOUN: Head noun of noun phrase of object (4) Adverb (a) When verb or adjective is modified Description method: (Modification destination “WORD”) WORD: Modification target verb or adjective
The pattern stored as the trigger condition is, for example,
Things. <Translation Processing in Translation Unit 6> Next, in the translation unit 6,
The details of the translation process will be described. FIG. 6 shows the translation unit 6
It is a flowchart which shows the flow of a translation process. According to Figure 6
Explaining the translation process in the translation unit 6, first, the input sentence
Is sent to the translation unit 6,
Is the information of the vocabulary part 7a in the dictionary part 7 for this input sentence
Using the rule information of the morphological analysis rule 7c, the morphological analysis /
Performs a dictionary lookup process and translates the part of speech and translation
Various necessary information is obtained (step S601). next,
The translation unit 6 is information on the syntax / semantic analysis rule 7d in the dictionary unit 7.
Is used to perform a syntax / semantic analysis process (step S60
2), and subsequently, in step S603, conversion rule 7
Use e to convert the analysis results to the structure of the second language.
U. At this point, it is in the learning dictionary unit 7b of the dictionary unit 7.
The learning dictionary is not referenced, and each
The translation for the headword is determined. Then,
In step S604, the translation unit 6 determines the learning dictionary unit in the dictionary unit 7.
7b, refer to the translated word assigned to the entry word of the original sentence.
Check the translated words that should reflect the contents of the learning dictionary section 7b.
Then, the translation of the necessary words is replaced.
At this time, collation processing with the learning dictionary of the learning dictionary unit 7b,
For the process of determining the translation to replace, refer to FIG.
It will be described later. When the processing in step S604 is completed,
The structure on the translation side is completed. After that, in step S605
Then, the translation unit 6 sends information about the syntax generation rule 7f in the dictionary unit 7.
Is used to generate the one-dimensional structure of the translated sentence, and
The morpheme is generated in step S606, and a general translation process is performed.
End the process. <Collation process for reflecting learning content>
9 is a flowchart showing a flow of a matching process in the unit 6,
With the function of the learning content selection means 6a of the translation unit 6, learning
Reflect the learning contents while referring to the contents of the learning dictionary section 7b.
Flow of collation processing with the learning dictionary unit 7b
FIG. Basically, the matching process is within one original sentence.
For each word of the, refer to the learning dictionary unit 7b in order from the beginning of the sentence.
Go together. However, the flow shown in FIG.
This is the flow of processing for the word output. Reflect learning content
The following conditions are the prerequisites. Therefore, this condition is met.
First, it is determined whether to stand (steps S701 and S70).
2). If this assumption does not hold, reflect learning content
Since it is not the target to be processed, the collation processing is ended. content of study
The conditions for reflecting are: (1) the headwords match (2) the translated words generated without using the learning dictionary
It is the same as the “translation without learning”. And if this precondition is met
Matches the situation of the original text from the co-occurrence learning information
You will choose one. "[1,1]"->"[1,
2] ”-> ……“ [1, n] ”->“ [2,1] ”-
->“[2,2]”-> “[2, n]”-> ……
Proceed with individual co-occurrence information and collation in the order of "[m, n]".
I will try. It controls the collation order of this individual co-occurrence information.
This is the numerical control of "i" and "j" in the figure.
"I" is moved up in order (increment),
If the information “[i]” is lost (step S70)
4), go to step S706, "Learning default translation
It is decided to replace with the word "." A branch number "[i,
j] ”matches the situation of the original text (step
S709), the process proceeds to step S710, and the item number
The registered translation of "[i]" is determined as the replacement translation. Steps
After either S706 or S710, the replacement word is
When the determination is made, the collation process ends. The above is the opposite of the learning content
Collation processing with the learning dictionary unit 7b when displaying
It is the details of. <Processing for updating the learning dictionary unit 7b>
Interactively select the translated word from the beginning of the original document and edit it appropriately.
Learning by learning control unit 3 when performing translation processing while
The update process of the dictionary unit 7b will be described. Figure 8 is the translation
The selection result of the translation target word at the time when the selection is made
Show the flow of the process of updating the learning dictionary unit 7b based on
I have. The update process will be described with reference to FIG. Select word
The part that is sent as the translation selection result from Part 2 is the
Words, translated words without learning, translated words you want to learn,
And four types of co-occurrence information in the original text. this
When they are sent to the learning control unit 3, first, step S
In 801 the combination of the translation and the translation without learning
In addition, check whether it is already stored in the learning dictionary unit 7b.
Click. As a result, the unregistered combination
Then, the process proceeds to step S803, and the learning term
Register new eyes and learn the translation you want to learn Default
Register as a translation. Then, proceed to step S805.
Register "[1]" based on the co-occurrence information of the original text.
You. The co-occurrence information of the original text is a separately defined part-of-speech / co-occurrence clause.
If the number matches, register "[1,1]"
If outside, "[1, X]" is registered. However, "X"
Indicates the maximum value of "m" of "[1, m]". Explanation below
The same applies to the treatment of "X" in. In any case,
The value of the "frequency meter" of the item number "[1]" of a certain headword is
The first is "1". A set of the translation and the translation without learning
If the combination has already been registered, step S80.
Proceed to 2. Here, the specific part-of-speech (the part-of-speech group defined separately
I.e., noun, verb, adjective, adverb
It is determined whether or not As a result of the judgment,
If it is other than the fixed part of speech, the process proceeds to step S804. This
Here, the learning information registered in the co-occurrence learning information
Then, find "[i]" that matches the translated word you want to learn,
The "frequency" of "[i, X]" is added, and the "frequency" of "[i]" is added.
Add a "Pedometer". On the other hand, it matches the translated word you want to learn
If there is no "[i]", add a new "[i]",
“Frequency” of “[i, X]” and “Frequency meter” of “[i]”
Add values. In step S802, if it matches the specific part of speech
If it is determined that there is, the process proceeds to step S806. here
Now, the co-occurrence learning information about the translated word you want to learn is already
Check if it is registered. Not registered yet
If not, the process proceeds to step S807, and the co-occurrence information of the original text
If the report meets the co-occurrence conditions defined separately,
In step S809, a new item number “[i]” and a branch are added.
Register the number "[i, 1]" and enter the number of "frequency meter"
Set it to "1". The co-occurrence condition of the original text is not a specific condition
Then, in step S811, the new item number
Register "[i]" and branch number "[i, X]", and
Set "1" as the registered content of "Meter". Step S80
The co-occurrence learning information about the translated word you want to learn in 6 is not registered
If it is determined that there is one, the process proceeds to step S808. This
In step S808, the co-occurrence information of the original text is separately determined.
Check whether the co-occurrence conditions that you have defined are met. So
As a result, if it is not the specific condition, the process proceeds to step S813,
Update "[i, X]" of "[i]" for the same translation
Then, the “frequency meter” of “[i]” is added. Step S8
As a result of the check in 08, the co-occurrence information of the original text is determined separately.
If it is determined that the defined co-occurrence conditions are met, the step
Go to step S810. Here, the item number “[i]”
Co-occurrence of any branch number "[i, j]" described in Section
Determine if the information matches the co-occurrence information in the original text
I do. As a result, if it is determined that they match, the step
Proceed to S812, and co-occur with the matching branch number "[i, j]".
Update information content. Specifically, the headword is a noun
Then, in the word entry fields such as "Prefix" and "Postfix"
Add the word sent as co-occurrence information of the original text
(Do not add if already included), set the value of "Frequency"
Add (increment). On the other hand, the item number
Any branch number "[i," described in "[i]"
j] ”does not match the co-occurrence condition, step S81
5 and register a new branch number “[i, j]”,
Add the "frequency meter" of the number "[i]" (increment
To). The "frequency meter" of either "[i]" is added
If it is determined that is, that is, steps S813, S804, S
After 815, compare “frequency meter” of each “[i]”
And the item number that has the "frequency meter" whose value is the maximum
Find "[i]". And the item number found
The registered translation of “[i]” is different from the “learning default translation”
"Learn default translation" in the entry
Is the registered translation of the item number “[i]” with the maximum value of “frequency meter”.
New "learning default translation" by replacing the translation of the word
Then, the update processing of the learning dictionary unit 7b ends. "frequency
There are multiple registered translations for the item number “[i]” with the maximum value of “total”.
Time, new selection result in time, that is, this time selection
The translated words are updated and registered as "learning default translations".
Record. Update the learning dictionary 7b for the target sentence
After that, the learning control unit 3 further adds the same headword last time,
The current sentence is selected from the sentence after selecting the translation for the translation without learning.
Between the current sentence, the same headword and learning as this learning
Where the learning results for the translated word without
Detection, and for each of them, the same processing as described in FIG.
And then update the “frequency meter” of the corresponding item number “[i]”.
New, update the default translation of learning. At this time, this
For the same headword in, and translation without learning
If no selection has been made, the first sentence of the document
The above processing is performed between processing statements. However, this
The most documented part of the work done by
From behind the part close to (more document head side)
When the target point of the place moves and the translation is selected there,
Updates the learning dictionary unit 7b based only on the selection result.
Then, the update processing of the learning dictionary unit 7b based on the above backward part
There is no reason. In other words, at the latest time when the translation selection operation was performed
Do not perform the translation selection operation at the document position before the document position of
It does not learn the translated words of the part that has been set. Above is interactive
When performing the translation process on the
The writing part 7b is automatically updated, and the translated words output in the following sentences
Translation of subsequent sentences by increasing the expected value of appropriateness
The labor of post-processing work in
I have revealed. Translation in contrast to such interactive translation processing
In order to proceed with the work, first do a batch translation and then
After, there is a method of interactive post-editing. This
Also, the learning dictionary unit 7b is updated based on the translated word selection.
The flow of new processing is common. By doing this,
The "Xi default translation" has the highest expected value in terms of frequency.
A new translation will be set, and the translation will be determined in subsequent translations.
Will be reflected in the default. Moreover, the translation word is translated under certain co-occurrence conditions.
That are learned are stored separately, so conflicting
Even if the target word is selected, the previous selection result becomes the latter selection result.
It cannot be canceled. For example, the line containing the source text
To the "the number of computer
There is a character string "Terminals" and behind it
In the line of "identification number
Suppose there is a character string "r". And the learning dictionary section
Learning about the entry word "number" in 7b
As the contents of the dictionary, for example, "identificati
found for the text string "on number"
Item number of FIG. 5 in the word "number"
With the co-occurrence information such as the registered content of "[1]",
The word "number" has been learned and retained.
"The number of comp, which appeared later
For the character string "uter terminals"
Co-occurrence information such as the registered content of item number “[2]” in FIG.
The word "number" was both learned and retained as a translation.
, It is the translation with the larger value of “frequency meter”
The word “number” is set as the “learning default translation”
"The numberer that occurs in the original text again
"the nu" in "of terminals"
The word "number" is used as a translation for "mber".
It is guaranteed that the leaves will be output. In this example, co-occurrence
As learning information, describe the usage of the original word at the time of selection.
By remembering, even if the default learning translation changes
We try to guarantee the effect of the last choice. On this occasion
In this specific example, in order to simplify the process, the word
Should be memorized as the usage (co-occurrence with other words)
The patterns are limited. Be limited in this way
Also, the types of patterns that have been limited are merely examples of the present invention.
Since it is only an embodiment, it is described as processing speed and learning dictionary.
It can be implemented in various forms depending on the storage capacity. Also,
In this example, the entire original document is treated as one range,
The frequency was counted, but it was "parting chapters of original document" and "parting sections".
It is also possible to adjust the frequency recording by setting a delimiter such as "re"
Noh. By doing this, for example,
If the content is different and the desired translation is different, "
Is effective when As described above, this system is
To convert the original text of the first language in natural language into the second language
Using the knowledge information of the dictionary section that has the knowledge information of
To convert the original sentence word string to the second language word string
In the translation device adapted to perform the translation process,
In the dictionary part, as knowledge information, for each headword of the first language
Default translations provided in (the highest priority without learning)
(Translated word) and translated word with the highest usage record
Learning default translation and co-occurrence information registered by translation
More updatable learning dictionaries containing news and frequency information
Configured by holding and output the translation result of the translation unit
And an output unit based on the contents of the learning dictionary
Select and operate the desired translation word in the translation word candidate output to
When you make it, the translation in the translation section based on the selection instruction content
The translation selection part that selects and indicates the translation to be used for
Based on the content of the selection instruction,
Update learning information including co-occurrence information and frequency information
The translation with the most information is updated as the learning default translation.
The learning control unit is provided. And enter
Enter the sentence and word string of the natural language which is the first language from Rikibu
If you type, the input data will be sent to the translation department, which will
The translation result is processed using the knowledge in the writing department and the translation result is
Output to the output section. Start editing process for translation result
The translation word selection unit obtains translation word candidates from the learning dictionary for editing.
Displayed in the output section, and the user selects this candidate from the input section.
When you select the target word from the
The replacement is done. On the other hand, within the selection based on the translation selection operation
The content is also sent from the translated word selection unit to the learning control unit, and the learning control unit
Then, in addition to the translation selection contents this time, it is generated in the translated sentence.
If the user did not select a translation,
Record information about the learning dictionary. The content to record is
Basically, if the entry word and learning contents are empty,
Co-occurrence, selected translation, and co-occur in source text
There are four types of words (the target is defined by the part of speech / syntax)
You. Each time a different selection is made for the same headword,
Your selections will be recorded and the translation process will be
The relevant headword that occurs within the range of the document that has been edited
Regarding, the frequency for each type of translation is also recorded.
Originally one translated word from the learning translation group for the relevant headword
Is selected and recorded as a learning default translation in the learning dictionary.
You. In the translation process in the translation section, when determining the translation word
, Refer to the learning dictionary, but the sentence to be newly translated is
Match any of the translation condition selection conditions recorded in the dictionary
If it does, the translation selected under that condition
Determined as the translation of the sentence, the headwords match, but the co-occurrence
If they do not match, the learning default translation is determined as the translation.
Set. As a result of applying such processing,
Even if a translation word that conflicts with the headword is selected,
By making the co-occurrence a condition without being invalidated,
All can be used as learning knowledge for subsequent translations
You. Furthermore, the usage of words depends on the co-occurrence information in the learning dictionary.
For translations in new sentences that do not match the
Based on the translated word frequency information, the
Also estimated the translated word with a high expected value and used this as the default translated word.
The user's desired translation can be stored
Probability of getting results is high
It will be possible to reduce.

As described above, according to the present invention, when translating an original sentence into a sentence in a desired language, as co-occurrence learning information,
By storing the usage of the original word at the time of selection, even if the default learning translation changes, the effect of the past selection result is guaranteed, so that the user repeats the translation selection Even if the contents are selected, the previous selection result is not invalidated in time, and the past selection result is guaranteed in the later translation. Moreover, especially when the usage of the word in the original sentence does not match the past learning result, the translated word having the highest expected value is reflected as the learning result in frequency, so that the post-editing work of the user can be reduced. Further, conventionally, the learning result by “translated word selection” performed by the user without particular consideration may have a bad influence on another sentence, but according to the present invention, the influence is considered one by one. It also frees you from the hassle of having to use it. Therefore, according to the present invention, the contradictory learning content can be used, and the translated word with the highest expected value can be reflected in the subsequent translation, and as a result, the user can select the translated word.
It is possible to provide a machine translation method capable of reducing the trouble of post-editing.

[Brief description of drawings]

FIG. 1 is a block diagram showing the overall configuration of a machine translation system according to a specific example of the present invention.

FIG. 2 is a flowchart showing the flow of processing during dialogue translation in the control unit of the machine translation system configured as shown in FIG.

FIG. 3 is a flowchart showing a processing flow in a translation word selection unit of the machine translation system configured as shown in FIG.

FIG. 4 is a diagram showing an example of a screen which guides a user's translation word selection processing in the machine translation system configured as shown in FIG.

FIG. 5 is a diagram showing an example of contents of information held in a learning dictionary section of the machine translation system having the configuration of FIG. 1.

6 is a flowchart showing a flow of translation processing in a translation unit of the machine translation system having the configuration of FIG.

FIG. 7 is a flowchart showing the flow of collation processing with the learning dictionary unit in the translation unit of the machine translation system configured as shown in FIG.

8 is a flowchart showing the flow of processing for updating the learning dictionary unit in the translation unit of the machine translation system configured as shown in FIG.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Input part 2 ... Translation selection part 3 ... Learning control part 4 ... Output part 5 ... Control part 6 ... Translation part 6a ... Learning content selection means 7 ... Dictionary part 7a ... Word graze part 7b ... Learning dictionary part 7c ... Morphological analysis Rule 7d ... Syntax / semantic analysis rule 7e ... Conversion rule 7f ... Syntax generation rule 7g ... Morphological generation rule.

Claims (2)

[Claims] 1. An input step for inputting characters, and a learning dictionary for holding learning information, and holding knowledge information for converting the input natural language sentence in the first language into a second language. A translation step of converting the word string of the first language into the word string of the second language by using the knowledge information of the dictionary means, an output step of outputting the translation result in the translation step, and selecting a translated word It is possible to learn the translation word selection step of, and a plurality of selection contents in which the relationship between the headword and the translation word is contradictory based on the content of the translation word selection in this translation word selection step, and the selection content before the last selection for one headword. And a learning control step of controlling the contents of the learning dictionary of the dictionary means so as to be effective in the subsequent translation. 2. An input step for inputting characters, and a learning dictionary for holding learning information, and holding knowledge information for converting the input natural language sentence in the first language into a second language. The word information of the first language is converted into the word string of the second language by using the knowledge information of the dictionary means, and the content of the learning dictionary is used for this conversion, and the most expected word for a sentence is expected. A translation step that reflects the learning content with a high value, an output step that outputs the translation result in the translation step, a translation word selection step for selecting translation words, and a translation word selection content by this translation word selection step. Based on this, it is possible to learn a plurality of selection contents in which the relationship between the headword and the translated word is contradictory, and the selection contents before the last selection for one headword, and in the subsequent translation. In order to an effective translation method characterized by comprising a learning control step for controlling the learning dictionary contents of said dictionary means.