JPH07200602A - Machine translation device - Google Patents

Abstract

(57) [Summary] [Purpose] A machine translation device that calculates the certainty factor of a translation result for a natural language sentence and controls the presentation order of translated sentences to be post-edited by this value to make post-editing work more efficient. To provide. [Structure] In a machine translation device for translating a first natural language sentence into another different natural language sentence, a translation dictionary 105 storing knowledge used for performing a translation process, and a first natural language input. A translation processing unit 104 for translating a sentence of a language into the other natural language using the translation dictionary and obtaining a certainty factor for the translation result, and the translation process in order of low or high confidence factor for the translation result. Further, there is provided a rearrangement means 103 for each pair of the natural language and the first natural language sentence which is the original sentence thereof, and a display means 106 for sequentially displaying the rearranged sentence pairs at the time of post-editing.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a natural language processing technique and, more particularly, to a machine translation device for analyzing a document handling a plurality of languages.

[0002]

2. Description of the Related Art In recent years, many machine translation systems have been developed and used. However, it cannot be said that any machine translation system can always perform a correct translation under the present circumstances, and the user is forced to correct the translation result by post-editing. At that time, unless the translation fails obviously, it is not possible to know whether the output of the machine translation system is correct or incorrect without reading and comparing both the original sentence and the translated sentence, so it is said that post-editing will take time. There was a problem.

To solve this problem, it is necessary to provide a system that presents to the user how accurate the translation result is together with the result.

Reference “Kwangseob, Shim and Yung Taek K
im “Towards a MachineTranslation System with Se
lf-Critiquing Capability, ”Proceedings of the Fift
h International Conference on Theoretical and Meth
odologicalIssues in Machine Translation, July 1
4-16, 1993, pp. 276-299 ”explains the necessity of self-determining the certainty factor of machine translation and presenting it.

However, in the experiment of self-judgment with respect to the certainty factor of the translated text of machine translation shown in this document, structurally ambiguous English "to infinitive" is a very limited phenomenon called the determination of the contact destination. It only deals with it, not the validity of the translated text.
In other words, machine translation usually consists of multiple processes such as analysis (morpheme, syntax), conversion (structure, vocabulary), and generation, and each of them is related to the quality of the translated text. It is necessary to consider all the processes, but this is not touched upon, so it is not intended to judge the validity of the most important translated sentence and show its certainty.

Further, the authors' method of the above-mentioned document is limited to the calculation of the numerical value of the certainty factor for the translated text, and at the most, it can only show the numerical value of the certainty factor for each translated text. Then, the numerical value of the certainty factor of the translation result is presented in the order of the original document. Therefore, the post-editing worker can only know the certainty factor for each translated sentence, and cannot apply the obtained certainty factor to any other value, which does not lead to efficient editing work. .

For example, the exact same original sentence may appear repeatedly in the original document. In this case, if it is a machine translation system capable of batch processing of the same original sentence, post-editing should be easy by using this, but the method of the above document has a check function for checking whether the same original sentence is included or not. Because there is no
Even if the original text with exactly the same content repeatedly appeared, the translated text was repeatedly displayed during post-editing. Therefore, the worker has to perform the same correction work many times. Originally, the translation should be checked only once, but since it was repeatedly displayed during post-editing as described above, the operator had to perform the same correction work many times.

Especially when the translation target is a manual, the same sentence is used everywhere, so that tendency is strong. The same can be said for very similar sentences. For example, the sentence patterns are the same, the verbs are the same, and only the noun phrases such as the object and the subject are different.

When the time spent for post-editing is limited, when the results in the same working time are compared, it is better to correct the translated sentences in order from the one with lower certainty. It can be said that the quality of the translated document as a whole will be higher than that after post-editing, but if the confidence level for each translated sentence is simply displayed, the user will be shown a low confidence level based on the display result. It is difficult to contribute to the time-saving effect because it is necessary to search for translated texts one by one to find each one, and there are also problems such as oversights and poor work efficiency.

[0010]

Since the translated sentence obtained by the translation process by the machine translation device is not perfect, it is necessary to correct it by post-editing so as to conform to the meaning of the original sentence. And, after this, editing becomes a work relying on human labor. At that time, by calculating the certainty factor of the translated text and showing it by a numerical value, it becomes possible to easily distinguish the translated text that should be post-edited and the translated text that is not, even if the post-editor does not refer to the original text, It will be a backup for post-editing work.

In the above-mentioned conventional technique, the certainty factor for the translated sentence as the whole system is calculated and used. However, the use is merely the display of the certainty factor for each translated sentence, and at the time of the post-editing work, the certainty factor is displayed. The post-editor can only be used as a guideline for judging whether or not the translated text should be modified by a numerical value, and the system cannot directly use it for the translation optimization process. Also, how can it be used for optimization processing of translation on the system side,
Even that guideline is not shown.

Therefore, even if the certainty factor is obtained, it is only understood that the post-editing work is necessary. Carefully read back the pair of the original sentence and the translated sentence, and check the translated sentence while comparing the corresponding portions of the original sentence and the translated sentence. However, all post-editing work, such as rewriting to an appropriate expression, must be done manually.
The post-editor did little to improve mental and physical burden.

In particular, when the document to be edited has a long sentence of several tens of pages or more, a great deal of time and energy are required for the correction / editing work, which significantly reduces the work efficiency. The conventional technology could not expect any effect to achieve this improvement.

Therefore, the object of the present invention is to
To provide a machine translation device capable of streamlining post-editing work for a translated sentence to reduce the burden on an operator when translating a natural language sentence into another type of natural language sentence. It is in.

[0015]

The machine translation apparatus of the present invention for achieving the above object is configured as follows. That is, in a machine translation device that translates a first natural language sentence into another different natural language sentence, first, a translation dictionary that stores knowledge used for performing a translation process, and a first input dictionary. A translation processing means for translating a sentence in the natural language into another natural language using the translation dictionary, and obtaining a certainty factor for the translation result, and the translation process in ascending or descending order of the certainty factor for the translation result. Each of the pair of the natural language and the first natural language sentence which is the original sentence thereof is rearranged, and the pair of rearranged sentences is sequentially displayed at the time of post-editing.

Secondly, a translation dictionary in which knowledge used for performing the translation process is stored, a database holding the original sentence and the translated sentence, and the input first natural language sentence is stored in the translation dictionary. When there is a translation example of a sentence similar to the sentence to be translated and translated into another natural language using the database, the translation is imitated, and
Translation processing means for obtaining a certainty factor for the translation result,
The confidence measure of the system for the translation result is measured, and the high confidence factor is accumulated as data together with the original sentence of the first language to control the knowledge database, and the order of low confidence factor for the translation result. Or, in descending order, a means for rearranging each pair of the translated natural language and a sentence of the first natural language which is the original sentence thereof, and a display means for sequentially displaying the pair of rearranged sentences at the time of post-editing. To have.

[0017]

In such a structure, in the case of the first structure,
The translation processing means translates the input first natural language sentence into the other natural language by using the translation dictionary, obtains the certainty factor for the translation result, and the rearrangement means obtains the obtained translation result. The translated natural language and each original text sentence of the first natural language sentence are rearranged in the order of low or high degree of certainty with respect to each other, and the display means displays the rearranged sentence pairs during post-editing. Display in order.

Further, in the case of the second configuration, there is a database for holding the original sentence and the translated sentence, and the translating means converts the input first natural language sentence into the different natural language by using the translation dictionary. When a translation example of a sentence similar to the sentence to be translated and translated is present in the database, the translation is imitated and translated, and the certainty factor for the translation result is obtained. Then, the control means supports the creation of the knowledge database by accumulating, with high confidence, the original text of the first natural language as data in the database.
Further, the rearrangement means rearranges each pair of the translated natural language and the sentence of the first natural language which is the original sentence thereof in ascending or descending order of confidence in the translation result,
The display means sequentially displays the rearranged pairs of sentences at the time of post-editing.

As a result, according to the present invention, when the original sentence is translated into the second natural language, the pair of the original sentence and the translated sentence can be displayed according to the calculated value of the certainty factor for the translated sentence.
Significantly improve the efficiency of post-editing work of machine translation by significantly reducing the rate of human involvement in manipulating and editing documents and the labor that is imposed on humans, and by supporting the creation of parallel translation databases. Will be possible.

[0020]

Embodiments of the present invention will be described below with reference to the drawings.

The present invention calculates the certainty factor of the processing of the system for a natural language sentence, and assists in accumulating a pair of an original sentence and a translated sentence in a bilingual base according to this value, and at the same time, calculates the certainty factor value. Accordingly, a machine translation device that controls the presentation order of the translated sentences to be post-edited to improve the efficiency of the post-editing work is provided, and the details will be described below.

(First Embodiment) First, FIG. 1 shows a block diagram of a machine translation apparatus according to an embodiment of the present invention.
In FIG. 1, reference numeral 101 denotes an input unit including a keyboard, a mouse, an OCR, a magnetic tape device, a magnetic floppy disk device, etc. The input unit 101 is a document to be subjected to language processing (for example, Japanese sentence, English sentence, Chinese sentence). It is for inputting commands and data for various edits and corrections as well as inputting (words, etc.).

Reference numeral 102 denotes a system control unit, which plays a central role in controlling the entire system. The control contents include, for example, sending language data and dictionary data from the input unit 101 to each unit described later, selecting a document to be translated based on an instruction from the input unit 101, and selecting a dictionary.

Reference numeral 103 denotes an editing unit, which edits data in the first language selected by the system control unit 102,
The translation unit 104 edits the data of the second language translated.

Further, 104 is a translation unit, and this translation unit 10
4 performs translation processing on the data in the first language selected by the system control unit 102 while referring to the dictionary of the translation dictionary unit 105 selected by the system control unit 102.

The translation dictionary unit 105 accommodates various dictionary data used when the translation unit 104 performs translation processing.
The contents of the translation dictionary unit 105 will be described later.

Reference numeral 106 denotes a display unit, which is a sentence / phrase in the first language input by the input unit 101 and the translation unit 10.
The sentence / phrase in the second language translated by 4 is displayed. As the display unit 106, a CRT, a liquid crystal display or the like is usually used as a unit capable of displaying characters.

The document storage unit 107 stores the data of the first language and the data of the second language independently of each other, and stores the data of the first language and the data of the second language as the translation result in association with each other. can do.

Next, the contents of the translation dictionary unit 105 are shown in FIG.
This will be described with reference to the translation dictionary unit 105 of. Translation dictionary
Reference numeral 105 denotes an inflection conversion dictionary 105a for converting a word or phrase having a change in word ending or the like into its original form, and a word or phrase dictionary in which a translated word in the second language corresponding to the word or phrase in the first language is stored together with its part-of-speech information. 105b, an analysis grammar dictionary 105c in which a grammar for analyzing the first language is stored, a conversion grammar dictionary 105d in which conversion information from the first language to the second language is stored, and a converted sentence structure of the second language It contains a generative grammar dictionary 105e for deciding, a morpheme grammar dictionary 105f for changing a word form such as the ending of the second language to complete a translated sentence, and the like.

Next, the operation of the apparatus of this embodiment having the above-mentioned structure will be described.

Here, the flow of the translation process for the input document from the input unit 101 will be described with reference to the flowcharts of FIGS. 2 to 7, taking the case where Japanese is translated into English as an example.

Here, for simplification, the morphological analysis unit,
Consider a case where the certainty factor of the structure conversion unit among the syntax analysis unit, the structure conversion unit, and the syntax generation unit is mainly used to adjust the translation result for post-editing. As a specific example of the document, a document in which sentences including the verb “Ageru” as shown in FIG.
Consider the case of post-editing.

FIG. 9 is an example of knowledge registered in the inflection change dictionary 105a, FIG. 10 is an example of knowledge registered in the word / idiom dictionary 105b, and FIG. 11 is a conversion grammar dictionary 105d.
Each of the examples of knowledge registered in is shown.

Utilization change dictionary of the translation dictionary unit 105 shown in FIG.
In 105a, the headword "Aruru" is registered, and in the conversion grammar dictionary 105d, as an indispensable case of this verb,
It is assumed that “ga”, “o”, “ni”, “de”, and “kara” are registered, and further four conversion rules are registered.

These conversion rules are arranged in order of strict application conditions, and application of the rules shall be attempted in this order.

To briefly explain these, the first rule always includes the "wa" case in the original sentence, the "ga" case is arbitrary, and there are no other "ni", "de", and "kara" cases. If the object of the "wa" case is any of "performance", "performance", "ability", "efficiency", or "arm", it has a "ga"
The case is the subject, the "wa" case is the object, and the translation for "raise" is "improve" or "increase".

The second rule has a structure in which the "wa" and "ni" cases are always included, the "ga" case is arbitrary, and the other "de" and "from" cases are absent. If the object of the "ni" case means a person, the subject of the "ga" case,
The case of "wa" is structurally transformed into the direct object, the case of "ni" into the indirect object, and the translated word of "ageru" is defined as "give".

The third rule has the same structure as the structure referred to in the second rule, but the object of the "o" case is
If the object of the "ni" case means a place, the "ga" case is the subject, the "o" case is the object, and the "ni" case is transformed into a prepositional phrase starting with "on", The translation of "give up" is "put".

The last rule is the default rule, and the above five essential cases need not be included in the sentence, and if they are, the "ga" case is the subject and the "wo" case is Case is the object, "ni" is a preposition that starts with to, and "de" is "wi."
The prepositional phrase that begins with “th”, the “from” case is structurally transformed into the prepositional phrase that begins with “from”, and the translated words are “raise”, “lift”,
It is said to be one of "fry", "land", "get", and "vomit".

The input unit 101 inputs a document to be subjected to language processing. Then, the system control unit 102 stores this document in the document storage unit 107 as data in the first language.
Next, a translation command is input from the input unit 101. For example, the case of inputting a Japanese sentence and translating it into an English sentence will be described. With this command, the system control unit 102 causes the data in the first language stored in the document storage unit 107 to be edited by the editing unit 103.
Via the translation unit 104, and the translation unit 104 refers to the dictionary of the translation dictionary unit 105 selected by the system control unit 102 for the data of the first language selected by the system control unit 102. While performing the translation process.

FIG. 2 is a flow of the whole process of translating an original document (first language) into a designated language (second language).

First, when the translation target region of the original document is designated by the input unit 101, the system control unit 102 checks the sentences in the designated region to check the number of sentences. Then, the system control unit 102 sets the number of the examined sentences in the variable B in step S1, and sets "1" in the variable C in step S2. Then, in step S4, the system control unit 102 transfers the sentence to the translation unit 104 and gives a translation instruction in order to translate the Bth sentence.

As a result, the translation unit 104 selects the translation dictionary unit selected by the system control unit 102 for the first language data which is the sentence passed from the system control unit 102.
The translation process is performed with reference to the 105 dictionary. In order to repeat this translation processing for the number of input sentences, the system control unit 102 adds "1" to the variable C in step S5. In step S3, it is determined whether the variable C is larger than the variable B. As a result, if the variable C is smaller than the value of the variable B in which the number of sentences is set, the process proceeds to step S4 and the translation unit 104
, The sentence corresponding to the B-th (variable B) is read from the document storage unit 107 and given, and the translation unit 104 translates it by referring to the dictionary of the translation dictionary unit 105.

Then, the system control unit 102 adds "1" to the variable C and makes a determination in step S3. By performing this process until the variable C becomes larger than the variable B,
The specified original document can be completely translated by the translation unit 104. The system control unit 102 stores the second language translated by the translation unit 104 in the document storage unit 107.

Here, the translation process in step S4 will be described in detail with reference to FIGS.
Now, a language processing target sentence is directly input from the input unit 101, and a processing command is given, or a specified original sentence is read from the document storage unit 107, and a language processing target sentence is input. When a processing command is given, the system control unit 102 first passes these to the translation unit 104. The translation unit 104 sets the processing certainty factor to “0” in step S201, and the translation unit 104 then executes step S2.
In 02, morphological analysis is performed on the input sentence.

Here, the translation unit 104 refers to the inflection change dictionary 105a in the translation dictionary unit 105 to perform a morpheme analysis on each word or idiom constituting the input original sentence, and then at step S203, this morpheme analysis result. By searching the word / phrase dictionary 105b in accordance with the dictionary, information on the part of speech and the translated word of the word in the input sentence is obtained. Note that various morphological analysis methods have already been proposed, so description thereof will be omitted. The same applies to syntax analysis and the like described later.

If the input sentence includes an unknown word in step S204, the processing certainty factor is subtracted by the number of words of the unknown word in step S205. In the example of FIG. 03, since nothing is included in the unknown word, nothing is done.

Next, in step S206, the translation unit 104 performs syntax analysis of the input sentence using the part-of-speech information obtained in the dictionary search in step S203 and the analysis grammar dictionary 105c.

As a result, it can be seen that the sentence (1) has a structure of "[elderly person <was] [big sea bream <was>] [raises], and the sentence (2)" [me <ga>] [ You can see that it has the structure of "pencil <>>] [you <to>] [given]".
Sentence (3) is "[manufacturer <is>] [car performance <is>]
It turns out that it has a structure of "[Give up]", and sentence (4) reads "[He <is>] [Shelf <to>] [Object <">] [Give] "
It is understood that the structure has the following, and the sentence (5) is "[Company <is>] [Performance of general-purpose machine <">] [Raise].
You can see that it has the structure. Note that sentence (2) and sentence (6)
Are the same sentence.

When the syntactic analysis is completed, the translation unit 104 then moves to step S207 and determines whether the syntactic analysis process is successful. As a result, if the analysis is successful, the process proceeds to step S209, and if the analysis is unsuccessful, the process proceeds to step S208. In step S208, the value of the processing certainty factor is reduced by "20". And step S2
Go to 09.

Subsequently, the translation unit 104 determines in step S209 whether these processing targets are phrases (titles) or sentences. In general, since the phrase has a simpler structure, the probability that a correct translation is obtained is higher. Therefore, in the case of the phrase, the value of the processing certainty factor is incremented by "1" in step S232 (increment process).
After that, the structure conversion is performed in step S212b, and when the structure conversion is completed, it is checked in step S213b whether or not this structure conversion is successful.

As a result, if the conversion is successful, the translation unit 10
4 proceeds to step S233, and if the conversion is unsuccessful, the translation unit 104 performs processing of subtracting “20” from the certainty factor value at step S214b, and proceeds to step S233.
Since all the examples in FIG. 03 are sentences, the process proceeds to step S210.

When it is determined that the sentence is a sentence, the translation unit 104 next determines in step S210 the type of sentence, that is, whether the sentence is a compound sentence, a compound sentence, or a simple sentence. In this case, the number of nodes is counted, and the value of the processing certainty factor is deducted by the amount obtained by subtracting "1" from the number in step S211. Since all the examples in FIG. 8 are simple sentences, the processing certainty factor does not change.

Next, the translation unit 104 performs step S212a.
Proceed to and perform structural conversion. Here, according to the result of the syntactic analysis, the conversion grammar dictionary 105d of the translation dictionary unit 105 is used to perform structure conversion on the input original sentence to obtain a structure candidate of the translated sentence. Then, in step S213a, it is determined whether or not the structure of the translated sentence has been successfully generated in step S213a. If the structure conversion is successful, the process proceeds to step S215. If not, the value of the processing certainty factor is decreased by 20 in step S214a.

By the way, for the example sentence of FIG.
According to the rule of (1), sentence (1) does not apply to the conditions of application of the first rule, the second rule, and the third rule, and the fourth rule, which is the default rule, applies. Sentence (2) structurally matches the second rule and the third rule, but since the "you" shown in the "ni" case means a person, not a place, the second condition is Only those that match

Therefore, the second rule applies. Sentence (3)
This sentence applies because the sentence "(5)" and "performance" indicated by the case "wa" meet the conditions of application of the first rule. Sentence
In (4), the "thing" indicated by the "wo" case means the thing, and the "ni"
The third rule applies because the "shelf" shown on the case means a place. Thus, sentence (1) has the fourth rule, sentence (2) has the second rule, sentences (3) and (5) have the first rule, and sentence (4) has the third rule. Rules will be applied respectively.

In the subsequent steps S215 to S225, it is assumed that a more accurate translation will be obtained if the structure conversion is performed using a rule that refers to more information.
The score to be added to the certainty factor is calculated according to the rule applied to each.

First, the translation unit 104 calculates the number of predicates included in each sentence in step S215, and the number is used as a variable.
Set to i. Then, in step S216, the variable j is set to "0" and the variable k is set to "1". Then, for each predicate in the input sentence, the processing certainty factor is added as follows according to the applied rule.

That is, in step S218, it is determined whether or not the rule applied to the j-th predicate from the beginning of the input sentence is the default rule, and if it is not the default rule, "1" is added to the variable k. To do.

Next, the translation unit 104 checks in step S220 whether or not the original word itself is referred to when applying the rule. As a result, if it is referred to, the step S221 is performed.
Add "2" to the variable k with. Further step S222
In step S22, it is checked whether or not the semantic information is referred to when the rule is applied.
In step 3, "1" is added to the variable k.

Then, in order to repeat this series of operations for the number of predicates, the translation unit 104 determines the variable j in step S224.
Add “1” to. Then, the translation unit 104 performs this according to the determination condition in step S217 for comparing the value of the variable j with the value of the variable i until the value of the variable j becomes larger than the value of the variable i. After exiting the loop, the translation unit 104 performs step S.
In 225, the total number k is divided by the number i of predicates, the added average value is obtained, and this is taken as the processing certainty factor in the structural transformation,
It is added to the processing certainty factor obtained before the structural conversion.

When the above processing is performed on the example sentence of FIG.
The processing confidence of sentence (1) does not change, and sentence (2) and sentence (4)
(6) is not a default rule, and because semantic information is referenced, "2" is added, sentence (3) and sentence (5) are not a default rule, and the original language itself is referenced. Therefore, "3" points will be added.

Next, in step S233, syntax generation is performed based on the generation grammar dictionary 105e. Here, the word order of the translated sentence is determined from the structure of the original sentence, and then converted into a word string. Further, in step S234, it is checked whether or not the syntax generation processing has succeeded. If the result is that the processing has failed, the value of the processing certainty factor is reduced by "20" in step S235. Then, the process proceeds to step S236, and the morpheme grammar dictionary.
Morphological generation is performed by applying 105f, and the ending of the word is changed to complete the translated sentence.

The above is the translation processing operation for each sentence.

Returning to the explanation of FIG. System control unit
102 successively sends sentences to the translation unit 104 until C> B is reached,
Although the translation processing as described above is performed by the translation unit 104,
When C> B is reached, the system control unit 102 shifts the processing to the editing unit 103 and causes the editing unit 103 to perform the processing of step S6.

In step S6, under the control of the system control unit 102, the editing unit 103 collects the ones having the same original text, which are translated by the translation unit 104, so that the same ones are not repeatedly displayed. Perform adjustment processing.

When the original text is written in English or symbols, such as "X = A / (B + C)" or "1.1 Introduction" Since post-editing is unnecessary, it is processed so that it is not displayed on the post-editing screen.

Further, under the control of the system control unit 102, the editing unit 103 rearranges the translated sentences in the order of low translation accuracy according to the processing certainty factor for each sentence obtained by the translation unit 104 as described above, and finally, In step S7, control is performed so that the translated text together with the original document and its sentence number are displayed on the post-editing screen on the display unit 106.

For example, in the case of the document shown in FIG. 8, the final confidence value is "0" for sentence (1), sentence (2), sentence (4) and sentence (6).
12 is "2" and sentences (3) and (5) are "3".
The display state is as shown in.

On the other hand, when the confidence values are displayed in descending order, the result is as shown in FIG. Here, the operator may be inquired to select whether the confidence factor values are displayed in the ascending order or in the descending order.

After this, the worker performs post-editing in step S8.

As a result, according to the above-mentioned embodiment, the processing reliability of the translation in each processing stage of the translation is calculated, and according to the calculated value, priority is given to the one having a low possibility of mistranslation, or the mistranslation is selected. The operator can review the translated sentence efficiently by giving priority to the display with the highest possibility, displaying the same sentence only once, and excluding the sentence that does not need translation from the display target.

The present invention is not limited to the above embodiments, and can be carried out by appropriately modifying it within the scope of the invention. First, as shown in FIG. 14, the numerical value of the certainty factor may be displayed on the screen together with the reason for adding or subtracting points, for the convenience of the operator. Alternatively, FIG.
As shown in, it is possible to display the original sentence and the translated sentence having the same reason according to the reason of the point addition / deduction according to the order of the confidence value.

When editing the translated text so that it is easy to read,
Figure 16 to help operators create better translations
As shown in, the context before and after the Japanese sentence can be displayed on the screen. Furthermore, not in order, but with confidence of “-10” or less, “-9 to 0”, and “1 to 9”
It is also possible to select and display the translated sentence according to the numerical range of the certainty factor. An example thereof is shown in FIG.

In addition, there are additional points in calculating the certainty factor.
The actual points deducted are not absolute, and may be changed by the user depending on the importance of which part of the translation is required to be accurate.

Next, in calculating the certainty factor, other parameters not mentioned in the embodiment can be added as long as they are related to the accuracy of translation.

First, although it has been assumed in the above description that only the standard term dictionary is used as the word / idiom dictionary, a user dictionary or a technical term dictionary may be used together, and various relationships between concepts may be used. It is also possible to use the described concept dictionary as a semantic classification dictionary. Therefore, the certainty factor may be changed depending on the type of dictionary used.

The user dictionary is applied if there is a match in the sentence to be translated at the time of parsing. In this dictionary, it is possible to add detailed translation information unique to the user, such as the specification of arbitrary cases not included in the standard term dictionary.

FIG. 18 shows an example of knowledge registered in the user dictionary.

In the first example, the original sentence "... apply oil to a machine ..." is translated into English by "apply oil to the machin".
This is a rule for translating as “e.” In this way, by using the user dictionary to generate a specific translated sentence for a specific word, it is possible for the user to have more preference than using the standard term dictionary. Therefore, if the rules in the user dictionary are applied, it is conceivable to add, for example, 2 points to the value of the processing certainty factor, and the technical term dictionary specified by the user is used. In this case, the accuracy of the translated word is likely to be improved, and therefore the processing certainty factor may be added at this time as well.

Further, regarding the display, by displaying the sentences which are similar to each other in that the same rule is applied or the sentence patterns are the same, only the individual words are different from each other, they are displayed close to each other. The person can efficiently review the translated text.

As described above, in the first embodiment, when translating the original sentence,
For each translated sentence, the certainty factor, which is the degree of possibility that the content was translated correctly, is calculated as a numerical value, and at the time of post-editing, the original sentence, the translated sentence, and the processing certainty value are displayed in ascending or descending order. Therefore, at the time of post-editing, it is possible to perform high-quality translations by displaying the numerical values of the processing certainty factors in order from the one with the lowest processing certainty factor, and making adjustments. Depending on the, the items with a high possibility of mistranslation can be displayed with priority and reworked, or the items with a low possibility of mistranslation can be displayed with priority and reworked. It is possible to utilize the numerical value of the processing certainty factor in various selections, such as displaying only once and excluding sentences that do not need to be translated. Review of translated text It has been able to be, so that it is possible to reduce the burden.

Therefore, when translating a natural language sentence into another type of natural language sentence, it is possible to make the post-editing work for the translated sentence efficient and reduce the burden on the operator. The device is obtained.

(Second Embodiment) As a second embodiment, consider the case where a translation based on an example is incorporated.

Currently, the machine translation methods are rule-based and example-base.
There are two streams of d).

All of the above examples were rule-based. Here, the translation based on an example means that a translation example of a sentence similar to a sentence to be translated is searched without depending on an abstract rule, and the translation is imitated and translated. The basic structure of the software is the same as that of the first embodiment.

The flow in the case where the translation process for the input document from the input unit 101 is performed based on an example will be described with reference to the flowchart of FIG.

Now, a sentence to be language processed is directly input from the input unit 101, and a processing command is given, or an original sentence to be designated is read from the document storage unit 107, and a sentence to be language processed is obtained. When input and a processing command is given, first, the system control unit 102 causes the translation unit 10 to
Pass these to 4.

The translation unit 104 executes morphological analysis in step A1. Here, the use dictionary of the word dictionary 105
By searching 105a and the word / phrase dictionary 105b, the part of speech, original form, and utilization information of the word in the input document can be obtained.
Then, in step A2, the syntax analysis is performed.

At this step A2, the grammatical dictionary 105c is analyzed by using the part-of-speech information obtained by the morphological analysis at step A1.
While parsing the input sentence with reference to, the syntactic structure of the first language is generated. The processing up to this point is common to the case of the rule-based translation method.

Next, in the translation unit 104, in step A3,
Perform an optimum matching search. Here, a sentence similar to the given input sentence and its translated sentence are searched from the parallel translation database 1201. If similar sentences are found, the process proceeds to step A4.

In step A4, application adjustment is performed. That is, the translation is performed with reference to the retrieved similar bilingual translation, adjusting the sentence to be translated.

If a similar sentence is not retrieved, translation according to the rules described in the flowcharts of FIGS. 3 to 7 is performed in step A6.

The translated sentence is completed as described above.

Since the parallel translation database used here is a translation example translated by humans, the expression is not influenced by the translation tone of the machine output, and the translation can be output by using the example. In other words, the example system translates using a finer level rule or a more precise rule than the conventional system.

On the other hand, a drawback is that there is not always a similar parallel translation example. Therefore, here, when there is no similar sentence, it is translated based on the rule. Therefore, when a similar sentence can be searched, it is conceivable to add 10 points to the processing certainty factor, for example. Furthermore, the following may be considered as a modification of the translation method based on an actual example.

The flow of the processing performed based on this method will be described with reference to the flowchart of FIG.

When a document to be language processed is input, the translation unit 104 performs the same steps as Steps A1 and A2 described with reference to FIG. 19 in Steps B1 and B2.
Done in.

Next, the translation unit 104 uses the transformation grammar dictionary 105d in step B3 to perform the structural transformation on the input original sentence,
Find a candidate structure for a translated sentence. In step B4, it is judged whether or not there are a plurality of structure candidates, and if there is, a step B5.
Proceed to. In step B5, the one having the structure found in the parallel translation database 1201 is set as the first translated sentence structure.
Generate syntax.

For example, the sentence "Introducing a new closed CSCW environment increases the overhead associated with cooperative work." Is a "subordinate clause + main clause" structure (in this case, the translation is "If we newly introduce closed CSCW enviro
nments, overhead incooperation will increase. ”and" inanimate subject + predicate "structure (in this case,
For example, “Introducing closed CSCW environments w
ill increaseoverhead in cooperation. ”), there are two possibilities.

Then, the translation unit 104 searches the parallel translation database 1201 in step B5, and the Japanese sentence is "-is-is", and the original sentence and the translated sentence are "inanimate subject + predicate" as the translated sentence. If there is a pair of, the latter translation will be output.

If it is determined in step B4 that there is one structure candidate, syntax generation is performed based on the generation grammar dictionary 105e in step B6. Finally, in step B7, morpheme generation is performed by applying the morpheme generation grammar dictionary 105f, the ending of the word is changed, and the translated sentence is completed.

Also in the case where translation is performed through step B5 as described above, it is conceivable to add 10 points to the processing certainty factor.

Finally, the confidence factor of the system for the translation result generated by the translation unit 104 in this way is measured, and if the confidence factor exceeds a specific reference value, it is stored as data together with the original text of the first natural language. By doing so, the flow of processing for creating a knowledge database will be described.

In this case, the processing performed is exactly the same as the post-editing in the first embodiment except for the final stage. The differences are as follows. The point is that, when a pair of a translated sentence and an original sentence that is completed in a natural and easy-to-read sentence is designated by a human according to the flowchart of FIG. 21, it is accumulated in the parallel translation database.

Therefore, the input section 101 is newly provided with an accumulation confirmation key, an accumulation cancellation key, and a registration switching mode key.

The accumulation confirmation key in the input unit 101 is for designating the parallel translations to be stored in the parallel translation database, and the accumulation cancel key is for deleting the parallel translations once accumulated in the parallel translation database from the parallel translation database, or the parallel translations for which post-editing has been completed. However, it is for designating parallel translations that are not desired to be stored in the parallel translation database.

The registration switching mode key in the input section 101 is an automatic mode that is automatically sent to the bilingual database when post-editing is completed, and an operator's confirmation is required after post-editing to register in the bilingual database. It is for switching to the confirmation mode.

That is, when the input section 101 is operated to issue an instruction for post-editing, the system control section 102 operates the editing section 103 and the post-editing processing is carried out by the editing section 103 (step C1). This is based on the contents displayed on the display unit 106 and is edited by the post-editor by operating the input unit 101.
Repair work is performed under the processing control of. When the post-editing process by the editing unit 103 is completed, the system control unit
At step C2, whether or not the registration mode is automatic is checked at step C2. If the registration mode is automatic, the system control unit 102 at step C3 stores the post-editing processing result as it is in the parallel translation database.

On the other hand, in the confirmation mode, the system control unit
The step 102 performs an operation for asking the post editor to confirm whether or not to register in step C4. This is performed by displaying a confirmation message asking whether or not to register on the display unit 106 and waiting for an instruction. As a result, the post-editor
If the confirmation key is pressed by the operation of 101 (step C5), the system control unit 102 stores it in the parallel translation database.

If the confirm key is not pressed and the cancel key is pressed in step C6, the system control unit 102 ends without storing anything in the parallel translation database.

As described above, in the second embodiment, a translation example of a sentence similar to the sentence to be translated is searched from the database, and the translation is imitated and translated.
In addition, it is possible to calculate the certainty factor in each process of translation and display the pairs of the original sentence and the translated sentence in descending numerical order or in descending numerical order in consideration of the similarity and identity between the numerical value and the sentence. It is the one. As a result, it is possible to save the trouble of searching for a sentence that may be mistranslated, and to perform correction work efficiently, such as minimizing unnecessary translation check.

Specifically, when a certain degree of accuracy is required within a certain period of time, it is a good idea to display the processing in ascending order. In addition, even if the developer of the machine translation system knows the weaknesses of his own system and improves the weaknesses, it is possible to expect efficient development by displaying the weakest processing confidences in order.

On the other hand, when the amount of translation is large and it is required to deliver the translation after it has been completed,
It is a good idea to display in order from the one with the highest processing certainty, which does not take a long time to correct.

In conventional machine translation, post-editing, which human translation does not have, sometimes lengthened the work time. However, by improving the efficiency of the translation work by this method, the operator can perform highly productive translation. Can be engaged in.

Further, as an application of the editing work after this, by displaying in order from the one with the highest processing certainty, the pair of the original sentence and the translated sentence edited and corrected by the worker can be efficiently and easily accumulated in the parallel translation database. . By accumulating the bilingual database in this way, it is possible to strengthen the user dictionary, increase the number of actual examples that are indispensable for the actual example-based translation method, and reduce the post-editing work.

The present invention is not limited to the embodiments described above and shown in the drawings, but can be appropriately modified and implemented within the scope of the invention.

[0118]

As described above in detail, according to the present invention,
Input means for inputting texts in at least two languages and instructions for the system, translation processing means for performing morphological analysis of the input texts in the first language and performing translation processing, and knowledge used for performing the translation processing. It has a stored translation dictionary and a display means for displaying the translated sentence written in the second language and information from the system, measures the certainty factor for the obtained translation result, and confirms it at the time of post-editing. Since it is a machine translation device configured to be able to display the original sentence and the translated sentence as a pair on the display unit in the ascending order or the high order, and to perform the post-editing process based on this,
Minimize the physical and mental burden on post-editing workers,
The post-editing work can be performed efficiently, and the edited / corrected translated text can be efficiently and easily stored together with the original text in the parallel translation database.

[Brief description of drawings]

FIG. 1 is a diagram for explaining an embodiment of the present invention,
The block diagram which shows the system configuration of the device of this invention.

FIG. 2 is a diagram for explaining an embodiment of the present invention,
3 is a flowchart showing the overall processing flow of the system in the first embodiment of the present invention.

FIG. 3 is a diagram for explaining an embodiment of the present invention,
The flowchart which shows the flow of the translation process of the same Example when translating Japanese into English.

FIG. 4 is a diagram for explaining an embodiment of the present invention,
The flowchart which shows the flow of the translation process of the same Example when translating Japanese into English.

FIG. 5 is a diagram for explaining an embodiment of the present invention,
The flowchart which shows the flow of the translation process of the same Example when translating Japanese into English.

FIG. 6 is a diagram for explaining an embodiment of the present invention,
The flowchart which shows the flow of the translation process of the same Example when translating Japanese into English.

FIG. 7 is a diagram for explaining an embodiment of the present invention,
The flowchart which shows the flow of the translation process of the same Example when translating Japanese into English.

FIG. 8 is a diagram for explaining an embodiment of the present invention,
The figure which shows the specific example of the document for describing the Example of this invention.

FIG. 9 is a view for explaining the embodiment of the present invention,
The figure which shows an example of the utilization change dictionary in the word dictionary used for the Example of this invention.

FIG. 10 is a diagram for explaining the embodiment of the present invention, showing an example of the word / idiom dictionary in the word dictionary used in the embodiment of the present invention.

FIG. 11 is a diagram for explaining an embodiment of the present invention, showing an example of a conversion grammar dictionary in a word dictionary used in the embodiment of the present invention.

FIG. 12 is a diagram for explaining the embodiment of the present invention, showing an example of the layout of the screen of the display unit used in the embodiment of the present invention.

FIG. 13 is a diagram for explaining the embodiment of the present invention, showing an example of the layout of the screen of the display unit used in the embodiment of the present invention.

FIG. 14 is a diagram for explaining the embodiment of the present invention, showing an example of the layout of the screen of the display unit used in the embodiment of the present invention.

FIG. 15 is a diagram for explaining the embodiment of the present invention, showing an example of the layout of the screen of the display unit used in the embodiment of the present invention.

FIG. 16 is a diagram for explaining the embodiment of the present invention, showing an example of the layout of the screen of the display unit used in the embodiment of the present invention.

FIG. 17 is a diagram for explaining the embodiment of the present invention, showing an example of the layout of the screen of the display unit used in the embodiment of the present invention.

FIG. 18 is a diagram for explaining the embodiment of the present invention and is a diagram showing an example of a user dictionary used in the embodiment of the present invention.

FIG. 19 is a diagram for explaining the embodiment of the present invention and is a flowchart showing a flow of a translation process when translating Japanese into English in the embodiment of the present invention.

FIG. 20 is a diagram for explaining the embodiment of the present invention and is a flowchart showing a flow of a translation process when translating Japanese into English in the embodiment of the present invention.

FIG. 21 is a diagram for explaining the second embodiment of the present invention and is a flowchart showing the flow of processing of the same embodiment when performing translation and storing the result in a database.

[Explanation of symbols]

 101 ... input section 102 ... system control section 103 ... editing section 104 ... translation section 105 ... translation dictionary section 106 ... display section 107 ... document storage section 1201 ... parallel translation database

Claims (2)

[Claims] 1. A machine translation device for translating a first natural language sentence into another different natural language sentence, and a translation dictionary storing knowledge used for performing a translation process, and a first input dictionary. A translation processing unit that translates a sentence in a natural language into the other natural language by using the translation dictionary, obtains a certainty factor for the translation result, and the translation is performed in ascending or descending order of the certainty factor for the translation result. A natural language and a first natural language sentence that is the original sentence thereof are rearranged, and a display unit that sequentially displays the rearranged sentence pairs during post-editing. Machine translation device. 2. A machine translation device for translating a first natural language sentence into another different natural language sentence, which holds a translation dictionary storing knowledge used for performing a translation process, and an original sentence and a translated sentence. If a database and a translation example of a sentence similar to the sentence to be translated and translated from the input first natural language sentence into the other natural language using the translation dictionary are stored in the database, A translation processing unit that imitates a translation and obtains a certainty factor for the translation result, and a certainty factor of the system for the translation result is measured, and a high certainty factor is stored as data together with the original text of the first language. By accumulating, a control means for supporting the creation of the knowledge database, and the translated natural language and its original sentence are arranged in ascending or descending order of confidence in the translation result. Means reorders each pair of sentences in the first natural language, machine translation apparatus that was characterized by having a display means for displaying sequentially at the rear edit pairs of reordered sentence.