JPH02148261A - Machine translation system with syntax learning function - Google Patents

Abstract

PURPOSE:To reduce the need for a user to subject each original sentence to the same preprocessing by registering the change of modifier/modificand relations and the change of translated words due to post-editing as patterns and preferentially subjecting the same patterns of original sentences to the same processing hereafter. CONSTITUTION:A modifier/modificand relation changing device 7 which leads out information from a translation information preserving device 6 and changes modifier/ modificand relations in accordance with grammar 4 and a learning pattern setting device 10 which sets the application condition for preferential use of the change processing of modifier/modificand relations and translated words changed by the modifier/modificand relation changing device 7 and a post-editing device 8 are provided. The change of modifier/modificand relations and the change of translated words due to post-editing are registered as patterns in a learning information preserving device 9, and the same patterns of original sentences are preferentially subjected to the same processing hereafter to extend the application range of learning informa tion up to the modifier/modificand relation structure. Thus, the work volume of pre- editing is reduced, and translation results of higher quality are obtained as the system learns more.

Description

[Detailed Description of the Invention] [Industrial Application Field] This invention relates to a machine translation system with a syntax learning function, especially -
This article relates to a machine translation system with a syntax learning function that incorporates corrected dependencies as a priority process during analysis and generation.

[Prior art] Figure 5 shows nine examples of literature, Yoshi Tsutsumi's [Editor for @machine translation with simple learning function] (Information Processing Society of Japan Vol. 3).
3rd National Conference, pp. 17711772) and [Translation Work Environment in the English-Japanese Machine Translation System S HALT J
(Information Processing Society of Japan 34th National Convention, pp. 1269-1
．． 270) is a block diagram showing the configuration of the conventional machine translation system shown in FIG. In the figure, (1) is the input device for the original text written in the first language, and (2) is the input device (1).
(3) is a dictionary that has word information between the languages to be translated; (4) is a grammar that describes the relationships between words; (5) is a dictionary (3) and a translation device that uses the grammar (4) to translate the original text in the first language into a translated text in the second language; (8) is the post-editing device that corrects the translated text translated by the translation device (5); 9) is a learning information storage device that stores learning information up to that point, (11) is a translation output device, and (12> uses information from the f & editing device (8) and learning information from the learning information storage device (9). and the 6th section described below.
This is a learning device that learns according to the steps shown in the figure.

FIG. 6 is a diagram showing the learning procedure for edited information, where (24) is the post-edited information edited by the post-editing device (8);
(25) is the learning information learned so far stored in the learning information storage device (9), and (26> to (28) are the ranges to be learned from the post-editing information (24) and learning information (25). The judgment step, <29) is the judgment step 26
) to judgment step (28), which is a process step for forming a rule when the judgment in step (28) is established.

The conventional machine translation system is configured as described above, and the user who inputs the input sentence (41) with the input device (1) inputs the input sentence (41) with the pre-editing device (1) because there is no syntactic ambiguity.
Translate using the translation device (5) without editing in step 2).

Next, the first user looks at the translated text and uses the post-editing device (8) to read the translated text (
43), the translation of 'pou+er facy' is corrected as 'power function' to 'power supply function'.

Next, one system collects post-editing information (
24) and learning information (25) in the learning information storage device (9)
Let's move on to judgment. In the initial state, the learning information (25) does not have any information, so the 2nd size (new step (26)
), from the post-editing information (24), the rule 45) that if ``potuer'' is followed by a part of speech - ``noun'', ``p Ou e r'' is translated as ``power source'' (45) is generated in the rule processing step (29). and register it in the learning information storage bag 7i (9). In other words, the tree of rule (45) is traced from the right.

Next, when the system translates the input sentence (42), the power of the input sentence (42) is calculated from the information in the learning information storage device (9).
u+1ndou+" is translated as "power window," and the user corrects it to "electric window" as in the translated sentence (44). The system takes this information as information, and this time, since it is not possible to classify using only the nine parts of speech, in the judgment step (27), the rule (46) in which the classification has been changed up to the meaning marker is re-registered in the learning information storage device (9).

[Problem to be solved by the invention] The above-mentioned conventional lf! In the lIl'1Ili translation system.

It works effectively when the syntactic structure is deterministic, but in a sentence like ``I'm going to Tabe's house in Tokyo,''``Tabe'' is 1 Tokyo.''
[For sentences with syntactic ambiguity that relate to house J, it is necessary to learn the post-edited information, and after that, it is not possible to give priority to the information at the time of translation, so post-editing is performed every time it is translated.1
For example, literature, Hirai et al. [A Study on Pre-editing System for Japanese-English Machine Translation] (Information Processing Society of Japan 35th National Conference,
pp1243-1244) is a well-known technique for removing dependency ambiguity, that is, by pre-editing °[
, “]°, etc., to specify the range of dependencies, such as “I am going to Abe's Tokyo house.” There was a problem that it was troublesome for the user.

This invention was devised to solve these problems, and the present invention provides a syntax learning function that allows users to easily change dependencies and to reduce the amount of preprocessing required by the user by providing a means for storing these changes. The purpose is to obtain an integrated machine translation system.

[Means for Solving the Problems] A machine translation system with a syntax learning function according to the present invention includes a storage means for storing dictionary information, grammar information, and dependency information used in one translation, and an analysis method after translation. a display/changing means for displaying and changing dependency information; a determining means for determining whether the modification is grammatically correct; and a retranslation means for retranslating in accordance with the modification.
The present invention is provided with a registration means for registering changes in the modification and changes in translated words by post-editing as a button, and a processing means for subsequently performing similar processing preferentially on similar patterns in the original text.

[Operation] In this invention, a processing pattern is registered on the original text along with application conditions, and after registration, the processing pattern is applied preferentially, thereby eliminating the need for the user to perform similar preprocessing for each original text. reduce.

[Embodiment] Figure 1 shows a machine with a syntax learning function according to an embodiment of the present invention.
FIG. 1 is a block diagram showing the configuration of the i-translation system. In the figure, (1) to (5), (9), and (11) are the same as the conventional one. (6) is a translation information storage device that stores the processing of the translation device (5); (7) is a translation information storage device (6)
Extract the information from 1 grammar (4) and change the dependency.The modification device is the modification device, and in (10), the modification device is the 1 system that was changed by the modification device (7) and the post-editing device (8). A learning pattern setting device that sets application conditions for preferentially using the Uke and translation change processing, (30) and (31) are examples of input sentences input to the input device (1), (32), (3
3) is an example of translation information stored in the translation information storage device (6), (36) is an example of pre-editing, (34), (35), (3)
7) is an example of the translation result.

FIG. 2 is a diagram illustrating the plain state of modification processing. [
In the figure, (13) is a step in which the user determines whether to change the dependency, (1, 4) is a setting step in which the user sets the modification part of the dependency, and (15) is a setting step (14). ), the necessary modification information is retrieved from the translation information storage device (6) and displayed in the display step, and (16) is the change step (17) in which the user changes the modification information. Judgment step 118) for determining whether the modification of the dependency made by ``17'' is grammatically correct is the re-translation step for retranslating the changed part of the modification ゛τ of the judgment step 17), and (19) is In step (38) of updating translation information that has been changed due to changes in these dependencies, the user can change the scope of the dependencies by 117T.
If, an example of an input sentence, (39) is an example of information where the dependency displayed by the system according to the specification of an example input sentence (38), (4
0) is an example in which the user has changed the dependency, and (37) is an example after retranslation.

Figure 3 shows 7 changes: change device (7) and post-edit device (8)
FIG. 4 is a diagram showing a procedure for setting application conditions for preferentially using the translation processing pattern changed in . In the figure, (20) is dependent on the change device (7) and the post-edit device (8).
), and the pre-editing device (2) displays the modified translation information of the modified part (21) to the user. The modification is a decision step in which the results of changes made in the change device (7) and post-editing device (8) and pre-editing in the pre-editing device (2) are registered, and it is determined whether or not to be reflected in the processing after registration.

(22) is affected by the change device (7) and the post-edit device (8)
(23) is a setting step (22) in which the user is asked to set attributes and attribute values in order to apply changes in the process and pre-editing processing patterns in the pre-editing device (2);
a storage step of storing the processing pattern in the learning information storage device (9) using the conditions set in step (41) as check conditions;
) is an example of the translation information displayed in the display step (20), (42) is an example of the attributes and attribute values selected by the user as processing conditions, and (43) is the learning pattern saved in the save step (23). This is an example.

FIG. 4 is a table showing details of examples (41) to (43) in FIG. 3.

In the machine translation system with a syntax learning function configured as described above, the user who manually created the input sentence (30) in Figure 1 pre-edited the input sentence (30) using the pre-editing device (2). It will be translated later with a translation device (3) or translated directly with a translation device (3).

First, the translation A position (
Consider the case where the input sentence (30) is passed to 3). The processing result in the translation device (3) is stored in the translation information storage device (6) as a set of translation information (32) and the structure, word attributes, and attribute values. The user receives the translation result (34)
Look at ``the house in Taro's
Knowing that the 114 structure in the ``Osa ka'' part is incorrect, he requests a modification of the dependency at the 'PI cut step (13) in Figure 2.This system starts at the setting step (14).
allows the user to specify the part for which the dependency is to be changed, as shown in example sentence (38). In this system, the modification of the specified part is performed by extracting the information from the translation information storage device (6) and displaying the information in step (15).
). The user looks at this information example (39), learns that "Abe" is related to "Osaka", judges that "Aabe" must be related to "House", and performs the change step. In (16), change to the modification example (40). After this, the system in decision step (17)
Using the grammar (4), it is determined whether the modification of the modification example (40) is grammatically correct. If it is correct, retranslation is performed in a retranslation step (18) to obtain a retranslated example (37).

If it is not correct, the process proceeds to a display step (15) and a change step (16) to request modification of the dependency again.

Information that the dependency has been changed and translation information after the change (
40) is stored in the translation information storage device (6), and the processing in the modification device (7) ends.

On the other hand, when the input sentence (30) is pre-edited by the pre-editing device (2), the sentence in which the range of dependency is specified with '[',']' in the pre-editing example (36) is passed to the translation device (3). The sentence of retranslation example (37) can be obtained directly.

Subsequently, post-editing is performed in the post-editing device (8), but since it is not necessary in the re-translation example (37), the process moves to the learning pattern setting device (10). Learning button setting device (10)
In this case, the translation information is retrieved from the translation information storage device (6), and it is determined whether the modification is specified in the pre-editing device (2) or the modification device (7), and if there is a modification, the It is displayed in a display step (20). For sentence example (30), in either of the above processes, translation information display example (41
) is displayed. The user determines whether or not to incorporate the displayed content as a learning button in a decision step (21). When importing the changed part of the translation information display example (41) as a learning pattern, the user should refer to the display example (41) in Figure 4.
From the table in 41), apply conditions to the example of attributes and attribute values (42)
Specify the arbitrary side as t■t of the attribute and attribute value surrounded by . In this example (42), the word sequence is “noun, t
8 particles 1 noun, case particles 1 noun”, 1
゛At number 1α, the pair (attribute, attribute value) is r (meaning 1 person's name),
(m-word, of), (meaning, place name) (word, of) 2〈meaning,
ken'PA) Indicates that J is the applicable condition. This is stored in the learning pattern example (43) in the storage step (23).
) form, [arrangement of parts of speech, application conditions, processing], is registered in the learning information storage device (9>), and the processing in the learning pattern setting device (10) is completed.

Next, when trying to translate the input sentence example (31), the user can divide it into words using the translation device (5) and convert it into 9 grammar words (4) without pre-editing it using the pre-editing device (2). By performing translation processing using the information in the learning information storage device (9), the translation information storage device (6) stores translation information examples (33
) is saved, and the translation result 35) is obtained as the 1g11 translation result.

In addition, in the above example, the criteria for setting conditions was set for each word in the display example (41), but for a language such as 1 Japanese where the bunsetsu is considered to be the u4 component of the syntax, the display example (41) is based on the arrangement of nine clauses (
The application condition may be set as the display unit of 41).

[Effects of the Invention] As explained above, the present invention includes a storage means for storing dictionary information, grammatical information, and dependency information used for translation, and a display for displaying and changing dependency information by analysis after translation. a changing means, a determining means for determining whether or not this modification of the dependency is grammatically correct, a retranslation means for retranslating according to the modification of the modification, and a pattern for changing the translated word by changing the modification and post-editing. Since it is equipped with a registration means for registering as a , and a processing means for subsequently performing similar processing on similar patterns in the original text, the scope of adaptation of learning information has been expanded to include the dependency structure. Front N
It has the effect of reducing the amount of work done by 4ffi, and that the more you learn, the better the quality of translation results will be obtained.

[Brief explanation of the drawing]

Figure 1 is a block diagram of the configuration of a machine translation system according to an embodiment of the present invention, Figure 2 is a flowchart of the dependency change process, Figure 3 is a flowchart of the learning pattern registration process, and Figure 4 is the learning of Figure 3. FIG. 5 is a block diagram of a conventional machine translation system configuration, and FIG. 6 is a flowchart of learning in FIG. 5. In the figure, (]) - input device, (2) pre-editing device, (3)...dictionary, (4)... grammar (5)...
Translation device, (6) ・Translation information storage device, (7) ・
・The modification is a change device, (8)...post-editing device, lol)
. . . Learning information storage device, (10) Learning pattern setting device 1ll) . . . Output device. In addition, the same reference numerals in each figure indicate the same or equivalent parts W) Figure 4 W-) Figure 5 Procedural Amendment Statement February 27, 1990

Claims (1)

[Claims] an input means for inputting an original text written in a first language; a pre-editing means for editing the original text input by the input means; A conversion means for converting an original text into a translated text written in a second language using a word dictionary, a grammar dictionary, etc., an output means for outputting the translated text converted into the second language by this conversion means, and this output means In a machine translation system having post-editing means for editing the translated text output by a display/change means for displaying and changing the modification; a judgment means for determining whether this modification of the modification is grammatically correct; a retranslation means for retranslating according to the modification of the modification; A machine translation system with a syntax learning function, comprising a registration means for registering a change in a translated word due to editing as a pattern, and a processing means for thereafter preferentially performing similar processing on similar patterns in the original text.