JPH02294779A - Machine translation system - Google Patents

Abstract

PURPOSE:To automatically acquire language knowledge from a translated sentence by finding out translating relation in each sentence between the 2nd language writing obtained by post-editing the 1st language writing and the 1st language writing. CONSTITUTION:The machine translation system is constituted of a Japanese dictionary storage device 04, an English dictionary storage device 05, a Japanese grammar frame storage device 06, an English grammar frame storage device 07, Japanese-English/English-Japanese conversion dictionary storage device 08, a Japanese coocurrence relation dictionary storage device 09, and English occurrence relation dictionary storage device 10, a Japanese writing file storage device 11, an English writing file storage device 12, and so on. The set of translated sentences of the 1st language sentences included in the 1st language writing can be found out from the 2nd language writing in the 1st and 2nd language writings having translating relation. Consequently, the translating relation in each sentence can be automatically determined from the input writing and the writing post-edited in each writing after translation, so that the knowledge of various levels can be automatically registered in a translation dictionary.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to a natural language processing system such as a machine translation system, and particularly to a machine translation system having an automatic learning function. [Prior Art] Several methods have been proposed for automatic knowledge learning methods used in machine translation systems.

For example, Japanese Patent Application No. 63-192751 proposes a method for acquiring and utilizing knowledge of co-occurrence relationships between words by extracting unambiguous intermediate expressions that are the result of input sentence analysis. Furthermore, Japanese Patent Application Laid-Open No. 62-219165 is a method of acquiring knowledge for selecting translations, such as co-occurrence relationships and idioms, by inputting fixed sentences. In addition, the special request for
-263934 is a method for extracting bilingual relationships between words, idioms, co-occurrence relationships between words, etc. from bilingual example sentences. Furthermore, “Feedback of post-editing information in Japanese-English translation systems” (Information Processing Society of Japan, Vol. 3)
7th National Convention, pp. The method described in 976-977) is a technique that learns conversion rules for Japanese-English intermediate representations and English sentence production rules from the results of post-editing of output English sentences from a Japanese-English translation system. [Problems to be Solved by the Invention] However, in the method of Japanese Patent Application No. 1982-19275, the co-occurrence relationship knowledge that can be acquired is only related to the source language. In addition, Japanese Patent Application Laid-open No. 62-219165 uses a method to learn co-occurrence relationship knowledge and idioms for selecting translations by inputting fixed sentences, but it is necessary to prepare a large set of fixed sentences. Therefore, it can be said that learning co-occurrence relationships is virtually impossible. Also, patent application No. 63-263934
The paper "Feedback of post-editing information in a Japanese-English translation system" describes a method for acquiring linguistic knowledge from first language sentences and second language sentences whose bilingual relationship has been determined. It cannot be used if the file is divided or combined. Post-editing is usually done on a sentence-by-sentence basis, so second language sentences are frequently divided and combined. The purpose of the present invention is to use the learning function of a machine translation system to determine the bilingual relationship of a post-edited second language text and a human-written first language text on a sentence-by-sentence basis, and to The goal is to enable automatic acquisition of language knowledge. [Means for solving the problem] The above purpose is to convert a set of parallel sentences of the first language sentence included in the first language sentence into a second language sentence from a first language sentence and a second language sentence that have a bilingual relationship. This is achieved by providing a step to find from among. This step includes counting the number M of first language sentences included in the first language sentences and the number N of second language sentences included in the second language sentences, and counting the first language sentences S(1) to Provisional parallel translations of S (M) T(1)~T (
M) and for m=1 to M, T (
Centering on S(m), from the bilingual relationship between the words included in the second language sentences before and after it and the words included in S(m),
), a step of finding a bilingual sentence set tr(m) of n =
1 to N, t r(1) to t r(M
) T(n) that is not included in any of t r (1)
~tr(M). [Operation] Find the number M of sentences included in the first language text and the number N of sentences included in the second language text that are in a bilingual relationship, respectively. Let the set of first language sentences be {S(1),...S (M)), and the set of second language sentences be ('1'(1)...T(N
)).

next. For m=1~M, F(m)=[m×N/M
Using the function F(m) defined by l, find the tentative translation sentence T(F(m)) of S(m) ([] means rounding up to the nearest whole number). Next, for m = 1 to M, the second language sentences containing the translations of the words constituting the sentence S(m) are defined as the set of second language sentences (T(w
ax(F(m)-r,1)),-,T(win(F(a)
+)+r, N))), and let tr(m) be the set of second language sentences containing translations of the words that make up the sentence S(m). Here, r is an integer satisfying 0≦r≦N. tr
If (m)=φ, then t r(m)= (T(F(m)
)).

Next, for n = 1 to N - tr (1) to tr
Second language sentence T (n) that is not included in any of (M)
If so, t r (m 1) = t r (m 1) LJ
(T(n)), t r(m2)=t r(m2)LJ(
T(n)), T(n) is tr(ml) and tr(
m2). here. tr(ml) is a set containing T(nl) for the maximum 01 in which there is a tr(m) containing T(nl) in the sentence before T(n). Also, tr(m2) is . In the sentence after T(n), T(n2)
For the smallest n2 such that there is a tr(yn.) containing .
It is a set containing T(n2). If there are multiple ml, m2, the above tr(ml
), tr(m2) is enlarged.

As described above, tr(M) is determined from the bilingual sentence set tr(1) of S(1) to S(M).

〔Example〕

The following describes a Japanese-English and English-Japanese bidirectional machine translation system that is an embodiment of the present invention. FIG. 2 is a hardware configuration diagram of the embodiment, showing a processing unit [201, input device 02, output device 03, Japanese dictionary storage device 04, English dictionary storage device 1!05, Japanese case frame storage device 06, English case frame storage device 07, El English/English-Japanese conversion dictionary storage device 08
, a Japanese co-occurrence relationship dictionary storage device 09, an English co-occurrence relationship dictionary storage device 10, a Japanese sentence file storage device 11, and an English sentence file storage device 12. The Japanese dictionary in the Japanese dictionary storage device consists of records as shown in Figure 3. The record is Japanese heading 041, part of speech 04
2, Semantic code 043, Japanese case frame code 044
Consists of. Records can be searched using Japanese heading 041 as a key. The English dictionary in the English dictionary storage device is composed of records as shown in Figure 4. The record is English heading 051
, part of speech 052, meaning code 053, and English case frame code 054. Records can be searched using English heading 051 as a key.

The Japanese case frame in the Japanese case frame storage device is composed of records as shown in FIG.

The record consists of Japanese case frame code name 061, deep case 062, surface case o63, and case element meaning code 064. Deep cases are represented by codes. Codes A, O, and I represent the nominative, object, and instrumental cases, respectively. Surface cases represent Japanese particles that correspond to deep cases.

The semantic code of a case element is a code that represents the semantic features that the case element should have. HUM, OBJ, and INST represent a person, an object, and a tool, respectively. Records can be searched using the Japanese case frame code name 061 as a key. The English case frame in the English case frame storage device consists of records as shown in Figure 6. The record is English case frame code name 071, deep case o72, surface case o7
3. Consists of case element semantic code 074. Deep cases are expressed using the same codes as for Japanese case frames. The surface case represents the English syntactic role or preposition that corresponds to the deep case, S represents the subject, and D○ represents the direct object. The semantic codes of case elements are expressed using the same codes as for Japanese case frames. Records can be searched using the English case frame code name 071 as a key. The Japanese-English/English-Japanese conversion dictionary in the Japanese-English/English-Japanese conversion dictionary storage device consists of records as shown in Figure 7. The record consists of a Japanese conceptual structure 081 and an English conceptual structure 082. Japanese conceptual structures and English conceptual structures in the same record have a bilingual relationship. Each conceptual structure may be a single word, or it may consist of a set of two words and a semantic relationship (expressed by a deep case code) between them. In Figure 7, for simplicity, [he] and rheJ, and "tool" and rtoo
Only bilingual relationships between words such as lJ are registered. Records can be searched using either the Japanese conceptual structure 081 or the English conceptual structure 082 as a key.

The Japanese co-occurrence relationship dictionary in the Japanese co-occurrence relationship dictionary storage device is composed of records as shown in Figure 8. The records are Japanese verbs 091, Japanese nouns 092, deep cases 09
Consists of 3. Records can be searched using the conjunction of Japanese verb 091 and Japanese noun 092 as a key. The English co-occurrence relationship dictionary in the English co-occurrence relationship dictionary storage device consists of records as shown in Figure 9. The record is
It consists of 101 English verbs, 102 English nouns, and 103 deep cases. Records can be searched using the conjunction of English verbs 101 and Japanese nouns 102 as keys. The Japanese co-occurrence relationship dictionary and the English co-occurrence relationship dictionary register verb-specific co-occurrence relationships that cannot be expressed in case frames as ternary relationships of verbs, nouns, and deep cases. Used in preference to frames. The Japanese text file in the Japanese text storage device is the 10th
It consists of records as shown in the figure. The record consists of Japanese sentence number 111, Japanese sentence 112, and bilingual English sentence number 113. The English text file in the English text storage device consists of records as shown in Figure 11. The record consists of English sentence number 121, English sentence 122, and parallel Japanese sentence number 123. The Japanese sentence storage device and the English sentence storage device each record the sentences that make up the nine sentences in the sentence in the cursed order. Each sentence is numbered and stored in one record unit. Next, let's talk about the processing executed by processing device 01. No. lrj
! Explain by I. First, process selection parameters are input from the input device 02. As the process selection parameter, enter "translation" if you want to use this system for Japanese-English machine translation, or "maintenance" if you want to use it for translation dictionary maintenance (0101) - Next, from the Japanese text file , reads the Japanese text. That is, all Japanese sentence records stored in the Japanese sentence file are read (0102). The set of Japanese sentences read in step 0102 is (SN(1),
..., SN(M)), the following step 0103
Step 0106 is repeated for m=l~M.
Parse the Japanese sentence SN(m) by referring to the Japanese dictionary, Japanese case frame, and Japanese co-occurrence relationship dictionary. The analysis result is that nodes represent content words, and arcs that connect pairs of semantically related words represent their semantic relationships (deep case codes).
It is expressed as a tree-structured graph representing . If multiple analyzes are possible, find all solutions. The set of analysis results of the Japanese sentence SN(m) is expressed as (N(m, 1), ++, N(m, I)) (0103). The analysis results are stored in the working memory within the processing device (0104). The input processing selection parameter is determined in step 0101, and if the parameter is "translation", the process proceeds to step 0106, and the parameter is "translation".
If it is 'Maintenance', the process returns to the beginning of the iterative process of steps 0103 to 0106 (0105). Japanese-English/English-Japanese conversion dictionary,
Refer to the English co-occurrence relationship dictionary and the English case frame dictionary to generate English sentences for each analysis result from the Japanese analysis results N (m, 1)...N (m, I) (010
6). The process selection parameter input in step 0101 is determined, and if the parameter is "translation", step 010
Proceed to step 8, and if it is "maintenance", proceed to step 0111 (0107).The English translated text output in step 0106 is output from the output device (0108).Post-editing of the English translated text is accepted (0109). Output the post-edited English text to the English text file (0110). Input the English text from the English text file. In the case of translation processing, this English text is the translated text that was post-edited in step 0109. , the bilingual text of the Japanese text input in step 0102.In the case of dictionary maintenance processing, this is also the bilingual text of the Japanese text input in step 0102, which is stored in advance in the English text file (0111).
The set of English sentences read in step 0111 is (SE(1
),...,SE (N)), the following step 0
Steps 0 and 3 from step 112 are repeated for n = 1 to N. Parse the English sentence SE(n) by referring to the English dictionary, English case frame, and English co-occurrence relationship dictionary. The analysis results are expressed as a tree structure graph in which nodes represent content words and arcs connecting pairs of semantically related words represent their semantic relationships (aI stratification codes). If multiple analyzes are possible, find all solutions. The set of analysis results of English sentence SE(n) is expressed as (E(n, 1),...E (m, J
)) (0 1 1 2). The analysis results are stored in the working memory within the processing device (0113). Next, the bilingual sentence determination routine described later is used to find the bilingual relationship between the Japanese sentence and the English sentence on a sentence-by-sentence basis. As described later, the bilingual relationship is a set of bilingual English sentences for Japanese sentences SN(1) to SN(M), respectively. Expressed as tr(1) to tr(M)<(0114). Based on the bilingual sentence relationship obtained in step 0114, the correspondence relationship between the Japanese sentence analysis result stored in step 0104 and the English sentence analysis result stored in step 0113 is determined, and the above-mentioned patent application No. 1988-263 is obtained.
Using the method described in 934, knowledge such as bilingual relationships between words, co-occurrence relationships between words, and idioms is extracted. That is, for m=1 to M, the analysis results of SN(m) obtained in step 0103 are N (m, 1) to N
(m, I) and the analysis result E (n, 1) to E (no
J) and based on the pattern matching process,
Analysis results N(m,i) and E with the best pattern matching
Knowledge is extracted by comparing (n, j). t
When r(m) includes multiple English sentences and when one English sentence is included in multiple tr(m), in other words, the bilingual sentence relationship is 1.
If the ratio is not 1, it is necessary to slightly improve the method described in Japanese Patent Application No. 63-263934, but as will be explained later in a specific example, it is possible to easily Can be improved (0115). Based on the knowledge extracted in step o115, the Japanese-English/English-Japanese conversion dictionary memory, the Japanese co-occurrence relationship dictionary, and the English co-occurrence relationship dictionary are updated (0116). The bilingual sentence relationship obtained in step 0116 is output to the Japanese text file and the English text file (0 1 1 7). The above is the process flow of the present invention. Next, the processing contents of the bilingual sentence determination routine in step 0114 will be described in detail with reference to FIG. first. m-1~M
, find the bilingual sentence candidate SE(F(m)) of the Japanese sentence SN(m) using the function F(m) defined by F(m) = [m×N/M] ([kowa/ I mean rounding up to the next few points). As shown in Figure 13, this is a combination of Japanese numbers (1,...,M) and a set of English numbers (1,...,M).
1+mtN) is represented by an M×N matrix, the bilingual sentence relationship is based on the fact that there is a high possibility that the set will be a subset distributed approximately on the diagonal of the matrix (01
141). Next, for m = 1 to M, the English sentences containing the translations of the words constituting the sentence SN(m) are divided into a set of English sentences, { S E (+
iax(F(m)-r,1)), S E(wax(
F(m)-r,1)÷1),-, S E (s
in (F(+*)+r,N))), sentence SN
tr(
m). This process means finding a set of bilingual sentences for sentence SN(m) from ±r sentences before and after the bilingual sentence candidate obtained in step 01141. Let r be an integer satisfying 0≦r≦N. Weight 11N and win are used to deal with cases where m is close to 1 or M. Word segmentation is performed in steps 0103 and 0112, and the results are used. The bilingual relationship between words is determined by searching Japanese-English and English-Japanese conversion dictionaries. Here, if tr(m)=φ, then tr(m)=(SE(F(m)))(0
1142), Next, for n=1 to N, t r (1) to t r
If there is an English sentence SE(n) that is not included in any of (m), then
t r (ml) = t r (ml) U (SE (n)), t
r(m2)=t r(m2)U(SE(n)). As, SE(n) is included in tr(ml) and tr(m2). Here, tr(ml) is the maximum n1 in English sentences before sE(n) where tr(m) including SE(nl) exists.
is a set including SE(nl). Also, tr
(m2) is an English sentence after SE(n), and SE(n2)
For the smallest n2 such that there exists a tr(m) containing SE
This is a set containing (n2). If there are multiple m1 and m2, the above t r(ml
), t r (m2) is expanded (01143
). As described above, tr(1) to tr(M) obtained
Let be the translation sets of SN(1) to SN(M), respectively. Next, the Japanese dictionary in Figure 3, the English dictionary in Figure 4, the Japanese case frame in Figure 5, the English case frame in Figure 6, the Japanese-English/English-Japanese conversion dictionary in Figure 7, and the English-Japanese conversion dictionary in Figure 8. Japanese co-occurrence relationship dictionary,
The process of automatic learning machine translation when using the English co-occurrence relationship dictionary shown in Figure 9, the Japanese text file shown in Figure 10, and the English text file shown in Figure 11 will be explained in detail using examples.
First, we will explain what happens when "maintenance" is input as the process selection parameter. In the Japanese text storage device, it says, ``After I made a printer with that tool, he made a keyboard with that tool.''
Only one sentence is registered, so this one sentence is SN(1
) using a Japanese dictionary, Japanese case frame, and Japanese co-occurrence relationship dictionary to obtain the analysis results shown in Figure 14. The deep case code M represents a modification relationship, and L represents a continuous relationship between predicates. In the English sentence storage device, there is a bilingual sentence of {SN(1))
with the tool. J, rThe
n,ha made the keyboard wi
th the tool. J) is registered. Therefore, these two sentences are SE(1). Enter as SE(2),
Analyze the English text and obtain the analysis results shown in Figures 15 and 16. The verb r @ake J has two case frames, E1 and E2, so S E (1) and S E (2
) are shown in Figures 15(a), (b), and 15(b), respectively.
Two methods are obtained in each of Figures 16 (a) and (b). Next, proceed to the bilingual sentence determination routine. Now, the number of Japanese sentences M = 1
, the number of English sentences N=2, so F(1)=[1x2/1
1 = 2. Therefore, the parallel sentence candidate for SN(1) is S
E (2). next. From the English sentences before and after SE (2). Find the set of translated sentences for SN(1) based on the translated relationships between words. If r=1, the search range is (S E
(1), S E (2)). Bilingual relations registered in the Japanese-English and English-Japanese conversion dictionaries (I, 工), (He, he),
(make, make) + (tool), (
t r(1)=
becomes {SE(1), SE(2)). Since all the English sentences are included in tr(1), the 5-parallel sentence determination routine ends. Next, we move on to knowledge extraction processing. At this time, as mentioned in the explanation of the process of step 0115, since the bilingual sentence relationship is 1:2, the SN(1) shown in FIG.
It is necessary to pattern match the graph representing the analysis results of SE(1) and SE(2) shown in Figure 15 and Figure 16, respectively.・Bilingual translations registered in the English-Japanese conversion dictionary (and
then), the graphs in Figures 15 and 16 can be combined into the graph in Figure 17,
Pattern matching is performed using the graph in Figure 14 and Figure 17 (a).
You can do this for the graphs in (b), (c), and (d). 2. The pattern matching method described in the above-mentioned Japanese Patent Application No. 63-263934. In other words, the graph in FIG. 17(a) that best matches the graph in FIG. 14,
By comparing the conceptual relationships in the graph in Figure 14, we can determine the bilingual relationship between words (keyboard, keyboard),
Japanese co-occurrence relationship (make, keyboard, 0), (make, him, A), (make, tool, I), (make, printer, 0)
and English co-occurrence relationship (make, keyboardL O
) + (maketI, A), (mak
8, tool, I), (make, pr
inter, 0) is extracted and registered in the dictionary (Wi
For simplicity, only the co-occurrence relationships between predicates and case elements are extracted). Finally, the sentence-by-sentence bilingual relationship obtained in the bilingual sentence determination routine is registered in the Japanese sentence storage device and the English sentence storage device, respectively, as shown in FIGS. 18 and 19, and the process ends. In the above example, we have taken up the case where "r maintenance" is input as the process selection parameter, but even if "translation" is selected, the bilingual text will be stored in the English text memory by going through the steps of English sentence generation and Volume IIi. Since the information is stored in the device, the processing after bilingual text determination can be performed in exactly the same way as in the case of "maintenance".An embodiment of the present invention has been described above.According to the above embodiment, the machine translation system Automatically determines the bilingual relationship on a sentence-by-sentence basis from the translated English text that has undergone manual post-editing and the Japanese text input into the machine translation system, and calculates the bilingual relationship as Japanese and English words, and the Japanese text input into the machine translation system. Origin relations can be automatically and appropriately extracted and registered in a dictionary.For this reason, a machine translation system can be provided with a dictionary self-propagation function. [Extended example 1] Also, an extended example of the above embodiment. In the process of determining parallel sentences, the parallel sentence candidates SE(F(m)
), select SN(
m) In the step of detecting parallel sentences, the accuracy and efficiency of determining parallel sentences is improved by adding a step of determining the value of r based on the magnitude relationship between the number M of Japanese sentences and the number N of English sentences. There are ways to do this. In other words. If it is MAN, the number of English sentences corresponding to one Japanese sentence will be small, so let r be small<L, . If M<N, increase r. For example, provide a step to calculate r = [N/M] ([] means rounding up to the nearest whole number). [Extension Example 2] Furthermore, in the above embodiment and the method of changing r in [Extension Example 1], in the case of M)N. There are many cases where the enlarging operation of tr(m) (step 01143) is used. It is thought that the accuracy of determining bilingual sentences will be slightly lower. Therefore, before the bilingual sentence determination process, a step is added in which the magnitude relationship between M and N is determined, and if M>N, the direction of the translation sentence determination is changed from Japanese to English to English. In other words. For MAN, English S E
The above problem can be solved by providing a step of changing the bilingual sentence determination in the direction of finding the set of Japanese bilingual sentences from (1) to SE(N). It is clear from the symmetry of the system that it is possible to change the direction of bilingual sentence determination. [Extended Example 3] Furthermore, in the above embodiment, a step (step 0114) of determining the bilingual sentence of the Japanese sentence SN(m) from among the bilingual sentence candidates SE (F(m)) is performed based on the bilingual relationship between words.
In 2), the bilingual relationships are found for all content words, but by limiting the parts of speech of the words (for example, nouns) and detecting the bilingual relationships only for the limited words,
I can't think of a way to speed up the processing. [Extended example 4] Also, in the above example, paragraph management information is provided in the records of the Japanese text file and the English text file as shown in Figures 20 and 21, and the parallel sentence determination process is performed not on a sentence basis but on a paragraph basis. One possible method is to improve the decision accuracy by doing it in units. [Expansion Example 5] Furthermore, in the English sentence generation process and knowledge acquisition process of the above embodiment, a step is provided to select an alphabetic character string enclosed in parentheses after a word in the Japanese sentence as a translation of the word, and the translation is If it is not registered in the Japanese-English/English-Japanese conversion dictionary, by providing a step to register the bilingual relationship when changing the dictionary, the ability to select translated words and acquire knowledge can be improved. For example, if the Japanese sentence "I created a program (progra■)." is input, "
Select rprogramJ as the translation word for program J, and if the bilingual relationship of (program, prograya) is not registered in the Japanese-English/English-Japanese conversion dictionary, register the bilingual relationship. [Extended Example 6] Furthermore, in the above embodiment, in step 01142 of the bilingual sentence determination routine, the Japanese sentence S is determined based on the bilingual relationship of the words.
It is also possible to perform the process of finding N(m) bilingual sentence sets as follows. In other words, by setting a threshold value θ≧O, the number of content words Y(n) included in the English sentence SE(n) is determined by the SE
The ratio of the number T (m, n) of things included in (n), R (m, n) = T (m, n) / Y (n) is. R(m, n
) > θ, SE(n) is included in tr(m). The above example corresponds to the case where θ=0 in this extended example. Also, in this extended example,
Since accuracy is improved by providing the threshold value θ, step 01141 in the above embodiment can be omitted.

That is, the bilingual sentence search range for SN(m) in step 01142 is set to the entire English sentence, that is, (SE(1
), ..., SE(N)). [Extension Example 7] Furthermore, in the above embodiment, in the step of extracting knowledge from bilingual sentences, only bilingual sentences with a one-to-one bilingual sentence relationship are obtained, and knowledge is extracted only from the one-to-one bilingual sentences. It is also possible to improve the accuracy of knowledge extraction. [Modification 1] Furthermore, the bilingual sentence determination method of the above embodiment is a method based on the bilingual relationship between words. However, since it is thought that a sentence set and a sentence set in a bilingual relationship contain almost the same number of content words, bilingual sentence determination is performed by selecting the sentence set that has the same number of content words at the beginning of the bilingual sentence, as described below. It is also possible to find it from the sentence . First, for m = 1 to M, the Japanese sentence S
Let the number of content words included in N(m) be respectively . X (m)
It is defined as Also, for n = 1 to N, English SE (n
) is defined as Y(n). Therefore, the bilingual sentence determination process is performed as explained below in accordance with FIG. 22. First, let SRT(1)=1. SRT(m) is SN(
This is the number of the first English sentence in the bilingual sentence set tr(m) of rn) (0114A). Set the variable ENDN=O. ENDN
is a flag indicating the end of the loop of steps 0114C to 0114J below (01 14B). m=1～
For M (until m=M or ENDN=1),
Repeat steps 0 1 14C to 0114L below. Let variable SUM=O. SUM is a variable that counts the number of words in an English sentence (0114G). n=SRT(
m) (0 1 1 4 D). X(m)≧SU
N, and while n≦N, the next two steps 011
Repeat 4E, 0114F.

Add Y(n) to SUM (0114E). Add 1 to n to k, (0 114 F) . 0 1 14E,01
14F (7) After passing through step, variable END (m)
=n-1. END(m) is the last English number included in tr(rn) (0 1 14G). t
r(m.)=(S E (S R T(m)), ・
=, S E (E ND(m))) (0114
H). Determine whether n < N. If n<N,
Proceed to step 0114J, and if n < N, proceed to step 0114K t/ (0114I). S.R.
Set T (m + 1) = n (0114J)
．． Set ENDN=1 (0 1 1 4 K). tr(
m+1)=,−−−,= t r (M)= t r
(m) (0 1 1 4 L). Above, 011
When the loop from 4C to 0114J ends, the process ends. [Modification 2] The bilingual sentence determination method in Modification 1 above is a method based on the number of content words included in a sentence. It is also possible to calculate from the beginning of the bilingual text. The execution method is based on the definitions of X(m) and Y(n) in Modification 1 above. By changing the number of characters included in SN (m) and the number of characters included in SE (n), the method of Modification 1 above can be used as is. Although the accuracy of this modification is considered to be slightly lower than that of modification 1 above, it is effective in determining parallel sentences between relatively similar languages such as English and French. [Effects of the Invention] According to the present invention, in a bidirectional machine translation system that acquires linguistic knowledge from bilingual sentences, it is possible to determine the bilingual relationship on a sentence-by-sentence basis from an input sentence and a post-edited sentence on a sentence-by-sentence basis after translation. Since it can be determined automatically, by pattern matching the analysis results for each sentence, the bilingual relationships between words, bilingual relationships as idioms, co-occurrence relationships, and semantic properties of words can be extracted and registered in the dictionary. This makes it possible to automatically and appropriately register knowledge at various levels in a translation dictionary, and automatically improve the accuracy of the dictionary. It is particularly effective for batch translation of large amounts of text such as manuals.

[Brief explanation of drawings]

Figure 1 is a PAD showing an outline of the processing of a Japanese-English/English-Japanese machine translation system as an example of the present invention, and Figure 2 is the hardware of a Japanese-English/English-Japanese machine translation system as an example of the present invention. Configuration diagram, Figure 3 shows an example of record contents of a Japanese dictionary, Figure 4 shows an example of English dictionary record contents, and Figure 5 shows an example of record contents of a Japanese case frame. , Figure 6 is a diagram showing an example of the record contents of an English case frame, @ Figure 7 is a diagram showing an example of the record contents of a Japanese-English/English-Japanese conversion dictionary, and Figure 8 is a diagram showing an example of record contents of an English case frame.
The figure shows an example of the record contents of a Japanese co-occurrence relationship dictionary.
Figure 9 is a diagram showing an example of the record contents of the English co-occurrence relationship dictionary, Figures 10, 18, and 20 are diagrams showing examples of the record contents of the Japanese text file, Figures 11, and 19. , Figure 21 is a diagram showing an example of the record contents of a file of English sentences, Figure 12 is a PAD showing details of the bilingual sentence determination process based on the bilingual relationship of words, and Figure 13 is a set of Japanese sentence numbers and English sentences. A conceptual diagram showing a Cartesian product set of a set of numbers. Figures 14 to 17 are diagrams showing examples of tree structure graphs representing the analysis results of Japanese and English sentences. Figure 22 is a diagram showing the number of content words included in a sentence. This is a PAD showing details of the bilingual sentence determination process. 01... Input device, 02... Processing device, 03...
Output device, 04...Japanese dictionary storage device. 05...
English dictionary storage device, 06...Japanese case frame storage device, 07...English case frame storage device, 08...Japanese-English/English-Japanese conversion dictionary storage device, 09...Japanese co-occurrence relationship dictionary Storage device, 10...English co-occurrence relationship dictionary storage device,
11...Japanese text file storage device, 12...English ahi fi Azusa Takuwari Yuka dai Kasiriku ■ m edict circulation bar ρ81 〆ρ82 th lθ ro Z mouth order // closed circulation mouth 2 1 heyt. hexJetkeyb
oa>tj WltHhetpa/, Ittakuroton l7 Mouth (a) Yukuchi Z I4 Prisoner) No. 15 (b) (of (b invitation/g Kuchitakutq ■ 2θ Kuchirokuzu)

Claims (1)

[Scope of Claims] 1. A machine translation system characterized by the step of determining a sentence-by-sentence bilingual relationship between a second language sentence obtained by post-editing the translation result of the first language sentence and the first language sentence. 2. In the machine translation system according to claim 1, the step of determining the sentence-by-sentence bilingual relationship between the first language sentence and the second language sentence is performed for each first language sentence SN in the first language sentence. In the second language sentence SE in a sentence, S among the translations of content words included in SN relative to the number of content words included in SE
A machine translation characterized by comprising a substep of calculating the ratio of the number of items included in E, and a substep of including SE in a bilingual sentence set of SN if the ratio exceeds a certain threshold. system. 3. In the machine translation system according to claim 1, the step of determining the sentence-by-sentence bilingual relationship between the first language sentence and the second language sentence includes determining the number M of first language sentences and the number N of second language sentences;
For the mth first language sentence of the first language sentences, m×N/M
a substep of determining a first language sentence having an integer number as a bilingual sentence candidate for the m-th first language sentence;
From several sentences before and after the parallel translation candidate in the second language sentence,
a substep of determining a set of sentences containing translations of words included in the m-th first language sentence and making this set a bilingual sentence set of the m-th first language sentence; It is determined whether there is a second language sentence in the second language sentence that is not included in the bilingual sentence set of any first language sentence in the sentence, and It consists of a substep of including a second language sentence in a first language sentence that is not included in a set of parallel sentences of a monolingual sentence into a set of parallel sentences of any first language sentence in the first language sentence. A machine translation system characterized by: 4. In the machine translation system according to claim 3, in the substep of obtaining a parallel translation set of the first language sentence from several sentences before and after a parallel translation candidate of the first language sentence, the range for obtaining the parallel translation set is A machine translation system characterized in that the determination is made based on the magnitude relationship between the number M of monolingual sentences and the number N of second language sentences. 5. In the machine translation system according to claim 3, the number M of first language sentences and the number N of second language sentences are determined before the step of determining the sentence-by-sentence bilingual relationship between the first language sentence and the second language sentence. If M>N, the direction of determining the bilingual relationship for each sentence is determined by changing the bilingual sentence set of the second language sentences in the second language sentence to the first language sentence in the first language sentence. A machine translation system characterized by providing a step of converting in a direction determined from among. 6. In the machine translation system according to claim 3, in the step of obtaining a bilingual sentence set of the first language sentence from among the second language sentences including translations of words included in the first language sentence, the part of speech in the critical state is determined. A machine translation system characterized by: 7. The machine translation system according to claim 1, wherein the bilingual translation relationship in sentence units is determined from the bilingual translation relationship in paragraph units. 8. A machine translation system, characterized in that the machine translation system includes a step of determining a sentence-by-sentence bilingual relationship between a second language sentence and a first language sentence, which are obtained by manually post-editing the translation result of the first language sentence. dictionary maintenance method. 9. A dictionary maintenance method for a machine translation system, comprising a step of determining a sentence-by-sentence bilingual relationship between a first language sentence and a second language sentence that are in a bilingual relationship. 10. In a machine translation system, a bilingual sentence file characterized by providing a step of obtaining a sentence-by-sentence bilingual relationship between a second language sentence and a first language sentence, which are obtained by manually post-editing the translation result of the first language sentence. Creation method. 11. A method for creating a bilingual text file, which is characterized by providing a step of obtaining a bilingual relationship in sentence units between a first language sentence and a second language sentence that are in a bilingual relationship. 12. In the machine translation system according to claim 1, when a second language string enclosed in parentheses is present in the first language sentence after a word in the first language sentence when selecting a translation word when generating a second language sentence. A machine translation system comprising a step of selecting the character string as a translated word. 13. The machine translation system according to claim 12, wherein a second language character string enclosed in parentheses exists in the first language sentence after a word in the first language sentence, and the character string is a translation word of the word. , determine whether it is registered in the dictionary,
A machine translation system comprising the step of registering the character string as a translation of the word if the character string is not registered. 14. In the machine translation system according to claim 1, the step of determining the bilingual relationship between sentence units of the first language sentence and the second language sentence includes a sentence set in the first language sentence and a sentence set in the second language sentence. , a step of finding pairs of sentence sets in which the number of content words included in the sentence sets matches, starting from the beginning of a first language sentence and a second language sentence. 15. In the machine translation system according to claim 1, the step of determining the bilingual relationship between sentence units of the first language text and the second language text includes a sentence set in the first language text and a sentence set in the second language text. , a machine translation system characterized in that the step is to find a pair of sentence sets that have the same number of characters in the sentence sets, starting from the beginning of a first language sentence and a second language sentence. 16. The machine translation system according to claim 1, wherein knowledge is acquired only for sentences with a one-to-one bilingual relationship.