JPS61128372A - English-japanese translation apparatus - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to machine translation of natural languages, particularly machine translation from English to Japanese.

[Prior art] In recent years, using dictionaries and grammar rules registered in storage devices,
An English-Japanese machine translation method that analyzes and converts an English input sentence to generate a Japanese translated sentence is attracting attention.

Generally, most practical English-Japanese machine translation methods use the transfer method. The Transf7 method first analyzes an input sentence according to the analysis dictionary and analysis grammar rules registered in the storage device, obtains an intermediate representation of the input sentence, and then processes it according to the conversion dictionary and conversion rules registered in the storage device. , convert this intermediate representation into an intermediate representation of the output sentence (translated sentence) (
Finally, this intermediate representation is used to generate a target translated sentence according to a generative dictionary and generative grammar rules.

[Problems to be Solved by the Invention] In the transfer method of the conventional English-Japanese machine translation system, an input English sentence is parsed to some extent to obtain an intermediate representation representing the syntax, which is then converted.

There is a method that outputs a Japanese sentence, and a method that deeply parses the input English sentence and also performs semantic analysis to obtain an intermediate expression that expresses the meaning, converts it, and outputs a single Japanese sentence. According to the former method, the structure of the English-Japanese machine translation device is simpler and smaller, and the translation speed is faster, but it has the disadvantage that the quality of the translated text is poor. On the other hand, according to the latter method.

Although the quality of the translated text improves, the structure of the English-Japanese machine translation device becomes more complex and larger, and the amount of language data that must be registered and stored in the storage device becomes enormous, resulting in slower translation speed and even slower translation. The disadvantage is that it is not easy to maintain and manage the device. In addition, because the analysis is done deeply into the intermediate expressions of the English sentence that express the meaning, it is difficult to reflect the subtle nuances and idiomatic expressions of the surface English sentence into the Japanese sentence.

Therefore, one object of the present invention is to parse an input English sentence. The advantages of the Katsu-English-Japanese machine translation method, which uses a parse tree as an English sentence intermediate expression and transfers it to the corresponding Japanese intermediate expression, are that it has a relatively simple system configuration and processing procedure;
Based on the characteristics that this method can be realized using a relatively small amount of linguistic data and requires almost no processing related to writing style, we aim to significantly improve the poor quality of translated sentences, which is a drawback of this method. Before English-Japanese conversion, we insert a processing step to transform the English sentence parse tree into another English sentence parse tree as necessary, and convert unique English expressions, including idiomatic phrases, from English sentences directly into Japanese. This improves the accuracy of subsequent English-Japanese conversion, makes it easier to maintain and manage the English-Japanese transfer part, and speeds up the English-Japanese transfer process. The object of the present invention is to provide an English-Japanese machine translation method that outputs high-quality translated sentences at high speed as a whole and realizes a system that is easy to maintain and manage.

[Means for Solving the Problems] In the English-Japanese transfer method of the present invention, an English input sentence is first parsed according to an analysis dictionary and syntactic analysis rules registered in a storage device, and an intermediate representation of the input English sentence is used as an intermediate representation. Obtaining a parse tree representing the syntax. Starting from the top node of this English sentence parse tree and moving toward the lower nodes, along the structure of this parse tree, the nodes of the parse tree are grouped by type. By applying a group of transformation rules at each node, structural transformation is performed from an English sentence parsing tree to another English sentence parsing tree, and at the same time, at each node, transformation procedure rules grouped by each node type are applied. By applying the group, we can perform the English-Japanese conversion process of the Lebel range between the node and its child node group, that is, the process of converting the English sentence parse tree to the corresponding Japanese tree structure. The above transformation rules, which are determined and stored in the storage device, are rules for transforming idiomatic expressions and special constructions of English sentences into basic sentence patterns. This transformation process is performed to facilitate the later English-Japanese conversion process.

In addition, the actual conversion process is not performed at the time the conversion process procedure is determined, but the transformation process from upper to lower direction, the determination of the conversion procedure, and the storage operation are performed at the lowest node of the English parse tree. Once reached, the actual conversion process is performed according to the already memorized English-Japanese conversion process. This conversion process is specialized depending on the type of each node and the status of its child nodes. A Japanese intermediate representation corresponding to the input English sentence parse tree, that is, a Japanese tree structure, is created one after another from lower nodes to upper nodes, and finally the input Perform English-Japanese conversion processing on the top node of the English sentence parse tree to obtain the desired Japanese intermediate expression.

Next, from this intermediate representation of the Japanese sentence, a Japanese sentence generating device outputs a surface Japanese sentence using a Japanese sentence generating dictionary.

[Embodiment] An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 shows an overview of the process of automatically translating an English sentence 100 into a Japanese sentence 120 using the English-Japanese machine translation method of the present invention.

When an English sentence is input using the input bag W101, first, the English sentence parsing tree 8106 is used to recognize words, idioms, and technical terms, that is, morphological processing is performed. Then, according to the analysis rules of the English sentence analysis rule dictionary 107, the English sentence analysis device 1
02, the input English sentence is analyzed and an intermediate representation of the English sentence, that is, an English sentence parse tree is output and stored in the English sentence intermediate representation storage device 112. Next, the English-Japanese intermediate representation transformation conversion device 103 converts this English sentence parse tree into a Japanese It is transferred to an intermediate representation of the sentence and stored in the Japanese sentence intermediate representation storage device 113. This intermediate representation is also a tree structure expressing the syntax of a Japanese sentence. Subsequently, a Japanese sentence is generated from this intermediate representation of the Japanese sentence by the Japanese sentence generating device 104 using the Japanese sentence generating dictionary 111.

Finally, the output device 105 outputs the target Japanese sentence 12.
0 is output.

In the following, the English-Japanese machine translation method according to the present invention will be explained in more detail using one embodiment, focusing on the processing by the English-Japanese intermediate representation transformation conversion device, which is the central part of the present invention, that is, the English-Japanese transfer processing. do.

The program provides these
a functi6ns," (1) - As an example, we will explain the case of translating the simple English sentence shown in sentence (1) above into a Japanese sentence. First, the input English sentence shown in sentence (1) is translated into English as shown in Figure 1. When the sentence analysis device 102 performs the analysis, the English sentence parse tree shown in FIG. 2 is obtained and stored in the English sentence intermediate expression storage bag M112 of FIG. 1. In FIG.

rDEcLJ is a sentence, rNPJ is a noun phrase, rVPJ is a verb phrase, rDETJ is a determiner, and "N" is a noun.

rVERBJti verb, "ADJ" represents an adjective,
Mata lowest rTHEJ, rPROGRAMJ, rPR
OVIDEsJ, rTHESEJ and rFUNcTI
ONsJ represents the English word itself. Further, the number of each node is added to identify each node. Next, the English sentence intermediate representation storage device 112 in FIG.
The English sentence parsing tree shown in FIG. 2 stored in is the input to the English-Japanese intermediate representation transformation conversion device 103 shown in FIG.

FIG. 3-1 shows an outline of the English-Japanese intermediate representation transformation conversion device.

The English sentence parse tree input by the English sentence parse tree input section T1 is stored in the transformation/conversion tree structure storage section T2, and the contents of T2 at this time are as shown in FIG. 3-2. This T
2 is a first-in, last-out stack in which subtree structures that have not yet been transformed are stored. The next step T3 is a tree structure selection section which takes out the data in the storage section T2 in order from the most recently entered data.

In this example, the tree structure shown in FIG. 2 is used. The next step T4 is a transformation rule selection section, which selects only one applicable rule from the transformation rule group corresponding to the type of vertex node of the tree structure selected at T3. used in this step. The transformation rule group corresponding to each node is T
51 DECL transformation rule storage unit, T52 VP transformation rule storage unit, T53 NP transformation rule storage unit, and T54.
The rules are stored in advance in the PP transformation rule storage unit or the like. In the example shown in Figure 2, the vertex node is rDECLJ
Therefore, a transformation rule applicable to the tree structure shown in FIG. 2 is searched from among the DECL transformation rule group stored in the DECL transformation rule storage section of T51. However, in this example, there are no transformation rules to apply, and the next step T
In the transformation executing section 6, the tree structure shown in FIG. 2 is sent as it is to the transformation result storage section of the next step T7 without being transformed.

Here, a specific example of transformation processing for the "DECL" structure will be explained.

It is assumed that the re
cord is 5tored onths disk
, (2) For example, in the case of the input English sentence shown in sentence (2) above, the English sentence parse tree shown in Figure 11(a) will be obtained, and in this,
The rule shown in the following equation (3) is applied from among the transformation rule group for the rDEcLJ structure.

<(NP(PRON IT))(VP(BIE-V l
5) (VERB 11) (NP(COMP THAT
) Umbrella 2))>→<(NP(COMP THAT)-2
)(VP(B[1i-V l5)(VERB fist 1
))> (3) In the transformation rule of formula (3), the left side of the → mark describes the condition for whether or not to apply this rule, and this condition is for the structure of all children of the rDEcLJ node, Satisfied if the patterns match. On the other hand, on the right side of the → mark, rDE in Fig. 11(a)
It describes a subtree structure sequence that is newly generated when the subbook structure below the cLJ node satisfies the matching condition of the pattern on the left side of the → mark. For illustration purposes, the formula (
If the contents of 3) are expressed graphically, the result will be as shown in FIG. 11(b).The transformation processing here will be explained based on FIG. 11(b). First, the subtree structure 2005 in FIG. 11(b) is different from the subtree structure 2005 in FIG. 11(a).
1 whose pattern matches 1 Next, the structure 2006 of FIG. 11(b) is matched with the structure 2002 of FIG. 11(a), but the symbols skewer 1 and umbrella 2 in the structure 2006 are variables , and when matching a pattern, it matches any subtree structure existing at that location, and at the same time, the matched structure is assigned to these variables. Therefore, the pattern of the structure 2006 also matches the structure 2002 in FIG. 11(a), the structure 2003 in FIG. 11(a) is substituted for *1, and the structure 2004 for the fist 2 , i.e. the first
1(c) The structure shown in Figure 1(c) is substituted. Here, the 11th (
b) All subtree structures on the left side of the figure are processed, and at the same time,
Since the processing has also been completed for all child nodes of the rD E CLJ node in FIG. 11(a), the conditions of this transformation rule are satisfied. Therefore, the first
1(b) Generate structures below the rDEcLJ structure according to the right side of the → mark in the diagram. First, according to IIZO 2007, the 11th
(d) A diagram structure is generated. However, here, the structure that has already been assigned to umbrella 2 is used.
The structure 2008 in figure t(b) generates the structure in figure 11(e). In this case as well, the structure assigned to umbrella 1 is used. Finally, these structures were converted into rDEcL
Connect as a child node of the J node to obtain the result shown in FIG. 11(f). This transformation is for translating the English IT-THAT syntax into proper Japanese.

Next, rules for transforming the English THERE-BE syntax and examples of their application will be shown.

There are three types
of commands, (4) The English sentence parse tree of the input English sentence of sentence (4) is as shown in FIG. 12(a), and the applied transformation rule is the following equation (5).

ku(＾VP(ADV THERE))(VP(BH-
V BE) (NP skewer i))>-e<(NP $1
) (VPmRB EXIST))> (5) 12th
Figure (b) shows the deformation results. In this variant, rTH
EREJ, rAREJ were removed and the 12th (b
) The verb rEXISTJ as shown in 2101 in the figure has been introduced.

As a final example of deformation processing for the rDEcLJ structure, 1
Next, an example for a transformation of the English Too-To syntax is shown.

The file is too large for
you to use, (6) The parse tree for the input English sentence (6) is shown in Figure 13(a), and the transformation rules to be applied are:
It is the following.

<(NP ml)(VP(VERB Fist 2)(AJP(
ADV Too) (ADJ-3) (PP (PROP
FOR) (NP Umbrella 4) (INFCL (INFTO)
To) Umbrella 5))))>-')<(SUBCL(CONJ
BFiCAIJSE) CNP Fist 1) (VP(V
l, RB,, -2) (AJP (ADV VERY) (
ADJ $3))))(NP Umbrella 4)(VP(AU)
XV CAN) (ADV N0T) $5)>
()) Skewers 3 and 4 described in transformation rule (7) whose transformation results are shown in Figure 13(b)
．． *5 is.

This is a variable similar to umbrella 1 and *2 in the previous explanation. In this case, the Too-To syntax is transformed into a syntax using "rBE'cAUsE".

By the way, the data sent to the transformation result storage section T7 in FIG. 3-1 has the tree structure shown in FIG. 2, and is stored there. The next step T8 is a conversion procedure rule selection section, which is a conversion procedure for performing English-Japanese conversion of the rubel range between the vertex node of the tree structure and its child node group stored in the transformation result storage section T7. Rules are retrieved from a conversion procedure rule storage section prepared in advance for each type of node. These storage units include the DECL conversion procedure rule storage unit of T91, the VP conversion procedure rule storage unit of T92, the NP conversion unit of T93, the PP storage unit of T94, and the AJP of T95.
(for adjective phrases), etc. In this example, the apex node of the tree structure in Figure 2 is rDEcLJ, so T9
D stored in the DECL conversion procedure rule storage unit of No. 1
Search the ECL conversion procedure rule group and obtain the following applicable rules.

((DECL umbrella 1)>→<(P-DECL・1)>
The left side (DECL umbrella 1) of (8) → represents a matching pattern with the input tree structure, and its contents are a tree structure with a vertex node called rDEcLJ, and the structure of its child node group is arbitrary. This means that the child node group is set to a variable called ``1''. On the other hand, on the right side of →, rP-DE is
In the conversion procedure rule application section of the zero-order step T10, which describes a procedure to perform a conversion called CLJ.

This conversion procedure rule is applied to obtain the conversion procedure data shown in FIG. 3-3.

The conversion procedure data shown in FIG. 3-3 has a list structure including three data, and the first data 340 is the second data 3°41 and the third koji data 342.
This represents the name of the procedure for creating a Japanese intermediate representation corresponding to the rDECLJ level in the English sentence parse tree using the respective conversion execution results. That is, the procedure name here is rP-DECLJ. Generally, the format of conversion procedure data is such that the first data represents a conversion procedure name, and the second and subsequent data are data that will be used in the conversion procedure represented by this conversion procedure after the conversion is executed. .

This data is stored in the conversion procedure storage section of the Tll.

The storage section of T11 is a first-in, last-out stack, and the contents at this point are as shown in Figure 3-4. In Figure 3-4, the number added to the leftmost end is: This conversion procedure data represents the node number of the corresponding node in the original English sentence parse tree, and this number is the node number of the corresponding node on the Japanese intermediate representation created by executing the conversion procedure data. becomes.

, The next step T12 is a conversion execution determination unit, and Tl
The contents of the conversion procedure storage unit of T1 and the conversion execution result storage unit of T16 are searched, and if the conversion procedure data stored most recently in T11 is executable, that data is retrieved and the data is retrieved from the next T13. The data is sent to the conversion procedure judgment unit. In this example, the conversion execution determination unit at T12 first examines the procedure data most recently stored in Tll, that is, the data shown in FIG. 3-3. Then, the conversion execution result storage unit of T16 is searched for conversion execution data regarding node number 2.3, but since nothing is stored in T16 yet, next, the child node of step 17 is searched for the third - 4 is selected from the procedure data shown in FIG. 4, and stored in the deformation conversion tree structure storage unit of T2. The contents of the storage section T2 at this point are as shown in FIG. 3-5.

Next, the tree structure selection section T3 first outputs the third 0-order transformation rule selection section T4, since the apex node of the target tree structure is rNPJ, the N of T53 is
The NP transformation rule group stored in the P transformation rule storage unit is searched, and only one applicable transformation rule is selected. However, in this example, there are no transformation rules to apply;
The target-binding tree structure passes through the transformation execution unit T6 and is stored as it is in the transformation result storage unit T7, and the contents of T7 are as shown in FIG. 3-6.

Here, a specific example of deformation processing for the rNPJ structure will be explained.

It ensures proper executive
on of the functions, (
9) For example, in the case of the input English sentence above, Section 17(a)
The English sentence parse tree shown in the figure is now rNPJ structure 260
Consider the case where 1 is transformed. "NP" structure 2901
, the following rules are applied from the group of transformation rules for the rNPJ structure.

<(ADJ 11)(N EXECUTION)(
PP (PREP OF) Umbrella 2)>4<(COMP
THA Ding) (NP $2) (BE-V BE) (A
DV (11), (VERBEXECUTHD)>
(10) The form and application of this transformation rule are the same as those described in the description of the transformation of the rDECLJ structure, but in this case, the partial apricot! The child structure column of structure 2601, that is, structure 26
, 92.2603 and 2604 are subjected to pattern matching processing. As a result of applying the rule, the cholerano structure 2
6o2.2603.260i1 is transformed into the structure 2606.2607.2608.2609.2610 in Figure 17(b), and when these are combined as a child structure of the rNPJ node, the rNPJ structure shown in Figure 17(c) is obtained. can get.

When the rNPJ structure subjected to this deformation is restored to its original structure, it becomes as shown in FIG. 17(d). This variant is rprop
er execution of the function
This is a process to translate ionsJ into a more preferable Japanese expression, such as ``the correct execution of the function,'' rather than the literal translation of ``the correct execution of the function'' in a stiff tone.

Next, an example of rNPJ structure modification for the conventional expression ra high 1 level of ~J will be shown.

The facility implants
a ”high 1level 0fdata
5ecurity, (11) The parse tree of this input English sentence becomes as shown in Figure 18(a), and now r
Consider the case where the NPJ structure 2701 is deformed.

<(DET A)(ADJ HIGH)(N L
EVEL) (PP (PREP OF) Umbrella 1)>-)<
(ADJ A-HIGH-LEVEL-OF) Umbrella 1>
(12) According to the above transformation rule, the first
8(b) is transformed like the "NP" structure 2702 in the figure, and rahighlevslof~J are grouped together as one adjective ra-high-1level-ofJ. This transformation greatly simplifies the subsequent English-Japanese conversion process.

As a final example of the modification of the rNPJ structure, we will show an example of processing expressions using present participles.

You use the file
ng the record, (13) The parse tree of the above input English sentence becomes as shown in Fig. 19(a), and when the rNPJ structure 2801 is transformed according to the following transformation rule (14), the rNPJ in Fig. 19(b) is obtained. 6 villages (PRPRTCL(VERB C0N
TAINING) Fist 2)〉→＜111(RELCL(P
RON Tl(AT)(VERB C0NTAIN)(
NP -z))> In this modification, the expression of the present participle is replaced with the expression of a relative pronoun that has the same meaning.

By the way, the next step T8 in FIG. 3-1 is as follows.

Since the apex node of the tree structure data of the transformation procedure rule selection section T7 is rNPJ, it searches the NP transformation procedure rule group stored in the NP transformation procedure rule storage section T93. , get only one applicable rule. In the NP conversion procedure rule group, as shown in equation (15) below, there is a prepositional phrase (P
As shown in applicable rules and formula (16) for the NP groove structure including P), the relational noun phrase (R
Although there are rules applicable to the NP groove structure including ELCL), in this example, such rules are not applicable, and the rule of formula (17) below is selected.

< (NP (NP skewer 1) (PP -2))>→
<CP-NP-PP ・1-2))< (NP (N
P-1) (RELCL 62))>-+<(P-NP-
RELCL Monkey 1 *2)>
(16) <(NP ml)>→<(P-SNP
Umbrella 1)> (17
) In the conversion procedure rule application section of the next step TIO, this rule is applied to obtain one procedure data shown in FIGS. 3-8. This data is stored in the conversion procedure storage section of Tll, and the contents of this storage section Tll are as shown in FIG. 3-9. The conversion execution determination unit T12 examines the most recently stored procedure data, ie, the data shown in FIGS. 3-8, among the contents of the storage unit Tll, and determines whether conversion is executable. The code and conversion execution data are searched for in the conversion execution storage unit at T16, but since nothing is stored in T16, this procedure data is not executed, and in the child node tree structure selection unit at the next step T17, these nodes.

The data is selected and stored in the modification transformation tree structure storage unit of T2. The contents of T2 at this point are as shown in Figure 3-10.

Next, the tree structure selection unit T3 selects the third-most recently stored content of the transformation conversion tree structure storage unit T2.
The transformation rule selection section of T4, which selects and extracts the tree structure in Figure 11, attempts to select a transformation rule for the vertex node of this tree structure, but since no transformation rule group for DET is prepared, The structure passes through the transformation execution unit at T6 as it is and is sent to the transformation result storage unit at T70. In the next T8 conversion procedure rule selection section, the previous rD E CL
As in the case of J, rNPJ, the DET conversion procedure rule group is searched to obtain the first rule (18), which is applied by the conversion procedure rule application unit TIO, and the result is stored in the conversion procedure storage unit. Store in Tll.

<(DtET -1)>→<CP-DEET Umbrella 1)> (18) The contents of Tll at this point are as shown in Figure 3-12.

The feasibility of this will be investigated. In the case of this procedure data, the data used by the conversion procedure base, P-DET, is the one that has been converted from the English word rTHEJ, so this is determined to be executable and is determined by the conversion procedure judgment unit of primary T13. , T
The DET structure conversion execution unit of 149 determines what should be processed and sends the English word rTHEJ there. In T149,
The determiner English-Japanese conversion dictionary storage unit of T155 is searched to extract one Japanese translation word "sono" for the English word rTHEJ, and a subtree structure of the Japanese intermediate expression shown in FIG. 4(a) is created. This result is stored in the conversion execution result storage section of T16, and T16 is used in FIG. 3-13.

Here, the conversion execution determination unit T12 stores the most recent procedure data in the conversion procedure storage unit Tll in the state shown in FIG. 3-14 and the conversion execution result storage in the state shown in FIG. 3-13. Search the contents of part T16, check that the conversion result of the tree structure whose apex node is EEEEI used by the procedure name P-5NP of this procedure data is already obtained in T16, and convert this procedure data. Replace the = structure part with the conversion result and change the contents of the conversion procedure storage section to 3-1.
Do as shown in Figure 5. Furthermore, it is checked whether the conversion result of the tree structure with Roro as the apex node of this procedure data is in the conversion execution result storage unit of T16, but if the converted tree structure with n as the apex node is retrieved at end number E. hand,
Since nothing is stored in T16, the procedure P-
3NP becomes unfeasible. In this case, originally
Since the feasibility of P-8NF has been investigated, there is no tree structure group for child nodes to be newly transformed, and the process proceeds to the primary tree structure selection section T3.At this point, the tree structure storage section for transformation transformation is performed. The contents of T2 are as shown in Figure 3-16.
At T3, a tree structure having the most recently stored Roro as its apex node is selected and sent to the transformation rule selection section at T4. Since a transformation rule group for rNJ is not prepared, this tree structure data passes through the transformation execution section of T6 and is stored as is in the transformation result storage section of T7. In the next conversion procedure rule selection section T8.

From the N conversion procedure rule storage part of T97, the following rule (1
9) is selected.

<(N Ko 1)> → <(P-N $1)>
(19) The conversion procedure rule application unit TIO applies the rule and stores it in the conversion procedure storage unit of Tll, as shown in FIG. 3-17.

Next, the feasibility of the most recent conversion is checked at T12 from the conversion procedure storage unit at T11. In the case of this procedure data, the data used by the conversion procedure name P-N is the English word r
Since it is a converted version of PROGRAM4, it is judged to be executable.

Next, the conversion procedure determination unit at T13 determines what should be processed by the N structure conversion execution unit at T145, and inputs the English word r.
Send PROGRAMJ. At T145, the noun English-Japanese conversion dictionary storage unit at T151 is searched and the English word rPROGR is searched.
We extracted the Japanese translation for AMJ, "program," and the semantic marker rLEJ, which is one of the semantic markers for nouns classified in a specific field, and is added to this "program." Figure 4 (b) shows the result. Create a subtree structure of the Japanese intermediate representation shown. This result is stored in the conversion execution result storage unit of T16, and T16 becomes as shown in FIG. 3-18.

Here, the conversion execution determination unit T12 selects the most recent procedure data in the conversion procedure storage unit Tll which is in a state similar to that shown in FIG. 3-15, and the conversion which is as shown in FIG. 3-18. Search the contents of the execution result storage unit T16, confirm that the transformation result of the tree structure whose apex node is GE used by the procedure name P-8NP of this procedure data has already been obtained in T16, and execute this procedure. The Rolo structure part of the data is replaced with the conversion result to make the contents of the conversion procedure storage section Tll as shown in FIG. 3-19. At this point, the procedure name P-8NP of this procedure data uses the φ conversion data “■I
Since Maro and EEE are available, this procedure data is executed. Namely.

Since the procedure name of this procedure data is P-3NP, it is sent by the conversion procedure determining section T13 to the SNP structure conversion execution section T142, where the conversion is executed. here,
Using the structure shown in FIGS. 4(a) and 4(b),
Since there is no ambiguity in the window layer that determines the translation and word order, a Japanese intermediate expression as shown in Figure 5 is created.

This result is.

The conversion execution result is stored in the conversion execution result storage unit T16, and T16 is the third
-20 It becomes as shown in figure. In addition, the conversion procedure storage unit Tl
l is as shown in Figure 3-21.

Next, the conversion execution determination unit T12 uses the most recent procedure data in the conversion procedure storage unit Tll shown in FIG.
The contents of the conversion execution result storage unit T16 in FIG.
Replace the part obtained in 6 with the conversion result,
The contents of the conversion procedure storage section Tll are set as shown in FIG. 3-22.

It is checked whether the conversion result for the tree structure exists in the conversion execution result storage section of T16, but since nothing is stored in T16, it becomes impossible to execute the conversion using one-step P-DECL. . In this case, since there is no tree structure group for child nodes to be newly transformed, the process proceeds to the tree structure selection section T3.

At this point, the tree structure to be the content of the tree structure storage section T2 for transformation conversion is selected and sent to the transformation rule selection section T4. Here, a transformation rule applicable to this tree structure is searched from the vP transformation rule storage T52 corresponding to the type of vertex node of this tree structure, that is, vP. But in this example case.

Since there are no transformation rules to be applied to this tree structure, Section 3-2
The tree structure shown in FIG. 3 passes through the transformation execution unit in the first step T6 and is sent as is to the transformation result storage unit in the next step T7.

Here, a specific example of deformation processing for the rVPJ structure will be described. For example, in the case of the first-order input English sentence (20), the English sentence parse tree becomes as shown in No. 14 (a), and now let us consider the case where the rVPJ structure 2301 is transformed.

The program makeS effic
ient use of the file,
(
20)<(VERB MAKE)(NP(ADJ $1
)(N USE)(PP(PREP 0F)-2))>
→＜(Aovsu)(VERB USE)(NP, 2)
> (21) The rVPJ structure 2301 is
The above rule (21) is applied from among the transformation rule group for the general "■P" structure. The form and application of this rule are the same as described in the description of the rDEcLJ structure transformation, but in this case
(a) Pattern matching processing is performed on a child structure example of structure 2301 in the figure, that is, structures 2302 and 2303. The result of applying this rule is rVPJ structure 2304 in FIG. 14(b). This transformation is
This process is for the idiomatic expression "rmake use of~" in English, and as a result, as shown in Figure 14(b), the noun rUSEl becomes the verb rUsEJ, and the adjective rEFFIC:IENTJ becomes the adverb rEFFICI.
It has been replaced by ENTLYJ. As a result, “+II
The expression "akeuseof~" can be translated into a more appropriate expression in Japanese.

Next, an example of a modification of the rVPJ structure for the idiomatic expression "rbsto~" will be shown.

The file is to be created
, (22)<(BE-V BE)(I
NFCL (INFTOTO)11)>4<(AUX
V WILL)-1)> (23)
The parse tree of the input English sentence (22) is as shown in FIG. 15(a). Let us now consider the case where the rVPJ structure 2401 is transformed. According to the transformation rule (23), (7) r in Fig. 15(b)
VPJ structure 2402 (7) is transformed as rbeto
~j are grouped together as one auxiliary verb rvillJ. This transformation simplifies subsequent English-Japanese conversions.

As a final example of a modification of the rVPJ structure, an example of a modification of the IT-TO syntax is shown.

You found it dafficult to
modify therecard,
(24) <(VERB FI
ND)(NP(PRON IT)(AJP(ADJ)
$1)(INFCL(INFTOTO)$2)))>
−+<(VERB FIND)(NP(COMP TH
AT) (INFCLCINFTOTO) Monkey 2) (VER
B l5) (^[)J41))>
(25) The parse tree of the input English sentence (24) is shown in Figure 16 (a), and rVPJ
When the lfl structure 2501 is deformed according to the deformation rule (25), a result like the rVPJ structure 2502 in FIG. 16(b) is obtained.

This transformation also greatly facilitates the subsequent English-Japanese conversion process.

As specific examples of transformation processing, we have so far explained those for the "rDECLJ structure, rNPJ structure, and Oyobi rVP" structure.
A specific example of deformation processing for a structure will be described.

The program is compiled
d in general, (26) For example, in the case of the input English sentence (26) shown above, the 20th
(a) The English sentence parse tree shown in the figure is obtained, and now consider the case where the rPPI structure 2901 is transformed.

<(PREP IN)(N GENERAL)>4<(
ADV GENERALLY)> (2
7) For rPPJ structure 2901, general "PP"
Rule (27) shown above from the group of transformation rules for the structure
will be applied. The form and application of this rule are the same as those described in the description of the rDECLJ structure transformation, but in this case the structure sequence of the child of structure 2901 in Figure 20(a), i.e. 2902.
2093 is subjected to pattern matching processing. However, on the right side of the → mark in rule (27) (ADV GE
For this rPPJ structure, the description ``NERALLY'' means that if the conditions are met, instead of replacing only the child structure of the rPPJ structure, the entire rPPJ clause, including the rPPJ clause, is replaced as shown in Figure 20(b): ” structure. The result of applying this rule is the “ADVJ structure 2094” in FIG. 20(c).
The idiomatic expression is changed to the simpler adverb r general.
Replace it as y J to make future English-Japanese-English conversions easier.

Next, an example of a modification of the rPPJ structure for the idiomatic expression rto some 5xtentJ will be shown.

The file is used to some
extent, (28) This input English sentence (28
) becomes as shown in Figure 21(a), and now rP
Consider the case where the PJ structure 3001 is deformed.

<(PROP To)(ADJ SOME)(N EX
TENT)>4< (ADV TO-5OME-EXT
ENT)> (29) This transformation rule (29
), rAovJn structure 3002 in Fig. 21(b)
It is transformed like.

rTo SOME EXTENTJ is summarized as one horn adverb equivalent phrase rTo-5OME-EXTENTJ. This transformation simplifies subsequent English-Japanese conversions.

As a final example of a modification of the rPPJ structure, ronpurp
An example of transforming the idiomatic expression of osaJ is shown.

You grass the file on pur
pose, (30)<(PREP 0N)(N
PURPO5H)>+((ADV INTENTIO
NALLYI> (31) Above input English sentence (30
) is shown in Figure 22(a), and rPP
When the J structure 3101 is transformed according to the above transformation rule (31), a result like the "ADv" structure 3102 in FIG. 22(b) is obtained. This transformation also facilitates subsequent English-Japanese conversion.

By the way, the data sent to the transformation result storage unit T7 in Figure 3-1 has a tree structure in Figure 3-24, and in the transformation procedure rule selection unit T8 of the 0th step stored in T7. , T7, the vertex node rVP of the tree structure stored in
According to J, the following applicable modification rules are obtained from the vP conversion procedure rule storage T92.

<(VP umbrella 1)>4<CP-VP at)>
(32) Next, in the conversion procedure rule application section of T10, this rule is applied to the target tree structure, and the result is stored in the conversion procedure storage section of Tll.

Tll at this point is as shown in Figure 3-25.

Next, the conversion execution determination unit T12 searches the conversion execution storage unit of the 6th for the most recently stored procedure data, -ta, among the contents of the storage unit Tll, but since it is not the T1 word itself, this For the procedure data, a tree structure with fruits as vertex nodes is selected and stored in the transformation/conversion tree structure storage section of T2, and the contents of T2 become as shown in FIG. 3-26.

Next, the tree structure selection section of T3 selects the most recently stored third--
Select the tree structure shown in Figure 27. The transformation rule selection part of T4 tries to select the transformation rule for the vertex node of this tree structure, but since the transformation rule group for VERB is not prepared, this tree structure passes through the transformation execution part of To as it is. death.

The conversion procedure rule selection unit of 0-order T8, which is sent to the transformation result storage unit of T7, searches for a group of conversion procedure rules for VERB, obtains the following rule (33), and applies this to the conversion procedure rule application unit TIO. and stores the result in the conversion procedure storage unit Tll.

<(Vll!RB Fist 1)>-)<(P-VERB
-1)> (33) Tl at this point
The contents of l are as shown in Figure 3-28.

Next, from the conversion procedure storage unit of T11, the most recent section T1
2, the feasibility of the conversion is examined. Procedure name P-VER
The data used by B is the English word rPROVIDEsJ.
Since this is a conversion of
The RB structure conversion execution unit determines what should be processed and sends the English word rPROVIDEsJ there. T148-1F
searches the verb English-to-Japanese conversion dictionary storage unit of T154 and retrieves the dictionary description for the English word rPROVIDEJ,
The main structure shown in FIG. 6 is made.

In block 6-01 of Figure 6, E
The contents of the English-Japanese conversion dictionary included with J are shown. Here, in an English sentence with rPROVIDEJ as the main verb, the subject, direct object, indirect object, and prepositional phrase are:
Which case particles to select when translated into Japanese sentences, in what order of Japanese noun clauses, and even rPROVIDE
It describes which translation word to select for J.

Four types of selections are described in blocks 601 (a) to (d), and an appropriate one that meets the conditions is selected. The rsBJ line selects a case particle based on the main noun of the noun phrase that constitutes the English subject (line "H"), and also selects a case particle based on the semantic marker (rLE, rUDJ, etc.) of the noun phrase. (line of “S”). This rS
One noun clause in a Japanese sentence is formed by the information in the line BJ. In the line rDOJ, a noun phrase that constitutes a direct object in an English sentence is described in the same way as rSBJ, and thus one noun clause is also formed in a Japanese sentence. Also, block 60177)(b
) When a specific preposition is written, as in the line ``rWITHJ''.

Used here with the verb rPROVIDEJ.

It represents a description for converting the expression of a prepositional phrase expressing a specific meaning into a corresponding Japanese noun clause.

That is, for the noun phrase located next to this preposition, r
As in the case of SBJ and rDOJ, case particles that meet one condition are selected to form one Japanese noun clause. The rPPJ O) line in the (C) part of block 601 describes the English-Japanese conversion of prepositional phrases that are often used with the verb pl rPROVIDEJ, and rFORJ.
The description in the line indicates processing for a prepositional phrase using the preposition rFORJ. The existence of a prepositional phrase in this case is (b
) is optional in English sentences, unlike the case of rWITHJ in the part. This description also makes it clear that r in part (b)
WITI (A single Japanese noun clause is formed in the same way as in the case of J.rsB", rDO").In addition to this, there is a line description of rIOJ. This is a description of the English-Japanese conversion of indirect objects in sentences, rsBJ.

One Japanese noun clause is formed using a method similar to "DO". The last line of each description part (a) to (d) of block 601, the line rPJ, is a description for forming a pragmatic part in one Japanese sentence, and is ' rPROVIDE
The translation for J and accompanying information for generating the auxiliary verb part of this Japanese verb in later Japanese generation, for example in part (a).

Contains descriptions such as ``lower first step,''``continuation,'' and ``transitive verb.''

By the way, this structure shown in FIG. 6 is stored in the conversion execution result storage section of T16, and T16 becomes as shown in FIG. 3-29.

Here, the conversion execution determination unit T12 replaces the tree structure conversion execution result with the 3-30th vertex node, and changes the contents of the conversion procedure storage unit Tll as shown in FIG. 3-31.

For the currently targeted procedure, the transformation execution result has not yet been obtained in the transformation execution result storage unit T16, so it is impossible to execute this procedure data, and the modification transformation tree structure storage unit T2 is From the state whose contents are shown in FIG. 3-32, the tree structure selection unit of T3 again extracts the tree structure for transformation. At the time when the transformation process for the tree structure to be the subsequent processing node and the transformation of the tree structure to be performed are applied, and the transformation procedure rules are applied and stored in the transformation procedure storage section of the Tll, the contents of the Tll are as follows: 3-33
It will look like the figure.

Among the contents of Tll, the tree structure conversion execution result with the most recently stored procedure is still stored in the conversion execution result storage unit T.
16, the tree structure of these child nodes is sent to the tree structure storage unit for transformation transformation of T2. At this point, the contents of T2 are as shown in Figure 3-34.

The specific processing procedure has already been described, so it will be omitted, but first, the result of the water groove deformation conversion processing shown in Fig. 3-34 is as shown in Fig. 7(a), and this result is ,
The contents shown in FIG. 3-33 are changed, and the contents of the conversion procedure storage unit at T11 become as shown in FIG. 3-35.

For the structure, after the deformation conversion process, the result is the seventh (
b) It will look like the figure. In FIG. 7(b), rFAJ
is a type of noun semantic marker in a specific field, and represents a certain function or ability. Here, the third
Among the procedure data shown in Figure 35, the procedure data of procedure base P-5NP becomes executable, and the execution result is shown in the third-
As shown in FIG. 36, the result is stored in the conversion execution result storage section of T16.

At this point, the conversion procedure storage unit Tll is as shown in Fig. 3-37, but as shown in Fig. 3-36, the conversion execution result of the tree structure with one card is stored, so take it out. Then, the procedure data of the currently targeted procedure base p-vp is replaced with a tree structure in which 7NP is the top note. Here, the procedure data of the procedure-based P-vP becomes executable, and the execution result is stored in the conversion execution result storage section 716 as shown in FIG. Further, the contents of the conversion procedure storage section T11 are as shown in FIG. 3-38.

Here, the structure stored in the conversion procedure storage unit T11 is as follows:
It is replaced with the conversion execution result already stored in the conversion execution result storage unit T16, and eventually the conversion procedure storage unit Tl
The contents of l are shown in Figure 3-39.

At this point, the conversion execution determination unit T12 determines that the procedure data of the procedure base P-DECL stored in the conversion procedure storage unit Tll is executable and sends it to the DECL structure conversion execution unit T141. At T141, semantic processing is performed using the sent tree structure data shown in FIGS. 5 and 8 to create an appropriate Japanese intermediate representation for the input tree structure shown in FIG. 3-2.

That is, in FIGS. 5 and 8, the subject of this input English sentence is block 501 in FIG.

Its semantic marker is rLEJ, and its direct object is block 802 of FIG. 8, whose semantic marker is rFAJ.
Under the condition that , only part (a) is selected from the verb conversion dictionary description 801 in FIG. As a result, the case particle taken by the subject "that program" is "ga", and the case particle taken by the direct object "these functions" is "ga".
In addition, the translation of the verb rPROVIDEJ is determined to be "preparation." Furthermore, the order of these words in Japanese follows the order of descriptions such as rSBJ, rDOJ, etc. in part (a) of block 801 in FIG.

If the subject or object is a polysemous word, an appropriate translation can be selected by performing semantic processing in two places.

For example, when targeting the field of information processing, in the case of an input English sentence (34) as shown below, in Figure 5 and
9(a) and 9(b) as structures corresponding to FIG. 8.
) get the figure.

The operator
key, (34) Block 9 in FIG. 9(a)
01 shows two translations for rTHEOPERA'rORJ, each separated by a semantic marker. In other words, the translation as rHMJ (semantic marker for humans) is "operator" and rLE
The translation as J (semantic marker representing a symbol etc.) is "operator", in this case block 90 of FIG. 9(b)
According to the condition described in 2, only the semantic marker rHMJ is allowed as the subject of this English sentence, so as a result, "operator" is selected as the translation of rTHEOPERATORJ.

Also, at the stage of processing the structures shown in FIGS. 5 and 8, if the options (a) to (d) described in block 801
, if block 501 and block 802 do not satisfy the conditions of any of the choices, the structure of FIGS. 5 and 8, that is, the English parse tree of FIG. 2 has no meaning in the target field. excluded as something that cannot be done.

Next, a Japanese sentence intermediate representation is formed based on the results obtained from the above semantic processing, and the structure shown in FIG. 10 is obtained. In Figure 10, rJDEcLJ is a Japanese sentence, rJN
cJ is a Japanese noun clause.

"JPC" is a Japanese pragmatic part, rJNPJ is a Japanese noun phrase, rJPARTJ is a Japanese particle, rJVERBJ
is a Japanese verb, and rJ VAUXJ is a clause identifier representing a Japanese auxiliary verb part.

゛ The result obtained here is stored in the conversion execution result storage section of T16. At this point, the conversion procedure storage unit Tll
is in a state where nothing is stored in it. Also, T
Nothing is stored in the transformation/transformation tree structure storage section 2. Therefore, the transformation process for the English sentence parse tree shown in FIG. 3-2 input from the English sentence parse tree input section T1 ends. The resulting tree structure shown in FIG. 10 is taken out from the conversion execution result storage unit at T16 and output by the Japanese sentence intermediate expression output unit at T18.

At this stage, the English-Japanese intermediate representation transformation conversion device 103 shown in FIG.
After the processing in FIG. 10 is completed, the Japanese intermediate expression shown in FIG. 10 is stored in the Japanese sentence intermediate expression storage device 113 shown in FIG.

The details of the processing procedure of the conversion execution determination unit T12 in FIG. 3-1 are shown in FIG. 3-40. Next, in the Japanese sentence generation device 104 shown in FIG. Generates a Japanese sentence by inputting a word-sentence intermediate expression. first,
According to the description of section rJVAUXJ in Figure 10,
0 Next, by extracting and arranging the terminal Japanese words in the structure of this Japanese sentence intermediate expression in the same order as they are, and adding a period mark "," at the end, we can create the following: Obtain the Japanese sentence shown in .

The program has these functions.

Finally, by outputting this result to a printing device or display device using the output device 105 in FIG.

゛The entire process using the English-Japanese machine translation method according to the present invention is completed.

Although preferred embodiments of the present invention have been described above, it goes without saying that various modifications can be made in carrying out the present invention.

Although the above embodiments can be implemented using a LISP program, the present invention can also be implemented using programs written in other languages or by hardware devices.

[Effects of the Invention] Hereinafter, the present invention will be compared with a conventional method, and advantages specific to the method of the present invention will be explained.

None of the conventional English-Japanese machine translation methods had the practical functionality to translate a practical scale of idiomatic expressions into appropriate Japanese expressions.

- For example, there is a method in which a dictionary of idiomatic expressions is searched for each English word during syntactic analysis of an English sentence, and in the case of an idiomatic expression, one exception processing is performed. In this type of method, idiomatic expressions are exceptionally processed during analysis, so if you try to process almost all idiomatic expressions commonly used in English sentences, the processing itself becomes complicated. This not only makes maintenance and management difficult, but also makes the analysis itself less efficient. Therefore, Tlt is actually in a situation where it can only process a limited number of idiomatic expressions. On the other hand, after parsing an English sentence and obtaining an English sentence parse tree,
Although there are conventional systems that process idiomatic expressions, they mainly process English expressions related to the selection of translation words such as Japanese verbs, nouns, and particles, and most do not process English expressions related to Japanese writing style and word order. There is no processing function.

In the method according to the present invention, appropriate processing is performed for English expressions related to Japanese writing style and word order by transforming the English parse tree, and by processing English-to-Japanese conversion processing, Japanese verbs, nouns, particles, etc. It is designed to appropriately process English expressions related to the selection of translation words, and can handle a variety of idiomatic expressions on a practical scale.

When trying to process English idiomatic expressions of practical scale types by traditional transfer methods.

From things related to Japanese writing style and word order, to verbs.

This is done by repeatedly applying a set of transfer rules, including those related to the translation of nouns and particles, but because the types of conversion are diverse, the conversion steps are complicated, and it is difficult to perform transfer rules on a practical scale. It becomes difficult to maintain and manage rules.

In the method according to the present invention, the English parse tree is first transformed into a simpler English basic sentence pattern, and then the English-Japanese This is dealt with by performing a conversion. This transformation process.

Transformation processing within the same English parse tree, ie.

Since it is a process of transforming idiomatic English sentence patterns into basic sentence patterns expressing the same meaning, it is very easy to understand and easy to maintain and manage. In addition, after transformation, English-Japanese conversion is performed, focusing on translating Japanese verbs, nouns, particles, etc., but the input is an English parse tree that has been transformed into a simple basic sentence pattern, and the processing is simple. As a result, maintenance and management become easier.

In conventional general tree structure transformation/conversion methods, a method is adopted in which a single transformation/conversion rule group is prepared as a so-called produlsion rule and is applied to an input tree structure. According to this method, the applicability of a tree structure is checked sequentially from the first rule in a group of transformation/transformation rules by pattern matching, and if applicable, the transformation - Apply transformation rules to transform and transform the tree structure. Next, the applicability of the next rule in the group of transformation/conversion rules is examined again for the tree structure obtained by this transformation/transformation. Then, when an applicable rule is found, it is applied to transform and transform the input tree structure,
Create another tree structure. When this process of applying rules is completed up to the last rule in the group of transformation/conversion rules, the process returns to the beginning of the group of transformation/conversion rules and continues applying rules in the same way. Then, when none of the rules in the group of transformation/conversion rules are applied to a certain input tree structure, the transformation/conversion process for the entire input tree structure is terminated.

The process of checking the applicability of rules and transforming/transforming them by pattern matching of tree structures inherently requires a lot of processing time, but in the above method, one transformation/conversion rule Because the group stores the rules for all node types of the tree structure together.

When examining the applicability of a rule, use an analogy.

Even if the rule is for a specific node, all nodes in the input tree structure must be examined, resulting in extra processing. moreover.

Since the application of rules to the input tree structure is repeated until the rules of the transformation/transformation rule group can no longer be applied, pattern matching is often performed a large number of times during that time.

In the method according to the present invention, transformation/transformation rules are grouped and stored for each type of target node in the tree structure, thereby reducing the number of unnecessary pattern matching operations. In addition, when applying each rule group, the applicability is checked starting from the first rule, but only the first rule that becomes applicable is applied once, and then the transformation/conversion rule group is applied for that node. end the application.

Furthermore, in the method according to the present invention, in order to minimize the number of searches for substructures of one structure in the transfer of the input English parse tree, that is, in the transformation/conversion process, the English parse tree is first transformed from the top node to the lower nodes. At the same time, at each node, following this transformation process, the procedure of English-Japanese conversion process is determined and stored. After the transformation process has finished up to the lowest node of the English parse tree, According to the procedure of the English-Japanese conversion processing that has already been stored, the Japanese intermediate expression is constructed in order from the lower-level partial structure toward the upper-level structure. In this way, the search for the partial structure of the input English parse tree only needs to be performed once, which is the minimum necessary tree structure search process.

In general, when searching a tree structure, starting from the highest node and proceeding toward lower nodes is much simpler than starting from the lowest node and proceeding toward higher nodes. Also, when transforming from a higher level to a lower level, the subtree structure that has already been transformed at the higher level will be transformed again later.

Since the transformation process is performed in the same way, there is also the advantage that the transformation rules can be written flexibly.

For the above reasons, the English-Japanese transfer method according to the present invention can perform English-Japanese transfer processing at high speed. .

[Brief explanation of drawings]

Figure 1 is a diagram showing an overview of one embodiment of the English-Japanese translation system of the present invention, and Figure 2 is an English sentence parsing tree obtained by parsing example sentences of English sentences used for detailed explanation of the present invention. illustration. FIG. 3-1 is a diagram showing in detail the English-Japanese intermediate expression transformation conversion device in the system of FIG. 1. Figures 3-2 to 3-39 are diagrams for explaining the processing process that the English sentence parse tree in Figure 2 undergoes in the English-Japanese intermediate representation and transformation conversion device in Figure 3-1. Figure 3-40 is 3-
FIG. 2 is a diagram showing details of the processing procedure of the first conversion execution determination part T12 in FIG. 1; 4(a) and 4(b), and FIGS. 5 to 8 are diagrams showing the process of converting the English sentence parse tree of FIG. 2 into a Japanese intermediate representation. Figures 9(a) and 9(b) are diagrams explaining an example of selecting translated words using semantic markers, and Figure 10 shows the intermediate representation of a Japanese sentence corresponding to the English parse tree in Figure 2. Figures 11(a) to 11(f) are diagrams for explaining an example of modification of the DECL structure. FIGS. 12(a) and 12(b) are diagrams for explaining a second example of modification of the DECL structure. 13(a) and 13(b) are diagrams for explaining a third example of deformation of the DECL structure, and 14(a)
14(b) are diagrams for explaining an example of a modification of the VP tank structure. 15(a) and 15(b) are diagrams for explaining a second example of VP tank structure modification, and 16(a) and 16(b) are diagrams for explaining a second example of VP tank structure modification. FIG. 3 is a diagram for explaining example No. 3; FIG. 17(a) to FIG. 17(d) are diagrams for explaining an example of a modification of the NP groove structure. 18(a) and 18(b) are diagrams for explaining a second example of NP groove structure deformation, and 19(a) and 19(b) are diagrams for explaining a second example of NP groove structure deformation. 20(a) to 20(Q) are diagrams for explaining an example of modification of the PP structure. FIG. 21(a) and FIG. 21(b) are diagrams for explaining a second example of modification of the pp structure. FIG. 22'(a) and FIG. 22(b) are diagrams for explaining a third example of modification of the pp structure. Applicant International Business Machines Corporation Agent Patent Attorney Oka 1) Tsugi Sei (1 other person) 2nd WJ Iku 3-3 Figure 3-6 Figure 3-8 Figure 3-9t12 Figure 3-10 Fang Figure 3-14 Figure 3-15 Figure 3-16 ÷ 3-18 factor -! d-3-23 Figure 3-24 +: Woll 3-27 Figure 3-29 Off (α) Figure Fang 7 (b) 1!1 Sweet 9 (Q) Figure Sweet 11 (b) llffflI Fang 11 (c) Figure fil (d) Zuto 11 (e) Zu Sai 14 (α) Quick appetizer 14 (b) Figure + 15 (a) Zu sweet 15 (b) Figure threat 16 (b) 17th (a) Figure 17 (b) Figure 17 (c) Figure + 18 (b) Figure 19 (.) Figure + 19 (b) Figure 20 (α) Figure 20 (b) Figure 20 (c) Figure + 21 (a) Figure 21 (b) Figure 22 (a) Figure + 22 (b)

Claims (2)

[Claims] (1) An English sentence analysis device that parses an input English sentence and outputs a parse tree of English sentence intermediate expressions, and for transforming idiomatic expressions and special constructions existing in the English sentence intermediate expressions into basic sentence patterns. A transformation rule dictionary that stores a group of transformation rules for each node type of the above parse tree, and a transformation rule dictionary that stores a group of English-Japanese conversion rules for converting the parse tree into a Japanese sentence intermediate representation for each type of node of the parse tree. an English-Japanese conversion rule dictionary that is maintained separately in an English-Japanese intermediate expression transformation conversion device including means for forming a Japanese sentence intermediate expression by applying the English-Japanese conversion rules from the lower to the upper transformed branches of the parse tree along the structure of the parse tree; An English-Japanese machine translation device comprising a Japanese sentence generation device that generates a Japanese sentence from a word-sentence intermediate representation. (2) The above-mentioned English-Japanese intermediate expression transformation conversion device has means for simultaneously determining the conversion procedure to Japanese sentences at each node when transforming the above-mentioned analytic tree from the upper to the lower direction, and reaches the terminal node. The English sentence recited in claim (1) is characterized in that, when the determination of the transformation and conversion procedure is completed, the Japanese sentence intermediate representation forming means executes the conversion operation from the lower to the upper part of the parse tree according to the conversion procedure. Japanese machine translation device.