JP2608384B2 - Machine translation apparatus and method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a machine translation apparatus and method for translating an input source language sentence into a target language sentence in an automatic translation system or an information retrieval system. In particular, the present invention relates to a machine translation apparatus and method for realizing an expression intended by the author in a target sentence faithfully in the meaning of the original sentence.

[0002]

2. Description of the Related Art As a conventional machine translation system, as described in "Language Processing and Machine Translation" (Hirogo Nomura / ed.), "Direct system", "transfer system", "intermediate language system" And so on.

[0003] The "direct method" is one that appears in the original text.
In this method, each word is replaced with a word in the target language, and then the word order of the translated word is replaced to complete the translated sentence.

[0004] The "transfer method" is based on the formal conversion of the expression of the structure of a sentence between languages, and the result of grammatically analyzing an original sentence is expressed as a syntax structure expression of the original language. The structural expression is converted into the syntactic structural expression of the target language, and a translated sentence is created from the converted syntactic structural expression. In this way, the translation is performed via an indirect intermediate expression called a syntax structure expression.

In the "intermediate language system", one expression means independent of language characteristics is set in place of the syntax structure expression, and a target is directly obtained from the intermediate expression of the original text described using the expression means. This is a method that considers creating sentences.

[0006] In any of the above three methods, the translation process is performed through the common processes of morphological analysis and morphological generation as the translation process. In the "transfer method" and the "intermediate language method", there are processing steps of syntactic structure analysis, semantic structure analysis, vocabulary / structure conversion, semantic structure generation, and syntactic structure generation. Therefore, in the conventional machine translation method, there is a common processing step of once dividing a sentence into minimum units of words and rearranging translated words according to the syntax structure of a target language.

According to this method, in a phrase expression or an idiom expression, the parsing rule or the generation rule is not skillfully applied, or even if fired, it is faithfully and accurately translated according to the original text. In many cases, the expression becomes awkward, and when there are a plurality of dependencies, there is no guarantee that the result can be obtained with high accuracy in eliminating polysemy.

As a supplement to this drawback, a "vertical search method" and a "horizontal search method" have been considered. However, these two are discussed only between the embedded sentence and the candidate noun to which the embedded sentence is to be modified, and the range of use is limited to noun phrases. In addition, there is a problem in eliminating polysemy of the embedded sentence and polysemy of the candidate noun to which the embedded sentence is modified, and it is necessary to analyze with high precision and evaluate the goodness of the entire generated noun phrase.

To solve this drawback, Japanese Patent Application Laid-Open No. 63-2130
Japanese Patent Laid-Open No. 65 further discloses that the structure of an embedded sentence and a noun to be modified are expressed as a two-dimensional matrix, and the modification relation is analyzed for a candidate noun.

[0010]

However, the method disclosed in Japanese Patent Application Laid-Open No. 63-213065 does not hold information on the location of the noun phrase generated by automatic translation in the target sentence. One of the drawbacks of the machine translation system itself is that there is no guarantee that the semantic relationship of the generated translation itself is destroyed and the word order of the generated translation is determined with high accuracy.
It did not solve the possibility of significantly lowering the sensitivity of the translated sentence as a sentence.

As described above, in the conventional machine translation system,
After the morphological analysis, the group of connected words is once divided into the minimum units of words, so that information indicating that the words are connected cannot be held until the translated sentence is generated. For example, in the conversion of a Japanese sentence to an English sentence, the conjunctive modifier and the adnominal modifier are converted into an adverb phrase and an adjective phrase, formed as a connected word group, and translated as a connected word group in a correct translation. You.

However, in a machine translation process, when there are a plurality of modifiers composed of a word group, since the coherent information is not used as the word group, the part-of-speech information at the stage of the morphological analysis is accurately determined. The word that should be combined among multiple word groups is nested in a different group of words, due to the lack of ambiguity when there are multiple destinations of word dependency, However, there is a problem that a translation generation result of unknown meaning often appears.

[0013] In this case, a translated sentence may be generated as a word order or a modification position that does not look like a target sentence.

Further, in the ordinary processing of conventional machine translation,
Since the translation is generated by rearranging the words in units of words, the translation rules and syntax generation rules of the machine translation system are restricted, and conversely, it may be difficult to derive a translation that matches the actual situation.

For example, the part of speech in the original sentence is a noun, but in the translated sentence, a syntactic structure is formed between the original sentence and the translated sentence, such as generating a clause or converting a verb into an adverb and one of the compound nouns into an adjective. If the conversion pattern cannot be generalized, it is bound by the general conversion rules and syntax generation rules of a normal machine translation system. Will be generated.

For example, the Japanese sentence "Today is fine weather. Is translated into English by machine translation.
day is good weather. Also, "Switch ON"
Becomes "switch ON". In the above, the commonly used expression is "It's fine today" and "switch
is turned ON], it is difficult with normal machine translation processing.

That is, a specific wording expression is difficult to follow the usual conversion rules and syntax generation rules of machine translation. When it is difficult to associate the pattern of the conversion of the syntactic structure, or the difference in the part of speech between the original sentence and the translated sentence, not only a correct translation cannot be generated, but also a mistranslation may occur.

Although a machine translation system using the embedding method provides a solution to these problems, as described above, the conventional embedding method uses an embedding sentence and a candidate noun to which the embedding sentence is to be modified. All the while, the ambiguity of selecting candidate nouns between them has been resolved, and the above-mentioned problem has not been solved. Furthermore, in the conventional embedding method, the target is limited to a noun phrase composed of a noun and a word that modifies the noun, but in the original sentence, the adjunct modifier, the repeated modifier, the independent sentence, etc. In a translated sentence, it is necessary to solve the above-described problem for all of the sentences such as a noun phrase, an adverb phrase, an adjective phrase, a verb phrase, and a clause.

[0019]

According to the present invention, a machine translation device stores a specific word used in a specific expression of an original sentence as an original sentence frame, a combination of the original sentence frame, and the original sentence frame. And for the attribute of a word or a group of words forming the specific expression, a translation frame corresponding to the original text frame and a syntax from the specific expression to a translation expression including the translation frame corresponding to the specific expression. A conversion rule storage unit that stores the conversion pattern of the structure as a rule, a dictionary unit that stores the attributes and translations of the words of the original sentence, an input unit that inputs the original sentence to be translated, and an input from the input unit. From the original text,
A detecting means for detecting an original sentence frame by referring to the frame storing means; and, in the input original sentence, a word or a group of words constituting a specific expression in combination with the original sentence frame detected by the detecting means. A search unit for searching attributes and translations from the dictionary unit, an original frame detected by the detection unit, and an attribute of the word or word group searched by the search unit, applied from the conversion rule storage unit. Selecting means for selecting a rule to be selected, a translation frame corresponding to the detected original text frame, and a translation from a specific expression including the detected original text frame to a translation expression based on the rule selected by the selecting means. Deciding means for deciding a conversion pattern of the syntax structure, and a translation frame determined by the deciding means, Generating means for embedding a translation of the searched word or group of words at a position based on the replacement pattern and generating a translation expression corresponding to the specific expression in the input original text; Determining means for determining whether a part other than an expression is included, and
When it is determined that a part other than the specific expression is included, a translation unit that translates the input original text as a unit of the specific expression, and a part corresponding to the specific expression in a translation result by the translation unit, Replacement means for replacing with the translation expression generated by the generation means.

According to another aspect of the present invention, in the machine translation method, an inputting step of inputting an original sentence to be translated, and an original sentence input from the inputting step are used to specify an original sentence. A detecting step of detecting an original sentence frame by referring to a frame storage unit in which a specific phrase is registered as an original sentence frame, and combining the input sentence with the original sentence frame detected by the detecting step to form a specific expression. A retrieval step of retrieving the attributes and translations of the constituent words or word groups from a dictionary in which the attributes and translations of the original sentence are registered; and forming the specific expression in combination with the original sentence frame and the original sentence frame. Translation table including the translation frame corresponding to the original text frame and the translation frame corresponding to the specific expression from the specific expression for the attribute of the word or word group to be changed From the conversion rule storage unit storing the conversion pattern of the syntactic structure into the rules based on the original frame detected in the detection step and the attribute of the word or word group searched in the search step. A selecting step of selecting a power rule, a translation frame corresponding to the detected original sentence frame based on the rule selected by the selecting step, and a syntax from a specific expression including the detected original sentence frame to a translated expression. A determining step of determining a conversion pattern of a structure; and a translated word of the searched word or group of words at a position based on the conversion pattern determined by the determining step with respect to the translation frame determined by the determining step. Generating a translation expression corresponding to the specific expression in the input original text; and And a translation step of translating the input original text as a unit of the specific expression when it is determined by the determination step that the part includes a part other than the specific expression. And a replacement step of replacing a portion corresponding to the specific expression in the translation result of the translation step with the translation expression generated by the generation step.

[0021]

According to the machine translation apparatus of the present invention, the detecting means refers to the frame storage means to detect an original sentence frame from the original sentence inputted from the input means, and the searching means finds the original sentence frame in the inputted original sentence. The dictionary unit searches for attributes and translations of a word or a group of words constituting the specific expression in combination with the original sentence frame detected by the detection unit, and searches the original sentence frame detected by the detection unit and the search unit. When a rule to be applied is selected from the conversion rule storage unit by the selection unit based on the attribute of the selected word or word group, the identification unit including the detected original text frame is determined by the determination unit based on the selected rule. A conversion pattern of the syntax structure from the expression to the translation expression is determined, and the determined conversion pattern is generated by the generation unit with respect to the determined translation frame. A position based on the over down, burying the translation of the retrieved words or word groups, to produce a translation representation corresponding to a particular representation in textual the inputted. When the judgment unit judges that the input original sentence includes a part other than the specific expression, the translation unit translates the input original sentence as a unit by the translation unit, and The part corresponding to the specific expression is replaced by the generated translation expression by the replacing means.

Further, in the machine translation method of the present invention, an original sentence frame is detected by referring to a frame storage unit in a detection step from an original sentence input in an input step, and is searched for in the input original sentence in a search step. Searching the dictionary process for the attributes and translations of the words or word groups that make up the specific expression in combination with the original text frame detected by the detection process, and the original text frame detected by the detection process;
When a rule to be applied is selected from the conversion rule storage step by the selecting step based on the attribute of the word or the word group searched by the searching step, based on the selected rule, the rule is detected by the determining step. A conversion pattern of a syntax structure from a specific expression including an original text frame to a translation expression is determined, and the search step is performed by the generation step at the position based on the determined conversion pattern with respect to the determined translation frame. The translated word of the word or group of words is embedded, and a translated expression corresponding to the specific expression in the input original sentence is generated. When it is determined by the determining step that the input original sentence includes a part other than the specific expression, the input original text is translated as a unit by the translation step, and the translation result in the translation result by the translation step is obtained. The part corresponding to the specific expression is replaced with the translation expression generated in the replacing step.

[0023]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, a preferred embodiment according to the present invention will be described in detail with reference to the accompanying drawings.

FIG. 1 is a block diagram showing a basic configuration of one embodiment of a machine translation apparatus according to the present invention.

In FIG. 1, reference numeral 1 denotes an input device for inputting an original sentence to be translated, for example, a keyboard for inputting characters. Alternatively, a communication device that receives character information from another device, a device that inputs and recognizes voice, and an OCR that optically recognizes characters may be used. Reference numeral 2 denotes a processing unit for embedding and translating an input original sentence.
The PU is realized by executing a program stored in a memory such as a ROM or a RAM.

Reference numeral 3 denotes an output device for outputting a translated sentence obtained through the processing unit 2, for example, a display for displaying character information. A device for synthesizing and outputting voice;
A communication device that transmits character information to another device may be used. Reference numerals 4 and 5 are an auxiliary storage device and a main storage device used during the operation of the processing section 2, respectively. The auxiliary storage device 4 uses a nonvolatile storage medium such as a floppy disk or a hard disk. The contents of the auxiliary storage device 4 may be obtained by reading necessary portions in the course of the processing, or may be loaded in the RAM in advance before the processing. Further, the contents may be stored in the ROM instead of the auxiliary storage device. The main storage device 5 is R
AM, and temporarily stores data generated in the process.

The processing section 2 has a frame extracting section 6, a code embedding section 7, a translation processing section 8, an embedded sentence analyzing section 9, an embedded sentence converting section 10, and an embedded sentence generating section 11. Also,
The auxiliary storage device 4 includes a frame expression dictionary 12, an embedding pattern 13, an embedding translation word dictionary 14, and embedding rules 15. The main storage device 5 stores an original sentence → translation sentence conversion table 16, an embedded sentence table 17, * input18, and * output19.

An arbitrary document is input from the input device 1 as a translation target. Hereinafter, the input document to be translated is referred to as an original sentence. The input text is * i
Stored in nput18. The frame extracting unit 6 extracts the frame expressions registered in the frame expression dictionary 12 from the original text of * input 18 with reference to the frame expression dictionary 12 in which specific phrase expressions and idioms are registered as frame expressions. .

The embedded sentence analyzing unit 9 determines an embedded pattern 13 and an embedded translated word dictionary 14 based on the arrangement of words, phrases, sections, and sentences used in combination with the extracted frames.
, A rule matching the embedding rule 15 is fired, the embedding pattern of the original sentence is determined, and the syntax structure is determined according to the embedding pattern in the translated sentence.

The embedded sentence conversion unit 10 converts the words, phrases, sections, and sentences used in combination with the frames of the original sentence into translated words while referring to the rule that fired and the embedded translated word dictionary 14. The syntax structure is converted according to the determined embedding pattern. The converted result is the original →
Output as the translation conversion table 16.

Here, when the document input by the input device 1 is composed of a registered original sentence frame and words, phrases, sections and sentences combined with this frame, the embedded sentence generation unit 11 According to the determined sentence embedded syntax structure, words, phrases,
Embed clauses and sentences inside or outside to generate translated sentences. Add morpheme generation to this and output * output19
Is stored. The contents are output by the output device 3.

On the other hand, if the input original text does not depend only on the registered frame and the words, phrases, clauses, and sentences combined with the registered frame, the code embedding unit 7
Referring to the original-to-translation conversion table 16, the portion to be embedded in the original is converted into an embedded code. At the same time, an embedded sentence table 17 that associates the position of the code embedded in the original sentence with the translation of the portion to be embedded is stored in the main storage device 5.
Expand on top. The embedded sentence table 17 and the original sentence (for example, FIG. 8) in which the embedded code is embedded are delivered to the translation processing unit 8, and are translated through a normal machine translation process.

The embedded sentence generation unit 11 refers to the translation result obtained by the translation processing unit 8 and the embedded sentence table 17, and converts the embed code in the translation result into the translated sentence of the corresponding embedding target portion of the embedded sentence table 17. And add morpheme generation over the entire input text. * Outp
Generate a translation on ut19. The result is output from the output device 3.

FIG. 12 is a flowchart of the translation operation of the apparatus of this embodiment. The translation procedure of this embodiment will be described with reference to FIG.

Step S1: Input of Original Sentence An arbitrary document is input by the input device 1 as an original sentence to be translated. This is stored in the main storage device 5 as * input 18, and the process proceeds to step S2. Here, it is assumed that the example sentence shown in FIG. 2 has been input.

Step S2: Detect registered original text frame
Sentences of the outgoing Figure 2 includes an embedded object part. With the expressions in this example sentence, the headings of the frames of the frame expression dictionary 12 of FIG. 3 stored in the auxiliary storage device 4 are sequentially scanned, and the embedding target portions “according to” and “from to "" And "is the same" are detected. The first and second frames of the i-th headline of the original text frame detected from * input18 are stored in areas Fr1 (i) and Fr2 (i).

"According to" will be described with reference to FIG.
It can be understood from the frame expression dictionary 12 shown in (1) that the words (first and second embedded words) used in combination are nouns. The embedding pattern is 2, and from the second pattern of the embedding pattern 13 in FIG. 4, the pattern in the original sentence is “first embedding word + first frame (+ second embedding word)”. The candidate for the embedding pattern is “first frame + first embedding word”. To be precise, the embedding pattern is not determined until the embedding rule is fired. As described above, when a frame can be detected, the process proceeds to step S3, and when the frame cannot be detected, the process proceeds to step S9.

Step S3: At step S3, the process goes to the embedded sentence analyzing unit 9, and
Performs dictionary lookup for sentences containing "according to". In "Determining bias according to measured value" in the example sentence of FIG.
Since the adjacent words of “according to” are “measurement value” and “bias”, the dictionary translation of the embedded translation dictionary 14 shown in FIG. 5 is performed, and from the third and sixth lines, the part of speech of “bias” is In the noun, the semantic attribute is understood to be “degree”. Similarly, it can be seen that the part of speech of “measurement value” is a noun and the semantic attribute is “environment”. Here, the result of dictionary lookup is stored in each area, and the headword area: JWd2 = bias / JWd1 = measured value, semantic attribute: IAtr2 = approximately / IAtr1 = environment, translation: EWd1 = m
stored in the main storage device 5 as an easurement, and
Go to S4.

Step S4: From the semantic attributes IAtr1 and IAtr2 obtained in the firing step S3 of the embedding rule
IAtr1 == RAtr1, IAtr2 == RAtr2 is satisfied, and based on the frame header obtained in step S2, Fr1 (i) == JFrm1
(j), a rule where Fr2 (i) == JFrm2 (j) is fired from the embedding rule 15 in FIG.

Referring to “determining bias according to measured value” in the example sentence of FIG. 2, in the frame translation word selection condition of embedding rule 15 in FIG. 6, the semantic attribute of the first embedded word “measured value” is “ Environment ”, and the semantic attribute of the second embedded word“ bias ”is“ degree ”. Therefore,
The first rule that satisfies the translation selection condition is fired. As described above, when the embedding rule is fired, the process proceeds to step S5, and when not fired, the process proceeds to step S9.

Step S5: Determination of embedded syntax and translated text
It is determined that the embedding pattern is 2 by the firing of the rule in step S4 for determining the inheritance information until generation . Referring to the embedding pattern 13 (FIG. 4) in the auxiliary storage device 4, since n (j) = 2, the original sentence frame and the translated sentence frame are as follows.

JPatern (j) = first embedded word + first frame (+ second embedded word) EPatern (j) = first frame + first embedded word

The part-of-speech role of the part to be embedded in the sentence is a continuous modifier according to the firing rule of the embedding rule 15 (FIG. 6).

In the example sentence shown in FIG. 2, there are three identical original sentence frames "according to".

The bias is determined according to the measured value.

The target potential is adjusted according to the standard conditions.

The microcomputer calculates the size of the original according to the magnification.

The portions to be embedded are as follows. (Hereinafter, the above example sentence is shown.) According to the measured value, according to the standard condition, according to the magnification, the first rule is fired in step S4, so the frame translation of “according to Is "to sui
t ”. And also from step S2
After the same procedure in step S4, the second rule fires, and the frame translation is "according to"
, The third rule fires and the frame translation is determined to be "based on".

According to the present embodiment, as shown in the example sentence of FIG. 2, even if a plurality of the same frame headings exist in the original sentence, the translation is divided by the firing of the translation word selection condition of the frame. Conversely, even if a polysemy exists in the frame translation, the polysemy selection can be eliminated through the same procedure. Furthermore, the role of the portion to be embedded in the translation is to generate an adverb phrase, the generation position is at the end of the sentence, and it is determined that the verb is modified from the end of the sentence.

These pieces of information are held together with the embedded code as inheritance information until a translated sentence is generated.

Example of inheritance information up to the time of generation of translated sentence Example "Determine bias according to measured value" Word order of embedded translation Position (j) = reverse order Part of speech in original sentence JPart (j) = Modified word of continuous use Part of speech in translated sentence EPart (j) = Adverb phrase: end of sentence Embed code Code (j) = # 200 && When the syntax and inheritance information are determined, the process proceeds to step S6.

Step S6: Inserting the part to be embedded into the translated sentence
In the conversion step S6, the process proceeds to the embedded sentence conversion unit 10, and the embedded translation word determined in step S3 and the translation pattern determined in step S4 and step S5 convert the original word into a translation word and convert the syntax structure. Do. Example: “Determine bias according to measured value” I) Conversion of embedded word from original text to translated text JWd1 = measured value → EWd1 = measurement (JWd2 = bias) Used for firing rules II) Translated text from original text pattern Conversion to pattern n (j) = 2 JWd1 + JFrm1 (+ JWd2) = According to measured value + (+ bias) ↓ EFrm1 + EWd1 = to suit + measurement For conversion result obtained by embedded sentence conversion unit 10 Add morpheme generation. In this example, "the" is added, resulting in "to suit the measurement". Hereinafter, only results are shown, and are as follows.

The target potential is adjusted according to the standard conditions.

Embedding part: → ac according to standard conditions
The microcomputer calculates the size of the document according to the magnification.

Embedding part: → based on magnification
On the reduction ratio The other embedding target portions processed by the embedding sentence conversion unit 10 are similarly converted. Here, the conversion result of the portion to be embedded obtained by the above-described conversion and the embedding code are associated with each other on a table, and are developed on the main storage device 5 as an original sentence → translation sentence conversion table 16 shown in FIG.

At this stage, the modification destination in the original text may not be obtained as information. The reason is that the part-of-speech information is given by the grouped words, so that in the subsequent machine translation processing, the combination of candidates for the modification destination is limited, and the modification destination can be easily determined.

About the contents stored in association with the embed code Embed code Code (j) = # 200 && Target part to be embedded in original text JWd1 + JFrm1 = Conversion result in translation according to measured value EFrm1 + EWd1 = to suit the
measurement Part of speech in translation Epart (j) = adverb phrase: end of sentence Next, the process proceeds to step S7.

Step S7: sentences to be embedded and sentences not to be embedded
If there is no frame that matches the header of the original text frame determined as a chapter , all of the input document is processed through normal machine translation processing. Conversely, when the original sentence input by the input device 1 is composed of words to be combined with the registered frames, morpheme generation is added to the entire generated document, and the output device 3 outputs the generation result.

Actually, it is rare that the whole sentence is composed of the sentences to be embedded. In the example sentence of FIG. 2, there are five sentences to be embedded, and the non-target includes five sentences. Therefore,
It is necessary to synthesize a translation result obtained by the embedding method and a translation result obtained through normal machine translation processing.
Here, the above is determined.

As a result of the determination, if all of the input documents are composed of sentences to be embedded, the process proceeds from step S7 to step S10, where the embedded sentence generation unit 11 performs processing. On the other hand, if the input document is composed of a sentence to be embedded and a sentence that is not
Example sentence) proceeds to step S8.

Step S8: Embedding the embed code in the original text
Referring to original → translation conversion table of FIG. 7 writing, to convert the embedding target portion in a textual embedded code. By determining the embedding position of the embed code in the original text,
The embedding sentence table 17 of FIG. 9 in which the embedding position and the embedding code are associated with each other is developed on the main storage device 5.

In this table 17, the row and the digit are used as the embedding position information. Alternatively, even if the appearance order of the embedding code is given, the same function can be performed in the transition processing. it can. Example: " Determine bias according to measured value " → "# 200 &&
Determine the bias. Embedding position information: No: BdN (j) = 1 Line: Rw (j) = 1 Digit: clm (j) = 1 Embedding code: Code (j) = # 200 && Translation target part generated: EFrm1 + EW
d1 = to switch the measure
nt In this example, since the embedding code is # 200 &&, the part-of-speech role in the translated sentence is an adverb phrase with reference to the original sentence → translation sentence conversion table 16, and the generation position information modifies the verb from the end at the end of the sentence. Can be retained as inheritance information.

The other embedding target parts are similarly replaced with the embedding code, and the original text in which the “embedding code” is embedded is obtained as shown in FIG. This is stored in * input on main memory 5.
Then, the original sentence → translation sentence conversion table 16 and the embedded sentence table 17 are delivered to the translation processing unit 8. The code embedding unit 7 is responsible for the above processing. Subsequently, the process proceeds to step S9.

Step S9: Processing in the translation processing section 8 The translation processing section 8 refers to the original-to-translation sentence conversion table 16 and the embedded sentence table 17 to determine the part of speech and the generation position in the translated sentence of the embedded code. Perform translation processing.

Inheritance information based on original-to-translation translation table 16 Embedding code: Code (j) = # 200 && Part of speech in original: Jpart (j) = Continuous modifier Modifier: ModifyPart (j) = determine Part of speech in translated ： Epart (j) = Adverb phrase: end of sentence * Before translation of example sentence expanded on input (shown in Fig. 8)

[0066]

[Outside 1] Example sentences generated by translation processing unit 8

[0067]

[Outside 2]

Referring to the inheritance information described above, a translation result of the embedded code is generated at the end of the sentence, and a translation is generated by the translation processing unit 8 so as to modify the verb from the end of the sentence. The same processing is performed on other embedded codes, and the result generated by the translation processing unit 8 is expanded on * output19.
FIG. 10 shows a translated sentence during the generation of this result. Step S10
Proceed to.

Step S10: Convert the embedded code into a translated word In step S10, the generation result (* outp
ut19) is passed to the embedded sentence generation unit 11. The embedded sentence generation unit 11 detects an embedded code from the generation result,
Referring to the embedded sentence table 17 shown in FIG. 9, the embedded code is replaced with the generated translation of the corresponding portion to be embedded.

Generation result passed from translation processing unit 8
(Shown in Figure 10) * Contents of output19: Determines the bias# 200 &
&Inheritance information of embedded sentence table 17 in FIG. 9 Embedding position No: BdN (j) = 1 Embedding code: Code (j) = # 200 && Translation of part to be embedded EFrm1 + EWd1 = to suit the measurement Embedding sentence generation Result generated by part 11 `` Determines the bias to suit the measuremenet. '' * Output 19 embed code # 200 && detected from above,
Referring to the embedded sentence table 17, the embedding position BdN (j)
EFrm1 + EWd1 = to suit the measurement
Convert. As a result, the embedded sentence generation unit 11
To get the result generated by About other embed codes
However, by performing the same processing, the
Get embedded sentence. In this processing, * output1
9 The necessary morpheme generation is performed on the generation result in 9.

For example, # 400 &&. Before embedding code conversion after end of LSTR until the COPY START KEY is turn
ed ON. After morpheme generation After end of LSTR until the COPY START KEY is turn
ed ON. In this way, there are tasks such as capitalizing the beginning of sentences. Of course, the processing of three simplexes and the processing of a single simplex are also executed as necessary.

Step S11: The output device 3 outputs the contents of the output * output19 from the output device.

In the above embodiment, the frame expression dictionary
12 is the heading of the frame, the part of speech attribute of the embedded word,
With the number of the embedding pattern, the embedding rule 15
Has the heading of the frame, the translation of the frame, the semantic attribute of the embedding word as the selection condition of the translation of the frame, the word order of the embedding translation word, the conversion code of the embedding target part, and the number of the embedding pattern. ing.

However, the configurations of the frame expression dictionary 12 and the embedding rule 15 are not limited to the above. A frame can be extracted by referring to both, and the translation of the frame and the embedded translation can be extracted based on the part of speech / semantic attribute of the embedding word. , The conversion code of the embedding target portion, and the embedding pattern are determined, even if there is a difference in some of the items between the two, even if it is divided into three or more, They may be combined into one. Further, even when there are three or more embedding portions, it is possible to cope with the same method.

As long as the functions of the present invention are performed, a single device or a system including a plurality of devices may be used. Further, it goes without saying that the present invention is applied even when the processing is performed by supplying a program to an apparatus or a system.

According to the present embodiment described above, in expressions used repeatedly or in specific wording expressions, expressions of the original sentence and the translated sentence are registered as frames, and words, phrases, and combinations used inside or outside the frames are registered. Sections and sentences are registered and stored as a translation dictionary.

This is because just () is provided in the text,
The frames outside the parentheses are registered as frames, and the words embedded in the parentheses are registered as a translation word dictionary in advance. Then, from the flow of words that appeared before and after (), (
Embed the words in parentheses). This allows
When a frame is detected in a sentence, a word to be used in combination, that is, a word to be embedded in parentheses, is determined by performing a dictionary lookup of the translation dictionary, and when generating a translation, the translation of the frame and the embedded translation are converted from the original to the translation. A target translation is generated based on the syntax conversion pattern.

This method does not perform the analysis and generation that is complicated and requires high precision in the conventional machine translation processing, but uses the mutual relation of the syntactic structure and vocabulary and the conversion pattern to group words. For groups, one-to-one mappin
How to do g. In other words, if there is a fixed expression or a specific expression registered in the original sentence, it is detected and embedded based on the information of the word, phrase, clause, sentence to be embedded used in combination with this expression. The rule is fired, the pattern of syntax conversion is determined, the translation frame of the target sentence and the embedded translation are determined by dictionary lookup of the translation dictionary, and the embedded translation is embedded in the translation frame.

Here, if the frame and the embedding word are combined and referred to as an embedding target portion, the target portion includes the part of speech information from the original sentence to the translated word and the word existing in the embedding target as a group of words related to each other. By inheriting as, it is possible to retain information on connected words until the translated sentence is generated, and to generate a translated sentence from the original sentence group as the target sentence group when the translated sentence is generated. Can be.

In a method in which a sentence is divided into words of the minimum unit and rearranged once in a random manner, in the determination of the modification destination of a word, etc., there is a problem in eliminating polysemy of candidate modification destinations. As described above, by processing the part-of-speech information collectively with the grouped words, the combinations of the candidates for the modification destination are also limited, and the ambiguity of the modification destination can be easily eliminated. Therefore, it can be generated at a correct generation position according to the word order of the original sentence from the word order of the target sentence, and it becomes easy to modify from which position to modify, thereby contributing to the generation of a word order suitable for the target language.

For example, when a sentence to be translated is converted from Japanese into English, it is divided into a main part and a predicate, and
The structure is divided into adnominal modifiers and repeated modifiers that modify them. The converted English structure is composed of a phrase structure, and is converted into an adverb phrase, a noun phrase, a verb phrase, an adjective phrase, and the like, and is often converted in units of this word group.

By paying attention to this, by giving the part of speech information in the original sentence and the part of speech information in the target sentence to the embedding target part in units of the embedding target part, the generation position in the target sentence can be determined according to the grammatical information. Of course, the generation position can be specified at the specific position. As a result, it is possible to prevent words to be combined among a plurality of word groups from being nested in different word groups.

As described above, since parsing rules are not based on parsing rules or generating rules of the syntactic structure, there are cases where nouns are used in the original sentence and clauses are generated in the translated sentence, or parts of speech are converted into verbs in the original sentence and adverbs in the target sentence. In the case of heterogeneous conversion or part-of-speech conversion in a syntactic structure, by giving part-of-speech information to a frame translation or an embedded translation, a translation can be generated in accordance with a pattern conversion rule.

Further, paying attention to the difference in structure between the source language and the target language, the skeleton forming the structure is registered as an original sentence frame and a translated sentence frame, and part-of-speech information is given to each unit to be embedded, thereby giving a syntactic structure. It is possible to generate one-to-one correspondence between conversion patterns and generate a translation that matches the actual situation without being bound by general conversion rules and syntax generation rules of the machine translation system.

Further, a specific expression such as an expression used repeatedly or a phrase expression can be registered as a frame, and words, phrases, clauses, and sentences used in combination inside or outside the frame can be registered as an embedded word. Thus, the mutual relationship between the embedding pattern and the vocabulary is held as a connected word group, and can be inherited until the translated sentence is generated.

Further, by forming rules of the embedding pattern and the syntax structure conversion pattern as a conversion pattern from the original sentence to the translated sentence, this rule can be fired and embedded as an embedded translated word inside or outside the translated sentence frame. It is now possible to generate a generalized word group. The frame extracting unit can determine a portion to be embedded in the input text by extracting the registered frame from the input text.

As described above, it has become possible to translate a wording expression, which is difficult to realize with a conventional machine translation system, and a specific expression such as "in this case, such an expression".

Further, when there are a plurality of candidate words, the words to be combined among a plurality of word groups are nested in a different group of words. The probability of occurrence of such a thing is reduced by inheriting the relevance of each other until the translation is generated. Furthermore, even if there is no correspondence between the syntactic structure and the part of speech in the original sentence and the part of speech in the vocabulary are different, one-to-one correspondence should be achieved by firing the rule of the conversion pattern from the original to the translated sentence. Can be.

Further, in the conventional machine translation system, in order to generate a target sentence by applying a general rule, an expression that is not appropriate for each expression is generated, and a sentence with poor sensitivity is generated. In the present embodiment, since the translation is performed while maintaining the connection between the words, the entire sentence becomes a tight sentence,
The specific expression "translate in this case" can be converted one-to-one from the original expression to the translated expression by firing the rules of the embedding rules and conversion patterns stored on the auxiliary storage device. Therefore, it is possible to increase the precision of expression appropriate for the situation, and to reduce the swaying factor in determining the syntactic structure and eliminating the polysemy of dependency.

[0090]

As described above, according to the present invention,
The effect is that the sensitivity of the original sentence can be reflected in the translated sentence, the accuracy of the translated result can be improved, and the completeness of the translated result as a sentence can be further increased.

By converting the structural unit mapping into a pattern, the case of converting from Japanese to English will be described.
It can be translated in units of continuous modifiers and continuous modifiers, and through a process such as morphological analysis like conventional machine translation, the sentence is divided into individual words and the translated words are rearranged. Since there is no translation, the intended result of the original text can be clearly expressed in the translation result.

Further, in the case of a wording expression or the like, in this case, such translation means that the intended expression can be realized with a wording more suitable for reality by utilizing information of a chain of adjacent words. .

[Brief description of the drawings]

FIG. 1 is a basic block diagram of an embedded machine translation system to which the present invention is applied.

FIG. 2 is a diagram illustrating an example sentence including an embedding target sentence and a process of extracting an embedding target portion.

FIG. 3 is a diagram showing a configuration of a frame expression dictionary.

FIG. 4 is a diagram showing an embedding pattern of an original sentence and a translated sentence.

FIG. 5 is a diagram showing a configuration of an embedded translation word dictionary.

FIG. 6 is a diagram showing a configuration of an embedding rule.

FIG. 7 is a diagram showing a configuration of an original sentence → translation sentence conversion table.

FIG. 8 is a diagram showing an example of an original sentence in which an embedded code is embedded.

FIG. 9 is a diagram showing a configuration of an embedded sentence table.

FIG. 10 is a diagram showing an example of a translated sentence being generated.

FIG. 11 is a diagram illustrating an example of a generated embedded sentence.

FIG. 12 is a flowchart illustrating a translation processing procedure of the machine translation apparatus according to the embodiment.

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 input device 2 processing unit 3 output device 4 auxiliary storage device 5 main storage device 6 frame extraction unit 7 code embedding unit 8 translation processing unit 9 embedded sentence analysis unit 10 embedded sentence conversion unit 11 embedded sentence generation unit 12 frame expression dictionary 13 embedding Pattern 14 Embedding translated word dictionary 15 Embedding rules 16 Original sentence to translation translation table 17 Embedded sentence table 18 * input 19 * output

Claims (14)

(57) [Claims] 1. A frame storage means for storing a specific phrase used for a specific expression of an original sentence as an original sentence frame, an attribute of a word or a group of words constituting the specific expression in combination with the original sentence frame and the original sentence frame. Conversion rule storage means for storing, as rules, a translated sentence frame corresponding to the original sentence frame, and a conversion pattern of a syntax structure from the specific expression to a translation expression including the translated sentence frame corresponding to the specific expression. Dictionary means for storing the attributes and translations of the words of the original sentence; input means for inputting the original sentence to be translated; and, from the original sentence input from the input means, referencing the frame storage means to read the original frame. Detecting means for detecting, in the input original sentence, combining with the original sentence frame detected by the detecting means to form a specific expression A search unit for searching for the attribute and translation of the word or word group from the dictionary unit, an original text frame detected by the detection unit, and the attribute of the word or word group searched for by the search unit, A selection unit for selecting a rule to be applied from the conversion rule storage unit; a translation frame corresponding to the detected original text frame based on the rule selected by the selection unit; and a specification including the detected original text frame. Determining means for determining a conversion pattern of a syntactic structure from an expression to a translation expression; and determining the translation frame determined by the determining means at a position based on the conversion pattern determined by the determining means. Generating means for embedding a translated word of a word or group of words and generating a translated expression corresponding to the specific expression in the input original sentence; Determining means for determining whether the sentence includes a part other than the specific expression; and if the determining means determines that the sentence includes a part other than the specific expression, the input original text And a translating means for translating a part corresponding to the specific expression in the translation result by the translating means with the translated expression generated by the generating means. 2. The machine translation apparatus according to claim 1, wherein the word group is a phrase, a clause, or a sentence. 3. The rule storage means includes, as the rule, grammatical information of the translated expression independently of grammatical information of a corresponding specific expression, and the translating means, based on grammatical information of the translated expression, 2. The machine translation apparatus according to claim 1, wherein the input original text is translated. 4. The machine translation apparatus according to claim 3, wherein the grammatical information is part of speech information. 5. The machine translation apparatus according to claim 3, wherein the grammatical information is positional information in a sentence. 6. The machine translation apparatus according to claim 1, wherein the generation unit includes a morpheme generation unit. 7. The translation unit stores the translation expression generated by the generation unit in a memory in association with a specific code, and identifies the specific expression in the input original sentence corresponding to the translation expression. 2. The machine translation apparatus according to claim 1, wherein the translation is performed by substituting and translating with the code, and the substitution unit obtains the translation expression by referring to the memory using the specific code. 8. An inputting step of inputting an original sentence to be translated, and from the original sentence input in the inputting step, refer to a frame storage unit in which a specific phrase used for a specific expression of the original sentence is registered as an original sentence frame. Detecting an original sentence frame, and, in the input original sentence, an attribute and a translated word of a word or a group of words constituting a specific expression in combination with the original sentence frame detected by the detection step are written in the original sentence. A search step of searching from a dictionary in which the attributes and translations of the word are registered, the source text frame and the attributes of the words or word groups that constitute the specific expression in combination with the source text frame are compared with the source text frame. And a translation pattern of a syntax structure from the specific expression to a translation expression including the translation frame corresponding to the specific expression are recorded as rules. A selection step of selecting a rule to be applied from the stored conversion rule storage unit based on the original frame detected in the detection step and the attribute of the word or word group searched in the search step; A determining step of determining a translation frame corresponding to the detected original sentence frame and a conversion pattern of a syntax structure from a specific expression including the detected original sentence frame to a translated expression, based on the rule selected in the step; Embedded in the translation frame determined in the determining step, at a position based on the conversion pattern determined in the determining step, a translated word of the searched word or word group, A generating step of generating a translation expression corresponding to the expression; a determining step of determining whether the input original sentence includes a part other than the specific expression; A step of translating the input original sentence as a group when the specific expression is determined to include a portion other than the specific expression; and a process corresponding to the specific expression in a translation result by the translation step. A replacement step of replacing a part with the translation expression generated by the generation step. 9. The method according to claim 8, wherein the word group is a phrase, a clause, or a sentence. 10. The rule storage unit includes, as the rule, grammatical information of the translated expression independently of grammatical information of a corresponding specific expression. In the translation step, based on grammatical information of the translated expression, 9. The machine translation method according to claim 8, wherein the input original text is translated. 11. The machine translation method according to claim 10, wherein the grammatical information is part of speech information. 12. The machine translation method according to claim 10, wherein the grammar information is position information in a sentence. 13. The machine translation method according to claim 8, wherein the generation step includes a morpheme generation step. 14. The translation step includes: storing a translation expression generated in the generation step in a memory in association with a specific code; and storing a specific expression in the input original sentence corresponding to the translation expression. 9. A code replacement step of replacing with a specific code, wherein the replacement step includes a code search step of obtaining the translated expression by referring to the memory with the specific code. Machine translation method.