JP3015509B2 - Automatic learning device for machine translation - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an automatic learning apparatus for machine translation, which automatically generates data used for polysemy word selection and relational element selection in accordance with a correction operation of a machine-translated translated sentence. A function for learning an appropriate translation is provided for selecting a polysemous word at the time of machine translation for automatically translating a Japanese sentence into an English sentence using a machine translation system.

However, the conventional learning function of polysemy word selection allows a user to indicate the position of an erroneous word in a translated English sentence and a correct translation word, or to add data of semantic relations directly used for polysemy word selection. Need, the burden on the user is large,
Improvement in this respect is desired.

[0003]

2. Description of the Related Art A conventional machine translation system is provided with a function of learning an appropriate translation of an ambiguous word. When learning an ambiguous word, a user explicitly specifies a position of an erroneous word and a correct translated word. There is a need. For example,

[0004]

(Equation 1)

[0005] In order to make the translation system learn the translation of "cut" (polysemy word selection), the "cut"
Must be specified to indicate that the correct translation is “shuffle”. When such polysemy learning is performed according to the user's instruction, semantic relation data for polysemy word selection is generated by structural analysis of the original sentence and registered in the semantic relation dictionary. Therefore, by translating after learning, by referring to the contents of the created semantic relation dictionary,

[0006]

(Equation 2)

Thus, an appropriate polysemy can be selected.

[0008]

However, in the polysemy word learning function of such a conventional machine translation system, a user mistakenly selects a translated sentence from a machine translation in a learning mode setting state. There is a problem that a learning operation of finding a polysemous word and instructing a correct translation word must be repeated, and a user burden required for learning is large.

[0009] If there is no appropriate translation to be learned, the user must directly add semantic relation data for polysemous word selection, and there is still a problem that the user burden is large. A similar problem also occurs in learning for generating the relationship knowledge data.

The present invention has been made in view of such conventional problems, and in a machine translation system, paying attention to the fact that a user always performs an operation of correctly correcting an incorrect translation, It is an object of the present invention to provide an automatic learning apparatus for machine translation, which automatically generates semantic knowledge data used for selecting a polysemy word and selecting a relation in association with a correction work, thereby greatly reducing a user's burden.

[0011]

FIG. 1 is a diagram illustrating the principle of the present invention. First, the present invention is directed to a machine translation system that translates an original sentence such as a Japanese sentence and outputs a translated sentence, for example, a translated English sentence. In the present invention, as an automatic learning device for machine translation intended for such a machine translation system, the present invention analyzes the structure of the original sentence to be corrected when performing a correction operation on the translated sentence machine translated. An original sentence structure analysis unit 1 for generating a conceptual structure, a translated sentence structure analysis unit 2 for analyzing a corrected correct translated sentence structure of the original sentence and generating a conceptual structure, and analysis by an original sentence structure analysis unit 1 and a translated sentence structure analysis unit 2 The result of the comparison is compared, the semantic relation data is generated based on the difference in the conceptual structure between the original sentence and the translated sentence, and the semantic relation dictionary 4 used for machine translation.
And a structure comparing unit 3 for registering the information.

For example, in the learning for polysemy word selection, the structure comparing unit 3 is based on the difference between the conceptual structure of the original sentence analyzed by the original sentence structure analyzing unit 1 and the modified translated sentence analyzed by the translated sentence structure analyzing unit 2. To generate semantic relation data used for selecting polysemous words during machine translation. Taking an example of polysemy selection when translating a Japanese sentence into an English sentence, the structure comparing unit 3 is a concept of the Japanese sentence analyzed by the original sentence structure analyzing unit 1 and the concept of the corrected translated English sentence analyzed by the translated sentence structure analyzing unit 2. Generates semantic relation data used for selecting polysemous words during machine translation based on the difference in structure.

In the learning of the relationship, the structure comparison unit 3 determines the difference between the correlator of the original sentence analyzed by the original sentence structure analysis unit 1 and the correlator of the corrected translated sentence analyzed by the translated sentence structure analysis unit 2. And generates semantic relationship data of the relationship based on the relationship. For example, taking a translation from a Japanese sentence to an English sentence as an example, the structure comparison unit 3 generates semantic relationship data of a relationship based on the difference between the analyzed Japanese sentence and the corrected translated English sentence.

[0014]

According to the automatic learning apparatus for machine translation of the present invention having such a configuration, when a user performs an operation of correcting a translated sentence that has been machine-translated, a structure of a corrected correct translated sentence and an original sentence are respectively set. Analysis is performed, and by comparing structural concepts obtained as a result of analysis, semantic relation knowledge (semantic relation data) of polysemy word selection or dependency relation selection is generated from the differences determined to be usable and registered in the semantic relation dictionary. Is automatically performed.

By automatically generating the semantic relation data based on the bidirectional analysis of the original sentence and the correct translation accompanying the operation of correcting the translated sentence, the user can select polysemous words and change relations without being conscious of the learning operation at all. Can be registered by learning the semantic relation data that determines the language, and the learning is performed through daily correction operations. Therefore, the more the machine translation system is used, the more the ability can be improved.

[0016]

FIG. 2 is a block diagram showing an embodiment of the present invention. In FIG. 2, reference numeral 10 denotes a source language morphological analysis unit, and reference numeral 12 denotes a source language syntax analysis unit, which is provided to analyze the conceptual structure of the source language as an original sentence. On the other hand, 14 is a target language morphological analysis unit, and 16 is a target language syntax analysis unit, which is provided to analyze the concept structure of the target language as a corrected correct translation.

The input to the source language morphological analysis unit 10 and the target language morphological analysis unit 14 is a set of "source text" and "correct corrected text" when the translation of the machine translated by the machine translation system is performed. Is entered as Of course, this "original text-corrected translation"
May not be input in real time, but may be stored in a memory for storing learning data, and read and learned by designating a learning mode after the translation correction operation is completed.

In the morphological analysis of each language by the source language morphological analysis unit 10 and the target language morphological analysis unit 14, words constituting a sentence are defined as one morpheme, and each morpheme is converted into an intermediate language having a data format shown in FIG. Convert. The data format of the intermediate language in FIG. 3 is composed of notations such as “cut” and “cut” and concept symbols, and further provided with attribute information as needed.

Here, the concept symbol is a language in which the meaning can be expressed in any language in common. For example, if "cut" in Japanese and "cut" in English have the same concept, the same concept symbol (code ) Is assigned. The following structural analysis is all processed using an intermediate language using concept symbols. The source language syntax analysis unit 12 and the target language syntax analysis unit 16 perform syntax analysis using the morphemes obtained from the source language morphological analysis unit 10 and the target language morphological analysis unit 14, and construct the conceptual structure of the source language and the target language. Convert to

FIG. 4 shows the conceptual structure of the source language and the target language of the original Japanese sentence "I cut cards in the classroom.""I cut cards in the classroom." It shows the result of parsing when this is corrected to the correct translation "I shuffle cards in the classroom." Of course, each is expressed in an intermediate language (concept symbol) corresponding to each language.

The original side (source language side) shown in FIG. 4 has the same conceptual structure in both the syntax analysis of the Japanese sentence and the syntax analysis of the original sentence before correction, and thus shows the translated original sentence. The target side (target language side) indicates the conceptual structure of the corrected translated sentence.
This conceptual structure is composed of the verbs “cut”, “shuffle”
Is a from node, and other morphemes are to nodes, and a relational element indicating a semantic relationship is shown between the nodes.

Referring again to FIG. 2, the analysis results of the source language syntax analysis section 12 and the target language syntax analysis section 16 are sent to the conceptual structure difference analysis section 18 where, for example, the conceptual structure of the original text shown in FIG. The concept structure of the already translated sentence is compared, and if there is a difference, it is notified to the next semantic relationship knowledge forming unit 20. In the case of FIG. 4, “cut” and “shuf” configuring the from node
fle "is different.

The semantic relation knowledge constructing unit 20 analyzes the difference between the conceptual structures of the source language and the target language and extracts semantic relation knowledge. If there is any extracted knowledge, it constructs it in a dictionary format. At 22, it is registered in the semantic relation dictionary.
Further, a registration information inquiry unit 24 is provided, and the semantic relation data registered by the semantic relation dictionary registration unit 22 is shown to the user, and if there is incorrect data, the user can point out the data and delete it.

The analysis process performed by the semantic relationship knowledge forming unit 20 will be described below with reference to the flowchart of FIG. The process of FIG. 5 is based on the general form of the conceptual structure of the original side and the target side shown in FIG. In FIG. 5, first, in step S1, the focus is set on the node (A) of the target verb.
Next, in step S2, the focal point is set to an arbitrary essential point, for example, a node (B).

Subsequently, in step S3, the same node (C) as the node (B) that has become the focus on the target side on the original side is searched for. In this case, since there is the same node, the process proceeds from step S4 to step S5, where the verb V1 of the target side from node (A) and the verb V2 of the target side from node (D) are different, and the relations α and β Check if are the same. That is,

[0026]

(Equation 3)

It is checked whether or not the following holds. In the case of FIG. 6, since this condition is satisfied, the process proceeds to step S6,
Focus on the other prerequisite node (E). Subsequently, in step S7, it is checked whether or not there is any other required node. In this case, since there is a node (E), step S3 is performed.
To find the node (F) on the same original side as the node (E), and through steps S4 and S5, to step S6.
To step S7 since there is no remaining node at this time.
Then, the process proceeds to step S10 to generate one set of semantic relation data.

Further, in step S8, another verb node is searched for, but there is no case in FIG. 6, so that a series of processing ends through step S9. FIG. 7 is an explanatory diagram of semantic relation data for selecting a polysemy word created through the processing of FIG. 6 with respect to the conceptual structure of FIG. 4. For the target language, the from node is set to “SHUFFFLE” and the to nodes “I” and “CAR” are used.
D "and" CLASSROOM "are provided.
Relations <agent>, <o between m node and to node
bj> and <place> are registered.

Next, the processing of selecting a polysemy in a machine translation system using the semantic relation data created by the present invention will be briefly described as follows. FIG. 8 shows the contents of the Japanese dictionary and the semantic relation dictionary of the machine translation system related to Japanese "cut". First, in the Japanese dictionary, English “cut”, “shuffle”, and “turn” are stored in the data format shown in FIG. 3 corresponding to the Japanese name notation “cut”.
Concept symbols having the same meaning as are registered, and priorities are set.

With this Japanese dictionary alone, polysemy selection based on priority is performed, so "cut" is translated into "cut". On the other hand, in the semantic relation dictionary, the from node is set to “s
The meaning relation data having “card” in the to node as “huffle” is registered together with the relational element <obj>. Regarding “turn” of the from node, semantic relation data having “steering wheel” is registered in the to node together with the relational element <obj>.

FIG. 9 shows polysemy selection in the case where "paper cut" is translated using the Japanese dictionary and semantic relation dictionary of FIG. First, the original text is analyzed for structure, and the target to be cut is “pap
er ", and referring to the semantic relation dictionary,
There is no relevant meaning. In this case, “cut”, which has the highest priority among the polysemy words of “cut” in the Japanese dictionary,
Is selected.

FIG. 10 shows the polysemy selection when "card cut" is translated using the Japanese dictionary and semantic relation dictionary of FIG. First, we analyze the structure of the original sentence and find out that the "cut" target is "card".

[0033]

(Equation 4)

Because there is, "shuffle" is chosen as the correct translation. In FIG. 9 and FIG. 10, <obj> is used as a criterion for polysemy word selection, that is,
Is used as an example, but in addition to the above, relational elements that serve as translation criteria include <inst>, <goal>, and <manne.
r> and <role> etc., and <subj> and <pl>
ace> is rarely used as a reference. Therefore, the semantic relation knowledge generated by the machine translation system according to the present invention for translation is <obj>, <inst>, <goal>, <
manager> and <role>.

Next, a description will be given of a learning process of the semantic relationship data of the relationship according to the present invention. Figure 11 is the original text "Manufacturing costs"
Here, the parsing result for obtaining the knowledge of the semantic relation of the relation is shown for the set of the correct translation "Cost for manufacture". In the parsing, the original sentence is expressed in COS because it is expressed by the intermediate word of the same conceptual symbol as the translated sentence.
In this syntax analysis expressed as T, MANUFACTURE, the relational element indicating the relational relation between "COST" and "MANUFACTURE" in the original text "Manufacturing cost" is <MOD> indicating the modification relation, while the relation in the correct translated sentence. The child is <PURPOSE> indicating the purpose relationship.

In this case, semantic relation data of the relation shown in FIG. 12 is generated. In FIG.
"COST" is stored in the node, "MANUFACTURE" is stored in the to node, and a relational element <PURPO obtained by parsing the correct translation is stored in the relational element.
SE> is stored. In other words, when the concept structure obtained by parsing the original sentence Japanese and the concept structure obtained by parsing the correct translated sentence English are different from each other, the relation between the tubes of the from node and the to node at the same position is different. In, if the correct answer is replaced by a relational element obtained by the English analysis, the possibility that a correct translation will be obtained increases.

FIG. 13 is a flowchart showing a process of generating the semantic relation data of the relation by the conceptual structure difference analyzing section 18 and the semantic relation knowledge forming section 20 shown in FIG. In FIG. 13, first, in step S1, an arbitrary relational element on the target side (correct answer translation side) is set to α. Next, in step S2, the from node of the relation α is F, and the to node is T. Further, in step S3, on the original side (original side),

[0038]

(Equation 5)

Search for the structure If it is determined in step S4 that this structure is on the original side, the process proceeds to step S5 to check whether or not the relational elements α and β on the original target side are equal. If the two are different, the process proceeds to step S6 to create semantic relationship data. When the creation of the semantic knowledge data of one relationship is completed, it is checked in step S7 whether there is a next relationship. If there is another relationship, the process returns to step S2 as α in step S8 and repeats the same processing. .
When all the relations have been processed, this is determined in step S7, and a series of processing ends.

In the above-described embodiment, the translation from Japanese to English is taken as an example. However, the reverse is also possible, and the translation between any languages can be applied as it is.

[0041]

As described above, according to the present invention, when a user corrects a machine-translated translated sentence, bidirectional structural analysis is performed on a set of the original sentence and the corrected correct translated sentence, and a conceptual structure is obtained. The semantic relation data of polysemy word selection and the semantic relation data of the relation are automatically generated and registered in the semantic relation dictionary based on the difference between the words, and automatic learning can be performed without making the user aware of the learning operation. The burden can be greatly reduced.

[Brief description of the drawings]

FIG. 1 is a diagram illustrating the principle of the present invention.

FIG. 2 is a configuration diagram of an embodiment of the present invention.

FIG. 3 is an explanatory diagram of a data format of a concept symbol used as an intermediate language in the present invention.

FIG. 4 is an explanatory diagram showing a result of parsing an original sentence and a corrected translated sentence used for generating semantic relation data for polysemy selection according to the present invention;

FIG. 5 is a flowchart showing a concept structure difference analysis and a semantic relationship knowledge generation process performed for polysemy word selection in the embodiment of FIG. 2;

6 is an explanatory diagram showing a general form of a conceptual structure of an original side and a target side to be processed in FIG. 5;

FIG. 7 is an explanatory diagram of semantic relation data generated based on a difference between the syntax analysis results of FIG. 4;

FIG. 8 is an explanatory diagram showing an example of the contents of a Japanese dictionary used in a machine translation system and a semantic dictionary generated in the present invention.

FIG. 9 is an explanatory diagram showing polysemy selection in the translation of “cut paper”.

FIG. 10 is an explanatory diagram showing polysemy selection in the translation of “cutting cards”

FIG. 11 is an explanatory diagram showing an example of a conceptual structure obtained by syntactic analysis of an original sentence and a correct translation used for automatic learning of a relation according to the present invention.

FIG. 12 is an explanatory diagram of semantic relation data of a relation created from differences in relations in the conceptual structure of FIG. 11;

FIG. 13 is a flowchart showing a difference analysis of a conceptual structure and a process of generating semantic relationship knowledge performed on the relationship in the embodiment of FIG. 2;

[Explanation of symbols]

 1: Source sentence structure analysis unit 2: Translated sentence structure analysis unit 3: Structure comparison unit 4: Semantic relation dictionary 10: Source language morphological analysis unit 12: Source language syntax analysis unit 14: Target language morphological analysis unit 16: Target language syntax analysis unit 18: Concept structure difference analysis unit 20: Semantic relation knowledge composition unit 22: Semantic relation dictionary registration unit 24: Registration information inquiry unit

────────────────────────────────────────────────── ─── Continued on the front page (58) Fields surveyed (Int. Cl. ⁷ , DB name) G06F 17/27 G06F 17/28 INSPEC (DIALOG) JICST file (JOIS)

Claims (4)

(57) [Claims] In a machine translation system for machine-translating an original sentence and outputting a translated sentence, when a correction operation of a machine-translated translated sentence is performed, the structure of the original sentence to be modified is analyzed to construct a conceptual structure. , A translated sentence analysis unit 2 that analyzes the structure of a corrected and correct translated sentence for the original sentence to generate a conceptual structure, and an analysis performed by the original sentence structure analysis unit 1 and the translated sentence structure analysis unit 2. And a structure comparison unit 3 that compares the results obtained, generates semantic relation data based on the difference in the conceptual structure between the original sentence and the translated sentence, and registers the data in a semantic relation dictionary 4 used for machine translation. Automatic learning device for machine translation. 2. An automatic learning apparatus for machine translation according to claim 1, wherein said structure comparing section is used for selecting polysemous words at the time of machine translation based on a difference in conceptual structure between said original sentence and a corrected translation. An automatic learning device for machine translation, which generates semantic relation data. 3. The automatic learning apparatus for machine translation according to claim 1, wherein said structure comparing section 3 generates semantic relation data of a relation based on a difference between a relational element of said original sentence and a relational element of a corrected translation. An automatic learning device for machine translation characterized by generating. 4. An automatic learning apparatus for machine translation according to claim 3, wherein said structure comparing section is analyzed by said translated sentence structure analyzing section and a relational element of the original sentence analyzed by said original sentence structure analyzing section. An automatic learning device for machine translation, wherein the automatic translation device generates semantic relationship data of a relationship based on a difference between the modified translation and a relative.