JPS6344276A - Automatic generator for generated syntax - Google Patents

Abstract

PURPOSE:To prevent the level of syntax from dropping and to cope with complicated syntax by correctly generating generated syntax by a machine. CONSTITUTION:A natural language sentence analyzer 2 and a natural language sentence generator 3 translate an original text inputted to a reader 1 for a parallel translation sentence as a parallel translation sentence. A natural language sentence comparator 4 compares the translation with a translation given to the reader 1. If they disagree with each other, the comparator 4 decides that there are errors in the generated syntax or its shortage. A generated syntax error recognizer 5 recognizes the syntax deriving the errors, and a generated syntax correcting device 6 corrects the generated syntax. A generated syntax inspecting device 7 inspects whether the corrected generated syntax correctly functions with respect to translations on a right answer file 9, thereby automatically generating the generated syntax.

Description

[Detailed Description of the Invention] [Summary] In order to solve the problem of having to develop a generative grammar for machines in the field of natural language processing, we have developed a method that automatically identifies errors in generative grammars and automatically improves the grammar. By modifying it, humans no longer need to develop a generative grammar for machines.

[Industrial application field]

The present invention relates to a generative grammar generating device for generating natural language sentences in natural language processing.

Natural language sentence generation in natural language processing is used to generate bilingual sentences in machine translation or to output knowledge from a database, and requires accurate sentence generation. Generative grammars tend to become larger and more complex. Therefore, there is a need for a development method that can easily create such generative grammars with high accuracy.

[Prior art and problems]

FIG. 4 is a diagram showing an outline of the translation method. As shown in Figure 4, a Japanese sentence is analyzed and a conceptual structure is created.

The conceptual structure has a tree or network structure. In the process of generating a target language (English in this case) from a conceptual structure, words are extracted from the word dictionary of the target language based on the conceptual symbol expressing the concept, and the node where the arc of the entry point indicated in the conceptual structure enters is extracted. Processing starts from , and that node determines the syntax while looking at the surrounding arcs and nodes ahead of it, as well as the grammatical attributes of the selected word, and performs morphological processing. These rules are called generative grammars.

FIG. 5 is a diagram showing an example of translation. Suppose that a Japanese sentence such as ``He gave her a book'' is manually input into a Japanese-English machine translation device. When a Japanese sentence such as ``He gave her a book'' is input, the translation device generates a conceptual structure representing the meaning of the human-powered Japanese sentence. In this conceptual structure, the attribute "partner" is added to the arc from "give" to "her," the attribute "doer" is added to the arc from "give" to "him," and the attribute "doer" is added to the arc from "give" to "he." The arc heading toward the "book" has an attribute called an object attached to it.

If processing is performed from ``give'', the English word for ``give'' is retrieved from the word dictionary. The English word "give", which corresponds to "give", has a construction called Dl and a construction called T1. In the syntax Dl, the word that indicates the actor (in this case, he) comes first, then the person himself (in this case, give), and then the word that indicates the other party (in this case, the purpose of she). case), and finally the word representing the object (in this case, book).

In the construction T1, the word representing the actor comes first, then the word itself, then the word representing the object, then to, and finally the word representing the other party. It says it will come.

FIG. 6 is a diagram illustrating a conventional method for developing a generative grammar. As shown in FIG. 6, the generative grammar is modified by humans through an editor.

[Problem to be solved]

In the conventional generation-grammar development method, humans look at the output sentences to determine whether they are incorrect, search for incorrect grammar, and correct it, which takes time and effort to find incorrect grammar, and requires correction. Even if the resulting grammar fits the sentence well, it is unclear whether it will be able to correctly generate sentences that were previously generated correctly.

The present invention was created in view of this point, and an object of the present invention is to provide an automatic generative grammar generation device that can correctly generate generative grammars by machine.

Means for Solving Problem C] FIG. 1 is a diagram showing the principle of an automatic generative grammar generation device of the present invention. In Figure 1, 1 is a bilingual text reading device, 2 is a natural language sentence analysis device, 3 is a natural language sentence generation device, 4 is a natural language sentence comparison device, 5 is a generative grammar error recognition device, and 6 is a generative grammar correction device. 7 is a generative grammar checking device, 8 is a generative grammar, and 9 is a correct answer file.

In the present invention, an original text input as a bilingual text into a bilingual text reading device 1 is translated by a natural language sentence analysis device 2 and a natural language sentence generation device 3. Then, the translated sentence and the translated sentence given to the bilingual sentence reading device 1 are compared by the natural language sentence comparison device 4, and if the results are different, it is determined that there is an error or deficiency in the generative grammar, and this error is corrected. The derived grammar is certified by the generative grammar error recognition device 5, and the generative grammar correcting device 6 corrects the generative grammar. Then, the generative grammar checking device 7 checks whether the modified generative grammar functions correctly for the sentences in the correct answer file 9 so far, thereby automatically creating a generative grammar. Note that the generative grammar checker 7
The line from to the natural language sentence analysis device 2 shows the flow of data, and the sentences inspected by the generative grammar inspection device 7 are re-executed from the analysis as necessary.

〔Example〕

FIG. 2 is a diagram showing the generation process of an Inuit language generation grammar according to the present invention, and FIG. 3 is a diagram showing an Inuit language dictionary.

In Figure 2, <agent> indicates the actor, and ■
indicates a verb, and <object) indicates an object. Furthermore, <st> indicates the central concept of the sentence. Furthermore, ■, ■, . . . represent the order of generation processing.

First, the English sentence r11ikeph is loaded into the bilingual sentence reading device 1.
ysical education J and Inuit language rIqaijarumaqattaqtur+8a
J is input, the semantic structure of English is analyzed by the natural language analysis device 2, and based on the semantic structure and the Inuit dictionary, 1ike is rumaq by the natural language sentence generation device 3.
attaq, l is nga, physical edu
Using the fact that cation corresponds to 1qaija as a clue, ngarumaqattaqiqaija is generated and output.

By comparing this generated sentence with the Inuit sentence read by the bilingual sentence reading device 1 using the natural language sentence comparison device 4, it is recognized that this generated sentence is different from the correct answer. Next, the grammar error recognition device 5 examines the grammar rules used by the natural language sentence generation device 3 to determine that the application of the generative grammar <agent>V <object> is incorrect. . The generative grammar correction device 6 converts the grammar into <object>Vtu
It is recognized that it is sufficient to modify the grammar to <agent>, and generative grammar 8 is modified. After correcting this grammar, the original sentence is given again to the natural language sentence analysis device 2 to confirm that a correct translation can be obtained through retranslation, and the bilingual sentence is registered in the correct answer file 9. Next rl don't h
ave a wife J and rNuliaqanngi
If you give ttunga J, perform English analysis, and generate an Inuit sentence, it becomes rNuliaqatungaj, and the grammar is changed to <object>V [<
Correct it as not>kotu<agent>. Then this sentence will be generated correctly and also the previous rlqaija
Confirm that it is generated correctly for rumaqattaqtunga J as well. In this way, an Inuit language generative grammar is created by providing bilingual words and repeating translation and grammar correction.

〔Effect of the invention〕

As is clear from the above explanation, according to the present invention, there is no need for humans to generate generative grammar, and the generation of sentences that have been correctly generated up to now is guaranteed, thus preventing the level of grammar from decreasing. , can also handle complex grammar.

[Brief explanation of drawings]

Figure 1 is a diagram explaining the present invention in detail, Figure 2 is a diagram showing a generation example of an Inuit language production grammar, Figure 3 is a diagram showing an example of an Inuit language dictionary, and Figure 4 is an overview of the translation method. FIG. 5 is a diagram showing an example of translation, and FIG. 6 is a diagram showing a conventional method for developing generative grammar. DESCRIPTION OF SYMBOLS 1... Bilingual text reading device, 2... Natural language sentence analysis device, 3... Natural language sentence generation device, 4... Natural language sentence comparison device, 5... Generative grammar error recognition device, 6 ...Generative grammar correction device, 7...Generative grammar check device, 8...
Generative grammar, 9... Correct answer file.

Claims (1)

[Claims] A bilingual text reading device (1), a natural language sentence analyzing device (2) that analyzes the original text of the text read by the bilingual text reading device (1), and this natural language sentence analyzing device A natural language sentence generation device (3) that generates a bilingual sentence from the results analyzed by (2), and a natural language sentence obtained by this natural language sentence generation device (3) and the bilingual sentence reading device (1). A natural language sentence comparison device (4) that compares the translated text of the read parallel text.
), and the natural language sentence comparison device (4) converts the sentence obtained by the natural language sentence generation device (3) into the bilingual sentence reading device (
When it is determined that the sentence is different from the translated text read by 1), the generative grammar error recognition device (3) examines the process by which the sentence was generated by the natural language sentence generation device (3).
5), a generative grammar correction device (6) that automatically corrects or deletes incorrect grammar or adds grammar, and a generative grammar correction device (6) that automatically corrects or deletes incorrect grammar, and that corrects the corrected grammar without adversely affecting sentences that were previously generated correctly. 1. A generative grammar automatic generation device comprising: a generative grammar checking device (7) for checking whether a generative grammar can be generated.