KR100641053B1 - Apparatus and method for restoration of ellipsis statement constituent - Google Patents

Abstract

A device and a method for restoring an omitted component of a sentence are provided to prevent an error caused from an ellipsis of the sentence component, offer correct sentence structure analysis information, and recognize/restore the omitted component of the Hangul sentence by properly using rule and statistics information. A sentence structure(10) analyzer analyzes a structure of the inputted sentence based on predefined grammar. An ellipsis candidate recognizer(20) detects restoration candidates of the ellipsis in the inputted sentence if the ellipsis of the sentence is determined by checking a necessary component of each inflected word appeared in the analyzed sentence. An ellipsis restorer(30) restores the ellipsis in the detected restoration candidates by using the predefined rule/statistics information(32).

Description

Apparatus and method for restoration of ellipsis statement constituent}

1 is a view showing a sentence component restoration apparatus according to the present invention

2 is a flowchart illustrating a sentence component restoration method according to the present invention.

3 is a flowchart illustrating a sentence structure analysis method in FIG. 2 in detail;

4 is a flowchart illustrating a method of setting an omitted component candidate in FIG. 2;

5 is a flowchart illustrating a method of restoring a skipped component in FIG. 2;

* Explanation of symbols for main parts of the drawings

10: sentence structure analysis unit 20: omitted candidate recognition unit

22: child / transitive verb dictionary 30: skipped component restoring unit

32: Rule / statistics information

The present invention relates to an apparatus and method for restoring omitted sentence components, and more particularly, to an apparatus and method for restoring sentence components omitted from Korean using rules and statistical information.

In general, ellipsis that occurs in natural language processing mainly occurs to avoid repetition, and thus the omission may be restored by inferring from the same sentence or a previous sentence.

In the case of English, the abbreviation is a form in which the verb phrase is omitted as follows.

"Helen saw the movie and Mary did too."

However, in the case of Korean, a verb phrase is rarely omitted, and a noun phrase mainly serving as a subject or object is omitted. That is, there are many omissions of subjects in the same sentence and headings in special documents such as encyclopedias.

[Example 1]

If the subject in the same sentence is omitted,

"Study Prasapadapah wrote <Old Law Enforcement (句 義 法 句 義 要)> ( Study Prasapadapaga) established the theory of this wave."

As a result of this phenomenon, "Scholastic Prasapadaga" is omitted in the actual sentence.

This phenomenon also occurs in the question-and-answer system that presents only correct answers to questions.

In the Q & A system, various techniques are used. One of them is the method of using the predicate-argument relationship of question and answer sentences as follows.

[Example Example 2]

If a heading is omitted in a special document such as an encyclopedia,

“Who received the Nobel Peace Prize in 2000?”

(Subj: person, obj: Nobel Peace Prize, adj: 2000)

Title: Kim Dae Jung

“… For the 2000 Nobel Peace Prize. ”

(Subj: Kim Dae-jung, obj: Nobel Peace Prize, adv: 2000, Merit)-①

(Subj: NULL, obj: Nobel Peace Prize, adj: 2000, merit)-②

In this case, the Korean encyclopedia document with frequent omissions is composed of the omitted sentences (②) with all of the headings in the actual sentences (①) omitted, so the predicate argument relations are formed as they are. You will not be able to find the correct answer.

The restoration of the omitted sentence component is limited to the form of the sentence that can be restored because it mainly attempts to restore the omitted component by applying only man-made rules. In addition, there are cases in which statistical information is used in other ways, which does not utilize the advantages of the rule, but uses only statistics, which causes a difference in performance.

Accordingly, the present invention has been made to solve the above problems, a sentence component restoration apparatus that prevents errors that may occur due to the omission of sentence components frequently occurring in Korean and provides more accurate sentence structure analysis information and The purpose is to provide a method.

Another object of the present invention is to provide a sentence component restoration apparatus and method for recognizing and restoring sentence components omitted from Korean sentences using appropriate rules and statistical information.

A feature of the sentence component restoring apparatus according to the present invention for achieving the above object is a sentence structure analysis unit for analyzing the structure of the input sentence based on a predefined grammar, and each of the words appearing in the analyzed sentence If it is determined that the sentence component is omitted by determining the essential components of each of the input sentence, the skip candidate recognition unit for detecting the reconstruction candidate of the omitted component in the input sentence, and the rule information and statistical information predefined in the detected reconstruction candidate And an omitted component restoring unit for restoring the omitted components.

Preferably, if the essential ingredient for each verb is identified as the target word, it is set as an omission candidate of the verb only when it is determined to be a transitive verb using the child / other verb dictionary of the verb.

Preferably, the structure of the sentence is a connection component representing a sentence component consisting of any one of a subject, a bore, a kind of a noun clause and an adjective clause, and a connection relationship between a sentence and a sentence, a clause and a clause, or a sentence and a clause. It is characterized in that the configuration.

Preferably, the candidate used in the omitted candidate recognizer is at least one of sentence components.

A feature of the sentence component restoring method according to the present invention for achieving the above object is a first step of analyzing the structure of the sentence, if the sentence including the omission of the candidate is input, and accepts the structure-analyzed sentence A second step of generating reconstruction candidates for each word appearing in a sentence among candidates of the omitted element by setting a required component for each element; and regular battle information among the omitted element candidates for each word generated by the skip candidate setting And a third step of completing the sentence in which the omitted component is restored by determining the actual restoration candidate using the statistical information.

Preferably, in the first step, when a sentence including an omitted sentence component is input, sequentially performing at least one or more of a part-of-speech tagging, entity name recognition, and vocabulary tagging process of the input sentence; And analyzing the structure of the sentence having a sentence component, a kind of a clause, and a linking component, by contrasting a grammar commonly defined in a sentence that has undergone one or more processes.

Preferably, in the second step, when the structure-analyzed sentence is input, identifying an essential component for each term appearing in the sentence through the inputted structured sentence, and the same sentence among the identified essential elements for each term Setting all the subjects of a verb appearing within as candidates for reconstruction of a subject of a verb that does not exist, and all the object words of the verb appearing in the same sentence among the identified essential elements of the verb, wherein the object does not exist And setting the candidates for reconstruction of the object of the verb.

Preferably, the candidate for restoration of the object to be set may be set as the object candidate of the corresponding word only when the corresponding word is determined to be the other verb using a pre-stored child / verb dictionary.

Preferably, the third step includes determining whether the omitted component candidate is a subject or a target word when the generated omitted element candidate for each term is input, and if the candidate is determined to be a subject, a regularized meaning of the omitted component candidate. Restoring the candidate to the omitted sentence as a subject if the sentence component having semantic information using the information is a candidate element of a specific term, and if the candidate is determined to be the subject, If it is determined that it is an omitted component candidate that is not determined to be restored or if the candidate is determined to be an object, performing matching to a predefined regular frame, and detecting the matching frame as a result of the matching; Restoring when a frame other than the omitted subject or object is detected through the statistical information defined in the restored sentence; And re-validating the sentence components already selected for restoration by using numerical statistics calculated by applying the beam.

Preferably, the statistical information is characterized by a numerical value representing a score between 0 and 1, respectively, when the omitted sentence is not restored, when it is restored to the subject, and when it is restored to the object as numerical statistics. .

Preferably, the verifying may include determining whether the restoration through sentence structure analysis and the restoration through the frame matching are the same as the verification through the statistical information. Maintaining the restored sentence as it is, and if it is not the same as a result of the determination, restoring the restored candidate by the restoration through the sentence structure analysis and the matching by a candidate defined in the statistical information It is characterized by.

Preferably, the verifying step includes a boundary of a value calculated by applying to the statistical information 1) when not restored, 2) when restored to a jurisdiction, and 3) when restored to a destination for verification through the statistical information. a user's predefined predefined threshold value, and if the restored candidate is a candidate whose numerical statistics calculated by the application of the statistical information do not exceed the threshold, the candidate recommended by the statistical information Selecting and restoring the final candidate to be restored; and if the restored candidate does not have a numerical value calculated by applying the statistical information below the threshold value, maintaining the previously restored candidate as it is. It is characterized by comprising.

Preferably, the statistical information is characterized by using any one or more of a lexical feature, a POS feature, a semantic feature, a structural analysis feature, and a composite feature.

Other objects, features and advantages of the present invention will become apparent from the following detailed description of embodiments with reference to the accompanying drawings.

A preferred embodiment of the sentence component restoring apparatus and method according to the present invention will be described with reference to the accompanying drawings.

1 is a diagram showing a sentence component restoration apparatus according to the present invention.

As shown in Figure 1, sentence structure analysis unit 10 for analyzing the structure of the input sentence on the basis of a predefined grammar, and by identifying the essential components of each term appearing in the analyzed sentence omission of sentence components If it is determined, by using the omitted candidate recognition unit 20 for detecting the reconstruction candidate of the omitted element in the input sentence, the rule information 31 and the statistical information 32 predefined in the detected reconstruction candidates It is configured to include an omitted component restorer 30 for restoring the omitted components.

In this case, the sentence structure analysis unit 10 analyzes the structure of the sentence through at least one process of morphological analysis, entity name recognition, vocabulary meaning tagging.

In addition, the structure of the sentence is composed of a sentence component such as a subject, an object, a bore, etc., a kind of a clause such as a noun clause, an adjective clause, and the like, and a connection component representing a connection relationship between a sentence and a sentence, a clause and a clause, or a sentence and a clause. In addition, the candidate used in the abbreviation candidate recognizing unit 20 refers to at least one of sentence elements such as a subject, an object, a bore, and the like that are omitted.

The abbreviation candidate recognition unit 20 is given the essential components for each verb, the target word, and if the essential component for each verb is identified as the target language, omitting the verb only if it is determined to be a transitive verb using the child / other verb dictionary of the verb Set as a candidate.

The method using the sentence component restoration apparatus according to the present invention configured as described above will be described in detail with reference to the accompanying drawings.

2 is a flowchart illustrating a sentence component restoration method according to the present invention. 3 is a flowchart illustrating a method of analyzing a sentence structure in FIG. 2, FIG. 4 is a flowchart illustrating a method of setting an omitted component candidate in FIG. 2, and FIG. 5 is a flowchart illustrating a method of restoring an omitted component in FIG. 2.

As shown in FIG. 2, when the sentence 11 including the omission of the candidate is input as the first step, the structure of the sentence is analyzed (S100).

Referring to FIG. 3, when a user inputs a sentence including an omitted sentence component (S110), the part-of-speech tagging (S120), the entity name recognition (S130), and the lexical meaning tagging ( S140) is performed sequentially.

Then, the structure of the sentence having a sentence component, a kind of a clause, and a linking component is analyzed while comparing the grammar that is normally defined in the sentence which has undergone the above processes (S150).

The sentence component indicates a subject, an object, a bore, and the like, and the kind of a clause indicates a noun clause, an adjective clause, and the like. In addition, the connection component indicates a connection relationship between a sentence and a sentence, a clause and a clause, or a sentence and a clause.

An embodiment of the input and the result obtained by performing the sentence structure analysis is as follows.

 Example 1

"Study Prassa Stafada wrote <the law of the old law" and established the theory of this wave. "

 Establish

<obj: This par theory:

Write

<subj: Prastapada: A>

<obj: <Compulsory Law (를 法 綱要):

In the next second step, the structure-analyzed sentence is input, and essential components for each verb are set, and reconstruction candidates are generated for each verb appearing in the sentence among candidates of the omitted component. At this time, for the sake of simplicity, the essential components for each term are given and limited to the target words (S200).

In more detail with reference to Figure 4, when the structure-analyzed sentence is input in the first step (S210), through the input structure-analyzed sentence to grasp the essential components for each word appearing in the sentence (S220) ).

In the present specification, given as essential components for each word, and limited to the target word, first, the candidate reconstruction candidates are generated for each word appearing in the sentence (S230). In this case, as the subject restoration candidate, all the subjects of the verb appearing in the same letter are set as the subject of the verb in which the subject does not exist.

Subsequently, object restoration candidates are generated for each verb appearing in the sentence (S240). In this case, since the target word has only the transitive verb differently from the subject, it is first determined whether the corresponding verb is an automatic verb or a transitive verb using a pre-stored child / transitive verb. Therefore, the candidate for restoration of the object is set as the object candidates of the verb only when it is determined that the verb is a transitive verb among sentence components used in the verb appearing in the same sentence.

In a final third step, the sentence for which the skipped component is restored is completed by determining the actual restore candidate using the skipped component candidate for each word generated by the skip candidate setting (S300).

Referring to FIG. 5, if the generated constituent skip component candidates are input (S310), it is first determined whether the candidate is a subject or a target object (S320).

In a case where the determination result (S320) determines that the candidate is a subject, regularized semantic information of an omitted component candidate is used (S330).

In this case, in the case of the defined specific semantic information, since the tendency to be restored is 50% or more, if the sentence component having such semantic information is determined to be a candidate element for omitting specific words (S340), the candidate is given to the omitted sentence. Restore to (S350).

Subsequently, when the candidate is determined to be the subject, it is determined as an omitted component candidate that does not determine whether to restore the ordered semantic information (S340), or when the candidate is determined as an object (S320). The battle information is used (S360).

In this case, when it is determined that the omitted component candidate is the target word (S320), it is preferable to use the ordered battle information without using the ordered semantic information. This is because, when the omitted component candidate is an object, there is almost no result matching the ordered semantic information when the ordered semantic information is used.

The regularized battlefield information used is the same as in the second embodiment.

Example 2

A = meaning code! Prospective words!

A = People! Go to B = Artificial place!> [He [A] goes to sea [B]]

The method using such a regularized framework is shown in Table 1.

 input   -Heading: Along Bay [sense] is located east of Hai Phong.  Parsing sentence structure   Position (subj: NULL, obj: NULL, ADV: East: In Battle: Direction!)   Battle match   24265-2 A = Where! B = Where! To C = Direction! 24265-4 A = Where! To B = Direction! 24265-8 A = Weather! Go = B = Direction! 24265-12 A = Direction 24265-17 A = site! To B = direction!  Restore or not   Restore to subject

As shown in Table 1, the method using the regularized framework is as follows.

First, if the input is inputted as “located east of Haiphong” with the headword Alangman, the input sentence is analyzed through the analysis of the sentence structure (S362).

The structured analysis results are used to match a predefined regular frame (S364), and the matched frame is detected (S366).

When the frame other than the omitted subject or object is detected through the detection process, the frame is restored (S368). In other words, if the detected frame is not present in the sentence structure analysis terminology and the detected frame is registered as an essential component of the input sentence, the component of the sentence to be restored along with whether or not to be restored is defined with reference to the detected regular frame. This will restore the omitted sentences.

However, these reconstructed sentences cannot determine the correct reconstruction for all omitted sentence component candidates because of the limitations of the regularized semantic information and regularized strife information.

therefore. The predefined statistical information is applied to the restored sentence (S370).

The statistical information refers to information obtained by numerically defining three cases through an input sentence structure. The three cases are first, when the omitted sentence is not restored, second, when the subject is restored, and third, when the object is restored.

In this case, the numerical statistics are each represented by a score between 0 and 1, and the larger the value is, the closer the sentence component candidate is to be omitted.

Accordingly, by applying the regularized semantic information by selecting one represented by the largest value among the values displayed by applying the statistical information, the sentence component already selected for restoration is applied once more to the rule / statistic information. Has the result of verification.

By applying the statistical information defined in this way to the reconstructed sentence, it is possible to more accurately reconstruct each omitted sentence component candidate.

Therefore, it is determined whether the reconstruction (subject) through the sentence structure analysis (S350) and the reconstruction through subframe matching (subject, object) (S360) are the same as the verification through the statistical information (S380).

As a result of the determination (S380), if it is the same, the sentence is restored (main language) through the sentence structure analysis (S350) and the restored (major object) (S360) through the matching of the frame (S390).

If the result of the determination (S380) is not the same, the candidates restored through the sentence structure analysis (subject) (S350) and the restored through the pertinent matching (subject, object) (S360) are changed into candidates defined in the statistical information. To restore (S400).

In this case, for verification through the statistical information, a user arbitrarily defines a threshold value for each of the three cases in advance.

That is, the threshold value is a value that sets a criterion for a value calculated by applying to the statistical information when not restored, when restored to the subject, when restored to the object, etc., which are the three cases. The case of more than the value becomes a candidate of the abbreviated sentence.

Therefore, the candidates restored through the sentence structure analysis (main) (S350) and the restored through the pertinent matching (subject, object) (S360) are candidates that cannot exceed a threshold value arbitrarily determined by the user in the application of statistical information. In the case, the candidate determined in the restoration through regularized semantic information and frame matching is ignored and the candidate recommended in the statistical information is selected as a target to be finally restored (S400).

If the candidates restored through the sentence structure analysis (S350) and the candidates restored through the matching of the frame (S360) do not show below a threshold value arbitrarily determined by the user in the application of statistical information, The restored candidate is maintained as it is (S390).

Through the above process, the presence of the omitted component and the omitted component candidate are recognized for the sentence including the omitted component, and the restored candidate is actually restored using the rule and statistical information after reconstructing the recognized candidate. Output (S410).

In this case, in the present invention, important features are lexical features, POS features, semantic features, structural analysis features, and composite features.

As described above, the preferred embodiment of the present invention has been disclosed through the detailed description and the drawings. The terms are used only for the purpose of describing the present invention and are not used to limit the scope of the present invention as defined in the meaning or claims. Therefore, those skilled in the art will understand that various modifications and equivalent other embodiments are possible from this. Therefore, the true technical protection scope of the present invention will be defined by the technical spirit of the appended claims.

The sentence component restoring apparatus and method according to the present invention as described above has the following effects.

First, it is possible to improve the performance of automatic translation, retrieval, information extraction, and question and answer system using sentence structure analysis.

Second, the sentence structure analysis was performed to recognize the candidates that could be omitted correctly by sentence structure analysis, and the ability of sentence structure analysis can be improved by appropriately using rules and statistics in the task of determining the restoration target among the omitted candidates. .

Claims (12)

A sentence structure analysis unit for analyzing a structure of an input sentence based on a predefined grammar, An omission candidate recognition unit that detects essential components for each term appearing in the analyzed sentence and detects reconstruction candidates of an element omitted in the input sentence if it is determined that the sentence component is omitted; And a skipped component restorer for restoring the omitted components by using the predetermined rule information and statistical information defined in the detected restored candidates. A first step of analyzing the structure of the sentence including the omission of the candidate; A second step of detecting reconstruction candidates of the omitted components in the input sentence, if it is determined that the essential components of each term appearing in the analyzed sentence are omitted; And a third step of determining the actual restoration candidate by using the ordered battle information and the statistical information among the omitted element candidates generated by the omission candidate setting, and completing a sentence in which the omitted elements are restored. Omitted sentence component restoration method. The method of claim 2, wherein the first step Sequentially performing at least one of a part-of-speech tagging, an entity name recognition, and a vocabulary tagging process on a sentence including an omitted sentence component; And analyzing the structure of a sentence while comparing a grammar predefined to a sentence that has undergone at least one process. The structure of claim 2, wherein the structure of the sentence A sentence component composed of any one of an object, a bore, a noun clause, and an adjective clause, and at least one of a connection component indicating a connection relationship between a sentence and a sentence, a clause and a clause, or a sentence and a clause. Omitted sentence component restoration method. The method of claim 2, wherein the second step Identifying essential components for each term appearing in the structurally analyzed sentence; Setting all the subjects of the verb appearing in the sentence among the identified essential ingredients for the verb as a reconstruction candidate for the subject of the verb without the subject; And restoring all object words of the verb appearing in the sentence among the identified essential components for the verb as a candidate for restoration of the object of the verb without the object. The method of claim 5, The restored candidate of the sentence, characterized in that it is set to the restored candidate according to the object of the verb only when the determined verb is determined to be another verb using the pre-stored child / verb verb dictionary. The method of claim 2, wherein the third step A determination step of determining whether restoration candidates of the omitted elements are given or object when the generated candidates for each omitted element are input; A first reconstruction step of restoring the candidate to the omitted sentence if the sentence component having the semantic information is a candidate element of a specific verb using the ordered semantic information of the omitted component candidate if the candidate is the subject; Wow, If it is determined that the candidate is a subject and is determined to be an omitted component candidate whose restoration is not determined by the ordered semantic information, or if the candidate is determined as an object, matching is performed on a predefined regularity frame. And a second restoring step of detecting and restoring the matching gap as a result of the matching; And a re-verification step of re-validating a sentence component that is already selected for restoration by using numerical statistics calculated by applying predefined statistical information to the restored sentence. . The method of claim 7, In the second restoring step, the detected sentence component restoring method may include detecting the same frame except for the omitted subject or object. The method of claim 7, wherein The statistical information is omitted when the omitted sentence is not restored, when restored to the subject, the case of restoring the object as a numerical value (score) between 0 and 1, respectively How to restore sentence components. 8. The method of claim 7, wherein the revalidation step is Determining whether reconstruction through sentence structure analysis and reconstruction through frame matching are the same as verification through the statistical information; If the determination result is the same, reconstructing the sentence through analysis of the sentence structure and maintaining the restored sentence as it is; And if it is not the same, restoring the sentence component by restoring through the sentence structure analysis and reconstructing the candidate restored through the stratification matching to the candidate defined in the statistical information. 8. The method of claim 7, wherein the revalidation step is For verification through the statistical information, the user applies a threshold value calculated by applying 1) not restored, 2) restored to the jury, and 3) restored to the destination. Optionally predefined steps, If the restored candidate is a candidate whose numerical statistics calculated by the application of the statistical information do not exceed the threshold value, selecting and restoring a candidate recommended in the statistical information as a final restoration target; If the restored candidate does not have a numerical value calculated by applying the statistical information below the threshold value, maintaining the previously restored candidate as it is; . The method of claim 7, wherein The statistical information may be at least one of a lexical feature, a POS feature, a semantic feature, a structural analysis feature, and a compound feature.

Images