KR101176772B1 - An automatic annotation system for generating sentence ontology and a method thereof - Google Patents

Abstract

The present invention relates to an automatic annotation system and method for generating a sentence ontology, the automatic annotation system for generating a sentence ontology according to the present invention is a decomposition storage unit, input sentence is decomposed and stored in the appropriate unit of the corpus annotated sentences When it comes in, it includes an annotation unit that finds a semantic and structurally similar fragment among sentence fragments decomposed from the decomposition storage unit and annotates an input sentence with reference to the contents annotated to the fragment.

In addition, the partial similarity-based annotation method of a sentence according to the present invention, the first process of decomposing the sentence into comparable units, the second process of measuring the structural similarity of the sentence fragments, in order to find out how similar the arguments are semantically And a third process of measuring semantic similarity of the sentence fragment, and a fourth process of performing annotation based on the sentence similarity.

Ontology, annotation, corpus, domain, sentence similarity, sentence decomposition

Description

An automatic annotation system for generating sentence ontology and a method

The present invention relates to an automatic annotation system and method for generating a sentence ontology, and more particularly, to find a sentence similar to the input sentence in a particular corpus, and to import the predicate-discussion structure of the similar sentence as it is and annotate the input sentence. An automatic annotation system and method for ontology generation.

The knowledge of human life is expressed and stored in many forms. It exists only in the head of an individual and can be lost or it can be spoken like a traditional fairy tale.

However, knowledge that is useful to many people remains mostly in the form of pictures and texts for future generations. It is the desire of scholars to study natural language processing that corresponds to a double writing, that is, to express the knowledge contained in a document in a logical form that can be processed by a computer.

Broadly speaking, attaching part-of-speech to words in a document and analyzing the syntax structure are all examples of ways that computers can understand the content of a document.

Ontology is a form of computer-readable representation that combines with the World Wide Web, which currently stores vast amounts of information.

1 is a diagram illustrating an example of an ontology.

As shown in FIG. 1, an ontology basically consists of a concept and a relationship between the concepts. In FIG. 1, all elliptical figures correspond to Concept, and arrows correspond to relation between concepts.

For example, in the case of (Person, wears, Apparel)-a person wears clothing-the person and the apparel each correspond to the concept and there is a relationship of wears between the person and the clothing. will be.

There are many documents on the web, and if these documents can be represented ontologies and connected to each other through higher ontology, inference can provide intelligent services that have not been possible until now. In this respect, converting documents to ontology with high accuracy is very meaningful.

However, finding concepts in sentences and finding relationships among them is a very difficult task. It is also difficult to find out which parts of a sentence form a unit of concept.

Chunking (see Table 1) and terminology extraction (Term Recognition, see Table 2) can be thought of as finding a unit of concept, each forming one difficult problem.

Table 1

(S (NP-SBJ Westinghouse Electric Corp))
(VP said
(SBAR (S (NP-SBJ it) (VP will (VP buy (NP Shaw-Walker Co)))))) (..))

Table 2

Westinghouse Electric Corp. _Term said it will buy [Shaw-Walker Co] _Term .

Finding relationships between concepts is also a difficult problem. Relationship Extraction currently occupies one big area of Information Extraction.

2 illustrates the representation of each sentence of the document in an ontology form when a natural language document is received. In other words, the concept is to find a concept for each sentence and to find a relationship between the concepts.

A predicate is one of the two main components of a sentence with a subject in classical English grammar, representing all parts of the sentence except the subject. However, predicates in the predicate-dissertation structure generally refer to verbs, in contrast to their classical meanings, and arguments refer to subjects, objects, or modifiers that correspond to predicates. That is, as shown in Table 3 below, the predicate-dissertation structure refers to a specific verb and the subject, object, and modifier corresponding to the verb.

TABLE 3

PRED: predicate, ARG: argument
(One) [John] _ARG0  broke _PRED  [the window] _ARG1
(2) [The window] _ARG1 broke _PRED

Table 3 shows an example of a predicate-dissertation structure. The verb 'break', which means 'break', can usually have a 'breaker' and 'breaking' argument.

In the example above, 'breaker' is labeled ARG0 and 'breaking' is labeled ARG1. In example (1) of Table 3, 'break' becomes a predicate, 'John' becomes 'broken person' and 'the window' becomes 'broken'. In the example of (2) of Table 3, 'broke' is used to mean 'broken' as an intransitive verb, and 'the window' comes to the position of the subject, meaning that it corresponds to 'broken'. It is marked ARG1.

Predicate-argument structures can be used as the basis for creating document-based ontology. Basically, arguments have properties similar to the concepts used in ontology, and in the case of predicates, they can be considered in relation to ontology relationships.

By associating predicates and arguments with concepts and relationships, it is possible to lay the groundwork for transforming each sentence in a document into an ontology. Also, the predicate-argument structure is particularly meaningful as Bootstrap data for annotating event structures [P. Kingsbury et al., "PropBank as a Bootstrap for Richer Annotation Schemes" The 6 ^th Workshop on Interlinguas: Annotations and Translations ,, in conjunction with MT Summit 2003, New Orleans, Louisiana, 9 pages, 2003].

If you can annotate the predicate-discussion structure automatically, you have the basic form to annotate the event structure of a number of natural language documents being produced on the Web. This means that current news can also be automatically analyzed to some degree. In other words, the predicate-argument structure may be used to improve performance such as cause-effect analysis between events and retrieval of various events through event annotation.

Propbank is one of the predicate-annotated structured corpus, annotating the predicate-argument for Penn-Treebank's Wall-Street journal article [M. Palmer et al., "The Proposition Bank: An Annotated of Semantic Roles", Computational Linguistics , Vol. 31. Pp. 71-106, 2005].

There are two different aspects to text-based ontology generation.

The first approach is to obtain a well refined ontology from a particular field of text. If the ontology corresponding to the 'sport' domain is to be created, the document corresponding to the 'sport' domain is collected, the hierarchical structure of the terminology is obtained from the document, and the relationship between the terminology is extracted.

This methodology is typical of domain ontology construction [N. Noy and DL McGuiness, "Ontology development 101: A guide to creating your first ontology", Technical Report KSL-01-05 and SMI-2001-0880, Stanford Knowledge Systems Laboratory and Stanford Medical Informatics, 2001] In this way, it is a way to create a single, consistent form of ontology, regardless of whether the sentence has a semantic loss.

The second method is an ontology building method that focuses on expressing the structure and meaning of the sentence itself. It understands ontology as one way of logicalizing a sentence. This methodology aims to represent the sentence as it is, without losing the meaning of the sentence, and enables the sentence to be understood through the mapping process with the higher ontology. Omega Ontology is an ontology that extends and restructures WordNet and provides grammatical elements to understand sentences [A. Philpot et al, "The omega ontology", In Proceedings of the ONTOLEX Workshop at the International Conference on Natural Language Processing, pp. 59-66, 2005].

TextToOntto is a program that generates domain ontology from a specific document set [D. Oberle, R. Volz, B. Motik, and S. Staab, "An extensible ontology software environment", In S. Staab and R. Studer, editors, Handbook on Ontologies, International Handbooks on Information Systems, chapter III, pp. 311-333, 2004.

It uses text mining techniques to help the steps defined in ontology engineering. According to the typical ontology building methodology, the ontology building process is classified as follows.

First, the domain is confirmed. Second, look for existing ontology. Third, collect terminology that frequently appears in domains. Fourth, create a hierarchy of terms. Fifth, define the property of class. Sixth, establish relationships between classes.

TextToOnto is a tool that semi-automates the above processes and shortens the time of ontology construction. In this case, only the parts of the document set that can be used appropriately for the domain are converted into ontology.

As can be seen in Figure 3, the sentence 'CheesyPizza is a kind of pizza.' Is reflected in the hierarchy of the pizza ontology. The second sentence, Joo-won likes it very much., Is an individual fact and did not affect the pizza domain ontology.

Omega Ontology is an ontology that reorganizes and expands WordNet (Christiane Fellbaum, editor., "WordNet: An Electronic Lexical Database", MIT Press , 1998.). In addition to WordNet, it has the frameworks of verbs such as FrameNet (CF Baker, CJ Fillmore, and JB Lowe, "The Berkeley FrameNet Project", In proceedings of COLLING / ACL, pp 86-90, 1998], Propbank, etc. Semcor [M. Sundeanu et al., Using Predicate-Argument Structures for Information Extraction ", In Proceedings of the 41 ^st Annual Conference of the Association for Computational Linguistics (ACL-03), pp. 8-15, 200]. Word Sense Disambiguation results are also linked to the omega ontology, with the rich concepts contained in the omega ontology can be a fundamental resource for interpreting natural language.

Among the many corpus, the most conducive to automatic ontology building is the Penn-Treebank and Propbank. Penn-Treebank is the corpus of parsed sentences.

As the name Bankbank suggests, the syntax is represented in the form of a tree. After the automatic analysis, these results were verified manually by two or more annotators to verify their accuracy.

Penn-Treebank currently contains not only simple syntax, but also semantic information about sub-ject and temporal and location. The Pen-Treebank is widely used as learning data for part-of-speech attachment, syntax analysis, and the like.

Propbank is a predicate-dissertation annotation on Penn-Treebank's Wall Street Journal. Frames were created for each verb, and arguments matching the skeleton of the verb were found and marked. Probbank's argument is one basic concept that can be transformed into an ontology class or instance.

If you divide the similarity measure of sentences into two categories, it can be divided into two branches. One focuses on the semantic similarity of sentences, and the other focuses on the structural similarity of sentences.

"It is a dog", "That must be a dog.", "It is a log." When three sentences are found, the method of focusing on the semantic similarity of the sentences is "It is a dog" and "That must be a dog." The similarity of is judged high.

On the other hand, positions that value the structure of sentences are "It is a dog" and "It is a log." Is judged to be more similar. The similarity of sentences can be largely divided into these two aspects, and each case also shows different aspects in detail.

Structural similarity measurement of sentences uses tree similarity.

4 illustrates three operations used in a method using an edit distance among tree similarity measurement methods. In this method, three operations, "insert", "delete", and "change", are presented to change the tree, and the similarity of the tree is obtained by penalizing each operation.

Syntactic Analysis of Sentences When the tree is compared using the Edit Distance, the structural similarity of the sentences can be measured.

The semantic similarity measurement of sentences takes a method of extending word similarity to sentence similarity. Similarity between words uses semantic hierarchy between words.

5 is a part of WordNet, and the similarity between words can be easily measured by measuring the distance between words under such a hierarchical structure. In order to extend this to the similarity of a sentence, the sentence is expressed in the form of a semantic vector.

In this case, by placing similar words on the same axis, a semantic vector capable of expressing semantic similarities of sentences is formed. In addition, the semantic similarity of the sentences is measured by finally considering the order between words not considered in the semantic vector.

An automatic predicate-discussion annotation is typically a case where training data is collected in argument units and trained through a specific learning model. In this case, the path of the tree from the predicate to the argument, the tone of the verb, the tag of the phrase or clause is used as the feature.

There was a phenomenon in which the accuracy rate and the reproducibility were inversely proportional to the use of specific features. Since then, automatic predicate-article annotations are usually divided into argument boundary analysis and role labeling, and each process is solved through machine learning. The double treekernel is a representative example of learning tree paths.

An ontology is a format that describes concepts and their relationships. In order to express sentences in ontology form, we must finally be able to capture the boundaries of concepts, and also be able to establish relationships between the captured concepts. However, in the case of the conventional complex sentence, there was a problem that the above operations are not easy.

In addition, the present invention is similar to the existing machine-learning (machine-learning) in terms of splitting the sentence to use as a feature. However, in the case of machine learning, it was difficult to trace back the example.

The present invention devised to solve the above problems provides an automatic annotation system and method for generating a sentence ontology that uses a corpus (eg, pen-tree bank) to annotate structure-predicate structure of natural language sentences. The purpose.

An automatic annotation system for generating a sentence ontology according to the present invention for achieving the above object includes a decomposition storage unit, input sentence to decompose and store the annotated sentences of the corpus (for example, pen-treebank) in a suitable unit It includes an annotation unit for finding the meaning, structurally similar fragments among the fragments of sentences decomposed from the decomposition storage unit, and annotating the input sentence with reference to the contents annotated to the fragments.

In addition, the partial similarity-based annotation method of a sentence according to the present invention, the first process of decomposing the sentence into comparable units, the second process of measuring the structural similarity of the sentence fragments, in order to find out how similar the arguments are semantically And a third process of measuring semantic similarity of the sentence fragment, and a fourth process of performing annotation based on the sentence similarity.

The system according to the invention has the advantage that it is easy to track the example and inform the feedback.

Current sentence comparison is a relatively time-consuming task. According to the present invention, if the key parts of the predicate-dissertation structure are extracted and compared, the effect of more accurate and faster predicate-dissertation structure is possible. have.

The system proposed in the present invention shines more when other additional problems are solved. First of all, the ontology mapping task is the most important part. Basically, ontology mapping is made through many lexical comparisons including Edit Distance, which is not important according to the system proposed in the present invention.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

6 is a diagram illustrating an example of an annotation using a similar sentence according to the present invention.

The present invention uses a propbank to annotate predicate-discussion structures of natural language sentences. The basic approach is to find sentences that are similar to the input sentences in the Propbank, and annotate the input sentences by bringing the predicate-argument structure of the similar sentences as they are.

In the example of FIG. 6, the phrase 'The prices hit bottom.' When I looked for similar sentences, I found out that the two sentences on the right are similar to the following.

Therefore, like the relevant sentence, the subject portion of "hit" was annotated with ARG0, and the object portion was annotated with ARG1. An annotation based on the similarity of sentences as described above is a method proposed by the present invention.

7 is a diagram illustrating a configuration of an automatic annotation system for generating a sentence ontology according to the present invention.

As shown in FIG. 7, the automatic annotation system for generating a sentence ontology according to the present invention includes a decomposition storage unit and an input sentence for decomposing and storing the annotated sentences of the Pen-Treebank into suitable units. When it comes to, it includes an annotation unit for finding the semantic, structurally similar fragments among the fragments of sentences decomposed from the decomposition storage unit, and annotating the input sentence with reference to the contents annotated to the fragments.

In addition, since only the parts of the predicate-determination structure centered on the verb in the annotation unit are compared in similarity, when several verbs appear in a sentence, the sentence is a sentence decomposition step which is the first step in the decomposition storage unit. It is characterized by decomposing into pieces centered around each verb.

Next, a partial similarity based annotation method of a sentence according to the present invention will be described in detail.

An annotation method based on partial similarity of a sentence according to the present invention includes a first process of decomposing a sentence into a comparable unit, a second process of measuring structural similarity of the sentence fragment, and how the arguments are semantically similar. A third process of measuring semantic similarity of the pieces, and a fourth process of performing annotations based on the sentence similarity.

More detailed description is as follows.

Sentence decomposition

8 is a diagram illustrating a sentence fragment generation according to the present invention.

In the present invention, for sentence similarity based annotation, the sentences are broken down into smaller units for all sentences in the propbank. As can be seen in Figure 8, there are two types of decomposed pieces, one is called a multinomial fragm-ent, the other is called a binomial Fragment.

The example shown in FIG. 8 illustrates the predicate-dispute relationship of the sentence "The market hits bottom" as a sentence fragment. Each fragment has a tree of paths from the predicate to the argument as fragments of the sentence. Polynomial fragments are intended to preserve both the predicate-discourse structure for a verb.

In other words, ([The market] _ARG0 , hit, [bottom] _ARG1 ) should be treated as a unit that can no longer be decomposed. In contrast, a binomial piece breaks down ([The marke -t] _ARG0 , hit, [bottom] _ARG1 ) into smaller units of ([The market] _ARG0 , hit), (hit, [bottom] _ARG1 ). Decompose into

The criterion of decomposition is to treat the path from the verb to each argument as one piece.

Therefore, since a sentence includes several verbs, the sentence is represented by a polynomial relation fragment as many as the number of verbs. In addition, each polynomial piece is represented by as many binomial pieces as the number of arguments. This sentence decomposition process is shown in Table 4 below.

Table 4: sentence decomposition algorithm

TreeDecomposition ( ParseAndPropTree T )
1.Find the predicate verb ( VBD in Figure 6)
Loop every argument, Arg _i
a. Find the parent node of Arg _i and the ancestor of VBZ
b. Decide the direction of the branch to Arg _i
c. Record the trace from VBZ to Arg _i and the internal of Arg _i
d. Produce one decomposed element from the record
3.Return the set of the decomposed elements

Also, the fragmented sentence fragment can be expressed as follows.

td becomes each binomial piece, and TD is a polynomial piece. TD can be generated by collecting binary binary pieces and vice versa. Td consists of arg , n , dir , and so on. n represents the path from the verb to the argument, dir is the record of the redirection of the path, and arg represents the tree of the argument itself. This fragmented sentence fragment is used in the similarity comparison process.

Measure structural similarity of sentence fragments

9 is a view showing the structural similarity of the polynomial relational fragments according to the present invention.

The structural similarity measurement of sentence fragments follows the tree similarity measurement method. Of these, the score of the 'change' operation was modified to fit the Treebank. The structural similarity measure of a statement is made only for the tree path of each piece, not for elements such as Arg. In the case of polynomial pieces, the similarity is measured as shown in FIG. 9.

As can be seen in Figure 9, to change the left tree to match the right tree, one 'Change', one 'Delete' and one 'Insert' were required. 50 penalties were imposed for changing 'VBG' to 'VBD', and 200 penalties were charged for 'Insert' and 'Delete'.

Thus, comparing the trees requires a total of seven operations, of which 450 penalties are imposed, resulting in a similarity of (200 * 7-450) / (200 * 7) = 0.68, that is, 0.68.

Binomial fragments can be compared in the same way, assuming a tree. In essence, however, a binomial fragment can be viewed as a string, which is equivalent to a typical Levenshtein distance.

Measure semantic similarity of sentence fragments

10 is a diagram illustrating the semantic similarity measurement of the argument according to the present invention.

As shown in FIG. 10, the semantic similarity measure is to determine how similarly the arguments are semantically, based on the word similarity of WordNet. Word similarity in WordNet is usually measured using the length of the path [T. Pedersen et al., "WordNet :: Similarity-Measuring the Relatedness of Concepts" Proceedin -gs of the Nineteenth National Conference on Artificial Intelligence (AAAI-04), July 2004. Words that are the same word or are included in the same Synset have a score of 1.0. In general, the number of paths is calculated while searching the network, and the score is inversely proportional to the inverse of the number of paths.

In the above equations, "dog" and "fox" have a WordNet path similarity of 0.33 points, and the semantic similarity of the two arguments finally has a value of 0.8325. Similarity of arguments is compared using the similarity of paths on WordNet. In addition, the similarity of arguments is finally obtained by additionally using the order of parts of speech and the height of the tree.

Sentence Similarity Based Annotation

Statement similarity-based annotation combines the aforementioned polynomial fragments, binary relation fragments with structural similarity, and semantic similarity. However, since the sentence is not divided into pieces, the sentence to be annotated calculates the similarity between pieces after inferring the pieces. The basic process of annotating is as shown in FIG.

The sentence to be annotated "The prices severely depressed levels" does not have a sentence fragment to compare. Therefore, the polynomial relation fragment is found in the parsing tree on the right using 'NP', 'S', 'hit', and 'NP' as the boundary nodes of the polynomial relation fragment in the red part on the left. We also assume that the part of each boundary of the found piece is a dispute and then calculate the similarity between the two pieces. In this case, the structure of the fragments is completely identical. In this case, the input sentence can be annotated using the argument information on the left.

There are six kinds of statement similarity-based annotation methods. In the case of using only structural similarity of sentences, the first two combinations are using structural similarity and semantic similarity. In addition, in the use of sentence fragments, there may be three cases of using only polynomial pieces, using only binomial pieces, and using polynomial pieces and binomial pieces at the same time. The combination of these two factors allows for a total of six sentence similarities.

12 shows a process of annotating using structural similarity between a polynomial piece and a sentence. First of all, the boundary of the fragment to be annotated and the structural similarity are measured.

However, there are limitations in the annotation method in FIG. 12 that depend only on the structural similarity and polynomial pieces of such sentences.

The example of FIG. 13 tells the first problem. "The prices hit bottom" is structurally similar to "I hit the ball". However, the "hit" in the previous two examples is used in a different way, rather than "Sterling plunged and hit its low."

In this case, if you can see that the terms "Sterling" and "Prices" and "low" and "bottom" are used in a similar way, it is more similar to the above sentence. In other words, the semantic similarity of sentences is necessary.

Figure 14 illustrates the limits of polynomial pieces.

If the example sentence for the "abort" verb has only two sentences, and the input "The company aborted its bid" is input, you can use both example sentences to annotate the predicate-discussion structure. That is, the lack of argument in the example should be compensated for as part of the other examples. Not only polynomial pieces, but also binomial pieces can fill this gap.

Due to the limitations of structural similarity and polynomial fragments described above, the semantic similarity and binomial fragments of sentences are needed, and the annotations when using them are as follows.

15 is a diagram illustrating a predicate-nonsense structure annotation using semantic similarity according to the present invention.

As can be seen in Figure 15, the limitation of Figure 14 could be solved using the semantic similarity of the argument. The similarity between "The prices hit bottom." And "I hit the ball" is 1.0 + 0.01 = 1.01, and in the case of "Sterling plunged and hit its low.", It is 0.83 + 0.365 = 1.195 to select the correct sentence. Could know.

16 is an annotation using a polynomial piece and a binomial piece at the same time, the polynomial piece, annotated ARG0 by selecting "She completely aborted" and annotating the empty part using the binomial piece of "Aborts command". That's the process. Using both at the same time helps solve the sparseness problem of the example.

In addition, in the case of using only the binary relational fragments without using the polynomial relational fragments, the partial correspondence of the sentences can be brought to the highest. However, there is a side that loses the whole of predicate-argument structure. 17, only the binomial relation fragment is used.

In annotations using binomial sculpting, to annotate "The prices hit bottom", only the "The prices" portion of "The prices hit severely depressed levels" and the "hit bottom" portion of "The market didn't hit bottom yet" Will be combined. This approach allows you to find the structure, semantically the most similar sentence fragment.

In terms of performance of the present invention, the accuracy was found to be higher than that of the existing studies, and the reproducibility was found to be low. The low reproducibility problem can be solved by partially using various machine learning through integration with existing research.

In self-assessment, the similarity of sentence was divided into structural similarity and semantic similarity to measure performance.

When using only structural similarities, there was a 60% chance that the predicate-argument structure could be annotated. The use of the semantic similarity of the arguments has been improved by about 3%, and when the predicate-dissertation structure is further decomposed and used as a binary relational fragment, the performance is improved by about 15%.

In addition, using binomial fragments only showed about 4% higher performance than using both polynomial and binomial fragments, which is more advantageous than decomposing predicate-non-term structure over polynomial relations. It is shown.

Clustering and network composition of sentences using Propbank is one improvement for faster, more accurate annotation.

Words in a sentence are mapped to WordNet, so as long as most words exist in WordNet, there will be few cases where there are no words to map. Rather, WordNet is composed of the meaning of words, so it is important to grasp the sense of each word of sentence ontology.

In other words, Word Sense Disambiguation (WSD) replaces ontology mapping problem. However, the condition of the WSD is that each concept has more specific information such as thematic role rather than a flat word.

In this particular situation, playing WSD can be a topic. In addition, the ontology construction process does not contain information at the discourse level. If the research to find the sentence analysis (Sentimental Analysis) goes well, it will make the sentence ontology richer and further develop the ontology that can grasp the whole document well.

Finally, the problem of finding an application of sentence ontology is also an interesting problem. After each newspaper article is in the form of a sentence ontology, it is possible to summarize the common opinions by analyzing what is the common ontology and what is different about the articles of the same subject, and to highlight the differences between newspapers. will be.

1 is a view showing an example of an ontology,

2 is a diagram illustrating an example of a sentence ontology;

3 is a diagram illustrating an example of creating a hierarchy of TextToOnto;

4 illustrates three Edit Distance operations for measuring tree similarity;

5 illustrates a semantic hierarchy of words (part of WordNet);

6 illustrates an example of an annotation using a similar sentence according to the present invention;

7 is a diagram illustrating automatic annotation for sentence ontology generation according to the present invention;

8 is a diagram illustrating a sentence fragment generation according to the present invention;

9 is a view showing the structural similarity of the polynomial relational fragments according to the present invention,

10 is a diagram illustrating the semantic similarity measurement of arguments according to the present invention;

11 is a diagram illustrating a baseline-sentence similarity based annotation according to the present invention;

12 is a diagram showing an annotation using structural similarity between polynomial relational fragments according to the present invention;

13 illustrates a limit of structural similarity according to the present invention;

14 is a view showing the limits of polynomial relational fragments according to the present invention;

15 is a diagram illustrating a predicate-discussion structure annotation using semantic similarity according to the present invention;

16 is a view showing the annotation using the polynomial relation and the binary relational fragments in accordance with the present invention, and

17 illustrates an annotation using a binary relational fragment according to the present invention.

Claims (8)

A decomposition storage unit for decomposing and storing corpus annotated sentences including Penn-Treebank into sentence units including verbs, and When an input sentence comes in, an automatic annotation for generating a sentence ontology including an annotation unit for finding a meaningful, structurally similar fragment among the fragments of the sentence decomposed from the decomposition storage unit and annotating the input sentence with reference to the contents annotated in the fragment. system. The method of claim 1, Since only similar parts of the predicate-discussion structure centered on the verb in the annotation unit are compared, if multiple verbs appear in a sentence, the sentence may be divided through a sentence decomposition step, which is the first step in the decomposition storage unit. Automatic annotation system for sentence ontology generation, characterized in that it is decomposed into fragments centered on the verb. A first process of decomposing the sentence into comparable units, A second process of measuring structural similarity of the sentence fragments, A third process of measuring semantic similarity of the sentence fragments to find out how similarly the arguments are semantically, and An automatic annotation method for generating a sentence ontology comprising a fourth process of performing an annotation based on the sentence similarity. The method of claim 3, The sentence fragment decomposed in the first process is classified into two types, one of which is a polynomial relational fragment, and the other is a binomial relational fragment. The method of claim 3, In the first process, since a sentence includes a plurality of verbs, the sentence ontology is represented by a polynomial relation fragment as many as the number of verbs, and each polynomial relation fragment is represented by a binary relation relation fragment as many as the number of arguments. Auto Annotation Method. The method of claim 3, The structural similarity measurement of the sentence in the second process is performed only for the tree path of each fragment, the automatic annotation method for generating a statement ontology that is not performed for elements such as Arg. The method of claim 3, The semantic similarity comparison of the arguments in the third process uses the path similarity of words on WordNet, and further generates the sentence ontology to finally obtain the similarity of the arguments by additionally using the order of parts of speech and the height of the tree. Auto Annotation Method. The method of claim 3, The annotation in the fourth process is an automatic annotation method for generating a sentence ontology to combine a polynomial relationship, a binary relationship fragment and structural similarity, semantic similarity.

Images