KR20170089142A - Generating method and system for triple data - Google Patents
Generating method and system for triple data Download PDFInfo
- Publication number
- KR20170089142A KR20170089142A KR1020160009168A KR20160009168A KR20170089142A KR 20170089142 A KR20170089142 A KR 20170089142A KR 1020160009168 A KR1020160009168 A KR 1020160009168A KR 20160009168 A KR20160009168 A KR 20160009168A KR 20170089142 A KR20170089142 A KR 20170089142A
- Authority
- KR
- South Korea
- Prior art keywords
- pattern
- subject
- predicate
- property
- extracted
- Prior art date
Links
Images
Classifications
-
- G06F17/30731—
-
- G06F17/2705—
-
- G06F17/277—
-
- G06F17/30684—
Landscapes
- Machine Translation (AREA)
Abstract
The present invention relates to a method and system for generating triple data, and more particularly to a method and system for generating triple data by receiving a knowledge base and a corpus composed of triple data including a subject word and an object word in a natural language sentence, Generating unit; A pattern learning unit for extracting and learning a pattern candidate for each lexicon indicating a relationship between the subject and the object among the generated patterns; And a triple generator for generating new triple data based on the learned pattern, wherein the pattern includes a subject search, an object search, and a predicate present in the natural language sentence, or between the subject and the object in the natural language sentence And at least one vocabulary to be located.
According to an aspect of the present invention, data of a document expressed in a natural language in a semantic web field can be easily and quickly structured, and the data processing speed of a computer can be improved.
According to another aspect of the present invention, it is possible to structure Korean data and improve the satisfaction of Korean users.
Description
BACKGROUND OF THE
Today, there is a large amount of information created by millions of users around the world, and new information is being added and updated on the internet every day. Such a large amount of information is provided to users through various web services.
The information that occupies most of the current Internet is represented by the unstructured natural language, and the documents expressed in this natural language are very common for the general users. However, there are many difficulties for a non-human computer to utilize data expressed in such natural language, that is, unstructured data.
Accordingly, in recent years, there has been an increasing number of cases in which information constructed in a non-structural representation as described above is structured so that the computer can construct information in a form that can be calculated by a computer, and representative examples thereof include various knowledge such as Wikidata, DBpedia, and YAGO There is a base.
However, since the cost of creating such structured data is high, the amount of structured data is very small compared to the vast amount of unstructured data on the Internet.
In addition, since such a small amount of structured data is mostly based on the English language, there is almost no Korean data among the structured data, and it is difficult to process the structured data using structured Korean data.
One aspect of the present invention discloses a method and system for generating triple data that can easily and quickly structure large amounts of data expressed in natural language in the field of semantic web.
According to an aspect of the present invention, there is provided a method of generating triple data, comprising: receiving a knowledge base and a corpus composed of triple data including a subject word and an object word in a natural language sentence and generating a pattern based on the received word; Extracting and learning a pattern candidate for each vocabulary representing a relationship between the subject and the object among the patterns in which the pattern learning unit is generated; And generating a new triple data based on a pattern in which the triple generating unit has been learned, wherein the pattern includes a subject search, an object search, and a predicate existing in the natural language sentence, or between the subject and the object in the natural language sentence Lt; RTI ID = 0.0 > vocabulary < / RTI >
In particular, the step of generating a pattern based on a knowledge base and a corpus composed of triple data including subject and object in the natural language sentence and generating a pattern based on the knowledge base and the corpus, Receiving a knowledge base composed of data and a corpus; Extracting at least one sentence including a subject and an object from the knowledge base and the corpus; Extracting at least one word phrase including an extracted subject or object, respectively; Extracting a subject search and an object search based on survey information present in the extracted word; Extracting a predicate present in the extracted sentence; And extracting at least one vocabulary located between a subject and an object in the extracted sentence to generate a pattern including the extracted subject search, object word search, and predicate.
In particular, the extracting of the predicates present in the extracted sentences may include extracting a predicate expressing a relationship between a subject corresponding to the extracted subject search and an object corresponding to the object search when a plurality of extracted in- .
In particular, the step of extracting a predicate present in the extracted sentence may analyze a dependency relationship between a subject corresponding to the subject search and an object corresponding to the subject search, and extract the predicate according to the analysis result.
In particular, the step of extracting a predicate present in the extracted sentence may include: generating a dependency tree structure based on the dependency information between the subject and the target word; extracting, from among the plurality of predicate nodes existing in the generated dependency tree structure, A predicate node corresponding to each object and a predicate node positioned closest to the object node can be selected and a predicate corresponding to the selected predicate node can be extracted.
In particular, the step of generating new triple data based on the learned pattern of the triple generator extracts a partial tree structure based on a predicate extracted from the dependency tree structure generated based on the dependency information between subject and object , New nodes corresponding to subject and object among the extracted partial tree structure are selected, and new triple data including subject, object, and extracted predicate corresponding to the selected node can be generated.
In particular, the step of extracting and learning a pattern candidate for each lexeme indicating a relationship between the subject and the object among the patterns in which the pattern learning unit is generated may further include removing an error pattern among the learned pattern candidates.
In particular, the step of extracting and learning a pattern candidate for each lexeme indicating a relationship between the subject and the object among the patterns in which the pattern learning unit is generated includes a step of obtaining a relation between the subject and the object among at least one word located between the subject and the object in the extracted sentence Or a lexicon representing a label, an identifier, or an attribute in the sentence, as a property; Measuring semantic similarity between the pattern and the property; And determining an error pattern based on the result of the measurement of the semantic similarity between the pattern and the property, and removing the determined error pattern.
In particular, the step of measuring semantic similarity between the pattern and the property may calculate a vector similarity between the pattern and the property mapped to the word embedding space, respectively.
In particular, the step of measuring the semantic similarity between the pattern and the property may be performed such that when the pattern and the property are in different languages, a correlation coefficient between the pattern and the property is calculated based on predetermined lexical pairs having the same meaning as the pattern and the property And projecting the pattern and property into the same word embedding space by learning the projection matrix so as to be higher.
In particular, the pattern may be in Korean, and the property may be in English.
In particular, the step of measuring the semantic similarity between the pattern and the property may calculate the degree of cosine similarity between the pattern and the property when the pattern and the property are each made up of one vocabulary.
In particular, when the pattern and property are composed of a plurality of words or a plurality of words, the step of measuring semantic similarity between the pattern and the property defines an average vector of the elements constituting the pattern and the property, The vector similarity degree can be computed between the pattern and the property based on the above-described relationship.
In particular, the word embedding space may be represented as a distributed representation by mapping a plurality of vocabularies into a vector space of N dimensions (where N is a natural number).
In particular, the storing unit may store the pattern generated from the pattern generating unit and the new triple data generated from the triple generating unit.
According to another aspect of the present invention, there is provided a system for generating triple data comprising: a pattern generator for receiving a knowledge base and a corpus composed of triple data including a subject word and an object word in a natural language sentence and generating a pattern based on the received knowledge base; A pattern learning unit for extracting and learning a pattern candidate for each lexicon indicating a relationship between the subject and the object among the generated patterns; And a triple generator for generating new triple data based on the learned pattern, wherein the pattern includes a subject search, an object search, and a predicate present in the natural language sentence, or between the subject and the object in the natural language sentence Can represent at least one vocabulary that is located.
In particular, the pattern generator may analyze a dependency relationship between a subject corresponding to the subject search and an object corresponding to the subject search, and extract a predicate based on the analysis result.
In particular, the pattern generator may generate a dependency tree structure based on the dependency information between the subject and the object, and may include a subject node corresponding to the subject and object among a plurality of predicate nodes existing in the generated dependency tree structure, It is possible to extract a predicate corresponding to the selected predicate node after selecting one predicate node located closest to the object node.
In particular, the pattern learning unit may generate at least one of a predicate expressing a relationship between the subject and the object among at least one vocabulary located between the extracted subject and the object, or a vocabulary representing the label, the identifier, or the attribute in the sentence as a property , The degree of semantic similarity between the pattern and the property is measured, an error pattern is determined based on the result of the semantic similarity measurement between the pattern and the property, and the determined error pattern can be removed.
In particular, the triple generation unit extracts a partial tree structure based on a predicate extracted from the dependency tree structure generated based on the dependency information between the subject and the target word, and extracts the partial tree structure based on the subject and object And generate new triple data including subject, object, and extracted predicate corresponding to the selected node, respectively.
The storage unit may further store a pattern generated by the pattern generation unit and a new triple data generated from the triple generation unit.
According to an aspect of the present invention, data of a document expressed in a natural language in a semantic web field can be easily and quickly structured, and the data processing speed of a computer can be improved.
According to another aspect of the present invention, it is possible to structure Korean data and improve the satisfaction of Korean users.
1 is a block diagram illustrating a system for generating triple data according to an embodiment of the present invention.
2 is a flowchart illustrating a method of generating triple data according to an embodiment of the present invention.
FIG. 3 is a flowchart showing detailed steps of generating a pattern based on a knowledge base and a corpus among the method of generating triple data according to FIG.
FIG. 4 is a flowchart illustrating the detailed steps of a predicate extraction process in a sentence according to FIG.
5 is a flowchart showing the detailed steps of generating new triple data based on the learned pattern of FIG.
6 is a schematic diagram illustrating a framework illustrating self-knowledge learning according to an embodiment of the present invention.
7 is a diagram illustrating a process of selecting a target node for pattern generation.
8 is a diagram showing a target node selected for instance generation.
9 is a graph showing the number of pattern generations.
10 is a graph showing the accuracy of the generated pattern.
11 is a graph showing the number of instances generated.
12 is a graph showing the accuracy of the generated instance.
13 is a graph showing the number of pattern generations after repeated execution.
FIG. 14 is a graph showing the number of instances generated after repeated execution.
FIG. 15 is a flowchart illustrating a detailed step of extracting and learning a pattern candidate for each lexicon representing the relationship between the subject and object among the patterns in which the pattern learning unit according to FIG. 2 is generated.
FIG. 16 is a flow chart showing detailed steps for calculating the semantic similarity between the pattern and the property in FIG.
17 is a schematic diagram showing a process for expanding a knowledge base.
18 is a diagram illustrating a dependency tree-based pattern generation process.
19 is a diagram showing an example of combining independent learned word embedding spaces.
20 is a diagram showing the projected word embedding space through projection matrix learning.
21 is a diagram showing the Top-K accuracy rate of the generated triple data.
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. It will be apparent to those skilled in the art, however, that these examples are provided to further illustrate the present invention, and the scope of the present invention is not limited thereto.
BRIEF DESCRIPTION OF THE DRAWINGS The above and other objects, features and advantages of the present invention will become more apparent from the following detailed description of the present invention when taken in conjunction with the accompanying drawings, in which: It is to be noted that components are denoted by the same reference numerals even though they are shown in different drawings, and components of different drawings can be cited when necessary in describing the drawings. It is to be understood, however, that the invention is not intended to be limited to the particular forms disclosed, but on the contrary, is intended to cover all modifications, equivalents, and alternatives falling within the spirit and scope of the invention.
In the following detailed description of the principles of operation of the preferred embodiments of the present invention, it is to be understood that the present invention is not limited to the details of the known functions and configurations, and other matters may be unnecessarily obscured, A detailed description thereof will be omitted.
In addition, in the entire specification, when a part is referred to as being 'connected' to another part, it may be referred to as 'indirectly connected' not only with 'directly connected' . Also, to include an element does not exclude other elements unless specifically stated otherwise, but may also include other elements.
Also, the terms first, second, etc. may be used to describe various components, but the components should not be limited by the terms. The terms may be used for the purpose of distinguishing one component from another. For example, without departing from the scope of the present invention, the first component may be referred to as a second component, and similarly, the second component may also be referred to as a first component.
The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular expressions include plural expressions unless the context clearly dictates otherwise. In the present application, the terms "comprises", "having", and the like are intended to specify the presence of stated features, integers, steps, operations, elements, components, or combinations thereof, , Steps, operations, components, parts, or combinations thereof, as a matter of principle.
Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. Terms such as those defined in commonly used dictionaries should be construed as meaning consistent with meaning in the context of the relevant art and are not to be construed as ideal or overly formal in meaning unless expressly defined in the present application .
The emergence of next-generation intelligent web-based semantic web that allows computers to read, understand and process information on behalf of people and create new information. And generates structured data. Extracts triple data such as <subject, property, object> from general natural language sentences and expands existing knowledge base using it. The knowledge base used in the present invention includes a concept representing a specific concept and a knowledge base in general, a concept representing a specific concept and a relation defining a relation between the concepts. This relationship is generally expressed in a triple form composed of a subject, a relation, and an object. A pattern representing the triple relation expressed in this manner can be generated, and a triple instance can be created using the pattern.
The Distant supervision learning method is a ring learning method and learns the model under the following assumptions. The assumption used here is that the sentence containing the subject and object object of a particular triple represents the meaning of the given triple property. As a result, it is possible to learn without having to construct separate learning data if only some seed triple and corpus of the property to learn the model are needed.
However, the distant supervision learning method is used in various studies because of convenience of application. However, it is not always true that the sentence containing the two objects of the triple will represent the meaning of the triple property. There is a problem in this.
For example, when a triplet data <Gloria Stewart, birthPlace, California> is input, the distant supervision learning method extracts the pattern from the following sentence.
"Gloria Stuart died on September 26, 2010, in California, while struggling with lung cancer."
However, the above sentence means that Gloria Stewart died in California. It has a different meaning from birthPlace, a property of the given triple data. Therefore, the pattern generated from the sentence has a problem that it is not suitable for use in generating new triple data having a birthPlace relationship. A process of generating new triple data using the present invention to solve such a problem will be described in detail.
1 is a block diagram illustrating a triple data generation system according to an embodiment of the present invention.
1, a triple
The
The
The
The
Hereinafter, a triple data generating method according to an embodiment of the present invention will be described in detail with reference to FIG. 2 to FIG.
2 is a flowchart illustrating a triple data generation method according to an embodiment of the present invention.
As shown in FIG. 2, the triple data generating method according to an embodiment of the present invention includes a
The
The
Thereafter, the
More specifically, the
Hereinafter, a case will be described in which the
FIG. 3 is a flowchart illustrating a detailed process of generating a pattern based on a knowledge base and a corpus among the triple data generation method of the present invention shown in FIG.
As shown in FIG. 3, the
At least one sentence including subject and object is extracted from the input knowledge base and the corpus (S222), and at least one word containing each of the extracted subject or object is extracted (S223).
The subject search and the object search are extracted based on the search information present in the extracted word (S224).
Subsequently, a predicate present in the extracted sentence is extracted (S225).
Thereafter, a pattern including the extracted subject search, object search, and predicate is generated, or all vocabularies located between the subject and the extracted word are extracted and generated as a pattern (S226).
FIG. 4 is a flowchart illustrating details of a process of extracting a predicate present in a sentence according to FIG.
As shown in FIG. 4, in order to extract a predicate present in a sentence, it is checked whether the extracted sentence predicate exists singularly (S225a).
If there is an odd number of the extracted in-sentence predicates, one predicate present in the sentence is extracted (S225b).
However, if there are a plurality of the in-sentence predicates extracted in step S224, it is possible to extract a predicate expressing a relation between a subject corresponding to the subject search extracted in step S224 of FIG. 3 and an object corresponding to the subject search That is, the dependency relation between the subject corresponding to the subject search and the object corresponding to the object search can be analyzed (S225c).
In step S225d, a dependency tree structure is created based on the dependency information between the subject and the object, which is generated according to the result of the dependency relationship analysis between the subject and the object analyzed as described above.
There may be a plurality of predicate nodes in the generated dependency tree structure. In the dependency tree structure, a predicate node corresponding to the subject and object and a predicate node closest to the object node are selected S225e).
Thereafter, one predicate corresponding to the selected predicate node is extracted (S225f).
The step of generating new triple data based on the learned pattern described above with reference to FIG. 2 will be described in more detail with reference to FIG.
5 is a flowchart showing the detailed steps of generating new triple data based on the learned pattern of FIG.
As shown in FIG. 5, the step of generating new triple data based on the learned patterns in the triple data generation method of the present invention is performed based on the subject-target word dependency information through the execution of step S225d of FIG. 4 In the generated dependency tree structure, a partial tree structure in which a predicate extracted through step S225b or S225f in FIG. 4 is used as a root is extracted (S241).
And selects a node corresponding to the subject and a node corresponding to the object among the nodes existing in the extracted partial tree structure (S242).
The new triple data including the subject, the object corresponding to the selected node and the predicate extracted through the step S225b or the step S225f of FIG. 4 is generated (S243).
The present invention can be applied to self-knowledge learning, wherein the self-knowledge learning is a method of self-generating triple instances having a relationship defined in a knowledge base. This self-knowledge learning consists of a pattern generation process and a knowledge generation process, and is repeatedly performed to generate a triple instance.
As shown in FIG. 6, there are two learning directions in the self-knowledge learning framework, the first of which is a pattern learning process.
The pattern learning process is a process of collecting natural language expressions expressing a specific relationship of a given knowledge base, thereby collecting patterns representing relationships between respective objects and helping to create new triple instances.
The second learning direction is a knowledge learning process, which is a process of creating new triple instances having the same relationship from a given pattern given a pattern of relationships between the objects.
The triple instances thus generated are stored in the knowledge base again and can be utilized as knowledge for learning a new pattern.
The most important part of the self-knowledge learning framework described above is to generate a pattern that can accurately extract the relationship between objects without being too specific. In other words, if there is an error in the generated pattern, it will create false knowledge. As a result, these errors can accumulate continuously and the learning can be performed in a completely wrong direction. Therefore, it is very important to create an accurate pattern that will not generate the error as much as possible.
Particularly, in the present invention, in generating patterns, it is one of great features to generate a pattern suitable for Korean as well as English, a word order is very free and a lot of usage types of vocabulary are utilized. Therefore, in order to apply the pattern generation process of the present invention to Korean, And three tuples for the condition for determining the object and the part of the predicate indicating the relationship between the objects, that is, the relation between subject and object.
In order to generate a pattern, a process of selecting a condition for determining subject and object and a step of selecting a predicate expression may be performed.
First, research information can be used to extract conditions for subject and object discrimination. That is, in the sentence, after selecting the word containing the word corresponding to the subject and the object, extracting the investigation information of each word and building each tuple of the pattern.
Second, the step of selecting a predicate expression is to select a predicate expressing the relation between target objects. In general, the relationship between target objects in a knowledge base is expressed as a predicate in a natural language sentence. Unlike English, the position of a predicate describing a relationship between two target objects is not standardized in Korean. Therefore, if there are a plurality of predicates in the sentence, the predicate expressing the relation of the two target objects is selected.
If there are a plurality of predicates in the sentence, dependency information on the words in the sentence is used. In particular, in the dependency tree structure generated based on the dependency information, a node corresponding to a predicate representing the relationship between two target objects in a sentence is located closest to a node corresponding to the two target objects. That is, through the following expression (1), the predicate in the dependency tree structure
[Equation 1]
In
Hereinafter, the pattern generation process will be described in more detail.
In order to generate a pattern for the relation 'isSpouseOf', the triple data (triangle, isSpouseOf, and triangle) with the relation is given. In order to retrieve a sentence containing such a relation, Yun Sara, who was a peer similar to Yunho, a sentence containing the subject and object of the corresponding triple instance (Yunaro, Jeonghyeonjeong) from the web, I am married to Jung-hyeon-jung, and became a supporter of the younger sister. " The dependency tree structure shown in FIG. 7 can be generated by analyzing the dependency relationship between the target objects on the extracted sentence.
In order to extract the condition for determining the subject and the object from the dependency tree structure, a node 'YunSaro' including the subject and a node 'JungHyeonJungju' including the object are extracted. Accordingly, the subject search is generated first of the patterns including (a), (b), (c), and (c).
Then, in order to select the predicate of the pattern, the two nodes selected above, namely, 'Yoon Sangro' and 'Jeongyeongjeong', are selected as 'Married' corresponding to the nearest predicate node from the node.
As a result, in the result of the above extracted sentence, 'Yun saro, who was a peer similar to Yunho, was married to Jung Hyeon-jong, a young woman of Sejong, .
A new triple instance corresponding to the triple data can be created using the pattern thus generated.
The process of generating the novel triple data of the present invention is performed in the opposite direction to the above-described pattern generation step.
That is, if the generated pattern is given, a sentence containing a word corresponding to the predicate is found in the generated pattern, and a sentence that is a triple data candidate including the relation corresponding to the predicate is searched.
In the second step, a target object having a relationship corresponding to the predicate is extracted.
In order to extract such a target object, the same method as the predicate selection process may be applied in the pattern generation step. In the dependency tree structure, a subject and an object having a relation of a specific predicate are composed of descendants of the corresponding predicate node. Therefore, in order to restrict the candidates of the subject node and the object node, a partial tree structure in which a matching predicate in the dependency tree structure is defined as a root is extracted. Then, in order to finally extract subject and object, nodes corresponding to each condition of the pattern are selected. If there are a plurality of nodes matching the condition of the subject and object, nodes closer to the node corresponding to the predicate can be selected.
FIG. 8 is a diagram illustrating a process of generating new triple data in the relationship of the learned 'isSpouseOf' pattern.
The daughter, Jo Gye-jeong, a sentence containing 'marrying', which corresponds to the predicate of the pattern learned in the previous (and marriage, and) pattern, was married to the son of the independence activist Lee Hoi-young, resulting in Lee Jong-chan. do.
In the extracted sentence, extract the partial tree structure which is the root of the matched predicate 'marriage'. Then, we extract the words 'Kyungjak' and 'Jogyakin', which correspond to the conditions of the following subject and object, and create a triple instance of the new triple data (Kyujak, isSpouseOf, Chojungjin).
Hereinafter, a method of generating triple data and an experimental procedure of a system according to an embodiment of the present invention will be described.
The initial knowledge for the first stage of pattern generation was performed for a total of six relationships related to the person. The instances of each relationship were randomly generated triple data. Table 1 below shows the relationships selected for performance experiments and the statistics of the triple instances provided for each relationship.
[Table 1]
Table 2 below shows an example of a base triple instance used as input for each relationship of the target object.
[Table 2]
In addition, unstructured documents, which are intended for pattern generation and new triple data generation, are targeted at documents existing in Wikipedia. A total of 25,000 documents of Korean Wikipedia were randomly selected. These documents were preprocessed through ETRI 's Natural Language Processing Analysis Tool, and morpheme, parts of speech, object name, and dependency information were used in the analysis. In a document about a person in a wiki document, the text often appears as a pronoun in place of the name of the person in the text. Thus, the extracted triple instance may appear as a pronoun rather than a proper noun. Since these triple instances can not have value as information, they must be restored to appropriate proper nouns. However, in this experiment, we simply applied the method of replacing the pronouns representing the same person with the title of the corresponding wiki document .
In this experiment, we did not utilize the parse tree information for performance comparison, but compared with methods for selecting predicates based on word order only.
The comparison method finds a noun phrase corresponding to the subject and object of the given instance in the sentence, and then selects the verb that appears closest to the word after the two words to generate the pattern. The survey information for subject and object discrimination was applied to both methods.
FIG. 9 is a graph showing the number of generated patterns for each relationship generated through the above two methods.
According to the experimental results, it can be seen that the comparison method generates a larger number of patterns than the present invention for the five relationships except for 'born' representing the relationship.
This result can be analyzed as a result of generating a large number of error patterns because the comparison method does not consider the semantic relation at all, but generates the pattern only based on the surface information. Therefore, in order to confirm these analysis results, we measured the accuracy rate of patterns generated by relation.
10 is a graph showing the accuracy of the generated pattern.
According to the evaluation results shown in FIG. 10, the pattern generating method using the present invention generates a small number of patterns but has a much higher accuracy rate.
That is, most of the patterns generated by the comparison method are error patterns, whereas in the present invention, error patterns are excluded because the semantic relation of the sentence is utilized through the dependency relationship analysis result.
Table 3 below shows an example of the patterns generated in the methods to be compared.
[Table 3]
Then, after learning the generated patterns, experiments for generating new triple data, that is, triple instances using the learned patterns, were conducted.
The comparative model extracts the sentences containing the pattern predicates in a manner similar to the pattern generation method, and selects the nearest phrases matching the subject and object conditions in the matched predicate.
11 shows the number of instances generated through the present invention and the comparison method, respectively.
As shown in FIG. 11, similar to the experiment on the pattern generation result performed earlier, it can be seen that the comparison method generates significantly more triple instances than the present invention for all the relations. This is because the comparison method generates a significantly larger number of patterns than the present invention. However, since the accuracy is more important than the number generated as in the case of the knowledge pattern, the triple instances generated are evaluated to check the accuracy rate.
In the case of the comparison method, since the number of generated knowledge is too large to evaluate all the results, the accuracy of the 'make', 'isChildOf', and 'isSpouseOf' is measured by extracting 200 instances from each sample.
According to the evaluation result, it can be seen that the present invention has a much higher accuracy rate than the comparison method. In the case of the comparison method, a large number of instances are generated, but most of them are wrong instances. In particular, it was found that the accuracy of the generated knowledge is very low for the relationships in which the accuracy of the generated patterns is low.
Accordingly, experiments have been conducted to confirm that the low recall rate can be overcome when the present invention actually undergoes an iterative learning process of the self-knowledge learning framework. For this, pattern generation and knowledge generation are performed by reusing knowledge generated through the present invention.
FIGS. 13 to 14 are graphs showing patterns and knowledge numbers generated through an iterative process in comparison with previous results. FIG.
According to the experimental results, it can be seen that the number of patterns and knowledge generated in each step of the present invention is small, but it is possible to overcome a sufficiently low recall rate through an iterative process. From these results, And the system can be usefully applied to self - knowledge learning.
Thereafter, the pattern generated in the above-described pattern generation step of the present invention does not include the subject search, the object search and the predicate but is located between the subject and the object in the natural language sentence of the corpus, The case where at least one vocabulary is represented will be described in detail.
In the case where the
FIG. 15 is a flowchart showing sub-steps of extracting and learning pattern candidates for each lexeme indicating a relationship between the subject and object among patterns in which the pattern learning unit is generated.
(S231) at least one of a predicate indicating a relation between the subject and the object among at least one vocabulary located between the subject in the extracted sentence and the object, or a vocabulary indicating the label, the identifier or the attribute in the sentence. The property generated at this time represents the relationship of pairs of objects shown together in one sentence.
Next, the similarity degree between the pattern and the property is measured (S232). That is, the vector similarity can be calculated between the pattern and the property mapped to the word embedding space, respectively. At this time, the word embedding space maps a plurality of vocabularies into a vector space of N dimensions (where N is a natural number), and expresses them in a distributed representation.
An error pattern is determined based on the result of the similarity measurement between the pattern and the property (S233), and the determined error pattern is removed (S234). For example, a pattern similarity degree value lower than the reference semantic similarity value may be determined as an error pattern by comparing the result of measurement of semantic similarity between the pattern and the property with a predetermined reference semantic similarity value, and the determined error pattern may be removed .
FIG. 16 is a flow chart showing detailed steps for calculating the semantic similarity between the pattern and the property in FIG.
As shown in FIG. 16, it is checked whether the pattern and the property are in different languages (S232a).
For example, when the pattern and the property are in different languages, for example, when the pattern is in Korean and the property is in English, based on predetermined lexical pairs having the same meaning as the pattern and property The projection matrix is learned so that the correlation coefficient between the pattern and the property is high, and the pattern and property are projected into the same word embedding space (S232b).
Then, it is checked whether the pattern and the property are constituted by one word (S232c).
If the pattern and the property are each composed of one word, the similarity degree between the pattern and the property is calculated (S232d).
However, if the pattern and the property are composed of a plurality of words or a plurality of words, an average vector of the elements constituting the pattern and the property is defined (S232e).
The vector similarity is calculated between the pattern and the property based on the defined average vector (S232f).
Hereinafter, a pattern generation process among the methods for generating triple data of the present invention will be described in detail.
First, patterns and properties generated from the knowledge base and the corpus used in one embodiment of the present invention may be in different languages. For example, patterns are constructed in Korean because they are generated from Korean sentences, while most properties can be expressed in English.
Therefore, the word embedding spaces of patterns and properties in different languages are learned independently, making it difficult to directly measure their semantic similarity. To do this, independently learned word embedding spaces are projected to the same low dimension using Canonical Correlation Analysis method, so that the degree of similarity between patterns and properties constructed in different languages can be calculated.
According to the experimental results performed using DividePedia and Wikipedia among the knowledge bases, the reliability similarity-based reliability measurement process used in the present invention contributes to more accurate triple data generation than the conventional statistical-based reliability measurement methods Able to know. As a result of evaluating the upper 2,000 triple data according to each reliability value, it can be seen that the present invention has better performance than the statistical method.
In the present invention, pattern generation, pattern filtering, and new triple data generation are performed through the procedure shown in FIG. We study pattern candidates for each property for use in extracting new triple data using given knowledge base and corpus (Corpus). At this time, the sentence including the subject and object word of the specific triple data is classified into the subject of the given triple data and the object corresponding to the object object using the distant supervision assumption that the meaning of the property of the triple data constituting the corpus is expressed Extracts sentences, and extracts pattern candidates using these sentences.
However, in this case, the learned patterns may contain a large number of errors due to the limit of the distant supervision assumption, which may reduce the accuracy of the final triple data. In order to solve this problem, the error pattern is removed by filtering the generated pattern. Finally, new triple data can be generated by using the filtered pattern and corpus, and the knowledge base and corpus can be expanded by this.
Generally, since the part corresponding to the property of the triple data is expressed by the sentence in the sentence, in the conventional English study, the pattern is generated by using the vocabularies contained in the subject and subject objects of the given triple data. However, unlike English, it is not appropriate to use a vocabulary between subject words given that the predicate is given and is not located between the objects. Also, because Korean is free of word order unlike English, it is difficult to find a predicate expressing the relation between two objects.
Therefore, in the present invention, all vocabularies existing between the subject word and object vocabulary nodes are used as a pattern based on the dependency tree. Thus, only the vocabulary expressing the relationship between the two objects corresponding to the subject and object vocabulary can be extracted, and the remaining vocabularies can be removed.
18 is a diagram illustrating a dependency tree-based pattern generation process.
As shown in Figure 18, an example of the pattern generated from the example sentence "Gloria Stewart died on September 26, 2010 at his home in California, during his illness with lung cancer. &Quot;
Distant supervision The most important thing in the home-based self-knowledge learning process is to effectively filter the patterns generated from sentences that do not contain the meaning of the given triple data. Accordingly, in the present invention, in order to overcome the limit of the statistical-based reliability function, a method of directly measuring the similarity degree of the pattern and the target property and using the method as the reliability is used.
As mentioned above, a pattern expresses meaning by a set of vocabulary extracted from a sentence, and a property represents a meaning through attributes such as a label and an identifier. Therefore, by measuring the semantic similarity between the pattern and the property vocabularies, it is possible to measure how appropriate each pattern is for the target property. These patterns
&Quot; (2) "
In order to measure the semantic similarity of vocabulary, we use the word embedding method which is a non-ambiguity learning method using large-scale data. The word embedding method is a method of mapping a vocabulary to a vector space of N dimensions (where N is a natural number) and expressing it in a distributed representation. The word embedding space is a space in which vocabularies are mapped . Using air information between words using a large amount of corpus, we can see that semantically similar words are learned in a similar form of vector in the learned word embedding space. Through this, the semantic similarity between vocabularies can be measured by obtaining the vector similarity degree of each vocabulary mapped to the word embedding space. The semantic similarity between patterns and properties through word embedding is expressed as vector expressions in each word embedding space
&Quot; (3) "
The vector similarity can be calculated by the cosine similarity as shown in Equation (4) below.
&Quot; (4) "
When the pattern and the property are constituted by one word, the similarity degree of the two vocabulary vectors is calculated using Equation (4).
But patterns and properties can consist of more than one vocabulary. For example, a pattern can be composed of several words, and in the case of properties it is represented by a single word, but sometimes more than two words, such as hasChild, are combined.
Word-embedding learns vectors by vocabulary, so vocabulary combining two or more words can not be expressed as a vector. Accordingly, in the present invention, a vector of a pattern and a property composed of two or more words is defined as an average vector of the elements. For example, if the pattern p consists of n words
&Quot; (5) "
&Quot; (6) "
The vector of the pattern and the property can be obtained through
&Quot; (7) "
In other words, while patterns generated in Korean sentences are composed of Korean vocabularies, most of the widely used knowledge base properties such as divipedia and wikipedia are mostly expressed in English. For this reason, the pattern requires a word embedding space learned from Korean corpus, and a property requires English word embedding space learned from English corpus. Since these two spaces are learned from independent corpus, the final learned word embedding space is also independent of each other. In addition, these two spaces can be learned differently in the number of dimensions, and even if the number of dimensions is the same, Korean and English vocabularies with the same meaning can be learned as completely different vectors.
FIG. 19 shows three pairs of vocabulary vectors having the same meaning in the independent Korean and English word embedding spaces. It can be seen from FIG. 19 that the vocabularies having the same meaning are learned with completely different types of vectors.
In this way, the pattern similarity and the property vector extracted from the independently-learned word embedding space can not properly measure the semantic similarity between the two.
Therefore, in order to solve such a problem, the present invention can measure semantic similarity using a heterogeneous language word embedding space projection method based on canonical correlation analysis. That is, the projection matrices are increased so that the correlation coefficients of the vocabulary pairs having the same meaning given in advance are increased, and the vectors of the heterogeneous space are projected to the same space. In the projection space thus obtained, heterogeneous language vectors having the same meaning appear in a similar form, so that the similarity of words in different languages can be measured.
In other words,
&Quot; (8) "
The matrix obtained by canonical correlation analysis
&Quot; (9) "
FIG. 20 shows a heterogeneous language word embedding space projected through the above process. The word embedding space shown in FIG. 20 can confirm that vocabularies having the same meaning are arranged close to each other, unlike the independent word embedding space shown in FIG. In order to measure the degree of similarity between the pattern and the property vector projected in the low dimension, Equations (5) to (6) are changed to Equation (10).
&Quot; (10) "
Hereinafter, performance test results of the features of the present invention described above are performed.
In the present invention, pattern learning and triple generation are performed for performance evaluation of the above-described semantic similarity measurement process. Korean language divide was used as a seed knowledge base, and Korean Wikipedia was used as a corpus for a pattern generation and a new triple data generation. We used ETRI language analyzer for natural language processing of Wikipedia.
Table 3 below shows a brief statistical data of the knowledge base and the corpus used in the pattern and triple generation experiments.
[Table 3]
In order to measure the above-described semantic similarity of the present invention, learning of the Korean and English word embedding spaces is required. To do this, we used the Korean and English Wikipedia to learn the word embedding space for each language. The word embedding space learning was performed using open source word2vec. Table 2 below shows a summary of the corpus used in the word embedding learning and the simplified statistics of the learned word embedding space.
[Table 4]
If the patterns and properties are more than two words, they are separated to separate them. In the case of Korean, they are divided into spaces. In the case of properties, vocabularies are separated based on a predefined simple rule.
The reliability measurement method using the semantic similarity according to the present invention is compared with the reliability measurement method based on the statistical information used in the prior art.
The reliability of statistical information based on each pattern was used for each property. In addition to reliability measurement, pattern generation and triple data generation are all performed in the same manner.
When triple data is generated using the whole data of dividea and wikipedia, it is difficult to evaluate all the triple data because it is too much. In the present invention, Top-K accuracy is measured for performance comparison of two reliability functions. To do this, triple data is generated with all the patterns without performing pattern filtering first. Then, the reliability of all patterns is measured using each reliability measurement function, and then the patterns are sorted based on the reliability value. Thus, for the patterns arranged by the respective reliability, the upper K pieces of the triple data generated by the respective patterns are extracted. The accuracy of the triple data thus extracted was checked by checking whether it was true or not.
A total of 25,784 patterns and 422,733 triple data were generated for Divipedia and Korean Wikipedia. The performance comparison of the two methods was performed by extracting the top 2,000 triples from the data measured by each reliability method and manually evaluating the accuracy of the triples.
Figure 21 shows the accuracy of Top-K triple data.
As shown in FIG. 21, it can be seen that the performance difference is about 10% in all intervals except for Top-200, compared with the conventional frequency-based reliability measurement method. Finally, when the top 2,000 triple data are extracted, the present invention shows an accuracy of about 71%, whereas the conventional frequency-based reliability measurement method shows a performance of about 62% and a performance difference of 9%.
The reason for the performance degradation of the Top-200 can be judged to be that the learned semantic relationship in the word embedding space is not always synonymous. A suitable pattern is a pattern that indicates the meaning of a property, so it should be expressed in vocabularies with synonyms. In general, the corpus-based methods show excellent performance in finding the thesaurus, but the word-based learning method based on corpus can learn a completely different embedding space depending on the context, and the similarity between the vocabularies is completely different It is possible.
In fact, if we look at the embedded English word embedding, we can confirm that spouse and grandparent are learned to have very similarity value. These two vocabularies can be interpreted as a similar vocabulary set of family vocabulary and have a high degree of similarity, but in fact they have a completely different meaning. It is considered that the performance of the proposed method is degraded due to these features of word embedding.
According to these experimental results, it can be seen that although it has a limitation on the similarity measurement based on the word embedding, it contributes greatly to generating more accurate triple data. This implies that it is more appropriate to reflect the direct semantic similarity of patterns and properties to the pattern reliability measurement than to the statistical based indirect measurement method.
According to an aspect of the present invention, data of a document expressed in a natural language in a semantic web field can be easily and quickly structured, and the data processing speed of a computer can be improved.
According to another aspect of the present invention, it is possible to structure Korean data and improve the satisfaction of Korean users.
Embodiments of the present invention may be implemented in the form of program instructions that can be executed on various computer means and recorded on a computer readable medium. The computer-readable medium may include program instructions, data files, data structures, and the like, alone or in combination. The program instructions recorded on the medium may be those specially designed and configured for the present invention or may be available to those skilled in the art of computer software. Examples of computer-readable media include magnetic media such as hard disks, floppy disks and magnetic tape, optical media such as CD-ROMs and DVDs, magnets such as floptical disks, Examples of program instructions, such as magneto-optical and ROM, RAM, flash memory and the like, can be executed by a computer using an interpreter or the like, as well as machine code, Includes a high-level language code. The hardware devices described above may be configured to operate as at least one software module to perform operations of one embodiment of the present invention, and vice versa.
It will be apparent to those skilled in the relevant art that various modifications, additions and substitutions are possible, without departing from the spirit and scope of the invention as defined by the appended claims. The appended claims are to be considered as falling within the scope of the following claims.
100: Triple data generation system
10: Database
120: pattern generator
140: pattern learning unit
160: Triple generating unit
180:
Claims (22)
Receiving a knowledge base and a corpus composed of triple data including a subject word and an object word in a natural language sentence, and generating a pattern based on the received knowledge base;
Extracting and learning a pattern candidate for each vocabulary representing a relationship between the subject and the object among the patterns in which the pattern learning unit is generated; And
Generating new triple data based on the triple generator-learned pattern;
, ≪ / RTI &
The pattern
Wherein the at least one vocabulary includes at least one vocabulary that is included in the natural language sentence and includes a subject word search, an object word search, and a predicate, or located between a subject and an object in the natural language sentence.
Wherein the pattern generating unit receives a knowledge base and a corpus composed of triple data including subject and object in a natural language sentence and generates a pattern based on the received knowledge base and corpus
Receiving a knowledge base and a corpus composed of triple data including subject and object in the natural language sentence;
Extracting at least one sentence including a subject and an object from the knowledge base and the corpus;
Extracting at least one word phrase including an extracted subject or object, respectively;
Extracting a subject search and an object search based on survey information present in the extracted word;
Extracting a predicate present in the extracted sentence; And
Extracting at least one vocabulary located between the subject and the extracted word to generate a pattern including the extracted subject search, object search, and predicate;
And generating the triple data.
The step of extracting the predicate present in the extracted sentence
And extracting a predicate expressing a relationship between a subject corresponding to the extracted subject search and an object corresponding to the object searching if there are a plurality of the in-sentience predicates extracted.
The step of extracting the predicate present in the extracted sentence
Analyzing a dependency relationship between a subject corresponding to the subject search and an object corresponding to the subject search, and extracting a predicate according to the analysis result.
The step of extracting the predicate present in the extracted sentence
A dependency tree structure is generated based on the dependency information between the subject and the object, and a subject node and a subject node corresponding to the subject and object among a plurality of predicate nodes existing in the generated dependency tree structure Selecting one predicate node to be located, and extracting a predicate corresponding to the selected predicate node.
The step of generating new triple data based on the learned pattern of the triple generator
Extracts a partial tree structure based on a predicate extracted from the dependency tree structure generated on the basis of the dependency information between the subject and the object, selects each node corresponding to the subject and object among the extracted partial tree structure, New triple data including a subject, an object, and an extracted predicate corresponding to a node are generated.
The step of extracting and learning a pattern candidate for each lexeme indicating a relationship between the subject and the object among the patterns generated by the pattern learning unit
And removing the error pattern from the learned pattern candidates.
The step of extracting and learning a pattern candidate for each lexeme indicating a relationship between the subject and the object among the patterns generated by the pattern learning unit
Generating at least one of a predicate expressing a relation between the subject and the object among at least one vocabulary located between the subject in the extracted sentence and the object or a vocabulary representing the label, the identifier or the attribute in the sentence as a property;
Measuring semantic similarity between the pattern and the property; And
Determining an error pattern based on a result of measurement of semantic similarity between the pattern and the property, and removing the determined error pattern;
And generating the triple data.
The step of measuring semantic similarity between the pattern and the property
And a vector similarity degree is calculated between the pattern and the property mapped to the word embedding space, respectively.
The step of measuring semantic similarity between the pattern and the property
Learning the projection matrix so that the correlation coefficient between the pattern and the property is high based on predetermined lexical pairs having the same meaning as the pattern and property when the pattern and the property are in different languages, And projecting the image into the embedding space.
Wherein the pattern is in Korean, and the property is in English.
The step of measuring semantic similarity between the pattern and the property
And calculating a cosine similarity between the pattern and the property when the pattern and the property each have a single vocabulary.
The step of measuring semantic similarity between the pattern and the property
When the pattern and the property are composed of a plurality of words or a plurality of words, an average vector of the elements constituting the pattern and the property is defined, and the vector similarity is calculated between the pattern and the property based on the defined average vector / RTI >
The word embedding space
Wherein a plurality of vocabularies are mapped to a vector space of N dimensions (where N is a natural number) and expressed as a Distributed Representation.
Storing a pattern generated from the pattern generation unit and new triple data generated from the triple generation unit;
And generating the triple data.
A pattern generator for receiving a knowledge base and a corpus composed of triple data including a subject and an object in a natural language sentence and generating a pattern based on the received knowledge base and corpus;
A pattern learning unit for extracting and learning a pattern candidate for each lexicon indicating a relationship between the subject and the object among the generated patterns; And
A triple generator for generating new triple data based on the learned pattern;
, ≪ / RTI &
The pattern
Wherein the at least one vocabulary indicates at least one vocabulary that includes subject search, object search, and predicate present in the natural language sentence or located between a subject and an object in the natural language sentence.
The pattern generation unit
Wherein a dependency relationship between a subject corresponding to the subject search and an object corresponding to the subject search is analyzed and a predicate is extracted according to the analysis result.
The pattern generation unit
A dependency tree structure is generated based on the dependency information between the subject and the object, and a subject node and a subject node corresponding to the subject and object among a plurality of predicate nodes existing in the generated dependency tree structure Selects a predicate node to be located, and extracts a predicate corresponding to the selected predicate node.
The pattern learning unit
The method comprising the steps of: generating at least one of a predicate expressing a relation between the subject and an object among at least one vocabulary located between an extracted subject and an object and a lexicon representing a label, an identifier or an attribute in the sentence as a property; Determining a similarity measure, determining an error pattern based on a result of measuring a similarity degree between the pattern and the property, and eliminating the determined error pattern.
The triple-
Extracts a partial tree structure based on a predicate extracted from the dependency tree structure generated on the basis of the dependency information between the subject and the object, selects each node corresponding to the subject and object among the extracted partial tree structure, And generates new triple data including a subject, an object, and an extracted predicate corresponding to each node.
A storage unit for storing a pattern generated by the pattern generation unit and new triple data generated from the triple generation unit;
And generating the triple data.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| KR1020160009168A KR101768852B1 (en) | 2016-01-26 | 2016-01-26 | Generating method and system for triple data |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| KR1020160009168A KR101768852B1 (en) | 2016-01-26 | 2016-01-26 | Generating method and system for triple data |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| KR20170089142A true KR20170089142A (en) | 2017-08-03 |
| KR101768852B1 KR101768852B1 (en) | 2017-08-17 |
Family
ID=59655624
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| KR1020160009168A KR101768852B1 (en) | 2016-01-26 | 2016-01-26 | Generating method and system for triple data |
Country Status (1)
| Country | Link |
|---|---|
| KR (1) | KR101768852B1 (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2019212289A (en) * | 2018-06-05 | 2019-12-12 | ベイジン バイドゥ ネットコム サイエンス テクノロジー カンパニー リミテッド | Method and device for generating information |
| CN110674637A (en) * | 2019-09-06 | 2020-01-10 | 腾讯科技（深圳）有限公司 | Character relation recognition model training method, device, equipment and medium |
| KR102111733B1 (en) * | 2018-11-29 | 2020-05-15 | 주식회사 솔트룩스 | System and method for integrating knowledge based on data reinforcement |
| KR20200094627A (en) * | 2019-01-30 | 2020-08-07 | 베이징 바이두 넷컴 사이언스 앤 테크놀로지 코., 엘티디. | Method, apparatus, device and medium for determining text relevance |
| WO2021054512A1 (en) * | 2019-09-18 | 2021-03-25 | 주식회사 솔트룩스 | System and method for reinforcing knowledge base |
| CN112749549A (en) * | 2021-01-22 | 2021-05-04 | 中国科学院电子学研究所苏州研究院 | Chinese entity relation extraction method based on incremental learning and multi-model fusion |
| CN113127669A (en) * | 2020-01-15 | 2021-07-16 | 百度在线网络技术（北京）有限公司 | Advertisement matching method, device, equipment and storage medium |
| KR20220063005A (en) * | 2020-11-09 | 2022-05-17 | 경희대학교 산학협력단 | Apparatus for generating patient health summary and method thererof |
| CN116129321A (en) * | 2023-02-17 | 2023-05-16 | 哈尔滨工业大学 | Video description generation method based on long-order dynamic host-object visual relationship modeling |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR102358048B1 (en) | 2017-12-28 | 2022-02-03 | 주식회사 엔씨소프트 | Apparatus and method for generating storyline from documents |
| KR102469712B1 (en) | 2018-02-22 | 2022-11-22 | 삼성전자주식회사 | Electronic device and Method for generating Natural Language thereof |
| KR20220134218A (en) | 2021-03-26 | 2022-10-05 | 한국전자통신연구원 | Method and apparatus for recognizing spo tuple relationship based on deep learning |
-
2016
- 2016-01-26 KR KR1020160009168A patent/KR101768852B1/en active IP Right Grant
Cited By (16)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11494420B2 (en) | 2018-06-05 | 2022-11-08 | Beijing Baidu Netcom Science And Technology Co., Ltd. | Method and apparatus for generating information |
| KR20190138562A (en) * | 2018-06-05 | 2019-12-13 | 베이징 바이두 넷컴 사이언스 앤 테크놀로지 코., 엘티디. | Method and apparatus for information generation |
| JP2019212289A (en) * | 2018-06-05 | 2019-12-12 | ベイジン バイドゥ ネットコム サイエンス テクノロジー カンパニー リミテッド | Method and device for generating information |
| KR102111733B1 (en) * | 2018-11-29 | 2020-05-15 | 주식회사 솔트룩스 | System and method for integrating knowledge based on data reinforcement |
| KR20200094627A (en) * | 2019-01-30 | 2020-08-07 | 베이징 바이두 넷컴 사이언스 앤 테크놀로지 코., 엘티디. | Method, apparatus, device and medium for determining text relevance |
| CN110674637B (en) * | 2019-09-06 | 2023-07-11 | 腾讯科技（深圳）有限公司 | Character relationship recognition model training method, device, equipment and medium |
| CN110674637A (en) * | 2019-09-06 | 2020-01-10 | 腾讯科技（深圳）有限公司 | Character relation recognition model training method, device, equipment and medium |
| KR20210033345A (en) * | 2019-09-18 | 2021-03-26 | 주식회사 솔트룩스 | System and method for consolidating knowledge base |
| WO2021054512A1 (en) * | 2019-09-18 | 2021-03-25 | 주식회사 솔트룩스 | System and method for reinforcing knowledge base |
| CN113127669A (en) * | 2020-01-15 | 2021-07-16 | 百度在线网络技术（北京）有限公司 | Advertisement matching method, device, equipment and storage medium |
| CN113127669B (en) * | 2020-01-15 | 2023-10-10 | 百度在线网络技术（北京）有限公司 | Advertisement mapping method, device, equipment and storage medium |
| KR20220063005A (en) * | 2020-11-09 | 2022-05-17 | 경희대학교 산학협력단 | Apparatus for generating patient health summary and method thererof |
| CN112749549A (en) * | 2021-01-22 | 2021-05-04 | 中国科学院电子学研究所苏州研究院 | Chinese entity relation extraction method based on incremental learning and multi-model fusion |
| CN112749549B (en) * | 2021-01-22 | 2023-10-13 | 中国科学院电子学研究所苏州研究院 | Chinese entity relation extraction method based on incremental learning and multi-model fusion |
| CN116129321A (en) * | 2023-02-17 | 2023-05-16 | 哈尔滨工业大学 | Video description generation method based on long-order dynamic host-object visual relationship modeling |
| CN116129321B (en) * | 2023-02-17 | 2023-10-27 | 哈尔滨工业大学 | Video description generation method based on long-order dynamic host-object visual relationship modeling |
Also Published As
| Publication number | Publication date |
|---|---|
| KR101768852B1 (en) | 2017-08-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| KR101768852B1 (en) | Generating method and system for triple data | |
| US10025819B2 (en) | Generating a query statement based on unstructured input | |
| US9727553B2 (en) | System and method for generating and using user semantic dictionaries for natural language processing of user-provided text | |
| RU2607975C2 (en) | Constructing corpus of comparable documents based on universal measure of similarity | |
| CN105988990B (en) | Chinese zero-reference resolution device and method, model training method and storage medium | |
| US9588962B2 (en) | System and method for generating and using user ontological models for natural language processing of user-provided text | |
| US9892111B2 (en) | Method and device to estimate similarity between documents having multiple segments | |
| KR101629415B1 (en) | Method for detecting grammar error and apparatus thereof | |
| CN102253930B (en) | A kind of method of text translation and device | |
| US9600469B2 (en) | Method for detecting grammatical errors, error detection device for same and computer-readable recording medium having method recorded thereon | |
| EP3014481A2 (en) | Methods and apparatuses for mining synonymous phrases, and for searching related content | |
| US8812504B2 (en) | Keyword presentation apparatus and method | |
| Atia et al. | Increasing the accuracy of opinion mining in Arabic | |
| CN110019820B (en) | Method for detecting time consistency of complaints and symptoms of current medical history in medical records | |
| CN109783806A (en) | A kind of text matching technique using semantic analytic structure | |
| TW201826145A (en) | Method and system for knowledge extraction from Chinese corpus useful for extracting knowledge from source corpuses mainly written in Chinese | |
| KR20140051606A (en) | Method and apparatus for coherence evaluation using rdf-based sentence ontology | |
| CN114706951A (en) | Temporal knowledge graph question-answering method based on subgraph | |
| JP2005122231A (en) | Screen display system and screen display method | |
| JP5812534B2 (en) | Question answering apparatus, method, and program | |
| Angeli et al. | Stanford’s distantly supervised slot filling systems for KBP 2014 | |
| KR101358614B1 (en) | Korean Morphological Analysis Apparatus and Method based on Tagged Corpus | |
| KR101983477B1 (en) | Method and System for zero subject resolution in Korean using a paragraph-based pivotal entity identification | |
| Garrido et al. | UNED Slot Filling and Temporal Slot Filling systems at TAC KBP 2013: System description. | |
| KR101489606B1 (en) | Method for detecting inconsistent code identifier and medium recording the same |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| E701 | Decision to grant or registration of patent right | ||
| GRNT | Written decision to grant |