KR20200071689A - Apparatus and method for named entity disambiguation based on rdf knowledge base - Google Patents

Abstract

A method for overcoming named entity disambiguation based on an RDF knowledge base comprises the steps of: selecting words recognized as entities among words of an input sentence; determining an entity candidate group for each of the selected words based on data learning representing a relationship between the entities using a knowledge base; constructing combination sets in which candidates are arbitrarily combined in consideration of the number of all cases from each of the individual candidate groups; and determining meaning of named entity of each of the selected words by calculating a ranking of the similarity of each of the combination sets.

Description

Method and apparatus for resolving object name neutrality based on RDF knowledge base {APPARATUS AND METHOD FOR NAMED ENTITY DISAMBIGUATION BASED ON RDF KNOWLEDGE BASE}

The present invention relates to a method and apparatus for resolving neutrality of an entity name based on a Resource Description Framework (RDF) knowledge base designed to be applied to a question-and-answer system, specifically, a word in which an entity name having two or more meanings is simultaneously used in a sentence It relates to a method and apparatus for resolving object name neutrality based on the RDF knowledge base by a comprehensive interdependence pair linking approach for linking to the ones that are related to the meaning.

With the recent development of Internet and computing technologies, the evolution of mobile devices and sensors, and the emergence of networks, the amount of information is rapidly increasing. Accordingly, various studies are being conducted to find necessary information among increasing information. Researches on object name recognition, which is a field of information extraction, and linking recognized object names to specific objects are actively being attempted to extract meaningful knowledge from vast information. Entity Linking is the task of mapping the name of an entity appearing in text to a specific entry in a knowledge base such as Wikipedia.

Various individual names are used in natural language expression. The names of entities that exist in reality, such as people, organizations, places, and products, are called entity names. One entity name can have various meanings.

Therefore, the neutrality problem of such an individual name is a very challenging subject in the field of natural language processing, and is one of difficult problems.

The present invention has been devised in view of the above-mentioned necessity, and the RDF knowledge base is based on a comprehensive interdependence pair linking approach for linking two or more meanings of entity names to the meanings of words used simultaneously in a sentence. It is an object of the present invention to provide a method and apparatus for resolving neutrality based on object names.

The technical problem to be achieved by the present invention is not limited to the technical problems mentioned above, and other technical problems that are not mentioned will be clearly understood by those skilled in the art from the following description. Will be able to.

In order to achieve the above object, according to an aspect of the present invention, as a method of resolving the neutrality of an entity name based on a knowledge base, selecting words recognized as entities among words in an input sentence, entities and entities using the knowledge base Determining an individual candidate group for each of the selected words based on data learning expressing a relationship between them, constructing a combination set in which candidates are arbitrarily combined considering the number of all cases from each of the individual candidate groups, and the combination set A method of resolving the neutrality of an individual name is provided, comprising calculating the ranking of each similarity and determining the meaning of the individual name of each of the selected words.

According to an embodiment of the present invention, the step of selecting words recognized as an object among words in the input sentence may include selecting words that are nouns, verbs or adjectives among words in the input sentence. .

According to an embodiment of the present invention, the step of determining an individual candidate group of each of the selected words may include matching an individual type of each of the selected words as a candidate group while matching each of the selected words. And determining.

According to an embodiment of the present invention, the object type may be any one of instances, classes, and properties.

According to an embodiment of the present invention, the similarity of each combination set is calculated based on the following [Equation],

[Mathematics]

는 u 노드셋의 i 번째 노드,

는 ν 노드셋의 j 번째 노드를 나타낼 수 있다.Here, V is the number of node sets that are a set of candidate objects for a word,

Is the i th node of the u nodeset,

May denote the j th node of the ν node set.

According to another aspect of the present invention, as a knowledge base-based object name neutralization resolution device, a word recognized as an object among words in an input sentence is selected, and data learning expressing a relationship between an object and an object using the knowledge base An individual candidate group of each of the selected words is determined based on the result, and a combination set of candidates is arbitrarily considered in consideration of the number of cases in each of the individual candidate groups, and the ranking of the similarity of each combination set is calculated. An apparatus for resolving deuteria of an individual name is provided, comprising one or more processors for determining the meaning of the individual name of each of the selected words.

According to an embodiment of the present invention, the one or more processors may select words that are nouns, verbs, or adjectives among words in the input sentence.

According to an embodiment of the present invention, while the one or more processors are matched with each of the selected words, an object matching the individual type of each of the selected words may be determined as a candidate group.

The apparatus and method proposed in the present invention can efficiently resolve an object name neutrality using an embedded vector by generating an object name neutrality resolution model based on a knowledge base-based approach that considers comprehensive interdependence. In addition, by using a comprehensive interdependence approach that utilizes whole words, it is possible to obtain better overall performance than the conventional approach that considers each word's interdependence. In addition, among the approaches considering comprehensive interdependence, better performance can be obtained than applying the individualized page rank algorithm using the pair linking approach. Through this, the method and apparatus proposed in the present invention can be applied to a question-and-answer system and an information extraction system as a method of resolving the neutrality of entity names based on the RDF knowledge base.

The effects obtained in the present invention are not limited to the above-mentioned effects, and other effects not mentioned may be clearly understood by those skilled in the art from the following description. will be.

1 is a conceptual diagram of an apparatus for resolving neutrality of entity names based on a knowledge base according to an embodiment of the present invention.
2 illustrates an application example of an Entity Transition Probabilities (ETP) algorithm according to an embodiment of the present invention.
3 is a flowchart of a method for resolving neutrality of an entity name based on a knowledge base according to an embodiment of the present invention.

Advantages and features of the present invention, and methods for achieving them will be clarified with reference to embodiments described below in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but may be implemented in various different forms, and only the embodiments allow the disclosure of the present invention to be complete, and common knowledge in the technical field to which the present invention pertains. It is provided to completely inform the person having the scope of the invention, and the present invention is only defined by the scope of the claims. The same reference numerals refer to the same components throughout the specification.

In the description of the embodiments of the present invention, when it is determined that a detailed description of known functions or configurations may unnecessarily obscure the subject matter of the present invention, the detailed description will be omitted, and the terms described below will be described in the embodiments of the present invention. These terms are defined in consideration of the function of the user, and may vary depending on the user's or operator's intention or custom. Therefore, the definition should be made based on the contents throughout this specification.

Combinations of each block in the accompanying block diagrams and steps in the flow charts may be performed by computer program instructions (execution engines), which are executed by a general purpose computer, special purpose computer, or other programmable data processing equipment processor. Since it can be mounted, its instructions, which are executed through a processor of a computer or other programmable data processing equipment, create a means to perform the functions described in each block of the block diagram or in each step of the flowchart.

These computer program instructions can also be stored in computer readable or computer readable memory that can be oriented to a computer or other programmable data processing equipment to implement a function in a particular way, so that computer readable or computer readable memory The instructions stored in it are also possible to produce an article of manufacture containing instructions means for performing the functions described in each block of the block diagram or in each step of the flowchart.

And since computer program instructions may be mounted on a computer or other programmable data processing equipment, a series of operation steps are performed on a computer or other programmable data processing equipment to create a process that is executed by the computer to generate a computer or other programmable It is also possible for instructions to perform data processing equipment to provide steps for performing the functions described in each block of the block diagram and in each step of the flowchart.

Further, each block or each step may represent a module, segment, or portion of code that includes one or more executable instructions for executing specified logical functions, and in some alternative embodiments referred to in blocks or steps It should be noted that it is also possible for the functions to be generated out of sequence. For example, two blocks or steps shown in succession may in fact be performed substantially simultaneously, and it is also possible that the blocks or steps are performed in the reverse order of the corresponding function as necessary.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. However, the embodiments of the present invention exemplified below may be modified in various other forms, and the scope of the present invention is not limited to the embodiments described below. Embodiments of the present invention are provided to more fully describe the present invention to those skilled in the art.

Named Entity Disambiguation is the task of linking the name of an entity appearing in text to an appropriate entity in the knowledge base. This object name neutralization is used in fields such as a question-and-answer system and an information extraction system. For example,'Leon' used in the question,'Is Matilda's role in film Leone Natalie Portman?', means'Leon'. By contrast, "the song of Leon released in 2015, IU and Park Myeong-su sang that? 'The question' Leon 'in the sense that the song' Leon 'released in 2015. As described above, the meaning of the individual name is determined to be related to the meaning of the words used in the sentence at the same time.

In the present invention, a model was created based on a knowledge base in the form of Resource Description Framework (RDF) to resolve the neutrality of an individual name. Then, using the generated model, ranking was attempted to resolve the neutrality of individual names, and the results were compared.

In the present invention, an Entity2vec model using Word2vec was generated as a model for resolving the neutrality of an entity name. Word2vec, a study published by Google in 2013, is one of the word embedding methodologies for vectorizing words. In the present invention, the learning technique of the Word2vec model was used to generate the Entity2vec model. The Entity2vec model is trained on the basis of the association between each entity in the knowledge base and consists of embedding vectors for each entity. Therefore, the learning data of Entity2vec, unlike Word2vec, trained the RDF knowledge base, not sentences. Although there are words necessary to understand the meaning of the object name in the natural language sentence, unnecessary words may also be included because it is a natural language. However, the RDF knowledge base is constructed based on the relationship with other entities in order to grasp the meaning of the corresponding entity name, thereby compensating for these shortcomings of natural language sentences. Therefore, in the present invention, in the model generation, the RDF knowledge base, not the sentence, was trained to create an Entity2vec model with embedding information for each entity name.

Pairwise based method, which is a representative methodology for resolving the identity of individuals, is a method of resolving the identity of individuals by using the correlation of each other when there are two or more individuals in a sentence. That is, it is a method of linking two different entity names, assigning weights, and selecting the pair with the highest weight. Although this method considers interdependence between the individual names appearing in the same sentence, there is a limitation that global interdependence cannot be considered because only two pairs are connected and used. For example,'Does the length of the age of 10 include pine trees?' 'Pine', a Korean seven-membered female music group in Korea, because the traditional pairing approach to'pine', which is the name of the neutrality, does not use the information of'kind', but the previous word'kind'. You will find In the present invention, in order to compensate for the limitations of such a pair linking approach, a pair linking approach was devised based on comprehensive interdependencies, and compared with a traditional pair linking approach through implementation and experimentation.

As described above, in the present invention, a comprehensive interdependence based pairing approach is used to overcome the limitations of the conventional pairing approach. In addition, the personalized page rank algorithm using the concept of comprehensive interdependence was compared with the algorithm presented in the present invention. The page rank algorithm is an algorithm that assigns weights according to the relative importance of other pages connected to the target page. In the present invention, by using the conceptual meaning of the page rank algorithm, an individualized page rank algorithm is applied, assuming that each entity is a page and applying association with the entities that appear together.

In the present invention, it was confirmed that the pair linking approach proposed based on the comprehensive interdependence for resolving object name neutrality has better performance than the results of applying the individualized page rank algorithm as well as the traditional pair linking approach.

1 shows a conceptual diagram of an apparatus for resolving neutrality of entity names based on a knowledge base according to an embodiment of the present invention.

Referring to Figure 1, the object name neutrality resolution device 100 includes a candidate group extraction unit 110, Entity2vec model unit 130, the object name neutrality resolution unit 140. According to an embodiment of the present invention, the RDF knowledge base 120 may be a database stored in a memory included in the object name neutralization resolution device 100 or a database stored in an external storage device.

The process of resolving the neutrality of the individual name is shown in FIG. 1. The object name neutrality cancellation process may be performed by one or more processors of the object name neutrality resolution device 100.

First, in step 1, an Entity2vec model using the Word2vec algorithm is generated. In step 2, the word used in the sentence is searched in the knowledge base to extract candidate groups for the object. The third step is to resolve the neutrality of the individual name, and a combination set is constructed by combining each candidate from the entire candidate group extracted in the second step. Then, using the embedded vector in the Entity2vec model, object neutrality can be resolved based on the similarity between each candidate.

Create Entity2vec model

First, the Entity2vec model unit 130 of the object name neutralization resolution apparatus 100 generates an Entity2vec model proposed in the present invention based on the RDF knowledge base 120 (①).

RDF was proposed to model metadata at the World Wide Web Consortium. The RDF knowledge base is in the form of a triple that is a subject-property-object. The RDF knowledge base expresses relationships between entities as descriptors, making it easy to intuitively grasp the relations between entities. Accordingly, the triple form of the RDF knowledge base can be written in natural language sentences. For example, when there is a triple called'Leon-Opening Year-1994', this triple can be written in a natural language sentence such as'Lyon's opening year is 1994.'

The present invention proposes an Entity2vec model using the Word2vec algorithm based on the RDF knowledge base. Word2vec, one of the word embedding methodologies for vectorizing words, extracts vectors for each word by learning words that appear at the same time in the sentence. The Entity2vec model uses the RDF knowledge base that can be expressed in natural language sentences by quoting the concept of Word2vec as learning data. That is, a vector is generated for each individual by learning co-occurence between the entities. At this time, among the various triple types in the knowledge base, only data in the form expressing the relationship between the individual and the individual were extracted and used as learning data. In other words, only the data in which the data type of the subject and object are URI is extracted and trained. This is excluded because data in the form of Datatype rather than URI is not an entity defined in the present invention and may interfere with data learning.

Candidate extraction

In step 2, the candidate group extracting unit 110 of the object name neutralization resolution apparatus 100 connects the words that can be recognized as objects among the words appearing in the sentence to the RDF knowledge base 120 to connect each word to the object. The candidate group is selected (②). At this time, the word that can be recognized as an individual means a noun, a verb, and an adjective, except for stopwords such as articles, prepositions, investigations, and conjunctions among all the words appearing in a sentence. For example, when there is a sentence such as'Are the roles of Matilda in the movie Leon, is my reporter?', the input words of the sentence are'Movie','Leon','Matilda','Role','Natal Reporter' to be. These words are first given an object type that a word can have through a dictionary analysis. 'Leon, Matilda, Natal Reporter' can be mapped to an instance,'Movie' to a class or instance, and'Role' to properties.

According to an embodiment of the present invention, two pieces of information are used to select an individual candidate group, and if both information matches, an individual candidate group is selected. In the RDF knowledge base 120, a word can be matched 1:1, and an object matching the word's object type and the object's object type can be selected as a candidate group while matching the word.

For example, in the case of the word'Leon', the RDF knowledge base 120 searches for'Leon', and at the same time, only an instance instance can be selected as a candidate group. In this case, the candidate group may be selected as a candidate for the movie'Leon', the character'Leon', and the song title'Leon'.

Eliminate the name neutrality

In step 3, the object name neutrality canceling unit 140 of the object name neutrality cancellation apparatus 100 performed ranking on the object candidate groups created in step 2. For ranking, candidates were extracted from each candidate group and a combination set was constructed (③).

Subsequently, the cosine similarity between candidates extracted from each candidate group is calculated for the constructed combination set, and an average is obtained (④). In the present invention, unlike the traditional pair linking approach, by using the similarity of the entire set of combinations, it is intended to resolve individual neutrality based on comprehensive interdependence. The algorithm formula of the pair linking approach based on comprehensive interdependence is as shown in <Equation 1>.

는 u 노드셋의 i 번째 노드를 의미하고,

는 ν 노드셋의 j 번째 노드를 의미한다. 즉, <수학식 1>의 알고리즘은 조합셋에 있는 개체 후보들 간의 코사인 유사도를 계산하여 모두 합한 뒤, 단어의 개수로 나누어 조합셋의 평균값을 구하는 것이다. 이후, 전체 조합셋을 대상으로 값을 구한 뒤, 그 중 가장 큰 값을 가진 조합셋을 선택하여 개체 중의성을 해소할 수 있다.The algorithm of <Equation 1> means the score of the combination set. The algorithm of Equation 1 has V node sets according to the number of words. A nodeset is a set of candidate entities for a word.

Is the i th node of the u node set,

Is the j th node of the ν node set. That is, the algorithm of <Equation 1> calculates the cosine similarity between the individual candidates in the combination set, sums them, and divides them by the number of words to obtain the average value of the combination set. Subsequently, after obtaining a value for the entire combination set, the object neutrality can be resolved by selecting the combination set having the largest value.

According to one embodiment of the present invention, an Adam knowledge base was used to create learning data. Adam Knowledgebase is an RDF knowledgebase built by Saltlux, Inc. to develop a knowledge base-based question-and-answer system. It consists of approximately 17 million instances, 1,000 properties, and 200 million triples.

The method for resolving the neutrality of the present invention can be applied to a question-and-answer system, etc., and a set of questions provided by Saltlux can be used as test data for evaluation. The three questions provided by Saltlux consist of questions with the correct answer of'Yes/No', such as'Irene is a member of the girl group Red Velvet?' and'Are you Matilda's role in the movie Leon? have. In one embodiment of the present invention, among the total of 198 questions, there were 51 questions with errors in word extraction because parsing was not accurately performed, and 94 questions except 53 questions in which candidates could not be found due to lack of words in the knowledge base. The experiment was conducted on the subject. When parsing 94 questions, a total of 370 words and object type information of the words are extracted. In order to evaluate the question-and-answer system to which the object name neutrality resolution method of the present invention was applied, a set of correct answers was manually constructed for 370 words. One example of the form of the correct answer set is shown in Table 1 below.

Word Entity type Answer My Reporter Resource http://kb.adams.ai/resource/0000410364 Leon Resource http://kb.adams.ai/resource/0000093989 movie Class http://kb.adams.ai/schema/class/movie_06205452 role Property http://kb.adams.ai/schema/property/role Matilda Resource http://kb.adams.ai/resource/0030163951

In order to evaluate the performance of the method for resolving the object name neutrality based on the global interdependence based pairwise approach (GIPW) proposed in the present invention, an interdependence based pairwise (IPW) ) And a personalized pagerank (PPR) designed based on comprehensive interdependence were compared and tested.

In the case of IPW, neutrality is resolved by using cosine similarity between individual candidates. In the case of PPR, in order to weight the edge when moving from one node to another, the algorithm is'DoSeR-a knowledge-base-agnostic framework for entity disambiguation using semantic embeddings.' The ETP (Entity Transition Probabilities) algorithm of the paper was used. This algorithm has V node sets according to the number of words. The node set refers to a set of candidate entities for a word, and the node refers to k candidates in the node set. The algorithm formula is as follows <Equation 2>.

는 u 노드셋의 i 번째 노드를 의미하고,

는 ν 노드셋의 j 번째 노드를 의미한다. here,

Is the i th node of the u node set,

Is the j th node of the ν node set.

An example of application of such an ETP algorithm is illustrated with reference to FIG. 2.

In FIG. 2, an example in which the ETP algorithm is applied to'Leon' and'Natural Reporter' used in sentences is illustrated. As shown in Fig. 2, there is a node set made of the word'Leon', and there are three individual candidates under the title of the movie'Leon', the character'Leon', and the song title'Leon' as the node, and'My' There is a node set made of the word'Turtle Reporter', and underneath there is one individual candidate, actor'Natal Reporter'. When the vector of each node in the Entity2vec model is applied to the ETP algorithm, the actor'Natural Reporter' receives 1 point each from the nodes in the'Leon' node set. Conversely, the actor'Natural Reporter' nodeset divides the total of 1 point that can be given to other nodes to the nodes in the'Leon' nodeset, and applies the score to the ETP algorithm. The score is 0.4 for',' the character'Leon' is 0.4, and the song title'Leon' is 0.2. In this way, two node sets can accumulate the value of applying the ETP algorithm between each node, and finally, the node with the highest score in each node set can resolve the object neutrality.

The performance results of the three methodologies are obtained as shown in Table 2 below. <Table 2> is the result of using the basic option (Baseline) of learning the Word2Vec model provided by DL4J in training the Entity2Vec model.

As shown in <Table 2>, it can be seen that the response rates of all three methodologies are 100%. This is because the set of test questions of the present invention excludes questions with words in which the candidate group cannot be found due to parsing and lack of words in the knowledge base, data evaluated in the present invention is targeted for words in which the candidate groups in the knowledge base exist Therefore, the response rate was 100%. The precision was 63.2% for IPW, 65.7% for GIPW, and 64.3% for PPR. GIPW and PPR considering comprehensive interdependence had higher results than IPW, of which GIPW of the present invention was 1.4% higher than PPR.

Through these results, it can be confirmed that the methodology considering comprehensive interdependence is more effective in resolving object name neutrality compared to IPW considering only interdependence, and among them, pair linkage considering comprehensive interdependence proposed in the present invention The approach (GIPW) was found to be very effective in resolving individual neutrality.

Baseline IPW GIPW PPR Response rate 100% 100% 100% Precision 63.2% 65.7% 64.3% F-score 77.5% 79.3% 78.3%

<Table 3> below shows the result of creating and testing the Entity2vec model by increasing Iteration from 1 to 10 in the basic option. As a result of applying the trained model by increasing the iteration from 1 to 10, it can be seen that GIPW is more effective than IPW and PPR with 70.5% precision and 82.7% F-score.

Iteration increase (1 -> 10) IPW GIPW PPR Response rate 100% 100% 100% Precision 67% 70.5% 65.6% F-score 80.2% 82.7% 79.2%

<Table 4> shows the result of testing after creating Entity2vec model by increasing Epoch from 1 to 10 in the basic option. As a result of applying the trained model by increasing the epoch from 1 to 10, it was confirmed that GIPW and PPR showed similar performance with an accuracy of 67.3% and an F-score of 80.5%.

Epoch increase (1 -> 10) IPW GIPW PPR Response rate 100% 100% 100% Precision 66.5% 67.3% 67.3% F-score 79.6% 80.5% 80.5%

<Table 5> shows the result of creating and testing the Entity2vec model by increasing the Layersize from 100 to 200 in the basic option. As a result of applying the trained model by increasing the layer size from 100 to 200, it can be seen that the precision of GIPW is 65.1% and the F-score is 78.9%, which is more effective than IPW and PPR. Although the performance result is lower than the basic option, it can be seen that the overall performance of the GIPW is excellent.

Increased Layersize (100 -> 200) IPW GIPW PPR Response rate 100% 100% 100% Precision 62.4% 65.1% 64.6% F-score 76.9% 78.9% 78.5%

Among the results of models that increased Iteration, Epoch, and Layersize respectively, the results of the model created by increasing Iteration only showed the best results. In addition, it was confirmed that the performance of the trained model by increasing Iteration and Epoch is superior to the model trained as the basic option. This indicates that training the model by increasing the Iteration and Epoch options is an important factor in creating a more advanced model compared to training the model by increasing the Layersize.

3 is a flowchart of a method for resolving neutrality of an entity name based on a knowledge base according to an embodiment of the present invention. For example, the knowledge base based object name neutrality resolution method may be performed by the object name neutrality resolution device 100.

Referring to FIG. 3, the knowledge base-based object name neutralization method includes selecting words recognized as objects among words in the input sentence (S310) and determining individual candidate groups for each of the selected words (S320). ), constructing a combination set in which candidates are randomly combined from each of the individual candidate groups (S330) and calculating the ranking of the similarity of each combination set to determine the meaning of the individual name of each word (S340 ).

First, the apparatus 100 for resolving the neutrality of individual names selects words recognized as entities among words of the input sentence (S310). For example, the candidate group extracting unit 110 may select words that can be recognized as objects among words appearing in the sentence. In this case, the word that can be recognized as an individual may mean a noun, a verb, or an adjective, excluding stopwords such as articles, prepositions, investigations, and conjunctions among all words appearing in a sentence.

Next, the apparatus 100 for resolving the neutrality of an individual name determines an individual candidate group for each of the selected words based on data learning expressing the relationship between the individual and the entity using the knowledge base (S320). For example, the candidate group extracting unit 110 may select an individual candidate group for each word by connecting the RDF knowledge base 120 to words that can be recognized as objects among the words. Words are first given an object type that a word can have through parsing. For example,'Leon, Matilda, Natal Reporter' may be mapped to an instance,'Movie' to a class or instance, and'Role' to properties. The individual candidate group may select an individual that matches the word and the object type of the word and the individual object type of the object as the candidate group.

Subsequently, the apparatus 100 for resolving the neutrality of an individual name constructs a combination set in which candidates are arbitrarily combined by considering the number of all cases from each of the individual candidate groups (S330). For example, as shown in FIG. 1, the object name neutrality canceling unit 140 may construct a combination set by extracting candidates from each candidate group for the object candidate groups created in the previous step.

Lastly, the apparatus 100 for resolving the neutrality of individual names calculates the ranking of the similarity of each combination set and determines the meaning of each word (S340). For example, the object name neutralization resolver 140 calculates cosine similarities between individual candidates in a combination set based on an algorithm of a pair linking approach based on a comprehensive interdependency, sums them all, and then counts the words. Divide by to get the average value of the combination set. Subsequently, after obtaining a value for the entire set of combinations, the object set can be resolved by selecting the combination set having the largest value. The algorithm formula of the pair linking approach proposed on the basis of the comprehensive interdependence proposed in the present invention can be expressed as Equation 1 described above.

In the above-described specific embodiments, components included in the present invention are expressed in singular or plural according to the specific embodiments presented. However, the singular or plural expressions are appropriately selected for the situation presented for convenience of explanation, and the above-described embodiments are not limited to the singular or plural components, and even the components expressed in plural are composed of the singular or , Even a component represented by a singular number may be composed of a plurality.

On the other hand, in the description of the invention, specific embodiments have been described, but various modifications are possible without departing from the scope of the technical spirit of the various embodiments. Therefore, the scope of the present invention should not be limited to the described embodiments, but should be defined not only by the claims described below, but also by the claims and equivalents.

Claims (10)

As a method of resolving the neutrality of individual names based on knowledge base,
Selecting words recognized as entities among words in the input sentence;
Determining an individual candidate group for each of the selected words based on data learning expressing the relationship between the individual and the individual using the knowledge base;
Constructing a combination set in which candidates are randomly combined considering the number of all cases from each of the individual candidate groups; And
And determining the meaning of each of the selected words by calculating the ranking of the similarity of each combination set. According to claim 1,
The step of selecting words recognized as entities among words of the input sentence,
And selecting words that are nouns, verbs, or adjectives among words in the input sentence. According to claim 1,
Determining the individual candidate group of each of the selected words,
And matching with each of the selected words, determining an individual matching the individual type of each of the selected words as a candidate group. According to claim 3,
The object type, characterized in that any one of instances, classes, properties, method of resolving object name neutrality. According to claim 1,
Similarity of each combination set is calculated based on the following [Equation],
[Mathematics]

Here, V is the number of node sets that are a set of candidate objects for a word,

Is the i th node of the u nodeset,

Is a j-th node of the ν node set. As a knowledge base based object name neutralization resolution device,
Among words in the input sentence, words recognized as individuals are selected, and based on data learning expressing the relationship between the individuals and individuals using the knowledge base, individual candidate groups for each of the selected words are determined, and in each case from each of the individual candidate groups And one or more processors for constructing a combination set in which candidates are arbitrarily considered in consideration of the number of, and determining the meaning of each of the selected words by calculating the ranking of the similarity of each combination set. , Deuteronization device. The method of claim 6,
The one or more processors,
Device for resolving the neutrality of an individual name, characterized by selecting words that are nouns, verbs or adjectives among words of the input sentence. The method of claim 6,
The one or more processors,
And matching each of the selected words, determining an individual matching the individual type of each of the selected words as a candidate group. The method of claim 8,
The object type is an instance, class, property, characterized in that any one of, object name neutrality resolution device. The method of claim 6,
Similarity of each combination set is calculated based on the following [Equation],
[Mathematics]

Here, V is the number of node sets that are a set of candidate objects for a word,

Is the i th node of the u nodeset,

Is a j-th node of the ν node set, characterized in that the object name neutrality cancellation device.

Images