KR20210130976A - Device, method and computer program for deriving response based on knowledge graph - Google Patents

Abstract

An apparatus for deriving response on basis of knowledge graph comprises: a collection unit for collecting semantic information including subject information, object information, and predicate information between subject information and object information from an open domain; a knowledge graph generating unit that generates a knowledge graph on the basis of subject information, predicate information, and purpose information; a query analysis unit that receives query data and analyzes the input query data; a subgraph derivation unit that searches for at least one subgraph from the knowledge graph generated based on an analysis result of the query data, and derives a final subgraph from among the at least one searched subgraph; and a response derivation unit that selects at least one document from among a plurality of documents collected from the open domain on the basis of the derived final subgraph, and derives response data to the query data on the basis of a selected document. A final subgraph among at least one subgraph corresponding to an analysis result of searched query data is derived from a knowledge graph.

Description

DEVICE, METHOD AND COMPUTER PROGRAM FOR DERIVING RESPONSE BASED ON KNOWLEDGE GRAPH

The present invention relates to an apparatus, method and computer program for deriving a response based on a knowledge graph.

Existing question-and-answer methods include a question-and-answer method using a Knowledge Based Question Answering (KBQA) and a Q&A method using an Information Retrieval Question Answering (IRQA) algorithm. Recently, research on a machine reading comprehension question answering (MRCQA) method in which deep learning technology is applied to an information search-based question and answer algorithm is being conducted.

The question-and-answer method using the knowledge base-based algorithm uses the ontology method. Here, the ontology method is a knowledge management technique that defines and creates relationships between words. The question-and-answer method using a knowledge base-based algorithm is a method of identifying the core meaning of a question and searching for an answer by querying the knowledge base. Since this question-and-answer method provides only Q&A for knowledge categories existing in the knowledge base, it is impossible to derive Q&A for open domain or real-time (trend) data.

The question-and-answer method using an information search-based algorithm builds expected question and answer data in advance, searches for questions similar to user questions, and provides answers matched to the searched questions. In this question-and-answer method, it is difficult to provide an answer to the expected question and the question for which the answer data is not established.

The machine reading-based question-and-answer method is a method that trains a machine to read documents, find answers to questions, and present them without a person directly building knowledge. Since this question-and-answer method learns the pattern for the correct answer in a document, it is difficult to derive an accurate answer unless an accurate document search is premised. The method of deriving a response based on such a correct answer pattern has a problem in that a response that is not related to the context of the query is derived.

Conventionally, when performing an indexing process for document search, the document search ranking is re-ranked using an index for entity names, a semantic paragraph unit search, or a paragraph re-ranking-based technology using deep learning technology. However, when searching through the entity name index is used, search results that do not match the query and the context semantically often come out. It is not possible to apply a search through the object name index of

Japanese Patent Publication No. 4746439 (Registered on May 20, 2011)

The present invention is to solve the problems of the prior art, and derives a final subgraph among at least one subgraph corresponding to an analysis result for query data searched from a knowledge graph, and based on the derived final subgraph We want to derive response data to the query data from the selected document.

However, the technical problems to be achieved by the present embodiment are not limited to the technical problems described above, and other technical problems may exist.

As a technical means for achieving the above-described technical problem, the apparatus for deriving a response based on the knowledge graph according to the first aspect of the present invention includes subject information, object information, and predicate information between the subject information and the object information from an open domain. a collection unit for collecting semantic information including; a knowledge graph generator configured to generate a knowledge graph based on the subject information, the predicate information, and the object information; a query analysis unit that receives query data and analyzes the input query data; a subgraph derivation unit that searches for at least one subgraph from the generated knowledge graph based on an analysis result of the query data and derives a final subgraph from among the found at least one subgraph; and a response derivation unit that selects at least one document from among the plurality of documents collected from the open domain based on the derived final subgraph, and derives response data to the query data based on the selected document. have.

A method for deriving a response based on a knowledge graph according to a second aspect of the present invention comprises: collecting semantic information including subject information, object information, and predicate information between the subject information and the object information from an open domain; generating a knowledge graph based on the subject information, the predicate information, and the object information; receiving query data; analyzing the input query data; searching for at least one subgraph from the generated knowledge graph based on an analysis result of the query data; deriving a final subgraph from among the searched at least one subgraph; selecting at least one document from among a plurality of documents collected from the open domain based on the derived final subgraph; and deriving response data to the query data based on the selected document.

A computer program stored in a computer readable recording medium including a sequence of instructions for deriving a response based on the knowledge graph according to the third aspect of the present invention, when executed by a computing device, includes subject information, object information and the above information from an open domain. Collect semantic information including predicate information between subject information and the target information, generate a knowledge graph based on the subject information, the predicate information, and the target information, and based on the analysis result of the query data, the generated Searching for at least one subgraph from the knowledge graph, deriving a final subgraph among the found at least one subgraph, and at least one of a plurality of documents collected from the open domain based on the derived final subgraph and a sequence of instructions for selecting a document and deriving response data to the query data based on the selected document.

The above-described problem solving means are merely exemplary, and should not be construed as limiting the present invention. In addition to the exemplary embodiments described above, there may be additional embodiments described in the drawings and detailed description.

According to any one of the above-described problem solving means of the present invention, the present invention derives the final subgraph among at least one subgraph corresponding to the analysis result for the query data searched from the knowledge graph, and adds it to the derived final subgraph. Based on the selected document, response data to the query data may be derived.

Also, according to the present invention, a document most suitable for the context of the query data may be searched for based on the final subgraph information (object name and relationship name), and response data to the query data may be derived from the searched document.

In addition, according to the present invention, the type of response data to the query data in the document may be derived using the final subgraph information, and the response data may be verified according to the type of the response data.

In addition, since the present invention can limit the domain of the entity name through semantic recognition of the relation name even in a situation where the entity name is not recognized, the domain for the entity name and the extracted relation name are searched even in a search where the entity name is not recognized. can increase the accuracy of

1 is a block diagram of an apparatus for deriving a response, according to an embodiment of the present invention.
2A to 2G are diagrams for explaining a method of deriving a response based on a knowledge graph, according to an embodiment of the present invention.
3 is a diagram for explaining a method of deriving a response based on a knowledge graph, according to another embodiment of the present invention.
4 is a flowchart illustrating a method of deriving a response based on a knowledge graph according to an embodiment of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those of ordinary skill in the art can easily implement them. However, the present invention may be embodied in many different forms and is not limited to the embodiments described herein. And in order to clearly explain the present invention in the drawings, parts irrelevant to the description are omitted, and similar reference numerals are attached to similar parts throughout the specification.

Throughout the specification, when a part is "connected" with another part, this includes not only the case of being "directly connected" but also the case of being "electrically connected" with another element interposed therebetween. . In addition, when a part "includes" a certain component, this means that other components may be further included, rather than excluding other components, unless otherwise stated.

In this specification, a "part" includes a unit realized by hardware, a unit realized by software, and a unit realized using both. In addition, one unit may be implemented using two or more hardware, and two or more units may be implemented by one hardware.

Some of the operations or functions described as being performed by the terminal or device in the present specification may be instead performed by a server connected to the terminal or device. Similarly, some of the operations or functions described as being performed by the server may also be performed in a terminal or device connected to the server.

Hereinafter, detailed contents for carrying out the present invention will be described with reference to the accompanying configuration diagram or process flow diagram.

1 is a block diagram of an apparatus 10 for deriving a response, according to an embodiment of the present invention.

Referring to FIG. 1 , the response derivation apparatus 10 includes a collection unit 100 , a knowledge graph generation unit 110 , a query analysis unit 120 , a subgraph derivation unit 130 , a response derivation unit 140 , and an index. part 150 may be included. However, the response deriving apparatus 10 shown in FIG. 1 is only one implementation example of the present invention, and various modifications are possible based on the components shown in FIG. 1 .

The indexing unit 150 includes a plurality of documents ( For example, an indexing process may be performed on sentences included in real-time/trend news, web documents, etc.). Also, the indexing unit 150 may index entity name and vocabulary information into subject information and object information of each sentence according to the entity name recognition result, and index entity name and vocabulary information into predicate information of each sentence. Here, the entity name dictionary is a dictionary defining entity names (eg, nouns having unique meanings such as person names, organization names, and geographical names), and the relation name dictionary is a dictionary defining relationships (eg, verbs) between entities. The entity name dictionary and relation name dictionary may be built around categories for, for example, people, places, organizations, food, cultural contents, terms, products, events, artifacts, and the like.

For example, the result of recognition of individual names in the first sentence of 'Trump's election made Melania the second immigrant to the United States and the first first lady from a communist country' is [Donald Trump/Person, Melania Trump/Person, USA/Nation, Immigrant/RQCWho, First Lady/PoliticalPerson, Melanie/Singer. second/Sequence]. In this case, the indexing unit 150 may index an entity name (eg, Person) and vocabulary information (eg, Donald Trump) to each of the subject information and the object information of the first sentence based on the entity name recognition result. Also, the indexing unit 150 may index the entity name (eg, IsA) and vocabulary information (eg, it was) in the predicate information of the first sentence based on the entity name recognition result.

The indexing unit 150 may perform an indexing process on sentences included in a plurality of collected documents based on the morpheme analysis result and the tokenizer result. For example, the morphological analysis result for the first sentence is [Trump/NNP+'s/JKG+Elected/NNG+/JKB+Melania/NNG+/JX+US/NNP+'s/JKG+Double/MM+No./NNB+th/XSN+Immigrant/ From NNG+/NNG+Lee/VCP+Ja/EC+First/MM+Communist/NNG+Kwon/XSN+Country/NNG+From/NNG+First/NNG+Lady/NNG+A/JKS+Be/VV+Had/EP+Da/EF+./SF] and Talk The Niger result is [Trump's election made Melania the second immigrant to the United States and the first First Lady from a communist country]. In this case, '_' in the tokenizer result means a space in the input sentence.

When indexing an entity name corresponding to each of subject information, object information, and predicate information of a sentence, the indexing unit 150 may perform an indexing process using keywords included in a pre-registered thesaurus. Here, the thesaurus is a dictionary in which synonyms (eg, words having similar meanings) for each of the entity name and the relation name are collected.

When a plurality of entity names are recognized among subject information, object information, and predicate information of a single sentence, the indexing unit 150 provides distance information between a plurality of entity names to index the corresponding entity name among information, object information, and predicate information. The indexing process can be performed. In this case, the distance information between the object names may be used when calculating a ranking for a document search later.

The collection unit 100 may collect subject information, object information, and semantic information including predicate information between subject information and object information from the open domain. For example, the collection unit 100 may collect semantic information including 'Donald Trump (subject information) was born in Queens, New York (purpose information) (predicate information)'.

The knowledge graph generating unit 110 may generate the knowledge graph as a main graph based on subject information, predicate information, and object information included in the collected semantic information. For example, the knowledge graph generating unit 110 extracts 'Donald Trump' from semantic information including 'Donald Trump married Melania Trump' as information, and extracts 'Melania Trump' as target information. and 'married' can be extracted as predicate information.

The knowledge graph generating unit 110 may generate a node corresponding to the subject information and the object information, generate an edge indicating predicate information between each node, and generate a knowledge graph based on the generated node and the edge. Here, the knowledge graph has a structure in which a triple set (subject information, object information, and predicate information) constituting one knowledge is expressed by nodes and edges. Here, a node represents subject information and object information, and an edge represents predicate information. Nodes can be divided into instances and literal values. For example, referring to FIG. 2B , an instance means an entity name such as 'Donald Trump, Slovenia, USA, etc.', and a literal value means a constant value such as '1946-06-14'.

For example, referring to FIG. 2A , the knowledge graph generating unit 110 generates 'Donald Trump' corresponding to subject information and 'Melania Trump' corresponding to target information as nodes 201 and 205, respectively, 'isMarriedTo' corresponding to predicate information between the node 201 corresponding to 'Donald Trump' and the node 205 corresponding to 'Melania Trump' may be generated as the edge 203 . In addition, the knowledge graph generating unit 110 uses the node 201 corresponding to 'Donald Trump', the node 205 corresponding to 'Melania Trump', and the edge 203 corresponding to 'isMarriedTo' to the knowledge graph. can create

The knowledge graph generating unit 110 may generate a knowledge graph based on semantic information included in a news article (news document) including real-time/trend information (refer to FIG. 2B ). For example, the knowledge graph generating unit 110 generates a knowledge graph corresponding to each domain by classifying the news into each field such as politics, economy, society, life, culture, world, IT/science, sports, etc. can In addition, the knowledge graph generating unit 110 may generate a knowledge graph by dividing the sports domain into baseball, soccer, volleyball, basketball, golf, e-sports, and the like.

The knowledge graph generating unit 110 may map an entity name or a relation name to schema information about a node or an edge included in the knowledge graph based on the previously registered entity name dictionary and relation name dictionary. Specifically, the knowledge graph generating unit 110 may map the entity name to the class information for the node included in the knowledge graph, and map the relation name to the attribute/relationship information (ie, the relationship between class information) about the edge. have. For example, if you look at 'Melania Trump/Person' in the entity name recognition result, Person corresponds to both the entity name and the class information of the instance 'Melania Trump' in the knowledge graph.

The knowledge graph generating unit 110 may manage the synonyms associated with each of the entity names mapped to the nodes of the knowledge graph and the relation names mapped to the edges based on the pre-registered thesaurus through the knowledge graph. For example, a synonym for 'plot' may be 'summary', and a synonym for 'author' may be 'writer'.

The query analysis unit 120 may receive query data and analyze the input query data. Also, the query analysis unit 120 may analyze the natural language-processed query data after performing natural language processing on the input query data. Here, the natural language processing may be performed using the entity name recognition result using the pre-registered entity name dictionary and relation name dictionary, morpheme analysis result, and tokenizer result.

The query analysis unit 120 may analyze the query data based on the entity name recognition result. Also, the query analyzer 120 may further analyze the query data based on the morpheme analysis result and the tokenizer result.

For example, if the query data is 'Melania Trump is the first lady from which immigrant', the object name recognition result is [Melania Trump/Person, Melanie/Singer, Donald Trump/Person, Immigrant/RQCWho, First Lady. /PoliticalPerson, number/Sequence], and the result of morphological analysis is [Melanie/NNP+ah/JKV+Trump/NNG+Ka/JKS+Several/MM+#/NNB+th/XSN+Immigrant/NNG+ native/NNG+First/NNG+Lady/NNG+Lee/ VCP+Ya/EF], and the tokenizer result is [Melania Trump is the First Lady of the First Immigrants].

The query analysis unit 120 may extract a plurality of tokens from the analysis result of the query data processed by the natural language. Specifically, the query analysis unit 120 may extract a plurality of tokens from the entity name recognition result, the morpheme analysis result, and the tokenizer result for the query data. Here, the plurality of tokens are used when searching for nodes and edges of the knowledge graph.

For example, the query analysis unit 120 extracts a 'Melania Trump' token from [Melania Trump/Person] corresponding to the entity name recognition result, and 'Trump' from [Trump/NNG] corresponding to the morphological analysis result. ' tokens can be extracted, and 'Trump' tokens can be extracted from the tokenizer result.

Also, the query analysis unit 120 may extract synonyms related to a plurality of tokens from the thesaurus. In this case, synonyms related to the plurality of extracted tokens may be utilized when searching for nodes and edges of the knowledge graph.

The search unit (not shown) may search for nodes and edges of the knowledge graph using the plurality of tokens extracted by the query analysis unit 120 . For example, referring to FIG. 2C , the search unit (not shown) corresponds to a node by using a plurality of tokens extracted from the analysis result of the query data composed of 'Melania Trump is the first lady from which immigrant is' You can search for {'donald trump', 'trump', 'melania trump', 'woman', 'first lady'} and edge {'hasCountry', 'immigrator'}.

The sub-graph derivation unit 130 may search for at least one sub-graph from the knowledge graph based on the analysis result of the query data, and derive a final sub-graph from among the at least one searched sub-graph. . For example, referring to FIGS. 2C and 2D together, the subgraph derivation unit 130 determines the nodes searched through a plurality of tokens ('Donald Trump', 'Trump', 'Melania Trump', 'Female', 'First Part 1 consisting of {Donald Trump-FamilyName-Trump} from the knowledge graph using 'Lady') and Edge ('hasCountry', 'immigrator'), Part 2 consisting of {Donald Trump-isMarriedTo-Melania Trump} Graph, subgraph 3 composed of {Melania trump-hasCountry-Slovenia}, subgraph 4 composed of {melania trump-gender-woman}, subgraph 5 composed of {melania trump-position-first lady} , {Melania Trump-immigrator-yes} can be searched for the sixth subgraph.

For example, the subgraph derivation unit 130 may sort nodes corresponding to instances among at least one node of the knowledge graph according to the analysis result of the query data in alphabetical order. Also, the subgraph derivation unit 130 may search for an edge connected to the first node in the process of searching the subgraph in the knowledge graph based on the first node among the nodes arranged in alphabetical order. In addition, when there is an edge connected to the first node found in the edge list according to the analysis result of the query data, the subgraph derivation unit 130 may extend the subgraph by connecting the first node, the edge, and the extension node connected to the edge. . If the edge connected to the first node is not in the edge list according to the analysis result of the query data, the subgraph derivation unit 130 may extend all edges connected to the first node and nodes connected to the edge to a temporary subgraph. In addition, the subgraph derivation unit 130 searches for another edge connected to the extension node with respect to the extension node, and based on whether the searched other edge exists in the edge list according to the analysis result of the query data, the extension node and the other edge It is possible to determine the subgraph extension of . Also, when the extension node is included in the current subgraph, the subgraph derivation unit 130 may end the subgraph extension. The subgraph derivation unit 130 may perform the subgraph search in the knowledge graph based on the node corresponding to the instance in the extended subgraph in the manner described above.

The subgraph derivation unit 130 may derive the final subgraph by inputting at least one found subgraph to the pre-learned deep learning network. For example, referring to FIG. 2E , the subgraph derivation unit 130 inputs a plurality of subgraphs found from the knowledge graph according to the analysis result of the query data into the deep learning network, and the subgraph extracted through the deep learning network. A subgraph may be determined as the final subgraph.

The subgraph derivation unit 130 may extract a triple set (ie, node-edge-node) from each of the plurality of extracted subgraphs. Here, the triple set may include information on at least two nodes included in each subgraph and information on an edge connecting the at least two nodes. For example, the triple set may be Melania Trump (node)-position (edge)-first lady (node).

Here, at least two nodes constituting the triple set and an edge connecting the at least two nodes are mapped to schema information for the knowledge graph. Specifically, (node, edge, node), which are components of the triple set, is mapped to (class information, attribute/relationship information, class information) of schema information. As another example, a triple set (Melania Trump, position, first lady) is mapped to schema information (Person class information, attribute/relationship information, and literal class information).

The subgraph derivation unit 130 may configure the extracted triple set as an input vector in a two-dimensional form. Such an input vector may correspond to a subgraph from which a triple set constituting the corresponding input vector is extracted.

The subgraph derivation unit 130 may map schema information for the knowledge graph to a row or column of an input vector corresponding to each subgraph. Here, the row and column of the input vector store information on class and relationship names. For example, class information (eg, row 1 is a politician class, row 2 is an actor class, row 3 is a singer class, etc.) is stored in each row of the input vector, and relationship information (eg, Column 1 is the name, column 2 is the nationality, column 3 is the date of birth, etc.).

Specifically, the subgraph derivation unit 130 maps class information among the schema information for the knowledge graph to the row of the input vector corresponding to the subgraph, and provides attribute/relationship information among the schema information for the knowledge graph to the column of the input vector. can be mapped. For example, referring to FIG. 2F , the subgraph derivation unit 130 places the Person class in the first row of the first input vector 207 corresponding to the first subgraph composed of {Melania Trump-position-First Lady} Information can be mapped, and position property/relationship information can be mapped to column 1. The subgraph derivation unit 130 maps hasCountry attribute/relationship information to column 2 of the second input vector 209 corresponding to the second subgraph composed of {Melania Trump-hasCountry-Slovenia}, and Nation class in row 3 Information can be mapped, and hasPeople attribute/relation information can be mapped in column 3. In this way, each subgraph may be expressed as a two-dimensional matrix based on triple information (or an input vector composed of it) extracted from each subgraph. Here, all subgraphs for the user's query can be expressed as input vectors.

For example, if query data including 'When did the corona start?' A two-dimensional matrix may be generated based on the class and relationship name to which the node and edge of each of the plurality of subgraphs belong. The subgraph derivation unit 130 may set an initial value for embedding at the matrix positions of class and relation names corresponding to nodes and edges of each extracted subgraph as shown in FIG. 2F .

Also, the subgraph derivation unit 130 may calculate a two-dimensional vector value by summing the two-dimensional matrix for each subgraph. When the plurality of extracted graphs is N, the subgraph derivation unit 130 may represent N two-dimensional matrices and generate a final union of the N two-dimensional matrices. At this time, the two-dimensional vector value in each two-dimensional matrix does not occupy a separate value even if the triple set overlaps, for example, it can be maintained as a preset value (eg, 1) as in the rightmost matrix of FIG. 2F. have. Through this, since the present invention uses a two-dimensional vector value as an input of a deep neural network (eg, CNN), it is possible to infer a subgraph that best matches a user query.

The subgraph derivation unit 130 may map a set of triples extracted from each subgraph to an input vector corresponding to each subgraph, and input the input vector to which the set of triples is mapped to the deep learning network. In addition, the subgraph derivation unit 130 may input a two-dimensional vector value obtained by summing the two-dimensional matrix for each subgraph to the deep learning network. In this case, the deep learning network may derive an output value based on the input vector to which the triple set for each subgraph is mapped. In this case, the derived output value may be an output value corresponding to the final subgraph (ie, the final subgraph for the query data).

The subgraph derivation unit 130 derives a degree of similarity between a triple set (a triple set extracted from each of a plurality of subgraphs) and an analysis result for query data based on the previously learned deep learning network, and based on the derived similarity The final subgraph may be derived from the subgraph with the highest similarity among the plurality of extracted subgraphs.

The subgraph derivation unit 130 may extract information to be used for document search from the derived final subgraph. Specifically, the sub-graph derivation unit 130 is an entity name for at least two nodes constituting the final sub-graph from the final sub-graph, a relation name for an edge connecting at least two nodes, and a corresponding entity name and relation extracted from the thesaurus. Synonyms related to each name and schema information (class information, attribute relationship) mapped to the final subgraph can be extracted. For example, the subgraph derivation unit 130 may derive a final subgraph for the query data including 'What immigrant first lady is Melania Trump?' as shown in FIG. 2G . Also, the subgraph derivation unit 130 may extract schema information including 'Nation, Person, hasCountry, gender, position, immigrator' from the final subgraph for the corresponding query data.

The response derivation unit 140 may select at least one document from among a plurality of documents collected from the open domain based on the derived final subgraph, and may derive response data for the query data based on the selected document.

The response derivation unit 140 may select at least one document from among the plurality of documents based on at least one of a keyword, an entity name, a relation name, and schema information for a node or edge included in the final subgraph. Here, the plurality of documents is a document on which an indexing process is performed on sentences included in each document. Entity name recognition result, morpheme analysis result, tokenizer result, entity name information, and relationship name information are indexed in the sentences of each document.

The document ranking unit (not shown) calculates the ratio (specific gravity) of the information extracted from the final subgraph for each of the plurality of documents (keyword, entity name, relation name, schema information for the node or edge of the final subgraph) can do. For example, the document ranking unit (not shown) calculates the first frequency of occurrence of keywords and synonyms for nodes or edges of the final subgraph for each document, and the entity name and relationship name of the final subgraph appear. The second frequency may be calculated, and the third frequency at which the schema information of the final subgraph appears may be calculated. For example, the document ranking unit (not shown) may calculate a ratio in which character strings included in the node and edge lists of the final subgraph appear as a pair in the sentence for each document. Alternatively, the document ranking unit (not shown) may calculate a ratio in which the entity name and the relation name of the final subgraph appear as a pair for each document.

The document ranking unit (not shown) learns the machine learning model by inputting the frequency of appearance (ratio) of the information extracted from the final subgraph calculated for each document into the machine learning model (eg, random forest model, etc.) as feature information. and a weight for each feature information can be calculated through the learned machine learning model.

The document ranking unit (not shown) calculates a ranking score for each document by using the weight for each characteristic information calculated by vomiting numerical information about the query data and a machine learning model, and based on the calculated ranking score for each document Thus, a plurality of documents can be ranked.

The response derivation unit 140 may select a document having a high ranking score among a plurality of documents, and may derive response data for the query data based on the selected document.

The response derivation unit 140 may verify response data for the derived query data. For example, the response derivation unit 140 may verify the response data by using information included in the final subgraph for the query data. Because the final subgraph has keyword, entity name, relation name, and schema information for a node or edge, response data (ie, purpose) based on the subject information (object name) and predicate information (relation name) constituting the query data information) can be inferred, and response data can be verified through type verification for response data.

For example, referring to FIG. 3 , the query data is 'What immigrant first lady is Melania Trump?' Assuming that she became the first lady from a communist country, the schema information of the final subgraph for the query data is 'Melania Trump/Person, Melanie/Singer, Donald Trump/Person, Immigrant/RQCWho, First Lady/PolicalPerson. . Second/Sequence'.

The correct answer for the query data is 'second'. The type of response data can be verified because the entity name corresponding to the 'number' in the query data and the entity name corresponding to the 'second' in the response data are the same sequence type. In addition, since it is a sequence type, it is possible to verify the response data even if it is expressed in a numeric format instead of a Korean character string.

On the other hand, if the entity name is not recognized through the entity name dictionary, the indexing unit 150 limits the domain of the entity name through semantic recognition of the relation name using the relation name dictionary, and a plurality of An indexing process can be performed on the sentences included in the document. The search accuracy can be improved by using the domain of the limited entity name through semantic recognition of the relation name.

For example, in the case of query data containing an abbreviation of a specific drama (eg, 'Gourmi', which is an abbreviation of the drama of moonlight drawn by clouds) or substrings, the relationship between the entity names included in the query data is not recognized through the entity name dictionary. The answer to the query can be derived through the domain of the entity name inferred through the semantic recognition of the name.

As another example, in the case of query data including 'When did the corona start?', it is difficult to determine which entity name the 'corona' entity corresponds to (eg, virus entity name, beer entity name, plasma atmospheric entity name, etc.) . In addition, even when searching using the 'corona' keyword, the probability of searching for a document describing the exact entity name corresponding to the 'corona' keyword is not high.

Accordingly, the indexing unit 150 may include 'When did the corona start?' Based on the query expression information, it is possible to derive the relationship name of 'occurrence date' from the relationship name dictionary. In addition, the indexing unit 150 may extract schema information related to the relationship name of the 'occurrence date' from the relation name dictionary by checking which class in the relation name dictionary the 'occurrence date' relation name is connected to.

Thereafter, the query analysis unit 120 performs a 'corona' (keyword, not an entity name) token, a 'occurrence date' relation name token (included in the relation name dictionary) from the analysis result of the query data including 'when did corona start?' class and relationship semantics) and 'when to start' tokens can be extracted.

The search unit (not shown) may search the nodes and edges of the knowledge graph using the 'corona' token, the 'occurrence date' relation name token, and the 'when start' token extracted by the query analysis unit 120 .

The subgraph derivation unit 130 may generate a plurality of subgraphs that are a set of nodes and edges that are semantically connected through the search of the subgraph using the knowledge graph based on the analysis result of the query data. At this time, nodes and edges that are not used to generate the subgraph are deleted.

The query analysis unit 120 may derive a result according to the semantic question analysis of the query data by deriving an optimal class, entity name (instance), and relationship name related to the query data through the deep neural network.

In addition, the document ranking unit (not shown) may perform a document search and ranking for each searched document through the result of semantic query analysis and keywords. In this process, it is possible to improve the search accuracy by reducing the rate of false searches that may appear when searching for the 'corona' keyword by calculating the documents that exist within an appropriate distance between the 'corona' keyword and the 'date of occurrence' as a higher ranking.

The response derivation unit 140 may derive an answer to the query data by inputting the query data and a document having a higher ranking to the machine learning model (eg, an MRC model, etc.).

As such, the present invention can provide an accurate answer to query data for open domain and real-time information.

Meanwhile, for those skilled in the art, the collection unit 100, the knowledge graph generation unit 110, the query analysis unit 120, the subgraph derivation unit 130, the response derivation unit 140, and the index unit 150 are each separated. It will be fully understood that it may be implemented, or one or more of these may be integrated and implemented.

4 is a flowchart illustrating a method of deriving a response based on a knowledge graph according to an embodiment of the present invention.

Referring to FIG. 4 , in step S401 , the response derivation apparatus 10 may collect subject information, object information, and semantic information including predicate information between subject information and object information from the open domain.

In step S403, the response derivation apparatus 10 may generate a knowledge graph based on subject information, predicate information, and object information.

In step S405, the response derivation apparatus 10 may receive query data.

In step S407 , the response derivation apparatus 10 may analyze the input query data.

In step S409 , the response derivation apparatus 10 may search for at least one subgraph from the knowledge graph generated based on the analysis result of the query data.

In step S411, the response deriving apparatus 10 may derive a final subgraph from among at least one searched subgraph.

In step S413 , the response derivation apparatus 10 may select at least one document from among a plurality of documents collected from the open domain based on the derived final subgraph.

In operation S415, the response deriving apparatus 10 may derive response data for the query data based on the selected document.

In the above description, steps S401 to S415 may be further divided into additional steps or combined into fewer steps, according to an embodiment of the present invention. In addition, some steps may be omitted as necessary, and the order between steps may be changed.

An embodiment of the present invention may also be implemented in the form of a recording medium including instructions executable by a computer, such as a program module executed by a computer. Computer-readable media can be any available media that can be accessed by a computer and includes both volatile and nonvolatile media, removable and non-removable media. Also, computer-readable media may include all computer storage media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data.

The above description of the present invention is for illustration, and those of ordinary skill in the art to which the present invention pertains can understand that it can be easily modified into other specific forms without changing the technical spirit or essential features of the present invention. will be. Therefore, it should be understood that the embodiments described above are illustrative in all respects and not restrictive. For example, each component described as a single type may be implemented in a dispersed form, and likewise components described as distributed may be implemented in a combined form.

The scope of the present invention is indicated by the following claims rather than the detailed description, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts should be construed as being included in the scope of the present invention. .

10: response eliciting device
100: collection unit
110: knowledge graph generating unit
120: query analysis unit
130: subgraph derivation unit
140: response elicitation unit
150: index

Claims (18)

An apparatus for deriving a response based on a knowledge graph, the apparatus comprising:
a collecting unit for collecting subject information, object information, and semantic information including predicate information between the subject information and the object information from the open domain;
a knowledge graph generator configured to generate a knowledge graph based on the subject information, the predicate information, and the object information;
a query analysis unit that receives query data and analyzes the input query data;
a sub-graph derivation unit that searches for at least one sub-graph from the generated knowledge graph based on an analysis result of the query data, and derives a final sub-graph from among the searched at least one sub-graph; and
A response derivation unit that selects at least one document from among the plurality of documents collected from the open domain based on the derived final subgraph, and derives response data to the query data based on the selected document
Which includes, a response derivation device.
The method of claim 1,
The knowledge graph generating unit generates a node corresponding to the subject information and the object information, generates an edge indicating predicate information between each node, and generates a knowledge graph based on the node and the edge, Response derivation Device.
3. The method of claim 2,
The knowledge graph generating unit is to map the entity name or relation name to schema information for a node or edge included in the knowledge graph based on the previously registered entity name dictionary and relation name dictionary.
4. The method of claim 3,
The knowledge graph generating unit will manage the thesaurus related to the mapped entity name or relationship name through the knowledge graph based on a pre-registered thesaurus.
4. The method of claim 3,
An indexing unit that performs an indexing process on the sentences included in the plurality of collected documents based on the entity name recognition result using the entity name dictionary and the relation name dictionary
further comprising,
The response derivation apparatus, wherein the query analyzer analyzes the query data based on the entity name recognition result.
6. The method of claim 5,
The indexing unit performs an indexing process on the sentences included in the plurality of collected documents based on the morpheme analysis result and the tokenizer result,
and the query analyzing unit further analyzes the query data based on the morpheme analysis result and the tokenizer result.
6. The method of claim 5,
The subgraph derivation unit searches for at least one node or edge from the knowledge graph based on subject information or predicate information included in the analysis result for the query data, and generates a plurality of subgraphs based on the found node or edge. to extract, a response derivation device.
8. The method of claim 7,
The subgraph derivation unit extracts a triple set from each of the plurality of extracted subgraphs, derives a similarity between the extracted triple set and an analysis result for the query data based on a previously learned deep learning network, and the derivation Deriving a final subgraph among the plurality of extracted subgraphs based on the obtained similarity,
The triple set includes information on at least two nodes included in each subgraph and information on an edge connecting between the at least two nodes,
The at least two nodes and an edge connecting the at least two nodes are mapped with schema information for the knowledge graph.
9. The method of claim 8,
The subgraph derivation unit maps schema information for the knowledge graph to a row or column of an input vector corresponding to each subgraph. The triple set extracted from each subgraph is mapped to the input vector, and the input vector to which the triple set is mapped is input to the deep learning network, the response derivation apparatus.
The method of claim 1,
The response derivation unit selects at least one document from among the plurality of documents based on at least one of a keyword for a node or edge included in the final subgraph, an entity name, a relation name, and schema information. .
A method for deriving a response based on a knowledge graph, the method comprising:
collecting semantic information including subject information, object information, and predicate information between the subject information and the object information from the open domain;
generating a knowledge graph based on the subject information, the predicate information, and the object information;
receiving query data;
analyzing the input query data;
searching for at least one subgraph from the generated knowledge graph based on an analysis result of the query data;
deriving a final subgraph from among the searched at least one subgraph;
selecting at least one document from among a plurality of documents collected from the open domain based on the derived final subgraph; and
deriving response data to the query data based on the selected document
A method for deriving a response that includes.
12. The method of claim 11,
generating the knowledge graph.
generating a node corresponding to the subject information and the object information;
generating an edge representing predicate information between each node; and
generating a knowledge graph based on the node and the edge
A method for deriving a response that includes.
13. The method of claim 12,
generating the knowledge graph.
mapping the entity name or relation name to the node or edge included in the knowledge graph based on the previously registered entity name dictionary and relation name dictionary
Which further comprises, the response derivation method.
14. The method of claim 13,
performing an indexing process on the sentences included in the plurality of collected documents based on the entity name recognition result using the entity name dictionary and the relation name dictionary
further comprising,
The analyzing of the input query data may include analyzing the query data based on the entity name recognition result.
15. The method of claim 14,
The step of searching the subgraph.
searching for at least one node or edge from the knowledge graph based on subject information or predicate information included in an analysis result of the query data; and
extracting a plurality of subgraphs based on the found node or edge
A method for deriving a response that includes.
16. The method of claim 15,
The step of deriving the final subgraph.
extracting a triple set from each of the plurality of extracted subgraphs;
deriving a similarity between the extracted triple set and the analysis result for the query data based on the previously learned deep learning network; and
deriving a final subgraph from among the plurality of extracted subgraphs based on the derived similarity
including,
The triple set includes information on at least two nodes included in each subgraph and information on an edge connecting the at least two nodes.
12. The method of claim 11,
Selecting the document includes:
The method for deriving a response, in which at least one of the plurality of documents is selected based on at least one of a keyword, an entity name, a relation name, and schema information for a node or edge included in the final subgraph.
A computer program stored in a computer-readable medium comprising a sequence of instructions for deriving a response based on a knowledge graph, the computer program comprising:
When the computer program is executed by a computing device,
Collect semantic information including subject information, object information, and predicate information between the subject information and the object information from the open domain,
generating a knowledge graph based on the subject information, the predicate information, and the object information;
Searching for at least one subgraph from the generated knowledge graph based on the analysis result of the query data,
Deriving a final subgraph among the searched at least one subgraph,
selecting at least one document from among the plurality of documents collected from the open domain based on the derived final subgraph;
and a sequence of instructions for deriving response data to the query data based on the selected document.

Images