KR102143157B1 - System and method for generating paraphrase sentence based on ontology - Google Patents

Abstract

A system for generating a paraphrase sentence, according to an exemplary embodiment of the present invention, acquires semantic resources including each of ontology components corresponding to tokens extracted from a query sentence, and semantic A characteristic extracting unit for generating a frame having a structure in which tokens can be replaced in a pattern and a query sentence, which is a combination of resources, a pattern index unit for correlating the patterns and frames, and storing the corresponding patterns and frames in the pattern storage unit, and It may include a sentence generating unit that generates a paraphrase sentence of the query sentence based on the frames corresponding to the pattern stored in the pattern storage unit.

Description

A system and method for generating paraphrase sentences based on ontology {SYSTEM AND METHOD FOR GENERATING PARAPHRASE SENTENCE BASED ON ONTOLOGY}

The technical idea of the present invention relates to a paraphrase sentence generation, and more particularly, to a system and method for generating an ontology-based paraphrase sentence.

The present invention is derived from research conducted by IPL Co., Ltd. and joint research conducted by Saltlux Co., Ltd. as part of the Robot Industry Core Technology Development Project-Artificial Intelligence Convergence Robot System Technology of the Ministry of Commerce, Industry and Energy. [Research period: 2018.01.01~2018.12.31, Research management professional institution: Korea Institute of Industrial Technology Evaluation and Management, research project name: home social robot and service development system, project serial number: 10077633]

Recognizing and processing natural language so that a machine understands human language can be referred to as natural language understand, and natural language understanding can be used in various fields. For example, natural language understanding can be used in a question-and-answer system that automatically provides an answer to a query by recognizing a user's query.

In spite of the same meaning, human language can be expressed in various ways, so in natural language understanding, it may be required to grasp the same meaning from various expressed sentences. Accordingly, preparing a plurality of sentences corresponding to the same meaning, that is, paraphrase sentences, may be an important basis for understanding natural language.

The technical idea of the present invention provides a system and method for automatically generating paraphrase sentences based on an ontology.

In order to achieve the above object, a system for generating a paraphrase sentence according to the technical idea of the present invention includes semantic components including each of the ontology components corresponding to tokens extracted from the query sentence. A feature extraction unit that acquires resources and generates a frame having a structure in which tokens can be replaced in a pattern and query sentence, which is a combination of semantic resources, correlates the pattern and frame, and stores the corresponding pattern and frame in the pattern storage unit The pattern indexing unit may include a sentence generating unit that generates a paraphrase sentence of the query sentence based on the frames corresponding to the pattern stored in the pattern storage unit.

According to an exemplary embodiment of the present invention, the feature extractor may acquire a plurality of query sentences through a network.

According to an exemplary embodiment of the present invention, the feature extraction unit receives a natural language processing unit that extracts tokens by performing natural language processing on a query sentence, and semantic resources corresponding to the tokens from the knowledge base, and generates a pattern and a frame. It may include a pattern generator.

According to an exemplary embodiment of the present invention, the pattern indexing unit may store patterns each including semantic resources of a different order as different patterns in the pattern storage unit.

According to an exemplary embodiment of the present invention, the sentence generation unit includes a pattern search unit that searches frames corresponding to the pattern generated by the feature extraction unit from the pattern storage unit, and at least one preliminary parametric by applying tokens to the searched frames. A vocabulary replacement unit for generating a phrase sentence, and a sentence post-processing unit for generating a paraphrase sentence by modifying at least one preliminary paraphrase sentence according to a natural language rule.

According to an exemplary embodiment of the present invention, the vocabulary replacement unit may generate at least one preliminary paraphrase sentence by applying synonyms of tokens to retrieved frames based on semantic resources.

According to an exemplary embodiment of the present invention, the vocabulary replacement unit may generate at least one preliminary paraphrase sentence by changing the structure of the searched frame.

According to an exemplary embodiment of the present invention, the vocabulary replacement unit may generate an additional preliminary paraphrase sentence by changing the position of the adverb phrase in the searched frame.

According to an exemplary embodiment of the present invention, when the searched frame ends with a noun phrase, the vocabulary replacement unit may generate an additional preliminary paraphrase sentence by adding a query terminating phrase.

According to the system and method according to the technical idea of the present invention, accurate and rich paraphrase sentences can be automatically generated by analyzing sentences based on the verified ontology.

In addition, according to the system and method according to the technical idea of the present invention, the performance of tasks based on natural language understanding, such as a query response system, knowledge extraction, etc., can be significantly improved due to the abundantly provided paraphrase sentences. I can.

In addition, according to the system and method according to the technical idea of the present invention, high-quality data for machine learning can be prepared due to the abundantly provided paraphrase sentences, and accordingly, the utilization of machine learning can be remarkably expanded. I can.

The effects obtained in the exemplary embodiments of the present invention are not limited to the above-mentioned effects, and other effects that are not mentioned are common knowledge in the technical field to which the exemplary embodiments of the present invention belong from the following description. It can be clearly derived and understood by those who have. That is, unintended effects of implementing the exemplary embodiments of the present invention may also be derived from the exemplary embodiments of the present invention by a person of ordinary skill in the art.

1 is a block diagram showing a system and an input/output relationship thereof according to an exemplary embodiment of the present invention.
FIG. 2 is a block diagram illustrating an example of a feature extraction unit of FIG. 1 according to an exemplary embodiment of the present invention.
3A and 3B illustrate examples of query sentences, patterns, and frames according to exemplary embodiments of the present invention.
4 shows examples of patterns and frames stored in the pattern storage unit of FIG. 1 according to an exemplary embodiment of the present invention.
Fig. 5 is a block diagram showing an example of the sentence generating unit of Fig. 1 according to an exemplary embodiment of the present invention.
6A and 6B are diagrams illustrating examples of operations of the vocabulary replacement unit of FIG. 5 according to exemplary embodiments of the present invention.
Fig. 7 is a flow chart showing a method of generating a paraphrase sentence according to an exemplary embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. Embodiments of the present invention are provided to more fully describe the present invention to those with average knowledge in the art. The present invention can be applied to various changes and may have various forms, and specific embodiments will be illustrated in the drawings and described in detail. However, this is not intended to limit the present invention to a specific form disclosed, it should be understood to include all changes, equivalents, and substitutes included in the spirit and scope of the present invention. In describing each drawing, similar reference numerals are used for similar elements. In the accompanying drawings, the dimensions of the structures are enlarged or reduced than actual ones for clarity of the present invention.

Terms used in the present application are only used to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions, unless the context clearly indicates otherwise. In this application, the terms "comprises" or "have" are intended to indicate the presence of features, numbers, steps, actions, elements, parts or combinations thereof described in the specification, one or more other features. It should be understood that the existence or addition possibilities of fields or numbers, steps, actions, components, parts or combinations thereof are not excluded in advance.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by a person skilled in the art to which the present invention pertains. Terms, such as those defined in the dictionary used in general, should be interpreted as having meanings consistent with meanings in the context of related technologies, and should be interpreted as ideal or excessively formal meanings unless explicitly defined in the present application. Does not work.

In the following drawings and description, a component indicated or described as a single block may be a hardware block or a software block. For example, each of the components may be an independent hardware block that exchanges signals with each other, or may be a software block executed in one processor. Also, in the present specification, "system" or "database" may refer to a computing system including at least one processor and a memory accessed by the processor.

1 is a block diagram showing a system and an input/output relationship thereof according to an exemplary embodiment of the present invention. As shown in FIG. 1, the paraphrase generation system 100 may be communicatively connected to the knowledge base 300 and the pattern storage unit 500. In addition, the paraphrase generation system 100 may receive a query sentence as a natural language, and may generate a paraphrase sentence having the same meaning as the query sentence by communicating with the knowledge base 300 and the pattern storage unit 500. As will be described later, the paraphrase generation system 100 may extract a pattern from the query sentence by referring to the knowledge base 300, and store a plurality of frames corresponding to the pattern in the pattern storage unit 500 together with the pattern. I can. The blocks 100, 300, and 500 shown in FIG. 1 may communicate with each other through a network or through a dedicated channel for one-to-one communication. In addition, two or more of the blocks 100, 300, and 500 shown in FIG. 1 may be included in one system (eg, a computing system), and in some embodiments, the pattern storage unit 500 generates a paraphrase. It may be included in the system 100.

The paraphrase generation system 100 may receive a query sentence in various ways. In some embodiments, the paraphrase generation system 100 may include a user interface, and may receive a query sentence from a person, that is, a user through the user interface. In some embodiments, the paraphrase generation system 100 may access a network, such as the Internet, and may actively collect multiple query sentences from the network. As shown in FIG. 1, the paraphrase generation system 100 may include a feature extracting unit 120, a pattern indexing unit 140, and a sentence generating unit 160.

The feature extractor 120 may acquire semantic resources including each of the ontology components corresponding to tokens extracted from the query sentence. Ontology is a representation of things that exist or can be recognized by humans in a form that can be handled by a computer. Ontology components are entities (E) (or instances), classes (C), properties (property). ; P), value (V) can be included. Additionally, the ontology components may further include a relationship (attribute between entities or between classes), a function term, a restriction, a rule, an event, and the like. The knowledge base 300 may store vast amounts of knowledge data based on an ontology. For example, the knowledge base 300 may include knowledge data expressed using a Resource Description Framework (RDF), and a triple ( triple) can be used. The knowledge base 300 may return a triple in response to a query, such as a SPARQL Protocol and RDF Query Language (SPARQL) query. In some embodiments, the feature extraction unit 120 may transmit a query sentence to a natural language recognition system (not shown) external to the paraphrase generation system 100, and may transfer tokens extracted from the query sentence from the natural language recognition system. Can receive. In some embodiments, the feature extraction unit 120 may generate semantic resources by directly processing a query sentence, as described later with reference to FIG. 2.

The feature extractor 120 may generate a pattern and a frame from semantic resources obtained from a query sentence based on the ontology. As will be described later with reference to FIGS. 3A and 3B, the pattern is a combination of semantic resources and may correspond to the meaning indicated by the query sentence. That is, sentences corresponding to the same pattern may be referred to as paraphrase sentences. In addition, as will be described later with reference to FIGS. 3A and 3B, the frame may have a structure in which tokens extracted from the query sentence can be replaced in the query sentence. As will be described later with reference to FIG. 4, one pattern may correspond to at least one frame, and patterns and frames corresponding to each other may be stored in the pattern storage unit 500 by the pattern index unit 140. The characteristic extracting unit 120 may provide a pattern and a frame to the pattern indexing unit 140, and may provide a token list and a pattern including the extracted tokens to the sentence generating unit 160.

The pattern index unit 140 may receive a pattern and a frame generated from the query sentence from the feature extraction unit 120. The pattern index unit 140 may correlate patterns and frames, and may store patterns and frames corresponding to each other in the pattern storage unit 500. Accordingly, when a pattern stored in the pattern storage unit 500 is searched (eg, by the sentence generating unit 160), corresponding frames may be searched together. In some embodiments, the pattern index unit 140 may search the pattern received from the feature extraction unit 120 in the pattern storage unit 500, and the pattern received from the feature extraction unit 120 is a pattern storage unit. If already stored in 500 and the frame received from the feature extraction unit 120 is also stored in the pattern storage unit 500, storage of the pattern and frame received from the feature extraction unit 120 can be omitted. have. Structures of patterns and frames stored in the pattern storage unit 500 by the pattern index unit 140 will be described later with reference to FIG. 4.

The sentence generating unit 160 may receive a pattern and a token list corresponding to the query sentence from the feature extracting unit 120, and may generate a paraphrase sentence of the query sentence. The sentence generating unit 160 may search for a pattern in the pattern storage unit 500, and may obtain the searched pattern and frames corresponding thereto from the pattern storage unit 500. For example, the sentence generation unit 160 may search the pattern received from the feature extraction unit 120 in the pattern storage unit 500, and based on the token list and frames corresponding to the searched pattern, the query sentence and the A sentence having the same meaning, that is, a paraphrase sentence, can be generated. An example of the sentence generating unit 160 will be described later with reference to FIG. 5.

As will be described later with reference to the drawings, the paraphrase generation system 100 may generate a paraphrase sentence of a query sentence by using the ontology-based knowledge base 300. Accordingly, the query sentence can be analyzed based on the verified ontology, and accurate and rich paraphrase sentences can be automatically generated. In addition, due to the paraphrase sentences generated by the paraphrase generation system 100, the performance of tasks based on natural language understanding, such as a question-and-answer system and knowledge extraction, can be remarkably improved. Quality data can be properly prepared.

2 is a block diagram showing an example of the feature extraction unit 120 of FIG. 1 according to an exemplary embodiment of the present invention, and FIGS. 3A and 3B are query sentences, patterns, and patterns according to exemplary embodiments of the present invention. Examples of frames are shown. As described above with reference to FIG. 1, the feature extraction unit 120 ′ of FIG. 2 may receive a query sentence and may generate a pattern and a frame corresponding to the query sentence with reference to the knowledge base 300. . As shown in FIG. 2, the feature extraction unit 120 ′ may include a natural language processing unit 122 and a pattern generation unit 124, and FIGS. 2, 3A and 3B hereinafter will be described with reference to FIG. 1. Will be explained.

In some embodiments, the feature extraction unit 120 ′ may extract tokens by processing a query sentence in natural language, and may obtain semantic resources corresponding to the tokens from the knowledge base 300. As described above with reference to FIG. 1, in some embodiments, different from the example of FIG. 2, the feature extraction unit 120 of FIG. 1 directly processes the query sentence, but the paraphrase generation system 100 It is also possible to pass a query sentence to the natural language processing system of the company and receive tokens from it.

The natural language processing unit 122 may extract tokens from the query sentence. For example, as shown in FIG. 3A, for the first query sentence Q1 of "What is the population of Seoul?", the natural language processing unit 122 may be used to "Seoul", "Us", "Population", and "A". You can create a list of tokens including "?". In some embodiments, the token extracted by the natural language processing unit 122 may generate not only words but also characteristics. For example, the token may include parts of speech (eg, nouns, adverbs of adjectives, etc.), or may include sentence elements (eg, subjects, objects, adverbs, etc.) in a query sentence. The natural language processing unit 122 may extract tokens in an arbitrary manner. For example, the natural language processing unit 122 may extract tokens by referring to the knowledge base 300. For example, the natural language processing unit 122 may determine the part-of-speech of a word included in the query sentence based on entities and attributes of the knowledge base 300. Further, in some embodiments, as described later with reference to FIGS. 6A and 6B, the extracted tokens may have a tree structure.

The pattern generator 124 may receive tokens from the natural language processing unit 122 and may acquire semantic resources corresponding to the received tokens from the knowledge base 300. For example, as shown in FIG. 3A, for the first query sentence Q1 of “What is the population of Seoul?”, an entity “<Entity-City>” may be obtained as a semantic resource of “Seoul” and , As a semantic resource of "population", the property "<Property_Population>" may be acquired. Accordingly, the first pattern P1 is semantic resources and may include “<Entity-City>” and “<Property_Population>”. In addition, as shown in FIG. 3B, with respect to the second query sentence Q2 of "Please tell me the number of population of Seoul?", an entity "<Entity_City>" may be obtained as a semantic resource of "LA", and " The attribute "<Peoperty_Population>" can be acquired as a semantic resource of "number of population". The second pattern P2 may include “<Entity-City>” and “<Property_Population>”, and thus may be the same as the first pattern P1.

In some embodiments, the pattern may define a combination of semantic resources as well as an order of semantic resources. That is, even if patterns including the same semantic resources have different orders of semantic resources, corresponding patterns may be stored in the pattern storage unit 500 by the pattern index unit 140 as different patterns.

In addition, the pattern generation unit 124 may generate a frame based on the query sentence and the pattern. For example, as shown in FIG. 3A, based on the first query sentence Q1 and the first pattern P1, the pattern generating unit 124 is the first frame F11, that is, “<Entity_City>”. <Property_Population>?" can be created. In the first frame F11, an instance such as "Seoul" and "LA" may be applied (or substituted) to a semantic resource corresponding to an entity, that is, "<Entity_City>", and a semantic resource corresponding to an attribute, That is, for "<Property_Population>", one of synonyms such as "population" and "population number" may be applied. Accordingly, when the tokens included in the query sentence analyzed in the first pattern P1 in the first frame F11 as a query sentence different from the first query sentence Q1 are applied to the first frame F11, the corresponding A paraphrase sentence of the query sentence can be generated. Similarly, as shown in FIG. 3B, based on the second query sentence Q2 and the second pattern P2, the pattern generating unit 124 performs a second frame F12, that is, <Entity_City>. Property_Population> please tell me" can be created. Similar to the first frame F11, in the second frame F12, an instance such as "Seoul", "LA", etc. may be applied to the semantic resource corresponding to the entity, that is, "<Entity_City>". One of synonyms such as "population" and "population count" may be applied to the corresponding semantic resource, that is, "<Property_Population>". Accordingly, when the tokens included in the query sentence analyzed in the first pattern P1 in the second frame F12 as a query sentence different from the second query sentence Q2 are applied to the second frame F12, the corresponding A paraphrase sentence of the query sentence can be generated.

In the examples of FIGS. 3A and 3B, since the first query sentence Q1 and the second query sentence Q2 correspond to the same pattern (that is, because the first pattern P1 and the second pattern P2 are identical) , By applying the tokens of the first query sentence Q1 to the second frame F12, a paraphrase sentence of the first query sentence Q1, that is, “Please tell me the population of Seoul” may be generated. Similarly, by applying the tokens of the second query sentence Q2 to the first frame F11, a paraphrase sentence of the second query sentence Q2, that is, “What is the population of LA?” may be generated.

4 shows examples of patterns and frames stored in the pattern storage unit 500 of FIG. 1 according to an exemplary embodiment of the present invention. As described above with reference to FIG. 1, the pattern index unit 140 may correspond to a pattern and a frame, and may store a pattern and a frame corresponding to each other in the pattern storage unit 500. In the following, FIG. 4 will be described with reference to FIG. 1.

As described above with reference to FIGS. 3A and 3B, a plurality of frames may correspond to one pattern, and after analyzing a pattern corresponding to a given query sentence, tokens of the query sentence are added to a plurality of frames corresponding to the pattern. By applying, multiple paraphrase sentences can be generated. Accordingly, as shown in FIG. 4, the pattern storage unit 500 may include a plurality of groups G1 to Gn corresponding to each of the plurality of patterns P1 to Pn (n is greater than 1). Large integer). Each of the plurality of groups G1 to Gn may include one pattern and a plurality of frames corresponding to the pattern. For example, the first group G1 may include the first pattern P1 and a plurality of frames F11 to F1x corresponding thereto (x is an integer greater than 1), and the second group G2 May include a second pattern P2 and a plurality of frames F21 to F2y corresponding thereto (y is an integer greater than 1), and the third group G3 is a third pattern P3 and corresponding thereto. It may include a plurality of frames (Fn1 to Fnz) (z is an integer greater than 1).

The pattern index unit 140 may store patterns and frames in the pattern storage unit 500 so as to have a structure as shown in FIG. 4, while the sentence generation unit 160 is due to the structure shown in FIG. 4. Thus, a plurality of frames corresponding to the searched pattern may be obtained. For example, when the pattern received from the feature extraction unit 120 by the sentence generating unit 160 corresponds to the second pattern P2, the sentence generating unit 160 stores the second pattern P2 as a pattern storage unit. It is possible to search at 500 and obtain a plurality of frames F21 to F2y corresponding to the second pattern P2. It will be appreciated that the example shown in FIG. 4 is only to show a pattern and a correspondence relationship between a plurality of frames, and that arbitrary data structures may be used to implement the structure shown in FIG. 4. In addition, in some embodiments, the frame may have a tree structure as will be described later with reference to FIG. 6A.

5 is a block diagram showing an example of the sentence generator 160 of FIG. 1 according to an exemplary embodiment of the present invention. As described above with reference to FIG. 1, the sentence generating unit 160 ′ of FIG. 5 may receive a pattern and a frame corresponding to a query sentence from the feature extracting unit 120, and stored in the pattern storage unit 500. A paraphrase sentence of the query sentence can be generated by referring to the pattern and frame. As shown in FIG. 5, the sentence generating unit 160 ′ may include a pattern search unit 162, a vocabulary replacement unit 164, and a sentence post-processing unit 166. Hereinafter, FIG. It will be explained with reference.

The pattern search unit 162 may receive a pattern from the feature extraction unit 120 and may search frames corresponding to the pattern received from the feature extraction unit 120 in the pattern storage unit 500. As described above with reference to FIG. 4, since the pattern storage unit 500 may be structured to include patterns and frames corresponding to each other, the pattern search unit 162 searches the received pattern in the pattern storage unit 500. By doing so, frames corresponding to the pattern may be received from the pattern storage unit 500.

The vocabulary replacement unit 164 may receive a token list from the feature extraction unit 120 and may receive frames from the pattern search unit 162. As described above with reference to FIG. 1, the token list may include tokens extracted from the query sentence, and the vocabulary replacement unit 164 uses the frames received from the pattern search unit 162 for tokens included in the token list. By applying to, we can generate preliminary paraphrase sentences. For example, when a token list extracted from the first query sentence Q1 of FIG. 3A is received and the second frame F12 of FIG. 3B is received from the pattern search unit 162, the vocabulary replacement unit 164 May generate a preliminary paraphrase sentence "Please tell me the number of population in Seoul" by applying "Seoul" and "Population" included in the token list to the second frame F12. Similarly, when receiving the token list extracted from the second query sentence Q2 of FIG. 3B and receiving the first frame F11 of FIG. 3A from the pattern search unit 162, the vocabulary replacement unit 164 A preliminary paraphrase sentence "What is the number of populations of LA?" may be generated by applying "LA" and "the number of populations" included in the token list to the first frame F11.

In some embodiments, the vocabulary replacement unit 164 may generate a preliminary paraphrase sentence by applying a synonym of a word corresponding to the token to the frame. For example, when receiving a token list extracted from the second query sentence Q2 of FIG. 3B and receiving the first frame F11 of FIG. 3A from the pattern search unit 162, the vocabulary replacement unit 164 Is a preliminary paraphrase sentence, "What is LA's population?" In addition, "What is the number of people in Los Angeles?" or "What is the number of people in Los Angeles?" may be generated based on the second frame F12. In addition, in some embodiments, the vocabulary replacement unit 164 may generate a preliminary paraphrase sentence by changing the structure of the frame. Examples of the vocabulary replacement unit 164 that generates a preliminary paraphrase sentence by changing the structure of the frame will be described later with reference to FIGS. 6A and 6B.

The sentence post-processing unit 166 may receive a preliminary paraphrase sentence from the vocabulary replacement unit 164, and may generate a paraphrase sentence by modifying the preliminary paraphrase sentence according to natural language rules. The preliminary paraphrase sentence in which tokens are applied to the frame by the vocabulary replacement unit 164 may include a portion in which the natural language rule is violated. For example, the preliminary paraphrase sentence may include inappropriate investigations and/or endings, and the sentence post-processing unit 166 may generate the paraphrase sentences as natural sentences by modifying the investigations and/or endings based on natural language rules. I can.

6A and 6B are diagrams illustrating examples of the operation of the vocabulary replacement unit 164 of FIG. 5 according to exemplary embodiments of the present invention. Specifically, FIGS. 6A and 6B show examples of the vocabulary replacement unit 164 generating preliminary paraphrase sentences by changing the structure of the frame. The vocabulary replacement unit 164 may change the structure of the frame or may change the structure of a preliminary paraphrase sentence in which tokens are applied to the frame. Hereinafter, FIGS. 6A and 6B will be described with reference to FIG. 5, and overlapping contents in the description of FIGS. 6A and 6B will be omitted. For convenience of explanation, FIGS. 6A and 6B illustrate examples of changing the structure of a preliminary paraphrase sentence to which tokens are applied to a frame.

6A, the vocabulary replacement unit 164 may generate an additional preliminary paraphrase sentence by changing the position of the adverb phrase in the frame. For example, the preliminary paraphrase sentence "Which city is the Smithsonian Museum in the United States?" may be expressed in a tree structure as shown in the upper part of FIG. 6A. In FIG. 6A, reference numerals indicated under tokens are characteristics of tokens, and may indicate parts of speech and sentence elements. For example, "VNP" may represent "noun + is" as a positive designation phrase, "NP_SBJ" may represent that a body language (eg, noun, pronoun, rhetoric) is used as a subject, and "NP_AJT" is It can indicate that body language is used as an adverb (terminal modifier). In addition, "NP_MOD" may indicate that the body language is used as a guan-type word (a body language modifier), and "DP" may represent a guan-type phrase.

The vocabulary replacement unit 164 may change the position of the sub-tree corresponding to the adverb phrase. For example, as shown in the lower part of FIG. 6B, the vocabulary replacement unit 164 may change the position of the sub-tree corresponding to “in any city in the United States” and the “Smithsonian Art Museum”. Accordingly, a preliminary paraphrase sentence of "Which city in the United States has Smithsonian Art Museum?" may be generated. The sentence post-processing unit 166 of FIG. 5 may finally generate a paraphrase sentence, that is, “Which city in the United States has the Smithsonian Art Museum?” by modifying the investigation.

Referring to FIG. 6B, the vocabulary replacement unit 164 may generate an additional preliminary paraphrase sentence with one or more phrases. For example, a preliminary paraphrase sentence “Which city is the Smithsonian Museum?” may be structured as shown on the left side of FIG. 6B. When the searched frame ends with a noun phrase, the vocabulary replacement unit 164 may generate an additional preliminary paraphrase sentence by adding a query terminating phrase. For example, as shown on the right side of FIG. 6B, the vocabulary replacement unit 164 adds “Where” as a query terminating phrase in the frame ending with the noun phrase “shown”, so that a preliminary paraphrase sentence, namely “Smith "Where is the Soonian Museum of Art?"

Fig. 7 is a flow chart showing a method of generating a paraphrase sentence according to an exemplary embodiment of the present invention. For example, the method of FIG. 7 may be performed by the paraphrase generation system 100 of FIG. 1. Specifically, steps S10, S20, and S30 may be performed by the feature extraction unit 120 of FIG. 1, and steps S40, S50, and S60 may be performed by the sentence generation unit 160. In the following, FIG. 7 will be described with reference to FIG. 1.

Referring to FIG. 7, an operation of receiving a query sentence may be performed in step S10. For example, the feature extraction unit 120 may receive a query sentence from a user through a user interface, or may collect query sentences from a network.

In step S20, an operation of acquiring semantic resources may be performed. For example, the feature extraction unit 120 may acquire semantic resources corresponding to the extracted tokens from the knowledge base 300 by processing a query sentence in natural language. As described above with reference to the drawings, since the knowledge base 300 may be constructed based on an ontology and may include verified knowledge data, it may be used to generate a paraphrase sentence of a query sentence. The natural language processing of the query sentence may be performed outside the paraphrase generation system 100 in some embodiments, or may be performed by the feature extraction unit 120 in some embodiments.

In step S30, an operation of generating a pattern and a frame may be performed. For example, the feature extraction unit 120 may generate a pattern as a combination of semantic resources acquired based on the query sentence, and may generate a frame having a structure in which tokens can be replaced in the sentence sentence. In addition, the feature extraction unit 120 may generate a token list including the extracted tokens.

In step S40, an operation of searching for a pattern may be performed. For example, the pattern storage unit 500 may store patterns and frames corresponding to each other, and one pattern may correspond to a plurality of frames. The sentence generating unit 160 may search the pattern generated in step S30 and the pattern storage unit 500 may obtain frames corresponding to the searched pattern from the pattern storage unit 500.

In step S50, an operation of applying tokens to the frame may be performed. For example, the sentence generation unit 160 may generate preliminary paraphrase sentences by applying tokens included in the token list to the frame obtained from the pattern storage unit 500. In some embodiments, the sentence generator 160 may generate an additional preliminary paraphrase sentence by using a synonym of the token, or may generate an additional preliminary paraphrase sentence by changing the structure of a frame or a preliminary paraphrase sentence. have.

In step S60, post-processing of the sentence may be performed. For example, the preliminary paraphrase sentence generated in step S50 may include a part in which the natural language rule is violated, and the sentence generation unit 160 examines the preliminary paraphrase sentence and/or corrects the ending according to the natural language rule. Paraphrase sentences can be finally created.

As described above, exemplary embodiments have been disclosed in the drawings and the specification. In the present specification, embodiments have been described using specific terms, but these are only used for the purpose of describing the technical idea of the present invention, and are not used to limit the meaning or the scope of the present invention described in the claims. . Therefore, those of ordinary skill in the art will understand that various modifications and equivalent other embodiments are possible therefrom. Accordingly, the true technical scope of the present invention should be determined by the technical spirit of the appended claims.

Claims (9)

As a system for generating paraphrase sentences,
Acquires semantic resources including each of the ontology components corresponding to tokens extracted from the query sentence, and creates a frame having a pattern that is a combination of the semantic resources and a structure in which the tokens can be replaced in the query sentence A feature extraction unit configured to;
A pattern storage unit for storing a plurality of frames including a plurality of patterns and at least one frame corresponding to each of the plurality of patterns;
A pattern indexing unit configured to associate patterns and frames with each other and store the corresponding patterns and frames in the pattern storage unit; And
And a sentence generator configured to generate a paraphrase sentence of the query sentence based on frames corresponding to the pattern stored in the pattern storage unit,
And the feature extraction unit comprises a pattern generation unit configured to receive the semantic resources corresponding to the tokens from a knowledge base and generate the pattern and the frame. The method according to claim 1,
And the feature extraction unit is configured to obtain a plurality of query sentences through a network. The method according to claim 1,
The feature extraction unit,
And a natural language processing unit configured to extract the tokens by performing natural language processing on the query sentence. The method according to claim 1,
And the pattern indexing unit is configured to store patterns each including semantic resources of a different order as different patterns in the pattern storage unit. The method according to claim 1,
The sentence generating unit,
A pattern search unit configured to search in the pattern storage unit for frames corresponding to the pattern generated by the feature extraction unit;
A vocabulary replacement unit configured to generate at least one preliminary paraphrase sentence by applying the tokens to the retrieved frames; And
And a sentence post-processing unit configured to generate the paraphrase sentence by modifying the at least one preliminary paraphrase sentence according to a natural language rule. The method according to claim 5,
And the vocabulary replacement unit is configured to generate the at least one preliminary paraphrase sentence by applying synonyms of the tokens to the retrieved frames based on the semantic resources. The method according to claim 5,
And the vocabulary replacement unit is configured to generate the at least one preliminary paraphrase sentence by changing a structure of the searched frame. The method according to claim 7,
And the vocabulary replacement unit is configured to generate an additional preliminary paraphrase sentence by changing a position of an adverb phrase in the searched frame. The method according to claim 7,
And the vocabulary replacement unit is configured to generate an additional preliminary paraphrase sentence by adding a query terminating phrase when the searched frame ends with a noun phrase.

Images