KR102059743B1 - Method and system for providing biomedical passage retrieval using deep-learning based knowledge structure construction - Google Patents

Abstract

A medical literature passage search method and system using deep learning based knowledge structure generation method is presented. According to one or more exemplary embodiments, a method of retrieving a medical document phrase may include extracting a paragraph unit paragraph from a medical document; Indexing the extracted phrase and searching for an initial query and an appropriate phrase inputted to obtain an initial phrase search result; Extracting key concepts from a predetermined number of higher phrases through natural language processing from the initial phrase search result, and then extracting association information between the key concepts; Generating a knowledge structure using association information between the core concepts; Exploring key concepts of the constructed knowledge structure; Adding a concept between the core concept in the searched knowledge structure and the initial query to a query set; And performing a medical phrase search again by utilizing the expanded set of query terms.

Description

METHOD AND SYSTEM FOR PROVIDING BIOMEDICAL PASSAGE RETRIEVAL USING DEEP-LEARNING BASED KNOWLEDGE STRUCTURE CONSTRUCTION}

The following embodiments are related to a method and system for retrieving medical text passages using a deep learning based knowledge structure generation method, and more specifically, to use the association information between a knowledge structure generation technology using deep learning and a concept in the knowledge structure. One medical literature passage relates to methods and systems.

In general, a search website providing a search service has a search result corresponding to the query when the user inputs a query (for example, a web site including the query, an article including the query, and a file name including the query). To the user).

In particular, the query language used in the medical field has a very large number of medical jargons and abbreviations unlike general query terms (for example, alpha fetoprotein is widely used as AFP, AFP-L3, etc.). For example, serum has a different meaning when used with albumin and when it is used as a test item). Therefore, even if the document or phrase does not contain the corresponding query, the development of a search technique that can be included in the search results in the above case is necessary. Required.

Examples of medical passages described below include medical papers from PubMed Central or many medical textbooks used by practitioners in practice.As of 2018, more than 780,000 papers are available from PubMed Central. It is an online medical database with high quality, and serves as a space to provide healthcare workers with information needed for diagnosis and research.

However, most medical practitioners need to extract only the most essential phrases from the searched documents and provide them to the user in order to be an effective decision support system due to the time constraint that many diagnosis must be performed in a limited time.

Conventional phrase retrieval methods have been used to extract related words from external resources or ontologies, or to mechanically extract relevant words from documents searched above.

As an example of an existing phrase search method, a knowledge structure generation method may be provided. Korean Registration No. 10-1538998 relates to a method and apparatus for providing a search service based on this knowledge structure. It extracts major keywords in a document based on BoW (Bag-of-words) and Word frequency, We measure word association by measuring -occurrence information. Although it is possible to extract the knowledge structure from a single document, it is difficult to construct a single integrated knowledge structure from a plurality of documents.

In addition, in the case of using the knowledge structure retrieval method, the knowledge structures of the initially retrieved contents are constructed, the distances of the initial query words are measured within the knowledge structure, and the corresponding information is defined as proximity to rearrange the initial search results. However, there is a disadvantage in that other hidden documents other than the initially searched documents cannot be searched.

In addition, in the case of a service using a knowledge structure, there are a recommendation service for recommending various categories to a user and an education service for automatically generating a problem based on a knowledge structure built on Wikipedia. However, it lacks a function to provide a service utilizing a knowledge structure in the medical field.

Korea Registration No. 10-1538998

Embodiments describe a method and system for searching a medical literature passage using a deep learning based knowledge structure generation method, and more specifically, automatically generates an integrated knowledge structure from a plurality of medical related phrases searched above through an initial query. In addition, by automatically extracting keywords related to query expansion from the generated knowledge structure, it provides a technique to significantly improve the phrase search performance of medical documents.

Embodiments can automatically build a knowledge structure integrated from a plurality of resources through a deep learning-based text analysis technique, thereby eliminating statistical methods that need to be considered for building and integrating knowledge structures for each resource. In addition, the present invention provides a method and system for retrieving medical text passages using a deep learning-based knowledge structure generation method that enables the retrieval of an unretrievable document by implicitly identifying words related to an initial query.

According to one or more exemplary embodiments, a method of retrieving a medical document phrase may include extracting a paragraph unit paragraph from a medical document; Indexing the extracted phrase and searching for an initial query and an appropriate phrase inputted to obtain an initial phrase search result; Extracting key concepts from a predetermined number of higher phrases through natural language processing from the initial phrase search result, and then extracting association information between the key concepts; Generating a knowledge structure using association information between the core concepts; Exploring key concepts of the constructed knowledge structure; Adding a concept between the core concept in the searched knowledge structure and the initial query to a query set; And performing a medical phrase search again by utilizing the expanded set of query terms.

In the obtaining of the initial phrase search result, the initial query word and the appropriate phrase input by the user may be searched through TF-IDF, BM25, and LM techniques, and then the search result may be obtained by combining.

The extracting the correlation information between the key concepts may include extracting key concepts through natural language processing of a predetermined number of upper phrases in the initial phrase search result; And performing embedding to convert each core concept into a vector by performing a deep learning based Word2Vec algorithm using the extracted core concepts.

Generating a knowledge structure using the association information between the key concepts may include constructing a distance matrix between the key concepts extracted from the predetermined number of higher phrases and displaying the similarity between each distance obtained. Generating an association matrix using the numerical values of; And generating the knowledge structure using the association matrix.

The searching of the core concept of the knowledge structure may search for the most important keyword in the knowledge structure by using a pagerank algorithm.

Adding a concept between the core concept in the knowledge structure and the initial query word to the query set may include adding a concept word placed between the key concept in the knowledge structure and the initial query word to the query set. have.

The knowledge structure may be automatically generated from the plurality of medical related phrases searched above through the initial query, and the keywords related to the query expansion may be automatically extracted from the generated knowledge structure.

According to another embodiment of the present invention, a medical document phrase search system includes a phrase extracting unit that extracts a paragraph unit paragraph from a medical document; A phrase search unit for indexing the extracted phrase and searching for an initial query word and an appropriate phrase inputted to obtain an initial phrase search result; An association information extracting unit for extracting key concepts from a predetermined number of higher phrases through natural language processing from the initial phrase search result, and then extracting association information between the key concepts; A knowledge structure generation unit generating a knowledge structure using the association information between the core concepts; A core concept search unit for searching a core concept of the constructed knowledge structure; And a query extension unit for adding a concept between the core concept in the searched knowledge structure and the initial query to a query set, and re-search the medical phrase using the expanded query set.

The phrase search unit may search for the initial query input by the user and the appropriate phrases through the TF-IDF, BM25, and LM techniques, and then obtain the search result by combining them.

The correlation information extracting unit extracts key concepts from a predetermined number of upper phrases through natural language processing from the initial phrase search result, and then performs an embedding operation of converting each key concept into a vector by performing a deep learning based Word2Vec algorithm. can do.

The core concept search unit may search for a key keyword in the knowledge structure by using a pagerank algorithm.

The query extension unit may add a conceptual word placed between a core concept in the knowledge structure and the initial query to a query set.

The knowledge structure may be automatically generated from the plurality of medical related phrases searched above through the initial query, and the keywords related to the query expansion may be automatically extracted from the generated knowledge structure.

According to another exemplary embodiment, a phrase search method includes extracting a phrase of a paragraph unit from a plurality of documents; Indexing the extracted phrase and searching for an initial query and an appropriate phrase inputted to obtain an initial phrase search result; Extracting key concepts from a predetermined number of higher phrases through natural language processing from the initial phrase search result, and then extracting association information between the key concepts; Generating a knowledge structure using association information between the core concepts; Exploring key concepts of the constructed knowledge structure; Adding a concept between the core concept in the searched knowledge structure and the initial query to a query set; And re-doing the phrase search using the expanded set of query terms.

The knowledge structure may be automatically generated from the plurality of related phrases searched above through the initial query, and the keywords related to the query expansion may be automatically extracted from the generated knowledge structure.

According to embodiments, by automatically constructing an integrated knowledge structure from a plurality of resources through a deep learning-based text analysis technique, statistical methods that need to be considered for constructing and integrating knowledge structures for each resource may be omitted. It is possible to provide a method and system for searching a medical text passage using a deep learning based knowledge structure generation method.

According to embodiments, by implicitly identifying words related to the initial query word, it is possible to search for a document that has not been initially searched and to obtain more improved search performance. This may be used to provide a search service that can be actively used in the medical field, in particular for nomenclature and various forms of abbreviations and synonyms.

1 is a block diagram schematically illustrating a medical document phrase search system using a deep learning based knowledge structure generation method according to an exemplary embodiment.
2 is a flowchart illustrating a method of searching for a medical document phrase using a deep learning based knowledge structure generation method according to an exemplary embodiment.
3 illustrates an example of a phrase search result associated with an AST according to an exemplary embodiment.
4 is a diagram illustrating an example of query expansion from a constructed knowledge structure according to an embodiment.

Hereinafter, exemplary embodiments will be described with reference to the accompanying drawings. However, the described embodiments may be modified in many different forms, and the scope of the present invention is not limited to the embodiments described below. In addition, various embodiments are provided to more fully describe the present invention to those skilled in the art. Shape and size of the elements in the drawings may be exaggerated for more clear description.

The present embodiments automatically generate an integrated knowledge structure from a plurality of medical related phrases (eg, sentences or paragraphs related to a query in a medical textbook or medical journal) searched upward through an initial query and from the generated knowledge structure. The present invention relates to a method and system for automatically extracting keywords related to query expansion to significantly improve the phrase search performance of medical documents.

Unlike the method of extracting related words from existing external resources or ontology or the method of mechanically extracting related words from the documents searched at the upper level, the phrase searching method proposed in this embodiment constructs a knowledge structure from the documents searched at the upper level. There are differences from existing query expansion methods in improving the search performance by using words related to the query in the knowledge structure as an extended query word.

In addition, the method proposed in this embodiment has the following contrasts with the method using the existing knowledge structure.

The existing knowledge structure generation method extracts major keywords in a document based on BoW (Bag-of-words) and Word frequency, and measures the association of words by measuring co-occurrence information between the two keywords. Although it is possible to extract the knowledge structure from a single document, it is difficult to construct a single integrated knowledge structure from a plurality of documents. Accordingly, in the present embodiment, an integrated knowledge structure can be automatically constructed from a plurality of documents by utilizing a deep learning based technology for processing natural language based on CNN (Convolutional Neural Network) called Word2Vec. By using the knowledge structure generation method proposed in this embodiment, there is an advantage that the knowledge structure based on a plurality of documents or phrases can be automatically generated compared to the prior art.

In addition, the existing method of utilizing the knowledge structure search constructs the knowledge structures of the initially searched contents, measures the distance of the initial query words in the knowledge structure, and defines the corresponding information as proximity to rearrange the initial search results. However, there is a disadvantage in that other hidden documents other than the initially searched documents cannot be searched. In this embodiment, since an integrated knowledge structure is constructed from an initially searched phrase and used for query expansion rather than rearrangement of documents, it can be used to search for documents that were not initially searched compared to existing patents. .

In addition, the service using the existing knowledge structure is provided with a recommendation service for recommending various categories to users and an education service for automatically generating a problem based on a knowledge structure built on Wikipedia. However, it does not provide services utilizing knowledge structures in the medical field. In the present embodiment, to provide a service utilizing the knowledge structure in the medical field through a service that supports the decision-making in the medical field through a search.

An object of the present invention is to construct a knowledge structure by calculating the core conceptual words and the relationships between the conceptual words from natural texts of medical literature through deep learning-based algorithms, and to expand the conceptual words related to the query words from the constructed knowledge structures. It is to provide medical phrase search service to support decision making by using query expansion and extended query.

1 is a block diagram schematically illustrating a medical document phrase search system using a deep learning based knowledge structure generation method according to an exemplary embodiment.

Referring to FIG. 1, the medical document phrase search system 100 using the deep learning-based knowledge structure generation method according to an embodiment may include a phrase extractor 110, a phrase searcher 120, and an association information extractor ( 130, the knowledge structure generation unit 140, the core concept search unit 150, and the query expansion unit 160. In addition, according to an embodiment, the apparatus may further include a screen configuration unit 170 for constructing and providing a search page to a user and a database 180 for storing various information.

More specifically, the medical document phrase search system using the deep learning-based knowledge structure generation method according to an embodiment is a phrase extraction unit 110 for extracting paragraphs of paragraph units from medical documents, indexing the extracted phrase and input After the initial query is searched and the appropriate phrase, the phrase search unit 120 that obtains the initial phrase search result, and extracts the core concepts through the natural language processing of a predetermined number of upper phrases from the initial phrase search result, Association information extraction unit 130 for extracting the association information, knowledge structure generation unit 140 for generating a knowledge structure using the association information between the core concepts, core concept search unit for searching the core concepts of the built knowledge structure ( 150) and a query extension unit 160 for adding a concept between the core concept in the searched knowledge structure and the initial query to a query set; The medical phrase search may be performed again using the expanded query set through the phrase search unit 120 or a separate phrase research unit.

Accordingly, the user can search for various medical passages desired by the user by accessing the medical literature passage search system (or server) using the deep learning-based knowledge structure generation method through a network using a terminal. . Here, the search for a medical phrase is described as an example, but is not limited to a medical phrase, and may be configured to search a general phrase or a phrase of a specific field.

The medical literature phrase search system using the deep learning based knowledge structure generation method according to an embodiment will be described in more detail with reference to one example below.

2 is a flowchart illustrating a method of searching for a medical document phrase using a deep learning based knowledge structure generation method according to an exemplary embodiment.

Referring to FIG. 2, according to an embodiment, a method of retrieving medical literature passages using a deep learning based knowledge structure generation method may further include a medical passage using a correlation information between a knowledge structure generation technology using deep learning and a concept in a knowledge structure. As a search method, it can be performed as follows.

According to an embodiment of the present invention, a method of searching a medical document passage using a deep learning-based knowledge structure generation method includes extracting a paragraph of a paragraph unit from a medical document (210), indexing the extracted passage, and inputting an initial query and a suitable phrase. In step 220, obtaining an initial phrase search result, extracting key concepts through natural language processing of a predetermined number of upper phrases from the initial phrase search result, and then extracting correlation information between key concepts (230). ), Generating a knowledge structure using the association information between the core concepts (240), searching for the core concepts of the constructed knowledge structure (250), concepts between the core concepts in the searched knowledge structure and the initial query Adding 260 to the query set, and re-running the medical phrase search using the extended query set 270. All.

Unlike the conventional knowledge structure construction method, the present embodiments can automatically construct a knowledge structure integrated from a plurality of resources through a deep learning based text analysis technique. Thus, statistical methods that need to be considered to construct and integrate knowledge structures for each resource can be omitted.

In addition, the query expansion method through the knowledge structure enables to implicitly identify and provide words related to the initial query, thereby enabling the retrieval of a document that was not initially searched, and further improving search performance.

This may be used to provide a search service that can be actively used in the medical field, in particular for nomenclature and various forms of abbreviations and synonyms.

Hereinafter, a method of searching for a medical document phrase using a deep learning based knowledge structure generation method according to an embodiment will be described in more detail.

According to an embodiment of the present invention, a medical document phrase search method using a deep learning based knowledge structure generation method may be performed by using a medical document phrase search system using the deep learning based knowledge structure generation method described in FIG. 1. It can be explained concretely. Here, the medical document phrase search system 100 using the deep learning-based knowledge structure generation method according to an embodiment of the phrase extraction unit 110, phrase search unit 120, association information extraction unit 130, knowledge It may include a structure generator 140, a core concept search unit 150, and a query expansion unit 160.

In operation 210, the phrase extraction unit 110 may extract a paragraph unit of paragraph from the medical document. For example, the phrase extraction unit 110 may extract a paragraph unit paragraph from a plurality of medical documents collected online.

The method of extracting paragraph units from the medical literature can be performed as follows.

You can convert .azw3 and .azw formats to .pdf, .txt, and .doc for e-book file format conversion.

You can extract text / images from the PDFbox library for text and image extraction.

For text pre-processing, text sentences and paragraph break preprocessing can be performed through the OpenNLP library.

References can be extracted at the end of the book for reference extraction.

For reference web crawling, you can crawl the title and / or abstract of references on the web.

Through the above process, it is possible to extract a paragraph from the text extracted from the literature and define it as a phrase.

In operation 220, the phrase search unit 120 may index an extracted phrase and search for an input initial query and an appropriate phrase to obtain an initial phrase search result. For example, the phrase search unit 120 may obtain an initial phrase search result using a bag-of-words (BOW) based search engine.

The phrase search unit 120 indexes the phrases obtained by the paragraph unit obtained above, searches for the initial query input by the user and suitable phrases through the TF-IDF, BM25, and LM techniques, and aggregates them and returns a search result. . The final scores of the phrases retrieved by each technique can be calculated as follows.

[Equation 1]

Where S _j , k are scores of passages using IR model F _k , N is the total number of IR models used, and M is the total number of retrieved phrases. Here, N is 3 and M is set to 100.

3 illustrates an example of a phrase search result associated with an AST according to an exemplary embodiment.

Referring to FIG. 3, phrases related to AST obtained through Equation 1 above and images of the phrases are shown.

In operation 230, the correlation information extraction unit 130 may extract key concepts through natural language processing of a predetermined number of higher phrases in the initial phrase search result, and then extract correlation information between the key concepts.

For example, the correlation information extraction unit 130 may extract key concepts through natural language processing of the upper M phrases, and then perform embedding to convert each core concept into a vector space model based on deep learning.

More specifically, the association information extraction unit 130 may extract key concepts through natural language processing of a predetermined number of higher phrases in the initial phrase search result. Afterwards, the deep learning-based Word2Vec algorithm is performed using the extracted core concepts to perform embedding to convert each core concept into a vector.

Word2vec (mikolov et al., 2013) is one of the deep learning methods for natural language processing used to learn embeddings that take corpus as input and express words in corpus as vectors. In other words, after extracting nouns by processing the searched upper M phrases by natural language, and inputting them into Word2vec, the relationship between the core concepts belonging to the corpus can be learned and expressed as a vector.

Table 1 shows the most similar words, close to Albumin, as an example of Word2Vec implemented for one embodiment.

TABLE 1

Table 1 is an example of Word2Vec implemented for one embodiment, showing the most similar words close to Albumin, and showing a list of the top ten concepts most similar to Albumin.

In operation 240, the knowledge structure generator 140 may generate a knowledge structure by using correlation information between key concepts.

For example, the knowledge structure generation unit 140 may generate a knowledge structure for the upper M phrases by connecting similar concepts using the generated word-vector matrix.

More specifically, the knowledge structure generator 140 constructs a distance matrix between key concepts extracted from a predetermined number of higher phrases and uses the similarity between the obtained distances as numerical values for each row to generate an association matrix. And a knowledge structure using an association matrix.

The existing knowledge structure generation method and the method proposed in the present invention have the following clear differences.

1) Create Association Matrix

The existing technique is to determine the degree to which key nouns appear together in paragraph units. The Paragraph co-occurrences Cosine Similarity (PCS) method can be used to quantify how often the nouns extracted above appear simultaneously in the same paragraph.

[Equation 2]

On the other hand, the medical literature passage search method using the deep learning-based knowledge structure generation method according to an embodiment constructs a distance matrix between concepts extracted from the top M phrases and the similarity between the distances obtained through step 230. Can be used as the value for each row.

2) multidimensional scaling

The previously calculated correlation information may be converted into distance information of 1-7, where 1 represents a high association between two words, and 7 represents a low association between two words. Then, using the PathFinder (Schvaneveldt, 1989) algorithm, the dimension of the association matrix is scaled to remove noise and sharpen the proximity information between key words. Create knowledge structures

According to an embodiment of the present invention, the method of searching a medical document passage using the deep learning based knowledge structure generation method does not use a pathfinder algorithm, unlike the existing method. This is to prevent algorithms designed to remove noise eliminate the implicit relationship between the two concepts.

In operation 250, the core concept search unit 150 may search the core concepts of the constructed knowledge structure.

In particular, the core concept search unit 150 may search for the most important keyword in the knowledge structure by using a PageRank algorithm.

The knowledge structure constructed through the step 240 is in the form of a graph. By using the PageRank algorithm, it is possible to know what is the most important keyword in the knowledge structure. In the present embodiment, a page rank may be modified to be used in accordance with the form of the knowledge structure as shown in the following equation.

[Equation 3]

Where B _i is a set of vertices pointing to v _i , L _j is the number of outgoing links from vertex v _j , and O _i and O _k are vertices v respectively Indicate the number of outgoing edges of _i and v _k . As used herein, PageRank is a weighted version that does not consider the edge weight of the original PageRank.

In operation 260, the query extension unit 160 may add a concept between a core concept in the searched knowledge structure and an initial query to a query set.

More specifically, the query extension unit 160 may add conceptual words that are placed between the core concepts (or key keywords) and the initial query in the knowledge structure to the query set. At this point, an external ontology can be used to determine if it is relevant to medical terminology before it is finally extended.

4 is a diagram illustrating an example of query expansion from a constructed knowledge structure according to an embodiment.

Referring to FIG. 4, as an example of query expansion from the constructed knowledge structure 400, a part of the actual knowledge structure using Albumin 410 as a core concept may be represented. When Aspartate (420) enters the initial query, 'Albumin (410)', 'GOT (431)', 'Glutamic (432)', and 'Activity ()' are placed on the path of Albumin (410) and Asparate (420). 433), and 'Mitochondrial 434' may be included in the initial query set to expand the query.

However, incorporating concepts that are not related to the medical field or are unclear in the query set may result in noise in the search, thereby reducing the overall search performance. To avoid this, external ontologies (BioPortal, UMLS, MeSH, etc.) can be used to determine whether the concepts relate to medical terms with UMLS code before they are finally extended.

For example, the GOT 431 on the established knowledge structure may recognize through an external ontology search that the medical term having the UMLS code C0528721 is not the past tense of the GET. However, in the case of the Activity 433 added in FIG. 4, there is no basis for judging it as a medical term, and thus may be excluded from the final query expansion list.

That is, for the initial query Asparate 420, 'Albumin 410', 'GOT 431', 'Glutamic 432', and 'Mitochondrial () through the query expansion through the knowledge structure proposed in this embodiment. 434) may be added to the initial query set.

In operation 270, the phrase search unit 120 may re-search the medical phrase using the expanded query set. Here, the medical phrase search may be re-executed by using the expanded query set through a separate phrase research unit instead of the phrase search unit 120.

That is, the medical query phrase search can be performed again through the added query word. This allows you to search for additional phrases that were not detected in the initial search results, and allows you to consider the proximity between the query and knowledge structure keywords by adding words that lie on the path, allowing for improved phrase search. Lose.

Accordingly, an integrated knowledge structure can be automatically generated from a plurality of medical-related phrases searched above through an initial query, and keywords related to query expansion can be automatically extracted from the generated knowledge structure.

On the other hand, such a phrase search method can be used for a variety of phrase search as well as medical literature passages.

According to another exemplary embodiment, a phrase search method includes extracting a paragraph-based phrase from a plurality of documents, indexing the extracted phrase, searching for an initial query input and a suitable phrase, and obtaining an initial phrase search result; Extracting the core concepts from the verse search result through natural language processing, extracting the correlation information between the core concepts, generating the knowledge structure using the correlation information between the core concepts, and constructing Exploring the core concepts of the knowledge structure, adding concepts between the core concepts in the knowledge structure and the initial query to the query set, and re-executing the phrase search using the extended query set. It can be done by.

Through the initial query, an integrated knowledge structure can be automatically generated from a plurality of related phrases searched above, and keywords related to query expansion can be automatically extracted from the generated knowledge structure.

According to another embodiment of the present invention, a phrase search method uses a general or specific phrase instead of a medical document phrase, and has a similar structure to a medical document phrase search method using a deep learning based knowledge structure generation method according to the above-described embodiment. Duplicate description will be omitted.

In addition, the phrase search system according to another embodiment is a phrase extraction unit for extracting a paragraph of paragraph units from the literature, the phrase search to index the extracted phrases and to search the input initial query and the appropriate phrases, to obtain an initial phrase search results After retrieving key concepts through natural language processing from a predetermined number of upper phrases in the initial and initial phrase search results, the relevance information extraction unit extracts the relevance information between the key concepts and the knowledge structure using the relevance information between the key concepts. A knowledge structure generation unit for generating, a core concept search unit for searching the core concepts of the constructed knowledge structure, and a query expansion unit for adding a concept between the core concepts and the initial query in the searched knowledge structure to the query set; Using the extended set of queries, the phrase search can be rerun.

According to another embodiment of the present invention, a phrase search system uses a general or specific field instead of a medical document phrase, and uses a deep learning-based knowledge structure generation method and its configuration according to the above-described embodiment. This similar and redundant description will be omitted.

As described above, according to embodiments, unlike a method of extracting a related word from an existing external resource or ontology or a method of mechanically extracting a related word from a document searched at a higher level, a knowledge structure is constructed from a document searched at a higher level and corresponding Words associated with queries in the knowledge structure can be used as extended query words to improve search performance.

The apparatus described above may be implemented as a hardware component, a software component, and / or a combination of hardware components and software components. For example, the devices and components described in the embodiments may include, for example, processors, controllers, arithmetic logic units (ALUs), digital signal processors, microcomputers, field programmable arrays (FPAs), It may be implemented using one or more general purpose or special purpose computers, such as a programmable logic unit (PLU), microprocessor, or any other device capable of executing and responding to instructions. The processing device may execute an operating system (OS) and one or more software applications running on the operating system. The processing device may also access, store, manipulate, process, and generate data in response to the execution of the software. For the convenience of understanding, a processing device may be described as one being used, but a person skilled in the art will appreciate that the processing device includes a plurality of processing elements and / or a plurality of types of processing elements. It can be seen that it may include. For example, the processing device may include a plurality of processors or one processor and one controller. In addition, other processing configurations are possible, such as parallel processors.

The software may include a computer program, code, instructions, or a combination of one or more of the above, and configure the processing device to operate as desired, or process independently or collectively. You can command the device. Software and / or data may be any type of machine, component, physical device, virtual equipment, computer storage medium or device in order to be interpreted by or to provide instructions or data to the processing device. It can be embodied in. The software may be distributed over networked computer systems so that they are stored or executed in a distributed manner. Software and data may be stored on one or more computer readable recording media.

The method according to the embodiment may be embodied in the form of program instructions that can be executed by various computer means and recorded in a computer readable medium. The computer readable medium may include program instructions, data files, data structures, etc. alone or in combination. The program instructions recorded on the media may be those specially designed and constructed for the purposes of the embodiments, or they may be of the kind well-known and available to those having skill in the computer software arts. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tape, optical media such as CD-ROMs, DVDs, and magnetic disks, such as floppy disks. Magneto-optical media, and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include machine code, such as produced by a compiler, as well as high-level language code that can be executed by a computer using an interpreter or the like.

Although the embodiments have been described by the limited embodiments and the drawings as described above, various modifications and variations are possible to those skilled in the art from the above description. For example, the described techniques may be performed in a different order than the described method, and / or components of the described systems, structures, devices, circuits, etc. may be combined or combined in a different manner than the described method, or other components. Or even if replaced or substituted by equivalents, an appropriate result can be achieved.

Therefore, other implementations, other embodiments, and equivalents to the claims are within the scope of the claims that follow.

Claims (15)

Extracting paragraph-based phrases from the medical literature in the phrase extraction unit;
Indexing the phrase extracted by the phrase search unit and searching for an initial query and an appropriate phrase inputted to obtain an initial phrase search result;
Extracting, by a correlation information extraction unit, key concepts from a predetermined number of upper phrases in the initial phrase search result through natural language processing, and then extracting correlation information between the key concepts;
Generating a knowledge structure by using the association information between the core concepts in the knowledge structure generator;
Searching for a core concept of the knowledge structure constructed by a core concept search unit;
Adding a concept between the core concept in the knowledge structure found in the query expansion unit and the initial query to a query set; And
Re-executing a medical phrase search by using the query term set expanded by the phrase search unit;
Including,
Extracting a paragraph unit paragraph from the medical literature,
After converting a file format from a plurality of medical documents, extracting text and images using the PDFbox library, and performing text preprocessing to separate text sentences and paragraphs through the OpenNLP library, extracting paragraphs from the extracted text Define a paragraph as a phrase and extract the phrase of the paragraph unit,
Generating a knowledge structure using the association information between the core concepts,
Constructing a distance matrix between the key concepts extracted from the predetermined number of upper phrases and generating an association matrix using the similarity between each obtained distance as a numerical value for each row; And
Generating the knowledge structure using the association matrix
Including;
Adding a concept between the core concept in the knowledge structure and the initial query to a query set,
A conceptual word placed between the core concept in the knowledge structure and the initial query word is added to the query set, and before the query expansion, the external term is used to determine whether the conceptual word is a medical term having a UMLS code. Not adding to a query set if it is not considered a term
Characterized in that, the method of retrieval of medical literature passages. The method of claim 1,
Acquiring the initial phrase search results,
Acquiring the search result by searching the user's initial query and the appropriate phrases through TF-IDF, BM25, and LM techniques.
Characterized in that, the method of retrieval of medical literature passages. The method of claim 1,
Extracting association information between the core concepts,
Extracting key concepts through natural language processing of a predetermined number of upper phrases from the initial phrase search result; And
Performing deep embedding-based word2vec algorithm using the extracted core concepts to convert each core concept into a vector;
Including, medical literature phrase search method. delete The method of claim 1,
Exploring the core concepts of the knowledge structure,
Using the PageRank algorithm to search for the most critical keywords in the knowledge structure
Characterized in that, the method of retrieval of medical literature passages. delete The method of claim 1,
Automatically generating the knowledge structure integrated from the plurality of medical related phrases searched above through the initial query, and automatically extracting keywords related to the query expansion from the generated knowledge structure
Characterized in that, the method of retrieval of medical literature passages. A phrase extraction unit for extracting a paragraph unit paragraph from a medical document;
A phrase search unit for indexing the extracted phrase and searching for an initial query word and an appropriate phrase inputted to obtain an initial phrase search result;
An association information extracting unit for extracting key concepts from a predetermined number of higher phrases through natural language processing from the initial phrase search result, and then extracting association information between the key concepts;
A knowledge structure generation unit generating a knowledge structure using the association information between the core concepts;
A core concept search unit for searching a core concept of the constructed knowledge structure; And
A query extension unit that adds a concept between the core concept in the searched knowledge structure and the initial query to a query set.
Including,
Re-search the medical phrases using the expanded set of queries,
The phrase extraction unit,
After converting a file format from a plurality of medical documents, extracting text and images using the PDFbox library, and performing text preprocessing to separate text sentences and paragraphs through the OpenNLP library, extracting paragraphs from the extracted text Define a paragraph as a phrase and extract the phrase of the paragraph unit,
The knowledge structure generation unit,
Construct a distance matrix between the core concepts extracted from the predetermined number of upper phrases, and use the similarity between the obtained distances as numerical values for each row to generate an association matrix, and use the association matrix to Create a knowledge structure,
The query expansion unit,
A conceptual word placed between the core concept in the knowledge structure and the initial query word is added to the query set, and before the query expansion, the external term is used to determine whether the conceptual word is a medical term having a UMLS code. Not adding to a query set if it is not considered a term
A medical literature passage search system, characterized in that. The method of claim 8,
The phrase search unit,
Acquiring the search result by searching the user's initial query and the appropriate phrases through TF-IDF, BM25, and LM techniques.
A medical literature passage search system, characterized in that. The method of claim 8,
The association information extraction unit,
Extracting key concepts from natural phrases by using a predetermined number of upper phrases from the initial phrase search results, and then performing deep embedding-based Word2Vec algorithm to perform embedding to convert each key concept into a vector.
A medical literature passage search system, characterized in that. The method of claim 8,
The core concept search unit,
Using the PageRank algorithm to search for the most critical keywords in the knowledge structure
A medical literature passage search system, characterized in that. delete The method of claim 8,
Automatically generating the knowledge structure integrated from the plurality of medical related phrases searched above through the initial query, and automatically extracting keywords related to the query expansion from the generated knowledge structure
A medical literature passage search system, characterized in that. Extracting paragraph-based phrases from a plurality of documents in a phrase extraction unit;
Indexing the phrase extracted by the phrase search unit and searching for an initial query and an appropriate phrase inputted to obtain an initial phrase search result;
Extracting, by a correlation information extraction unit, key concepts from a predetermined number of upper phrases in the initial phrase search result through natural language processing, and then extracting correlation information between the key concepts;
Generating a knowledge structure by using the association information between the core concepts in the knowledge structure generator;
Searching for a core concept of the knowledge structure constructed by a core concept search unit;
Adding a concept between the core concept in the knowledge structure found in the query expansion unit and the initial query to a query set; And
Re-executing the phrase search using the query term set expanded by the phrase search unit;
Including,
Extracting a paragraph unit paragraph from the plurality of documents,
After converting a file format from the plurality of documents in a specific field, text and image extraction are performed using the PDFbox library, and text preprocessing is performed to separate text sentences and paragraphs using the OpenNLP library. Extract the paragraph of the paragraph unit by extracting and defining the paragraph as one phrase,
Generating a knowledge structure using the association information between the core concepts,
Constructing a distance matrix between the key concepts extracted from the predetermined number of upper phrases and generating an association matrix using the similarity between each obtained distance as a numerical value for each row; And
Generating the knowledge structure using the association matrix
Including;
Adding a concept between the core concept in the knowledge structure and the initial query to a query set,
A conceptual word placed between the core concept in the knowledge structure and the initial query word is added to a query set, and the term of a specific field is determined by determining whether the conceptual word is a term of a specific field by using an external ontology before expanding the query. Not being added to a query set if it is not determined
Characterized in that, the phrase search method. The method of claim 14,
Automatically generating the knowledge structure integrated from a plurality of related phrases searched higher through the initial query, and automatically extracting keywords related to query expansion from the generated knowledge structure
Characterized in that, the phrase search method.

Images