KR20130057715A - Method for providing deep domain knowledge based on massive science information and apparatus thereof - Google Patents

Abstract

PURPOSE: A deep knowledge providing method based on a scientific knowledge memory and a device thereof are provided to analyze literature of a specific science field and copy a complex process which learns knowledge by using natural language processing and mining technology, thereby automatically extracting and accumulating specialized knowledge. CONSTITUTION: A knowledge memory(304) stores relational knowledge, structural knowledge, and procedural knowledge for a document. A deep knowledge providing unit(306) inputs a query language. The deep knowledge providing unit searches and provides a triple which includes the query language and documents related to the triple. The deep knowledge providing unit uses a GCL(Generalized Concordance Lists) query which searches a specific word or a relation between word sets or between words. [Reference numerals] (302) Multidimensional knowledge generating technology; (304a) Relational knowledge memory; (304b) Structural knowledge memory; (304c) Procedural knowledge memory; (306) Deep knowledge providing technology; (AA) Large scholarly information; (BB) Deep knowledge delivery by field

Description

Method for providing deep knowledge based on academic science knowledge and apparatus suitable for this

The present invention relates to text analysis and knowledge processing technology based on large-scale digital academic resources, and in particular, to analyze multi-dimensional knowledge consisting of relational knowledge, structural knowledge, and procedural knowledge by analyzing target documents such as academic papers, patents, reports, etc. The present invention relates to a method for providing in-depth knowledge based on academic scientific knowledge memory and a device suitable for storing and utilizing multi-dimensional knowledge of a target document and providing in-depth knowledge corresponding to a query word.

The biggest concern for science and technology researchers and policy makers is to efficiently analyze technical data such as papers, patents and reports that are exploding and reflect them in their own research and policies or to set new research directions.

In this regard, Sophia Ananiadou, director of the UK's National Center for Text Mining (NaCTeM), commented in a 2008 Biomedical Computation Review that "scientists are now drowning in text." We emphasized the importance of large-scale resource-based text analysis and knowledge processing techniques to overcome this problem.

In-depth analysis of S & T information, which can be carried out to solve the above concerns of science and technology workers, is a large part of the entire R & D process, and most of them are performed by highly manual work.

As a result, inadequate identification of deep domain knowledge embedded in a vast amount of expertise produced in a wide range of science and technology sectors has led to difficulties in the development and implementation of R & D. There are many cases where errors or irrationalities such as inefficiency in establishment, (2) patent disputes between companies and countries, and (3) duplication and backwardness of research are caused.

Sectoral science and technology researchers (enterprise / public research institutes, graduate students, professors, etc.) want to quickly acquire and reflect the latest high-quality in-depth knowledge produced in their research fields. However, by deciding on their own hypothesis or originality and justification for ideas before or during the research process, they spend a great deal of time hindering the effectiveness of their R & D activities. It is a situation.

In addition, R & D decision makers and policy makers want to easily obtain reliable technical trends in the subject research area. The key issue is to secure intuitive and practical trend knowledge that can be easily identified using only their own knowledge without any additional work.

Until now, problems of obtaining in-depth expertise in science and technology through manual labor, which have been mainly performed by science and technology researchers, R & D decision makers, and policy makers, are: (1) excessive time and resource utilization, (2) securing comprehensiveness, ( 3) Difficulties in reuse, sharing, and diffusion of personally acquired in-depth expertise.

There are many retrieval systems based on information extracted from deep knowledge from documents. Three of these systems are briefly introduced.

The first search system was PubMed of the US National Library of Medicine, which attempted to differentiate itself from the general search system, including advanced search functions with various options and the presentation of documents related to the result documents.

The second search system is given by MEDIE of Tsujii Laboratories of Tokyo University, Japan, allowing advanced search using up to three query terms of verbs and objects. Unlike a general search system that indexes and searches searched documents, this system extracts in-depth knowledge acquired by applying advanced natural language processing techniques such as sentence separation, parsing, and object name recognition to all documents. This information is provided as content.

The last retrieval system is LEGENDA of the National Institute of Advanced Industrial of Science and Technology (AIST) in Japan, which grasps the air information between objects in advance, and when a single object is retrieved, the names of the objects with high air frequency Are listed in order of frequency as

These conventional retrieval systems use advanced natural language processing techniques to extract in-depth knowledge in a document. However, these results can only be analyzed in various ways because they present only fragmentary results without significant deviation from the information retrieval method for general documents. And ultimately cannot select research topics or analyze research trends.

The present invention has been devised to solve at least some of the above problems, and an object thereof is to provide a method for providing in-depth knowledge based on academic science memory for efficient sharing / diffusion of academic science and technology knowledge.

The present invention has been devised to solve at least some of the above problems, and an object thereof is to provide an apparatus for providing in-depth knowledge based on academic scientific knowledge for efficient sharing / diffusion of academic scientific and technological knowledge.

Method for providing in-depth knowledge based on academic scientific knowledge memory according to the present invention for achieving the above object

The multidimensional knowledge consists of relational knowledge representing the relationship between the objects contained in the target document, structural knowledge representing the pragmatic role between the sentences contained in the target document, and procedural knowledge including the purpose, behavior, and methods of the target document. In the knowledge providing method based on the knowledge memory stored in the form of (subject / event verb / object),

Searching for and presenting a triple including a query word input for searching from the knowledge memory; And

And searching for and presenting documents related to the selected triple among the presented triples.

Here, the process of searching for and presenting the triple

Search for triples with event verbs with air relations based on the query terms,

Presenting event verbs with air relationships; And

If one of the presented event verbs is selected, it may include a process of performing a search again with a search triple having a “subject-selected event verb”.

Here, the method may further include a step of further presenting a similar word for a subject or object among the entities included in the triples, and re-searching for the search triple having the selected similar word when the similar word is selected.

The method may further include searching for and presenting documents related to the search triple reconstructed from the knowledge memory by an instant search whenever the search triple is reconstructed.

On the other hand, the process of searching for and presenting the triple

Searching for a triple having a query word input from the knowledge memory as a subject;

Extracting an object (object object) corresponding to an object from the searched triples; And

Providing a network representing a triple relationship by a line connecting the subject and the object.

The method may further include expanding the network by searching for triples having the object as the subject from the knowledge memory.

On the other hand, the process of searching for and presenting the triple

Searching for triples having a query word input from the knowledge memory as a subject; And

The searched triples may be classified by object, and a list may be included according to the frequency of appearance of each object.

On the other hand, the process of searching for and presenting the triple

Searching for a triple including a query word input from the knowledge memory;

Extracting an object from the retrieved triples;

Presenting a list of extracted objects;

Reconstructing a query including an object selected from an object list.

In-depth knowledge providing apparatus based on academic scientific knowledge memory to achieve another object of the present invention

A knowledge memory for storing relational knowledge, structural knowledge, and procedural knowledge about the document; And

An in-depth knowledge provider for inputting a query and searching and presenting triples and documents related to the triples including the query word input from the knowledge memory;

Characterized in that it comprises a.

In this case, the in-depth knowledge provider may use Generalized Concordance Lists (GCL) Query to search for a specific term or to search for a term group or a relationship between terms.

In addition, the in-depth knowledge provider is characterized by providing a slide navigation to dynamically look at the document information for the triple, while moving the continuously connected triple-based knowledge information on the same screen with a mouse click.

In addition, the in-depth knowledge provider is a dynamic triple to express the statistical results with a chart in order of frequency (up to the top five ranks) the subject / event verb / target information associated with the keyword when the user enters the desired keyword and executes a search Characterized in that the analysis information browsing further.

In addition, the in-depth knowledge provider may further provide a dynamic table-based search for outputting the object list and the event list associated with each other based on the two input keywords.

In addition, the in-depth knowledge provider may further provide in-depth knowledge navigation for navigating detailed element knowledge based on a knowledge triple displayed in a radial form.

The method and apparatus for providing in-depth knowledge according to the present invention is an advanced natural language processing and mining process for the complex process of analyzing and acquiring / digesting knowledge in a specific scientific field in order to effectively construct academic scientific knowledge memory. Partial simulation with technology has the effect of enabling automatic extraction and accumulation of academic expertise.

The present invention enables the user to innovatively utilize / share in-depth scientific knowledge by field by constructing an academic scientific knowledge memory in which planar bibliographic information is structured in multiple dimensions and element knowledge embedded in individual unit information can be identified / managed. Has an effect.

1 schematically illustrates the concept of multidimensional knowledge in the present invention.
2 schematically illustrates the relationship between multidimensional knowledge in the present invention.
3 illustrates a configuration of an apparatus for providing in-depth knowledge according to the present invention.
Figure 4 shows the configuration of a unified language analysis module factory for surface field knowledge identification and extraction in the present invention.
5 diagrammatically illustrates a modeling process of procedural knowledge in accordance with the present invention.
6 shows a procedural knowledge extraction method according to the present invention.
Fig. 7 shows a search initial screen in the knowledge providing apparatus according to the present invention.
8 shows an example of a slide navigation service screen in the present invention.
9 to 10 show an example of the in-depth knowledge navigation service screen in the present invention.
11 shows an example of a dynamic triple analysis service screen in the present invention.
12 shows an example of a dynamic table search service screen in the present invention.

Hereinafter, the configuration and operation of the present invention will be described in detail with reference to the accompanying drawings.

An entity in information retrieval is defined as an object or concept that can generate information or classify it into a specific group, such as a person, organization, or location.

Examples of individuals include "Euljimunde", "Yi Soon Shin", "Bill Gates", etc., and individuals belonging to the institution include "Seoul National University" and "Ministry of Education, Science and Technology". The individuals belonging to are Manchurian Plate, Seoul, Daejeon, and Guro-gu.

Ontology refers to the fields and techniques that deal with human knowledge.Ontology, in particular, as a computer-based expression of knowledge, expresses the explicit conventions of conceptualization, that is, expressly and logically so that the knowledge of a certain field can be processed by a calculator. It is to enable knowledge sharing and reuse. In ontology, classes provide the ability to group resources with the same properties and to logically express common properties. The nature of a class can be expressed by specifying the condition of the property of the class. On the other hand, an instance refers to an entity belonging to the concept.

Ontology construction is accomplished by analyzing patterns and components from sampled documents in each field and modeling the ontology based on the analysis results.

Multifaceted knowledge defined in the present invention is divided into three levels according to the depth of knowledge that can be extracted from academic information, (1) Relational Knowledge, (2) Structural Knowledge (3) Knowledge information in a multi-layered structure that includes procedural knowledge.

First, relational knowledge is knowledge that identifies the relationships between entities in the literature. More specifically, relational knowledge identifies the relationships among key scientific and technological entities, such as various terminology and entity names, which serve as a focal point of content in the literature. Knowledge is relatively easy to identify and extract.

Second, structural knowledge is the knowledge that distinguishes discourse roles for various syntactic forms of sentences in a document. More specifically, the relationship between the properties of sentences and sentences within a specific document. Knowledge is structured through analysis.

For example, in the abstract of a scholarly article, a sentence (a goal) expressing a problem or a field to be solved by the article, a sentence briefly describing a research result (result), a sentence (method) explaining a research method, and the like are displayed. By separating the sentence-based pragmatic roles, it is very easy to extract various in-depth knowledge.

Third, procedural knowledge is knowledge derived from structured research and research activities (R & D activities) or the purpose, methods, and verification procedures described by academic papers or other specialized literature in a specific science and technology field. The purpose is to identify R & D activities through in-depth analysis of the knowledge base of each field, and it is structured knowledge about the purpose, method, and verification procedure of academic research.

1 schematically illustrates the concept of multidimensional knowledge in the present invention.

As can be seen with reference to FIG. 1, the relational knowledge, which is relatively easy to recognize and extract, is being actively promoted around the world, is called Surface Domain Knowledge (SDK). Structural and procedural knowledge that is difficult and requires a variety of infrastructure processes is classified as deep domain knowledge (DDK).

This stepwise division of knowledge according to the level of in-depth and processing difficulty enables a systematic approach in conducting research for in-depth knowledge based on academic scientific knowledge memory, and has the advantage of clearly setting targets for each step of research.

2 schematically illustrates the relationship between multidimensional knowledge in the present invention. Figure 2 shows a specific example of the knowledge of degenerative myelopathy that can be obtained by the user when using the individual knowledge in the biomedical field.

For example, in relational knowledge, information about the cause and effect of degenerative myelopathy and gene or protein information related to the disease can be easily obtained.

Structural knowledge provides information on methodologies for identifying endogenous / exogenous factors of degenerative myelopathy. For example, the approach applied to elucidate spondylopathy and endogenous / exogenous factors may be known.

In addition, procedural knowledge provides an opportunity to gain detailed knowledge about the treatment of myelopathy and experimental procedures. For example, research data on when patients have side effects on tramadlo, or information on experimental procedures for spontaneous spondylopathy with algae can be obtained.

As shown in FIG. 2, it can be seen that as the depth of knowledge deepens, the flexibility of utilization on the user side increases, and more opportunities for acquiring dynamic in-depth knowledge can be given.

3 shows a configuration of a knowledge providing apparatus according to the present invention. The knowledge providing apparatus 300 according to the present invention shown in FIG. 3 has a function of inputting large-scale academic information on one side, generating multidimensional knowledge, and storing the multidimensional knowledge in a knowledge memory. It has a function of retrieving the multidimensional knowledge described through 2 from the knowledge memory and providing it.

In order to generate multidimensional knowledge from the inputted large-scale academic information, the knowledge providing apparatus 30 performs natural language processing on the input target document, and uses relational knowledge, structural knowledge, and structural knowledge from the document using various qualities obtained by natural language processing. Generate procedural knowledge and store relational, structural and procedural knowledge about the document in knowledge memory.

In addition, in order to provide a search service for multidimensional knowledge, the knowledge providing apparatus 30 proposes a query automatic completion, various search and presentation methods.

Referring to FIG. 3, the knowledge providing apparatus 300 according to the present invention includes a multidimensional knowledge generating unit 302, a knowledge memory 304, and an in-depth knowledge providing unit 306.

The knowledge memory 304, also known as a virtual science brain, may be composed of a relational knowledge memory 304a, a structural knowledge memory 304b, and a procedural knowledge memory 304c. to provide.

The multidimensional knowledge generator 302 extracts multidimensional knowledge consisting of relational knowledge, structural knowledge, and procedural knowledge that can be extracted from academic information. The relational knowledge, structural knowledge, and procedural knowledge extracted from the multidimensional knowledge generator 302 are stored in the relational knowledge memory 304a, the structural knowledge memory 304b, and the procedural knowledge memory 304c, respectively.

The in-depth knowledge provider 306 extracts in-depth knowledge by referring to the knowledge memory 304 with respect to the query word. A query autocompletion function may be provided to assist with query entry. Various interfaces may be provided including a slide type interface in order to provide extracted deep knowledge.

The apparatus 300 for providing knowledge according to the present invention shown in FIG. 3 is a large-scale academic scientific knowledge memory in which a surface field knowledge (SDK) and a deep field knowledge (DDK) are fused together, and a deep knowledge providing technology that can utilize the same. It consists of a combined structure.

The apparatus 300 for providing knowledge according to the present invention shown in FIG. 3 is an advanced natural language processing method for a complicated process of analyzing a document of a specific scientific field and acquiring / digesting knowledge in order to effectively construct an academic scientific knowledge memory. Simulation with mining technology enabled the automatic extraction and accumulation of digital expertise.

In this way, by constructing such academic scientific knowledge memory in which planar bibliographic information is structured and element knowledge embedded in individual unit information can be identified / managed, users can innovatively utilize / share in-depth scientific knowledge of each field.

In the present invention, by giving a realistic definition of scientific and technological knowledge and classifying according to its depth and extraction difficulty step by step to secure the feasibility of the research results, derive a high level of international research results based on the differentiation from existing related research Make sure you have a chance to do it.

The multidimensional knowledge generation unit 302 includes a term dictionary, ontology, term dictionary and ontology based term recognition engine, machine learning based term learning engine, ENJU syntax analysis system, support vector machines (SVMs) and maximum entropy based Multidimensional knowledge is extracted from academic information using in-depth language analysis system. In the extraction of multidimensional knowledge, sharing and co-utilization of the underlying technologies (engines), databases, and language resources are very important.

In the multi-dimensional knowledge generation unit 302, a kernel-based model was implemented in detail to enable deep language analysis based on support vector machines (SVMs) and maximum entropy.

In the embodiment of the present invention, in order to introduce and utilize the MEDIE (Intelligent Search Engine for MEDLINE) developed by the University of Tokyo, Japan, the ontology-based terminology recognition engine and the Enju syntax analysis system in the field of biology were analyzed in detail.

The terminology recognition engine developed to construct the MEDIE database was developed based on a biological ontology and a set of terminology dictionaries, and identifies protein names, gene names and disease names that appear in the literature and links them with descriptors in MeSH or UMLS. Let's do it.

The Enju parsing system, which is necessary for estimating the relationship between terminology in the literature, is a high-speed linguistic analysis system based on HPSG (Head-driven Phrase Structure Grammar), which identifies various grammatical dependencies in a sentence and uses the predicate-argument structure. By expressing it as (Predicate Argument Structure), it is possible to extract various qualities that are important for relation extraction.

The basic query structure for accessing the MEDIE database uses a structure called Generalized Concordance Lists (GGL) Query. GCL has a structure similar to SQL, and includes a variety of similar comparison operators that make it easy to search for specific terms or even term groups or relationships between terms.

The multidimensional knowledge generator 302 performs a surface knowledge extraction function and a deep knowledge extraction function.

For the extraction of Surface Domain Knowledge (SDK),

(A) Language resource utilization techniques for surface knowledge extraction (b) Language processing / machine learning application techniques for SDK identification are used.

First, the language resource utilization technology for extracting surface knowledge is used to collect and analyze various corpus and dictionaries (MeSH, UMLS, OntoNotes 2.0, MUC, etc.) and structure them.

The Application of Language Processing / Machine Learning for Surface Knowledge Identification utilizes a universal interface and an integrated language analysis module factory that can utilize various language processing platforms that are actively used around the world as a single API for deep language analysis. .

Charniak Parser, Stanford Parser and Enju Parser are now included in this factory. Finally, the kernel-based model, which is known to be the highest performing model among relational knowledge extraction, is implemented in detail to enable deep language analysis based on support vector machines (SVMs) and maximum entropy.

Figure 4 shows the configuration of a unified language analysis module factory for surface field knowledge identification and extraction in the present invention.

Referring to FIG. 4, the integrated language analysis module factory uses a Java Virtual Machine (JNI) interface based on a Java Virtual Machine (JVM) to be widely used independently of a development environment and a language, and is currently the most widely used worldwide. It is driven by 3 different parsing modules.

Table 1 shows the capabilities of the parsers used in the unified language analysis module factory.

Parser Parse tree  offer Dependency syntax tree  offer Provide predicate-argument structure Charniak Parser (US) O Stanford Parser (US) O O Enju Parser (Sun) O O

Information extraction is a key area in natural language processing and text mining. The final goal of information extraction is to identify important and relevant information present in the text to extract and transform the tabular structured data within the unstructured text data.

Element technologies constituting such information extraction techniques include (1) Named-Entity Recognition, (2) Relationship Extraction, and (3) Co-reference Resolution.

In order to improve the performance of relationship extraction, various supervised learning-based relationship extraction techniques have been introduced. They can be classified into three types: (1) rule-based methods, (2) feature-based methods, and (3) kernel-based methods.

As a relatively recently developed method, a kernel-based method that newly constructs a kernel function specialized for relation extraction and applies it to a support vector machine (SVM) based on this is drawing attention.

The characteristics of kernel-based methods in the field of relation extraction are that the performance of the kernel is best represented by constructing a kernel that best expresses the relationship between two entities in a sentence, and calculates the similarity between two relation-containing statements that includes it most effectively. It is very high.

In the multi-dimensional knowledge extraction unit 302 shown in FIG. 3, a technique for extracting in-depth knowledge is largely based on (1) structured digital abstracts (SDA) based technology, (2) application and modification of large-scale activity ontology, and (3) specialized fields. It can be divided into in-depth expertise extraction and utilization model.

First, in the SDA-based technology part, support vector machines (SVM) and conditional randomization are used to automatically extract and classify key sentences that express the purpose or method of research in order to extract structural knowledge from abstracts of academic literature. Uses the field (Conditional Random Fields, CRF) model.

In the application and application of large-scale activity ontology, the evaluation set for procedural knowledge extraction, along with detailed modeling of the relationship between multiple processes in which procedural knowledge, that is, objects, behaviors, and methods, can be expressed in academic literature, set), procedural knowledge tagging support tools for evaluation sets, and automatic learning of procedural knowledge extraction based on machine learning. In particular, detailed modeling of automatic extraction of procedural knowledge based on academic information consisting of a total of seven steps was performed.

Researches that extract information about various terminology, objects, concepts, etc. using dictionaries, thesauruses, and ontologies have been steadily progressing until recently. Recently, researches have been conducted.

However, beyond this fragmentary knowledge, research on structured and procedural knowledge extraction is extremely difficult to find. However, there are studies that extract procedural information from web documents such as eHow and wikiHow, but this is only about extracting ontology from a set of sentences already structured and sequenced by tfka.

Procedural knowledge is a collection of sequential or structural acts or unit procedural knowledge to achieve a specific purpose. In other words, the structure of the document to be analyzed can be expressed as a set of sentences that make up the document and solution sentences to achieve the purpose. I can express it.

It can be assumed that the unit procedures in the solution are sequential, parallel, or independent among the procedures, and each unit procedure is composed of triples of target, method, and action for performing the procedure. Can be assumed.

Target is the part that is the subject of the experiment in the unit procedure, Method is the specific experimental method applied to the subject, and Act indicates how the experimental method is applied to the subject.

In the embodiment of the present invention, the target (Target) was limited to the disease name, disease name, symptoms, effects, values, etc., Method (Method) was defined as a treatment method, a surgical method, a medication method and the like. In addition, the act (Act) means a predicate portion connecting the object and the method, it can be interpreted as "act" by applying the "method" to the "object". At this time, universal predicates such as 'have' and 'be' are not considered.

This procedural knowledge modeling was carried out based on the analysis of the content and characteristics of the document along with the specialists who are the biggest beneficiaries of the analysis results.

5 is a schematic representation of procedural knowledge.

Referring to FIG. 5, procedural knowledge may be expressed as a relationship between multiple processes, and each process includes "target", "action", and "method". It can be seen.

A single process is characterized by three elements: "Target", "Action", and "Method" and is named TAM after the capital letters of these words.

Accordingly, procedural knowledge may be defined as a triple having a Process (TAM) -Relation-Process (TAM) structure.

There are seven types of relationships that can be specified between the two processes and their definitions are shown in Table 2.

relation Explanation Temporal relation Sequential relationship between two TAMs Parallel relation Two TAMs in Parallel Relationship Causal relation Relationship between two TAMs as cause and effect Comparable Relation Relationship comparing two TAMs Explanatory Relation Relationship where the second TAM describes the first TAM Targetable Relation The second TAM corresponds to the Target item in the first TAM Methodological Relation The second TAM corresponds to the Method item of the first TAM

6 diagrammatically illustrates a procedural knowledge extraction process according to the present invention.

Referring to FIG. 6, for a tagged abstract

(1) candidate sentence extraction including target (T), behavior (A), and method (M), (2) TAM extraction, (3) TAM normalization, (4) TAM interworking, and (5) process (TAM) Procedural knowledge is extracted through relationship extraction.

The extracted procedural knowledge is stored in the knowledge memory 304.

Various natural language processing techniques can be applied to the abstracts of academic papers such as pruning, syntactic analysis, predicate-argument structure analysis, and terminology extraction.

The predicate-dissertation structure is a structure that expresses a significant association between each word in a sentence by using a predicate and an argument.

Terminology extraction is information tagged with word or multi-term terms within a document based on well-known medical / life discipline ontology such as Unified Medical Language System (UMLS), UniProt, and Gene Ontology (GO), which are included in the target document. Given that these terms are terminology in the field, they are essential for the efficiency of in-depth knowledge extraction.

The candidate abstract including the target (T), action (A), and method (M) is extracted by analyzing the tagged abstract. Since the unit procedure consists of triples of Target / Action / Method, extracting candidate sentences means extracting sentences with such triples.

For example, in the literature, "Target" refers to the research task to be addressed in a particular study, such as disease name, symptoms, effects, etc., and "Method" refers to the method used to solve the research task. That is, the treatment method, the surgical method, the medication method and the like. Finally, "Action" acts as a predicate that connects "target" and "method".

Extracting semantic objects corresponding to Target, Action, and Method from the sentence is described below.

For the extraction of Target, Actin, and Method, which are the basic elements of unit procedure extraction, the features of word itself, context, predicate-argument, and terminology can be used.

● the qualities of the word itself; The word itself, the base form of the word, the part-of-speech tag, the part-of-speech classification information (ie, verbs, nouns, symbols, etc.)

Context qualities; Before / after N word and part-of-speech tag information

Predicate-argument structure; Predicates and Argument Corresponding Phrases

Terminology qualities; UMLS / Uniprot / GO Ontology Ontology Tagging Information

Available features are extracted from the tagged abstract, and CRFs model training is performed based on the extracted features. In CRFs, the input is the above feature information set, and the output is each unit procedure element (Target, Action, Method).

As described above, the knowledge providing apparatus 300 according to the present invention may provide a hint about an in-depth knowledge service in a specific field, thereby enabling a user who uses a science and technology knowledge discovery integrated platform to advance a service that he or she thinks of. Can provide a foundation.

The in-depth knowledge provider 310 retrieves in-depth knowledge corresponding to the input query word from the knowledge memory 308 and processes and outputs it according to various presentation methods.

The present invention provides various search methods such as (1) slide navigation, (2) deep knowledge navigation, (3) dynamic triple analysis, and (4) dynamic table based search. All services basically consist of semantic based triple set extracted from MEDIE database, and various types of deep relational knowledge can be provided.

The search method proposed in the present invention enables the user to discover and create new knowledge by presenting various analysis results as well as simple semantic triple search.

Fig. 7 shows a search initial screen in the knowledge providing apparatus according to the present invention.

The knowledge providing apparatus 300 according to the present invention is a document-based search system, and a search is started with one query word like a general search system. When the query is input, the candidate query may be presented by an autocomplete function as shown in FIG. 7.

(1) slide navigation

Slide navigation provides the ability to dynamically view detailed scholarly information on the corresponding triple while moving continuously connected triple-based knowledge information with the click of a mouse on the same screen. Such slide navigation makes it easy to start with a specific object and expand based on the relationship between the objects associated with the object and re-expand the object object.

Another feature is automatic query generation. For the first input query, this candidate candidate query can be presented in the form of slide navigation so that the user can simply expand the query to triple. The results can be reconstructed in the same way as Google's instant search and the search results can be presented.

8 shows an example of a slide navigation search screen in the present invention.

The search results show the ranked documents for the query words (ex, diabete) similar to the results of a general search system. The difference is two, one with six object type information highlighted in each color and access to the text or other services.

A search method using slide navigation according to the present invention is as follows.

First, a triple including a query word input from the knowledge memory is retrieved.

The navigation function is used to reconstruct the triple by selecting a similar word for each of the entities constituting the retrieved triple, that is, the subjects, event verbs, and objects constituting the triple. A navigation box is provided for receiving individual objects for navigation to the objects. The navigation box is provided with a slide bar from which similar words can be selected. The navigation box can be extended to include other entities in addition to those included in the triple.

Retrieve, list, and present documents that contain reconstructed triples.

The interface screen 900 for the slide navigation search shown in FIG. 8 includes a query input window 902, a navigation window 904, and a list window 906. The query input window 902 displays a query input by the user.

Meanwhile, the navigation window 904 displays navigation boxes 904a to 904d for expanding each element of a triple including a query input by a user. The navigation box includes a slide bar that allows movement between similar words.

The navigation boxes 904a to 904c correspond to subjects, event verbs, and objects of the triple 'diabete (main) -associate (event verb) -hypertension (object)' including the query word 'diabete', respectively. In the initial search, a triple having a query input by the user as a subject may be displayed. As another example, when a user inputs two query terms, a triple including the two query terms may be displayed. That is, the query query input by the user is provided by the query automatic generation function.

This automatic query generation function is based on multidimensional knowledge, especially structural knowledge. That is, the multi-dimensional knowledge, especially structural knowledge, of the triple structure including the query input by the user is searched, and the retrieved triple is presented using the query automatic generation function.

Each navigation box 904a-904d has a slide bar that allows navigation of similar words. For example, sliding the slide bar up or down can cause the words 'diabete' and words with similar meanings arranged in alphabetical order to be displayed one after the other. .

The structural knowledge of the search triple corresponding to the combination of words selected by the respective navigation boxes 904a to 904d is searched by the instant search function, and a list of documents corresponding to the searched structural knowledge is displayed in the list window 906. do.

Here, the list of each document includes a title and a sentence representing structural knowledge extracted from the document. The sentence representing the structural knowledge is the structural knowledge obtained by analyzing the document, that is, the sentence representing the unit process, that is, the knowledge of TAM triple structure. In sentences representing unit processes, words corresponding to Target, Action, and Method are displayed in different colors.

If you select a title from the list of documents, the original text is accessed.

The navigation boxes 904a-904d can be expanded. That is, it is expanded to include other words in addition to the subject / predicate / object that constitutes a triple. The navigation box 904d of FIG. 8 illustrates this example. 8 shows an example of searching for a document having a triple of "diadiabete-associate-hypertension" and having the word "treat".

By manipulating the horizontal slider under the navigation window 904, the navigation boxes can be expanded or reference can be made to the navigation boxes before / after.

When the object of the navigation box is selected, the following in-depth network browsing search screen pops up.

(2) Deep Network Browsing

In-depth network browsing provides the ability to search for a single entity in a network format.

Knowledge triples can be expressed radially by default, and by constructing systems and services that can navigate detailed element knowledge, users can find the object by searching (providing both keyword and identifier base) and based on it. Continuous navigation is possible while checking the relationship between objects on the screen.

This in-depth knowledge navigation includes the ability to adjust the degree of output of other entities and relationships associated with a particular entity at a particular location for the efficiency of navigation across a wide range of knowledge triples.

In the present invention, a search method using in-depth knowledge navigation is as follows.

First, a triple having a query word input from the knowledge memory as a subject is searched.

The object corresponding to the object (object object) is extracted from the retrieved triples.

It provides a network that expresses a triple relationship by a line connecting the input query object and the object object. Here, the line with the query word and the object word represents a predicate that associates the subject and the object, and the contents of the predicate can be confirmed by placing the mouse on the line.

The network can be extended to include sub-networks that have object objects as their subject.

9 to 10 show an example of the in-depth knowledge navigation service screen in the present invention.

In-depth knowledge navigation is a service that searches one object in a network format. When one object is selected in the search result screen, in-depth network browsing service is started based on the object. The network is expandable up to three levels, and the zoom view function allows you to check the results closely. 9 shows a first stage search network and FIG. 10 shows a two stage search network.

9, a network for an entity called UMLS: C0011849 (database; number) is shown. The subject UMLS; C0011849 is diabetes. This network may be for the entity selected in the slide navigation window 902 of FIG. 8. For example, the slide navigation box 904a may be selected by a double-click operation of a mouse in a state where 'daabetes' is displayed in the slide navigation window 904a of FIG. 8.

The selected objects 1002 are displayed in the center of the network and the radially associated objects are displayed around them. These associated objects 1004 may be objects included in triples that have the selected object 1002 as the subject. A line showing the connection relationship between the selected objects and the objects associated with them is displayed. This line may be an "associate" representing the relationship between the selected entity and its associated hypertension. That is, the triplet "diabetes associate hypertension" is represented by the selected diabetes, the hypertensions associated with them, and the lines connecting them. When the user places the mouse over the line, the relationship represented by the line, that is, the predicate of the triple, is displayed.

The category 1006 shown in the upper left of FIG. 9 shows that the attributes of entities included in the network are displayed by color. On the other hand, the network control box 1008 shown in the upper right of FIG. 8 allows to adjust the degree of output of other entities and relationships associated with a particular entity at a particular location for efficiency of navigation for a wide range of knowledge triples.

In addition, if you select an entity (node), synonyms are shown in order of frequency, and if you select a relationship (line), you can check the possible relationships between two entities and the sentence that proves them. In addition, access to the text is possible. All objects have ID information from external well-known databases and appear in the results, making it easy to access them.

Referring to FIG. 10, it can be seen that the network of the entities associated with the selected entity 1002 (stage 1 network), as well as other entities 1010 associated with the associated entity 1004 (network 2 levels) are displayed. .

(3) dynamic triple analysis

Dynamic triple analysis is a service that checks mutual entity or relationship information about a query. When one object is entered as a query, the results of analysis of other entities and relationships in the air are presented in order of frequency. When two queries are entered, The two query terms, along with other entity or relationship verbs, are presented in order of frequency.

In the present invention, a search method for dynamic triple analysis is as follows.

First, triples having a query word input from the knowledge memory as a subject are searched.

The retrieved triples are sorted by object and listed according to the frequency of appearance of each object. The frequency and share of occurrences of each individual are listed. The frequency of occurrence and share can be provided in chart form or in table form.

11 shows an example of a dynamic triple analysis service screen in the present invention.

Referring to FIG. 11, when a user inputs a desired query ('diabete') and executes a search, statistics of the subject / event verb / target word information associated with the query in frequency order (up to the top 5 ranks) are displayed along with the statistical results. Express.

The upper side of FIG. 11 shows a pie chart and a table showing the frequency and the occupancy rate of five verbs induce, associate, injibit treat, and affect the most frequently among verbs associated with the query term 'diabete'. .

Meanwhile, the lower side of FIG. 11 shows pie charts and tables showing the frequency and the occupancy rate of five objects insulin, abesity, hypertension, atherosclerosis, and cardiovascular disease, respectively, among the objects associated with the query 'diabete'. .

Here, various chart types may be designated according to the taste of the user. For example, you can get overall statistical information about all the treatments or therapies that can treat a particular disease. In the embodiment of the present invention, examples of a pie type and a bar type are shown. Korean verbs can also be identified for related verbs.

(4) dynamic table search

The dynamic table search is a search method that analyzes the related entity or event verbs of two input queries, and discovers new neighbor relationships by re-searching the semantic triples containing these common factors.

In the present invention, the dynamic table search method is as follows.

First, the triples including the query word input from the knowledge memory are retrieved.

The objects are extracted from the retrieved triples.

A list of extracted entities is presented. Each list contains two or more entities. Possible combinations of entities related to the query entered by this list are presented to the user, and by reconstructing the query by referencing these lists, it is possible to expand and browse to new relationships not found in network browsing or dynamic triple analysis. do.

Reconstruct the query that contains the selected object from the list of objects.

12 shows an example of a dynamic table search service screen in the present invention.

Referring to FIG. 12, a user inputs two keywords to search for and performs a search and outputs an object list and an event list related to each other based on the two keywords.

You can drag and drop the list of retrieved objects and events into a dynamic table box to find the object you want, or to find new object information.

300 ... Knowledge Feeder
302 Multidimensional Knowledge Generator 304 Knowledge Memory
306.In-depth knowledge provider

Claims (15)

The multidimensional knowledge consists of relational knowledge representing the relationship between the objects contained in the target document, structural knowledge representing the pragmatic role between the sentences contained in the target document, and procedural knowledge including the purpose, behavior, and methods of the target document. In the knowledge providing method based on the knowledge memory stored in the form of (subject / event verb / object),
Searching for and presenting a triple including a query word input for searching from the knowledge memory; And
And searching for and presenting documents related to the selected triple among the presented triples.
The process of claim 1, wherein the searching and presenting of the triples is performed.
The knowledge memory-based in-depth knowledge providing method comprising: searching for triples having event verbs having air relations based on the query terms.
The method of claim 2,
Presenting event verbs with air relationships; And
A method of providing in-depth knowledge based on knowledge memory, characterized in that if one of the presented event verbs is selected, the search is repeated with a search triple having a "subject-selected event verb".
The method of claim 3,
Providing further similar words for a subject or object among the entities included in the triples, and if the similar words are selected, further including re-searching for the search triple having the selected similar words. Way.
The method according to claim 3 or 4,
And searching for and presenting documents related to the search triple reconstructed from the knowledge memory by an instant search whenever the search triple is reconstructed.
The process of claim 1, wherein the searching and presenting of the triples is performed.
Searching for a triple having a query word input from the knowledge memory as a subject;
Extracting an object (object object) corresponding to an object from the searched triples; And
Providing a network expressing a triple relationship by a line connecting a subject and an object; a method of providing in-depth knowledge based on knowledge memory.
The method according to claim 6,
And searching for a triple having the object as the subject from the knowledge memory, and expanding the network.
The process of claim 1, wherein the searching and presenting of the triples is performed.
Searching for triples having a query word input from the knowledge memory as a subject; And
And classifying the searched triples by object and listing the searched triples according to the frequency of appearance of each object.
The process of claim 1, wherein the searching and presenting of the triples is performed.
Searching for a triple including a query word input from the knowledge memory;
Extracting an object from the retrieved triples;
Presenting a list of extracted objects;
Reconstructing a query containing a selected object from the list of objects; Deep knowledge-based knowledge providing method comprising a.
A knowledge memory for storing relational knowledge, structural knowledge, and procedural knowledge about the document; And
An in-depth knowledge provider for inputting a query and searching and presenting triples and documents related to the triples including the query word input from the knowledge memory;
Device for providing in-depth knowledge based on academic scientific knowledge memory.
The academic scientific knowledge memory of claim 10, wherein the in-depth knowledge provider uses Generalized Concordance Lists (GCL) queries to search for a specific term or to search for a term group or a relationship between terms. In-depth knowledge provision device.
The method of claim 11, wherein the in-depth knowledge providing unit provides a slide navigation to dynamically look at the document information for the triple, while moving the continuously connected triple-based knowledge information on the same screen with a mouse click easily An in-depth knowledge providing device based on academic scientific knowledge.
The method of claim 12, wherein the in-depth knowledge provider inputs a keyword desired by the user and executes a search to display statistical results along with a chart of subject / event verb / target word information associated with the keyword in order of frequency (up to the top five ranks). An apparatus for providing in-depth knowledge based on academic scientific knowledge, characterized by further providing dynamic triple analysis information browsing.
The method of claim 13, wherein the in-depth knowledge provider further provides a dynamic table-based search for outputting a list of entities and events related to each other based on two input keywords. In-depth knowledge provision device.
The apparatus of claim 14, wherein the in-depth knowledge provider further provides in-depth knowledge navigation for navigating detailed element knowledge based on a knowledge triple displayed in a radial manner.

Images