KR101100830B1 - Entity searching and opinion mining system of hybrid-based using internet and method thereof - Google Patents

Abstract

The present invention relates to an object search using the Internet and a hybrid-based opinion analysis system and method therefor, the method comprising: a first server for collecting web document data existing on the internet, and a web document collected from the first server A data analysis server that receives data and extracts meta information for each object, and analyzes positive / negative statistical information about each object of each object by using the meta information for each object, and for each object analyzed from the data analysis server A second server that indexes positive / negative statistical information about each object of each entity including meta information, and indexes the database to be stored; and receives a user search keyword transmitted from a user terminal connected through the Internet. Meta information or target key pre-stored in the user search keyword in conjunction with A web server for determining whether a card is present and displaying an object list result related to the corresponding meta information or the target keyword on the screen of the corresponding user terminal when the pre-stored meta information or the target keyword exists. By directly finding the entity indicated by the keyword, it is possible to easily search and monitor the statistical information including the entity result list at a glance.

 Object Search, Opinion Analysis, Internet, User Search Keyword, Meta Information Keyword, Target Keyword, Web Server

Description

ENTITY SEARCHING AND OPINION MINING SYSTEM OF HYBRID-BASED USING INTERNET AND METHOD THEREOF}

The present invention relates to an entity search using the Internet and a hybrid based opinion analysis system and method therefor, and more particularly, directly to an entity indicated by a specific search keyword inputted by users using the Internet. Find and display the results of opinions, not documents, and / or statistical information extracted in a hybrid-based manner using both rule-based and machine learning methods. By doing so, Internet users relate to object search using the Internet and hybrid based opinion analysis system and method for using the Internet to search and monitor objects related to a specific search keyword at a glance.

Recently, as the use of the Internet has become more active, many people are expressing their opinions on the Internet through media such as blogs and wikis. In addition, when evaluating the value of certain information, there is an increasing demand to refer to opinion information posted by these other people on the Internet.

For example, there are opinions of users on the Internet ranging from product reviews to movie reviews. The opinions of each of these users can be used when the general user wants to buy the goods or see other users' opinions before watching a movie. Can also be used if you want to know. In particular, general users tend to purchase other items after first reviewing other users' ratings before purchasing certain items.

However, the opinions that exist on the Internet exist only on individual websites, and when there is a desire to use such opinion information, the user has to manually search all these individual websites.

It is difficult for users to browse all of these websites, and if you want to find other users' opinions by general search, you can effectively mix and match web documents with feedback, web documents with positive feedback, and web documents with negative feedback. There is a problem that is difficult to find their opinion.

In order to solve these problems, technology for extracting user's opinions is actively researched in domestic and foreign academia. In the field of information retrieval, various technologies have been developed since early 2000.

However, the existing information retrieval technology simply provides a search based on the information in which the keyword exists, and provides a higher level search based on the positive / negative evaluation of the document or sentence in which each keyword appears. It is not. Recently, attempts have been made to apply user feedback extraction techniques to information retrieval, but they are still only at the level of sharing positive and negative documents.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-described problem, and an object of the present invention is to directly find an entity indicated by a specific search keyword input by users using the Internet, and to search for an entity result list and / or not a document. By displaying opinion statistics extracted by the hybrid-based method using both rule-based and machine learning methods through the corresponding user terminal, Internet users can search for specific search keywords and keywords. The present invention provides an object search using the Internet to search and monitor related objects at a glance, and provides a hybrid-based opinion analysis system and method thereof.

In order to achieve the above object, a first aspect of the present invention includes a first server for collecting web document data existing on the Internet; A data analysis server that receives the web document data collected from the first server, extracts meta information for each object, and analyzes positive / negative statistical information about each object of each object by using the meta data for each object; A second server for indexing positive / negative opinion statistics information for each object of each individual including the meta information for each entity analyzed from the data analysis server to be stored in a database; And receiving a user search keyword transmitted from a user terminal connected through the Internet to determine whether there is meta information or a target keyword pre-stored in the user search keyword in association with the second server, and the meta prestored in the user search keyword. If the information or the target keyword exists, the object search using the Internet including a web server displaying the result of the object list related to the meta information or the target keyword on the screen of the user terminal, and a hybrid based opinion analysis system and method therefor To provide.

Here, it is preferable that the first server collects and stores RSS addresses on the Internet, receives RSS files corresponding to the stored RSS addresses, and collects web document data using link information provided by each RSS file. .

Preferably, the data analysis server may extract meta information for each object by using a regular expression representing a string of a preset form in the web document data.

Preferably, the data analysis server, the first module for classifying by sector using a pre-set machine learning model for the collected web document data; A second module for performing language processing on the collected web document data to extract opinion sentences, and dividing the extracted opinion sentences into positive / negative opinion expressions; A third module for determining which object corresponds to the collected web document data using the object-specific meta information; And a fourth module for determining which object corresponds to a word and part-of-speech information around the opinion sentence extracted through the second module.

Preferably, the second module comprises: a language processor for separating the collected web document data into sentence units and performing linguistic processing on the separated sentences to extract linguistic features; An opinion / non-computation division unit for classifying opinion / non-comment sentences using the linguistic qualities of the extracted sentences; And an opinion expression division unit for dividing the linguistic qualities of the divided opinion sentences into positive / negative opinion expressions.

Preferably, when the predefined rule exists for the extracted opinion sentence, the second module applies a predetermined rule-based model to classify the opinion into positive / negative opinion expressions, and the extracted opinion. If there is no predefined rule for a sentence, the machine learning model can be divided into positive and negative opinions by applying a preset machine learning model.

Preferably, the second module, after applying a rule-based model and a machine learning model previously set to the extracted opinion sentence at the same time to determine a positive / negative opinion expression, the rule-based model and machine learning Different reliability scores may be given according to whether or not the results of applying the model are matched, and may be divided into positive / negative opinion expressions based on the reliability scores.

Preferably, the second module determines whether the extracted exception sentence is a predefined exception handling rule candidate sentence, and the predefined exception if the extracted opinion sentence is a predefined exception handling rule candidate sentence. Depending on the existence of processing rules, predefined exception handling rule-based models or machine learning models can be applied to classify as positive / negative / neutral.

Preferably, the second module simultaneously performs a rule-based model and a machine learning model preset for the extracted opinion sentence when the extracted opinion sentence is not a predefined exception handling rule candidate sentence. After determining the positive / negative opinion expression by applying, different reliability scores can be given according to whether the rule-based model and the machine learning model are matched, and can be divided into positive / negative opinion expressions based on the reliability score. .

Preferably, the second module is affirmative / negative / by applying a predefined exception handling rule-based model when the extracted opinion sentence is a predefined exception handling rule candidate sentence and a predefined exception handling rule exists. In the case of a neutral opinion expression, the extracted opinion sentence may be divided into positive / negative / neutral opinion expression by applying a preset machine learning model when a predefined exception handling rule does not exist.

Preferably, the positive / negative opinion statistics information for each object of each object stored in the second server may include an object ID, a target, the number of positive / negative opinion expressions for each object of each object, the total number of opinion expressions, or each object. At least one of the opinion expression content using the information can be made.

Preferably, when the meta information or target keyword stored in advance in the user search keyword does not exist, the web server may search for the corresponding user search keyword in the second server to which the object is mapped and display the related object result list. have.

Preferably, the web server may display the object result list in the order of positive / negative opinion according to the opinion analysis result for the user search keyword.

Preferably, the web server analyzes a user search keyword transmitted from the user terminal to determine whether there is meta information or a target keyword stored in the second server and classifies a keyword search method according to the determination result. module; And a keyword search module for searching the entities related to the corresponding meta information or the target keyword and displaying the object list result on the screen of the corresponding user terminal according to the keyword search method classified from the keyword analysis module. Can be.

Preferably, the web server searches for the entities related to the meta information keyword if the meta information and the target keyword pre-stored in the user search keyword exist, and then, affirmative / negative or all of the target keywords for the searched entities. The result list of objects can be rearranged and displayed in any order of the opinions.

Preferably, the affirmative / negative opinion statistics information for each object of each object stored in the second server is stored so that each object is mapped on a document / paragraph basis, and the web server may store meta in advance in the user search keyword. If the information and the target keyword do not exist, the corresponding user search keyword is morphologically analyzed, and then the morphologically analyzed user search keyword is compared and analyzed by comparing the stemmed user search keyword with documents / paragraphs stored in the second server. Search in a paragraph result list and display the entity result list mapped to the searched document / paragraph result list on the corresponding user terminal.

Preferably, the affirmative / negative opinion statistics information for each object of each object stored in the second server is stored so that each object is mapped on a document / paragraph basis, and the web server may store meta in advance in the user search keyword. If the information and the target keyword do not exist and the pre-stored meta information keyword exists, the corresponding user search keyword is stemmed and then the stemmed user search keyword except the meta information keyword and the document / paragraph stored in the second server are stored. Compare and analyze the search results of the stemmed user search keyword excluding the corresponding meta information keyword in the document / paragraph result list, search the entity result list mapped to the searched document / paragraph result list, and then search the retrieved entity result list. Lease object results related to the corresponding metainformation keyword By the filter can give a display in the user terminal.

Preferably, the affirmative / negative opinion statistics information for each object of each object stored in the second server is stored so that each object is mapped on a document / paragraph basis, and the web server may store meta in advance in the user search keyword. If the information and the target keyword do not exist and the pre-stored target keyword exists, the corresponding user search keyword is stemmed, and then the stemmed user search keyword excluding the target keyword is compared with the documents / paragraphs stored in the second server. After analyzing, search for the stemmed user search keyword except the target keyword in the document / paragraph result list, search the entity result list mapped to the searched document / paragraph result list, and then apply the searched entity result list to the corresponding search result. Collect objects in the order of positive or total opinions of the target keywords. By reordering the list that will be displayed on the user terminal.

Preferably, the affirmative / negative opinion statistics information for each object of each object stored in the second server is stored so that each object is mapped on a document / paragraph basis, and the web server may store meta in advance in the user search keyword. If the information and the target keyword do not exist and the pre-stored meta information and the target keyword exist, the corresponding user search keyword is stemmed, and then the stemmed user search keyword except the meta information and the target keyword is stored in the second server. Compare and analyze the stored documents / paragraphs, search for the stemmed user search keyword except the meta information and the target keyword in the document / paragraph result list, and search the object result list mapped to the searched document / paragraph result list. , The corresponding meta information keyword and the The related entity result list may be filtered, and the filtered entity result list may be rearranged and displayed on the corresponding user terminal in order of positive or total opinions of the corresponding target keywords.

Preferably, the web server may display opinion statistics information of each entity together with the entity list result on the corresponding user terminal.

A second aspect of the invention comprises the steps of: (a) collecting web document data residing on the internet; (b) receiving the collected web document data, extracting meta information for each object, and analyzing positive / negative statistical information about each object of each individual using the individual meta information; (c) indexing positive / negative statistical information about each object of each individual including the analyzed individual meta-information to be stored in a database; And (d) determining whether there is meta information or a target keyword stored in advance in the user search keyword transmitted from the user terminal connected through the Internet, and if the meta information or the target keyword is stored in the user search keyword in advance. The present invention provides a hybrid object-based opinion analysis system and method for retrieving an object using the Internet including the step of displaying an object list result related to corresponding meta information or a target keyword on a screen.

Here, the step (b) may include: (b-1) classifying by sector by using a preset machine learning model for each of the collected web document data; (b-2) extracting an opinion sentence by performing language processing on the collected web document data, and dividing the extracted opinion sentence into positive / negative opinion expressions; (b-3) determining which entity corresponds to the collected web document data using the entity-specific meta information; And (b-4) determining which object corresponds to the word and part-of-speech information around the opinion sentence extracted in the step (b-2).

Preferably, the step (b-2) may include: separating linguistic units from the collected web document data, and performing linguistic processing on the separated sentences to extract linguistic features; Classifying opinion / non-computation sentences using linguistic qualities of the extracted sentences; And dividing the linguistic qualities of the divided opinion sentences into positive / negative opinion expressions.

Preferably, in step (b-2), when a predefined rule exists for the extracted opinion sentence, a predetermined rule-based model is applied and divided into positive / negative opinion expressions. When there is no predefined rule for the extracted opinion sentence, a predetermined machine learning model may be applied and classified into positive / negative opinion expressions.

Preferably, in step (b-2), the rule-based model is determined by simultaneously applying a rule-based model and a machine learning model to the extracted opinion sentence at the same time. And different reliability scores according to whether or not the results of applying the machine learning model are matched, and may be divided into positive / negative opinion expressions based on the reliability scores.

Preferably, in the step (b-2), it is determined whether the extracted exception sentence is a predefined exception handling rule candidate sentence for the extracted opinion sentence, and in advance when the extracted opinion sentence is a predefined exception handling rule candidate sentence. Depending on the existence of the defined exception handling rules, a predefined exception handling rule-based model or machine learning model can be applied to classify as positive / negative / neutral.

Preferably, when the extracted opinion sentence is not a predefined exception handling rule candidate sentence, a positive / negative opinion is applied by simultaneously applying a rule-based model and a machine learning model preset to the extracted opinion sentence. After determining the expression, different reliability scores may be given according to whether the rule-based model and the machine learning model are matched, and classified into positive / negative expressions based on the reliability scores.

Preferably, the extracted comment sentence is a predefined exception handling rule candidate sentence, and when a predefined exception handling rule exists, a predefined exception handling rule-based model is applied and divided into positive / negative / neutral opinion expressions. Subsequently, when the extracted opinion sentence does not have a predefined exception handling rule, a predetermined machine learning model may be applied and classified into positive / negative / neutral opinion expressions.

Preferably, in the step (d), if there is no meta information or target keyword pre-stored in the user search keyword, the search results are searched with the corresponding user search keyword in the database (DB) to which the object is mapped to search for the related entity result list. Can display

Preferably, in step (d), the object result list may be displayed in the order of having a lot of positive / negative opinions according to the opinion analysis result for the user search keyword.

Preferably, in step (d), if the meta information and the target keyword pre-stored in the user search keyword exist, the entities related to the meta information keyword are searched, and then the positive / negative of the corresponding target keyword is searched for the searched entities. Alternatively, the object result list may be rearranged and displayed in any order of the entire opinion order.

Preferably, in step (c), affirmative / negative statistical information about each object of each object is stored so that each object is mapped in document / paragraph units, and in step (d), the user search keyword is previously stored. If the stored meta information and the target keyword do not exist, the corresponding user search keyword is morphologically analyzed, and then the morphologically analyzed user search keyword is compared with the stored document / paragraph to compare the stemmed user search keyword with the document / paragraph. The result list may be searched and the entity result list mapped to the searched document / paragraph result list may be displayed on the corresponding user terminal.

Preferably, in step (c), affirmative / negative statistical information about each object of each object is stored so that each object is mapped on a document / paragraph basis, and in step (d), in advance to the user search keyword If the stored meta information and the target keyword do not exist and the pre-stored meta information keyword exists, the corresponding user search keyword is morphologically analyzed, and then the stored user search keyword except the meta information keyword is compared with the stored document / paragraphs. Analyze the search results of the stemmed user search keyword except the meta information keyword in the document / paragraph result list, search the entity result list mapped to the searched document / paragraph result list, and then apply the corresponding search result in the searched entity result list. Filter the list of object results related to meta-information keywords to Can be displayed on the user terminal.

Preferably, in step (c), affirmative / negative statistical information about each object of each object is stored so that each object is mapped on a document / paragraph basis, and in step (d), in advance to the user search keyword If the stored meta information and the target keyword do not exist, and the pre-stored target keyword exists, the corresponding user search keyword is morphologically analyzed, and then the stemmed analysis of the user search keyword except the target keyword is compared with the stored documents / paragraphs. Search for the stemmed user search keyword excluding the target keyword in the document / paragraph result list, search the entity result list mapped to the searched document / paragraph result list, and search the searched entity result list for the target keyword Reorder the list of object results in order of positive or total feedback It can be displayed on the user terminal.

Preferably, in step (c), affirmative / negative statistical information about each object of each object is stored so that each object is mapped on a document / paragraph basis, and in step (d), in advance to the user search keyword If the stored meta information and the target keyword do not exist and the pre-stored meta information and the target keyword exist, the corresponding user search keyword is stemmed and then the stemmed user search keyword and the stored document except the meta information and the target keyword are Comparing / parsing the paragraphs, searching the document / paragraph result list for the stemmed user search keyword excluding the meta information and the target keyword, and searching the entity result list mapped to the searched document / paragraph result list. The object result lease associated with the corresponding metainformation keyword in the retrieved object result list. Filtering the list and rearranging the filtered result list of the object in the order of the number of positive or total opinions of the target keyword and displaying the filtered result list on the corresponding user terminal.

Preferably, in step (d), opinion statistics information of each entity may be displayed on the corresponding user terminal together with the entity list result.

Preferably, in the step (d), when displaying a search result for the corresponding user keyword on the screen of the corresponding user terminal, each object information and the target keyword statistical information automatically extracted for the object are positive, total opinion or negative. One of the opinions is arranged in order of many, and the positive / negative values of the target keywords and the symbols indicating them are displayed, and the positive / negative opinions for the found objects are displayed left and right, or the positive / negative tabs are selected. Can be.

According to the object search using the Internet of the present invention as described above, and a hybrid-based opinion analysis system and method therefor, the document is searched directly by searching for an object indicated by a specific search keyword inputted by users using the Internet. Internet users can search and search for objects related to a specific search keyword at a glance by displaying, on the user's device, a list of individual results rather than a list of individual results and / or a hybrid-based method that uses both rule-based and machine learning methods. There is an advantage to monitoring.

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

First of all, the term 'sub-theme' mentioned throughout the present invention is referred to as an essential property of an entity, that is, a property, and mainly expresses many opinions. It is a comprehensive meaning that can be applied across all fields (eg, cinema, politics, economy, games, sports, etc.).

For example, the sub-theme in the field of film may be composed of emotion, fun, actor, acting, story, inversion, graphics, music, scene, etc. It may be made of price, design, battery, A / S, and the like (Sub-theme) in the restaurant field may be made of taste, price, atmosphere and the like.

The term 'Entity' is a higher concept of the sub-theme, in which related objects are gathered to represent one information unit.

FIG. 1 is a block diagram illustrating an entity search using the Internet and a hybrid-based opinion analysis system for the same according to an embodiment of the present invention, and FIG. 2 is a data analysis server applied to an embodiment of the present invention. It is a block diagram for explaining in detail.

1 and 2, the object search using the Internet and a hybrid-based opinion analysis system for the same according to an embodiment of the present invention are largely a data collection server 100, a data analysis server 200, and indexing. The server 300 may include a web server 400 and a user terminal 500.

Here, the data collection server 100 is a server that collects web document data existing on the Internet 10, collects and stores RSS addresses on the Internet, and provides RSS files corresponding to the collected and stored RSS addresses. It collects web document data using link information provided by each RSS file.

The data collection server 100 is connected to the Internet 10 performs a function of transmitting a number of RSS addresses to be stored in a separate database (DB) by collecting a large number of RSS addresses in a normal automatic expansion method.

In this case, the automatic extension method includes a web (eg, general web in IPv4 and electronic products in IPv6) using a representative Internet resource (eg, RSS or ATOM) address representation form. Web addresses are extracted automatically from the document (HTML file), and the same way to extract the RSS address from the link (link) in the web document.

In other words, starting with a preconfigured main portal or blog web document, the web site automatically extracts the RSS address by visiting a web link that gradually goes outside, or the latest meta-site provides the web document to extract the RSS address. It is a method of extracting RSS address by visiting RSS file periodically and visiting the link address contained in it.

In addition, the data collection server 100 receives RSS files corresponding to RSS addresses stored in a separate database in advance and collects web document data using link information provided by each RSS file. .

That is, the data collection server 100 receives a list of RSS addresses periodically collected and stored in association with a separate database DB, visits each RSS address, downloads the corresponding RSS file, and then downloads each RSS file. Links that are present in the source link information among the RSS information provided (e.g. title, link, description, category, publication date information, etc.) To analyze the data by collecting the relevant web document data (e.g. RSS address, text link, date, title, body, tag, blog name, category, thumbnail, image, video, character count / image count / video count, etc.) The database 200 of the server 200 or the indexing server 300 is transmitted.

At this time, when visiting a link existing in the source link information, the RSS file list previously stored in a separate database (DB) is compared with the downloaded RSS file and present in the source link information of the updated RSS information among the contents of the RSS file. It is desirable to visit and collect links.

On the other hand, when collecting web document data by the data collection server 100, the activation check function for checking whether the link provided by each RSS file is activated, and spam RSS (eg, advertising and adult articles such as RSS feeds that only link to other sites, such as commercial RSS, report shops, and glyphs that update too quickly, and duplicate RSS feeds (for example, RSS1.0, RSS2.0, and Atom on the same blog) , RSS feeds via feedburner or meta blogs).

In addition, the data collection server 100 is a web application that exists on the Internet 10 in the manner proposed in Patent Application No. 2008-93125 (Opinion Retrieval System and Method Using the Internet) filed by the applicant. Downloading Hyper Text Markup Language (HTML) information of Web Sites in real time, and information required in the downloaded Web document data, for example, text, image, or video. At least one of the information data, such as information may be extracted and stored in a separate data storage means.

In addition, the data collection server 100 may collect and collect web document data including opinion information data (ie, general sentence / document data and information data given affirmative / negative evaluation thereof).

In this case, as a method for selectively collecting only web document data including the opinion information data, the specific web document data including the opinion information data is selected and machine learning algorithms (for example, SVM and K-) described later are selected. NN, Bayseian, etc.) to generate a web document screening model, and then use the generated web document screening model to selectively collect only web document data including opinion information data from the entire Internet web page. do.

Furthermore, the web document data collected by the data collection server 100 may be used immediately, but a method of classifying each domain by using a domain classification module (not shown) may be used.

On the other hand, the Internet (10) is a TCP / IP protocol and a number of services that exist in the upper layer, that is, Hyper Text Transfer Protocol (HTTP), Telnet, File Transfer Protocol (FTP), Domain Name System (DNS), SMTP (Simple Mail Transfer Protocol), Simple Network Management Protocol (SNMP), Network File Service (NFS), Network Information Service (NIS), and the like, a worldwide open computer network structure that provides the user terminal 500 will be described later It provides an environment that allows easy access to the web server 400. Meanwhile, the Internet 10 may be a wired or wireless internet, or may be a core network integrated with a wired public network, a wireless mobile communication network, or a portable internet.

The data analysis server 200 receives web document data collected from the data collection server 100, extracts meta information for each object, and affirmative / negative opinion statistics for each object of each object using the meta information for each object. It analyzes the information.

At this time, the data collection server 100 is a specific website (for example, http: // movie.daum) to periodically update the meta information (for example, actor, director, release date, etc.) for each individual on the Internet .net / moviedetail / moviedetailMain.do? movield = 52800, etc.) in advance to collect specific web document data, and the data analysis server 200 is a preset form from the specific web document data collected from the data collection server 100 Extract meta information for each object using regular expression that represents string of.

And, the regular expression is a symbol indicating a string of some form, for example, "a href =" (. *?) "" If a regular expression is made in this manner, where "()" represents a group of strings, ". Refers to an arbitrary string, and "*" indicates zero or more occurrences of the string, and "?)" "Means a string before the" "" string.

Therefore, the string "http://test.com" can be found in the string <a href="http://test.com">. By using this regular expression, meta information about each entity can be extracted. At this time, the extracted information is transmitted to the indexing server 300 to be stored immediately.

In addition, in an embodiment of the present invention, the meta data for each object is extracted from the data analysis server 200 and transmitted to the indexing server 300. However, the present invention is not limited thereto, and the meta information for each object is extracted from the data collection server 100. It may be transmitted to be stored in the indexing server 300 or a separate database (DB) immediately.

In addition, the data analysis server 200 uses a machine learning model for each field (eg, movie, politics, economy, game, sports, etc.) preset for web document data collected from the data collection server 100. The first module 210 to classify by field and language processing on the web document data collected from the data collection server 100 to extract opinion sentences, and affirmative / negative opinions on the extracted opinion sentences. A second module 220 for classifying the expression and the third module 230 for determining which object corresponds to the extracted object-specific meta information on the web document data collected from the data collection server 100. And a fourth module 240 for determining which sub-theme corresponds to the word and part-of-speech information around the opinion sentence extracted through the second module 220, and the extracted entity. Star meta information And a fifth module 250 for analyzing positive / negative statistical information about each object of each individual by using information output from the first to fourth modules 210 to 240.

Here, the second module 220 may include a language processor 221, an opinion / non-discrimination division 222, an opinion expression division 223, and the like.

The language processing unit 221 separates the web document data collected from the data collection server 100 or stored in a separate database (DB) in sentence units, and performs language processing on each of the separated sentences. ) To extract them.

In this case, the linguistic processing may be performed by, for example, Morpheme Analyze or Segmentation, but in addition to the irradiation process for extracting features (or index words), Korean refractive processing, or circular return processing, etc. It may be.

And, the opinion / disagreement separator 222 performs a function of distinguishing the opinion / disagreement sentences using linguistic features of each sentence extracted from the language processing unit 221.

That is, the sentences extracted from the language processing unit 221 include sentences with opinions, and general sentences without opinions. These sentences may be divided into sentences in which an opinion exists and sentences in which an opinion does not exist using the opinion / non-comment division unit 222.

The opinion / disagreement separator 222 can be easily implemented using the conventional machine learning algorithm described above. Specifically, first, a data set composed of opinions and a data set composed only of fact information are collected. Thereafter, for example, Morpheme Analyze or Segmentation is performed to extract an appropriate linguistic feature.

Here, the spacing is a process of dividing an input sentence into units having meanings. For example, if the input sentence says "I enjoyed the movie", the resulting sentence translates to "I enjoyed the movie".

In addition, the morpheme analysis (Morpheme Analyze) is a task for finding what part of speech information for each of the divided units. For example, if the input sentence says "I enjoyed the movie", the result sentence reads "I (CTP1 first person pronoun) + (fjb assistant) movie (CMCN secretary common noun) + (fjco purpose check) (YBDO general verbs) + crab (fmoca auxiliary verb) + (YBDO general verbs) + (fmbtp past tense first ending endings) + da (fmofd flat ending endings).

Next, using the extracted linguistic features, learning is performed by selecting a general machine learning algorithm, for example, Navve Baysian, SVM, K-NN or other models.

After the learning is completed, if any sentence or document is input, the opinion / non-dispute classification model that can distinguish whether the data is opinion data or fact data, that is, the opinion / non-comment separator 222 may be implemented. .

In addition, the opinion expression division unit 223 performs a function of dividing the language features (feature) of the opinion sentences separated from the opinion / non- opinion division unit 222 into positive / negative opinion expressions.

That is, the opinion expression division unit 223 finds a part that is a positive / negative opinion among the input comment sentences and displays the part. On the other hand, instead of using the opinion / disagreement divider 222, the opinion expression divider 223 may be used to display a positive / negative expression portion in the input sentence.

The opinion expression division unit 223 quantifies the degree of affirmation / negativeness of all words such as general self-help words and words as well as salmon, and utilizes them as a resource, and finds a machine learning model for finding positive / negative expressions in sentences. Can be used to generate

In addition, the second module 220 according to the appropriate application of the rule-based model and / or machine learning model when distinguishing positive / negative opinion expressions for the extracted opinion sentences. Opinion Mining (eg, Cascading, Interpolation, and Mixed Opinion Analysis) may be performed.

In the cascading opinion analysis method, a rule set having a very high accuracy is defined, and then a rule is first applied to the web document data crawled from the data collection server 100 to determine an opinion and then applied to the rule. In the case of the example that does not apply, the pre-statistic post-statistic method to apply the learning method. That is, the cascading method is a method of solving an exceptional case that cannot be covered by a rule through machine learning, and has a high usefulness in practical situations.

In case of applying such a cascading opinion analysis method, the second module 220 applies a rule-based model that is set in advance when a predefined rule exists for the extracted opinion sentence. Positive / negative opinion expressions may be divided, and if there is no predefined rule for the extracted opinion sentence, a predetermined machine learning model may be applied and divided into positive / negative opinion expressions.

In addition, the interpolation opinion analysis method uses two different methods, that is, the rule-based and machine learning methods to determine whether the opinion document is positive or negative, and then the polarity of the two methods. If this is the same, a method of giving high reliability is a method of increasing the reliability of the positive and negative judgment results through different methods. That is, by first presenting the opinion polarity having high reliability through the interpolation method, it is possible to obtain high reliability from high users.

When applying the interpolation opinion analysis method, the second module 220 simultaneously expresses positive / negative opinions by simultaneously applying a rule-based model and a machine learning model previously set to the extracted opinion sentences. After the determination, different reliability scores may be given according to whether the rule-based model and the machine learning model are matched, and classified into positive / negative opinion expressions based on the reliability scores.

In addition, the mixed opinion analysis method employs a cascading method of applying a rule first in an exceptional situation in which a rule may increase high performance, and in the general rule situation, two different methods (rule-based and It is a method that uses interpolation method using machine learning method, which is difficult to process by machine learning method, and has the strength in the part that can define certain rules and can use reliability concept as it is.

When the mixed opinion analysis method is applied, the second module 220 determines whether the exception processing rule candidate sentence is predefined for the extracted opinion sentence, and the extracted opinion sentence is defined in advance. In the case of an exception handling rule candidate sentence, it can be classified into positive / negative / neutral opinion expression by applying a predefined exception handling rule-based model or a machine learning model according to the existence of a predefined exception handling rule.

If the extracted opinion sentence is not a predefined exception handling rule candidate sentence, the second module 220 simultaneously performs a rule-based model and a machine learning model preset for the extracted opinion sentence. After determining the positive / negative opinion expression by applying, different reliability scores can be given according to the result of applying the rule-based model and the machine learning model, and can be divided into positive / negative expression based on the reliability score. have.

In addition, the second module 220 is positive / negative / by applying a predefined exception handling rule-based model when the extracted comment sentence is a predefined exception handling rule candidate sentence, if a predefined exception handling rule exists. In the case of a neutral opinion expression, the extracted opinion sentence may be divided into positive / negative / neutral opinion expression by applying a preset machine learning model when a predefined exception handling rule does not exist.

The indexing server 300 may store positive / negative statistical information about each object of each object, including meta information for each object collected and / or analyzed from the data collection server 100 and / or the data analysis server 200. DB) to index and store it.

In this case, the indexing process is generally used in a search engine, and the following rank uses an information retrieval algorithm using the number of reverse links and external links for the corresponding web page. This algorithm is more advantageous for links with many reverse links, and each link distributes its own scores to links extending from it.

In the case of a synthetic search keyword instead of a single search keyword, after extracting links commonly included among links of each morpheme of the search keyword input by the user, information about how close each morpheme exists in each link document and the corresponding information is given. The rank information of the link is calculated and displayed to the corresponding users in order of score.

To explain this briefly, an inverted list of morphemes for each link is created based on morphemes, stored in rank order, and the user enters a search keyword. After the links are obtained from the Inverted List, they are presented to users in the order of scores calculated using boolean, distance, TF (Term Frequency), and Inversed Document Frequency (IDF). Will be shown.

In addition, positive / negative opinion statistics information for each object of each object stored in the indexing server 300 may include, for example, object ID, sub-theme, number of positive / negative opinion expressions for each object of each object, At least one of the total number of opinion expressions or the content of opinion expression using each object may be included.

In addition, the indexing server 300 may group and store each sub-theme through clustering. In this case, the clustering is performed by using information on which opinions each sub-theme appeared like. For example, each sub-theme such as a story, content, story, plot, etc. may be grouped into a "story", and each sub-theme such as an action, a sight, a scene, etc. may be grouped and stored as a "scene". .

In addition, positive / negative consensus statistical information about each object of each object stored in the indexing server 300 may be stored so that each object is mapped on a document / paragraph basis.

In addition, the web server 400 receives user search keywords transmitted from each user terminal 500 through the Internet 10, and interoperates with the indexing server 300 to store meta information and / or stored in advance in the corresponding user search keywords. It is determined whether the target keyword exists, and if the meta information and / or target keyword pre-stored in the corresponding user search keyword exist, the object list result related to the corresponding meta information and / or target keyword is displayed on the screen of the corresponding user terminal 500 ( Display) function.

In addition, when the meta information or target keyword stored in advance in the user search keyword does not exist, the web server 400 searches the user search keyword in the indexing server 300 to which the object is mapped and displays a list of related object results. I can do it.

The web server 400 analyzes a user search keyword transmitted from the user terminal 500 to determine whether there is meta information or a target keyword stored in the indexing server 300, and classifies the keyword search method according to the determination result. In accordance with the keyword analysis module 410 and the keyword search method categorized from the keyword analysis module 410, the indexing server 300 may be linked to search for objects related to the corresponding meta information or the target keyword and may be screened on the user terminal 500. The keyword search module 420 for displaying the object list result may be included.

In addition, when the meta information and the target keyword pre-stored in the user search keyword transmitted from the user terminal 500 exist, the web server 400 interworks with the indexing server 300 to search for entities related to the meta information keyword. In response to the searched entities, the object result list may be rearranged and displayed in any one of affirmative / negative or total opinion order of the corresponding target keyword.

In addition, when the meta information and the target keyword stored in advance in the user search keyword transmitted from the user terminal 500 do not exist, the web server 400 performs stemming of the corresponding user search keyword and then analyzes the user search keyword. Compares and analyzes documents / paragraphs stored in the indexing server 300 to search for the stemmed user search keyword in the document / paragraph result list, and converts the individual result list mapped to the searched document / paragraph result list into the corresponding user terminal. Can be displayed on the 500.

In addition, the web server 400, when there is no meta information and a target keyword pre-stored in the user search keyword transmitted from the user terminal 500, and the pre-stored meta information keyword is present, the web server 400 performs stemming analysis of the corresponding user search keyword. Then, by comparing and analyzing the stemmed user search keyword excluding the corresponding meta information keyword and the documents / paragraphs stored in the indexing server 300, searching for the stemmed user search keyword excluding the meta information keyword in the document / paragraph result list. After searching the entity result list mapped to the searched document / paragraph result list, the entity result list related to the meta information keyword is filtered in the searched entity result list and displayed on the corresponding user terminal 500. Can be.

In addition, when the meta information and the target keyword pre-stored in the user search keyword transmitted from the user terminal 500 do not exist and the pre-stored target keyword exists, the web server 400 stems the corresponding user search keyword. Compare and analyze the stemmed user search keyword excluding the target keyword and the documents / paragraphs stored in the indexing server 300, and search for the stemmed user search keyword excluding the target keyword in the document / paragraph result list. After retrieving the entity result list mapped to the searched document / paragraph result list, the entity result list is rearranged in the order of positive or total opinions of the target keyword and displayed on the corresponding user terminal 500. I can do it.

In addition, the web server 400 stems from the user search keyword if the meta information and the target keyword are not stored in the user search keyword transmitted from the user terminal 500 and the meta information and the target keyword are stored in advance. Then, a comparison of the stemmed user search keyword excluding the meta information and the target keyword and documents / paragraphs stored in the indexing server 300 is performed, and the stemmed user search keyword excluding the meta information and the target keyword is analyzed. Search the paragraph result list, search the entity result list mapped to the searched document / paragraph result list, filter the entity result list related to the meta information keyword in the searched entity result list, and filter the entity result list Is the number of positive or total comments for that target keyword The result list of the objects may be rearranged in ascending order and displayed on the corresponding user terminal 500.

In addition, the web server 400 may display opinion statistics information of each entity on the corresponding user terminal 500 together with the entity list result.

In addition, the web server 400 may display the individual result list in the order of positive / negative opinions according to the opinion analysis result for the user search keyword transmitted from the user terminal 500.

In addition, the user terminal 500 is connected to the web server 400 through a wired or wireless communication network such as, for example, a network or the Internet, and the web server 400 through a conventional web browser. You can get a variety of services provided by).

The user terminal 500 is preferably implemented as a personal computer (PC), but is not limited thereto. A notebook, a personal digital assistant, or a PDA capable of connecting and communicating with the Internet may be used. ), A PDA phone or a DMB (Digital Multimedia Broadcasting) phone with a communication function can be implemented as any device having a communication function.

Hereinafter, an object search using the Internet and a hybrid-based opinion analysis method for the same according to an embodiment of the present invention will be described in detail.

3 is an overall flowchart illustrating an object search using the Internet and a hybrid-based opinion analysis method for the same according to an embodiment of the present invention, and FIGS. 4 and 5 are each entity applied to an embodiment of the present invention. This is a flow chart for explaining in detail the process of analyzing positive / negative opinion statistics information for each target of.

3 to 5, in order for users using the Internet to be provided with entity result lists and opinion statistics on specific user search keywords, first, meta information about each entity (eg, an actor for a movie) is provided. (Song Kang-ho, Lee Byung-hun, etc.), directors (Bong Joon-ho, compulsory rules, etc.), release date, etc.} should be extracted in advance and stored keywords for the sub-theme.

In this case, the meta information of each entity is information stored in advance by extracting all values in each field of the database DB of the indexing server 300 (see FIG. 1). The sub-theme is information that is held when performing clustering.

That is, the web document data existing on the Internet is collected through the data collection server 100 (refer to FIG. 1) (S100), and the web document data collected in step S100 is referred to as the data analysis server 200 (refer to FIG. 1). After extracting the meta information for each object through the object, the positive / negative opinion statistics information for each object of each individual is analyzed using the meta information for each object (S200).

In this case, the meta information for each object may designate a web site in which a web crawler of the data collection server 100 periodically updates such information on the Internet, and update the information using a regular expression in the designated web site. do.

In addition, the method of constructing the positive / negative opinion statistics information for each sub-theme of each object, first, after collecting the web document data through the data collection server 100, the data analysis server 200 The classified web document data belongs to which field (eg, movie, politics, economy, game, sports, etc.) using a pre-set automatic classification model based on machine learning, and classifies the collected web document data. Opinion Mining finds positive / negative expressions in each sentence.

In addition, the data analysis server 200 determines which entity the collected web document data indicates. That is, it determines which entity is close to using meta information (eg, movie-director, actor, etc.) of each entity.

For example, in the web document data (Doc 1) collected when the meta information of the entity A has "a, b, c", the entity name and the frequency of occurrence of each meta information, etc. The probability that the corresponding web document data Doc1 points to the entity A is obtained by using.

On the other hand, if the comparison in the document unit as described above, after dividing the web document data (Doc1) by each sentence / paragraph / arbitrary phrase unit meta-information as described above to determine the probability that each unit data is any entity (Entity) You can also get it.

Obtain entity candidates indicated by each unit data (document / sentence / paragraph / arbital phrase) and then select the entity with the highest probability, or select the highest N entities with the highest probability. ) Can be selected.

In addition, the opinion information displayed in the unit data is regarded as the opinion information on the entity for the entity selected as the unit data unit, and the opinion information is stored in the database DB of the indexing server 300. Done.

In addition, the data analysis server 200 finds out which object (sub-theme) each opinion represents in the collected web document data. At this time, the sub-theme is classified using meta information or words and parts-of-speech information around the target.

For example, in this sentence `` This / NNG movie / NNG + / JX story / NNG + / JKS really / MAG <positive> okay / VA + did / EP + da / EF </ positive> '' Sub-theme must be found. At this time, the noun "story / NNG" in front of the main investigation "JKS" is a sub-theme. I can guess.

More specifically, for a movie entity called "Shin Ki-jeon", if there are four meta-information of "Director: Kim Yu-jin", "Actress: Jae-Young Jung, Han Eun-jung, Hur Jun-Ho", document / paragraph A Determines whether there is a word / paragraph that is likely to refer to an entity called "new sensation" by identifying the surrounding words that represent any entity such as "new sensation", [new sensation], "new sensation", movie novel, late sensation, or appreciation of the epoch. do.

Then, after grasping through the surrounding context as above to indicate "negative mechanism", we can see how much of the four meta-informations we have above. If two meta-information exists, "Jeong Jae-Young and Kim Yu-Jin", two out of four exist. In other words, it is understood that 50% exists. Thus, if any threshold is set and more than that, the document / paragraph pointing to the entity is identified.

As described above, the sub-theme can predict and display the parts of speech and the parts of speech information surrounding it, and thus, the rule-based method that directly assigns each case, and the object under a specific context. Any machine learning method that finds words that are likely to be sub-theme is possible.

Also, before storing opinion statistics information, each sub-theme can be grouped through clustering. For example, you can group sub-themes such as stories, content, stories, plots, etc. into "story". In this case, the clustering is performed by using information on which sub-theme has appeared with which opinion expression.

On the other hand, affirmative / negative opinion statistics information for each object of each entity may include, for example, an entity ID, a sub-theme, the number of positive / negative opinion expressions for each object of each individual, the total number of opinion expressions, or each object. At least one of the opinion expression used content may be made.

For example, if you find an expression with a positive opinion in the "Emotions" section on an Entity A, the Sub- In the theme), the positive count expression is increased by 1 and stored.

That is, as shown in Figures 4 and 5, the step S200, the step of classifying by sector using the machine learning model for each field set in advance for the web document data collected in the step S100 (S210), and Performing language processing on the web document data collected in step S100 to extract a comment sentence, and dividing the extracted comment sentence into positive / negative expressions (S220) and the web collected in step S100; Determining which object corresponds to the document data using meta information for each object (S230), and determining which object corresponds to the object using word and part-of-speech information around the opinion sentence extracted in step S220. It may include the step (S240).

Here, in step S220, the web document data collected in step S100 is separated into sentence units, and language processing is performed on each of the separated sentences to extract linguistic features (S220-1), and A step S220-2 is used to classify the opinion / non-sentence sentences using the linguistic qualities of the sentences extracted in step S220-1, and affirms the linguistic qualities of the comment sentences separated in step S220-2. It may include a step (S220-3) to divide the negative opinion expression.

In addition, in step S220, hybrid opinion analysis according to appropriate application of a rule-based model and / or a machine learning model when distinguishing positive / negative opinion expressions on the extracted opinion sentences. (Opinion Mining) (eg, Cascading, Interpolation, and Mixed Opinion Analysis, etc.) may be performed.

6 to 9 are flowcharts illustrating a hybrid-based opinion analysis method according to an embodiment of the present invention in detail, FIG. 6 is a flowchart illustrating a cascading opinion analysis method, and FIG. 7 is interpolated. (Interpolation) A flowchart illustrating a method of analyzing opinions, FIG. 8 is a diagram illustrating a reliability score according to each opinion expression of FIG. 7, and FIG. 9 is a flowchart illustrating a mixed opinion analysis method.

Referring to FIG. 6, it is determined whether a predefined rule exists for the opinion sentence extracted in step S220 (S221), and as a result of the determination in step S221, a predefined rule exists for the extracted opinion sentence. In this case, a predetermined rule-based model is applied to distinguish positive / negative opinions (S222).

For example, a pre-set rule-based model (more + affirmations + no → affirmations) for the comment sentence "This movie can't be better" is categorized as a positive opinion.

Otherwise, if there is no predefined rule for the extracted opinion sentence as a result of the determination in step S221, a predetermined machine learning model is applied to classify affirmative / negative opinion representation (S223).

For example, if there is no pre-defined rule for the comment sentence "This movie is just like that", the opinion is analyzed in units of units using a preset machine learning model, and then the opinion is provided through the machine learning model. Probability sampling method for positive / negatives is divided into negative opinions.

Referring to FIGS. 7 and 8, after a rule-based model and a machine learning model are previously applied to the opinion sentences extracted in step S220, the positive / negative opinion expressions are respectively determined (S224a and S224b). ), Different reliability scores are assigned according to whether the rule-based model and the machine learning model are matched (S225), and classified into positive / negative opinion expressions based on the reliability scores (S226).

For example, a pre-set rule-based model for the comment sentence "This time A's overall story was too good" is divided into positive comments (e.g., this time A's overall story is too < positive> OK </ positive>.), the opinions are analyzed in units of units through a preset machine learning model, and then classified into positive opinions using the machine learning model as a probability extraction method for positive / negative opinions. , Movie A, this time, the overall story was <positive> too good </ positive>.).

Next, as shown in FIG. 8, different reliability scores (Positive-Positive or Negative-Negative (“Okay” in the example)) according to whether the result of applying the rule-based model and the machine learning model are matched. 2 points, neutral matched results 1 point for Positive-Objective, Negative-Objective ("too" in the example), and -1 for Positive-Negative and Negative-Positive if the two models yield completely opposite results ), And based on the reliability score can be divided into positive / negative opinion expression only when a certain reliability score or more.

Referring to FIG. 9, after it is determined whether or not the exception processing rule candidate sentence is predefined for the opinion sentence extracted in step S220 (S227), as a result of the determination in step S227, the extracted opinion sentence is an exception defined in advance. If you have a candidate sentence (e.g., it can't be better than this, you've seen the movie "The Best Moments of Our Lives", the main character A enjoyed this situation, ~~~~, etc.). Depending on the existence of the rule, a predefined exception handling rule-based model or machine learning model can be applied to classify as positive / negative / neutral.

If, as a result of the determination in step S227, the extracted opinion sentence is a predefined exception handling rule candidate sentence, and if a predefined exception handling rule exists, affirmative / negative by applying a predefined exception handling rule based model. After dividing by / neutral opinion expression, the extracted opinion sentence is divided into affirmative / negative / neutral opinion expression by applying a preset machine learning model when a predefined exception handling rule does not exist (S228).

On the other hand, as a result of the determination in step S227, when the extracted opinion sentence is not a predefined exception processing rule candidate sentence, a rule-based model and a machine learning model preset for the extracted opinion sentence are used. After determining the positive / negative opinion expression by applying it at the same time, different reliability scores are assigned according to whether the rule-based model and the machine learning model are matched, and classified into positive / negative expression based on the reliability score. (S229).

Next, affirmative / negative opinion statistics information for each object of each object including the meta data for each object analyzed in step S200 is indexed to be stored in a database through the indexing server 300 (see FIG. 1) (S300). .

At this time, in step S300, the positive / negative opinion statistics information for each object of each object may be stored so that each object is mapped in units of documents / paragraphs. That is, through the indexing server 300, a database (DB) may be stored as, for example, a document / paragraph ID, a title, a content, a tag, a classification (eg, a movie, a politics, an economy, etc.), an object name list, and the like.

Finally, after determining whether there is a meta information or a target keyword stored in advance in the user search keyword transmitted from the user terminal 500 (see FIG. 1) through the web server 400, the meta information or the target stored in advance in the user search keyword is determined. If there is a keyword, the result of displaying the object list related to the corresponding meta information or the target keyword is displayed on the screen of the corresponding user terminal 500 (S400). In this case, in step S400, the object result list may be displayed in order of a large number of positive / negative opinions according to the opinion analysis result of the user search keyword.

For example, it is possible to search based on each field name, that is, a meta information keyword such as a movie name, actor name, production year, genre, and country. That is, when the user searches for "Bong Joon Ho Director Movie", it is determined that the meta information keyword "Bong Joon Ho" exists, and the Bong Joon Ho Director movies are searched by the indexing server 300.

Similarly, when searching for "songgangho movie", it is determined that the meta information keyword "songgangho" exists, and the indexing server 300 searches for the movie starring "songgangho".

On the other hand, it is possible to search by sorting by affirmative / negative opinions for each target based on target keywords such as general comments, stories, acting, actors, fun, excitement, inversion, and music. In other words, it is possible to search for movies with the most positive opinions about "emotion" or movies with the most negative opinions about "fun".

For example, when searching for "story movie", it is determined that there is a target keyword "story", and the indexing server 300 selects a movie having a lot of opinions or a lot of positive / negative opinions in the "story" part. Search. In addition, the search will be sorted by the number of "story" total opinion / positive / negative feedback.

Similarly, if a search is made for a movie with a good story, it is determined that the target keyword "story" is included, and the search sentence itself is analyzed to include positive opinions through opinion analysis. Therefore, the indexing server 300 searches for the movies with the most positive opinions in the "story" part.

In addition, when searching for "the story is a messed-up movie," it is determined that the target keyword "story" is included, and the search sentence itself through the analysis of opinions to find that negative opinions are contained. Therefore, the indexing server 300 searches in the order of movies with a large number of negative opinions in the "story" part.

10 to 14 are detailed flowcharts illustrating various keyword search methods according to user search keywords.

Referring to FIG. 10, after determining whether a meta keyword and a target keyword pre-stored in the user search keyword transmitted from the user terminal 500 (see FIG. 1) through the web server 400 (see FIG. 1) are present (S401). As a result of the determination in step S401, when the meta information and the target keyword pre-stored in the corresponding user search keyword exist, the entities related to the meta information keyword are searched using the indexing server 300 (see FIG. 1) (S402). ), The object result list is rearranged in the order of affirmative / negative or total opinion order of the corresponding target keywords for the objects searched in step S402 and displayed on the screen of the corresponding user terminal 500 (S403).

For example, a search for "Lee Byung Hun's touching movie" is a case where a meta information keyword "Lee Byung Hun" and a sub-theme of "Impression" are combined. In the indexing server 300, "Lee Byung Hun" The searched entities are searched first, and then the results of the entities are rearranged in the order of positive / negative or total opinions of the subjects “inspired” with respect to the corresponding entities. Display on.

10 and 11, after determining whether the meta information and the target keyword stored in advance in the user search keyword transmitted from the user terminal 500 through the web server 400 are present (S401), in step S401. As a result of the determination, if neither the meta information previously stored in the user search keyword and the target keyword exist, the process proceeds to step (A) and the user search keyword is morphologically analyzed (S404).

Subsequently, the documents / paragraphs stored in the indexing server 300 to which the object candidate list is mapped in units of documents / paragraphs are compared and analyzed in step S404 (S405).

Then, after searching the user search keyword stemmed in step S404 from the document / paragraph result list (S406), the user result list mapped to the document / paragraph result list retrieved in step S406 is corresponding to the corresponding user terminal. The display on the screen 500 is displayed (S407).

For example, a search for "movie chased after" does not include both a meta information keyword and a target keyword, and as described above, on a database (DB) in which an object candidate list is mapped on a document / paragraph basis. Perform a search.

That is, in such a database DB, for example, a search keyword of a title, a tag, a text and a search term of the user's search keyword may be searched for a stemmed keyword (chase / VV + high / EC chasing / VV + / ETM movie / NNG). do. At this time, the ranking is based on the degree of relevance of the corresponding user search keyword and the document / paragraph.

To explain this in detail, if there is a user search keyword "movie chased after," a document containing many of these words "chase", "chased", and "movies" is more relevant (TF: Term Frequency-Document Frequency in the corresponding term), plus more weights that contain less frequent words (IDF: Inverse Document Frequency) than "movie", such as "chase" and "chase". Is given.

As described above, the document / paragraphs related to the user search keyword are searched, and the object list mapped to the corresponding result in the document / paragraph result list is returned as the object search result.

In this case, after N documents / paragraphs are searched, the first method of returning the objects mapped to each of the N results as a search result first and second, and calculating the frequency of each object in the top K, A second method of returning in high order and a third method of interpolating and returning a search result rank score and frequency are possible.

Here, the third method may normalize and return a rank score, for example. At this time, the rank score normalization may be converted to a value between [0-1] by normalizing to "(rank rank-lowest rank score) / (highest rank score-lowest rank score) of Entity [i]".

In addition, the frequency score normalization is normalized to "(frequency of Entity [i]-frequency of Entity with lowest frequency) / (frequency of Entity with highest frequency-frequency of Entity with lowest frequency)" between [0-1]. Can be converted to a value.

In addition, it can be converted by normalizing to "Score (Entity [i]) = (1-lambda) * (rank score of normalized Entity [i]) + lambda * (frequency of normalized Entity [i])".

Referring to FIG. 12, after determining whether the meta information and the target keyword stored in advance in the user search keyword transmitted from the user terminal 500 through the web server 400 exist (S408), as a result of the determination in step S408, the corresponding result is determined. If neither the meta information prestored in the user search keyword nor the target keyword exist, it is determined whether the meta information keyword pre-stored in the user search keyword exists (S409).

As a result of the determination in step S409, when there is a meta information keyword stored in advance in the user search keyword, after the stemming of the corresponding user search keyword is performed (S410), the stemming analysis of the user in step S410 except for the meta information keyword is performed. The document / paragraphs stored in the indexing server 300 to which the entity candidate list is mapped in terms of keywords and documents / paragraphs are compared and analyzed (S411).

Then, after searching for the user search keyword which has been stemmed in step S410 except for the meta information keyword in the document / paragraph result list (S412), the object mapped to the document / paragraph result list retrieved in step S412. The result list is retrieved (S413).

Thereafter, the entity result list related to the corresponding meta information keyword is filtered in the entity result list found in step S413 and displayed on the corresponding user terminal 500 (S414).

For example, a search for "Ha Jung-woo movie chased and chased" will find that there is a meta information keyword "Ha Jung-woo", and another search keyword "chased and chased". Meanwhile, "movie" is a common word in a movie entity search and may be included in the exception word list.

First, an entity search is performed by "tracking and chasing" to obtain an entity result list, and then, on the obtained entity result list, the entity result list is filtered by the meta information keyword "Hajeongwoo". Finally, return the filtered result.

12 and 13, after determining whether the meta information and the target keyword stored in advance in the user search keyword transmitted from the user terminal 500 through the web server 400 exist (S408), the determination in step S408 is performed. As a result, when neither the meta information prestored in the user search keyword nor the target keyword exist, it is determined whether the meta information keyword prestored in the user search keyword exists (S409).

As a result of the determination in step S409, if there is no meta information keyword previously stored in the user search keyword, the process proceeds to step (B) to determine whether the target keyword previously stored in the user search keyword exists (S415).

As a result of the determination in step S415, if there is a target keyword stored in advance in the corresponding user search keyword, the user search keyword is morphologically analyzed (S416), and the user search keyword which is morphologically analyzed in step S416 except for the target keyword is selected. The document / paragraphs stored in the indexing server 300 to which the entity candidate list is mapped on a document / paragraph basis are compared and analyzed (S417).

Then, after searching for the user search keyword stemmed in step S416 except for the target keyword in the document / paragraph result list (S418), the entity result mapped to the document / paragraph result list retrieved in step S418. Search the list (S419).

Thereafter, the entity result list retrieved in step S419 is rearranged in the order of positive or total opinions of the corresponding target keywords and displayed on the screen of the corresponding user terminal 500 (S420).

For example, a search for "inspiring war movie" identifies a sub-theme called "inspiration" and an entity search element called "war." That is, the keywords to search for the object are not meta information keywords or target keywords.

First, an entity search is performed with the word "war", and the resultant entity result list is rearranged in order of the total number of opinions or the number of positive opinions with the object of "inspiration". Finally, the reordered result is returned.

12 and 14, after determining whether the meta information and the target keyword stored in advance in the user search keyword transmitted from the user terminal 500 through the web server 400 exist (S408), the determination in step S408 is performed. As a result, when neither the meta information prestored in the user search keyword nor the target keyword exist, it is determined whether the meta information keyword prestored in the user search keyword exists (S409).

As a result of the determination in step S409, if the meta information keyword previously stored in the user search keyword exists, the process proceeds to step (C) to determine whether the target keyword previously stored in the user search keyword exists (S421).

As a result of the determination in step S421, if there is a target keyword stored in advance in the corresponding user search keyword, after the stemming of the corresponding user search keyword (S422), the user who has been stemmed in step S422 except for the corresponding meta information and the target keyword The document / paragraphs stored in the indexing server 300 to which the object candidate list is mapped in units of search keywords and documents / paragraphs are analyzed in operation S423.

Then, after searching the document / paragraph result list except for the meta information and the target keyword, the user search keyword searched in step S422 is mapped to the document / paragraph result list searched in step S424. The searched object result list is searched (S425).

Subsequently, after filtering the entity result list related to the corresponding meta information keyword from the entity result list retrieved in step S425 (S426), the object result list filtered in step S426 is positive or total opinion of the target keyword. The result list of the objects is rearranged in the order of increasing number and displayed on the screen of the corresponding user terminal 500 (S427).

For example, searching for "Emotional War Movies by Jeong Jae-young" identifies a sub-theme of "emotion" and a meta-information keyword of "Jeong Jae-young." It also knows that an entity search element called "war" is also included.

First, an entity search is performed with the word "war", and then the result list of the entity is filtered by the meta information keyword "Jeong Jae-young". Then, the filtered object result list is rearranged in order of the total number of opinions or the number of positive opinions.

15 to 19 are diagrams illustrating keyword search result screens displayed on a screen of a user terminal by using an entity search using the Internet and a hybrid-based opinion analysis method for the same according to an embodiment of the present invention.

15 and 16, a screen showing a search result for meta information keywords such as “transformer” and “Ha Jung-woo”, and basic searched movie information (eg, post, title, director, actor, etc.) are searched on the search result screen. ) And the opinion display information (eg, total rating, rating for each item and rating method) for each target item (eg, general rating, scene, story, fun, actor, impression, acting, etc.) (ex, graph), etc.), a second display window 2000 expressing a searched object result list and related image, video, and positive / negative statistical information, and a searched basic movie. A fourth display window 4000 that expresses the original text of the opinion of the movie (eg, the movie with the most overall opinion) may be included.

Here, the object result list searched on the third display window 3000 may be rearranged and displayed in the order of the highest total opinion or the order of the most positive / negative opinion.

Meanwhile, referring to FIG. 17, unlike FIG. 16, another screen showing a search result for the meta information keyword “Ha Jung-woo” is a display screen configured to be simpler and easier to see.

Referring to FIG. 18, a screen showing a search result of a sub-theme keyword of "movie with inversion", and the search results are sorted in order of the total opinion or positive opinion of the object of "inversion". You can display the object result list. In FIG. 17, a display screen is arranged in order of affirmative opinion about "inversion".

Referring to FIG. 19, a screen showing an individual search result for a user search keyword of “movie chased and chased”, and on the left screen, an object search list (eg, a chaser, a fighting skill, an elderly person) for the user search keyword. There is no country, nom nom, etc.) is displayed.

Although the above-described preferred embodiments of the object search using the Internet according to the present invention and a hybrid-based opinion analysis system and method therefor have been described, the present invention is not limited thereto, and the claims and the detailed description of the invention are described. And it is possible to carry out various modifications within the scope of the accompanying drawings, which also belongs to the present invention.

FIG. 1 is a block diagram illustrating an entity search using the Internet and a hybrid-based opinion analysis system for the same according to an embodiment of the present invention.

2 is a block diagram illustrating in detail a data analysis server applied to an embodiment of the present invention.

FIG. 3 is a flowchart illustrating an object search using the Internet and a hybrid analysis method based on the same according to an embodiment of the present invention.

4 and 5 are flowcharts illustrating in detail a process of analyzing affirmative / negative opinion statistics information for each object of each entity applied to an embodiment of the present invention.

6 to 9 are flowcharts illustrating a hybrid-based opinion analysis method according to an embodiment of the present invention in detail, FIG. 6 is a flowchart illustrating a cascading opinion analysis method, and FIG. 7 is interpolated. (Interpolation) A flowchart illustrating a method of analyzing opinions, FIG. 8 is a diagram illustrating a reliability score according to each opinion expression of FIG. 7, and FIG. 9 is a flowchart illustrating a mixed opinion analysis method.

10 to 14 are detailed flowcharts illustrating various keyword search methods according to user search keywords.

15 to 19 are diagrams illustrating keyword search result screens displayed on a screen of a user terminal by using an entity search using the Internet and a hybrid-based opinion analysis method for the same according to an embodiment of the present invention.

Claims (37)

delete A first server for collecting web document data present on the Internet; A data analysis server that receives the web document data collected from the first server, extracts meta information for each object, and analyzes positive / negative statistical information about each object of each object by using the meta data for each object; A second server for indexing positive / negative opinion statistics information for each object of each individual including the meta information for each entity analyzed from the data analysis server to be stored in a database; And Receives a user search keyword transmitted from a user terminal connected through the Internet, and interoperates with the second server to determine whether there is meta information or a target keyword pre-stored in the user search keyword, and meta information prestored in the user search keyword. Or, if the target keyword exists, the web server displaying the result of the object list related to the corresponding meta information or the target keyword on the screen of the corresponding user terminal. The first server collects and stores RSS addresses on the Internet, An object search using the Internet and a hybrid based opinion analysis system for collecting web document data using the link information provided by each RSS file by receiving RSS files corresponding to the stored RSS addresses. The method of claim 2, The data analysis server is a hybrid based opinion analysis system for object searching using the Internet, characterized in that for extracting meta information for each object by using a regular expression representing a string of a predetermined form in the web document data. A first server for collecting web document data present on the Internet; A data analysis server that receives the web document data collected from the first server, extracts meta information for each object, and analyzes positive / negative statistical information about each object of each object by using the meta data for each object; A second server for indexing positive / negative opinion statistics information for each object of each individual including the meta information for each entity analyzed from the data analysis server to be stored in a database; And Receives a user search keyword transmitted from a user terminal connected through the Internet, and interoperates with the second server to determine whether there is meta information or a target keyword pre-stored in the user search keyword, and meta information prestored in the user search keyword. Or, if the target keyword exists, the web server displaying the result of the object list related to the corresponding meta information or the target keyword on the screen of the corresponding user terminal. The data analysis server, A first module for classifying the collected web document data by sector using a preset machine learning model for each sector; A second module for performing language processing on the collected web document data to extract opinion sentences, and dividing the extracted opinion sentences into positive / negative opinion expressions; A third module for determining which object corresponds to the collected web document data using the object-specific meta information; And And a hybrid module for object searching using the Internet, characterized in that it comprises a fourth module for determining which object is used by using words and parts of speech information around the opinion sentence extracted through the second module. Opinion Analysis System. 5. The method of claim 4, The second module, A language processor that separates the collected web document data into sentence units and performs linguistic processing on each of the separated sentences to extract linguistic features; An opinion / non-computation division unit for classifying opinion / non-comment sentences using the linguistic qualities of the extracted sentences; And An object search using the Internet and a hybrid-based opinion analysis system for the same, characterized in that it comprises a opinion expression division for classifying the linguistic qualities of the divided opinion sentences into positive / negative opinion expression. 5. The method of claim 4, The second module applies a rule-based model that is set in advance when there is a predefined rule for the extracted opinion sentence, and divides it into positive / negative opinion expressions. If there is no predefined rule for the extracted opinion sentence, applying the preset machine learning model to classify the expression into positive / negative opinion expression. Analysis system. 5. The method of claim 4, The second module determines a positive / negative opinion expression by simultaneously applying a rule-based model and a machine learning model to the extracted opinion sentence at the same time, and then applying the rule-based model and the machine learning model. An object search using the Internet and a hybrid-based opinion analysis system for assigning different reliability scores according to whether or not the results are matched, and classifying them into positive / negative opinion expressions based on the reliability scores. 5. The method of claim 4, The second module determines whether the extracted exception sentence is a predefined exception handling rule candidate sentence, and if the extracted opinion sentence is a predefined exception handling rule candidate sentence, An object search using the Internet and a hybrid-based opinion analysis system for distinguishing positive / negative / neutral opinion expressions by applying a predefined exception handling rule-based model or a machine learning model according to existence. The method of claim 8, The second module is affirmed by simultaneously applying a rule-based model and a machine learning model preset to the extracted opinion sentence when the extracted opinion sentence is not a predefined exception handling rule candidate sentence. After determining the negative opinion expression, the different reliability scores are assigned according to whether the rule-based model and the machine learning model are matched, and classified into positive / negative opinion expressions based on the reliability scores. Object search using the Internet and hybrid-based opinion analysis system. The method of claim 8, The second module may express the positive / negative / neutral opinion by applying a predefined exception handling rule-based model when the extracted opinion sentence is a predefined exception handling rule candidate sentence and a predefined exception handling rule exists. Separated by When the extracted opinion sentence does not have a pre-defined exception handling rule, a predetermined machine learning model is applied to classify the object into positive / negative / neutral opinion expressions. Based opinion analysis system. The method according to claim 2 or 4, Positive / negative opinion statistical information about each object of each object stored in the second server is the object ID, the object, the number of positive / negative opinion expressions for each object of each object, the total number of opinion expressions, or opinions using each object. An object search using the Internet, characterized in that the information consists of at least one of the contents and a hybrid-based opinion analysis system for this. A first server for collecting web document data present on the Internet; A data analysis server that receives the web document data collected from the first server, extracts meta information for each object, and analyzes positive / negative statistical information about each object of each object by using the meta data for each object; A second server for indexing positive / negative opinion statistics information for each object of each individual including the meta information for each entity analyzed from the data analysis server to be stored in a database; And Receives a user search keyword transmitted from a user terminal connected through the Internet, and interoperates with the second server to determine whether there is meta information or a target keyword pre-stored in the user search keyword, and meta information prestored in the user search keyword. Or, if the target keyword exists, the web server displaying the result of the object list related to the corresponding meta information or the target keyword on the screen of the corresponding user terminal. The web server, when there is no meta information or a target keyword stored in advance in the user search keyword, the second server that is the object mapping is searched by the user search keyword to display a list of related object results. Object search using the Internet and hybrid-based opinion analysis system. The method according to any one of claims 2, 4 or 12, And the web server is a hybrid based opinion analysis system for object searching using the Internet, characterized in that to display a list of object results in the order of a lot of positive / negative opinions according to the opinion analysis results for the user search keyword. A first server for collecting web document data present on the Internet; A data analysis server that receives the web document data collected from the first server, extracts meta information for each object, and analyzes positive / negative statistical information about each object of each object by using the meta data for each object; A second server for indexing positive / negative opinion statistics information for each object of each individual including the meta information for each entity analyzed from the data analysis server to be stored in a database; And Receives a user search keyword transmitted from a user terminal connected through the Internet, and interoperates with the second server to determine whether there is meta information or a target keyword pre-stored in the user search keyword, and meta information prestored in the user search keyword. Or, if the target keyword exists, the web server displaying the result of the object list related to the corresponding meta information or the target keyword on the screen of the corresponding user terminal. The web server, A keyword analysis module that analyzes a user search keyword transmitted from the user terminal to determine whether there is meta information or a target keyword stored in the second server, and classifies a keyword search method according to the determination result; And And a keyword search module for interworking with the second server according to the keyword search method classified from the keyword analysis module to search for entities related to the corresponding meta information or target keyword and to display the object list result on the screen of the corresponding user terminal. Object search using the Internet and hybrid based opinion analysis system for it. The method according to any one of claims 2, 4, 12 or 14, The web server searches for entities related to the meta information keyword if meta information and target keyword pre-stored in the user search keyword exist, and then, among the affirmative / negative or total opinion order of the target keyword for the searched entities. An object search utilizing the Internet, characterized by rearranging and displaying the object result list in any one order, and a hybrid-based opinion analysis system for the same. A first server for collecting web document data present on the Internet; A data analysis server that receives the web document data collected from the first server, extracts meta information for each object, and analyzes positive / negative statistical information about each object of each object by using the meta data for each object; A second server for indexing positive / negative opinion statistics information for each object of each individual including the meta information for each entity analyzed from the data analysis server to be stored in a database; And Receives a user search keyword transmitted from a user terminal connected through the Internet, and interoperates with the second server to determine whether there is meta information or a target keyword pre-stored in the user search keyword, and meta information prestored in the user search keyword. Or, if the target keyword exists, the web server displaying the result of the object list related to the corresponding meta information or the target keyword on the screen of the corresponding user terminal. Positive / negative opinion statistics information for each object of each object stored in the second server is stored so that each object is mapped on a document / paragraph basis. If the meta information and the target keyword previously stored in the user search keyword do not exist, the web server may stem the corresponding user search keyword and then search the stemmed user search keyword and the documents / paragraphs stored in the second server. By comparing and analyzing the search results of the morphologically analyzed user search keyword in the document / paragraph result list, and displaying the entity result list mapped to the searched document / paragraph result list on the corresponding user terminal. Search and hybrid based opinion analysis system. A first server for collecting web document data present on the Internet; A data analysis server that receives the web document data collected from the first server, extracts meta information for each object, and analyzes positive / negative statistical information about each object of each object by using the meta data for each object; A second server for indexing positive / negative opinion statistics information for each object of each individual including the meta information for each entity analyzed from the data analysis server to be stored in a database; And Receives a user search keyword transmitted from a user terminal connected through the Internet, and interoperates with the second server to determine whether there is meta information or a target keyword pre-stored in the user search keyword, and meta information prestored in the user search keyword. Or, if the target keyword exists, the web server displaying the result of the object list related to the corresponding meta information or the target keyword on the screen of the corresponding user terminal. Positive / negative opinion statistics information for each object of each object stored in the second server is stored so that each object is mapped on a document / paragraph basis. When the meta information stored in the user search keyword and the target keyword do not exist together, and the entity keyword exists together with the stored meta information keyword, the web server may stem the corresponding user search keyword, and then the corresponding meta information keyword Compare and analyze the stemmed user search keyword excluding the document and paragraphs stored in the second server, search for the stemmed user search keyword except the meta information keyword in the document / paragraph result list, and search the searched document / paragraph. Search for an entity result list mapped to a result list, and then filter the entity result list related to the corresponding meta information keyword from the searched entity result list and display it on the corresponding user terminal. Hybrid based Feedback Analysis System. A first server for collecting web document data present on the Internet; A data analysis server that receives the web document data collected from the first server, extracts meta information for each object, and analyzes positive / negative statistical information about each object of each object by using the meta data for each object; A second server for indexing positive / negative opinion statistics information for each object of each individual including the meta information for each entity analyzed from the data analysis server to be stored in a database; And Receives a user search keyword transmitted from a user terminal connected through the Internet, and interoperates with the second server to determine whether there is meta information or a target keyword pre-stored in the user search keyword, and meta information prestored in the user search keyword. Or, if the target keyword exists, the web server displaying the result of the object list related to the corresponding meta information or the target keyword on the screen of the corresponding user terminal. Positive / negative opinion statistics information for each object of each object stored in the second server is stored so that each object is mapped on a document / paragraph basis. When the meta information and the target keyword which are pre-stored in the user search keyword do not exist together and the entity keyword exists with the pre-stored target keyword, the web server may stem the corresponding user search keyword and then exclude the target keyword. Compare and analyze the stemmed user search keyword and the documents / paragraphs stored in the second server, search for the stemmed user search keyword except the target keyword in the document / paragraph result list, and enter the searched document / paragraph result list. Search for the mapped entity result list, and then search for the entity result list by displaying the searched entity result list in the order of positive or total opinions of the corresponding target keywords and displaying the result list on the corresponding user terminal. And hybrids for this Based on the analyzed system. A first server for collecting web document data present on the Internet; A data analysis server that receives the web document data collected from the first server, extracts meta information for each object, and analyzes positive / negative statistical information about each object of each object by using the meta data for each object; A second server for indexing positive / negative opinion statistics information for each object of each individual including the meta information for each entity analyzed from the data analysis server to be stored in a database; And Receives a user search keyword transmitted from a user terminal connected through the Internet, and interoperates with the second server to determine whether there is meta information or a target keyword pre-stored in the user search keyword, and meta information prestored in the user search keyword. Or, if the target keyword exists, the web server displaying the result of the object list related to the corresponding meta information or the target keyword on the screen of the corresponding user terminal. Positive / negative opinion statistics information for each object of each object stored in the second server is stored so that each object is mapped on a document / paragraph basis. The web server may, if the entity keyword exists together with the meta information and the target keyword pre-stored in the user search keyword, after stemming the corresponding user search keyword, and the stemmed analysis of the user search keyword excluding the meta information and the target keyword; Comparatively analyze the documents / paragraphs stored in the second server, search for the stemmed user search keyword except the meta information and the target keyword in the document / paragraph result list, and the entity result mapped to the searched document / paragraph result list. After searching the list, the entity result list related to the corresponding meta information keyword is filtered in the searched entity result list, and the filtered entity result list is rearranged in the order of positive or total opinions of the target keyword. Display on the corresponding user terminal Using the Internet, wherein the object discovery that hybrid systems based on analyzed for this purpose. The method according to any one of claims 2, 4, 12, 14, 16, 17, 18, or 19, The web server, a hybrid based opinion analysis system for the object search using the Internet characterized in that to display the opinion information of each individual with the object list results on the user terminal. delete (a) collecting web document data residing on the internet; (b) receiving the collected web document data, extracting meta information for each object, and analyzing positive / negative statistical information about each object of each individual using the individual meta information; (c) indexing positive / negative statistical information about each object of each individual including the analyzed individual meta-information to be stored in a database; And (d) determining whether there is a meta information or a target keyword pre-stored in the user search keyword transmitted from the user terminal connected through the Internet, and if the meta information or a target keyword is pre-stored in the user search keyword, the screen of the user terminal Including displaying the result of the object list related to the corresponding meta information or the target keyword, Step (b) is, (b-1) classifying the collected web document data by sector using a preset machine learning model for each sector; (b-2) extracting an opinion sentence by performing language processing on the collected web document data, and dividing the extracted opinion sentence into positive / negative opinion expressions; (b-3) determining which entity corresponds to the collected web document data using the entity-specific meta information; And (b-4) searching for an object using the Internet, characterized in that it comprises a step of determining which object corresponds to the word and part-of-speech information around the opinion sentence extracted in the step (b-2); Hybrid based opinion analysis method for this. 23. The method of claim 22, Step (b-2), Dividing the collected web document data into sentence units, and performing linguistic processing on the separated sentences to extract linguistic features; Classifying opinion / non-computation sentences using linguistic qualities of the extracted sentences; And And a hybrid-based opinion analysis method for object searching using the Internet, comprising the step of dividing the linguistic qualities of the divided opinion sentences into positive / negative opinion expressions. 23. The method of claim 22, In step (b-2), if there is a predefined rule for the extracted opinion sentence, a rule-based model is applied and a positive / negative opinion expression is applied. If there is no predefined rule for a sentence, the object search using the Internet and hybrid-based opinion analysis method for distinguishing a positive / negative opinion expression by applying a preset machine learning model. 23. The method of claim 22, In step (b-2), the rule-based model and the machine learning model are applied to the extracted opinion sentence at the same time to determine positive / negative expression, and then the rule-based model and machine learning Individual search using the Internet and hybrid based opinion analysis method for assigning different reliability scores according to the results of applying the model, and classifying them into positive / negative opinion expressions based on the reliability scores. . 23. The method of claim 22, In step (b-2), it is determined whether the extracted exception sentence is a predefined exception handling rule candidate sentence with respect to the extracted opinion sentence, and the exception is predefined when the extracted opinion sentence is a predefined exception handling rule candidate sentence. Based on the existence of processing rules, applying the exception-based rule-based model or machine learning model to classify them into positive / negative / neutral opinion expressions. Way. The method of claim 26, When the extracted opinion sentence is not a predefined exception handling rule candidate sentence, a rule-based model and a machine learning model that are preset to the extracted opinion sentence are simultaneously applied to determine positive / negative opinion expression. After that, different reliability scores are given according to whether the rule-based model and the machine learning model are matched, and the object search using the Internet is divided into positive / negative opinion expressions based on the reliability scores. And hybrid based opinion analysis method. The method of claim 26, The extracted opinion sentence is a predefined exception handling rule candidate sentence, and if there is a predefined exception handling rule, it is divided into positive / negative / neutral opinion expression by applying a predefined exception handling rule-based model, and extracting the above. If there is no pre-defined exception handling rule, the specified opinion sentence is classified into positive / negative / neutral opinion expression by applying a preset machine learning model. Opinion analysis method. (a) collecting web document data residing on the internet; (b) receiving the collected web document data, extracting meta information for each object, and analyzing positive / negative statistical information about each object of each individual using the individual meta information; (c) indexing positive / negative statistical information about each object of each individual including the analyzed individual meta-information to be stored in a database; And (d) determining whether there is a meta information or a target keyword pre-stored in the user search keyword transmitted from the user terminal connected through the Internet, and if the meta information or a target keyword is pre-stored in the user search keyword, the screen of the user terminal Including displaying the result of the object list related to the corresponding meta information or the target keyword, In step (d), if there is no meta information or a target keyword stored in advance in the user search keyword, the corresponding user search keyword is searched in the database to which the object is mapped and the related entity result list is displayed. Object searching using the Internet and hybrid based opinion analysis method. The method of claim 22 or 29, In the step (d), the entity search using the Internet and hybrid based opinion analysis for displaying the object result list in the order of a lot of positive / negative opinions according to the opinion analysis result for the user search keyword. Way. The method of claim 22 or 29, In step (d), if the meta information and the target keyword pre-stored in the user search keyword exist, the entities related to the meta information keyword are searched, and then the positive / negative or overall opinion of the target keyword is searched for the searched entities. An object search using the Internet, and a hybrid-based opinion analysis method for displaying the object result list in any one order. (a) collecting web document data residing on the internet; (b) receiving the collected web document data, extracting meta information for each object, and analyzing positive / negative statistical information about each object of each individual using the individual meta information; (c) indexing positive / negative statistical information about each object of each individual including the analyzed individual meta-information to be stored in a database; And (d) determining whether there is a meta information or a target keyword pre-stored in the user search keyword transmitted from the user terminal connected through the Internet, and if the meta information or a target keyword is pre-stored in the user search keyword, the screen of the user terminal Including displaying the result of the object list related to the corresponding meta information or the target keyword, In the step (c), affirmative / negative statistical information about each object of each entity is stored so that each entity is mapped on a document / paragraph basis, In step (d), if there is no meta information and a target keyword previously stored in the user search keyword, the corresponding user search keyword is morphologically analyzed, and then the morphologically analyzed user search keyword is compared with the stored documents / paragraphs. Search for the morphologically analyzed user search keyword in the document / paragraph result list, and display the object result list mapped to the searched document / paragraph result list on the corresponding user terminal. Hybrid based opinion analysis method for this. (a) collecting web document data residing on the internet; (b) receiving the collected web document data, extracting meta information for each object, and analyzing positive / negative statistical information about each object of each individual using the individual meta information; (c) indexing positive / negative statistical information about each object of each individual including the analyzed individual meta-information to be stored in a database; And (d) determining whether there is a meta information or a target keyword pre-stored in the user search keyword transmitted from the user terminal connected through the Internet, and if the meta information or a target keyword is pre-stored in the user search keyword, the screen of the user terminal Including displaying the result of the object list related to the corresponding meta information or the target keyword, In the step (c), affirmative / negative statistical information about each object of each entity is stored so that each entity is mapped on a document / paragraph basis. In the step (d), if the meta information and the target keyword which are pre-stored in the user search keyword do not exist together and the entity keyword exists together with the pre-stored meta information keyword, the corresponding user search keyword is stemmed and then the meta Compare and analyze the stemmed user search keyword excluding the information keyword and the stored document / paragraph, and search the stemmed user search keyword excluding the meta information keyword in the document / paragraph result list, and search the searched document / paragraph result list. Search for an entity result list mapped to and then filter the entity result list related to the corresponding meta information keyword from the searched entity result list and display it on the corresponding user terminal. Opinion of Way. (a) collecting web document data residing on the internet; (b) receiving the collected web document data, extracting meta information for each object, and analyzing positive / negative statistical information about each object of each individual using the individual meta information; (c) indexing positive / negative statistical information about each object of each individual including the analyzed individual meta-information to be stored in a database; And (d) determining whether there is a meta information or a target keyword pre-stored in the user search keyword transmitted from the user terminal connected through the Internet, and if the meta information or a target keyword is pre-stored in the user search keyword, the screen of the user terminal Including displaying the result of the object list related to the corresponding meta information or the target keyword, In the step (c), affirmative / negative statistical information about each object of each entity is stored so that each entity is mapped on a document / paragraph basis. In the step (d), if the meta information and the target keyword pre-stored in the user search keyword do not exist together and the entity keyword exists together with the pre-stored target keyword, the corresponding user search keyword is stemmed and then the target keyword Compare and analyze the stemmed user search keyword except for the stored document / paragraph, and search for the stemmed user search keyword except the target keyword in the document / paragraph result list and map the searched document / paragraph result list. After searching the individual result list, the searched individual result list is rearranged in the order of positive or total opinions of the corresponding target keywords and displayed on the corresponding user terminal. Hybrid based for Dogs analysis method. (a) collecting web document data residing on the internet; (b) receiving the collected web document data, extracting meta information for each object, and analyzing positive / negative statistical information about each object of each individual using the individual meta information; (c) indexing positive / negative statistical information about each object of each individual including the analyzed individual meta-information to be stored in a database; And (d) determining whether there is a meta information or a target keyword pre-stored in the user search keyword transmitted from the user terminal connected through the Internet, and if the meta information or a target keyword is pre-stored in the user search keyword, the screen of the user terminal Including displaying the result of the object list related to the corresponding meta information or the target keyword, In the step (c), affirmative / negative statistical information about each object of each entity is stored so that each entity is mapped on a document / paragraph basis. In the step (d), if the entity keyword exists together with the meta information and the target keyword pre-stored in the user search keyword, after the stemming of the corresponding user search keyword, the stemmed analysis of the user search except for the meta information and the target keyword By comparing and analyzing a keyword and the stored document / paragraphs, a search for a stemmed user search keyword excluding corresponding meta information and a target keyword is performed in a document / paragraph result list, and an entity result list mapped to the searched document / paragraph result list. After searching, the entity result list related to the meta information keyword is filtered from the searched entity result list, and the filtered entity result list is rearranged in the order of positive or total opinions of the target keyword, and the corresponding user is rearranged. Display on the terminal That the use of the Internet featuring object retrieval and analysis method based on hybrid feedback for them. 36. The method according to any one of claims 22, 29, 32, 33, 34 or 35, In the step (d), the object search using the Internet and hybrid based opinion analysis method for this, characterized by displaying the opinion information of each individual with the result of the object list on the corresponding user terminal. (a) collecting web document data residing on the internet; (b) receiving the collected web document data, extracting meta information for each object, and analyzing positive / negative statistical information about each object of each individual using the individual meta information; (c) indexing positive / negative statistical information about each object of each individual including the analyzed individual meta-information to be stored in a database; And (d) determining whether there is a meta information or a target keyword pre-stored in the user search keyword transmitted from the user terminal connected through the Internet, and if the meta information or a target keyword is pre-stored in the user search keyword, the screen of the user terminal Including displaying the result of the object list related to the corresponding meta information or the target keyword, In the step (d), when displaying the search result for the corresponding user keyword on the screen of the corresponding user terminal, the individual keyword information and the target keyword statistical information automatically extracted for the individual are either positive, overall or negative. One comment can be arranged in many orders, display the positive / negative value of the target keyword and the symbol indicating it, and then display the positive / negative opinions about the found object left and right, or select the positive / negative tab. Object searching using the Internet and hybrid based opinion analysis method.

Images