KR20240057666A - Method of extraction and summary of categorical sentences in document - Google Patents

Abstract

The present invention relates to a system and method for determining whether a categorical sentence exists in a company or economy-related article or document and extracting and summarizing the corresponding sentence. The categorical sentence in the document according to an embodiment of the present invention The method of extracting and summarizing includes the steps of the collection unit of the system collecting documents containing text and storing them in the system's database, and the preprocessing unit of the system performing preprocessing to remove non-text and common phrases from the collected documents. The following steps are performed: a duplicate removal unit of the system calculates the degree of redundancy of two or more pre-processed documents and removes one document from the database; a learning unit of the system extracts one or more sentences from the document; and a summary unit of the system. It includes a step of converting the extracted sentence into a summary sentence and a step of the report generation unit of the system applying a report format to the converted summary sentence to generate a report.

Description

Method and system for extracting and summarizing categorical sentences in a document {METHOD OF EXTRACTION AND SUMMARY OF CATEGORICAL SENTENCES IN DOCUMENT}

The present invention relates to a system and method for determining whether a categorical sentence exists in an article or document related to a company or economy and extracting and summarizing the corresponding sentence.

Conventionally, sentence extraction methods use statistical techniques or are based on the TextRank algorithm. Statistical techniques extract sentences above a certain threshold by determining the weight of sentences and words through information such as the document title, sentence location, important words, and related sentences.

The TextRank algorithm focuses on calculating the sentence importance by considering the site as a word or sentence in PageRank. This method has the limitation that it is difficult to extract sentences of the desired category because it searches for and extracts the central sentence.

In addition, the method of extracting sentences using a sentence classification model and a threshold has limitations in that the classification results vary depending on the sentence separation method and it is difficult to accurately extract sentences belonging to the corresponding category. In addition, the existing method has the complexity of having to go through the steps of sentence separation, sentence classification, and threshold verification.

Looking at previous research cases on conventional sentence extraction to overcome the limitations of existing sentence classification and extraction methods, the embedding values of each sentence and title are obtained and compared based on a language model, or the score is weighted according to the position of the sentence. A method for extracting central sentences is presented.

However, this method also has the complexity of the process of comparing embedding values, separating sentences, and calculating positions, and is a method of separating sentences, so it is difficult to extract only the exactly necessary sentences.

This specification seeks to provide a method of extracting and summarizing sentences providing information from corporate or economic-related articles or documents.

This specification seeks to provide a method to determine the reliability of the source of a document and to selectively extract and summarize key sentences within the document.

The problems to be solved by the present invention are not limited to the problems mentioned above, and other problems not mentioned here will be clearly understood by those skilled in the art from the description below.

A method of extracting and summarizing categorical sentences in a document according to an embodiment of the present invention includes the steps of the collection unit of the system collecting documents containing text and storing them in the database of the system, and the preprocessing unit of the system collecting non-text sentences from the collected documents. A step of performing preprocessing to remove (Non-Text) and common phrases, a step of the duplicate removal unit of the system calculating the degree of redundancy of two or more preprocessed documents and removing one document from the database, and a step of the learning unit of the system removing the document. It includes extracting one or more sentences from the system, converting the extracted sentences into a summary sentence by the summary unit of the system, and generating a report by applying a report format to the converted summary sentence by the report generation unit of the system.

A system for extracting and summarizing categorical sentences in documents according to an embodiment of the present invention includes a collection unit that collects documents containing text and stores them in the system's database, and non-text and commercial information from the collected documents. A preprocessing unit that performs preprocessing to remove phrases, a duplicate removal unit that calculates the redundancy of two or more preprocessed documents and removes one document from the database, and a learning unit that extracts one or more sentences from the document and the extracted sentences. It includes a summary unit that converts a summary sentence into a summary sentence and a report generation unit that generates a report by applying a report format to the converted summary sentence.

When implementing an embodiment of the present invention, it is possible to extract sentences providing information from corporate or economic-related articles or documents and provide a summary report.

When implementing an embodiment of the present invention, a highly reliable document can be created by providing a summary report by selectively extracting the reliability of the source of the document and key sentences within the document.

The effects provided by the present invention are not limited to the effects mentioned above, and other effects not mentioned herein will be clearly understood by those skilled in the art from the description below.

1 is a diagram showing the configuration of the entire system according to an embodiment of the present invention.
Figure 2 is a diagram showing the operation process of the system according to an embodiment of the present invention.
Figure 3 is a flowchart showing a system for determining whether a categorical sentence exists in a company-related article and extracting and summarizing sentences in a company-related article according to an embodiment of the present invention.
Figure 4 is a detailed flowchart of the newspaper article web crawling step according to an embodiment of the present invention.
Figure 5 is a detailed flowchart of the newspaper article pre-processing step according to an embodiment of the present invention.
Figure 6 is a detailed flowchart of the title embedding-based duplicate data removal step according to an embodiment of the present invention.
Figure 7 is a detailed flowchart of the steps of learning and using a categorical sentence extraction model according to an embodiment of the present invention.
Figure 8 is a detailed flowchart of a sentence summary module according to an embodiment of the present invention.
Figure 9 is a detailed flowchart of the automatic report creation module according to an embodiment of the present invention.
Figure 10 is a diagram showing learning and use of a categorical sentence extraction model according to an embodiment of the present invention.
Figure 11 is a diagram of an automatic report creation module according to an embodiment of the present invention.
Figure 12 is a diagram showing an example of learning data according to an embodiment of the present invention.
Figure 13 is a diagram showing the creation of a new category of learning data according to an embodiment of the present invention.
Figure 14 is an overall configuration diagram of the sentence extraction and summary system for determining the presence of categorical sentences in company-related text and including the main embodiment of the invention described in the present invention.

The advantages and features of the present invention and methods for achieving them will become clear by referring to the embodiments described in detail below along with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below and will be implemented in various different forms. The present embodiments only serve to ensure that the disclosure of the present invention is complete and are within the scope of common knowledge in the technical field to which the present invention pertains. It is provided to fully inform those who have the scope of the invention, and the present invention is only defined by the scope of the claims. Like reference numerals refer to like elements throughout the specification.

For a detailed description of the invention, reference is made to the accompanying drawings, which show by way of example specific embodiments in which the invention may be practiced. These embodiments are described in sufficient detail to enable those skilled in the art to practice the invention. It should be understood that the various embodiments of the invention are different from one another but are not necessarily mutually exclusive. For example, specific shapes, structures and characteristics described herein may be implemented in one embodiment or another without departing from the spirit and scope of the invention. Additionally, it should be understood that the location or arrangement of individual components within each disclosed embodiment may be changed without departing from the spirit and scope of the invention.

Accordingly, the detailed description that follows is not intended to be taken in a limiting sense, and the scope of the invention is limited only by the appended claims, together with all equivalents to what those claims assert, if properly described. Similar reference numbers in the drawings refer to identical or similar functions across various aspects.

Hereinafter, this specification explains the method and system for extracting sentences describing the company's stock price and general status/prospects and the status/prospects of the industry to which the company belongs from online newspaper article texts, other reliable data, SNS, etc. .

The present invention overcomes the limitations of the conventional PageRank centrality-based core sentence extraction method and additionally applies a methodology that introduces a machine reading model learning method for a categorical sentence extraction system based on financial texts such as online newspaper articles and reports. Understand financial texts in depth and build an artificial intelligence extraction summary system that applies text extraction technology and summary models. Through this, sentences of the necessary categories are extracted from company-related online texts and a report is generated.

Hereinafter, this specification presents a method and system for extracting and summarizing sentences by determining whether a categorical sentence exists in a document related to a company or a specific field of the economy.

Figure 1 is a diagram showing the configuration of a system for extracting and summarizing sentences from information related to a company according to an embodiment of the present invention.

The system 1000 includes a collection unit 1100, a preprocessing unit 1200, a duplicate removal unit 1300, a learning unit 1400, a summary unit 1500, a report generation unit 1600, and a database 1700. .

The collection unit 1100 collects articles and writings containing sentences from newspaper articles, company-related papers or materials, and other major influencers' SNS (Social Network Service) or blogs. At this time, the collection unit 1100 may use the source information to assign a value for the reliability of the information to the collected information. The collection unit 1100 collects documents including text and stores them in the system database 1700.

The preprocessing unit 1200 performs preprocessing to remove non-text and common phrases from the documents collected by the collection unit 1100. In this process, the preprocessing unit 1200 uses the characteristics of HTML code to convert sentences containing words that are emphasized in bold, adjust font size, or set in italics to main sentences while performing preprocessing work to remove HTML code, etc. Weights can be set for processing.

The duplicate removal unit 1300 calculates the degree of duplication of two or more pre-processed documents and removes one document from the database. At this time, the reliability value of each document can be reflected in the duplicate removal process. For example, the duplicate removal unit 1300 uses the source information added by the collection unit 1100 during the document collection process to maintain documents with highly reliable source information among duplicate documents in the database 1700.

The learning unit 1400 extracts one or more sentences from documents in the database from which duplicates have been removed. One or more sentences corresponding to a document can be extracted by applying learning data, dictionary language model, and categorical sentence extraction model.

The summary unit 1500 may convert the sentences extracted from the learning unit 1400 into summary sentences.

The report generator 1600 may generate a report by applying a report format to the converted summary sentence. The report format includes not only the file format that stores the report (Excel, Word, PowerPoint, Hangul, etc.), but also XML format. The report may include the name of a specific company, the content of the sentence about it, information on whether the sentence is positive or negative about the company, source information, reliability, etc.

Figure 2 is a diagram showing the process of extracting sentences and generating a report according to an embodiment of the present invention.

The entire process includes the information collection step (S10), the collected information pre-processing step (S20), the duplicate data removal and reliability check step (S30), the learning step for sentence extraction (S40), the sentence summary step (S50), and the automatic report creation step. It consists of (S60).

Documents collected in the information collection step (S10) may include one or more of newspaper articles, SNS (Social Network Service) articles, and newsletter articles. Of course, the document of the present invention is not limited to the above example, and includes all documents that can be web crawled on the Internet, newsletters delivered through email, etc.

The collection unit 1100 may perform web crawling to collect information. In addition, the collection unit 1100 can subscribe to various newsletters related to companies or the economy and collect newsletters provided by email, etc. That is, the collection unit 1100 may collect documents by performing web crawling or collect documents received by a pre-registered email address.

The duplicate data removal and reliability check step (S30) includes when a large number of similar articles or contents are collected, the duplicate removal unit 1300 selects only articles or documents from highly reliable sources among them. Additionally, for this purpose, the duplicate removal unit 1300 can check the reliability of the source or document.

Afterwards, in the learning process for sentence extraction, existing learning data is used or new learning is performed, and then one or more sentences corresponding to the document are extracted. And the summary unit 1500 summarizes the sentences. In this process, the learning unit 1400 or the summarizing unit 1500 may divide one sentence into two or more sentences or combine two or more sentences into one.

As shown in Figure 2, the system 1000 provides information summarized in one or multiple sentences about a document through web crawling of newspaper articles, pre-processing of newspaper articles, removal of duplicate data based on title embedding, and learning and use of a categorical sentence extraction model. can be created. Additionally, through the sentence summary process, the system 1000 can automatically create a report and generate a report as shown in FIG. 11.

Hereinafter, with reference to the attached drawings, a summary system for determining the presence or absence of a categorical sentence in a corporate or economic-related text and extracting a sentence in a text according to an embodiment of the present invention will be described.

Figure 3 is a diagram showing the process of collecting newspaper articles and extracting sentences according to an embodiment of the present invention. The embodiment of FIG. 3 can be applied to a method of extracting sentences from other documents 15 whose reliability has been verified, such as SNS and papers, in addition to newspaper articles. That is, the embodiments of FIGS. 3 to 9 can be applied to various documents accessible on the Internet in addition to newspaper articles.

According to the embodiment of the present invention shown in Figure 3, the system 1000 includes a newspaper article web crawling step (S100), a newspaper article pre-processing step (S200), a title embedding-based duplicate data removal step (S300), and categorical sentence extraction. A model learning and use step (S400), a sentence summary step (S500), and an automatic report creation step (S600) are performed.

Referring to FIG. 3, the system 1000 uses the embeddings of the dictionary language model 35 when extracting newspaper article title embeddings to remove duplicate articles with similar titles and when transfer learning a categorical sentence extraction model. You can. The preprocessed newspaper article text is also embedded through the corresponding dictionary language model and used as input information in the learned categorical sentence extraction model to extract categorical sentences (45).

When learning a categorical sentence extraction model, the system 1000 can utilize a machine reading learning method that learns the positions of the first and last tokens of categorical sentences. The extracted categorical sentences are filtered according to length, and if they are longer than a certain length, they are summarized through a sentence generation summary model, and a report (65) is automatically generated based on the results.

In one embodiment, the report is generated in the form of an Excel file, but the present invention is not limited to this, and the system 1000 can generate a report in document formats such as MS-Word and Hangul, XML format, etc. there is.

Figure 4 is a detailed flowchart of a specific method of performing the newspaper article web crawling step according to an embodiment of the present invention.

Referring to the embodiment of FIG. 4, the detailed web crawling step (S100) may include a company-related article list crawling step (S110), a date-based filtering step (S120), and an article content crawling step (S130).

The collection unit 1100 of the system 1000 crawls a list of articles related to the company, excluding the article content. In addition, by filtering only the article list that meets the limited conditions of the period, such as the requested date or a specific range of dates from the extracted article list, and crawling the article contents, it is possible to crawl only the article data of the required period more efficiently.

The collection unit 1100 may refer to the company item code 17 to confirm the company name.

In addition, in the process of collecting documents, the collection unit 1100 can also collect articles uploaded to the newsletter list of services that provide newsletters, influencers' SNS or blogs, and homepages, in addition to the article list. . Information about media companies, influencers, newsletter services, etc. that are the subject of information to be collected by the collection unit 1100 may be set in advance in the system 1000.

The system 1000 or the collection unit 1100 may maintain or change reliability information about media outlets, influencers, and newsletter services that are the target of information to be collected. Additionally, the collection unit 1100 may not collect specific articles or newsletters depending on reliability if a lot of information about the company is collected repeatedly during a specific period.

In one embodiment, the collection unit 1100 may add a value for the reliability of the document to the reliability field of the document using the source information of the document and store it in the database 1700. The original link address and document content of the document are stored in the database 1700, and a separate field can be created to add a value for the reliability of the document.

As an example, assume that the reliability value is set from 0 to 9. The collection unit 1100 may set a reliability value according to the media company in the case of newspaper articles among documents collected within a certain period of time. The collection unit 1100 can set the reliability to 8 or 9 for articles from major media companies. For articles from new media outlets or those that receive and upload articles from other media outlets, the reliability can be set to 3 or 4, etc. The above number can be set in various ways depending on the system (1000).

Likewise, in the case of an influencer, the collection unit 1100 may set the reliability of the influencer's SNS post or blog post to 5 to 7, etc., depending on the past or present reliability or reputation of the information provided by the influencer. Even in the case of newsletters, the collection unit 1100 can set the reliability to 7 to 9 for newsletter articles with many subscribers or reputation in the industry.

To establish such reliability, the collection unit 1100 can assign a separate reliability rating to the source of the document, and the collection unit 1100 can determine the reliability of the document according to the reliability level of the source.

Figure 5 is a detailed flowchart of a specific method of performing the newspaper article pre-processing step according to an embodiment of the present invention.

According to one embodiment of FIG. 5, the newspaper article pre-processing step (S200) may include an HTML entity code removal step (S210), an unnecessary boilerplate removal step (S220), and an article title-based duplicate removal step (S230).

According to an exemplary embodiment, the pre-processing unit 1200 performs a pre-processing process for the HTML entity code literal expression included in the crawled newspaper article text in the HTML entity code removal step (S210). An entity code is a code used to input special characters in an HTML document.

In the unnecessary boilerplate removal step (S220), the preprocessor 1200 removes boilerplate text that is repeatedly added and is not related to the article content. Unnecessary boilerplate includes media outlets, reporter names, and advertisement parts, and by removing them, the input of the categorical sentence extraction model can be refined and extraction accuracy can be improved.

After the boilerplate removal step, the preprocessor 1200 may perform the article title duplicate removal step (S230), and all duplicate articles are removed based on the title text. In other words, information duplication can be prevented by excluding documents with the same article title.

The system 1000 or the preprocessor 1200 may remove documents with low reliability by reflecting reliability information about media companies, influencers, and newsletter services in the process of removing duplicate information.

The preprocessor 1200 can distinguish between main sentences and non-main sentences using HTML code attributes in the process of removing HTML entity codes to distinguish text information collected from the web from non-text information. For example, the preprocessor 1200 may increase the size of letters within a sentence or, in the case of a boldface sentence, may increase the weight of the sentence. The weight of the sentence can be used in the process of extracting the sentence in S400 or in the process of summarizing the sentence in S500.

The pre-processing unit 1200 can identify and remove HTML code within the document in order to remove non-text information or advertising information that does not contain information from the collected document 19, for example, corporate articles collected through web crawling. there is. At this time, if the HTML code is an italic code, bold code, or font size change code that emphasizes a specific sentence or paragraph, the preprocessor 1200 can identify words included at the beginning and end of these codes or sentences containing these words. You can.

In the case of boldface, it starts with <b> and ends with </b>. For italics, it starts with <i> and ends with </i>. In addition, codes such as underline (<u> start, </u> end), title (<h1>, <h2>, etc.), paragraph (<p>), and text italics (<em>) also affect the weight of the sentence. affects.

Therefore, the step (S210) in which the preprocessor 1200 identifies and removes the HTML code within the document is a step of identifying the code that controls the font among the HTML codes and setting the weight of the sentence included in the start and end points of the code. Includes.

For example, if the 5th sentence of the document is highlighted with a bold code in the HTML code, the preprocessor 1200 inserts pre-arranged symbols (e.g., square brackets, etc.) before and after the 5th sentence or Enter weight-related information such as “5-bold” or “5th = important” so that the sentence can be treated as important when extracting or summarizing sentences from the document later.

Then, the pre-processing unit 1200 removes advertising links or boilerplate text from the document (S220) and removes duplicate sentences or phrases from the document (S230).

More specifically, the pre-processing unit 1200 maintains pattern information about external advertising links and boilerplate text within the document and removes them from the document to prevent unnecessary information from being included in the subsequent sentence extraction process.

Figure 6 is a detailed flowchart of a specific method of performing the title embedding-based duplicate data removal step according to an embodiment of the present invention.

Applying the embodiment of FIG. 6, the duplicate data removal step (S300) may include an article title embedding extraction step (S310), an embedding cosine similarity comparison step (S320), and a similar embedding article removal step (S330).

According to an exemplary embodiment, the duplicate removal unit 1300 extracts embedding information about the article title based on a dictionary language model in the article title embedding extraction step (S310). In this embodiment, the duplicate removal unit 1300 retrieves embedding information about the article title using a dictionary language model.

According to an exemplary embodiment, the duplicate removal unit 1300 compares cosine similarity between the embedding information extracted in the embedding cosine similarity comparison step (S320). In the similar embedding article removal step (S330), the duplicate removal unit 1300 determines that articles with titles with similar embeddings are duplicate articles and removes them.

For example, the duplicate removal unit 1300 extracts a title sentence corresponding to a title from a document stored in the database 1700 and extracts a title embedding by applying the dictionary language model 35 to the extracted title sentence (S310 ). Additionally, the duplicate removal unit 1300 may compare the similarity of title embeddings of documents stored in the database 1700 (S320) and remove one document from the database (S330).

In one embodiment, when there are two or more articles with similar embedding titles, the duplicate removal unit 1300 may maintain articles from highly reliable media outlets and remove articles from non-reliable media outlets from the database 1700. That is, if the title embeddings of the first document and the second document are similar, and the reliability of the first document is higher than the reliability of the second document, the duplicate removal unit 1300 may remove the second document from the database 1700. there is.

Of course, the duplicate removal unit 1300 can remove articles judged to be duplicates not only for articles but also for newsletter articles or articles written by influencers.

In one embodiment, the collection unit 1100 may calculate the reliability value of the document using the document source information and enter the reliability value into the document reliability field. The duplicate removal unit 1300 can use this value to delete documents from unreliable sources among duplicate documents from the database 1700. However, if there are many similar documents, only one document is maintained, and the number of documents deleted based on the standard of identical content can be separately recorded in the field related to the document.

For example, if 10 articles were collected under the title “Company A’s free capital increase,” and only the articles from the most reliable media outlets were kept, the duplicate removal unit 1300 would remove the same (or similar) content for the documents. The information that there were 9 articles from different sources can be stored in a field related to the document in the database, and the information that there were 9 other articles with similar content can be included in the report when generating the report later. That is, the duplicate remover 1300 can calculate the cosine similarity between embeddings and remove articles with semantically similar titles.

Figure 7 is a detailed flowchart showing a specific method of performing the steps of learning and using a categorical sentence extraction model according to an embodiment of the present invention.

Referring to Figure 7, the step of learning and using the categorical sentence extraction model (S400) includes the step of automatically generating training data for non-existence of categorical sentences (no answer) (S410), the step of input embedding (S420), and the step of learning the machine reading model. (S430) and a categorical sentence extraction step (S440). In more detail, the learning unit 1400 automatically generates additional learning data for cases where categorical sentences do not exist through the input learning data, then embeds the learning data to learn the machine reading model, and learns from new newspaper articles. Categorical sentences can be searched and extracted. Please refer to the examples of Figures 12 and 13, which will be described later.

According to an exemplary embodiment, the learning unit 1400 determines the text existing in the learning data 32, the category of the categorical sentence existing in the text, and the sentence After checking the location, an additional learning set is automatically created with the category and sentence location of categorical sentences that are not in the text as 0.

Through this, when the learning unit 1400 creates data by tagging only categorical sentences that exist in the text when generating the learning data 32, a learning set is created that can learn that there are no categories that are not in the text through the corresponding step later. It is automatically generated, allowing you to create a learning set efficiently.

According to an exemplary embodiment, the learning unit 1400 converts the training data 32 and the article input for future prediction into an embedding value based on the dictionary language model 35 in the input embedding step (S420). This makes it possible to generate input information used in the machine reading model.

According to an exemplary embodiment, the learning unit 1400 learns the embedding value (generated in the previous input embedding step (S420)) into the model in the machine reading model learning step (S430). The information learned at this time is the input category, the embedding value of the entire text, and the start and end positions of the category sentence.

According to an exemplary embodiment, the learning unit 1400 receives a document (e.g., a newspaper article) stored in the database 1700 in the categorical sentence extraction step (S440) and embeds it as an input along with each category to create a vector. You can. The learning unit 1400 may output the position of the categorical sentence through the model built in the machine reading model learning step (S430). The learning unit 1400 may extract a categorical sentence by applying the categorical sentence extraction model 41 in S430.

Additionally, the learning unit 1400 may extract key sentences using the weight of the sentence added by the preprocessor 1200 using HTML code attributes in the process of removing the HTML entity code. That is, the learning unit 1400 preferentially extracts sentences with high weight.

In summary, the learning unit 1400 applies the dictionary language model 35 to extract one or more sentences 45 corresponding to the document.

Figure 8 is a detailed flowchart of a specific method of performing sentence summarization according to an embodiment of the present invention.

According to one embodiment, the sentence summarizing step (S500) may include the extracted sentence length checking step (S510), the sentence filtering step above the length threshold (S520), and the sentence summarizing step (S530).

In the extracted sentence length confirmation step (S510), the summary unit 1500 performs a process of checking the length of the sentence 45 extracted in the previous categorical sentence extraction step (S440).

According to an exemplary embodiment, in the step of filtering sentences above the length threshold (S520), the summary unit 1500 performs the task of filtering sentences above a certain threshold through the length of the sentence confirmed in the previous sentence length confirmation step (S510). .

In this process, the summary unit 1500 forms one sentence, but if it is in the form of a complex sentence, it can be separated into simple sentences. For example, the sentences before and after words such as “~but”, “~ did,” and “~but,” can be separated into two sentences. The summary unit 1500 can convert a complex sentence into two or more simple sentences to make the meaning of each sentence clear.

According to an exemplary embodiment, in the sentence summary step (S530), the summary unit 1500 performs the task of summarizing the sentences filtered in the previous length threshold or more sentence filtering step (S520) by applying the generation summary model 52. do.

Additionally, the summarizing unit 1500 may summarize sentences using the weight of sentences added by the pre-processing unit 1200 using HTML code attributes during the process of removing HTML entity codes.

In summary, the summary unit 1500 merges two or more sentences extracted from the learning unit 1400 and converts them into one or more summary sentences 55.

The summary unit 1500 checks the length of the extracted categorical sentences, filters sentences with a length exceeding a certain threshold, and outputs a summary of the sentences through a generative summary model.

Figure 9 is a detailed flowchart of a specific method of performing the automatic report creation step according to an embodiment of the present invention. The report generator 1600 converts the summary sentence 55 of FIG. 8 into a preset report format and generates a report.

The automatic report creation step (S600) may include a basic information creation step (S610), an extracted summary sentence creation step (S620), and a report style application step (S630).

According to an exemplary embodiment, the report generation unit 1600 extracts the current standard company market capitalization, market capitalization ranking, number of listed stocks, PER, same industry PER, and 52-week high/lowest price in the basic information creation step (S610). and carry out the writing process.

According to an exemplary embodiment, the report generator 1600 performs a process of creating the result of the sentence summary module (S500) in the extracted summary sentence creation step (S620).

According to an exemplary embodiment, the report generation unit 1600 performs a process of applying a style to previously written content in a report format in the report style application step (S630).

The report format includes not only the file format that stores the report (Excel, Word, PowerPoint, Hangul, etc.), but also XML format. The report may include the name of a specific company, the content of the sentence about it, information on whether the sentence is positive or negative about the company, source information, reliability, etc.

Additionally, when duplicate documents are deleted, the report generation unit 1600 also includes information on the number of articles providing the same or similar content in the report so that the importance of the content can be confirmed.

The report generation unit 1600 can create the basic financial information of the company, then write the extracted categorical sentences and apply the style in the form of a report.

The present invention can be applied not only to companies but also to specific industrial fields.

Figure 10 is a diagram showing input and output in the process of learning and using a machine reading model according to an embodiment of the present invention.

The input 401 and output 402 used when learning the machine reading model according to the embodiment in the S400 process are as shown in FIG. 10. It shows the type of output information of the machine reading model. The parts highlighted in squares in 401 and 402 indicate labeling for each sentence during the input and output process.

Each labeling can be added as a category in a sentence as follows:

“Corporate Status/Forecast”, “Industry Status/Forecast”, “Stock Status/Forecast”, “Stock Price Forecast”, “Sales Status”, and “Sales Forecast” indicate the categories added to each sentence in the document. 401 and 402 in Figure 10 are an example of the output (output result) of the model. Looking at this, if there is no sentence belonging to the corresponding category in the document, no separate content is included, but if there is corresponding content, the corresponding sentence is included in the document. It includes values for location and goodness of fit for categories. For example, see the following part of 401:

“'Sales status': ('It was analyzed that NCSoft recorded sales of 729.3 billion won and operating profit of 185.7 billion won in the first quarter.', '159~215', 49.480000000000004)”

It belongs to the sales status category, and the starting and ending points of the sentence are indicated by positions (159 to 215) in the document. The last number indicates the score associated with the labeling.

Figure 11 is a diagram showing a report generated by an embodiment of the present invention. Previously, in the automatic report creation step (S600), a report automatically generated according to the embodiment is shown.

601 is a summary section that explains the overall outline of the company. The summary section contains summary information such as the stock price of the company called “A_Company”.

602 is a section that extracts and lists sentences corresponding to each category from various articles related to companies. Sentences corresponding to one or more categories can be extracted from one article. The report lists the source of the article containing each sentence, and users who want to check the contents of the summary sentence can select the link.

Figure 12 is a diagram showing an example of learning data according to an embodiment of the present invention. The items that make up the learning data include the "context" item indicating the content, the "title" item indicating the title, the "url" item indicating the website address, and the "raw_html" item indicating the content of the web document.

Additionally, the "qas" entry can be configured to include key sentences. Additionally, “answer_start” and “html_answer_start”, which indicate the starting point of the corresponding sentence, can be used to identify the corresponding sentence in the later learning process.

In addition, “question” indicates the category for the document, and consists of “corporate status/prospect”, “industry status/prospect”, “stock price status”, “stock price outlook”, sales status”, “sales outlook”, etc. It can be.

And in the learning data, the category is “Corporate Status/Forecast,” but the system 1000 can automatically generate learning data in response to other categories other than the corresponding category. See Figure 13.

Figure 13 is a diagram showing the creation of a new category of learning data according to an embodiment of the present invention. It is the same as in Figure 12 above, with only the "question" item being set as a new category. In Figure 13, a category such as “stock price status” may be set, or a category such as “sales status” may be set. This is an example in which the system 1000 adds a new category in response to a category called corporate status/forecast as shown in FIG. 12 as learning data.

Model learning can be performed on the automatically generated training data shown in FIG. 13 through an input embedding step.

Figure 14 is a diagram showing the overall configuration of a system for determining the presence of categorical sentences in company-related text and extracting and summarizing sentences according to a main embodiment of the present invention. Please refer to Figures 1 and 2.

When the system and method of the present invention are applied, the limitations of conventional sentence extraction research can be complemented.

In an embodiment of the present invention, in the title embedding-based duplicate data removal step, the embedding of the article title is used to remove duplicate articles by reflecting the characteristics of newspaper articles in which articles with similar content are generally published with similar titles. Similarity can be used. As a result, even if the text does not match, newspaper articles with semantically similar titles can be removed to prevent duplication.

Additionally, embodiments of the present invention can apply a method of building a model that predicts the location of sentences for a specific category within a newspaper article by applying a supervised learning-based machine reading model. Therefore, the performance of the model can be further improved by introducing a learning procedure for cases where sentences for a specific category do not exist.

Additionally, in order to improve readability, if the length of the extracted sentence is longer than a certain length, a method of summarizing the sentence through a generative summary model can be applied. When implementing the system and method of the present invention, summarized sentences can be automatically provided as reports in various file formats such as Excel, Word, Hangul, PowerPoint, and PDF.

Methods according to embodiments of the present invention may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. Computer-readable media may include program instructions, data files, data structures, etc., singly or in combination. Program instructions recorded on the medium may be those specifically designed and constructed for the present invention, or may be known and available to those skilled in the art of computer software.

As described above, although the present invention has been described using limited embodiments and drawings, the present invention is not limited to the above embodiments, and various modifications and variations can be made from these descriptions by those skilled in the art. This is possible. The above-described embodiments should be understood in all respects as illustrative and not restrictive, and the scope of the present invention will be indicated by the claims to be described later rather than the detailed description given above. In addition, the meaning and scope of this patent claim, as well as all changes and modifications derived from the equivalent concept, should be construed as being included in the scope of the present invention.

Therefore, the scope of the present invention should not be limited to the described embodiments, but should be determined by the claims described later and equivalents to these claims.

1000: System 1100: Collection unit
1200: Preprocessing unit 1300: Duplicate removal unit
1400: Study section 1500: Summary section
1600: Report generation unit 1700: Database

Claims (20)

A collection unit of the system collects documents containing text and stores them in the database of the system;
A preprocessing unit of the system performing preprocessing to remove non-text and common phrases from the collected documents;
a step of the duplicate removal unit of the system calculating the degree of duplication of two or more documents on which the preprocessing has been performed and removing one document from the database;
A learning unit of the system extracting one or more sentences from the document;
A summary unit of the system converting the extracted sentence into a summary sentence; and
A method of extracting and summarizing categorical sentences in a document, comprising the step of generating a report by applying a report format to the converted summary sentence by the report generating unit of the system.
According to paragraph 1,
The above documents include one or more of newspaper articles, SNS (Social Network Service) articles, and newsletter articles,
The storing step includes the collection unit performing web crawling to collect the document or collecting the document received by a pre-registered email address.
According to paragraph 1,
The saving step is
A method of extracting and summarizing categorical sentences in a document, including the step of the collection unit adding a value for the reliability of the document to a reliability field of the document using source information of the document and storing it.
According to paragraph 1,
The step of performing the preprocessing is
The preprocessor identifying and removing HTML code from the document;
The preprocessor removing advertising links or boilerplate text from the document; and
A method of extracting and summarizing categorical sentences in a document, wherein the preprocessor includes removing duplicate sentences or phrases in the document.
According to clause 4,
The step of the preprocessor identifying and removing HTML code within the document is
The preprocessor identifying a code that controls a font among the HTML codes; and
A method of extracting and summarizing categorical sentences in a document, comprising setting weights of sentences included in the start and end points of the code.
According to paragraph 1,
The removal step is
The duplicate removal unit extracts a title sentence corresponding to a title from the document;
The duplicate removal unit extracts a title embedding by applying a dictionary language model to the extracted title sentence; and
The method of extracting and summarizing categorical sentences in a document, wherein the duplicate removal unit compares similarities between title embeddings of documents stored in the database and removes one document from the database.
According to clause 6,
The duplicate removal unit includes the step of removing the second document from the database when the title embeddings of the first document and the second document are similar and the reliability of the first document is higher than the reliability of the second document, A method to extract and summarize categorical sentences in a document.
According to paragraph 1,
The extraction step is
A method of extracting and summarizing categorical sentences in a document, including the step of the learning unit extracting one or more sentences corresponding to the document by applying a dictionary language model.
According to paragraph 1,
The conversion step is
A method of extracting and summarizing categorical sentences in a document, wherein the summary unit merges the two or more extracted sentences and converts them into a summary sentence.
According to paragraph 1,
The steps for generating the report are
A method of extracting and summarizing categorical sentences in a document, including the step of the report generating unit converting the summary sentences into a preset report format to generate a report.
a collection unit that collects documents containing text and stores them in a database of the system;
a pre-processing unit that performs pre-processing to remove non-text and commercial phrases from the collected documents;
a duplicate removal unit that calculates the degree of duplication of two or more pre-processed documents and removes one document from the database;
a learning unit that extracts one or more sentences from the document;
a summary unit converting the extracted sentence into a summary sentence; and
A system for extracting and summarizing categorical sentences in a document, including a report generator that generates a report by applying a report format to the converted summary sentences.
According to clause 11,
The above documents include one or more of newspaper articles, SNS (Social Network Service) articles, and newsletter articles,
A system for extracting and summarizing categorical sentences in a document, wherein the collection unit performs web crawling to collect the document or collects the document received by a pre-registered email address.
According to clause 11,
A system for extracting and summarizing categorical sentences in a document, wherein the collection unit uses the source information of the document to add and store a value for the reliability of the document in a reliability field of the document.
According to clause 11,
The preprocessor identifies and removes HTML code within the document and removes advertising links or boilerplate text from the document,
A system for extracting and summarizing categorical sentences in a document, wherein the preprocessor removes duplicate sentences or phrases in the document.
According to clause 14,
In order for the preprocessor to identify and remove HTML code within the document,
A system for extracting and summarizing categorical sentences in a document, wherein the preprocessor identifies the code that controls the font among the HTML codes and sets the weight of the sentence included in the start and end points of the code.
According to clause 11,
The duplicate removal unit extracts a title sentence corresponding to the title from the document, applies a dictionary language model to the extracted title sentence, and extracts a title embedding,
The duplicate removal unit compares the similarity of title embeddings of documents stored in the database and removes one document from the database. A system for extracting and summarizing categorical sentences in a document.
According to clause 16,
The duplicate removal unit removes the second document from the database when the title embeddings of the first document and the second document are similar and the reliability of the first document is higher than the reliability of the second document. A system for extracting and summarizing sentences.
According to clause 11,
A system for extracting and summarizing categorical sentences in a document, wherein the learning unit extracts one or more sentences corresponding to the document by applying a dictionary language model.
According to clause 11,
A system for extracting and summarizing categorical sentences in a document, wherein the summary unit merges the two or more extracted sentences and converts them into a summary sentence.
According to clause 11,
A system for extracting and summarizing categorical sentences in a document, wherein the report generator generates a report by converting the summary sentences into a preset report format.

Images