An extraction method for medical industry news events in the financial field

Abstract

This invention provides a method for extracting news events in the medical industry within the financial sector, relating to the field of information intelligence technology. The method uses web crawlers to collect publicly available medical news as a data source, manually summarizes templates for medical news event types in the financial sector and defines keywords, obtains a dataset of statements to be extracted using keyword localization technology, performs feature learning on the statements in the dataset, and utilizes a fusion pattern matching and convolutional neural network approach to extract events. This method, through the above steps, completes the extraction of news events in the medical industry within the financial sector, helping relevant practitioners analyze and predict the development trends of medical companies and related pharmaceutical products in the financial market.

Description

Technical Field

[0001] This invention relates to the field of information intelligence technology, and in particular to a method for extracting news events in the medical industry within the financial sector. Background Technology

[0002] With the rapid development of machine learning and artificial intelligence, all industries are trending towards and aiming for intelligentization. Practitioners in various industries also hope to use artificial intelligence technology to achieve more efficient information processing. The rapid extraction and analysis of information has become an urgent need for practitioners in the field.

[0003] An event refers to an objective fact of the interaction between specific people and things at a specific time and place; it is a dynamically changing form of information. In the face of massive amounts of information, to enable machines to possess functions such as "reasoning" and "prediction," it is necessary to discover the influence logic between various factors within an industry from among events, in order to achieve better business applications, especially in the pharmaceutical industry within the financial sector. Under the current impact of the pandemic, public health measures, vaccine-related biological products, and therapeutic drugs have become key areas of focus. To help industry practitioners grasp industry dynamics in advance, track industry development trends, and capture opportunities from the massive and constantly changing information center, the rapid acquisition and analysis of industry news events is essential. Therefore, an efficient and accurate method for extracting news events specific to the medical industry within the financial sector is needed.

[0004] Event extraction is a task that involves identifying and extracting key elements of an event relevant to a target from semi-structured or unstructured data. It comprises two core subtasks: event detection and type identification, and event argument extraction. Currently, typical event extraction methods fall into two categories: pattern matching-based and machine learning-based methods. Pattern matching-based methods extract events guided by patterns, achieving good results in specific domains. However, because the extraction process involves finding information within sentences that conforms to pattern constraints, it is relatively inflexible. Machine learning-based event extraction adopts the idea of ​​text classification, treating the trigger word matching and argument extraction subtasks of event extraction as two separate classification problems.

[0005] However, due to the limited training data in specific sectors such as finance and healthcare, the accuracy of the results extracted by this method cannot be guaranteed. Furthermore, not every sentence in a news article contains an event, and irrelevant sentences will reduce the efficiency of event extraction. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to address the shortcomings of the prior art by providing a method for extracting news events in the medical industry within the financial field. This method is an event extraction method that integrates pattern matching and convolutional neural networks, thus making up for the lack of methods for extracting news events in the medical industry within the financial field and ensuring the efficiency and accuracy of news event extraction in this specific field.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0008] A method for extracting news events in the medical industry within the financial sector is proposed. This method uses web crawlers to collect publicly available medical news as a data source, manually summarizes templates for medical news event types in the financial sector and defines keywords, obtains a dataset of sentences to be extracted through keyword localization technology, performs feature learning on the sentences in the dataset, and uses a fusion pattern matching and convolutional neural network method to extract events.

[0009] The specific steps of this method are as follows:

[0010] S1: Obtain publicly available medical news texts from websites using web crawling technology;

[0011] S2: By manually summarizing and categorizing templates for medical news event types and manually defining keywords, a keyword set is obtained;

[0012] S3: Constructs the keyword regular expression used for pattern matching, for keyword location;

[0013] S4: Using keyword positioning technology, locate sentences in medical news containing keywords to obtain a dataset of sentences to be extracted;

[0014] S5: Use Word2vec to learn features from the sentences in the dataset to be extracted, and learn word-level feature vector representations respectively. and word-level feature vector representation ;

[0015] S6: Since the word classification task and the argument extraction task have different requirements for character vectors and word vectors, different weights are assigned to character vectors and word vectors for different tasks to perform feature fusion, resulting in a feature representation CF with larger character weights and a feature representation WF with larger word weights.

[0016] S7: Use the feature representation CF with high word weights to perform trigger word matching. This process uses a dynamic multi-pooling convolutional neural network. If there is no trigger word, return to step S6 to select the next sentence for trigger word matching; if there is a trigger word, use pattern matching to optimize the matched trigger word.

[0017] S8: Use the word weighted feature representation WF to perform the next argument extraction task. This process uses a dynamic multi-pooling convolutional neural network. After extracting arguments, pattern matching is used to optimize the extraction results.

[0018] The beneficial effects of adopting the above technical solution are as follows: The method for extracting news events in the medical industry within the financial sector provided by this invention aims to extract events from medical industry news that will have an impact on the financial sector. Firstly, keyword positioning technology is applied to the news text to retain valuable sentences to the maximum extent, reducing the impact of irrelevant information on the extraction results and improving extraction efficiency. Secondly, targeted extraction is performed using manually summarized event type templates for the medical industry within the financial sector. Furthermore, by fusing pattern matching and convolutional neural network extraction methods, regular expressions are used to optimize and limit the extraction results, thereby improving the accuracy of event extraction in specific domains. Attached Figure Description

[0019] Figure 1 A flowchart illustrating a method for extracting news events in the medical industry within the financial sector, provided in an embodiment of the present invention.

[0020] Figure 2 This is a flowchart of an event extraction method provided in an embodiment of the present invention. Detailed Implementation

[0021] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0022] like Figure 1 The diagram shows a flowchart of a method for extracting news events in the medical industry within the financial sector. The method described in this embodiment is as follows.

[0023] S1: Obtain publicly available medical news texts from websites using web crawling technology. Websites that can be used include Sina Pharmaceutical News, Menet, and Yaozhi News, and the crawled news content includes pharmaceutical news updates, market trends, and industry information.

[0024] S2: A keyword set is obtained by manually summarizing and categorizing templates for medical news event types and manually defining keywords. The specific method is as follows:

[0025] S21: Manually summarize medical news event type templates. Medical news events published on news websites are obtained through web scraping. Event type templates for medical industry news in the financial sector are manually summarized. The templates include event type, trigger words, arguments, and argument roles. Trigger words are the core words of the event, arguments are the participants in the event, and argument roles are the roles the arguments play in practice. The obtained event type templates are used for event extraction.

[0026] S22: Manually define keywords. Summarize the patterns of event sentences in financial and medical industry news, identifying frequently occurring keywords within these sentences (note that these keywords do not include words related to event types or trigger words), thus creating a keyword set. The purpose of this step is that not every sentence in a news article contains an event; using keywords allows for preliminary categorization of event sentences, preserving valuable parts of the news text while removing irrelevant statements, achieving the effect of data cleaning.

[0027] S3: Construct a keyword pattern matching regular expression for keyword location.

[0028] S4: Using keyword localization technology, locate sentences containing keywords in medical news articles to obtain a dataset of sentences to be extracted. The specific method used is pattern matching. By constructing a regular expression for each keyword, the pattern matching method is used to locate keywords in the news text. The specific method is as follows:

[0029] S41: Text initialization, splitting the news text by periods and placing it into list S;

[0030] S42: Text preprocessing, removing newline characters, spaces, and images;

[0031] S43: Obtain the set M of regular expression templates for each keyword, where some keyword regular expression templates are shown in the table below.

[0032] Table 1 Keyword Regular Expression Templates

[0033]

[0034] S44: Iterate through each sentence in the sentence list S. With each template in the regular expression template set M ;

[0035] S45: If and Match the sentences and extract them into the dataset of sentences to be extracted.

[0036] S5: Use Word2vec to learn features from the sentences in the dataset to be extracted, and learn their word-level feature vectors respectively. and word-level feature vector representation .

[0037] S6: Since the requirements for character vectors and word vectors differ between the trigger word classification task and the argument extraction task, different weights are assigned to character vectors and word vectors for feature fusion for different tasks, resulting in a feature representation CF with a larger character weight and a feature representation WF with a larger word weight.

[0038] S7: Trigger word matching is performed using feature representations with high word weights (CF). This process uses a dynamic multi-pooling convolutional neural network. If no trigger word exists, the process returns to step S6 to select the next statement for trigger word matching. If a trigger word exists, pattern matching is used to optimize and limit the matched trigger word. The trigger words corresponding to each type of event are shown in Table 2.

[0039] Table 2. Trigger words for each type of event

[0040]

[0041] like Figure 2 The specific method for step S7 in the trigger word matching process is as follows:

[0042] S71: The vector feature CF with a large word weight after feature fusion and the position vector feature PF relative to the candidate trigger word position are concatenated to serve as the word-level feature for each word;

[0043] S72: Utilizes convolutional neural networks to capture multiple local semantic features;

[0044] S73: Perform dynamic multi-pooling on the captured semantic features to obtain sentence-level feature vectors;

[0045] S74: Input the learned sentence-level feature vectors into the classifier, calculate the probability of each word matching different event types using softmax. If there is no trigger word, return to step S6 to select the next sentence for trigger word matching. If there is a trigger word, take the trigger word with the highest probability and proceed to the next step.

[0046] S75: Use pattern matching to optimize the extraction results to ensure the correctness of word segmentation during the extraction process.

[0047] S8: The next step, argument extraction, is performed using the weighted feature representation (WF). This process employs a dynamic multi-pooling convolutional neural network. After argument extraction, pattern matching is used to optimize and limit the extraction results. The trigger words corresponding to each type of event are shown in Table 3.

[0048] Table 3. Argument Roles for Each Type of Event

[0049]

[0050] like Figure 2 The specific method for argument extraction in step S8 is as follows:

[0051] S81: The vector feature WF with the large word weight after feature fusion, the position vector feature PF relative to the position of the candidate trigger word, and the event type vector feature EF corresponding to the trigger word are concatenated to form the word-level feature of each word;

[0052] S82: Utilizes convolutional neural networks to capture multiple local semantic features;

[0053] S83: Perform dynamic multi-pooling on the captured semantic features to obtain sentence-level feature vectors;

[0054] S84: Input the learned sentence-level feature vectors into the classifier, and calculate the probability of each word matching different argument roles through softmax;

[0055] S85: The extracted event argument roles are optimized by using regular expressions in pattern matching to ensure the correctness of word segmentation during the extraction process, and finally the extraction result is obtained.

[0056] The extraction process is illustrated below using a news example.

[0057] S1: Using web crawling technology, obtain publicly available medical news texts from websites. Extract the event from the news article "On December 30, 2021, the National Healthcare Security Administration and the State Administration of Traditional Chinese Medicine jointly issued the 'Guiding Opinions on Medical Insurance Supporting the Inheritance and Innovative Development of Traditional Chinese Medicine.' The opinions clarify that general TCM medical services can continue to be paid for on a per-service basis, and TCM medical institutions can temporarily refrain from implementing payment based on Diagnosis Related Groups (DRG). For regions that have already implemented DRG and disease-based point-based payment, the coefficients and scores for TCM medical institutions and TCM diseases will be appropriately increased to fully reflect the characteristics and advantages of TCM services. For TCM-advantaged diseases requiring long-term hospitalization, such as rehabilitation medicine and palliative care, payment can be made on a per-diem basis."

[0058] S2: By manually summarizing and categorizing templates for medical news event types and manually defining keywords, a keyword set is obtained;

[0059] S21: Manually summarize medical news event type templates. Medical news events published on news websites are obtained through web scraping. Event type templates for medical industry news in the financial sector are manually summarized. The templates include event type, trigger words, arguments, and argument roles. Trigger words are the core words of the event, arguments are the participants in the event, and argument roles are the roles the arguments play in practice. The obtained event type templates are used for event extraction.

[0060] For example, in the event sentence "On December 30, the National Healthcare Security Administration and the State Administration of Traditional Chinese Medicine jointly issued the 'Guiding Opinions on Medical Insurance Supporting the Inheritance and Innovative Development of Traditional Chinese Medicine'", the event type is "issuance" time, the trigger word is "issuance", and the argument role correspondence is "issuance time - December 30", "issuing party - National Healthcare Security Administration and State Administration of Traditional Chinese Medicine", and "issuance content - 'Guiding Opinions on Medical Insurance Supporting the Inheritance and Innovative Development of Traditional Chinese Medicine'".

[0061] S22: Manually define keywords. Summarize the patterns of event sentences in news articles from the financial and medical industries, and find keywords that appear frequently in the event sentences (note that these keywords do not include words related to event types and trigger words), to obtain a keyword set. For example, "National Healthcare Security Administration" in the above example event sentence can be defined as a keyword. The purpose of this step is that not every sentence in a news article contains an event. Using this keyword, the event sentences in the news can be initially located and divided, retaining the valuable parts of the news text to the greatest extent and removing irrelevant sentences to achieve the effect of data cleaning.

[0062] S3: Construct a keyword pattern matching regular expression for keyword location.

[0063] S4: Using keyword targeting technology, locate sentences in medical news containing keywords to obtain a dataset of sentences to be extracted. The specific steps are as follows:

[0064] S41: Text initialization. The news text is split into lists S by periods. The statements in list S are shown in the table below:

[0065] Statements in List S of Table 4

[0066] s1 On December 30, 2021, the National Healthcare Security Administration and the State Administration of Traditional Chinese Medicine jointly issued the "Guiding Opinions on Medical Insurance Support for the Inheritance and Innovative Development of Traditional Chinese Medicine". s2 The guidelines clarify that general TCM medical services can continue to be paid on a per-service basis, and TCM medical institutions can temporarily refrain from paying based on Diagnosis Related Groups (DRG). For regions that have already implemented DRG and disease-based point-based payment, the coefficients and points for TCM medical institutions and TCM diseases will be appropriately increased to fully reflect the characteristics and advantages of TCM services. s3 For diseases where traditional Chinese medicine has advantages and requires long-term hospitalization, such as rehabilitation medicine and palliative care, payment can be made on a per-bed-day basis.

[0067] S42: Text preprocessing, removing unnecessary semicolons such as newlines and spaces;

[0068] S43: Get the set M of regular expression templates for each keyword;

[0069] S44: Iterate through each sentence in the sentence list S. With each template in the regular expression template set M ;

[0070] S45: If and The matching process extracts the sentence and adds it to the dataset of sentences to be extracted. The keyword regular expression ".*(Guarantee Bureau).*" matches s1 in list S, and s1 is added to the dataset of sentences to be extracted.

[0071] S5: Use Word2vec to learn features from the sentences in the dataset to be extracted, and learn their word-level feature vectors respectively. and word-level feature vector representation .

[0072] S6: Since the requirements for character vectors and word vectors differ between the trigger word classification task and the argument extraction task, different weights are assigned to character vectors and word vectors for feature fusion for different tasks, resulting in a feature representation CF with a larger character weight and a feature representation WF with a larger word weight.

[0073] S7: Trigger word matching is performed using word-weighted feature representation (CF). This process uses a dynamic multi-pooling convolutional neural network. The trigger word "publish" is matched, and the event type is determined to be "publishing event". Pattern matching is used to limit the trigger word, and no word segmentation errors are found. S8: The event type obtained from the previous step is "publishing event". Argument roles in this event type are obtained and argument extraction is performed. The argument roles of "publishing event" include "time, publisher, and published content". Word-weighted feature representation (WF) is used for the next argument extraction task. This process uses a dynamic multi-pooling convolutional neural network to extract the arguments "time - December 30, 2021", "publisher - National Healthcare Security Administration, State Administration of Traditional Chinese Medicine", and "published content - Guiding Opinions on Medical Insurance Support for the Inheritance and Innovative Development of Traditional Chinese Medicine". After extraction, pattern matching is used to limit and optimize the matching results.

[0074] The final results of the extraction of this news item are shown in Table 5.

[0075] Table 5 Sampling Results

[0076]

[0077] In summary, this invention improves extraction efficiency by first using keyword positioning technology to retain valuable sentences in news texts to the maximum extent; it then uses manually summarized event type templates from the financial and medical industries for targeted extraction; and by combining pattern matching and convolutional neural network extraction methods, it further improves the accuracy of event extraction in specific fields by using regular expressions for optimization after obtaining the extraction results.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. An extraction method for medical industry news events in the financial field, characterized in that: The method uses a web crawler to crawl public medical news as a data source, manually summarizes a medical news event type template in the financial field and defines keywords, acquires a to-be-extracted sentence dataset through keyword positioning technology, performs feature learning on the sentences in the to-be-extracted sentence dataset, and extracts events by using a method of fusing pattern matching and a convolutional neural network; The specific steps of the method are as follows: S1: acquiring public medical news text on a website through web crawler technology; S2: manually summarizing a medical news event type template and defining keywords to obtain a keyword set; S3: constructing a keyword regular expression used for pattern matching, for keyword positioning; S4: positioning sentences containing keywords in medical news by using keyword positioning technology to obtain a to-be-extracted sentence dataset; the specific method is as follows: S41: initializing text, and dividing news text into a list S according to a period; S42: text preprocessing, removing line breaks, spaces, and pictures; S43: acquiring a regular expression template set M of each keyword; wherein the regular expression template of the keyword "medical noun special class" includes ".*(patent).*", ".*(drug).*", ".*(research).*", ".*(preparation).*", ".*(product).*", ".*(treatment).*", and ".*(case).*"; the regular expression template of the keyword "company class" includes ".*(company).*" and ".*(enterprise).*"; the regular expression template of the keyword "related agency class" includes ".*(guarantee bureau).*", ".*(drug administration bureau).*", and ".*(supervision bureau).*"; and the regular expression template of the keyword "policy class" includes ".*(management bureau).*", ".*(health commission).*", ".*(policy).*", ".*(report).*", ".*(notice).*", and ".*(announcement).*"; S44: iterate over each sentence s in the list of sentences S i with each template m in the set of regular expression templates M i ; S45: If s i With m i Match, take the sentence into the extracted sentence dataset; S5: using Word2vec to learn features of the sentences in the pre-extracted sentence dataset, learning word-level feature vector representation f char and word-level feature vector representation f word ; S6: since the trigger word classification task and the argument extraction task have different requirements for word vectors and word vectors, the word vectors and the word vectors are respectively assigned different weights for feature fusion, to obtain a feature representation CF of a word weight and a feature representation WF of a word weight; S7: using the feature representation CF of the word weight to perform trigger word matching, and using a dynamic multi-pooling convolutional neural network in the process; if there is no trigger word, the next sentence is selected in step S6 for trigger word matching; if there is a trigger word, the matched trigger word is limited and optimized by using a pattern matching method; the specific method is as follows: S71: concatenating the vector feature CF of the word weight after feature fusion and the position vector feature PF relative to the position of the candidate trigger word, as the word-level feature of each word; S72: using a convolutional neural network to capture multiple local semantic features; S73: performing a dynamic multi-pooling operation on the captured semantic features to obtain a sentence-level feature vector; S74: input the learned sentence-level feature vector into the classifier, calculate the probability of each word matching different event types by softmax, if there is no trigger word, return to step S6 to select the next sentence for trigger word matching, if there is a trigger word, take out the trigger word with the maximum probability for the next step operation; S75: use pattern matching to limit the optimization of the extraction result to ensure the correctness of the word segmentation in the extraction process; S8: use the feature representation WF with large word weight for the next argument extraction task, which uses a dynamic multi-pooling convolutional neural network, and after extracting the argument, uses the regular expression of pattern matching to limit and optimize the extraction result; the specific method is as follows: S81: concatenate the vector feature WF with large word weight after feature fusion, the position vector feature PF relative to the position of the candidate trigger word, and the event type vector feature EF corresponding to the trigger word as the word-level feature of each word; S82: use a convolutional neural network to capture multiple local semantic features; S83: perform dynamic multi-pooling operation on the captured semantic features to obtain a sentence-level feature vector; S84: input the learned sentence-level feature vector into the classifier, and calculate the probability of each word matching different argument roles by softmax; S85: limit and optimize the extraction result by using the regular expression in pattern matching to ensure the correctness of the word segmentation in the extraction process, and finally obtain the extraction result.

2. The method according to claim 1, wherein the news event is a medical industry news event in the financial field. The specific method of step S2 is as follows: S21: manually summarize medical news event type templates, obtain medical event news from news information websites through network crawlers, and manually summarize event type templates of financial field medical industry news; the obtained event type templates are used for event extraction; the event type templates include event type, trigger word, argument, and argument role; the event trigger word is the core word of the event occurrence, the argument is the participant of the event, and the argument role is the role played by the argument in practice; S22: manually define keywords, manually summarize the occurrence rules of event sentences in financial field medical industry news, find keywords with high occurrence frequency in event sentences (note that this keyword does not include words involved in event type and trigger word), and obtain a keyword set; the purpose of this step is to preliminarily locate and divide the event sentences in the news by using the keywords, to maximize the retention of valuable parts in the news text and remove irrelevant sentences to achieve the effect of data cleaning.

Images