Knowledge graph establishment method and system and news event dialogue method and system

Abstract

A knowledge graph establishment method includes executing by a processing device: classifying a plurality of news articles into a main event associated with a plurality of sub-events; taking the main event as a first node in a knowledge graph, taking the sub-events as a plurality of second nodes in the knowledge graph respectively, connecting the second nodes to the first node, extracting a plurality of event summaries from the news articles according to a template rule respectively, taking the event summaries as a plurality of third nodes in the knowledge graph respectively, and connecting each of the third nodes to one of the second nodes according to the relevance of the event summaries and the sub-events; extracting a plurality of speaker identities from the event summaries as a plurality of fourth nodes in the knowledge graph, and connecting each of the fourth nodes to at least one of the third nodes.

Description

TECHNICAL FIELD

[0001] The present application relates to a knowledge graph establishing method and system and a news event dialogue method and system, in particular, to a knowledge graph establishing method and system and a news event dialogue method and system applicable to chatbots. BACKGROUND

[0002] In the era of information explosion, in addition to traditional paper newspapers and magazines, there are many news from different media on the network, and the same theme will also produce different information due to different interviewees. However, although the audience can receive multi-aspect information, not every interviewee has professional and detailed knowledge about the theme he / she interviewed, so the audience may receive incomplete or even incorrect information. SUMMARY

[0003] In view of the above, the present application provides a knowledge graph establishing method and system and a news event dialogue method and system.

[0004] According to a knowledge graph establishing method of an embodiment of the present application, comprising executing by a processing device: classifying multiple news articles into a main event associated with multiple sub-events; taking the main event as a first node in a knowledge graph; taking the sub-events as multiple second nodes in the knowledge graph respectively; connecting the second nodes to the first node; extracting multiple event summaries from the news articles according to a template rule respectively; taking the event summaries as multiple third nodes in the knowledge graph respectively, and connecting each of the third nodes to one of the second nodes according to the relevance of the event summaries and the sub-events; extracting multiple speaker identities from the event summaries; and taking the speaker identities as multiple fourth nodes in the knowledge graph, and connecting each of the fourth nodes to at least one of the third nodes.

[0005] According to a news event dialogue method of an embodiment of the present application, comprising executing by a processing device: analyzing a user question to obtain a target question; determining a target event knowledge graph matching the target question from multiple news event knowledge graphs generated by the knowledge graph establishing method of the above-mentioned embodiment; determining a target sub-event matching the target question from the sub-events of the target event knowledge graph; taking the event summary connected to the target sub-event as a target summary; and outputting a complete answer according to the target summary and one or more speaker identities corresponding to the target summary in the speaker identities.

[0006] A knowledge graph establishing system according to an embodiment of the present disclosure includes a storage and a processing device. The storage stores a main event and a plurality of sub-events, the main event being associated with the sub-events. The processing device is connected to the storage, and is configured to perform the following: classifying a plurality of news articles into the main event; taking the main event as a first node in a knowledge graph; taking the sub-events as a plurality of second nodes in the knowledge graph, respectively; connecting the second nodes to the first node; extracting a plurality of event summaries from the news articles according to a template rule, respectively; taking the event summaries as a plurality of third nodes in the knowledge graph, respectively, and connecting each of the third nodes to one of the second nodes according to the relevance between the event summaries and the sub-events; extracting a plurality of speaker identities from the event summaries; and taking the speaker identities as a plurality of fourth nodes in the knowledge graph, and connecting each of the fourth nodes to at least one of the third nodes.

[0007] A news event dialogue system according to an embodiment of the present disclosure includes a user interface, a storage, and a processing device. The user interface is configured to receive a user question, and the storage stores a plurality of news event knowledge graphs generated by the knowledge graph establishing method according to the above embodiment. The processing device is connected to the user interface and the storage, and is configured to perform the following: parsing the user question to obtain a target question; determining a target event knowledge graph matching the target question from the news event knowledge graphs; determining a target sub-event matching the target question from the sub-events of the target event knowledge graph; taking an event summary connected to the target sub-event as a target summary; and outputting a complete answer according to the target summary and one or more speaker identities corresponding to the target summary from the speaker identities.

[0008] In summary, the knowledge graph establishing method and system according to one or more embodiments of the present disclosure can provide structured and detailed information, so that a system using the knowledge graph (e.g., a search engine and the news event dialogue system according to the embodiments of the present disclosure) can quickly search for relevant information according to the knowledge graph. When a user raises a question through the user interface, the news event dialogue method and system according to one or more embodiments of the present disclosure can determine the field to which the question belongs and provide an event summary to reply to the user question, so that the user can obtain brief and relevant information without spending a lot of time searching.

[0009] The above description and the following description of the embodiments of the present disclosure are used to demonstrate and explain the spirit and principles of the present disclosure, and provide further explanation for the claims of the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0010] Figure 1A block diagram of a system based on a knowledge graph illustrated in an embodiment of the present invention.

[0011] Figure 2 A flowchart illustrating a knowledge graph creation method according to an embodiment of the present invention.

[0012] Figure 3 An example diagram of a knowledge graph drawn according to an embodiment of the present invention.

[0013] Figure 4 Draw Figure 2 The detailed flowchart of step S17.

[0014] Figure 5 An example diagram of a knowledge graph illustrated according to another embodiment of the present invention.

[0015] Figure 6 A block diagram of a news event dialogue system according to an embodiment of the present invention.

[0016] Figure 7 A flowchart illustrating a news event dialogue method according to an embodiment of the present invention.

[0017] Figure 8 Draw Figure 7 The detailed flowchart of step S25.

[0018] Figure 9 The diagram illustrates the application of a news event dialogue system and method according to one or more embodiments of the present invention to a chatbot. Detailed Implementation

[0019] The following detailed description of the features and advantages of the present invention in the embodiments is sufficient to enable any person skilled in the art to understand the technical content of the present invention and to implement it accordingly. Based on the disclosure of this specification, the claims, and the accompanying drawings, any person skilled in the art can easily understand the related objects and advantages of the present invention. The following embodiments further illustrate the points of the present invention in detail, but are not intended to limit the scope of the present invention in any way.

[0020] Please refer to Figure 1 and Figure 2 ,in Figure 1 A block diagram of a system based on a knowledge graph illustrated in an embodiment of the present invention is provided. Figure 2A flowchart of a knowledge graph establishment method according to an embodiment of the present application is shown. The knowledge graph establishment system 1 of an embodiment of the present application can include a processing device 11 and a memory 12. The processing device 11 can be electrically connected to or in communication with the memory 12. The processing device 11 can be implemented by a processor, a programmable logic device (PLD), an application specific integrated circuit (ASIC), or the like, but the present application is not limited thereto. The memory 12 is preferably a non-volatile memory (NVM), such as a read-only memory (ROM), an electrically-erasable programmable read-only memory (EEPROM), a flash memory, or the like, but the present application is not limited thereto.

[0021] The knowledge graph establishment method of an embodiment of the present application can be performed by the knowledge graph establishment system 1, in particular by the processing device 11 of the knowledge graph establishment system 1. As shown in Figure 2 The knowledge graph establishment method can include the following steps performed by the processing device 11: step S11: classifying a plurality of news articles into a main event, the main event being associated with a plurality of sub-events; step S12: taking the main event as a first node in a knowledge graph; step S13: taking the sub-events as a plurality of second nodes in the knowledge graph, respectively; step S14: connecting the second nodes to the first node; step S15: extracting a plurality of event summaries from the news articles according to a template rule, respectively; step S16: taking the event summaries as a plurality of third nodes in the knowledge graph, respectively, and connecting each of the third nodes to one of the second nodes according to the relevance of the event summaries to the sub-events; step S17: extracting a plurality of speaker identities from the event summaries; and step S18: taking the speaker identities as a plurality of fourth nodes in the knowledge graph, and connecting each of the fourth nodes to at least one of the third nodes. In this embodiment, step S12 and step S13 can be performed simultaneously, step S12 can be performed before step S13, and steps S12 to S14 can be performed before step S11, but the present application is not limited thereto.

[0022] For a detailed description of the above, please refer to Figures 1 to 3 wherein Figure 3 An example graph of a knowledge graph according to an embodiment of the present application is shown, and Figure 3The knowledge graph KG is illustrated with news about vaccines for the Coronavirus Disease 2019 (COVID-19) as an example. It is also noted that the memory 12 can pre-store the main event and the plurality of sub-events of step S11, and the main event is associated with the sub-events. Specifically, the processing device 11 or another processing device connected to the memory 12 can pre-establish the association between the main event and the sub-events, and the method of establishing the association between the main event and the sub-events can include extracting a plurality of sub-events from the plurality of news titles containing the main event and the first paragraph of the news, wherein the format of the sub-event can be a pair of “main word-verb” or a pair of “verb-object”, and the number of words between the main word and the verb and the number of spaces between the verb and the object are preferably no more than 10 words, and the present application does not limit the specific value of the number of spaces. For example, assuming that the news title of a news is “Parents don't worry! 77.76 million doses of children's vaccines arrived, and the fastest next week will be vaccinated”, and the first paragraph of the text contains “After signing a contract with the vaccine company to purchase 4 million doses of new coronavirus adult and children's vaccines, the first batch of 185.7 million doses of adult vaccines arrived, and the first batch of 77.76 million doses of children's vaccines also arrived this morning”, the sub-events extracted from the news title can include “vaccine arrived (main word-verb)” and “vaccine vaccination (main word-verb)”, and the sub-events extracted from the first paragraph of the text can include “vaccine company contract (main word-verb)”, “purchase vaccine (verb-object)”, and “vaccine arrived (main word-verb)”. In addition, since multiple sub-events can be extracted from the news, the sub-events can be sorted according to the frequency of each sub-event appearing in the text of the news before being stored in the memory 12, and sub-events with a frequency lower than a frequency threshold value can be deleted, wherein the frequency is, for example, the number of times each sub-event appears in the text of the news divided by the number of articles, and the frequency threshold value is, for example, 0.5, but the present application does not limit the specific value of the frequency threshold value.

[0023] At step Sll, the processing device 11 can use a web crawler to crawl a plurality of news articles related to the main event from the network, cloud, etc. For example, if the main event is a vaccine, and the sub-events are topics related to the vaccine, including that the protection of A brand vaccine has not been proven, mixing vaccines, and postponing the second dose of vaccine for teenagers, etc., the processing device 11 can use the web crawler to crawl a plurality of news articles related to the vaccine (main event) and vaccine topics (sub-events). Specifically, the processing device 11 can crawl the articles whose titles contain the words of the main event and / or the words of the sub-events, the articles whose texts contain the words of the main event and / or the words of the sub-events for a default number of times, etc. After determining the plurality of news articles related to the vaccine, the processing device 11 can classify the plurality of news articles into the main event “vaccine”. At steps S12, S13, and S14, the processing device 11 takes the main event “vaccine” as the first node Nl of the knowledge graph KG, takes the above-mentioned plurality of sub-events as a plurality of second nodes N2a, N2b, and N2c of the knowledge graph KG, and connects the second nodes N2a to N2c to the first node Nl. It should be noted that the “taking the main event as the first node of the knowledge graph” in this article means storing the main event as the first node of the knowledge graph, and the second node, the third node, and the fourth node described below are the same. In addition, the processing device 11 can first crawl a plurality of news articles whose titles contain the main event, and then obtain one or more sub-events from the titles and the first paragraphs of the texts of the articles in the above-mentioned manner to establish the relevance between the main event and the sub-events. After the above-mentioned frequency screening, if the remaining sub-events are consistent with the existing sub-events, the processing device 11 can perform the following steps on this news article; if the remaining sub-events are not consistent with the existing sub-events, the processing device 11 can establish new sub-events in the knowledge graph according to the remaining sub-events.

[0024] Next, at step S15, the processing device 11 extracts a plurality of event summaries from the news articles belonging to the main event according to a template rule, wherein the template rule can be “person name – said / indicated – word content”, and the content extracted according to the template rule can be used as the event summary. For example, if one of the news articles belonging to the main event records “… Zhang Luanying pointed out that adults over the age of forty, the elderly, or the ninth type of high-risk patients can consider mixing, and the protection will be much higher than two doses of AZ”, the processing device 11 can extract the person name “Zhang Luanying” and the word content “adults over the age of forty, the elderly, or the ninth type of high-risk patients can consider mixing, and the protection will be much higher than two doses of AZ” as the event summary according to the template rule. In other words, the event summary at this time contains the identity of the speaker and the content of the speech in the news article.

[0025] At step S16, the processing device 11 can use the above-mentioned template rule to extract a plurality of event summaries from the news articles belonging to the sub-events, and then use the above-mentioned method to establish the relevance between the main event and the sub-events. Figure 3For example, the processing device 11 can treat these event summaries as multiple third nodes N3a, N3b, N3c, and N3d in a knowledge graph KG, and connect the third nodes N3a to N3d of these event summaries to one of the second nodes N2a to N2c of the related sub-events. The correlation is, for example, that the event summary contains all the text of the sub-event, or contains a portion of the text of the sub-event, and the portion of the text accounts for a default proportion (e.g., 50%) of the total text of the sub-event, but this invention is not limited to this. Further, taking the third node N3c as an example, the event summary of the third node N3c contains a portion of the text of the sub-event of the second node N2b (i.e., "mixed typing") and reaches a default proportion. Therefore, the processing device 11 determines that the event summary of the third node N3c is related to the sub-event of the second node N2b and connects the third node N3c to the second node N2b.

[0026] In steps S17 and S18, the processing device 11 extracts the speaker's identity from the corresponding news article based on the event summary, uses the speaker's identity as the fourth node, and connects the fourth node to the corresponding third node. As mentioned above, the event summary extracted in step S15 contains the speaker's identity and their speech content in the news article, and in steps S17 and S18, the processing device 11 further extracts the speaker's identity from the event summary as the fourth node. Figure 3 For example, the event summary after extracting the speaker's identity would be as follows: Figure 3 The event summaries represented by the third nodes N3a to N3d shown can be represented by, for example, the speaker's identity (a person's name) as fourth nodes N4a, N4b, N4c, and N4d in the knowledge graph KG. Each of the fourth nodes N4a to N4d can be connected to its corresponding third node N3a to N3d. Similarly, taking the aforementioned news article as an example, if the event summary of the third node N3c corresponds to the speaker's identity "Zhang Luanying," the processing device 11 can use the speaker's identity "Zhang Luanying" as the fourth node N4c and connect the fourth node N4c to the third node N3c. Accordingly, the knowledge graph established through the above embodiments can provide structured and detailed information, allowing systems using the knowledge graph (e.g., search engines) to search for relevant information more quickly based on the knowledge graph.

[0027] In addition, after the knowledge graph KG is completed, the processing device 11 can also dynamically delete one or more of the second nodes N2a-N2c on the knowledge graph KG. Specifically, the processing device 12 can determine the total number of news related to the main event of the first node within a period of time (e.g., a week, a month, three months, etc.), and the number of occurrences of the sub-events of each second node in the news related to the main event, and divide the number of occurrences of the second node by the total number of the first node to obtain a ratio, if the ratio is less than a predetermined ratio, the processing device 12 can delete the second node and the third and fourth nodes connected after the second node, wherein the predetermined ratio may, for example, be 0.2, but the present application does not limit the actual value of the predetermined ratio.

[0028] To describe the content of extracting the speaker identity in more detail, please refer to Figure 1 and Figure 4 , wherein Figure 4 a detailed flowchart of step S17 of Figure 2 is shown. As shown in Figure 4 , step S17 of Figure 2 may include: step S171 of extracting a plurality of speaker names from the event summaries; step S172 of searching for a plurality of speaker positions of the speaker names from a professional database; step S173 of assigning a plurality of weight values corresponding to the speaker names according to a plurality of relevancies between the speaker positions and the main event, wherein the relevancy and the weight value are positively correlated; and step S174 of forming one of the speaker identities by each of the speaker names and the corresponding weight value in the weight values.

[0029] In steps S171 and S172, the processing device 11 extracts the speaker names from the event summaries, and searches for the positions corresponding to the speaker names as the speaker positions from a professional database, wherein the professional database stores a plurality of names and positions corresponding to the names, and the professional database can be a database stored in the memory 12, or an external database, such as Wikipedia, etc. Taking the news article as an example, the processing device 11 extracts the speaker name "Zhang Luanying", and searches for the positions of "Zhang Luanying" from the professional database, which includes the position of "pediatric infectious disease department attending physician", so the corresponding speaker position is "pediatric infectious disease department attending physician".

[0030] Next, at step S173, the processing device 11 assigns a weight value corresponding to the speaker's name according to the relevancy between the speaker's position and the main event. The higher the relevancy between the speaker's position and the main event, the higher the weight value corresponding to the speaker's name. The processing device 11 can pre-store a dictionary of relevant professional positions corresponding to the main event, which records positions having professional degrees in the field of the main event. For example, the dictionary of relevant professional positions corresponding to the main event "vaccine" can record the positions of the director of the official health unit, the pediatrician, and the professor of immunology department, etc. Therefore, when assigning a weight value corresponding to the speaker's name, the processing device 11 can assign a weight value corresponding to the speaker's name according to the following rules: if the speaker's position matches one of the relevant professional positions in the dictionary, and the speaker's position on the Wikipedia page matches the relevant professional position in the dictionary, the speaker is given the highest weight value; if the speaker's position matches one of the relevant professional positions in the dictionary, but the speaker has no corresponding Wikipedia page, the speaker is given the second highest weight value; and if the speaker's position does not match any of the relevant professional positions in the dictionary, the speaker is given the lowest weight value. In an embodiment, multiple speakers can have the same weight value, which is not limited by the present disclosure.

[0031] In addition, in the case where multiple third nodes are connected to the same second node, the processing device 11 can first filter out the irrelevant positions of the speakers from the dictionary according to the dictionary, and directly give the filtered speakers a preset lowest weight value (for example, the weight value is 0), so that the processing device 11 does not need to additionally search for the Wikipedia to assign a weight value corresponding to the speaker. Alternatively, for example, the speaker's position "pediatric infectious disease department attending physician", the processing device 11 can determine the number of times that "pediatric infectious disease department" or "pediatric infectious disease department attending physician" appears together with the main event "vaccine" by crawling the network, and divide the number of times by a preset value as the weight value; alternatively, the processing device 11 can search "pediatric infectious disease department" or "pediatric infectious disease department attending physician" and the main event "vaccine" through the Google search engine, and divide the total number of search results by a default value as the weight value. Conversely, if the processing device 11 extracts the speaker's name "Wang Xiaoming" from the event summary of the third node N3c, and the processing device 11 searches the professional database and finds that the speaker's position of "Wang Xiaoming" is "singer", then the weight value of "Wang Xiaoming" should be lower than that of "pediatric infectious disease department attending physician". In short, the weight value represents the professional degree of the speaker in the field of the speaker's speech content. The higher the weight value, the higher the professional degree of the speaker in the field of the speaker's speech content, and the more reference value of the speaker's speech content.

[0032] Next, in step S174, the processing device 11 matches the speaker's name with the corresponding weight value to form the speaker's identity as the fourth node. Figure 3 For example, the third speaker identity of the fourth node N4c includes the speaker's name and corresponding weight value for the statement that published the event summary of the third node N3c. This weight value represents the relevance between the speaker's position (corresponding to that name) and the main event of the first node N1. In other words, the third speaker identity of the fourth node N4c includes the speaker's name "Zhang Luanying" and its weight value related to the main event.

[0033] In another embodiment, in addition to performing the operations described in the above embodiments, the processing device 11 can further establish a speaker identity based on the speaking time and news URL. Please refer to... Figure 5 ,in Figure 5 An example diagram of a knowledge graph according to another embodiment of the present invention. Specifically, in Figure 2 The event summary retrieved in step S15 may include, in addition to the speaker's name, the speaking time (i.e., the time when the news article was published) and the news URL (i.e., the URL where the news article was published), and may also include the news media (i.e., the name of the media outlet that published the news article); alternatively, the event summary may also include one or more of the speaker's name, speaking time, news URL, and news media, and this invention is not limited thereto. Therefore, in Figure 2 In step S17, when the processing device 11 extracts the speaker's name from the event summary, it can also extract the speaking time, news media, and news URL from the event summary, and form the speaker's identity using the speaker's name and its corresponding speaking time, news media, and news URL. Figure 5 For example, the event summary of the third node N3c in the knowledge graph KG' connects to the second speaker's name SN2 of the fourth node N4c, which is "Zhang Luanying". The event summary of the third node N3c is taken from a news article with a speaking time of November 18, 2021, and a third-party URL. Since this news article was published by a third media outlet, the processing device 11 uses "Zhang Luanying" as the speaker's name SN2 of the fourth node N4c, the date "November 18, 2021" as the speaking time N41c of the fourth node N4c, "third-party media" as the news media N42c of the fourth node N4c, and the "third-party URL" that published the news article as the news URL N43c of the fourth node N4c. It should also be noted that, as... Figure 5 As shown, the same speaker's name can be linked to multiple event summaries, meaning multiple fourth nodes can indicate the same speaker's name but different speaking times, news media, and / or news URLs. Furthermore, the same sub-event can be linked to multiple different event summaries, and the speakers linked to each event summary can be the same or different.

[0034] Please refer to this as well. Figure 6 and Figure 7 ,in Figure 6 A block diagram of a news event dialogue system according to an embodiment of the present invention is shown. Figure 7 A flowchart illustrating a news event dialogue method according to an embodiment of the present invention is shown. The news event dialogue system 2 of an embodiment of the present invention may include a processing device 21, a memory 22, and a user interface 23. The processing device 21 may be electrically connected to the memory 22 and the user interface 23, or communicatively connected to the memory 22 and the user interface 23. The processing device 21 and the memory 22 may be connected to... Figure 1 The processing device 11 and memory 12 can also be connected with Figure 1 The processing device 11 and the memory 12 are different components but implemented in the same way, and the present invention is not limited thereto. The user interface 23 is used to receive instructions input by the user (e.g., user questions below) and present the corresponding results (e.g., complete answers below) to the user. The user interface 23 includes, for example, a microphone, speaker, keyboard, mouse and screen, and is preferably a touch screen, but the present invention is not limited thereto.

[0035] The news event dialogue system 2 and news event dialogue method of an embodiment of the present invention can be applied to a chatbot, providing corresponding answers based on questions input by users. The news event dialogue method of an embodiment of the present invention can be executed by the news event dialogue system 2, particularly by the processing device 21 of the news event dialogue system 2, and as... Figure 7 As shown, a news event dialogue method according to an embodiment of the present invention includes: step S21: analyzing the user's question to obtain the target question; step S22: determining the target event knowledge graph that matches the target question from multiple news event knowledge graphs generated by the knowledge graph building method; step S23: determining the target sub-event that matches the target question from the sub-events of the target event knowledge graph; step S24: using the event summary connected to the target sub-event as the target summary; and step S25: outputting a complete answer based on the target summary and one or more speaker identities corresponding to the target summary among the speaker identities.

[0036] At step S21, the processing device 21 can receive the user question through the user interface 23, and parse the user question to obtain the target question, for example, by natural language understanding (NLU) to parse the user question to obtain the target question. For example, if the user question is "why are teenagers not vaccinated with the second dose", the processing device 21 can obtain the target question "teenagers are not vaccinated with the second dose" by natural language understanding. Alternatively, the processing device 21 can determine the default keyword in the target question, and take the text connected after the default keyword as the target question. For example, the default keywords can include "why", "who", "when", "how", "where", etc. The processing device 21 can determine that the user question "why are teenagers not vaccinated with the second dose of vaccine" contains the default keyword "why", and take the text "teenagers are not vaccinated with the second dose of vaccine" connected after the default keyword as the target question.

[0037] At step S22, the processing device 21 determines the target event knowledge graph matched with the target question from a plurality of news event knowledge graphs, wherein the news event knowledge graphs are stored in the storage 22, and the news event knowledge graphs can be generated by one or more embodiments of the knowledge graph establishment method and system described above. The news event knowledge graphs can correspond to different main events respectively, for example, different main events such as vaccine, mask or Olympics, etc. The processing device 21 can compare the target question with the main events of each news event knowledge graph, and take the news event knowledge graph matched (for example, the target question covers the main event) as the selected target event knowledge graph. Taking the target question "why are teenagers not vaccinated with the second dose of vaccine" as an example, the processing device 21 determines that this target question covers the main event "vaccine", so the news event knowledge graph of the main event "vaccine" is taken as the target event knowledge graph, for example, the knowledge graph KG shown in FIG. 1. Figure 4

[0038] At steps S23 and S24, the processing device 21 compares a plurality of sub-events in the target event knowledge graph with the target question to select the target sub-event matched with the target question, and takes the event summary connected with the target sub-event as the target summary. The processing device 21 can select the sub-event matched with the target question by determining that the target question contains all the text of one of the sub-events, or contains part of the text of one of the sub-events, and the part of the text accounts for a default proportion (for example, 50%) of all the text of the target question, but the present application is not limited thereto. Taking the target question "why are teenagers not vaccinated with the second dose of vaccine" as an example, the processing device 21 can determine that the target question contains the sub-event "vaccine", and take the event summary connected with the sub-event "vaccine" as the target summary, for example, the event summary "vaccine" shown in FIG. 1. Figure 4 ​For example, the processing device 21 can determine from the sub-events of the second nodes N2a to N2c that one of them matches the target problem as the target sub-event of the second node N2c, and use the event summary of the third node N3d connected to the target sub-event of the second node N2c as the target summary.

[0039] Next, in step S25, the processing device 21 outputs a complete answer to the user interface 23 based on the target summary and the corresponding speaker's identity. The complete answer includes at least the speaker's identity and their speech content (i.e., the target summary). The processing device 21 can generate the complete answer using Natural Language Generation (NLG) technology. Similarly... Figure 4 For example, the target summary of the third node N3d corresponds to the identity of the fourth speaker N4d, so the complete answer could be "Fourth speaker identity: The epidemic is well controlled; the risk of myocarditis from the second dose is 5 to 10 times that of the first dose." Accordingly, when a user asks a question through the user interface, the processing device can determine the domain of the question and provide an event summary to answer the user's question, allowing the user to obtain concise and relevant information without spending a lot of time searching. Furthermore, by first determining the target event knowledge graph that matches the target question from multiple news event knowledge graphs, and then generating a complete answer based on the target sub-events of the target event knowledge graph, the processing device can quickly search for sub-events that match the target question, instead of searching for a matching sub-event from all sub-events of all knowledge graphs, effectively reducing the time it takes for the processing device to generate a complete answer.

[0040] Please continue to refer to this. Figure 6 and Figure 8 ,in Figure 8 Draw Figure 7 The detailed flowchart for step S25 is shown below. Figure 8 As shown, Figure 7 Step S25 may include: step S251: sorting the target summaries according to the weight values ​​corresponding to the speaker identities linked by the target summaries; step S252: generating multiple sub-answers according to each of the speaker identities and the corresponding target summaries; and step S253: sorting the sub-answers as a complete answer according to the sorting results of the target summaries.

[0041] As mentioned earlier, a speaker's identity can include their name and corresponding weight value. Furthermore, the same sub-event can be linked to multiple different event summaries, and the speaker identities linked to each event summary can be the same or different. Figure 8 The embodiments are described in the context of Figure 7The target summary obtained in step S24 is an implementation of multiple target summaries. Therefore, in step S251, the processing device 21 sorts the target summaries according to the weight values ​​corresponding to the speaker identities connected to them; in step S252, the processing device 21 matches the speaker identities and their corresponding target summaries one by one to generate multiple sub-answers; and in step S253, the processing device 21 sorts the sub-answers according to the sorting result of step S251 as a complete answer to be output to the user interface 23.

[0042] To explain in more detail Figure 8 Please refer to the following embodiments. Figure 6 , Figure 8 and Figure 9 ,in Figure 9 This diagram illustrates the application of a news event dialogue system and method according to one or more embodiments of the present invention to a chatbot. Assume that the user question Q obtained by user interface 23 is "Why are teenagers not receiving their second dose of vaccine yet?", and through... Figure 7 The target abstract matched by the example includes a first target abstract "The epidemic is well controlled, and the risk of myocarditis from the second dose is 5 to 10 times that of the first dose" and a second target abstract "There are currently no suitable vaccines for adolescents." The speaker identity corresponding to the first target abstract includes the first speaker's name and a first weight value, and the speaker identity corresponding to the second target abstract includes the second speaker's name and a second weight value. The first and second weight values ​​are determined based on... Figure 4 The implementation of the example is generated, and the first weight value is higher than the second weight value. Therefore, in step S251, the processing device 21 can sort the first target summary before the second target summary according to the first weight value and the second weight value; in step S252, the processing device 21 can generate the first sub-answer A1 "Name of the first spokesperson: The epidemic control is good, and the risk of myocarditis from the second dose is 5 to 10 times that of the first dose" and the second sub-answer A2 "Name of the second spokesperson: There is currently no suitable vaccine for adolescents"; and in step S253, the processing device 21 sorts the first sub-answer A1 before the second sub-answer A2 according to the sorting result of step S251, and outputs the sorted first sub-answer A1 and second sub-answer A2 as a complete answer A to the user interface 23 for presentation. Since the multiple sub-answers presented on the user interface 23 are sorted according to the weight value corresponding to the spokesperson, and the weight value indicates the professionalism of the spokesperson in the main event corresponding to the content of the event summary, the user can judge from the sorted multiple sub-answers that the sub-answers sorted earlier are more referential and correct than the sub-answers sorted later.

[0043] In summary, the knowledge graph establishment method and system according to one or more embodiments of the present application can provide structured and detailed information, so that the system using the knowledge graph (for example, a search engine and a news event dialogue system according to embodiments of the present application) can quickly search for relevant information according to the knowledge graph. When a user raises a question through a user interface, the news event dialogue method and system according to one or more embodiments of the present application can determine the field to which the question belongs and provide an event summary to reply to the user's question, so that the user can obtain brief and relevant information without spending a lot of time searching, and the user can determine that the sub-answer ranked first is more referential and correct than the sub-answer ranked last. In addition, the news event dialogue method and system according to one or more embodiments of the present application can quickly search for a sub-event matching the target question from the target event knowledge graph matching the target question according to the target sub-event of the target event knowledge graph, so that the system (for example, the processing device according to embodiments of the present application) can quickly search for a sub-event matching the target question without searching for one matching the target question from all sub-events of all knowledge graphs, effectively reducing the time for generating a complete answer.

[0044] SYMBOL DESCRIPTION

[0045] 1: knowledge graph establishment system

[0046] 2: news event dialogue system

[0047] 11, 21: processing device

[0048] 12, 22: memory

[0049] 23: user interface

[0050] KG, KG': knowledge graph

[0051] N1: first node

[0052] N2a, N2b, N2c: second node

[0053] N3a, N3b, N3c, N3d: third node

[0054] N4a, N4b, N4c, N4d: fourth node

[0055] N41a, N41b, N41c, N41d: speaking time

[0056] N42a, N42b, N42c, N42d: news media

[0057] N43a, N43b, N43c, N43d: news website

[0058] SN1, SN2: Speaker name

[0059] Q: User question

[0060] A1: First sub-answer

[0061] A2: Second sub-answer

[0062] A: Complete answer

[0063] S11, S12, S13, S14, S15, S16, S17, S18, S21, S22, S23, S24, S25, S171, S172, S173, S174, S251, S252, S253: Steps.

Claims

1. A method for establishing a knowledge graph, characterized in that, Includes execution by a processing device: Multiple news articles are categorized into a main event, which is related to multiple sub-events; This main event is used as the first node in the knowledge graph; These sub-events are each used as multiple second nodes in the knowledge graph; Connect these second nodes to the first node; Extract multiple event summaries from these news articles according to the template rules; These event summaries are used as multiple third nodes in the knowledge graph, and based on the relevance between these event summaries and these sub-events, each of these third nodes is connected to one of these second nodes; Extract the identities of multiple speakers from these event summaries; and These speaker identities are treated as multiple fourth nodes in the knowledge graph, and each of these fourth nodes is connected to at least one of the third nodes.

2. The knowledge graph construction method according to claim 1, wherein extracting the speaker identities from the event summaries includes: Extract the names of multiple speakers from these event summaries; Search for multiple spokesperson positions using the names of these spokespeople from a professional database; Based on multiple relevance levels between the speakers' positions and the main event, the speakers' names were assigned multiple weight values, where relevance and weight values ​​were positively correlated; and Each speaker's name is combined with its corresponding weight value to form one of their identities.

3. The knowledge graph construction method according to claim 2, wherein extracting the speaker identities from the event summaries further comprises: Extract multiple post times and multiple news URLs from these event summaries; and Each speaker's name, along with the corresponding speaking time and news URL, forms one of their identities.

4. A method for dialogue on news events, characterized in that, Includes execution by a processing device: Analyze user problems to obtain the target problem; Determine the target event knowledge graph that matches the target question from multiple news event knowledge graphs generated by the knowledge graph building method according to claim 1; Determine the target sub-events that match the target question from among the sub-events of the target event knowledge graph; The target summary is the summary of the event connected to the target sub-event; as well as Output a complete answer based on the target summary and one or more speaker identities that correspond to the target summary.

5. The news event dialogue method according to claim 4, wherein the target summary is multiple target summaries, and the complete answer is output based on the target summary and one or more speaker identities corresponding to the target summary, including: The target summaries are sorted according to multiple weight values ​​corresponding to the speaker identities they are linked to; Multiple sub-responses are generated based on the identities of the speakers and each of the corresponding target summaries; and The sub-answers are sorted according to the sorting results of these target summaries to form the complete answer.

6. A knowledge graph construction system, characterized in that, Include: The memory stores a main event and multiple sub-events, the main event being associated with the sub-events; and A processing device, connected to the memory, wherein the processing device is used to execute: Multiple news articles were categorized under this main event; This main event is used as the first node in the knowledge graph; These sub-events are each used as multiple second nodes in the knowledge graph; Connect these second nodes to the first node; Extract multiple event summaries from these news articles according to the template rules; These event summaries are used as multiple third nodes in the knowledge graph, and based on the relevance between these event summaries and these sub-events, each of these third nodes is connected to one of these second nodes; Extract the identities of multiple speakers from these event summaries; and These speaker identities are treated as multiple fourth nodes in the knowledge graph, and each of these fourth nodes is connected to at least one of the third nodes.

7. The knowledge graph building system according to claim 6, wherein the processing device is further connected to a professional database, and the processing device performs the extraction of speaker identities from the event summaries, including: Extract the names of multiple speakers from these event summaries; Search the database for multiple spokesperson positions based on the names of these spokespeople; Based on multiple relevance levels between the speakers' positions and the main event, the speakers' names were assigned multiple weight values, where relevance and weight values ​​were positively correlated; and Each speaker's name is combined with its corresponding weight value to form one of their identities.

8. The knowledge graph building system of claim 7, wherein the processing device performs the extraction of speaker identities from the event summaries, comprising: Extract multiple post times and multiple news URLs from these event summaries; and Each speaker's name, along with the corresponding speaking time and news URL, forms one of their identities.

9. A news event dialogue system, characterized in that, Include: The user interface is used to receive user questions; A memory, storing multiple news event knowledge graphs generated by the knowledge graph building method according to claim 1; and A processing device, connected to the user interface and the memory, wherein the processing device is used to perform: Analyze the user's problem to obtain the target problem; Determine the target event knowledge graph that matches the target question from these news event knowledge graphs; Determine the target sub-events that match the target question from among the sub-events of the target event knowledge graph; The target summary is the summary of the event connected to the target sub-event; as well as Output a complete answer based on the target summary and one or more speaker identities that correspond to the target summary.

10. The news event dialogue system of claim 9, wherein the target summary is a plurality of target summaries, and the processing device executes the output of the complete answer based on the target summary and one or more speaker identities corresponding to the target summary, comprising: The target summaries are sorted according to multiple weight values ​​corresponding to the speaker identities they are linked to; Multiple sub-responses are generated based on the identities of the speakers and each of the corresponding target summaries; and The sub-answers are sorted according to the sorting results of these target summaries to form the complete answer.

Images