An opinion event recommendation method, system and terminal combining content and atlas

Abstract

The application belongs to the technical field of network space cognition, and discloses a public opinion event recommendation method, system and terminal combined with content and atlas. The field knowledge atlas is introduced into the related public opinion event recommendation, the entity extraction and relation extraction knowledge atlas construction method is used to integrate the public opinion event into the existing field atlas data, and the knowledge embedding method of TransE is used to calculate the knowledge representation of the event. The final event representation obtained is used as additional supplementary information, and the event with the largest influence value is recommended as a supplement when the number of recommended events does not meet k. The application innovatively introduces the field knowledge atlas into the related public opinion event recommendation, uses the entity extraction and relation extraction knowledge atlas construction method to integrate the public opinion event into the existing field atlas data, and uses the knowledge embedding method of TransE to calculate the knowledge representation of the event.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of network space cognition, and particularly relates to an opinion event recommendation method, system and terminal combining text content and a knowledge graph. BACKGROUND

[0002] At present, with the development of technology, the Internet has become the main carrier of public opinion fermentation and dissemination, and an opinion event mainly refers to an event that arouses the discussion, attention and dissemination of netizens on the Internet.

[0003] In the absence of any user behavior data, the available recommendation means is still mainly a content-based related item (event) recommendation model. The existing technical solution adopts a language model to represent the semantic of event content text, and calculates the similarity between the current event representation and the historical event representation by using different similarity calculation methods, and finally inputs the k most similar related events.

[0004] Through the above analysis, the problems and defects of the prior art are:

[0005] (1) The event recommendation method adopted by the existing opinion event recommendation technology can only recommend according to the surface semantic information of the event text, and there is no way to mine the deep correlation between different event texts / feature words.

[0006] (2) When the recommendation scene is a certain specific field, the knowledge in the field, such as a certain important person in the field, should be repeatedly appeared in the opinion event. The existing technical solution cannot apply the inherent characteristics of each entity (such as the characteristics of a person) appearing in the event to the event recommendation. This makes the information accuracy rate low and the information processing efficiency low.

[0007] The difficulty of solving the above problems and defects is:

[0008] Mining the deep correlation between event texts / feature words cannot be obtained only from the analysis of the surface semantic information of the event text, and it is necessary to construct an event information library for storing all useful information of the event, including integrating the historical event domain knowledge graph, event description text, event keyword, event category, etc.

[0009] The significance of solving the above problems and defects is:

[0010] Through the analysis of the event text content and the event graph structure, multi-angle and multi-level consideration of event recommendation is realized. In a specific recommendation scene, the characteristics of the opinion event can play a role in the graph structure, thereby having a positive effect on information extraction in the event recommendation process.

[0011] Based on the problems existing in the existing public opinion event recommendation method, the scheme proposes an event recommendation method combining knowledge representation and text content representation, introduces external information of a domain knowledge graph, and considers the complex correlation behind different events, solving the problem of being unable to depict the deep semantic and domain features of event text. At the same time, by introducing a recall mode based on event influence, the problem of insufficient number of events recommended caused by event representation is solved. SUMMARY

[0012] To overcome the problems in the related art, the present application discloses an embodiment of a public opinion event recommendation method, system and terminal combining text content and a knowledge graph.

[0013] The technical solution is as follows: a public opinion event recommendation method combining text content and a knowledge graph, comprising: combining a domain knowledge graph and public opinion event recommendation, and when performing public opinion event recommendation, no longer limiting to single consideration of event text content, integrating public opinion events into a domain knowledge graph; considering the structure and attributes of public opinion events in the domain knowledge graph, and finally obtaining a recommendation list from the perspectives of public opinion event text and graph network structure through multi-angle analysis and calculation.

[0014] In an embodiment of the present application, the public opinion event recommendation method combining text content and a knowledge graph specifically comprises:

[0015] Step one, non-periodic updating of new event information extraction and warehousing;

[0016] Step two, periodic historical event information updating and warehousing;

[0017] Step three, periodic new event information updating and warehousing;

[0018] Step four, related public opinion event recommendation.

[0019] In an embodiment of the present application, the step one non-periodic updating of new event information extraction and warehousing specifically comprises:

[0020] First step, crawling all social platform public opinion text data in the current time period, wherein the public opinion text data includes articles and posts;

[0021] Second step, filtering the crawled public opinion data by using domain keywords, and retaining the public opinion text data in the domain;

[0022] Thirdly, all events of the field in the current time period are extracted by using an event extraction algorithm, and the events are assigned corresponding categories; meanwhile, keywords and weights of the corresponding event text data are calculated by using a textrank algorithm, and the word vectors of each keyword are calculated by using a language model, and then the event text vector is calculated by using the weighted average of the keyword word vectors; the influence index of each event is calculated by using the heat, danger and focus network public opinion indexes, the event text with the largest influence is selected for event description content extraction, and the description content is subjected to entity and relation extraction, and is integrated into the current field knowledge graph by using entity disambiguation and entity linking;

[0023] Fourthly, the event category, event keyword and weight, event text representation, event influence value, event text summary information and knowledge graph information of the integrated new event are updated into the event information database.

[0024] In an embodiment of the present application, the step two of periodically updating the historical event information into the database specifically comprises:

[0025] The data calculated by the periodic historical event information in the event information database comprises knowledge graph representation, event representation combined with the knowledge graph and event text, and a recommended list of all events in the current event database;

[0026] Knowledge graph representation calculation: the information of the entire field graph is embedded by using the knowledge representation method of TransE;

[0027] Event text representation calculation: the keywords and corresponding weights of the event text are extracted by using the textrank method, and the embedding of each keyword is calculated by using the word2vec model, and the event text vector is calculated by using formula (1):

[0028]

[0029] Wherein c represents the event text vector, n represents the number of keywords of the event text, x i represents the embedding of the i-th keyword, w i represents the weight of the i-th keyword;

[0030] Event representation calculation: the event representation for similarity calculation is combined with the knowledge representation of the event and the text representation of the event, and is obtained by weighted summation through a self-defined weight, and the formula is:

[0031] v = w1u + w2c (2)

[0032] Wherein v represents the final event vector, u represents the event knowledge representation vector, w1 represents the weight of the event knowledge representation vector, c represents the event text vector, and w2 represents the weight of the event text vector;

[0033] The recommendation list calculation is based on the similar event representation, and the similarity calculation, retrieval method and vector storage method of Faiss are used to store the vectors and calculate the similarity s of the current event and other historical events in the periodic update; the similarity calculation, retrieval method and vector storage method are selected according to the size of the data amount and the different requirements; then, the historical events with s < threshold value d are removed, and the k historical events with the largest similarity are selected from the remaining events as the recommendation results; if the final recommendation results do not satisfy k, then the k-n events with the largest influence value in the event category to which the current event belongs are selected as the supplements, wherein n is the number of the recommendation results based on the event representation.

[0034] In an embodiment of the present application, the step three of updating the new event information into the database in the period includes:

[0035] (1) judging whether the new event can be associated with the current knowledge graph, and whether the extracted entities and relations have the same places as the current knowledge graph, if there are overlapping places, step (2) is implemented, otherwise step (3) is implemented;

[0036] (2) when the event is associated with the existing graph, the knowledge representation of the current new event can be calculated through the connected entities and relations, the current event representation is calculated by using formula (2), and step (4) is implemented;

[0037] (3) when the event is not associated with the existing graph, the text representation of the event is taken as the current event representation, and step (4) is implemented;

[0038] (4) after obtaining the event representation, the recommendation list of the event is generated and stored in the event information database; when the event is not associated with the existing graph, the similarity between the new event representation and the text representation of all historical events is calculated to generate the recommendation results.

[0039] In an embodiment of the present application, the step three of updating the new event information into the database in the period further includes: calculating the event representation combined with the knowledge graph and the text content, specifically including the following steps:

[0040] 1) selecting the intersection of the extracted entities and relations of the new event and the entities and relations of the existing graph as (e, r), (E, R) is all entity and relation pairs of the knowledge graph, (V, L) is all entity and relation pairs extracted from the new event, and formula (3) is:

[0041] (e, r) = (E, R) ∩ (V, L) (3)

[0042] 2) using the knowledge representation method of TransE to calculate the knowledge representation of the new event with the connected entities and relations of the new event, and formula (4) is:

[0043]

[0044] wherein n represents the number of entity relation pairs of (e, r), e i represents the vector of the connected entity, r i represents the vector of the connected relation.

[0045] 3) The final new event representation vector is obtained by using the method of formula (2).

[0046] In an embodiment of the present application, the step four related public opinion event recommendation comprises:

[0047] When an event is recommended, it is judged whether the event is a new event, if yes, the recommended list finally calculated in the library is output after the new event information is updated, otherwise the recommended list stored in the library is directly output.

[0048] Another object of the present application is to provide a public opinion event recommendation system combining text content and knowledge graph, which implements the public opinion event recommendation method combining text content and knowledge graph, and is characterized in that the public opinion event recommendation system combining text content and knowledge graph is based on the embedding recommendation method, the similarity between the current event representation and the historical event representation is compared by calculating the event representation, and the topk events with the similarity d greater than the threshold value are selected as the recommended list; specifically comprising:

[0049] An event information library is used to store all information needed for event recommendation, including the integrated historical event domain knowledge graph, event description text, event keywords and corresponding weights, event category, event influence index, and is also used to store the calculated event representation and the recommended list corresponding to the historical event.

[0050] An event representation unit is used to form a recommended list by comparing the similarity between event representations, and the final event representation and event text representation are stored in the event information library.

[0051] An event recommendation list unit is used to generate three kinds of recommended lists, which are the recommended list based on the event representation, the recommended list based on the event text representation and the recommended list based on the event influence index.

[0052] Another object of the present application is to provide a receiving user input program storage medium, and the stored computer program makes the electronic device execute the public opinion event recommendation method combining text content and knowledge graph.

[0053] Another object of the present application is to provide an information data processing terminal comprising a memory and a processor, the memory storing a computer program, the computer program being executed by the processor to cause the processor to execute the public opinion event recommendation method combined with text content and a knowledge graph.

[0054] In combination with all the technical solutions described above, the present application has the following advantages and positive effects:

[0055] Based on the event recommendation similarity calculation of the present application, see Table 1.

[0056] Table 1

[0057] Event id to be recommended Top 5 event ids recommended Similarity distance 0 5553 0.1937712 0 5413 0.4555958 0 7145 0.5150939 0 58 0.5561578 0 2717 0.5599424

[0058] The average value of the similarity distance of the five recommended events starting from the event id_0 is 0.4561122.

[0059] Based on the event recommendation similarity calculation of the event text surface meaning, see Table 2.

[0060] Table 2

[0061] Event id to be recommended Top 5 event ids recommended Similarity distance 0 6564 0.5541926 0 6602 0.5561241 0 5413 0.5742827 0 1801 0.5797074 0 3895 0.5883735

[0062] The average value of the similarity distance of the five recommended events starting from the event id_0 is 0.5705360.

[0063] Technical effect and advantage analysis:

[0064] In addition to considering the surface semantics of the event text, the present application also extracts entities and relationships from the event itself, and further performs deep mining in the graph network structure. Based on the multi-level consideration, the event recommendation is more reasonable, and the average similarity distance is smaller, reflecting the better technicality.

[0065] The present application innovatively introduces the field knowledge graph into the related public opinion event recommendation, adopts the knowledge graph construction method such as entity extraction and relationship extraction to integrate the public opinion event into the existing field graph data, and adopts the knowledge embedding method of TransE to calculate the knowledge representation of the event. At the same time, in terms of shallow semantics, the scheme still calculates the semantic representation of the event text. By weighting the event knowledge representation and the semantic representation, the single event semantic representation is replaced as the final event representation. In addition, the present application also adopts the influence index of the public opinion event as additional supplementary information, and recommends a certain number of events with great influence as a supplement when the number of recommended events does not meet k.

[0066] It is understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and cannot limit the disclosure of the present application. Attached Figure Description

[0067] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0068] Figure 1 This is a flowchart of a public opinion event recommendation method that combines text content and knowledge graph, provided in an embodiment of the present invention.

[0069] Figure 2 This is a diagram of a public opinion event recommendation system that combines text content and knowledge graph, provided in an embodiment of the present invention.

[0070] In the diagram: 1. Event information database; 2. Event representation unit; 3. Event recommendation list unit.

[0071] Figure 3 This is a flowchart of the new event information extraction and storage process provided in the embodiments of the present invention.

[0072] Figure 4 This is a flowchart of the event information database update information calculation process provided in an embodiment of the present invention.

[0073] Figure 5 This is a flowchart of the periodic recommendation list calculation provided in an embodiment of the present invention.

[0074] Figure 6 This is a diagram showing the update and storage of new event information within a given period, provided by an embodiment of the present invention.

[0075] Figure 7 This is a flowchart of the relevant event recommendation process provided in the embodiments of the present invention. Detailed Implementation

[0076] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0077] like Figure 1As shown, the present application provides a public opinion event recommendation method combining text content and knowledge graph, which comprises the following steps: introducing a domain knowledge graph into the related public opinion event recommendation, integrating the public opinion event into the existing domain graph data by using an entity extraction and relationship extraction knowledge graph construction method, and calculating the knowledge representation of the event by using a TransE knowledge embedding method; at the same time, on the shallow semantic level, calculating the semantic representation of the event text, and taking the weighted event knowledge representation and semantic representation as the final event representation.

[0078] The final event representation obtained is supplemented by the influence index of the public opinion event, and k-n events with the largest influence values are recommended as supplements for the event recommendation not meeting k.

[0079] In a preferred embodiment of the present application, the public opinion event recommendation method combining text content and knowledge graph specifically comprises the following steps:

[0080] Step 1: Non-periodic updating of new event information extraction and storage.

[0081] Step 2: Periodic updating of historical event information storage.

[0082] Step 3: Updating of new event information storage in a period.

[0083] Step 4: Related public opinion event recommendation.

[0084] The technical solutions of the present application will be further described below in combination with specific embodiments.

[0085] Embodiment

[0086] 1. Public opinion event recommendation system combining text content and knowledge graph

[0087] The public opinion event recommendation system combining text content and knowledge graph is mainly a recommendation method based on embedding, which compares the similarity of the current event representation and the historical event representation by calculating the event representation, and selects the top k events with a similarity d greater than a threshold value as a recommendation list.

[0088] As shown in the public opinion event recommendation system combining text content and knowledge graph. It mainly comprises three parts: Figure 2 The public opinion event recommendation system combining text content and knowledge graph mainly comprises three parts:

[0089] Event information database 1: The event information database stores all the information required for event recommendation, including the domain knowledge graph integrated with historical events, event description text, event keywords and corresponding weights, event category, and event influence index. In addition, the event representation calculated in the subsequent steps and the corresponding recommendation list of the historical events are also stored in the event information database. At the same time, with the continuous generation of new events, the event information database is also continuously updated incrementally.

[0090] Event representation unit 2: The current recommendation scheme is still mainly formed by calculating the similarity between event representations to form a recommendation list. The final event representation and event text representation are stored in the event information library 1. In terms of event representation incremental update, the current scheme combines the method of periodic full update and incremental update within the period to ensure the freshness of the recommendation results.

[0091] Event recommendation list unit 3: The current recommendation scheme will generate three kinds of recommendation lists, namely event representation-based recommendation list, event text representation-based recommendation list and event influence index-based recommendation list, among which the event representation-based recommendation list is the main recommendation means. The event recommendation list mainly adopts the method of periodic update to ensure the freshness of the recommendation results.

[0092] 2. Specific calculation process

[0093] 2.1 Non-periodic update of new event information extraction and storage process, as shown in Figure 3 , including:

[0094] First, all social platform public opinion text data such as articles and posts in the current time period need to be crawled. Then, the domain keywords are used to filter the crawled public opinion data, and the domain public opinion text data is retained. Next, the event extraction algorithm is used to extract all events about the domain in the current time period, and the events are assigned corresponding categories. At the same time, the textrank algorithm is used to calculate the keywords and weights of the corresponding event text data, and the language model is used to calculate the word vector of each keyword. Finally, the event text vector is calculated by weighted average of keyword word vectors. In addition, the influence index of each event needs to be calculated through network public opinion indicators such as heat, danger and focus. The event text with the largest influence is selected for event description content extraction, and the description content is subjected to entity and relationship extraction. Through entity disambiguation and entity linking, it is integrated into the current domain knowledge graph.

[0095] Finally, all information of the new event, including event category, event keyword and weight, event text representation, event influence value, event text summary and knowledge graph integrated with the new event, are updated to the event information library.

[0096] 2.2 In the periodic historical event information update storage process, the event information library update information calculation process is as shown in Figure 4 .

[0097] In order to keep the freshness of the recommendation results, the information in the event information library needs to be updated periodically, and the data that needs to be updated includes the knowledge graph representation, the event representation combined with the knowledge graph and the event text, and the recommendation list of all events in the current event library. Other information in the event information library includes event id, event text, event category, event keyword, event text representation, etc. The update operation is completed when a new event is added to the event information library, and there is no need to repeat the update.

[0098] In the periodic update of public opinion event information library, four kinds of calculations are mainly involved, namely graph knowledge representation calculation, event text representation calculation, event representation calculation and recommendation list calculation.

[0099] (1) Graph knowledge representation calculation: the current scheme mainly uses the knowledge representation method of TransE to embed the information of the entire field graph;

[0100] (2) Event text representation calculation: the textrank method is used to extract the keywords of the event text and the corresponding weight, and the word2vec model is used to calculate the embedding of each keyword, and the event text vector is calculated by the following formula:

[0101]

[0102] Where c represents the event text vector, n represents the number of event text keywords, x i represents the embedding of the i-th keyword, w i represents the weight of the i-th keyword.

[0103] (3) Event representation calculation: the final event representation used for similarity calculation combines the event knowledge representation and the event text representation, and is obtained by weighted summation through a self-defined weight, as follows:

[0104] v = w1u + w2c (2)

[0105] Where v represents the final event vector, u represents the event knowledge representation vector, w1 represents the weight of the event knowledge representation vector, c represents the event text vector, and w2 represents the weight of the event text vector.

[0106] (4) Recommendation list calculation, as shown in the periodic recommendation list calculation process. Figure 5

[0107] ​In periodic update, the recommendation is mainly based on the similarity event representation, and the similarity calculation and retrieval method of Faiss and the vector storage method are used to store vectors and calculate the similarity s of the current event and other historical events. The similarity calculation, retrieval method and vector storage method can be selected according to the size of the data and the different needs, such as IndexFlatL2 (calculating the Euclidean distance between different vectors by exhaustive search and storing the full amount of vectors), IndexIVFFlat (faster search by k-means clustering and dividing the search space, which can use Euclidean distance or inner product to calculate the similarity and store the full amount of vectors), IndexIVFPQ (lossy compression of vectors by using product quantization based method and storage), etc.

[0108] Then, by removing the historical events with s < threshold d, and selecting the k historical events with the largest similarity from the remaining events as the recommendation result. If the final recommendation result does not meet k, this scheme will select the k-n events with the largest influence value in the event category to which the current event belongs as a supplement, where n is the number of recommendation results based on event representation.

[0109] 2.3 The information update process of new events in the period is shown in Figure 6

[0110] In the period, when a new event is input and the basic information of the event has been updated, some information specific to the period needs to be updated. The steps are as follows:

[0111] (1) Determine whether the event can be associated with the current knowledge graph, and whether the extracted entities and relationships have any overlap with the current knowledge graph. If there is an overlap, step (2) is performed, otherwise step (3) is performed;

[0112] (2) When the event is associated with the existing graph, the knowledge representation of the newly added event can be calculated through the connected entities and relationships, and the current event representation is calculated using formula (2), and step (4) is performed;

[0113] (3) When the event is not associated with the existing graph, the text representation of the event is taken as the current event representation, and step (4) is performed;

[0114] (4) After obtaining the event representation, the recommendation list of the event needs to be generated and stored in the event information database. The recommendation list calculation method is basically the same as that in the periodic update, but when the event is not associated with the existing graph, the similarity between the new event representation and the text representation of all historical events needs to be calculated to generate the recommendation result;

[0115] ​The most important step of periodic update is to calculate the event representation combined with the knowledge graph and the text content, and the calculation method is divided into the following steps:

[0116] 1) Select the intersection of the entities and relations extracted from the new event and the entities and relations of the existing graph as (e, r), and (E, R) is all entity and relation pairs of the knowledge graph, (V, L) is all entity and relation pairs of the new event extraction, and the specific formula is as follows:

[0117] (e, r) = (E, R) ∩ (V, L) (3)

[0118] 2) Since the knowledge representation method of TransE is adopted, the entity and relation connected with the new event are used to calculate the knowledge representation of the new event according to the idea, and the formula is as follows:

[0119]

[0120] Wherein, n represents the number of entity relation pairs of (e, r), e i represents the vector of the connected entity, and r i represents the vector of the connected relation.

[0121] 3) The final new event representation vector is obtained by using the method of formula (2).

[0122] 2.4 Related public opinion event recommendation, as Figure 7 shown. After obtaining the analysis process of the event to be recommended, the type judgment (new event \ historical event) of the event to be recommended is included, and then the event recommendation is performed through different processes. For example, for a new event, in addition to outputting the list of event recommendations, the event information library should also be updated according to the new event, and the expansion of the event information library is completed in real time.

[0123] After the above steps, when an event needs to be recommended, it is first necessary to judge whether the event is a new event, if yes, the new event information is updated and then the final calculated recommendation list in the library is output, otherwise the recommendation list stored in the library is directly output.

[0124] The positive effects of the present application are further described in combination with specific experimental data.

[0125] Comparison and analysis of event recommendation similarity calculation based on the present application and event text surface meaning based event recommendation similarity calculation:

[0126] For the event to be recommended: id_0, the present application scheme and other technical schemes are used for recommendation respectively, and the average similarity distance is given, see Table 3.

[0127] Table 3

[0128]

[0129] Other embodiments of the disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the concepts disclosed herein. It is intended that the specification and examples be considered as exemplary only, with the true scope and spirit of the disclosure being indicated by the following claims.

[0130] It will be understood that the present disclosure is not limited to the precise structures herein described and illustrated in the drawings, and that various modifications and changes can be made without departing from its scope. The scope of the present disclosure should be gauged by the following claims.

Claims

1. A method for recommending public opinion events that combines text content and knowledge graphs, characterized in that, The proposed method for recommending public opinion events by combining text content and knowledge graphs includes: combining domain knowledge graphs and public opinion event recommendations; integrating public opinion events into the domain knowledge graph during the recommendation process; and obtaining a recommendation list from the perspective of public opinion event text and graph network structure through multi-angle analysis and calculation based on the structure and attributes of public opinion events in the domain knowledge graph. The method for recommending public opinion events that combines text content and knowledge graphs specifically includes: Step 1: Extract and store new event information that is not periodically updated; Step 2: Periodically update and store historical event information in the database; Step 3: Update and store new event information within the cycle; Step 4: Recommendation of relevant public opinion events; The specific steps of extracting and storing the non-periodic updated new event information in step one include: The first step is to crawl the public opinion text data of all social media platforms within the current time period, including articles and posts; The second step is to filter the crawled public opinion data using domain keywords, while retaining the public opinion text data within the domain. The third step involves using an event extraction algorithm to extract all events related to the domain within the current time period and assigning each event to a corresponding category. Simultaneously, the TextRank algorithm is used to calculate the keywords and weights of the corresponding event text data, and a language model is used to calculate the word vector for each keyword. The event text vector is then calculated by weighted averaging of the keyword word vectors. Furthermore, the influence index of each event is calculated using online public opinion indicators such as popularity, risk level, and attention level. The event text with the greatest influence is selected for event description content extraction, and entity and relation extraction is performed on the description content. Entity disambiguation and entity linking are then used to integrate this information into the current domain knowledge graph. The fourth step is to update the event information database with the event category, event keywords and weights, event text representation, event influence value, event text summary information, and integrated knowledge graph information of the new event.

2. The method for recommending public opinion events by combining text content and knowledge graphs according to claim 1, characterized in that, Step two, periodic historical event information update and storage, specifically includes: The data used to calculate periodic historical event information in the event information database includes knowledge graph representations, event representations combining knowledge graphs and event texts, and a recommendation list of all events in the current event database. Graph knowledge representation computation: The TransE knowledge representation method is used to embed information from the entire domain graph; Event text representation calculation: Keywords and their corresponding weights are extracted from the event text using the TextRank method, and the embedding of each keyword is calculated using the Word2Vec model. The event text vector is then calculated using formula (1). （1） in Represents an event text vector. Indicates the number of key elements in the event text. Indicates the first Embedding of keywords, Indicates the first The weight of each keyword; Event representation calculation: The event representation used for similarity calculation combines the event's knowledge representation and textual representation, and is obtained through a weighted summation using custom weights. The formula is: （2） in Represents the final event vector. Represents event knowledge as a vector. The weights of the event knowledge representation vector are used to represent the event knowledge. Represents an event text vector. The weights of the event text vectors; The recommendation list is calculated and updated periodically. Recommendations are made based on similar event representations. Faiss's similarity calculation and retrieval methods, along with vector storage methods, are used to store vectors and calculate the similarity between the current event and other historical events. The similarity calculation, retrieval method, and vector storage method are selected based on the data volume and specific requirements; then, by removing... Historical events, and then select the events with the greatest similarity from the remaining events. A historical event is used as the recommendation result; if the final recommendation result does not meet the requirements... If there are 100, then select the one with the highest influence value within the event category to which the current event belongs. One event is provided as a supplement, among which The number of recommendation results based on event representation.

3. The method for recommending public opinion events by combining text content and knowledge graphs according to claim 1, characterized in that, The update and storage of new event information within the third cycle in step three includes: (1) Determine whether the new event can be associated with the current knowledge graph, and whether the extracted entities and relationships overlap with the current knowledge graph. If there is overlap, proceed to step (2); otherwise, proceed to step (3). (2) When the event is related to the existing graph, the knowledge representation of the newly added event can be calculated through the connected entities and relationships, and the current event representation can be calculated using formula (2), and step (4) can be implemented. (3) When the event is not associated with the existing graph, the text representation of the event is used as the current event representation, and step (4) is performed. (4) After obtaining the event representation, generate a recommendation list for the event and store it in the event information database; when the event is not related to the existing graph, the new event representation needs to calculate the similarity with the text representations of all historical events and generate recommendation results.

4. The method for recommending public opinion events by combining text content and knowledge graphs according to claim 1, characterized in that, The update and storage of new event information within the third cycle also includes: calculating the event representation combining the knowledge graph and text content, specifically including the following steps: 1) The intersection of the entities and relations extracted from the newly added event and the entities and relations in the existing graph is represented as: ,and For all entity and relation pairs in the knowledge graph, Formula (3) is used to extract all entity and relation pairs for the new event: （3） 2) The knowledge representation of the new event is calculated using the TransE knowledge representation method and the entities and relations connected to the new event. Formula (4) is as follows: （4） in, express The entity relationship with the quantity, A vector representing connected entities. A vector representing a connection relationship; 3) The final new event representation vector is obtained by using formula (2).

5. The method for recommending public opinion events by combining text content and knowledge graphs according to claim 1, characterized in that, The recommended public opinion events in step four include: When recommending an event, determine whether the event is a new event. If it is, update the new event information and output the final calculated recommendation list in the database. Otherwise, directly output the recommendation list stored in the database.

6. A public opinion event recommendation system combining text content and knowledge graph, implementing the public opinion event recommendation method combining text content and knowledge graph as described in any one of claims 1 to 5, characterized in that, The public opinion event recommendation system that combines text content and knowledge graphs uses an embedding-based recommendation method. It calculates event representations, compares the similarity between the current event representation and historical event representations, and selects the top k events with a similarity d greater than a threshold as the recommendation list. Specifically, it includes: The event information database stores all the information needed for event recommendation, including a domain knowledge graph integrating historical events, event description text, event keywords and their corresponding weights, event categories, and event influence index; it also stores the calculated event representations and the recommendation lists corresponding to historical events. Event representation units are used to form a recommendation list by calculating the similarity between event representations. The final event representations and event text representations are stored in the event information database. The event recommendation list unit is used to generate three types of recommendation lists: a recommendation list based on event representation, a recommendation list based on event text representation, and a recommendation list based on event influence index.

7. A user input program storage medium, wherein the stored computer program causes an electronic device to execute the public opinion event recommendation method combining text content and knowledge graph as described in any one of claims 1 to 5.

8. An information data processing terminal, characterized in that, The information data processing terminal includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor performs the public opinion event recommendation method combining text content and knowledge graph as described in any one of claims 1 to 5.

Images