A multi-threaded graph generation method and system based on motivation relationship

Abstract

The present application belongs to the technical field of multi-threaded graph generation, and particularly relates to a multi-threaded graph generation method and system based on motivation relationship. The method comprises: obtaining various event texts, and processing the obtained various event texts to obtain event entities corresponding to the various event texts; constructing a clue mining framework, inputting the various event entities into the clue mining framework to mine various clues including association information between various events, and calculating the confidence of the various clues; constructing a multi-layer convolutional neural network model, training the model by taking the clues with a confidence higher than a preset value as training samples to obtain a trained model; based on the trained model, predicting the confidence of the clues with a confidence lower than the preset value to replace the confidence of the mined clues with a confidence lower than the preset value; and based on the confidence of the various clues, constructing a multi-threaded graph to establish the connection between events.

Description

Technical Field

[0001] This invention belongs to the field of multi-clue graph generation technology, specifically relating to a multi-clue graph generation method and system based on motivational relationships. Background Technology

[0002] In the extraction of event relationships from event knowledge graphs, many relationships are proposed, including coreference, sub-event relationships, and causal relationships. Coreference refers to the relationship where two events point to the same thing; sub-event relationships reflect the inclusion relationship between two events; and causal relationships reflect the causal inheritance relationship between events, which has a certain temporal sequence.

[0003] However, in extracting events from listed companies, some events may not belong to any of the above relationships. One event of a listed company may trigger many events, and the triggered events may involve many different aspects of other events. Therefore, relying solely on coreference, sub-event relationships, causal relationships, etc., to extract the relationships between events of listed companies results in incomplete extraction. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for generating a multi-clue graph based on motivational relationships. By processing various event texts to obtain event entities corresponding to each event text, and mining various clues between events based on a clue mining framework and each event entity, the motivational relationships between each clue are extracted, the confidence of each clue is calculated, and a multi-clue graph of each event is constructed based on the confidence of each clue to establish the connection between events.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] The first aspect of this invention provides a method for generating a multi-clue graph based on motivational relationships, comprising the following steps:

[0007] S1. Obtain various event texts and process the obtained event texts to obtain event entities corresponding to various event texts. The event entity is a fragment containing information about the relationship between words.

[0008] S2. Construct a clue mining framework, input various event entities into the clue mining framework to mine various clues including the correlation information between various events, and calculate the confidence of various clues.

[0009] S3. Construct a multi-layer convolutional neural network model, and use clues with confidence scores higher than the preset value as training samples to train the model to obtain the trained model.

[0010] S4. Based on the trained model, predict the confidence of cues with confidence levels lower than the preset value, in place of the confidence of cues with confidence levels lower than the preset value in step S2.

[0011] S5. Based on the confidence levels of various clues, construct a multi-clue graph to establish connections between events.

[0012] As a preferred option, the event text consists of announcements published on the official websites of the listed companies and news about the listed companies on various media platforms.

[0013] As a preferred embodiment, step S1 includes:

[0014] S1.1 Clean up the various event texts obtained to obtain cleaned event texts;

[0015] S1.2. Use deep learning methods to extract events from the cleaned event texts and obtain the association information between words to obtain the event entities corresponding to the event texts.

[0016] As a preferred embodiment, step S1.1 includes deleting images and links from the event text using regular expressions.

[0017] As a preferred embodiment, step S2 includes:

[0018] S2.1 Build a web-based clue mining framework;

[0019] S2.2 Select at least two keywords as the purpose of mining, input various event entities into the Web model to obtain the search results corresponding to the current iteration. The search results are fragments that contain at least one selected keyword. Based on the number of keywords in all the fragments obtained from the current search, extract all the fragments obtained from the search and use the extracted results as multiple candidate clues corresponding to the current iteration, and calculate the confidence of each candidate clue.

[0020] S2.3. Repeat step S2.2 to obtain the confidence level of all candidate clues.

[0021] As a preferred option, in step S2.2, the confidence level of the clue is calculated based on the co-occurrence rate, point-by-point mutual information rate, and contextual similarity.

[0022] Co-occurrence rate, representing the percentage of all clues that contain the same keywords as the given clue;

[0023] Pointwise mutual information rate indicates the relevance of a clue containing two selected keywords to other clues;

[0024] Contextual similarity indicates the degree of similarity between the given clue text and the rest of the clue text.

[0025] As a preferred option, the formula for calculating the point-by-point mutual information rate (PMI) is:

[0026]

[0027] Where e1 and e2 represent two selected keywords, a represents the number of clues containing e1 and e2, b represents the number of clues containing e1 but not e2, c represents the number of clues containing e2 but not e1, and N represents the total number of clues.

[0028] As a preferred approach, after each iteration of step S2.2, it is necessary to determine whether the confidence level of all candidate clues is higher than the preset value. If so, the clue mining operation should be stopped.

[0029] As a preferred approach, starting from step S2.2 of the second iteration, after each iteration, it is necessary to determine whether the proportion of clues with confidence higher than the preset value among all candidate clues obtained in the current iteration is higher than the proportion in the previous iteration. If so, the clue mining operation is stopped.

[0030] The second aspect of the present invention provides a multi-clue graph generation system based on motivational relationships, and a multi-clue graph generation method based on motivational relationships according to the first aspect and any one of the first aspects of the present invention, comprising an event entity acquisition module, a clue mining module, a clue confidence prediction module, and a multi-clue graph construction module connected in sequence.

[0031] The event entity acquisition module is used to acquire various event texts of listed companies and process the acquired event texts to obtain various event entities. The event entities are fragments that include information about the relationships between words.

[0032] The clue mining module mines clues between various events based on the clue mining framework to obtain various clues including the correlation information between various events, and calculates the confidence of various clues.

[0033] The clue confidence prediction module trains the constructed multi-layer convolutional neural network model with clues ranked within a preset number to obtain the trained clue confidence prediction model. Based on the trained clue confidence prediction model, it predicts the confidence of clues ranked outside the preset number and replaces the confidence of clues with confidence values ​​lower than the preset value obtained in the clue mining module with the predicted confidence values.

[0034] The multi-clue graph construction module builds a multi-clue graph based on the confidence level of various clues to establish connections between events.

[0035] The beneficial effects of this invention are:

[0036] An event in a listed company may trigger many other events, which in turn involve various aspects of other events. Therefore, simply using coreference, sub-event relationships, and causal relationships to extract the relationships between events in a listed company results in incomplete extraction. This invention constructs a clue mining framework to mine various clues between events. The mined clues contain correlation information between various events and reflect the motivational relationships between events, thereby establishing connections between various events in a listed company.

[0037] At the same time, the confidence level of each clue is calculated, and a multi-clue map is constructed based on the confidence level of each clue. This not only establishes the connection between various events of the listed company, but also reflects the closeness of the connection between various events of the listed company and reflects the internal relationship of information between various events. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart of a multi-clue graph generation method based on motivational relationships provided in Embodiment 1.

[0040] Figure 2 This is a schematic diagram illustrating the use of an LSTM network to extract events from event text.

[0041] Figure 3 This is a module diagram of a multi-clue graph generation system based on motivational relationships provided in Embodiment 2. Detailed Implementation

[0042] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0043] Example 1:

[0044] Reference Figure 1This embodiment provides a method for generating a multi-clue graph based on motivational relationships, including the following steps:

[0045] S1. Obtain various event texts and process the obtained event texts to obtain event entities corresponding to various event texts. The event entity is a fragment containing information about the relationship between words.

[0046] S2. Construct a clue mining framework, input various event entities into the clue mining framework to mine various clues including the correlation information between various events, and calculate the confidence of various clues.

[0047] S3. Construct a multi-layer convolutional neural network model, and use clues with confidence scores higher than the preset value as training samples to train the model to obtain the trained model.

[0048] S4. Based on the trained model, predict the confidence of cues with confidence levels lower than the preset value, in place of the confidence of cues with confidence levels lower than the preset value in step S2.

[0049] S5. Based on the confidence levels of various clues, construct a multi-clue graph to establish connections between events.

[0050] Furthermore, the event texts refer to announcements published on the official websites of the listed companies and news about the listed companies on various media platforms.

[0051] Specifically:

[0052] The software crawls various news articles from listed companies on news websites and other media platforms, as well as announcements from the companies' official websites.

[0053] Furthermore, step S1 includes:

[0054] S1.1 Clean up the various event texts obtained to obtain cleaned event texts;

[0055] S1.2. Use deep learning methods to extract events from the cleaned event texts and obtain the association information between words to obtain the event entities corresponding to the event texts.

[0056] Furthermore, step S1.1 includes deleting some invalid event text and deleting images and links from the event text using regular expressions.

[0057] Specifically:

[0058] Event extraction is similar to event extraction in ordinary knowledge graphs. It is viewed as an end-to-end system that uses network structure to capture the relationships between words in the text.

[0059] The event extraction process is referenced. Figure 2 As shown.

[0060] In this embodiment, an LSTM network is selected. The basic language model TF-IDF is used to vectorize the text. The vectorized text is then input into the LSTM in both the forward and backward directions to capture the association information between words in both directions.

[0061] TF stands for Term Frequency, which is the frequency with which a word appears in a document. TF = Number of times a word appears in a document / Total number of words in the document. IDF, or Inverse Document Frequency, is an indicator of the importance of a word. IDF = log(Number of documents in the corpus / (Number of documents containing the word + 1)). The value of TF-IDF is the product of the value of TF and the value of IDF.

[0062] In this embodiment, the word vectors extracted by the LSTM network are illustrated in the table below:

[0063]

[0064] Furthermore, step S2 includes:

[0065] S2.1 Build a web-based clue mining framework;

[0066] S2.2 Select at least two keywords as the purpose of mining, input various event entities into the Web model to obtain the search results corresponding to the current iteration. The search results are fragments that contain at least one selected keyword. Based on the number of keywords in all the fragments obtained from the current search, extract all the fragments obtained from the search and use the extracted results as multiple candidate clues corresponding to the current iteration, and calculate the confidence of each candidate clue.

[0067] S2.3. Repeat step S2.2 to obtain the confidence level of all candidate clues.

[0068] Furthermore, in step S2.2, the confidence level of the clue is calculated based on the co-occurrence rate, point-by-point mutual information rate, and contextual similarity;

[0069] Co-occurrence rate, representing the percentage of all clues that contain the same keywords as the given clue;

[0070] Pointwise mutual information rate indicates the relevance of a clue containing two selected keywords to other clues;

[0071] Contextual similarity indicates the degree of similarity between the given clue text and the rest of the clue text.

[0072] Furthermore, the formula for calculating the point-to-point mutual information rate (PMI) is as follows:

[0073]

[0074] Where e1 and e2 represent two selected keywords, a represents the number of clues containing e1 and e2, b represents the number of clues containing e1 but not e2, c represents the number of clues containing e2 but not e1, and N represents the total number of clues.

[0075] Furthermore, the signal to stop excavation is:

[0076] After each iteration of step S2.2, it is necessary to determine whether the confidence level of all candidate clues is higher than the preset value. If so, the clue mining operation is stopped.

[0077] Alternatively, starting from step S2.2 of the second loop, after each loop is completed, it is necessary to determine whether the proportion of clues with confidence higher than the preset value among all candidate clues obtained in the current iteration is higher than the proportion in the previous iteration. If so, the clue mining operation is stopped.

[0078] Alternatively, starting from step S2.2 of the second loop, after each loop, it is necessary to determine whether the overlap between all candidate clues obtained in the current iteration and all candidate clues obtained in previous iterations is higher than a preset value. If so, the mining operation is stopped. When the overlap between all candidate clues obtained in the current iteration and all candidate clues obtained in previous iterations is higher than the preset value, it means that all clues obtained in the current iteration are unlikely to provide new information.

[0079] Specifically:

[0080] The purpose of the clue mining framework is to uncover effective clues that reflect the motivational relationships between events.

[0081] The reliability of a clue, or the confidence level of the clue, is calculated using co-occurrence rate, point-by-point mutual information rate, and contextual similarity.

[0082] In the process of clue mining, continuous cyclical mining can yield more clues, but the mining process also needs to consider efficiency. Therefore, setting judgment conditions can help avoid inefficient mining processes after finding valid clues.

[0083] Furthermore, using the clues with confidence levels higher than the preset value obtained in step S2 as training samples, the constructed multi-layer convolutional neural network model is trained. The trained neural network model then predicts the confidence levels of clues with confidence levels lower than the preset value obtained in step S2. The confidence levels predicted by the model are then used to replace the confidence levels of clues with confidence levels lower than the preset value obtained in step S2, thereby unifying the confidence levels of the clues obtained in step S2. Based on the confidence levels of each discovered clue, a multi-clue graph is constructed to establish the connection between time and events, while also reflecting the tightness of the connections between events.

[0084] Example 2:

[0085] Reference Figure 3 This embodiment provides a multi-clue graph generation system based on motivational relationships, which is based on the multi-clue graph generation method based on motivational relationships described in Embodiment 1 of the present invention. It includes an event entity acquisition module, a clue mining module, a clue confidence prediction module, and a multi-clue graph construction module connected in sequence.

[0086] The event entity acquisition module is used to acquire various event texts of listed companies and process the acquired event texts to obtain various event entities. The event entities are fragments that include information about the relationships between words.

[0087] The clue mining module mines clues between various events based on the clue mining framework to obtain various clues including the correlation information between various events, and calculates the confidence of various clues.

[0088] The clue confidence prediction module trains the constructed multi-layer convolutional neural network model with clues ranked within a preset number to obtain the trained clue confidence prediction model. Based on the trained clue confidence prediction model, it predicts the confidence of clues ranked outside the preset number and replaces the confidence of clues with confidence values ​​lower than the preset value obtained in the clue mining module with the predicted confidence values.

[0089] The multi-clue graph construction module builds a multi-clue graph based on the confidence level of various clues to establish connections between events.

[0090] It should be noted that the multi-clue graph generation system based on motivational relationships provided in this embodiment is similar to that in Embodiment 1, and will not be described in detail here.

[0091] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope of the present invention.

Claims

1. A multi-threaded graph generation method based on motivation relationship, characterized in that, The method comprises the following steps: S1, obtaining various event texts, and processing the obtained various event texts to obtain event entities corresponding to the various event texts, the event entities being segments containing association information between words; S2, constructing a clue mining framework, inputting the various event entities into the clue mining framework to mine various clues including association information between various events, and calculating the confidence of the various clues; S3, constructing a multi-layer convolutional neural network model, using clues with a confidence higher than a preset value as training samples to train the model, to obtain a trained model; S4, based on the trained model, predicting the confidence of clues with a confidence lower than the preset value to replace the confidence of the clues with a confidence lower than the preset value in step S2; S5, based on the confidence of the various clues, constructing a multi-clue graph to establish the relationship between events; Step S2 comprises: S2.1, constructing a Web-based clue mining framework; S2.2, selecting at least two keywords as the purpose of mining, inputting the various event entities into the Web model to obtain search results corresponding to the current iteration, the search results being segments containing at least one selected keyword, extracting all search segments based on the number of keywords in all search segments, and taking the extracted results as a plurality of candidate clues corresponding to the current iteration, and calculating the confidence of each candidate clue; S2.3, performing step S2.2 in a loop to obtain the confidence of all candidate clues; In step S2.2, the confidence of the clue is calculated based on co-occurrence rate, point-wise mutual information rate, and context similarity; The co-occurrence rate represents the percentage of clues containing the same keywords as the clue among all clues; The point-wise mutual information rate represents the relevance of a clue containing two selected keywords and other clues; The context similarity represents the similarity between the text of the clue and the texts of the remaining clues.

2. The method of claim 1, wherein, The event texts are announcements published on the official websites of listed companies and news about listed companies on various media platforms. 3.The method of claim 1, wherein, Step S1 comprises: S1.1, cleaning the obtained various event texts to obtain cleaned various event texts; S1.2, using a deep learning method to perform event extraction on the cleaned various event texts to obtain association information between words, to obtain various event entities corresponding to the various event texts.

4. The method of claim 3, wherein, In step S1.1, the pictures and links in the event texts are deleted by using regular expressions.

5. The method of claim 1, wherein, The calculation formula of the point-wise mutual information rate PMI is: Where e1 and e2 represent two selected keywords, a represents the number of clues containing e1 and e2, b represents the number of clues containing e1 but not e2, c represents the number of clues containing e2 but not e1, and N represents the total number of clues.

6. The method of claim 5, wherein, After each loop of step S2.2, it is necessary to determine whether the confidence of all candidate clues is higher than a preset value, and if so, the clue mining operation is stopped.

7. The method of claim 5, wherein, From the second cycle step S2.2, after each cycle is completed, it is necessary to judge whether the proportion of all candidate clues with confidence higher than the preset value in all candidate clues obtained after the current iteration is higher than the proportion in the previous iteration process, and if so, the clue mining operation is stopped.

8. A multi-threaded graph generation system based on motivation relationship, based on the multi-threaded graph generation method based on motivation relationship of any one of claims 1-7, characterized in that, The method comprises sequentially connected event entity acquisition module, clue mining module, clue confidence prediction module and multi-clue graph construction module. The event entity acquisition module is used for acquiring various event texts of listed companies and processing the acquired various event texts to obtain various event entities, wherein the event entity is a segment comprising the association information between words. The clue mining module mines the clues between various events based on a clue mining framework to obtain various clues comprising the association information between various events and calculates the confidence of various clues. The clue confidence prediction module trains the multi-layer convolutional neural network model constructed by the clues ranked within the preset number to obtain the trained clue confidence prediction model, predicts the confidence of the clues ranked outside the preset number based on the trained clue confidence prediction model, and replaces the confidence of the clues with confidence lower than the preset value obtained in the clue mining module with the predicted confidence. The multi-clue graph construction module constructs a multi-clue graph based on the confidence of various clues to establish the connection between events.

Images