Chinese information hotspot event generation system

By combining the method of determining the model and the Scann similarity model with the T5 model, the calculation complexity and data sparsity problems in the generation of information hotspot events are solved, and efficient and accurate hotspot event detection is achieved, which is suitable for the Chinese automotive information field.

CN118917304APending Publication Date: 2024-11-08CHEZHI HULIAN BEIJING SCI & TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411219744.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-30
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The existing information hotspot event generation technology has problems such as computational complexity, data sparsity, difficulty in interpreting results and noise impact, resulting in low efficiency and poor reliability of hotspot detection.

Method used

The judgment model is used to judge information hotspots through the cls vector and softmax algorithm, and cluster it with the Scann similarity model, and generate hotspot event themes through the T5 model, and use multiple information title vectors to add to generate top c.

Benefits of technology

It improves the detection efficiency and accuracy of hot-spot events, is suitable for the Chinese automotive information field, reduces the false alarm rate, and enhances the reliability and timeliness of the system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118917304A_ABST
    Figure CN118917304A_ABST
Patent Text Reader

Abstract

The invention belongs to the technical field of Internet, and discloses a Chinese information hotspot event generation system, which comprises the following specific steps of: 1, taking information materials generated by an automobile website in recent three days; and 2, respectively inputting the titles of the information materials into the judgment model, outputting a cls vector by the judgment model, and judging whether the information is a hot spot or not by the cls vector through softmax. The invention provides an automatic generation method for Chinese hotspot events, which comprises the following steps of: judging hotspots through an original hotspot judgment model, fusing professional knowledge in the field of automobiles in the scheme, recalling by using a vertical class similarity model, clustering by using a graph, adding a plurality of information title text vectors, averaging, and generating topic through a T5 model. In conclusion, according to the automatic generation method for the Chinese hotspot event, hotspot generation can be better carried out on the Chinese field information, and compared with traditional event extraction and graph learning, the method is more suitable for the field of Chinese automobile information.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of Internet technology, and specifically is a Chinese information hot event generation system. Background Art

[0002] Currently, all major information products display and recommend hot-spot related content on prominent pages. The recommendation and presentation of hot content has also become very diversified. The demand goal of generating hot topics is to obtain the current hot and finance-related topics in the world through rapid analysis of the latest news. The topic content includes but is not limited to corporate financial events, domestic and foreign political and military events, etc. Users can quickly perceive the current global hot topics through this functional module, providing them with public opinion intelligence when making investment decisions.

[0003] Existing technologies for generating news hot events include event extraction, graph learning, etc. Event extraction is to automatically extract structured knowledge from an unstructured text. The disadvantage is that the event type needs to be defined first, and it is difficult to cover all events manually. Graph learning hotspot algorithms are very useful in many scenarios, but they also have some disadvantages:

[0004] 1. Computational complexity: Many graph analysis algorithms, especially hotspot detection, may have high computational complexity on large-scale graphs; this may lead to slow running time and high consumption of computing resources;

[0005] 2. Data sparsity problem: For sparse graphs, hotspot detection may be affected. Data sparsity may cause unclear or inaccurate connections between nodes, thereby reducing the effectiveness of hotspot detection.

[0006] 3. Result interpretation and verification: In some cases, the output of a graph hotspot algorithm may be difficult to interpret because it may involve abstract nodes and relationships or may not directly correspond to real-world entities. In addition, verifying the accuracy of hotspot detection results may also be challenging.

[0007] 4. Noise and outliers: Hotspot detection may be affected by noise and outliers in the data. These factors may lead to false positives and reduce the reliability of the algorithm, so it needs to be improved. Summary of the invention

[0008] The purpose of the present invention is to provide a Chinese information hot event generation system to solve the problems raised in the above background technology.

[0009] In order to achieve the above-mentioned purpose, the present invention provides the following technical solution: a Chinese information hot event generation system, the specific steps are as follows:

[0010] Step 1:

[0011] Get the information materials generated by the automobile website in the past three days;

[0012] Step 2:

[0013] Use the titles of information materials to input the judgment model respectively, and the judgment model outputs the cls vector. The cls vector is passed through softmax to determine whether the information is a hot spot;

[0014] Step 3: Use the news titles that belong to the hot spots to cluster. Specifically, use the similarity model to recall each news title in Scann. The recalled news titles create new nodes in the graph and the thresholds of their corresponding edges.

[0015] Step 4:

[0016] Use multiple clustering tit les to input into the model, and use the vector addition of multiple tit les to generate topi c through the T5 model.

[0017] Preferably, in step 1, when obtaining automobile website information materials, it is necessary to ensure that stable data collection technology or tools are available, such as crawlers, API interfaces, etc., and stable and efficient data collection technology or tools, such as professional web crawlers, data interfaces, etc., can be selected.

[0018] Preferably, when the model in step 2 outputs the cls vector, it is necessary to take into account the model's pre-training method, architecture design, and characteristics of the input data to ensure that it can accurately reflect the semantic content of the text.

[0019] Preferably, in step 2, the threshold of the softmax output needs to be reasonably set according to the actual situation to accurately distinguish between hot information and non-hot information.

[0020] Preferably, in the process of recalling and clustering information titles using the similarity model in Scann as described in step 3, it is necessary to reasonably configure the parameters of Scann, such as vector dimension, search radius, etc., to optimize the recall effect and computational efficiency.

[0021] Preferably, as new information is continuously added as described in step 3, the graph needs to be dynamically updated to maintain the timeliness and accuracy of the clustering.

[0022] Preferably, the threshold in step three needs to be flexibly adjusted according to actual conditions when setting the threshold.

[0023] Preferably, when the vectors of the multiple titles are added in step 4, the vectors need to be normalized or weighted.

[0024] Preferably, in step 4, some of the tit les may be more representative or important than other tit les, and thus this can be reflected by assigning different weights.

[0025] The beneficial effects of the present invention are as follows:

[0026] This solution provides a method for automatically generating Chinese hot events. First, hot spots are determined through an original hot spot determination model. Then, this solution integrates professional knowledge in the automotive field, uses a vertical category similarity model for recall and combines it with a graph for clustering. Finally, multiple information title text vectors are added and averaged and the top ic is generated through the T5 model. In summary, the method for automatically generating Chinese hot events proposed in this solution can better generate hot spots for Chinese field information. Compared with traditional event extraction and graph learning, this method is more suitable for the field of Chinese automotive information. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a schematic diagram of hot spot identification of the present invention;

[0028] Figure 2 This is a schematic diagram of information title clustering of the present invention;

[0029] Figure 3 Generate a schematic diagram for the topic of the present invention. DETAILED DESCRIPTION

[0030] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] like Figures 1 to 3 As shown, the embodiment of the present invention provides a Chinese news hot event generation system, and the specific steps are as follows:

[0032] Step 1:

[0033] Get the information materials generated by the automobile website in the past three days;

[0034] Step 2:

[0035] Use the titles of information materials to input the judgment model respectively, and the judgment model outputs the cls vector. The cls vector is passed through softmax to determine whether the information is a hot spot;

[0036] Step 3: Use the news titles that belong to the hot spots to cluster. Specifically, use the similarity model to recall each news title in Scann. The recalled news titles create new nodes in the graph and the thresholds of their corresponding edges.

[0037] Step 4:

[0038] Use multiple clustering tit les to input into the model, and use the vector addition of multiple tit les to generate topi c through the T5 model.

[0039] First, we collect information materials generated by automobile websites in the past three days, then use the judgment model to analyze the information titles, identify hot information through cls vector and softmax algorithm, and then cluster the identified hot information titles. Use the similarity model and Scann technology to recall similar information, and build the relationship between nodes and edges in the graph. Finally, input multiple clustered titles into the model, and generate generalized hot event topics through the T5 model combined with the addition of title vectors, so as to efficiently and accurately capture and present hot information events in the automotive field.

[0040] Among them, in step one, when obtaining information materials from automobile websites, you need to ensure that you have stable data collection technology or tools, such as crawlers, API interfaces, etc. You can choose stable and efficient data collection technology or tools, such as professional web crawlers, data interfaces, etc.

[0041] By adopting appropriate and stable information acquisition methods, we can ensure that data can be acquired stably and efficiently.

[0042] Among them, when the model in step 2 outputs the cls vector, it is necessary to consider the model's pre-training method, architecture design, and input data characteristics to ensure that it can accurately reflect the semantic content of the text.

[0043] The output of CLS vectors not only helps the model learn context dependencies and vocabulary meanings, but also plays a key role in improving model efficiency and text classification tasks.

[0044] Among them, in step 2, it is necessary to reasonably set the threshold of softmax output according to the actual situation to accurately distinguish hot information from non-hot information.

[0045] By setting the softmax output threshold reasonably, we can avoid the situation where a too high threshold may cause the real hot information to be missed, while a too low threshold may introduce a large amount of non-hot information.

[0046] Among them, in step three, when using the similarity model in Scann to recall and cluster information titles, it is necessary to reasonably configure Scann parameters, such as vector dimension, search radius, etc., to optimize the recall effect and calculation efficiency.

[0047] By using Scann, similar information can be quickly found in large-scale data sets, improving the efficiency and accuracy of clustering.

[0048] In step three, as new information is continuously added, the graph needs to be dynamically updated to maintain the timeliness and accuracy of the clustering.

[0049] Through continuous updating, we can ensure the timeliness and accuracy of information.

[0050] Among them, the threshold in step three needs to be flexibly adjusted according to the actual situation when setting the threshold.

[0051] Avoid a threshold that is too high, which may cause the clustering to be too loose and unable to effectively reflect the correlation between information; a threshold that is too low may cause the clustering to be too tight and introduce unnecessary noise.

[0052] Among them, when adding the vectors of multiple titles in step 4, the vectors need to be normalized or weighted.

[0053] By normalizing or weighting the vectors, the impact of different titles on the final topic generation can be balanced.

[0054] Among them, in step 4, some tit les may be more representative or important than other tit les, so this can be reflected by assigning different weights.

[0055] By assigning different weights to different titles, the system can better identify which titles are more critical to the description of the entire hot event.

[0056] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0057] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A Chinese news hot event generation system, characterized in that: The specific steps are as follows: Step 1: Get the information materials generated by the automobile website in the past three days; Step 2: Use the titles of information materials to input the judgment model respectively, and the judgment model outputs the cls vector. The cls vector is passed through softmax to determine whether the information is a hot spot; Step 3: Use the news titles that belong to the hot spots to cluster. Specifically, use the similarity model to recall each news title in Scann. The recalled news titles create new nodes in the graph and the thresholds of their corresponding edges. Step 4: Use multiple clustered titles to input into the model, and use the T5 model to generate topics by adding the vectors of multiple titles.

2. A Chinese information hot event generation system according to claim 1, characterized in that: As mentioned in step one, when obtaining information materials from automobile websites, you need to ensure that you have stable data collection technology or tools, such as crawlers, API interfaces, etc. You can choose stable and efficient data collection technology or tools, such as professional web crawlers, data interfaces, etc.

3. A Chinese information hot event generation system according to claim 1, characterized in that: When the model described in step 2 outputs the cls vector, it is necessary to consider the model's pre-training method, architecture design, and input data characteristics to ensure that it can accurately reflect the semantic content of the text.

4. A Chinese information hot event generation system according to claim 1, characterized in that: As described in step 2, it is necessary to reasonably set the softmax output threshold according to the actual situation to accurately distinguish hot information from non-hot information.

5. A Chinese information hot event generation system according to claim 1, characterized in that: In the process of recalling and clustering news titles using the similarity model in Scann as described in step 3, it is necessary to reasonably configure Scann parameters, such as vector dimension, search radius, etc., to optimize the recall effect and computational efficiency.

6. A Chinese information hot event generation system according to claim 1, characterized in that: As described in step 3, as new information is continuously added, the graph needs to be dynamically updated to maintain the timeliness and accuracy of the clustering.

7. A Chinese information hot event generation system according to claim 1, characterized in that: The threshold in step 3 needs to be flexibly adjusted according to the actual situation when setting the threshold.

8. The Chinese information hot event generation system according to claim 1, characterized in that: When adding the vectors of multiple titles in step 4, the vectors need to be normalized or weighted.

9. A Chinese information hot event generation system according to claim 1, characterized in that: As mentioned in step 4, some titles may be more representative or important than others, so this can be reflected by assigning different weights.