A method for discovering hot topics on the Internet in the field of science and technology
Through large language models combined with FEW-SHOT prompts and RAG technology, network hot topics were found, which solved the traditional method of traditional methods in real-time and readability, and achieved efficient and accurate compatibility of articles and themes in the field of science and technology.
Patent Information
- Application Number
- CN202411397908.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-09
- Publication Date
- 2025-01-24
- Estimated Expiration
- 2044-10-09
AI Technical Summary
It is difficult for the existing technology to effectively use large language models to discover hot topics on the Internet, especially when dealing with massive scientific and technological literature, traditional clustering methods have problems such as low real-time requirements, large impact on parameter setting, and poor readability of clustering results.
Use large models to combine Few-SHOT prompts to subdivide the science and technology field, use RAG technology for multi-level retrieval and multi-road recall, realize the single-time cluster of the article, and generate theme descriptions and names through large models to improve the accuracy of clustering categories And readability.
It has greatly reduced the cost of manual labeling, improved the accuracy of classification of scientific and technological segments, realized the rapid clustering and theme matching of the article, improved the processing efficiency and resource utilization rate.
Smart Images
Figure CN118897894B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of natural language processing, and in particular to a method for discovering hot topics on the Internet in the field of science and technology. Background Art
[0002] Hot science and technology events will not only affect the public, but also have direct or indirect impacts on enterprises and other industries. Timely grasping and analyzing hot science and technology events can help enterprises and other industries make more informed decisions and reduce unnecessary losses.
[0003] The Big Language Model is an artificial intelligence technology based on deep learning. Through a stacked neural network structure, the model learns and simulates the complex laws of human language to achieve text generation capabilities close to human levels. It can not only generate natural language text, but also deeply understand the meaning of the text and handle various natural language tasks, such as text summarization, question answering, translation, etc. However, how to use the Big Language Model for online hot topic discovery, especially for processing massive amounts of scientific and technological literature, still faces severe challenges.
[0004] First, the amount of data on the Internet and social media is huge and updates rapidly. How to filter out scientific and technological articles in subdivided fields is one of the problems faced. Secondly, how to use the knowledge of scientific and technological subdivided fields to improve the effect of topic clustering needs further research. Thirdly, some traditional clustering methods such as kmeans, mean-shift, hierarchical clustering and other mainstream clustering models require all samples to be clustered at one time and gradually iterate. If the input samples are slightly different, the clustering results may also change dramatically, which is not suitable for hot topic discovery, which requires high real-time requirements; some other clustering models, such as DBscan and single-pass clustering, only need to scan the data set once to complete the clustering, but the density parameters and similarity thresholds need to be specified in advance. These artificially set parameters have a great impact on the quality of clustering. These reasons have led to the fact that traditional clustering models are not suitable for hot topic discovery tasks. In addition, traditional clustering algorithms can only obtain clustering results, and there is no general description of the articles in each cluster, which leads to poor readability and comprehensibility of clustering results.
[0005] Therefore, in order to more accurately discover hot scientific topics, it is urgent to find a new method that can fully utilize the advantages of large language models while solving the above-mentioned problems. Summary of the invention
[0006] The purpose of the present invention is to provide a method for discovering hot topics on the Internet in the field of science and technology, so as to solve the above-mentioned problems existing in the prior art.
[0007] In order to achieve the above object, the technical solution adopted by the present invention is as follows:
[0008] A method for discovering hot topics on the Internet in the field of science and technology, comprising the following steps:
[0009] S1. Define the classification standards of science and technology fields and construct few-shot prompts;
[0010] S2, multiple media articles are input into the big model, and the big model combines the few-shot prompts to segment the input articles into scientific and technological fields;
[0011] S3. Based on the results of the scientific and technological field segmentation, the abstracts of each scientific and technological article are extracted using a large model and the article abstracts are vectorized;
[0012] S4. Classify the first article as a topic in the field, and index it according to the subdivided field category and store it in the subject vector library of science and technology subdivided fields; for the articles after the first article, use the big model to rewrite multiple article abstracts, and use RAG technology to perform multi-level retrieval and multi-way recall on multiple rewritten article abstracts to recall the Top N most similar topics from the subject vector library of science and technology subdivided fields;
[0013] S5. Use the large model to determine whether the newly input article belongs to one of the Top N most similar topics, and generate the topic description and topic name in a corresponding manner based on the judgment result.
[0014] Preferably, step S1 is specifically to subdivide the science and technology field based on the "National Key High-tech Fields", use the primary, secondary, and tertiary field contents in the text as learning examples input to the large model, and construct few-shot prompts to guide the large model to output reasonable classification results.
[0015] Preferably, in step S3, the article summary is vectorized by calling the m3e-large model.
[0016] Preferably, in step S4, for the first article:
[0017] The first article is automatically classified as a topic in its field, and the article abstract is used as the topic description. The key phrases in the article are extracted as the topic name, and indexes are established according to the subdivided field categories and stored in the science and technology subdivided field topic vector library.
[0018] Preferably, in step S4, for other articles after the first article:
[0019] Use the big model to generate and rewrite article abstracts multiple times, input multiple rewritten article abstracts, use vector cosine similarity to search the first-level fields in turn, and then use the same method to search the corresponding sub-fields, and finally find similar topics in specific sub-fields, and then recall the Top N most similar topics from the subject vector library of the technology sub-field.
[0020] Preferably, the multi-way recall is to input a plurality of rewritten article abstracts and perform multi-level retrieval simultaneously.
[0021] Preferably, step S5 is specifically as follows: the large model determines whether the newly input article belongs to one of the Top N most similar topics. If so, the large model classifies the article into the corresponding topic and regenerates the topic description and topic name based on the latest M articles of the topic; otherwise, the article is classified as a new topic in the subdivided field, and the abstract of the article is used as the description of the new topic, and the key phrases of the article are extracted as the topic name of the topic.
[0022] The beneficial effects of the present invention are: 1. The emergence capability of large models is combined with few-shot prompt engineering to greatly reduce the cost of manual annotation and improve the accuracy of classification in science and technology sub-fields. 2. The retrieval enhancement generation RAG technology is effectively applied to cluster articles published in real time on the network, realizing single-pass clustering of articles and avoiding the one-to-one comparison of each article with each existing cluster in single-pass clustering, improving the speed and accuracy of clustering, and avoiding the bias and uncertainty caused by manual intervention. 3. Multi-level indexing divides huge data and information needs into categories and organizes them at different levels to achieve more effective management and retrieval. The multi-level routing mechanism ensures that each query is efficiently directed to the most appropriate index. It not only improves processing efficiency, but also optimizes resource allocation and use, ensuring accurate matching of various queries. 4. Abstract rewriting multi-way recall improves the recall rate of subject queries and makes subject matching more accurate. 5. Use the big model to generate topic names and descriptions, making the generated topics easier to understand and more readable; generate topic descriptions based on the abstracts of multiple articles in the cluster, making the descriptions more suitable for the topic and avoiding one-sidedness. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] Figure 1 is a flow chart of a method in an embodiment of the present invention. DETAILED DESCRIPTION
[0024] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings. It should be understood that the specific implementation methods described herein are only used to explain the present invention and are not used to limit the present invention.
[0025] like Figure 1As shown, the present invention provides a method for generating hot topics in the field of science and technology with high quality and high efficiency by using a large language model few-shot prompt to segment articles in the field of science and technology, and using RAG multi-level retrieval and multi-way recall to improve single-pass clustering. This method is different from traditional text classification and topic clustering methods. This method does not require manual annotation of a large amount of training corpus, nor does it require setting the number and parameters of clusters, which greatly reduces the workload of manual intervention and the bias introduced. The method of the present invention specifically includes the following parts:
[0026] S1. Define the classification standards of science and technology fields and construct few-shot prompts.
[0027] Specifically: Based on the "National Key High-tech Fields", the science and technology fields are subdivided, and the primary, secondary, and tertiary field contents in the text are used as input to the learning samples of the large model, and few-shot prompts are constructed to guide the large model to output reasonable classification results. The emergence ability of the large model enables it to use a small number of examples in the prompt to achieve the training effect of a large amount of corpus for the small model.
[0028] S2. Input the collected articles from official media, central media, social media, etc. into the big model. The big model uses few-shot prompts to segment the input articles into scientific and technological fields.
[0029] In this embodiment, the prompt input to the large model is: Please classify [Article] into the following scientific and technological fields according to the content of the article, and only output the field name. [Field]: Electronic information technology, biological and new medical technology, aerospace technology, new material technology, high-tech service industry, new energy and energy-saving technology, resources and environmental technology, high-tech improvement of traditional enterprises. [Article]: The specific content of the article.
[0030] When classifying sub-fields, you need to modify the field name to the corresponding sub-field.
[0031] S3. Based on the results of the scientific and technological field segmentation, the large model is used to extract the abstracts of the articles and vectorize the article abstracts.
[0032] After screening out scientific articles and segmenting them by field, the abstract is extracted from each scientific article and the abstract is vectorized.
[0033] In this embodiment, the large model is directly used for processing. Article Abstract: The prompt input to the large model is: extract the article abstract for the following text. [Article]: Article content; Abstract vectorization: call the m3e-large model, input text, and output vector.
[0034] S4. Use the improved single-pass topic clustering method to perform topic clustering for each article: Specifically: classify the first article as a topic in the field, and index it according to the subdivided field category and store it in the science and technology subdivided field topic vector library; for the articles after the first article, use the large model to rewrite multiple article abstracts, and use RAG technology to perform multi-level retrieval and multi-way recall on the multiple rewritten article abstracts to recall the Top N most similar topics from the science and technology subdivided field topic vector library.
[0035] (1) For the first article: Based on the article abstract vector, the first article is automatically classified as a topic in its field, and the article abstract of the article is used as the topic description. The key phrases in the article are extracted as the topic name, and an index is established according to the subdivided field category and stored in the milvus vector database (i.e., the subject vector library of science and technology subdivision fields).
[0036] (2) For other articles after the first article: Use the large model to generate and rewrite article abstracts multiple times, input multiple rewritten article abstracts, use vector cosine similarity to search the first-level fields in turn, and then use the same method to search the corresponding sub-fields, and finally find similar topics in specific sub-fields, and then recall the top N most similar topics from the subject vector library of technology sub-fields.
[0037] In this embodiment, the first article is automatically classified as a topic in its field, and the article abstract of the article is used as the topic description, the key phrases in the article are extracted as the topic name, and the index is established according to the subdivided field category and stored in the subject vector library of the science and technology subdivision field. When the subsequent articles are subject clustered, the large model is allowed to generate and rewrite the article abstract multiple times, and the multi-level routing mechanism is used to ensure that each abstract query is efficiently guided to the most appropriate science and technology subdivision field index, and the enhanced retrieval generation RAG technology is used to replace the distance comparison between the article and each cluster in the single-pass clustering. The RAG technology recalls the most similar TOP N topics from the vector library for each rewritten article abstract. The use of enhanced retrieval to generate RAG for subject recall replaces the one-to-one comparison with the existing clusters in the single-pass clustering, and the retrieval speed of the vector library is faster, which greatly reduces the clustering time. The multi-level routing mechanism enables the query to be routed to a specific index according to the science and technology subdivision field, which not only improves the processing efficiency, but also optimizes the allocation and use of resources, ensuring accurate matching of various queries. The multi-way recall of the abstract rewrite reduces the problem of low recall rate caused by differences in expression style.
[0038] In this embodiment, the large model is directly used to rewrite the abstract. The prompt input to the large model is: rewrite the following text. [Article]: article abstract.
[0039] Multi-level search is: input article abstracts, first use vector cosine similarity to search the first-level field, then use the same method to search the corresponding sub-fields, and finally find similar topics in specific sub-fields. Multi-way recall: input multiple rewritten article abstracts and search at multiple levels at the same time, which is multi-way recall.
[0040] S5. Use the large model to determine whether the newly input article belongs to one of the Top N most similar topics, and generate the topic description and topic name in a corresponding manner based on the judgment result.
[0041] Specifically: the big model is used to determine whether the newly input article belongs to one of the Top N most similar topics. If so, the big model classifies the article into the corresponding topic and regenerates the topic description and topic name based on the latest M articles on the topic; otherwise, the article is classified as a new topic in the niche field, and the abstract of the article is used as the description of the new topic, and the key phrases of the article are extracted as the topic name of the topic.
[0042] The large model is used to determine whether an article belongs to a certain topic based on the subject description and article abstract, avoiding the manual setting of parameters such as the center of the cluster, the number of clusters, and the threshold, thereby reducing the randomness and uncertainty of the clustering results.
[0043] In this embodiment, the prompt input to the large model is: please classify [article] into one of [topic 1], [topic 2], [topic N]. If it cannot be classified into any topic, then output [other]. [article]: article content.
[0044] By adopting the above technical solution disclosed in the present invention, the following beneficial effects are obtained:
[0045] The present invention provides a method for discovering hot topics on the Internet in the field of science and technology. The method utilizes the emergence capability of a large model combined with a few-shot prompting project to greatly reduce the cost of manual annotation and improve the accuracy of classification in the subdivided field of science and technology. The retrieval enhancement generation RAG technology is effectively applied to cluster articles published in real time on the network, realize single-pass clustering of articles, and avoid the one-to-one comparison of each article with each existing cluster in the single-pass clustering, thereby improving the speed and accuracy of clustering and avoiding the bias and uncertainty caused by manual intervention. The multi-level index divides the huge data and information requirements by category and organizes them at different levels to achieve more effective management and retrieval. The multi-level routing mechanism ensures that each query is efficiently guided to the most appropriate index. Not only does it improve processing efficiency, but it also optimizes resource allocation and use, ensuring accurate matching of various queries. Abstract rewriting multi-way recall improves the recall rate of subject queries and makes subject matching more accurate. The large model is used to generate subject names and subject descriptions, making the generated topics easier to understand and more readable; the subject descriptions are generated according to the abstracts of multiple articles in the clustering cluster, so that the descriptions are more adapted to the subject and avoid one-sidedness.
[0046] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be considered as the scope of protection of the present invention.
Claims
1. A method for discovering hot topics in the field of science and technology, characterized by: The following steps are included: S1. Define the classification standards of science and technology fields and construct few-shot prompts; S2, multiple media articles are input into the big model, and the big model combines the few-shot prompts to segment the input articles into scientific and technological fields; S3. Based on the results of the scientific and technological field segmentation, the abstracts of each scientific and technological article are extracted using a large model and the article abstracts are vectorized; S4. Classify the first article as a topic in the field, and index it according to the subdivided field category and store it in the subject vector library of science and technology subdivided fields; for the articles after the first article, use the big model to rewrite multiple article abstracts, and use RAG technology to perform multi-level retrieval and multi-way recall on multiple rewritten article abstracts to recall the Top N most similar topics from the subject vector library of science and technology subdivided fields; S5. Use the large model to determine whether the newly input article belongs to one of the Top N most similar topics, and generate the topic description and topic name in a corresponding manner based on the judgment result.
2. The method for discovering hot topics in the field of science and technology according to claim 1, characterized in that: In step S3, the article summary is vectorized by calling the m3e-large model.
3. The method for discovering hot topics in the field of science and technology according to claim 1, characterized in that: In step S4, for the first article: The first article is automatically classified as a topic in its field, and the article abstract is used as the topic description. The key phrases in the article are extracted as the topic name, and indexes are established according to the subdivided field categories and stored in the science and technology subdivided field topic vector library.
4. The method for discovering hot topics in the field of science and technology according to claim 1, characterized in that: In step S4, for other articles after the first article: Use the big model to generate and rewrite article abstracts multiple times, input multiple rewritten article abstracts, use vector cosine similarity to search the first-level fields in turn, and then use the same method to search the corresponding sub-fields, and finally find similar topics in specific sub-fields, and then recall the Top N most similar topics from the subject vector library of the technology sub-field.
5. The method for discovering hot topics in the science and technology field according to claim 4, characterized in that: The multi-way recall is to input a plurality of rewritten article abstracts and perform multi-level retrieval at the same time.
6. The method for discovering hot topics in the field of science and technology according to claim 1, characterized in that: Specifically, step S5 is that the big model determines whether the newly input article belongs to one of the Top N most similar topics. If so, the big model classifies the article into the corresponding topic and regenerates the topic description and topic name based on the latest M articles on the topic; otherwise, the article is classified as a new topic in the subdivided field, and the abstract of the article is used as the description of the new topic, and the key phrases of the article are extracted as the topic name of the topic.
Citation Information
Patent Citations
Text generation method and device, computer equipment, storage medium and program product
CN118520854A
Enterprise annual report analysis method based on LLM and RAG
CN118520867A