Enterprise technology label extraction method and system, and storable medium

Abstract

Disclosed in the present invention are a an enterprise technology label extraction method and system, and a storable medium. The method comprises: collecting enterprise data, wherein the enterprise data comprises patent application documents of an enterprise; cleaning and pre-processing the collected enterprise data, so as to form field data; a large language model extracting an industrial chain of the enterprise by means of the field data, and generating preliminary technology labels; extracting enterprise keywords by using a machine learning algorithm and on the basis of the field data; and matching and aligning the preliminary technology labels and the enterprise keywords by means of the large language model, so as to obtain final enterprise technology labels. The method can realize highly accurate, efficient and practical technology label extraction.

Description

Enterprise technology tag extraction method, system and storage medium TECHNICAL FIELD

[0001] The present application relates to the technical field of technology tag extraction, and particularly relates to an enterprise technology tag extraction method, system and storage medium. BACKGROUND

[0002] Enterprise technology tag extraction is an important research field, which involves extracting technology-related keywords or phrases from various types of information about enterprises. This technology tag extraction is of great significance for understanding the technical capabilities, competitiveness and market positioning of enterprises.

[0003] Traditional enterprise technology tag extraction methods mainly rely on manual text analysis and keyword extraction techniques. These methods usually require text preprocessing, including word segmentation, stop word removal, etc., and then use some statistical or machine learning algorithms to identify and extract keywords. Among them, the rule-based tag extraction method identifies technical keywords and phrases through predefined language rules and patterns. This method relies on rich industry knowledge and professional background, although it can improve the relevance and accuracy of the tags, but it is time-consuming and laborious to develop effective rules, and has poor flexibility. Statistical-based methods use text analysis techniques such as TF-IDF (Term Frequency-Inverse Document Frequency) algorithm, which determines the importance of each word by calculating its frequency in the document and its rarity in the entire corpus. In addition, there are some graph-based methods, such as TextRank, which identify keywords by constructing a co-occurrence network between words. This method is simple and efficient, but may ignore the semantic relationship between words and phrases, making it difficult to identify new technical vocabulary in context.

[0004] With the development of big data and deep learning, enterprise technology tag extraction methods based on deep learning and large language models have emerged. Deep learning-based methods use the powerful expression capabilities of neural networks to automatically learn the complex features of text, allowing for more accurate keyword extraction. For example, convolutional neural networks (CNN) and recurrent neural networks (RNN) are widely used in text classification and keyword extraction tasks. In particular, the emergence of the BERT (Bidirectional Encoder Representations from Transformers) model has significantly improved the performance of enterprise technology tag extraction. The BERT model learns rich contextual information through pre-training and can achieve excellent results in various downstream tasks. Using deep neural networks can learn deep features of text, allowing for the extraction of technology tags. Deep learning methods can better capture complex language patterns and semantic information, but model training requires a large amount of computing resources and is sensitive to parameter adjustment and model selection. Methods based on large language models involve setting templates for large language model inputs, including example information, output format, and other content, and inputting the data that needs to extract keywords into the large language model, and finally outputting the enterprise technology keywords. However, the cost of training and inference of large language models is relatively high, and the length of the input context is limited; sometimes large models cannot complete the task well.

[0005] In summary, whether based on rules, statistics or dictionaries, these methods require a lot of manual involvement, such as rule-making and dictionary construction, which to some extent limits the degree of automation of these methods. These methods can only identify and extract known technical vocabulary, and may not be able to accurately identify and extract new technical vocabulary, especially low-frequency new vocabulary. With the development of technology, new technical vocabulary will continue to emerge, which requires frequent updates and maintenance of these methods, undoubtedly increasing the operating costs of enterprises. SUMMARY

[0006] To overcome the above-mentioned shortcomings, the purpose of the present application is to provide an enterprise technology tag extraction method, system and storage medium, which has high accuracy, high efficiency and strong practicality.

[0007] To achieve the above purpose, the technical solution adopted by the present application is: an enterprise technology tag extraction method, characterized by comprising:

[0008] Collecting enterprise data, the enterprise data including the patent application documents of the enterprise;

[0009] Cleaning and preprocessing the collected enterprise data to form field data;

[0010] The large language model extracts the industrial chain of the enterprise through the field data and generates a preliminary technical label;

[0011] The machine learning algorithm is used to extract the enterprise keywords according to the field data;

[0012] The large language model is used to match and align the preliminary technical label and the enterprise keywords to obtain the final enterprise technical label.

[0013] The beneficial effects of the present application are:

[0014] First, the large language model extracts the industrial chain, and then automatically generates a preliminary technical label according to the industrial chain, so that the enterprise technical label system is more rich and detailed;

[0015] Second, the machine learning algorithm can better extract the main keywords and appropriately retain the context semantic information;

[0016] Third, by using the prompt word engineering to gradually ask the large language model, semantic analysis and alignment from keywords to technical labels can be realized, which helps to improve the accuracy of technical label extraction and reduces the labor cost;

[0017] Fourth, the method does not completely rely on the machine learning algorithm or the large language model, but combines the machine learning algorithm and the large language model. According to the existing enterprise data and the prepared technical label, the prompt word engineering is used to gradually ask the large language model, so that semantic analysis and alignment from enterprise keywords to technical labels can be realized, and the problems of low accuracy, low efficiency and poor practicability of enterprise technical label extraction are solved.

[0018] Further, the collected enterprise data is cleaned and preprocessed to form field data, which specifically includes:

[0019] The enterprise data is cut into corresponding field data through regular matching;

[0020] The segment data formed by cutting is subjected to a deduplication operation, and the deduplicated field data is the final field data.

[0021] Further, the industrial chain and the preliminary technical label form a tree structure. The closer to the leaf node, the more detailed, so that the enterprise technical label system is more rich and detailed.

[0022] Further, the machine learning algorithm is used to extract the enterprise keywords according to the field data, which specifically includes:

[0023] The importance of each word in the field data is analyzed using the TF-IDF algorithm;

[0024] The main keywords are extracted by matrix decomposition and unsupervised clustering algorithm, and the context semantic information is appropriately reserved.

[0025] Further, a prompt module is set for the large language model, and the prompt template includes:

[0026] An example template shows an example of the desired output format and correct keyword extraction to the large language model, which is used to prompt and constrain the output of the model;

[0027] An instruction template includes specific instructions for extracting keywords and extracted data, which is used to prompt the large language model to extract keywords.

[0028] Further, the large language model is used to match and align the preliminary technical labels and enterprise keywords to obtain the final enterprise technical labels, which specifically include:

[0029] The input preliminary technical labels and enterprise keywords are first segmented by a segmenter and converted into Token;

[0030] The Token is input into the Embedding layer of the large language model to convert the Token into corresponding vectors;

[0031] After passing through multiple transformer layers, the context semantic information is learned, and the hidden state is constantly updated;

[0032] The input is output to the linear layer through softmax, and then the matching and alignment of the enterprise technical labels are completed.

[0033] Further, the enterprise data is obtained by a web crawler algorithm. By using the web crawler algorithm to obtain a large amount of patent application documents and other related information, the accuracy and integrity of the enterprise data can be ensured.

[0034] The application also discloses an enterprise technical label extraction system, which includes:

[0035] A data acquisition module is used to collect and store enterprise data;

[0036] A preprocessing module is used to clean and preprocess the enterprise data to form field data;

[0037] A technical label formulation module is used to extract an industrial chain from the field data and generate preliminary technical labels;

[0038] A keyword extraction module uses a machine learning algorithm to extract enterprise keywords according to the field data;

[0039] A technology label generation module matches and aligns the preliminary technology label and the enterprise keyword to obtain a final enterprise technology label.

[0040] The application further discloses a storage medium, which stores a computer program. BRIEF DESCRIPTION OF DRAWINGS

[0041] Fig. 1 is a flowchart of an embodiment of the application.

[0042] Fig. 2 is a system block diagram of an embodiment of the application. DETAILED DESCRIPTION

[0043] The preferred embodiments of the application are described in detail below with reference to the accompanying drawings, so that the advantages and features of the application can be more easily understood by those skilled in the art, and the protection scope of the application can be more clearly defined.

[0044] Embodiment one

[0045] The enterprise technology label extraction method of the application is used for accurate extraction of enterprise technology labels, and comprises the following steps, as shown in Fig. 1:

[0046] S100, collecting enterprise data.

[0047] In an embodiment, the enterprise data comprises patent application documents of the enterprise, because the patent application documents can reflect the technical direction of the enterprise.

[0048] The enterprise data collection refers to obtaining a large amount of patent application documents and other related information of the enterprise by using a network crawler algorithm for the enterprise which needs to extract technology labels. By using the existing network crawler technology, the patent application documents and other public data can be quickly obtained.

[0049] S200, cleaning and preprocessing the collected enterprise data to form field data.

[0050] The cleaning and preprocessing specifically comprises regular matching of the crawled enterprise data, cutting the enterprise data into corresponding field data, and performing a deduplication operation on the field data formed by the cutting. The deduplicated field data is the final field data. The field data comprises patent title, patent content, web address, etc. The field data is the key information extracted from the enterprise data, which facilitates subsequent label extraction.

[0051] In this embodiment, a large amount of patent application documents and other related information are obtained by using a network crawler algorithm, and data cleaning and preprocessing are performed, so that the accuracy and integrity of the enterprise data can be ensured, and the accuracy and efficiency of the enterprise technology label extraction can be improved.

[0052] S300, the large language model extracts the industrial chain of the enterprise through the field data and generates a preliminary technology label.

[0053] The field data is input into the trained large language model, which preliminarily judges the technical industrial chain of the enterprise and generates different preliminary technology labels. The industrial chain and the preliminary technology label form a tree structure, and the closer to the leaf node, the more detailed, making the enterprise technology label system more rich and detailed. When the large language model extracts the industrial chain, it will automatically generate a preliminary technology label according to the industrial chain.

[0054] In one embodiment, the large language model (LLM) is a deep learning model trained based on massive text data. It can not only generate natural language text, but also deeply understand the meaning of the text. By setting templates to train the large language model, the trained large language model can directly extract the corresponding industrial chain according to the field data in step S200, and generate a preliminary technology label according to the industrial chain.

[0055] In this embodiment, the large language model preliminarily judges the technical industrial chain and generates different preliminary technology labels, forming a tree structure, making the enterprise technology label system more rich and detailed. The corresponding relationship between the preliminary technology label and the industrial chain has been formulated.

[0056] S400, using a machine learning algorithm, extracting enterprise keywords according to field data.

[0057] This step completes the extraction of enterprise keywords. For the prepared enterprise data, the method of combining statistics and machine learning is used to extract the keywords that can reflect the enterprise technology label. First, the importance of each word in the field data is analyzed using the TF-IDF algorithm. Then, through matrix decomposition and unsupervised clustering algorithm, the main keywords are extracted, and the context semantic information is appropriately retained. The field data can be each patent application document, that is, the importance of each word in each patent application document is analyzed.

[0058] Using the TF-IDF algorithm, the enterprise keywords according to the field data include:

[0059] S41, use the jieba library to segment the title and abstract information of the field data, define a stop word list to remove some punctuation marks, numerical strings and other data.

[0060] S42, create a TF-IDF vectorizer to fit and train these texts, and convert the title and abstract texts into a TF-IDF matrix.

[0061] S43, dimension reduction is performed on the TF-IDF matrix using an NMF (Nonnegative Matrix Factorization) algorithm to find a relationship matrix between the documents and the topics (document-topic matrix) and a matrix of word items contained in each topic (topic-word item matrix).

[0062] The document-topic matrix and the topic-word item matrix form an NMF component.

[0063] S44, the matrix obtained by the NMF algorithm is clustered using a k-means algorithm to obtain clustering labels and clustering centers of the documents.

[0064] S45, the clustering centers, the NMF component, and the documents are vectorized, the cosine similarity between each clustering center and the NMF component is calculated, and the several words with the highest similarity are selected as representative words of the topic, i.e., enterprise keywords.

[0065] In an embodiment, when the enterprise data is less, in order to preserve semantic information, no excessive processing is performed, and when the enterprise data is more and miscellaneous, the above steps are adopted. In the process of extracting enterprise keywords, different processing methods are adopted for enterprise data of different sizes, which not only preserves the semantic information of a small amount of data but also improves the processing efficiency of a large amount of data. In addition, through TF-IDF analysis, matrix decomposition, and unsupervised clustering algorithm, etc., the main keywords can be better extracted, and the context semantic information can be appropriately preserved.

[0066] In an embodiment, step S300 and step S400 are performed synchronously.

[0067] S500, the preliminary technical labels and the enterprise keywords are matched and aligned through a large language model to obtain the final enterprise technical labels.

[0068] According to the characteristics of the large language model and the task requirements, a suitable prompt template is set. The set prompt template includes two parts, an example template and an instruction template. The example template shows the expected output format and the example of correctly extracting keywords to the large language model, which is used to prompt and constrain the output of the model; and the instruction template includes specific instructions for extracting keywords and extracted data, which is used to prompt the large language model to extract keywords. The two parts of the template are combined to form a final template, which is input into the large language model.

[0069] The large language model is loaded, and the preliminary technical labels and the enterprise keywords are matched and aligned to obtain the final enterprise technical labels, which specifically include:

[0070] S51, first tokenize the input preliminary technical tags and enterprise keywords through tokenizer and convert them into Token.

[0071] S52, Token is input into the Embedding layer of large language model to convert these Token into corresponding vectors.

[0072] S53, after passing through multiple transformer layers, learn the context semantic information and constantly update the hidden state.

[0073] S54, input into the linear layer to output the results through softmax, and then complete the matching and alignment of enterprise technical tags.

[0074] In this embodiment, semantic analysis and alignment are performed using a large model, which helps to improve the accuracy of technical tag extraction and reduces the labor cost.

[0075] The enterprise technical tag extraction method in this embodiment does not completely rely on the method of machine learning algorithm or large language model, but combines the method of machine learning algorithm and large language model. According to the existing enterprise data and the prepared technical tags, the large language model is gradually inquired using the prompt word engineering, which can realize semantic analysis and alignment from enterprise keywords to technical tags, and solve the problems of insufficient accuracy, low efficiency and insufficient practicality of enterprise technical tag extraction.

[0076] In one embodiment, the enterprise data of enterprise A is collected, the relevant information extracted and the selectable industry chain are provided in the prompt word according to the title and detailed data of the achievement of enterprise A crawled, so that the large language model can preliminarily judge the industry chain of the company A. For example, for the achievement "a fast detachable hand-held support for auxiliary stable infusion puncture needle insertion", the large language model judges that it belongs to "biological medicine industry chain" and generates relevant preliminary technical tags according to "biological medicine industry chain", such as "medical instrument", "supporting rod", "stable rod", "auxiliary device" and the like. Then, the enterprise data is processed by using TF-IDF algorithm, and the key word list is extracted through the steps of tokenization, word frequency calculation and corpus construction. Then, the large language model is used to design a template, the achievement title and detailed data are input, the key word list is combined, and the final technical tags are generated, so as to complete the matching of keywords and technical tags, that is, the classification of technical tags. The existing enterprise technical tag extraction result is analyzed, the keyword extraction algorithm and the large model prompt word are adjusted, and the iteration optimization is repeated. The finally generated enterprise technical tags are ['stable rod', 'auxiliary device'].

[0077] Embodiment two

[0078] The embodiment of the present specification also provides an enterprise technology label extraction system, which adopts any one of the enterprise technology label extraction methods provided in embodiment one. The enterprise technology label extraction system comprises a data acquisition module, a preprocessing module, a technology label formulation module, a keyword extraction module and a technology label generation module.

[0079] The data acquisition module is used to collect and save enterprise data, and obtain a large number of patent application documents and other related information of the enterprise through a web crawler algorithm.

[0080] The preprocessing module is used to clean and preprocess the enterprise data to form field data. The cleaning and preprocessing specifically include regular matching of the crawled enterprise data, cutting the enterprise data into corresponding field data, and de-duplication of the cut field data. The de-duplicated field data is the final field data.

[0081] The technology label formulation module extracts an industrial chain and generates a preliminary technology label through the field data. The industrial chain and the preliminary technology label form a tree structure, and the closer to the leaf node, the more detailed, so that the enterprise technology label system is more rich and detailed.

[0082] The keyword extraction module extracts enterprise keywords according to the field data by using a machine learning algorithm. First, the importance of each word in the field data is analyzed by using a TF-IDF algorithm. Then, the main keywords are extracted by using matrix decomposition and unsupervised clustering algorithm, and the context semantic information is appropriately retained.

[0083] The technology label generation module matches and aligns the preliminary technology label and the enterprise keywords to obtain the final enterprise technology label.

[0084] Embodiment three

[0085] The embodiment of the present specification also provides a computer readable storage medium, which stores a computer program. When the computer program is executed, any one of the enterprise technology label extraction methods provided in the embodiment one is realized.

[0086] It is to be appreciated that the above-described logic and / or steps of a flow diagram, represented, for example, in a flow chart or otherwise described herein, can be embodied in non-transitory machine- executable instructions, which can be used to cause a general purpose or special purpose processor or computer to perform functions that constitute one or more steps of the logic and / or steps represented by the flow diagram. Accordingly, it should be understood that the logic and / or steps of a flow diagram can be implemented in hardware, software, or a combination of hardware and software, as desired. When implemented in software, the functions can be stored on or transmitted over as one or more instructions or code on a computer-readable medium. Computer-readable media include computer data storage media or communication media including any medium that facilitates transfer of a computer program from one place to another. In this manner, computer-readable media generally can be viewed as a computer-readable memory. By way of example, and not limitation, computer-readable media can include RAM, ROM, EEPROM, CD-ROM or other optical disc storage, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer. Also, any connection is properly termed a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using a coaxial cable, fiber optic cable, or twisted pair, then the coaxial cable, fiber optic cable, or twisted pair are included in the definition of medium. Disk and disc, as used herein, include compact

[0087] It should be understood that aspects of the application can be implemented in hardware, software, firmware or combinations thereof, as desired. In the above embodiments, various steps or methods can be implemented in software or firmware that is stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, any of the following techniques, known in the art, or combinations thereof, can be used: discrete logic circuitry having logic gates for implementing logic functions upon data signals, application specific integrated circuits having appropriate combinational logic gates, programmable gate arrays (PGA), field programmable gate arrays (FPGA), and so forth.

[0088] The above embodiments are only for illustrating the technical concept and characteristics of the present application, and the purpose is to enable those skilled in the art to understand the content of the present application and implement it, and cannot limit the protection scope of the present application, and any equivalent changes or modifications made according to the spirit and essence of the present application should be covered within the protection scope of the present application.

Claims

1. A method for extracting enterprise technology tags, characterized in that: include: Collect enterprise data, including the enterprise's patent application documents; The collected enterprise data is cleaned and preprocessed to form field data; The large language model extracts the enterprise's industrial chain and generates preliminary technology tags through the field data; Using machine learning algorithms, enterprise keywords are extracted based on the field data; The initial technical tags and corporate keywords are matched and aligned using a large language model to obtain the final corporate technical tags.

2. The method for extracting enterprise technology tags according to claim 1, characterized in that: The collected enterprise data is cleaned and preprocessed to form field data, specifically including: The enterprise data is matched using regular expressions, and the enterprise data is then segmented into corresponding field data. The segment data formed by the segmentation is deduplicated, and the deduplicated field data is the final field data.

3. The method for extracting enterprise technology tags according to claim 1, characterized in that: The aforementioned industrial chain and initial technology labels form a tree-like structure.

4. The method for extracting enterprise technology tags according to claim 1, characterized in that: Using machine learning algorithms, the extraction of enterprise keywords based on the aforementioned field data specifically includes: Use the TF-IDF algorithm to analyze the importance of each word in the field data; The main keywords are extracted using matrix factorization and unsupervised clustering algorithms, while retaining appropriate contextual semantic information.

5. The method for extracting enterprise technology tags according to claim 1, characterized in that: A prompting module is set for the large language model, and the prompting template includes: Example templates, which show the large language model the expected output format and examples of correctly extracted keywords, serve to prompt and constrain the model's output; The instruction template includes specific instructions for extracting keywords and the data to be extracted, which is used to prompt the large language model to extract keywords.

6. The method for extracting enterprise technology tags according to claim 1, characterized in that: By matching and aligning the preliminary technical tags and enterprise keywords using a large language model, the final enterprise technical tags are obtained, specifically including: First, the input preliminary technical tags and enterprise keywords are segmented by a word segmenter and converted into tokens; The tokens are input into the embedding layer of the language model and transformed into corresponding vectors. Through multiple transformer layers, the contextual semantic information is learned and the hidden state is continuously updated; The input is fed into a linear layer and the output is processed through softmax, thereby completing the matching and alignment of enterprise technology tags.

7. The method for extracting enterprise technology tags according to claim 1, characterized in that: The enterprise data was obtained through web crawler algorithms.

8. A system for extracting enterprise technology labels, characterized in that: include: A data acquisition module, which is used to collect and store enterprise data; A preprocessing module is used to clean and preprocess enterprise data to form field data; A technology labeling module, which extracts industry chain data from field data and generates preliminary technology labels; The keyword extraction module utilizes machine learning algorithms to extract enterprise keywords based on field data. The technology tag generation module matches and aligns preliminary technology tags with enterprise keywords to obtain the final enterprise technology tags.

9. A storage medium, characterized in that: It stores a computer program, which, when executed, implements the enterprise technology label extraction method as described in any one of claims 1-8.

Images