A scientific field-oriented intelligent technology dynamic tracking system

Abstract

The application provides a scientific field-oriented intelligent technology dynamic tracking system, and relates to the technical field of language models, and comprises a full-automatic field literature library maintenance module, an information extraction module based on a language processing model and a technology dynamic development graph generation module; the full-automatic field literature library maintenance module is used for text conversion of screened literature by using a data processing tool and storage into a scientific field literature library; the information extraction module based on the language processing model is used for extracting attribute information of scientific field literature; and the technology dynamic development graph generation module is used for drawing and graphically presenting a technology dynamic development graph of the scientific field. Through a full-automatic crawler tool and a field paper correlation judgment model, the application realizes dynamic updating of a field literature database, and extracts and summarizes key points in literature based on a large language model, so as to meet the needs of researchers in rapidly tracking field technology development.

Description

Technical Field

[0001] The present invention relates to the field of language model technology, and in particular to a dynamic tracking system for intelligent technology in the scientific field. Background Art

[0002] Currently, the most common approach to tracking technological trends is manual research and organization by researchers, combining literature search tools, knowledge graphs, and technology trees. Existing literature search tools can automate the retrieval and download of academic literature. Open-source tools such as PySciHub and Scholarly also offer similar capabilities. The primary goal of these systems is to automatically collect relevant literature.

[0003] In terms of knowledge graphs, the ConnectedPapers website provides a graph function for literature association analysis. Users simply enter the literature information to be analyzed, and the website automatically generates a literature analysis grid. It visualizes information such as citation counts through elements such as ball size, color, and line thickness and length.

[0004] A technology tree is a graphical tool used to illustrate the evolution and innovation path of technology within a specific field. It helps researchers understand the path of technological development, key milestones, and possible future research directions.

[0005] With the rapid development of various scientific fields, technology tracking has become an essential part of researchers' daily work. Researchers need to constantly track the latest technological advances and research trends to ensure their work remains aligned with the highest standards in their field. By understanding the latest scientific developments, researchers can identify new research hotspots and breakthroughs, adjust their research directions promptly, and make informed decisions regarding research direction, technology development, and scientific project planning. However, faced with a vast amount of scientific literature and information, researchers often struggle to quickly track the latest technological advances in their field.

[0006] At present, technical summaries in various scientific fields are usually manually sorted out by researchers. For example, the field technology tree is a graphical tool for displaying technological evolution and innovation routes. This work is very laborious and time-consuming. In order to improve the efficiency of technology tracking, some literature association analysis graph tools based on representation models and text similarity have appeared on the market. However, these graphs are only drawn based on a single document and cannot widely cover the literature in the entire field. They usually only provide the context of the development of the field, which requires researchers to read the content of the article. At the same time, compared with manual sorting methods, these tools have failed to significantly speed up the speed at which researchers track technological development. Therefore, faced with a huge amount of scientific research literature and information, researchers often find it difficult to quickly track the latest technological advances in the field, and the summaries of scientific field technology are usually manually sorted out by researchers, which is very laborious and time-consuming.

[0007] Currently, no effective solutions have been proposed for the problems in related technologies. Summary of the Invention

[0008] In view of this, the present invention provides an intelligent technology dynamic tracking system for the scientific field to solve the above-mentioned problems.

[0009] In order to solve the above problems, the specific technical solutions adopted by the present invention are as follows:

[0010] An intelligent technology dynamic tracking system for scientific fields, including: a fully automated field literature library maintenance module, an information extraction module based on language processing models, and a technology dynamic development map generation module;

[0011] A fully automated domain literature maintenance module is used to screen documents that meet the preset scientific fields through a document classifier, convert the screened documents into text using data processing tools, and store them in the scientific domain literature library;

[0012] An information extraction module based on a language processing model, which is used to extract attribute information of scientific literature by using the language processing model and combining it with information used to characterize the development attributes and technical points of the scientific field;

[0013] The technology dynamic development diagram generation module is used to draw and graphically display the technology dynamic development diagram in the scientific field based on the attribute information of the extracted scientific field literature.

[0014] Preferably, the fully automated field document library maintenance module includes: a document crawling submodule, a document processing submodule and a document storage submodule;

[0015] The document crawling submodule is used to use a fully automated crawler tool to regularly crawl documents from the online document library and retain the document abstract text;

[0016] The document processing submodule is used to screen and classify the crawled documents based on the document abstract text through data processing tools, and convert the documents related to the scientific field into text form;

[0017] The document storage submodule is used to store the text-converted documents into a pre-set scientific field document library.

[0018] Preferably, the crawled literature is screened and classified by data processing tools, and the literature related to the scientific field is converted into text form, including:

[0019] For the crawled documents, use data processing tools to read the metadata of the documents;

[0020] Based on the metadata of the document and the document abstract text as the classification basis, the document classifier is used to classify the crawled documents; based on the classification results of the documents, the documents that meet the scientific field are screened out;

[0021] The screened documents that meet the scientific field are converted into text form, and the reference information and random numerical values ​​of the text documents are matched using a pre-set regular expression; the matched reference information and random numerical values ​​are deleted and replaced to obtain the text for storing in the scientific field document library.

[0022] Preferably, the document classifier is obtained by training a classification model using pre-set document titles and abstract data as a training set.

[0023] Preferably, the information extraction module based on the language processing model includes: a representation attribute extraction submodule and a technical detail extraction submodule;

[0024] The representation attribute extraction submodule is used to analyze the attributes of scientific documents based on the language processing model, and evaluate and select the optimal representation attributes of the documents based on the similarity calculation results of the texts;

[0025] The technical details extraction submodule is used to obtain the technical details of the document through technical key point analysis based on the language processing model.

[0026] Preferably, based on the language processing model, attribute analysis is performed on scientific documents, and the optimal representation attributes of the documents are evaluated and selected based on the text similarity calculation results, including:

[0027] Perform word segmentation and segmentation on scientific literature to obtain literature paragraphs;

[0028] The document paragraph is spliced ​​with the pre-set model extraction prompt words and input into the language processing model to determine whether the document paragraph contains data with preset attributes;

[0029] Perform structured data extraction on the document paragraphs containing preset attribute data to obtain structured representation attributes;

[0030] Based on the abstract text of each scientific field document, the cosine similarity is used to calculate the similarity between the abstract text and the structured representation attribute text;

[0031] According to the similarity calculation results, the confidence is calculated by multi-dimensional scoring of the summary text and attribute text, and the representation attribute with the highest confidence is selected as the optimal representation attribute.

[0032] Preferably, the formula for calculating the similarity between the document abstract text and the structured attribute text using cosine similarity is:

[0033] ;

[0034] Where, Indicates the similarity between the document abstract text and the structured attribute text;

[0035] E abstract It represents the mapping of the document abstract text encoded by the text encoding model in the vector space, also known as the embedding vector of the abstract text;

[0036] E property It represents the mapping of the structured attribute text in the vector space obtained by encoding the structured attribute text through the text encoding model, also known as the embedding vector of the structured attribute text.

[0037] Preferably, the confidence is calculated by scoring the summary text and the attribute text in multiple dimensions as follows:

[0038] ;

[0039] Indicates the similarity between the document abstract text and the structured attribute text;

[0040] Represents the multi-dimensional scoring of summary text and attribute text;

[0041] Indicates confidence;

[0042] ω 1 represents the weight of the text similarity between the summary text and the attribute text;

[0043] ω2 represents the weight of the multi-dimensional score of the summary text and the attribute text, and the confidence is weighted calculated using the above formula based on the two weights.

[0044] Preferably, based on the language processing model, through technical point analysis, the technical details of the document include:

[0045] S1. Encode the literature paragraphs and preset technical query sentences into text vectors through the embedding model;

[0046] S2. Calculate the cosine similarity between the query and each document paragraph, and sort them according to the similarity;

[0047] S3. Based on the similarity ranking results, select a preset number of document paragraphs as the initial boundary, and use the language processing model to determine whether the selected document paragraphs contain the required technical details;

[0048] S4. If the selected document paragraph contains the required technical details, the boundary is gradually expanded and S3 is repeated until the model determines that it does not contain technical details, and the final text boundary is obtained;

[0049] S5. Splice the document paragraph text and the abstract text within the final boundary, extract structured data through the language processing model, and fill in the technical details template to obtain technical details data;

[0050] S6. Construct a statement to inquire about the correctness of the technical details. Use the language processing model to determine the correctness of the technical details. If incorrect, make corrections based on the model's modification suggestions and re-execute S5 until the model verifies the correctness of the technical details.

[0051] Preferably, the technology dynamic development diagram generation module includes: a representation attribute drawing submodule, a technology detail embedding submodule and a technology detail display submodule;

[0052] The characterization attribute drawing submodule is used to draw the characterization attributes into a technology development curve in chronological order;

[0053] The technology details embedding submodule is used to embed the technology details data of scientific literature into the technology development graph;

[0054] The technical details display submodule is used to display technical details at the corresponding time nodes of the technology development diagram in the form of marks or annotations.

[0055] The beneficial effects of the present invention are:

[0056] 1. The present invention uses a fully automated crawler tool and a field paper relevance judgment model to achieve dynamic updates of the field literature database, and extracts and summarizes the key points in the literature based on a large language model. This can continuously track and summarize the latest technologies and draw a dynamic development map of the technology in the field to meet the needs of researchers to quickly track the development of field technology.

[0057] 2. The present invention realizes the dynamic maintenance of the scientific literature database, which not only can continuously track the latest technological progress, but also can ensure the timeliness and comprehensiveness of the literature database, reducing the workload of researchers to manually update and screen literature.

[0058] 3. The present invention uses a large language model to automatically extract and summarize the key technical points in the literature. This process can quickly identify and extract key technical information from a large amount of literature and integrate it into key points in tabular form. In this way, the system can efficiently extract core technical content from massive data, greatly improving the speed and accuracy of literature analysis.

[0059] 4. The present invention extracts and summarizes a large amount of scientific literature, which can quickly draw a dynamic technological development map of the field and update it based on the dynamic literature library. This not only significantly improves the efficiency of researchers in tracking technological development, but also provides a detailed technological development context, which helps to make more accurate decisions in research directions, technology development and scientific research project planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0061] Figure 1 This is a principle block diagram of a dynamic tracking system for intelligent technology in the scientific field according to an embodiment of the present invention;

[0062] Figure 2 1 is a schematic diagram of the principle of a fully automated document pool maintenance module of a dynamic tracking system for intelligent technology in the scientific field according to an embodiment of the present invention;

[0063] Figure 3 2. It is a schematic structural diagram of an optimal attribute extraction system in a dynamic tracking system of intelligent technology for scientific fields according to an embodiment of the present invention;

[0064] Figure 41 is a schematic structural diagram of technical detail extraction in a dynamic tracking system of intelligent technology for scientific fields according to an embodiment of the present invention.

[0065] In the picture:

[0066] 1. Fully automated field document library maintenance module; 2. Information extraction module based on language processing model; 3. Technology dynamic development map generation module. DETAILED DESCRIPTION

[0067] In order to enable those skilled in the art to better understand the technical solutions in this application, the following will clearly and completely describe the technical solutions in the embodiments of this application in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this application.

[0068] According to an embodiment of the present invention, a scientific field-oriented intelligent technology dynamic tracking system is provided.

[0069] The present invention will now be further described with reference to the accompanying drawings and specific embodiments. Figure 1-Figure 4 As shown, the intelligent technology dynamic tracking system for the scientific field according to an embodiment of the present invention includes: a fully automated field document library maintenance module 1, an information extraction module 2 based on a language processing model, and a technology dynamic development map generation module 3;

[0070] Fully automated domain document library maintenance module 1, used to screen documents that meet the preset scientific fields through document classifiers, convert the screened documents into text using data processing tools, and store them in the scientific domain document library;

[0071] As a preferred embodiment, the fully automated field document library maintenance module includes: a document crawling submodule, a document processing submodule and a document storage submodule;

[0072] The document crawling submodule is used to use a fully automated crawler tool to regularly crawl documents from the online document library and retain the document abstract text;

[0073] The document processing submodule is used to screen and classify the crawled documents based on the document abstract text through data processing tools, and convert the documents related to the scientific field into text form;

[0074] As a preferred embodiment, the crawled literature is screened and classified by data processing tools, and the literature related to the scientific field is converted into text form, including:

[0075] For the crawled documents, use data processing tools to read the metadata of the documents;

[0076] Based on the metadata of the document and the document abstract text as the classification basis, the document classifier is used to classify the crawled documents; based on the classification results of the documents, the documents that meet the scientific field are screened out;

[0077] As a preferred embodiment, the document classifier is obtained by using pre-set document titles and abstract data as a training set to train a classification model.

[0078] The screened documents that meet the scientific field are converted into text form, and the reference information and random numerical values ​​of the text documents are matched using a pre-set regular expression; the matched reference information and random numerical values ​​are deleted and replaced to obtain the text for storing in the scientific field document library.

[0079] The document storage submodule is used to store the text-converted documents into a pre-set scientific field document library.

[0080] Specifically, in the fully automated domain document library maintenance module, a fully automated document library maintenance strategy has been developed: a fully automated crawler tool will periodically crawl and return specific documents in the field from the online document library. Simultaneously, tools such as pymupdf and Python are used to read the document PDFs, which can read the document's metadata. This metadata is combined with the document abstract text retained during the crawling process and used as the basis for document classification. A document classifier is used to classify the returned documents, screening out documents in the corresponding scientific fields. If the classifier determines that the document is relevant to the scientific field, it is automatically added to the document library, thereby dynamically updating the scientific field document library.

[0081] The document classifier is based on the BERT model, which is a classification model trained using pre-processed data on specific document titles and abstracts, as well as data on document titles and abstracts in other fields. Finally, data processing tools such as pymupdf are used to pre-process and clean the documents, including converting the documents from PDF to text format. At the same time, regular expression templates are pre-set, and regular expressions are used to match reference information and messy values ​​in the text. The matched information is deleted and replaced using a Python program, and the final text is stored in the document library.

[0082] An information extraction module 2 based on a language processing model is used to extract attribute information of scientific literature by using the language processing model and combining it with information used to characterize the development attributes and technical points of the scientific field;

[0083] As a preferred embodiment, the information extraction module based on the language processing model includes: a representation attribute extraction submodule and a technical details extraction submodule;

[0084] As a preferred embodiment, based on the language processing model, attribute analysis is performed on scientific documents. The optimal representation attributes of the documents are evaluated and selected based on the text similarity calculation results, including:

[0085] Perform word segmentation and segmentation on scientific literature to obtain literature paragraphs;

[0086] The document paragraph is spliced ​​with the pre-set model extraction prompt words and input into the language processing model to determine whether the document paragraph contains data with preset attributes;

[0087] Perform structured data extraction on the document paragraphs containing preset attribute data to obtain structured representation attributes;

[0088] Based on the abstract text of each scientific field document, the cosine similarity is used to calculate the similarity between the abstract text and the structured representation attribute text;

[0089] As a preferred embodiment, the formula for calculating the similarity between the document abstract text and the structured attribute text using cosine similarity is:

[0090] ;

[0091] Where, Indicates the similarity between the document abstract text and the structured attribute text;

[0092] E abstract It represents the mapping of the document abstract text encoded by the text encoding model in the vector space, also known as the embedding vector of the abstract text;

[0093] E property The mapping of the structured attribute text encoded by the text encoding model in the vector space is also called the embedding vector of the structured attribute text.

[0094] According to the similarity calculation results, the confidence is calculated by multi-dimensional scoring of the summary text and attribute text, and the representation attribute with the highest confidence is selected as the optimal representation attribute.

[0095] The representation attribute extraction submodule is used to analyze the attributes of scientific documents based on the language processing model, and evaluate and select the optimal representation attributes of the documents based on the similarity calculation results of the texts;

[0096] As a preferred embodiment, the confidence is calculated by scoring the summary text and attribute text in multiple dimensions using the following formula:

[0097] ;

[0098] Indicates the similarity between the document abstract text and the structured attribute text;

[0099] Represents the multi-dimensional scoring of summary text and attribute text;

[0100] Indicates confidence;

[0101] ω 1 represents the weight of the text similarity between the summary text and the attribute text;

[0102] ω 2 represents the weight of the multi-dimensional score of the summary text and the attribute text, and the confidence is weighted calculated using the above formula based on the two weights.

[0103] The technical details extraction submodule is used to obtain the technical details of the document through technical key point analysis based on the language processing model.

[0104] As a preferred implementation, based on the language processing model and through technical analysis, the technical details of the document include:

[0105] S1. Encode the literature paragraphs and preset technical query sentences into text vectors through the embedding model;

[0106] S2. Calculate the cosine similarity between the query and each document paragraph, and sort them according to the similarity;

[0107] S3. Based on the similarity ranking results, select a preset number of document paragraphs as the initial boundary, and use the language processing model to determine whether the selected document paragraphs contain the required technical details;

[0108] S4. If the selected document paragraph contains the required technical details, the boundary is gradually expanded and S3 is repeated until the model determines that it does not contain technical details, and the final text boundary is obtained;

[0109] S5. Splice the document paragraph text and the abstract text within the final boundary, extract structured data through the language processing model, and fill in the technical details template to obtain technical details data;

[0110] S6. Construct a statement to inquire about the correctness of the technical details. Use the language processing model to determine the correctness of the technical details. If incorrect, make corrections based on the model's modification suggestions and re-execute S5 until the model verifies the correctness of the technical details.

[0111] Specifically, in the information extraction module based on the language processing model, researchers need to specify attributes that can characterize the development of a specific scientific field and the corresponding technical points as the information to be extracted from the literature. The details are as follows:

[0112] (1) Representational attribute extraction

[0113] The Jieba model and Python program are used to segment the document, calculate the number of tokens (words) in the document, and segment the document at the level of 300 tokens. The document is divided into paragraphs of 300 words each. The document paragraph to be extracted is concatenated with the pre-set prompt words extracted by the model to obtain the corresponding string fragments as the input of the language processing model. The model uses the above prompts to determine whether the segmented document paragraphs contain data (values ​​and units) related to the attribute in question. For example: "Please determine whether the current text contains the {} attribute? Please only answer yes. If it does, answer yes, and if it does not, answer no." For the document paragraphs judged to be yes, further structured data extraction is performed, that is, template filling, "The {} attribute exists in this paragraph of text. Please extract the attribute and organize it into the following structure:

[0114] {

[0115] 'type': '',

[0116] 'value': '',

[0117] 'unit': '',

[0118] }

[0119] ”

[0120] This results in structured representational attributes. The resulting representational attributes fall into two categories: optimal attributes and test attributes. The optimal attributes are selected based on the similarity between the summary and the structured representational attribute text and the confidence level of the model's weighted multi-dimensional scores for the summary and attribute text.

[0121] The similarity between attribute text and summary text is calculated based on cosine similarity. First, the text is segmented, stop words are removed, and lowercase is performed. A pretrained text encoding model is then loaded. The processed text and the model's built-in vocabulary file are then converted into the token IDs required for model input. These are the corresponding IDs of the tokens (words) in the vocabulary. Special markers [CLS] and [SEP] are added to the text. After these token IDs are input to the model, the model maps the text to numerical vectors in vector space. Cosine similarity is then calculated to represent the similarity between the two texts.

[0122] Using pre-defined model scoring prompts, the language processing model is prompted to perform a multi-dimensional scoring of the summary text and the corresponding attribute text, resulting in a similarity score. This score encompasses several dimensions: semantic similarity between the summary text and the corresponding text, consistency of referential attributes between the summary and the attribute text, and consistency between the attributes and standard representational attributes. For example, "Please rate the semantic similarity between the summary text {} and the attribute text {} extracted from the attribute text, within a range of 0-10," is a scoring prompt for one dimension. After obtaining the scores from these three perspectives, they are weighted and averaged to obtain the model's actual score for the representational attribute extraction result. This is used to assess whether the two texts describe the same attribute and whether the attribute is in the representational attribute format, ultimately determining the representational attribute in the document.

[0123] (2) Extraction of technical details

[0124] A query statement with pre-set technical points is used. For example, in the electrolyte field, the query statement "What are the components of an electrolyte that achieves a Coulombic efficiency of 99.8%?" is set as the technical point. This statement and the segmented literature are encoded using the embedding model and converted into text vectors. The text similarity is then calculated based on cosine similarity, and a list of texts sorted by similarity is returned. The top five paragraphs are selected as the first-level boundary, prompting the language processing model to determine whether the top five paragraphs contain the required technical details, that is, "Does the current text contain all {} technical points?" If so, the boundary is expanded, and the next paragraph with less similar semantics is selected as the new boundary. The above steps are repeated until the model determines that no technical details are included, which is the final text boundary.

[0125] The text of the top N paragraphs is spliced ​​with the summary text to extract structured data, fill in the technical point template, and obtain the specific data of the technical point. For example, the prompt model fills in the technical point template: "Please find the specific technical point of attribute {} based on the above text and fill it into the following template:

[0126] {

[0127] "name": "",

[0128] "consist": "",

[0129] "condiction" : "",

[0130] }

[0131] ”

[0132] The specific details of the template are based on the designated scientific field. Based on the technical point template, the name of the technical point is used as an identifier to construct inquiries about the correctness of other attributes. For example, "Based on the above text, does {name} consist of {consist}? Please indicate yes or no. If yes, no action will be taken; if no, suggestions for revision will be provided." This is used to determine the correctness of the final technical point. If the language processing model returns "no," the revision suggestions output by the language processing model are rejoined to the original text, and the structured data extraction steps for the technical point are repeated. These steps are repeated until the language processing model returns "yes," verifying the correctness of the technical point. If not, revisions are made. This process is repeated until the model verifies the correctness of the technical details.

[0133] The technology dynamic development diagram generation module 3 is used to draw and graphically display the technology dynamic development diagram in the scientific field based on the attribute information of the extracted scientific field literature.

[0134] As a preferred embodiment, the technology dynamic development diagram generation module includes: a representation attribute drawing submodule, a technology detail embedding submodule and a technology detail display submodule;

[0135] The characterization attribute drawing submodule is used to draw the characterization attributes into a technology development curve in chronological order;

[0136] The technology details embedding submodule is used to embed the technology details data of scientific literature into the technology development graph;

[0137] The technical details display submodule is used to display technical details at the corresponding time nodes of the technology development diagram in the form of marks or annotations.

[0138] Specifically, in the Technology Dynamics Development Chart Generation module, the system generates a technology dynamics chart based on representational attributes and technical details, aiming to graphically and intuitively display technological developments within a scientific field. Specifically, this module first plots the representational attributes into a technology development curve in chronological order, demonstrating the evolution of technology over time. Simultaneously, the system embeds specific technical details from relevant literature into the chart, displaying them as markers or annotations at the corresponding time nodes.

[0139] This paper proposes an intelligent technology dynamic tracking system for scientific fields, with practical application in the electrolyte field. It aims to dynamically track and analyze technological development trends in the electrolyte field. The system can dynamically access the latest literature in the field and extract key technical points, enabling dynamic tracking of technologies within the electrolyte field. The intelligent technology dynamic tracking system for the electrolyte field consists of three main modules: a fully automated domain literature maintenance module, an information extraction module based on language processing model technology, and a technology dynamic development map generation module.

[0140] The present invention constructs a fully automated dynamic maintenance subsystem for a field literature database, the main purpose of which is to construct and dynamically maintain a literature database in the field of electrolytes. Specifically, it consists of two parts:

[0141] The first is the fully automated literature crawling part. Researchers specify literature keywords in the electrolyte field. The fully automated crawling system will regularly crawl and return specific literature in related fields from the online literature library, and use OCR and Python technologies to pre-process and clean the literature.

[0142] The second is the judgment of the relevance of the literature field. The present invention implements a BERT-based electrolyte field classification model, which can judge the relevance to the electrolyte field based on the title and abstract of the literature, and process the returned literature using the classification model, thereby dynamically updating the electrolyte field literature library.

[0143] In order to facilitate understanding of the above technical solutions of the present invention, the specific implementation of the present invention in the field of electrolytes is described in detail below.

[0144] This paper builds an electrolyte information extraction system based on language processing model technology. Its main purpose is to extract properties such as coulombic efficiency, capacity retention rate, cycle life, etc. that can characterize the development of the electrolyte field and the corresponding technical points - electrolyte formula from the literature.

[0145] The specific process is as follows:

[0146] 1. Optimal Attribute Extraction Module

[0147] 1. The document text is segmented and input into the language processing model in sequence to determine whether it has the specified attribute value. If it does, it is extracted and the paragraph text corresponding to the attribute value is saved. If it does not have the attribute value, it is skipped and organized into the following format:

[0148] {

[0149] 'type': 'coulombic efficiency',

[0150] 'value': '98.2% within 450 cycles',

[0151] 'unit': None,

[0152] 'sentence_id': 2

[0153] }

[0154] 2. Filter the top five attribute values ​​and their corresponding paragraphs. Calculate the similarity between each attribute value's corresponding paragraph and the abstract. Use the language processing model to determine the degree of similarity between the attribute value's corresponding paragraph and the optimal attribute of the document mentioned in the abstract. If similar, a score of 1-10 is assigned to describe the degree of similarity; if not, a score of 0 is assigned. The confidence level of the optimal attribute is determined by combining the similarity scores and the language processing model scores using specific weights. The attribute with the highest confidence level is selected as the optimal attribute, representing the technological advancement brought about by the technology of this document.

[0155] 2. Technical Points Extraction Module

[0156] 1. Calculate the similarity between the description of the electrolyte composition and the paragraphs in the literature, return the top five most similar paragraphs, and use the language processing model to determine whether the last paragraph contains the electrolyte composition. If so, then determine the top six paragraphs, and repeat this process until the model determines that it does not contain the electrolyte composition.

[0157] 2. Combine the description of the electrolyte composition with the largest paragraph containing the electrolyte composition found by the language processing model. Use the language processing model to extract a structured expression of the electrolyte composition. An example is shown below:

[0158] {

[0159] "name": "CNE-30",

[0160] "solute": "LiFSI",

[0161] "solute concentration" : "2 M",

[0162] "solvents": "EC / DMC",

[0163] "diluents": "CTAB",

[0164] "proportion": "EC / DMC 3:7 by volume; CTAB 30mM"

[0165] }

[0166] 3. The large language model uses a preset list of questions that can verify the correctness of the answers to questions to verify the correctness of the extracted electrolyte composition. That is, a list of questions that inquire whether the electrolyte composition is correct, such as "Is the salt solute in the CNE-30 electrolyte LiFSI?", to verify whether there is an illusion of the electrolyte composition. The verification answer, the description of the electrolyte composition, and the largest paragraph containing the electrolyte composition found by the language processing model are provided to the language processing model for correction or confirmation of the answer. The cycle continues until the verification stage is passed to obtain the final electrolyte composition.

[0167] This paper constructs a module for generating a technical development diagram for the electrolyte field based on optimal properties and key technical points. Its main purpose is to graphically display the technical trends in the electrolyte field. Specifically, it consists of two parts:

[0168] The first is the technology development trend display section, which presents the technological dynamics in the electrolyte field in a visual form by drawing a scattered line graph of the technology development trend. Each point in the line graph corresponds to the electrolyte properties and technical points of a document;

[0169] The second is the technical information display section, which displays the specific technical details of the corresponding literature. Researchers can quickly browse the detailed information of each technical point, thereby fully understanding the technological development status in the electrolyte field and the technical details of the latest technologies.

[0170] In summary, with the help of the above technical solutions of the present invention, the present invention realizes the dynamic update of the field literature database through a fully automated crawler tool and a field paper relevance judgment model, and extracts and summarizes the key points in the literature based on a large language model, so that the latest technology can be continuously tracked and summarized, and a technical dynamic development diagram of the field can be drawn to meet the needs of researchers to quickly track the development of field technology. The present invention realizes the dynamic maintenance of the scientific field literature database, which not only can continuously track the latest technological progress, but also can ensure the timeliness and comprehensiveness of the literature database, reducing the workload of researchers to manually update and screen literature. The present invention utilizes a large language model to automatically extract and summarize the technical key points in the literature. This process can quickly identify and refine key technical information from a large amount of literature, and integrate it into key points in the form of a table. In this way, the system can efficiently extract core technical content from massive data, greatly improving the speed and accuracy of literature analysis. The present invention extracts and summarizes a large amount of scientific literature, can quickly draw a dynamic technological development map of the field, and update it according to the dynamic literature library. This not only significantly improves the efficiency of researchers in tracking technological development, but also provides a detailed technological development context, which helps to make more accurate decisions in research directions, technology development and scientific research project planning.

[0171] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, optical storage, etc.) containing computer-usable program code.

[0172] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A dynamic tracking system for intelligent technology in the scientific field, characterized by: include: Fully automated domain literature database maintenance module, information extraction module based on language processing model, and technology dynamic development diagram generation module; The fully automated domain document library maintenance module is used to screen documents that meet the preset scientific fields through a document classifier, convert the screened documents into text using a data processing tool, and store them in the scientific domain document library; The information extraction module based on the language processing model is used to extract attribute information of scientific documents by using the language processing model and combining it with information used to characterize the development attributes and technical points of the scientific field; The technology dynamic development diagram generation module is used to draw and graphically display the technology dynamic development diagram of the scientific field based on the attribute information of the extracted scientific field literature; The information extraction module based on the language processing model includes: a representation attribute extraction submodule and a technical details extraction submodule; The representation attribute extraction submodule is used to perform attribute analysis on scientific documents based on the language processing model, and evaluate and select the optimal representation attributes of the documents based on the text similarity calculation results; The technical details extraction submodule is used to obtain the technical details of the document through technical key point analysis based on the language processing model; The language processing model is used to analyze the attributes of scientific literature, and the optimal representation attributes of the literature are evaluated and selected based on the text similarity calculation results, including: Perform word segmentation and segmentation on scientific literature to obtain literature paragraphs; The document paragraph is spliced ​​with the pre-set model extraction prompt words and input into the language processing model to determine whether the document paragraph contains data with preset attributes; Perform structured data extraction on the document paragraphs containing preset attribute data to obtain structured representation attributes; Based on the abstract text of each scientific field document, the cosine similarity is used to calculate the similarity between the abstract text and the structured representation attribute text; Based on the similarity calculation results, the confidence is calculated by multi-dimensionally scoring the summary text and attribute text, and the representation attribute with the highest confidence is selected as the optimal representation attribute; The formula for calculating the similarity between the document abstract text and the structured attribute text using cosine similarity is: Where, Indicates the similarity between the document abstract text and the structured attribute text; E abstract Represents the mapping of the document abstract text encoded by the text encoding model in the vector space; E property Represents the mapping of the structured attribute text encoded by the text encoding model into the vector space; The confidence calculation formula is calculated by multi-dimensional scoring of summary text and attribute text: Indicates the similarity between the document abstract text and the structured attribute text; Represents the multi-dimensional scoring of summary text and attribute text; Indicates confidence; ω1 represents the weight of the text similarity between the summary text and the attribute text; ω2 represents the weight of the multi-dimensional score of the summary text and the attribute text, and the confidence is weighted calculated using the above formula based on the two weights.

2. The intelligent technology dynamic tracking system for scientific fields according to claim 1, characterized in that: The fully automated domain document library maintenance module includes: a document crawling submodule, a document processing submodule and a document storage submodule; The document crawling submodule is used to crawl documents from the online document library at regular intervals using a fully automated crawler tool and retain document abstract texts; The document processing submodule is used to screen and classify the crawled documents based on the document abstract text through data processing tools, and convert the documents related to the scientific field into text form; The document storage submodule is used to store the text-converted documents into a pre-set scientific field document library.

3. The intelligent technology dynamic tracking system for scientific fields according to claim 2, characterized in that: The data processing tools are used to screen and classify the crawled documents and convert the documents related to the scientific field into text form, including: For the crawled documents, use data processing tools to read the metadata of the documents; Based on the metadata of the document and the document abstract text as the classification basis, the document classifier is used to classify the crawled documents; based on the classification results of the documents, the documents that meet the scientific field are screened out; The screened documents that meet the scientific field are converted into text form, and the reference information and random numerical values ​​of the text documents are matched using a pre-set regular expression; the matched reference information and random numerical values ​​are deleted and replaced to obtain the text for storing in the scientific field document library.

4. The intelligent technology dynamic tracking system for scientific fields according to claim 3, characterized in that: The document classifier is obtained by training a classification model using pre-set document titles and abstract data as a training set.

5. The intelligent technology dynamic tracking system for scientific fields according to claim 1, characterized in that: Based on the language processing model, the technical details of the literature obtained through technical analysis include: S1. Encode the literature paragraphs and preset technical query sentences into text vectors through the embedding model; S2. Calculate the cosine similarity between the query and each document paragraph, and sort them according to the similarity; S3. Based on the similarity ranking results, select a preset number of document paragraphs as the initial boundary, and use the language processing model to determine whether the selected document paragraphs contain the required technical details; S4. If the selected document paragraph contains the required technical details, the boundary is gradually expanded and S3 is repeated until the model determines that it does not contain technical details, and the final text boundary is obtained; S5. Splice the document paragraph text and the abstract text within the final boundary, extract structured data through the language processing model, and fill in the technical details template to obtain technical details data; S6. Construct a statement to inquire about the correctness of the technical details. Use the language processing model to determine the correctness of the technical details. If incorrect, make corrections based on the model's modification suggestions and re-execute S5 until the model verifies the correctness of the technical details.

6. The intelligent technology dynamic tracking system for scientific fields according to claim 1, characterized in that: The technology dynamic development diagram generation module includes: a representation attribute drawing submodule, a technology detail embedding submodule and a technology detail display submodule; The characterization attribute drawing submodule is used to draw the characterization attributes into a technology development curve in chronological order; The technical details embedding submodule is used to embed technical details data of scientific field literature into the technology development map; The technical details display submodule is used to display technical details at corresponding time nodes of the technology development diagram in the form of marks or annotations.

Images