Fluororubber field knowledge graph construction method based on large language model

Abstract

The invention provides a fluororubber field knowledge graph construction method based on a large language model, and the method comprises the steps: obtaining process related knowledge data of the fluororubber field, and carrying out the classification, and obtaining the classification knowledge data of the fluororubber field; based on a preset ChatGLM model, performing knowledge extraction according to the fluororubber field classification knowledge data to extract key entities and relationships; carrying out deep analysis on the key entity and the relationship by utilizing semantic correlation and context association analysis, and enhancing the connectivity of the key entity and knowledge in the relationship through semantic similarity and association strength to obtain fused knowledge data; explaining and expanding the fused knowledge data by using a ChatGLM model to obtain expanded triple data; and performing knowledge fusion on the triple data, and performing graph storage by using a Neo4j database to obtain a knowledge graph. According to the method, the knowledge graph construction efficiency is remarkably improved through automatic knowledge extraction and processing of the large language model, and the time cost of manual labeling and data processing is reduced.

Description

Technical Field

[0001] The present invention relates to the field of knowledge graph technology, and in particular to a method for constructing a knowledge graph in the field of fluororubber based on a large language model. Background Art

[0002] With the rapid development of the Internet and mobile Internet, various network service applications emerge in an endless stream, and the amount of network data shows an exponential growth. In 2012, Google first proposed the concept of knowledge graph, defining it as an expression of entities and their relationships. Knowledge graph can effectively cope with the challenges of large-scale data processing, promote data exploration, insight and decision-making in various fields, and play an important role in business, scientific research, social media and news. Knowledge graph can be divided into general knowledge graph and domain knowledge graph. General knowledge graph is oriented to open fields, and typical examples include Google's "knowledge graph", Sogou's "knowledge cube", Yago and DBpedia. However, general knowledge graph data mostly comes from Wikipedia, and it is difficult to build a complete global ontology. In contrast, domain knowledge graph focuses more on the description of professional knowledge and shows higher pertinence in specific industries, such as military, public security, transportation, medical and other fields, and is suitable for complex application analysis and decision-making assistance. The construction of this type of knowledge graph requires in-depth participation of experts, complex knowledge structure, high quality requirements and fine granularity. For example, the "Galaxy" knowledge graph, as a knowledge graph in the military field, covers multiple sources such as dark web data, Internet data, traditional databases and military books, and contains more than 100,000 equipment entity data and 330 military ontology categories. Other typical domain knowledge graphs include the IBM Watson Health medical knowledge graph, the Haizhi Xingtu financial knowledge graph and the Hisense Traffic Management Cloud Brain traffic knowledge graph.

[0003] Researchers in various industries have extensively explored various methods for building knowledge graphs. In the field of chemical engineering, although some relevant research has been done, it is still in its early stages. For example, in the intelligent question-answering system in the field of chemical safety, researchers combined the NB (Naive Bayes) algorithm and the BertBiLSTM-CRF model to build a multi-classifier to complete the question classification task, thereby building a knowledge graph question-answering system. Pranav Shetty trained the MaterialsBERT language model using 2.4 million material science abstracts and analyzed and applied the extracted data, covering the fields of fuel cells, supercapacitors, and polymer solar cells, which can be used to find material property data recorded in the abstracts. The rapid development of text mining and natural language processing (NLP) technology has made it possible to transform scientific texts into knowledge graph data sets. By applying the trained NER model to polymer abstracts and combining heuristic rules with the predictions of the NLP model, material property data can be extracted from these abstracts. Weiren Wang introduced an NLP model to automatically extract chemical composition and property data from journal articles on organic and inorganic compounds, superalloys, and aluminum alloys, and determined the optimal synthesis design route through data science to produce the desired alloy from the specified starting materials.

[0004] As modern military equipment develops towards high reliability, the demand for high-performance fluororubber and its products is becoming increasingly urgent. However, there is little basic research in the field of fluororubber, the product variety is limited, and there is a lack of systematic knowledge management, which leads to a slow development process. With its powerful semantic expression, storage and reasoning capabilities, knowledge graphs provide an effective solution for chemical data management and intelligent applications in the Internet era. As an important part of artificial intelligence technology, knowledge graphs have been widely used in professional fields such as medicine, electricity, and finance. Its advantages lie in processing huge data, building rich semantic relationships, ensuring high-quality expression, and providing user-friendly structures. Therefore, knowledge graphs can be used to express professional knowledge in fluororubber research. However, due to the complexity and diversity of knowledge in the field of fluororubber, there is currently no standardized methodology to construct relevant knowledge graphs. In addition, existing knowledge graphs do not fully reflect the correlation or similarity between entities, which makes it difficult to query relevant information. Summary of the invention

[0005] In order to overcome the shortcomings of the prior art, the purpose of the present invention is to provide a method for constructing a knowledge graph in the field of fluororubber based on a large language model. By automating knowledge extraction and processing through a large language model, the efficiency of knowledge graph construction is significantly improved, and the time cost of manual annotation and data processing is reduced.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A method for constructing a knowledge graph in the field of fluororubber based on a large language model, comprising:

[0008] Acquire process-related knowledge data in the field of fluororubber, and classify the process-related knowledge data to obtain classified knowledge data in the field of fluororubber;

[0009] Based on the preset ChatGLM model, knowledge extraction is performed according to the classified knowledge data in the fluororubber field to extract key entities and relationships;

[0010] Using semantic relevance and context association analysis to deeply analyze the key entities and relationships, and enhancing the connectivity of knowledge in the key entities and relationships through semantic similarity and association strength to obtain fused knowledge data;

[0011] Using the ChatGLM model to interpret and expand the fused knowledge data to obtain expanded triple data;

[0012] The triple data is subjected to knowledge fusion, and the graph is stored using the Neo4j database to obtain a constructed knowledge graph.

[0013] Preferably, the process-related knowledge data is obtained from professional books, professional databases and expert knowledge.

[0014] Preferably, the classified knowledge data in the field of fluororubber includes: fluororubber basic knowledge data, fluororubber formula data and fluororubber process data.

[0015] Preferably, the annotation categories of the key entities include: rubber, raw material, description, formula and embodiment.

[0016] Preferably, the annotation categories of the relationship include: performance, preparation, steps, formula, method and others.

[0017] Preferably, the key entities and relationships are deeply analyzed by using semantic relevance and context association analysis, and the connectivity of the knowledge in the key entities and relationships is enhanced by semantic similarity and association strength to obtain fused knowledge data, including:

[0018] The definition of the fluororubber keyword corresponding to the key entity is converted into a feature vector using the text2vec-large-chinese model;

[0019] Based on the feature vector, the semantic similarity between the fluororubber keywords is calculated by cosine similarity;

[0020] Constructing a keyword co-occurrence matrix based on the fluororubber keywords;

[0021] Based on the keyword co-occurrence matrix, the correlation strength between the fluororubber keywords is calculated using a point mutual information formula;

[0022] Knowledge fusion is performed according to the semantic similarity and the association strength to obtain the fused knowledge data after entity matching.

[0023] Preferably, the ChatGLM model is used to interpret and expand the fused knowledge data to obtain expanded triple data, including:

[0024] The fused knowledge data is interpreted using the ChatGLM model and preset prompt words to obtain an interpretation result;

[0025] Based on the interpretation results, generate relevant new entities and relationships according to the ChatGLM model;

[0026] The fused knowledge data is expanded according to the new entities and relationships to obtain expanded triple data.

[0027] Preferably, the triple data is fused with knowledge and stored in a graph using a Neo4j database to obtain a constructed knowledge graph, including:

[0028] Integrating the expanded triple data to obtain knowledge-fused triple data;

[0029] Based on the Neo4j database, the triple data after the knowledge fusion is stored in a graph to obtain a constructed knowledge graph.

[0030] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0031] The present invention provides a method for constructing a knowledge graph in the field of fluororubber based on a large language model, comprising: obtaining process-related knowledge data in the field of fluororubber, and classifying the process-related knowledge data to obtain classified knowledge data in the field of fluororubber; based on a preset ChatGLM model, performing knowledge extraction according to the classified knowledge data in the field of fluororubber to extract key entities and relationships; performing in-depth analysis of the key entities and relationships using semantic relevance and context association analysis, enhancing the connectivity of the knowledge in the key entities and relationships through semantic similarity and association strength, and obtaining fused knowledge data; interpreting and expanding the fused knowledge data using the ChatGLM model to obtain expanded triple data; performing knowledge fusion on the triple data, and using a Neo4j database for graph storage to obtain a constructed knowledge graph. The present invention automates knowledge extraction and processing through a large language model, significantly improving the efficiency of knowledge graph construction, and reducing the time cost of manual annotation and data processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0033] Figure 1 A flow chart of a method provided by an embodiment of the present invention;

[0034] Figure 2 A schematic diagram of a construction process provided by an embodiment of the present invention;

[0035] Figure 3 A schematic diagram of data categories provided by an embodiment of the present invention;

[0036] Figure 4 A schematic diagram comparing the previous knowledge extraction and large model knowledge extraction processes provided in an embodiment of the present invention;

[0037] Figure 5 A schematic diagram of a keyword similarity calculation process and a keyword context association calculation process provided by an embodiment of the present invention;

[0038] Figure 6 A schematic diagram of the large model knowledge identification, extraction and training process provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] The purpose of the present invention is to provide a method for constructing a knowledge graph in the field of fluororubber based on a large language model. By automating knowledge extraction and processing through a large language model, the efficiency of knowledge graph construction is significantly improved, and the time cost of manual annotation and data processing is reduced.

[0041] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0042] Figure 1 A flow chart of a method provided by an embodiment of the present invention, such as Figure 1As shown, the present invention provides a method for constructing a knowledge graph in the field of fluororubber based on a large language model, comprising:

[0043] Step 100: Acquire process-related knowledge data in the field of fluororubber, and classify the process-related knowledge data to obtain classified knowledge data in the field of fluororubber;

[0044] Step 200: Based on the preset ChatGLM model, knowledge extraction is performed according to the classified knowledge data in the fluororubber field to extract key entities and relationships;

[0045] Step 300: Performing in-depth analysis on the key entities and relationships using semantic relevance and context association analysis, enhancing the connectivity of the knowledge in the key entities and relationships through semantic similarity and association strength, and obtaining fused knowledge data;

[0046] Step 400: using the ChatGLM model to interpret and expand the fused knowledge data to obtain expanded triple data;

[0047] Step 500: perform knowledge fusion on the triple data and use the Neo4j database to store the graph to obtain a constructed knowledge graph.

[0048] In previous studies, obtaining the association information between entities mainly relies on deep learning models to extract entity relationships. However, this method is highly dependent on complex manual data labeling. To address this challenge, the latest large language model (LLM) breaks through the traditional data labeling method. LLM has demonstrated the ability to understand and respond to complex queries, and has shown excellent entity relationship extraction capabilities in new tasks, demonstrating their potential to provide a unified solution to various challenges. Through prompt word engineering, LLM can mine stored knowledge and reveal the association information between entities. In order to enrich the knowledge content in the field of fluororubber and realize information integration, this embodiment applies large model technology for the first time to construct a knowledge graph in the field of fluororubber. The construction process is as follows: Figure 2 As shown. Studies have shown that ChatGLM based on the GLM model can serve as a powerful assistant in the field of materials, helping researchers to accelerate the process of text mining and data analysis. The main research work of this embodiment includes the following aspects:

[0049] (1) The standardized storage and management methods of multi-category and multi-dimensional attribute data of fluororubber production process were studied, and a basic data platform for fluororubber production process was constructed to ensure the standardized storage and management of key data, providing reliable data guarantee for the knowledge graph in the fluororubber field.

[0050] (2) Compared with the traditional knowledge extraction method that relies on complex manual data labeling, this embodiment uses the ChatGLM model to directly extract the basic data, formula data and instance data of fluororubber, and constructs an entity relationship triple data set containing three major categories of fluororubber knowledge, thereby establishing a basic fluororubber knowledge graph.

[0051] (3) A method for calculating the semantic similarity and contextual relevance of fluororubber entity relationship knowledge using a text embedding model and a co-occurrence matrix was proposed, which effectively eliminated the ambiguity between entity relationships, helped train and fine-tune the ChatGLM model, promoted the integration and expansion of entity relationship knowledge, and further enhanced ChatGLM's professional analysis and research capabilities in the field of fluororubber.

[0052] Specifically, this embodiment locally builds ChatGLM (Chat General Language Model), which was proposed by Tsinghua University, based on the underlying architecture of the General Language Model (GLM), and pre-trained with more than 400 billion text tags. GLM is a general language model based on autoregressive filling-in-the-blank. By introducing two-dimensional position encoding and supporting predictions of arbitrary spans, GLM optimizes the pre-training process of blank filling, and performs well in natural language understanding (NLU) tasks, surpassing models such as BERT and T5. Compared with traditional position encoding methods, GLM uses RoPE (relative position encoding) technology. RoPE more accurately represents the relative distance between any two positions in a sequence by introducing a relative position encoding matrix. This method enables the model to learn and capture the relative relationship between different positions, not just the absolute position. The core concept of RoPE involves constructing a position-dependent projection matrix to calculate the attention weights, as shown below:

[0053]

[0054] Where q and k correspond to the query vector and key vector in the attention mechanism, m and n represent two positions, and R represents the projection matrix at position i. Given that Ri is usually sparse, in order to improve computational efficiency, the following method can be used:

[0055]

[0056] Where d represents the dimension of query and key, and θ is a hyperparameter. Since ChatGLM adopts RoPE technology, it has unique advantages in text generation and entity relationship extraction in the field of fluororubber.

[0057] Furthermore, in the data collection process, this embodiment cooperated with the Chengdu Documentation and Information Center of the Chinese Academy of Sciences to extensively investigate a large amount of information such as professional books, encyclopedia information resources, existing database systems, scientific and technological archives of cooperative units and expert knowledge, and sorted out the data resource directory in the field of fluororubber. The main data sources include the following aspects: 1. Professional books. For example, "Fluorine-containing Functional Materials", "Modern Organic Fluorine Chemistry: Synthesis, Reaction, Application", "Fluorine Chemical Production Technology", etc. 2. Encyclopedia information resources. Through the investigation and collection of relevant information such as the synthetic materials, production processes, product types, properties of fluororubber involved in Baidu Encyclopedia and Wikipedia, as well as the introduction of fluorine-related products on the website, news reports, etc. 3. Database system. Systematically sort out and analyze relevant information in public database systems, such as MatWeb. 4. Literature and expert knowledge. Through the collection and processing of data and information related to fluororubber in conference proceedings, patents, journals, research reports and other documents. Communicate with relevant researchers of cooperative units to understand the production process, process, formula design and other contents of fluororubber. In order to manage data more efficiently, this embodiment divides the collected data into three categories, namely, fluororubber basic knowledge data, fluororubber formula data and fluororubber process data, and manages and displays these data separately. Figure 3 shown.

[0058] Furthermore, the specific steps of the semantic relevance and context relevance technical route of this embodiment are as follows:

[0059] Due to the complexity and diversity of knowledge in the field of fluororubber, there is no standardized methodology for constructing a fluororubber knowledge graph, and there are no relationships reflecting relevance or similarity in the existing knowledge graph, which makes it difficult to query related information. Therefore, this embodiment introduces semantic relevance and context relevance to enhance the knowledge graph's query of related information.

[0060] Optionally, the steps of constructing a knowledge graph in the field of fluororubber in this embodiment are as follows:

[0061] 1) Knowledge Extraction

[0062] Based on the fluororubber data resources constructed in the early stage, this embodiment extracts entity relationships by training the large language model ChatGLM, extracts the structural information of entity relationships in domain knowledge, and constructs a fluororubber process knowledge base. The previous method usually extracts entities and their relationships from unstructured text, and then classifies the entity relationships after intermediate processing. The application of the large language model can directly output the hierarchical structure of the entities and their relationships we need. Figure 4 A comparison between traditional relationship extraction methods and large model knowledge extraction methods is shown.

[0063] There are two main steps:

[0064] Ⅰ. Named entity recognition and extraction. Knowledge extraction for process flow mainly identifies and extracts key knowledge points from process data, constructs knowledge graphs, and applies them to process optimization and intelligent control. The collected process data include the properties and definition data of fluororubber, various fluororubber formula data, and manufacturing method data of fluorine-containing elastomers. Based on these data, this embodiment extracts the key knowledge points in the fluororubber production process and forms several main entity categories. The specific information is shown in Table 1.

[0065] Table 1 Entity annotation categories

[0066]

[0067]

[0068] Ⅱ. Extraction of relations between entities. Knowledge extraction involves not only the identification of entities, but also the extraction of relations between entities. After completing the entity construction, the entity relations are studied based on the existing knowledge base data, and the association between relational semantics and entities is established. Similarly, this embodiment extracts the key knowledge points in the fluororubber production process from the data, and while extracting the entities, summarizes several main relationship categories. The specific information is shown in Table 2.

[0069] Table 2 Relationship annotation categories

[0070]

[0071] Specifically, the semantic vector model of this embodiment is as follows:

[0072] In the problem of semantic relevance, when the search query words are different from the keywords in the document but have the same meaning, the model often finds it difficult to understand the semantics and thus cannot effectively match. Therefore, this embodiment introduces the Text2Vec model. The Text2Vec model converts sentence information from keyword definitions into sentence vectors, and then uses these sentence vectors to represent word vectors, thereby enhancing the information richness contained in the vectors and improving the accuracy of similarity calculation. Text2Vec uses the text2vec-large-chinese model, which is based on the Transformer encoder architecture and can map Chinese text of any length to a fixed-length semantic vector. During the pre-training process, the model acquires deep text understanding capabilities by learning the grammatical and semantic knowledge of a large-scale Chinese corpus. After fine-tuning, the text2vec-large-chinese model can convert input Chinese sentences or documents into high-dimensional semantic vectors, fully capturing the semantic, grammatical, and word order information in the text sequence. These text vectors can be used to train downstream tasks such as text similarity calculation, text classification, and semantic matching. At the same time, it also serves as a feature representation of sentence semantics, significantly improving the overall performance of Chinese natural language processing tasks.

[0073] Furthermore, this embodiment utilizes the calculation formula of cosine similarity, which is an indicator used to measure the similarity between two vectors and is widely used in fields such as data science, information retrieval, and natural language processing. It evaluates similarity by calculating the cosine value of the angle between two vectors, capturing the similarity in direction rather than in size. This makes cosine similarity particularly suitable for comparing documents or vectors of different lengths or that require normalization. Its calculation formula is as follows:

[0074]

[0075] Among them, A i , B i The table is the i-th component of vector A and B. The calculated similarity value varies between -1 (representing completely opposite) and 1 (representing exactly the same), where 0 represents that the two vectors are orthogonal or have no correlation, and the intermediate values ​​represent different degrees of similarity or dissimilarity.

[0076] Furthermore, the present embodiment also uses a co-occurrence frequency analysis method, which believes that if two words often appear at the same time in a text, there may be some semantic or syntactic relationship between them. Therefore, by analyzing their co-occurrence, their relevance can be revealed. The method represents text data as a two-dimensional co-occurrence matrix, in which rows and columns correspond to words respectively, and the elements in the matrix represent the co-occurrence frequency between corresponding words. When two words appear in the same text at the same time, the corresponding counts in the matrix will increase. With the help of the co-occurrence matrix, the association score between words can be calculated. Point mutual information (PMI) is one of the most common association metrics.

[0077] PMI is often used to measure the correlation between two words. If the PMI value of two words is high, it means that the correlation between them is strong. The calculation formula of PMI is as follows:

[0078]

[0079] Wherein, w1 and w2 represent entity 1 and entity 2. PMI (w1, w2) represents the frequency of co-occurrence of entity 1 and entity 2. This embodiment constructs an entity co-occurrence matrix by extracting keywords from document information, and uses PMI values ​​to represent the relationship between these entities.

[0080] 2) Knowledge Fusion

[0081] Due to differences in entity representation, such as aliases, abbreviations, etc., different entity references may actually point to the same entity, causing redundancy in the knowledge graph and causing misunderstandings in subsequent use. Therefore, it is necessary to fuse knowledge from different sources. To this end, this embodiment uses a similarity-based entity matching algorithm to eliminate these problems. The core of the similarity algorithm lies in the representation of entities, which must cover the semantic information, attribute information and associated information of each entity reference as much as possible.

[0082] In terms of semantic information, the original corpus is first segmented and stop words are removed, and then word2vec is used to generate word vectors after segmentation. As for the attribute information and association information of the referent, they are represented by the relationship structure and attribute structure respectively, so as to obtain the attribute structure and relationship structure representation of the entity. Then, using these semantic vectors and the embedding vectors based on the attribute structure and relationship structure, the similarity of the entities in terms of semantics, attributes and relationships is calculated respectively, and the overall similarity between the entities is obtained by weighted summation.

[0083] Through previous research, this embodiment has defined the fluororubber keywords in detail and used the text2vec-large-chinese model to convert these definitions into feature vectors, which contain information related to the keywords. Then, the semantic similarity between the keywords is calculated by cosine similarity, such as Figure 5 shown.

[0084] The context information of keywords can also be used to represent the relationship between them. By using the fluororubber keyword, a keyword co-occurrence matrix can be constructed. Then, the PMI value between keywords is calculated using the PMI formula. If the PMI value of two keywords is greater than 2, it indicates that there is a strong correlation between them, such as Figure 5 shown.

[0085] 3) Knowledge interpretation and expansion

[0086] In order to further enrich the fluororubber knowledge base, this embodiment uses the large language model ChatGLM to interpret and expand the entity relationship. Through the real-time engineering of the large language model, the specific meaning or definition of the fluororubber keywords is obtained, so as to generate concise and clear explanations, which helps to more deeply understand the importance of these keywords in the field of fluororubber. Prompt word engineering is an emerging research field that focuses on the creation and optimization of prompts to help the large language model (LLM) solve problems in various applications and research fields more efficiently. Prompt words are natural language input sequences of LLM, which need to be customized for specific tasks and usually include multiple elements, such as instructions, background information, and input text. This embodiment completes the content expansion and generation of domain information through the prompt word project, as shown in Table 3.

[0087] Table 3 Knowledge explanation examples

[0088]

[0089] Knowledge expansion is a key means to expand the knowledge graph. By integrating knowledge expansion into the knowledge base system, the graph expansion is achieved with the help of knowledge organization and association. The filtered fluororubber entity information is passed to ChatGLM for further expansion, as shown in Table 4.

[0090] Table 4 Knowledge extension examples

[0091]

[0092] 4) Construction of knowledge graph

[0093] The method of the large model used in this embodiment in the named entity recognition and relationship extraction tasks is summarized as follows: Figure 6As shown. First, data in JSON format is prepared according to a predefined schema and the large language model ChatGLM is trained. In the second step, the preparation of additional training data is accelerated through the partially trained model and manually corrected annotations. This process may be repeated many times, and each subsequent fine-tuning will further improve the model performance. In the last step, the required information is extended from the new text to obtain the specific meaning or definition of the fluororubber keywords. This method can generate concise and clear explanations for the fluororubber keywords, and through these extended and explained contents, it helps ChatGLM to fine-tune on the complete dataset, thereby obtaining a model suitable for the fluororubber field.

[0094] Based on the above research, this embodiment has successfully extracted the properties and definition data of fluororubber through the ChatGLM large model, further helping researchers understand the basic characteristics of fluororubber. Based on the three major categories of data divided previously, the ternary data are also divided into three categories, namely, fluororubber basic knowledge ternary data, fluororubber formula ternary data and fluororubber process ternary data. Among them, the basic knowledge data provides basic support for formula design and process parameter adjustment, and some ternary data are shown in Table 5. The fluororubber formula data lists in detail the formula of low hardness fluororubber, covering various raw materials and their proportions. These data are of great significance to the optimization of production processes and product performance. Some ternary data are shown in Table 6. The manufacturing method data of fluorinated elastomers contains specific manufacturing steps and parameters, which provides a detailed operating guide for process control in the actual production process. Some ternary data are shown in Table 7.

[0095] Table 5 Properties and definition data of some fluororubbers

[0096]

[0097] Table 6 Part of the fluororubber formula data

[0098]

[0099]

[0100] Table 7 Data on the manufacturing methods of some fluoroelastomers

[0101] entity relation entity Method for producing fluorinated elastomer Example 1 step 1. Add 0.99L pure water Method for producing fluorinated elastomer Example 1 step 2. Add 0.2g emulsifier A solution Method for producing fluorinated elastomer Example 1 step <![CDATA[3. Add 1.54 g of F(CF2)5COONH4 solution]]> Method for producing fluorinated elastomer Example 1 step 4. Nitrogen replacement Example 1 of the method for producing fluorinated elastomer step 5. Stirring and heating at 600rpm at 80℃

[0102] In order to better display the data in these knowledge bases, this embodiment uses data visualization to help intuitively understand and display the knowledge structure. Through the large language model technology, the triple information extracted from the fluororubber knowledge is stored in Neo4j, thereby constructing a complete knowledge graph.

[0103] Furthermore, the model comparison experiment steps of this embodiment are as follows:

[0104] 1. Experimental evaluation indicators

[0105] This embodiment uses three evaluation indicators, namely precision, recall and F1 score, to judge the performance of the ChatGLM improved model in extracting triple relationships.

[0106] Precision. Precision is used to measure the accuracy of positive example predictions in binary classification problems. It represents the proportion of triples predicted as correct relations that are actually correct relations. A higher precision means that the model has a lower chance of false positives (i.e., incorrectly predicting an incorrect relation as a correct relation) when predicting positive examples. The formula for calculating precision is as follows:

[0107]

[0108] Among them, TP represents the number of triplets of correctly extracted relations, and FP represents the number of triplets of incorrectly extracted relations. The precision reflects the proportion of relations correctly extracted by the model among all the triple samples extracted. The higher the precision, the lower the error rate of the model in extracting relations.

[0109] Recall rate. Recall rate is one of the important indicators for evaluating the performance of classification models, and measures the model's ability to correctly extract relations. It indicates the proportion of triple samples that the model successfully extracts among the triple samples that should actually be correctly extracted. A higher recall rate means that the model has a lower chance of missing (i.e., incorrectly predicting a correct relation as an incorrect relation) when the extraction is accurate. The calculation formula for recall rate is as follows:

[0110] recall=TP / (TP+FN) (6)

[0111] Among them, TP represents the number of triplets of correctly extracted relations, and FN represents the number of triplets that misclassify correct relations as wrong relations. The recall rate reflects the proportion of the model actually extracted among all triple samples of correctly extracted relations. The higher the recall rate, the better the model can capture the triple samples of the correct relations.

[0112] F1 score. The F1 score combines precision and recall and is a comprehensive evaluation indicator. It is the harmonic mean of precision and recall and is often used to comprehensively evaluate model performance. The higher the F1 score, the better the overall performance of the model. The calculation formula for the F1 score is as follows:

[0113] F1=2*(precision*recall) / (precision+recall) (7)

[0114] The F1 score ranges from 0 to 1, with larger values ​​indicating better model performance. The F1 score is suitable for scenarios with imbalanced categories and can balance the precision and recall of the model. In some scenarios, an ideal model should have both high precision and high recall, and the F1 score provides a comprehensive evaluation capability for this purpose.

[0115] 2. Experimental results

[0116] This embodiment locally builds and tests the ChatGLM (Chat general language model), Llama2-Chat model, and ChatGPT model. ChatGLM was proposed by Tsinghua University. It is based on the underlying architecture of the general language model (GLM) and has been pre-trained with more than 400 billion text tags. Llama 2 is a large autoregressive Transformer language model (LLM) that has been pre-trained and fine-tuned, with a parameter scale ranging from 7 billion to 70 billion. Llama 2-Chat is a version fine-tuned specifically for the dialogue field. It outperforms other open source chat models in multiple open benchmarks, and has also been evaluated for usability and security. ChatGPT is a generative pre-trained model (Generative Pre-trained Transformer) developed by OpenAI. It belongs to the field of natural language processing. It is based on the Transformer deep learning architecture. Through pre-training on large-scale text data, it can understand and generate natural language and provide a human-like dialogue experience.

[0117] This example conducts comparative experiments on knowledge extraction on three large models: ChatGLM fine-tuning model, Llama 2 and ChatGPT, and takes their average scores. The experimental results are shown in Table 8.

[0118] Table 8 Average score results of knowledge extraction of large models

[0119]

[0120]

[0121] By comparing the performance of the three models in the experiment, it can be seen that on the data set in the field of fluororubber, the ChatGLM model, which adds semantic similarity and context relevance calculations, outperforms the Llama 2 and ChatGPT models in terms of precision, recall, and F1 score. This shows that the ChatGLM large language model has shown good performance in constructing knowledge graphs in the field of fluororubber.

[0122] Specifically, traditional knowledge extraction methods usually require manual query of various information, while direct extraction of entity relationships based on large language models (LLMs) has many advantages. In the current era of large models, zero-shot learning and few-shot learning have become particularly important. For large-scale complex tasks, models often require a large amount of data for training, but it is often difficult to obtain enough data. In this case, zero-shot and few-shot learning can use existing data to accelerate model training and improve the generalization ability of the model. In addition, this method can also reduce the workload of manually annotating data, significantly reducing manpower and time costs.

[0123] In the field of fluororubber, scattered knowledge points can be integrated into a systematic knowledge network through knowledge graphs, and can also be used for process optimization and intelligent control in the future. By analyzing the relationship between formula, process parameters and product performance, the optimal production conditions can be found to improve product quality and production efficiency. For example, in low-hardness fluororubber formula 2, the ratio of FC-2123, N-990, calcium hydroxide, magnesium oxide and dibutyl sebacate can be optimized and adjusted based on experimental data. With the help of manufacturing method data, the effects of different process parameters (such as temperature and stirring speed) on product performance are analyzed, and the process parameters are further adjusted to achieve the best effect. Taking Example 1 of the manufacturing method of fluorine-containing elastomer as an example, the operating conditions of each step can be optimized to improve the consistency and performance of the product.

[0124] In addition, through the associations in the knowledge graph, the product performance under different formulas and process conditions can be predicted to guide actual production. For example, by analyzing the formula and manufacturing process of fluororubber, the Mooney viscosity and glass transition temperature can be predicted, and then the process conditions can be adjusted in advance in production to avoid potential problems. These methods will greatly help R&D personnel to deeply understand and grasp the impact of process differences on product performance, more effectively guide the improvement of formula design and production processes, enhance the independent R&D capabilities of fluororubber, and promote the capacity upgrade of my country's fluororubber industry.

[0125] This embodiment integrates structured data and unstructured knowledge, and uses the large language model ChatGLM to conduct in-depth research and extraction of relevant knowledge in the field of fluororubber, focusing on the terms and concepts in the field of fluororubber. With the help of the powerful prior knowledge of ChatGLM, this embodiment successfully extracts content related to keywords from the knowledge base and comprehensively explains the keywords. On this basis, this embodiment introduces word embedding models, similarity analysis technology and co-occurrence frequency analysis methods, further revealing the complex relationship between terms and concepts in the field of fluororubber, especially in the study of semantic similarity and contextual relevance. Through the systematic processing and extraction of data in the field of fluororubber, this embodiment constructs a comprehensive and in-depth fluororubber knowledge graph. This graph can not only effectively integrate scattered data, but also realize accurate description of different entities and their relationships. Compared with traditional manual data annotation and extraction methods, the automatic extraction method based on a large language model proposed in this embodiment significantly improves data processing efficiency, reduces manpower and time costs, and improves the generalization ability of the model. The research in this example also verifies the superior performance of the ChatGLM model in processing semantic similarity and contextual relevance, and demonstrates the great potential of large models in knowledge extraction and knowledge fusion.

[0126] The knowledge graph in the field of fluororubber constructed in this embodiment not only provides an important knowledge base for scientific research in this field, but also provides strong support for related technological innovation and industrial production. Through the application of this knowledge graph, R&D personnel can better understand the production process and optimization method of formula design of fluororubber, thereby improving product quality and production efficiency. In addition, the method of this embodiment can be extended to other professional fields, further promoting the application and development of knowledge graph technology in various industries.

[0127] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0128] In this embodiment, specific examples are used to illustrate the principle and implementation of the present invention. The above description of the embodiment is only used to help understand the method and core idea of ​​the present invention; at the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for constructing a knowledge graph in the field of fluororubber based on a large language model, characterized in that: include: Acquire process-related knowledge data in the field of fluororubber, and classify the process-related knowledge data to obtain classified knowledge data in the field of fluororubber; Based on the preset ChatGLM model, knowledge extraction is performed according to the classified knowledge data in the fluororubber field to extract key entities and relationships; Using semantic relevance and context association analysis to deeply analyze the key entities and relationships, and enhancing the connectivity of knowledge in the key entities and relationships through semantic similarity and association strength to obtain fused knowledge data; Using the ChatGLM model to interpret and expand the fused knowledge data to obtain expanded triple data; The triple data is subjected to knowledge fusion, and the graph is stored using the Neo4j database to obtain a constructed knowledge graph.

2. The method for constructing a knowledge graph in the field of fluororubber based on a large language model according to claim 1, characterized in that: The process-related knowledge data is obtained through professional books, professional databases and expert knowledge.

3. The method for constructing a knowledge graph in the field of fluororubber based on a large language model according to claim 1, characterized in that: The classified knowledge data in the field of fluororubber include: fluororubber basic knowledge data, fluororubber formula data and fluororubber process data.

4. The method for constructing a knowledge graph in the field of fluororubber based on a large language model according to claim 1, characterized in that: The key entity annotation categories include: rubber, raw material, description, formula and embodiment.

5. The method for constructing a knowledge graph in the field of fluororubber based on a large language model according to claim 1, characterized in that: The annotation categories of the relationship include: performance, preparation, steps, formula, method and others.

6. The method for constructing a knowledge graph in the field of fluororubber based on a large language model according to claim 1, characterized in that: The key entities and relationships are deeply analyzed by using semantic relevance and context association analysis, and the connectivity of the knowledge in the key entities and relationships is enhanced by semantic similarity and association strength to obtain fused knowledge data, including: The definition of the fluororubber keyword corresponding to the key entity is converted into a feature vector using the text2vec-large-chinese model; Based on the feature vector, the semantic similarity between the fluororubber keywords is calculated by cosine similarity; Constructing a keyword co-occurrence matrix based on the fluororubber keywords; Based on the keyword co-occurrence matrix, the correlation strength between the fluororubber keywords is calculated using a point mutual information formula; Knowledge fusion is performed according to the semantic similarity and the association strength to obtain the fused knowledge data after entity matching.

7. The method for constructing a knowledge graph in the field of fluororubber based on a large language model according to claim 6, characterized in that: The ChatGLM model is used to interpret and expand the fused knowledge data to obtain expanded triple data, including: The fused knowledge data is interpreted using the ChatGLM model and preset prompt words to obtain an interpretation result; Based on the interpretation results, generate relevant new entities and relationships according to the ChatGLM model; The fused knowledge data is expanded according to the new entities and relationships to obtain expanded triple data.

8. The method for constructing a knowledge graph in the field of fluororubber based on a large language model according to claim 6, characterized in that: The triple data is fused and stored in a graph using a Neo4j database to obtain a constructed knowledge graph, including: Integrating the expanded triple data to obtain knowledge-fused triple data; Based on the Neo4j database, the triple data after the knowledge fusion is stored in a graph to obtain a constructed knowledge graph.

Images