A material synthesis data extraction method and system based on knowledge-enhanced large models

Abstract

This invention discloses a method and system for extracting materials synthesis data based on a knowledge-enhanced large-scale model, relating to the field of artificial intelligence. The method includes: retrieving literature data; constructing materials synthesis knowledge text by performing data denoising and OCR text conversion; fine-tuning the basic large-scale model using LoRA, introducing a domain instruction dataset for fine-tuning learning, and determining the extraction large-scale model; performing semantic segmentation and quantization on the materials synthesis knowledge text, constructing a multi-level retrieval architecture, and enhancing the retrieval by combining it with a materials science knowledge base to determine the enhanced knowledge text; constructing a data extraction prompt, combining the enhanced knowledge text and the materials synthesis knowledge text, inputting them into the extraction large-scale model, and performing knowledge extraction processing. This addresses the problem of insufficient accuracy in data extraction in existing technologies, achieving improved accuracy. Simultaneously, it can replace manual work in literature analysis and extraction, and domain knowledge association, providing support for materials synthesis process recommendations.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to a method and system for extracting material synthesis data based on a knowledge-enhanced large model. Background Technology

[0002] Accurate extraction of materials synthesis data is crucial for accelerating materials research and development, reducing costs, and promoting innovation in materials science. Efficiently and accurately acquiring materials synthesis-related data from massive amounts of literature is a critical problem that urgently needs to be solved in the field of materials science research. Currently, the main methods for addressing this problem are based on traditional rules or deep learning models that require extensive labeled data for training, to extract data from materials synthesis literature. However, current methods suffer from insufficient understanding of the complex semantics and professional knowledge in materials science due to the difficulty in obtaining large amounts of labeled data, and struggle to handle the diverse expressions and implicit information in the literature.

[0003] At present, the extraction of material synthesis data faces technical problems, such as the difficulty in obtaining a large amount of labeled data, resulting in insufficient accuracy, completeness, and generalization ability of the extracted data. Summary of the Invention

[0004] This application provides a method and system for extracting materials synthesis data based on a knowledge-enhanced large-scale model. It employs data denoising and OCR text conversion on materials synthesis literature data to generate structured materials synthesis knowledge text. The basic large-scale model is fine-tuned using LoRA and combined with a domain instruction dataset for further fine-tuning to obtain a dedicated extraction model. The knowledge text undergoes semantic segmentation and vectorization, a multi-level retrieval architecture is constructed and integrated with a materials science knowledge base to generate enhanced retrieval knowledge text. A data extraction prompt template is designed, and the enhanced knowledge text and the original knowledge text are input into the extraction model. Through knowledge extraction processing, structured materials synthesis data is finally output. These techniques address the technical problems of existing materials synthesis data extraction methods, such as the difficulty in obtaining large amounts of labeled data, leading to insufficient accuracy, completeness, and generalization ability. This method effectively improves the accuracy, completeness, and generalization ability of data extraction.

[0005] This application provides a method for extracting materials synthesis data based on a knowledge-enhanced large model, comprising: retrieving literature data on materials synthesis, constructing materials synthesis knowledge text by performing data denoising and OCR text conversion; fine-tuning the basic large model using LoRA, introducing a domain instruction dataset for fine-tuning learning, and determining the extraction large model; traversing the materials synthesis knowledge text, performing semantic segmentation and quantization, constructing a multi-level retrieval architecture, combining it with a materials science knowledge base for retrieval enhancement, and determining the enhanced knowledge text; constructing a data extraction prompt, combining the enhanced knowledge text and the materials synthesis knowledge text, inputting it into the extraction large model, performing knowledge extraction processing, and determining the extracted materials synthesis data.

[0006] In a possible implementation, a material synthesis knowledge text is constructed, and the following processing is performed: by constructing a search query, a targeted search of the literature pool is performed to determine the literature data; the literature data is parsed, and if the literature data is in text format, text denoising preprocessing is performed to determine the material synthesis knowledge text, wherein the preprocessing method includes at least special element transformation and redundancy denoising; if the literature data is in graph format, first-order processing based on OCR text conversion and second-order processing based on text denoising are performed to determine the material synthesis knowledge text.

[0007] In a possible implementation, a domain instruction dataset is introduced for fine-tuning learning to determine the extraction large model, and the following processes are performed: designing prompt words based on material synthesis terminology; constructing an instruction dataset based on the prompt words, wherein the instruction dataset is in the format of a structured triple of instruction-input-output; and fine-tuning the basic large model based on the instruction dataset to determine the extraction large model.

[0008] In a possible implementation, the following processing is performed: a transfer call is made to the base model, a LoRA module is inserted into the base model, and fine-tuning is performed using the instruction dataset; wherein the learning steps include: freezing the weights of the base model, using the LoRA module as the training target, using a preset convergence as the training iteration condition, introducing a learning rate, and performing iterative training learning steps based on the instruction dataset.

[0009] In a possible implementation, the enhanced knowledge text is determined, and the following processing is performed: The material synthesis knowledge text is segmented into blocks using contextual semantic awareness to determine a knowledge block set, wherein complete technical semantic units serve as the segmentation constraint; the knowledge block set is traversed, embedded words are introduced and vectorized to determine a semantic vector library, wherein the embedded words are domain-enhanced feature words; a candidate set is retrieved from the semantic vector library using vector similarity retrieval, and text semantic reordering is performed to determine the text ranking result, wherein those in the semantic vector library that satisfy vector similarity are taken as a candidate set, and ranking is constrained by the weight of technical keywords in the material synthesis domain; based on the text ranking result, knowledge query and text enhancement processing are performed to determine the enhanced knowledge text.

[0010] In a possible implementation, based on the text sorting results, knowledge query and text enhancement processing are performed to determine the enhanced knowledge text. The following processing is then performed: traversing the text sorting results, extracting technical keywords, and determining the query vector; based on the query vector, database distribution and retrieval are performed to determine highly relevant texts that meet a preset relevance, wherein the database is the source database for knowledge enhancement; semantic alignment is performed between the highly relevant texts and the knowledge block set, and additional information is generated to determine the enhanced knowledge text, wherein the additional information includes at least background information and focus areas.

[0011] In a possible implementation, a data extraction prompt is constructed, and the following processing is performed: based on the material synthesis process, key feature words of material synthesis are extracted; the key feature words are integrated with role and task prompt words to determine the data extraction prompt; wherein, the role-based task prompt word integration includes at least material synthesis feature prompt words, special data prompt words, and return result prompt words.

[0012] In a possible implementation, knowledge extraction processing is performed to determine the material synthesis extraction data, and the following processing is performed: the material synthesis knowledge text, the data extraction prompt, and the enhanced knowledge text are fused to determine the fusion input data; the fusion input data is imported into the extraction model for knowledge extraction processing, and the material synthesis extraction data is output.

[0013] In a possible implementation, the following processing is performed: associating with an external domain knowledge base, aligning the material synthesis extraction data with domain knowledge, compensating for missing fields through knowledge reasoning, and performing logical verification and correction through domain rules, which serves as the material synthesis extraction result; wherein, the alignment method includes entity disambiguation and coreference resolution.

[0014] This application also provides a material synthesis data extraction system based on a knowledge-enhanced large model, comprising: a material synthesis knowledge text construction module, used to construct material synthesis knowledge text by retrieving literature data on material synthesis and performing data denoising and OCR text conversion; a large model fine-tuning module, used to perform LoRA fine-tuning on the basic large model, introduce a domain instruction dataset for fine-tuning learning, and determine the extraction large model; a knowledge enhancement module, used to traverse the material synthesis knowledge text, perform semantic segmentation and quantization, construct a multi-level retrieval architecture, combine with a materials science knowledge base for retrieval enhancement, and determine the enhanced knowledge text; and a knowledge extraction processing module, used to construct a data extraction prompt, combine the enhanced knowledge text and the material synthesis knowledge text, input them into the extraction large model, perform knowledge extraction processing, and determine the material synthesis extracted data.

[0015] This application proposes a method and system for extracting materials synthesis data based on a knowledge-enhanced large-scale model. First, it retrieves literature data on materials synthesis, performs data denoising and OCR text conversion to construct materials synthesis knowledge text. Then, it fine-tunes the basic large-scale model using LoRA, introducing a domain instruction dataset for fine-tuning learning to determine the extraction large-scale model. Next, it traverses the materials synthesis knowledge text, performs semantic segmentation and quantization, constructs a multi-level retrieval architecture, and enhances the retrieval by combining it with a materials science knowledge base to determine the enhanced knowledge text. Finally, it constructs a data extraction prompt, combines the enhanced knowledge text and the materials synthesis knowledge text, inputs it into the extraction large-scale model, performs knowledge extraction processing, and determines the extracted materials synthesis data. This achieves the technical effect of improving the accuracy, completeness, and generalization ability of data extraction. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1 This is a flowchart illustrating a method for extracting material synthesis data based on a knowledge-enhanced large model, as provided in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of a material synthesis data extraction system based on a knowledge-enhanced large model, provided in an embodiment of this application.

[0019] Figure labeling: Material synthesis knowledge text construction module 10, large model fine-tuning module 20, knowledge enhancement module 30, knowledge extraction and processing module 40. Detailed Implementation

[0020] The above description is merely an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below.

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application will be provided in conjunction with the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] In the following description, references to "some embodiments" describe a subset of all possible embodiments; however, it is understood that "some embodiments" may be the same or different subsets of all possible embodiments and may be combined with each other without conflict. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only.

[0023] This application provides a method for extracting material synthesis data based on a knowledge-enhanced large model, such as... Figure 1 As shown, the method includes:

[0024] Step S100: By retrieving literature data on material synthesis, and by performing data denoising processing and OCR text conversion, a knowledge text on material synthesis is constructed.

[0025] Specifically, materials synthesis literature refers to academic documents, experimental reports, and other textual materials containing information such as descriptions of the materials synthesis process, process parameters, and material properties. OCR text conversion refers to the process of converting text content in images into an editable text format using optical character recognition technology.

[0026] Based on keyword combinations in the field of materials synthesis, such as "(material synthesis) AND (process parameters)", searches are conducted in academic databases (e.g., Web of Science, Materials Science Citation Index). A web crawler script is written to repeatedly send HTTP requests to retrieve search results pages and extract literature links. For each literature link, a request is sent to retrieve the literature content, which is then saved as a local file (e.g., HTML, PDF, image format). For text data files, the content is read line by line, and regular expressions are used to remove noise, including special formatting, special characters, headers, footers, references, and other irrelevant information. For image files, image preprocessing is performed using OpenCV (e.g., grayscale conversion, binarization), followed by text recognition using an OCR tool (e.g., Tesseract-OCR software). The processed text content is then stored in a materials synthesis knowledge text database (e.g., in JSON format, including fields such as literature ID, title, author, and text content).

[0027] In one possible implementation, a knowledge text on materials synthesis is constructed. Step S100 further includes step S110, which involves constructing a search query and performing a targeted search of a literature pool to determine literature data. Specifically, a Boolean logic combination search query is constructed based on keywords and phrases in the field of materials synthesis. For example, keywords such as "material synthesis," "process parameters," and "material properties" are combined with logical operators (such as AND, OR, and NOT). The search query is then executed in specified academic databases (such as Web of Science, Springer Materials, and the Materials Science Citation Index) to obtain relevant literature data. The Python requests library can be used to send an HTTP request to retrieve the search results page and parse and extract the literature links, storing them in a local file (such as CSV or JSON format).

[0028] Step S120: Parse the literature text of the said literature data. If the literature data is in text format, perform text denoising preprocessing to determine the material synthesis knowledge text, where the preprocessing methods at least include special element conversion and redundant noise reduction. Specifically, download the literature content from the obtained literature link and save it as a local text file. If the literature data is in text format, perform structured parsing on the text-format literature, identify and segment the logical parts such as the title, author, abstract, body, charts, references, etc. of the literature. Use regular expressions to remove special formats (such as HTML tags, LaTeX formulas) in the text, remove punctuation marks and special characters in the text, identify and remove the header, footer, and reference parts. For example, the header and footer are usually located at the top and bottom of each page and can be extracted and deleted by analyzing the text pattern; the reference part usually starts with "References" or "参考文献" and can be removed by splitting the text. Convert special formats (such as tables, formulas) in the text into a processable format (such as converting a table into JSON or CSV format, converting a formula into LaTeX format). For example, use the pandas library to convert a table into a DataFrame and save it as a CSV file; use the latex library to convert a formula into LaTeX format and save it as a separate file. Save the denoised text as the material synthesis knowledge text (such as JSON format, including fields such as literature ID, title, author, text content, etc.).

[0029] Step S130: If the literature data is in graphic format, perform a first-order processing based on OCR text conversion and a second-order processing based on text denoising to determine the material synthesis knowledge text. Specifically, download the literature content in graphic format from the obtained literature link and save it as a local graphic file (such as PNG, JPG format). Use the OpenCV library for image preprocessing (such as grayscale conversion, binarization, denoising), and use an OCR tool (such as Tesseract-OCR) to recognize the text content in the image as text. Perform denoising processing on the text after OCR recognition, and use regular expressions to remove redundant information and correct OCR recognition errors. For example, correct "teqmperture" to "temperature". Save the denoised text as the material synthesis knowledge text (such as JSON format, including fields such as literature ID, title, author, text content, etc.).

[0030] Step S200: Perform LoRA fine-tuning on the base large model, introduce the domain instruction dataset for fine-tuning learning, and determine the extraction large model.

[0031] Specifically, the base model refers to a model with general language understanding capabilities, pre-trained on large-scale text data, such as BERT and GPT, which serves as the basis for fine-tuning. LoRA fine-tuning is a model fine-tuning method that inserts low-rank matrix modules (LoRA modules) into the Transformer layer of the original model, training only the low-rank parameters to adapt to specific domain tasks, thus reducing computational resource consumption. Domain instruction datasets are datasets containing instruction-response pairs within a specific domain, used to guide models in learning knowledge and tasks within that domain.

[0032] Load the base large model using the transformers library, add a LoRA layer to the Transformer layer of the model, and add a low-rank matrix to the original weight matrix through matrix factorization. Extract domain-specific instruction-response pairs from materials science databases (such as MaterialsProject) and labeled data to form a domain instruction dataset. For example, the instruction could be: "Extract the reaction temperature and time in the following synthesis steps: heat the precursor at 300°C for 2 hours"; the response could be: "Temperature: 300°C, Time: 2 hours". Input the domain instruction dataset into the model for training, and optimize the LoRA parameters using the cross-entropy loss function. Set appropriate batch size (e.g., 32), learning rate (e.g., 1e-4), and number of training epochs (e.g., 3 epochs). Evaluate the model's accuracy, recall, and other metrics on a materials science test set, and adjust the LoRA configuration or data augmentation strategies (e.g., data augmentation, regularization, etc.) based on the results. Finally, compress and deploy the resulting large extracted model.

[0033] In one possible implementation, a domain instruction dataset is introduced for fine-tuning learning to determine the extraction model. Step S200 further includes step S210, designing prompt words based on materials synthesis terminology. Specifically, materials synthesis terminology is extracted from materials science databases (such as Materials Project and ICSD), including process parameters such as reaction temperature, time, and pressure, as well as material properties such as chemical formulas and crystal structures. Standardized prompt templates are constructed based on these terms, for example: "Please extract {parameter type} from the following materials synthesis text and return it in JSON format," where {parameter type} can be replaced with specific parameter categories such as temperature and time. At least three different natural language expressions are designed for each term; for example, "reaction temperature" can correspond to "synthesis temperature," "heating temperature," and "thermal condition," ensuring semantic coverage of the prompt words. Finally, the designed prompt words are stored as a JSON file in a structure of "parameter type - prompt template - thesaurus."

[0034] Step S220: Based on the prompt words, construct an instruction dataset, wherein the instruction dataset is in the format of a structured triplet of instruction-input-output. Specifically, collect material synthesis description texts from materials science literature and databases, and construct instruction fields for each text sample using the prompt template generated in step S210. For example, the instruction is "Please extract the reaction temperature and reaction time from the following text," the input field is the original text "Anneal the TiO2 precursor at 450℃ for 3 hours," and the output field is structured data {"temperature":"450℃","time":"3 hours"}. Apply different prompt templates to the same text sample to generate multiple training data, for example, change the parameter type to generate the instruction "Please extract the reaction environment." Clean all data samples, removing records whose output does not conform to the JSON format or whose parameter values ​​are abnormal. The final dataset contains multiple instruction-input-output triples and is stored in JSONL format.

[0035] Step S230: Based on the instruction dataset, fine-tune the basic large model to determine the extraction large model. Specifically, load a pre-trained BERT or GPT model as the basic large model, insert a LoRA adaptation module into the encoder layer of the model, and set the rank of the LoRA matrix. During training, freeze all parameters of the basic model and optimize only the weights of the LoRA module and the output layer. Use the input field from the instruction dataset as the model input and the output field as the target output, and optimize using the cross-entropy loss function. Set the initial learning rate and batch size, and adopt a linear learning rate decay strategy. Monitor the model's generalization ability to unseen prompt templates on the validation set. The final extraction large model can accept new material synthesis text input and output structured process parameters and material property data.

[0036] In one possible implementation, step S230 further includes: transferring the basic large model, inserting a LoRA module into the basic large model, and fine-tuning the learning with the instruction dataset; wherein the learning step includes: freezing the weights of the basic large model, using the LoRA module as the training target, using a preset convergence as the training iteration condition, introducing a learning rate, and performing iterative training learning steps based on the instruction dataset.

[0037] Specifically, pre-trained basic large models such as BERT and GPT are selected. A LoRA module is inserted into the Transformer layer of the model, and a low-rank matrix is ​​added to the original weight matrix through matrix factorization. By leveraging the advantages of transfer learning, general language knowledge is transferred to the material synthesis data extraction task.

[0038] The weights of the basic large model are frozen, and the learning rate, batch size, number of training epochs, and preset convergence threshold (e.g., loss function value less than 0.01) are set. The LoRA module is used as the training target, and training is conducted iteratively, using the input-output mapping relationship in the instruction dataset for supervised learning. In each iteration, the model processes the input text and generates a predicted output, which is compared with the true response to calculate the loss function. The LoRA module weights are updated through backpropagation. Training continues until the loss function converges to the preset threshold, ensuring the model stably learns domain knowledge.

[0039] Step S300: Traverse the material synthesis knowledge text, perform semantic segmentation and quantization, construct a multi-level retrieval architecture, combine with the materials science knowledge base for retrieval enhancement, and determine the enhanced knowledge text.

[0040] Specifically, semantic chunking refers to dividing text into multiple blocks with complete semantic units based on its semantic features, for better understanding and processing of the text. Text vectorization refers to converting text into a fixed-dimensional vector representation, enabling computation and comparison in a mathematical space for easier retrieval and analysis. A multi-level retrieval architecture is a hierarchical retrieval system architecture that progressively filters out the most relevant text or information to the query through multiple stages of retrieval and sorting.

[0041] Based on the semantic features of the synthesized text, such as keywords and sentence structure, the text is segmented into fine-grained semantic blocks. For example, the sentence segmentation and keyword extraction functions of the spaCy library can be used to segment the text into multiple blocks according to semantic coherence. Each block contains a complete technical semantic unit (such as a description of a synthesis step). Each block of text is then encoded using a domain-enhanced word embedding model (such as BERT-MAT, GPT-MOL, etc.) to obtain a fixed-dimensional vector representation, which is then stored in a vector library.

[0042] A multi-level retrieval architecture is constructed using information retrieval techniques such as inverted indexes. For example, at the first level, categorized retrieval is performed based on the type of semantic blocks (e.g., experimental method blocks), while at the second level, more refined retrieval is performed based on the similarity between vectors. Simultaneously, a materials science knowledge base is integrated; for instance, when a material name is retrieved, information such as the material's properties and synthesis methods is queried using a knowledge graph to enhance the knowledge text and determine the enhanced knowledge text.

[0043] For example, the material synthesis text "Dissolve the precursor in a solvent, stir evenly, heat to 80℃ and react for 12 hours, then cool to room temperature" is segmented into three semantic blocks: "Dissolve the precursor in a solvent," "Stir evenly and heat to 80℃ and react for 12 hours," and "Then cool to room temperature." The BERT-MAT ​​model is used to encode the segmented text "Stir evenly and heat to 80℃ and react for 12 hours," resulting in a 768-dimensional vector representation. A large number of material synthesis text vectors are stored in a vector database. After inputting a query vector, a candidate set of text blocks semantically similar to the query is first recalled. Then, a reordering model optimizes the ranking, ultimately returning the most relevant enhanced knowledge text, such as "Heating to 80℃ and reacting for 12 hours is a key synthesis step that affects the crystallinity of the material."

[0044] In one possible implementation, the enhanced knowledge text is determined, and step S300 further includes step S310, which involves segmenting the material synthesis knowledge text into blocks using contextual semantic awareness to determine a knowledge block set, wherein complete technical semantic units are used as the block segmentation constraint. Specifically, text content is read from a stored material synthesis knowledge text library, and the text is processed by sentence segmentation and word segmentation. The text is segmented into sentences using spaCy's sentence segmentation function. Keywords and entities in each sentence are extracted using the TF-IDF algorithm or spaCy's named entity recognition function. Based on the keywords and sentence structure, the sentences are combined into semantic blocks to ensure that each semantic block expresses a complete technical semantic unit. For example, consecutive sentences describing the same synthesis step are divided into a semantic block. The segmented semantic blocks are organized into a knowledge block set and stored in JSON format, containing fields such as "semantic block ID" and "semantic block content".

[0045] Step S320: Traverse the knowledge block set, introduce embedded words and vectorize them to determine the semantic vector library, where embedded words are domain-enhanced feature words. Specifically, load a pre-trained domain-enhanced model, such as BERT-MAT, using Hugging Face's transformers library. Extract embedded words from the domain expert knowledge base or domain corpus and organize them into a vocabulary. These embedded words are key terms in the field of materials synthesis, such as "hydrothermal synthesis," "sol-gel method," "reaction temperature," and "purity." Traverse the knowledge block set, process each semantic block in the knowledge block set, input each semantic block into the word embedding model, and generate corresponding semantic vectors. Simultaneously, incorporate embedded words as auxiliary features into the vectors to enhance the domain specificity of the semantic vectors. Store the generated semantic vectors in the vector library and use the FAISS library to build an efficient vector retrieval index.

[0046] Step S330 involves using vector similarity retrieval to recall candidate sets from the semantic vector library, performing text semantic rearrangement, and determining the text ranking result. Specifically, vectors in the semantic vector library that satisfy the vector similarity criteria are grouped into a candidate set, and ranking is constrained by the weights of technical keywords in the field of materials synthesis. Specifically, the cosine similarity between all vectors in the semantic vector library and the query vector is calculated, and vectors that meet a similarity threshold (e.g., 0.7) are grouped into a candidate set. The weights of technical keywords in the field of materials synthesis are determined using the TF-IDF algorithm or domain expert annotation; for example, keywords such as "reaction temperature" and "crystal transformation" have higher weights than ordinary words. The semantic vectors in the candidate set are weighted and ranked according to the keyword weights to ensure that text more relevant to the query intent is displayed first. The semantic blocks in the knowledge block set are rearranged according to the ranked vector order, and semantically related semantic blocks are combined to form a preliminary text ranking result.

[0047] Step S340: Based on the text sorting result, perform knowledge query and text enhancement processing to determine the enhanced knowledge text. Specifically, the sorted text is used as a basis, and a knowledge query is performed in conjunction with a materials science knowledge base. For example, when a key process parameter is retrieved, the normal range of the parameter and its impact on material properties are queried using a knowledge graph. Semantic enhancement is performed on the rearranged semantic blocks to supplement missing process parameters or material properties and correct erroneous information. At the same time, based on the domain knowledge base, background information and focus areas related to user needs are generated, such as supplementing background information on material properties and application scenarios, as well as suggestions for process optimization or material application. The rearranged, supplemented, and corrected semantic blocks are integrated into enhanced knowledge text and stored in JSON format, containing fields such as "enhanced knowledge text ID," "enhanced knowledge text content," "background information," and "focus areas." For example, the sorted text "Heating to 80°C and reacting for 12 hours is a key synthetic step" can be enhanced to "Heating to 80°C and reacting for 12 hours is a key synthetic step. This temperature range is applicable to most organic synthesis reactions and can affect the crystallinity and purity of the material. It is recommended to optimize temperature control to improve material performance."

[0048] In one possible implementation, based on the text ranking results, knowledge query and text enhancement processing are performed to determine the enhanced knowledge text. Step S340 further includes step S341, which involves traversing the text ranking results, extracting technical keywords, and determining the query vector. Specifically, keyword extraction is performed on each semantic block. The TF-IDF algorithm is used to calculate the term frequency-inverse document frequency value to filter out high-frequency and important technical keywords. SpaCy is used for part-of-speech tagging, selecting nouns, verbs, etc., as candidate keywords. Combining the TF-IDF values ​​and part-of-speech tagging results, the most representative keywords, such as "hydrothermal synthesis" and "reaction temperature," are selected. The technical keywords are input into a pre-trained word embedding model to generate corresponding vectors, and the average is calculated to obtain the query vector.

[0049] Step S342: Based on the query vector, perform database distribution and retrieval to determine highly relevant texts that meet a preset relevance score. The database in question is a knowledge-enhanced source database. Specifically, analyze the technical keywords in the query vector, identify its respective domain subclass (e.g., "ceramic synthesis"), and distribute the query vector to the corresponding sub-database. In the target sub-database, calculate the cosine similarity between the query vector and the text vector. Set a relevance threshold (e.g., 0.8) to filter out texts that meet the criteria. The target sub-database is a knowledge-enhanced source database that stores text vectors enhanced with domain knowledge. These text vectors integrate entity information and semantic relationships from the materials science knowledge base, enabling them to more accurately reflect professional knowledge in the field of materials synthesis.

[0050] Step S343: Semantically align the highly relevant text with the knowledge block set and generate additional information to determine the enhanced knowledge text. The additional information includes at least background information and a focus area. Specifically, sequence alignment is performed between the highly relevant text and the knowledge block set to identify similar parts and correct inconsistencies. Background information such as material properties and applications is supplemented based on the domain knowledge base. For example, if the original text lacks performance parameters of the synthesized material, information such as its strength and conductivity is supplemented from the knowledge base. User needs and usage scenarios are analyzed to generate a focus area. For example, if the user is concerned with optimizing the synthesis process, the focus area is suggestions for optimizing process parameters; if they are concerned with material applications, the focus area is the material's application scenarios and advantages. The background information and focus areas are integrated into the knowledge block set to generate enhanced knowledge text. This is stored in JSON format, containing the original text content, background information, and focus area fields.

[0051] Step S400: Construct a data extraction prompt, combine the enhanced knowledge text and the material synthesis knowledge text, input them into the extraction model, perform knowledge extraction processing, and determine the material synthesis extraction data.

[0052] Specifically, a data extraction prompt is a prompt or instruction designed to guide a large extraction model to extract specific information from text, clarifying the model's task and output format.

[0053] Design structured prompts that include role / task information and process characteristics, based on material synthesis process and attribute information. For example, a prompt could be designed as: "You are an expert in the field of materials synthesis. Please extract information such as pretreatment steps, synthesis steps, post-treatment steps, synthesis equipment, environmental solvents, reaction time, and reaction temperature from the following text: {input text}".

[0054] The enhanced knowledge text and material synthesis knowledge text are preprocessed and integrated into a format suitable for model input. The model API is invoked via an HTTP request, sending the prompt and input data as the request body, and setting request parameters (such as maximum generation length, Temperature, etc.). The model's returned JSON data is received, and the material synthesis information within it is parsed.

[0055] In one possible implementation, a data extraction prompt is constructed. Step S400 further includes step S410, which extracts key feature words for material synthesis based on the material synthesis process. Specifically, the material synthesis process flow is analyzed and broken down into a sequence of synthesis steps for the target material. For example, "hydrothermal synthesis of ceramic materials" is divided into steps such as precursor dissolution, sol formation, gel aging, drying, and calcination. Preliminary key feature words are extracted, combining the core verbs and nouns of each step into feature words, such as "precursor dissolution conditions" and "calcination temperature and time." Feature elements are refined by combining domain knowledge. Starting from materials science theory, feature elements closely related to the process steps are identified, such as "reaction atmosphere," "cooling rate," and "product crystal form." Text mining methods such as TF-IDF and word frequency statistics are used to screen high-frequency professional terms from domain literature to supplement the feature word library. All key feature words are classified and organized into a structured list, labeled with their corresponding process steps and feature types.

[0056] Step S420 involves integrating the key feature words using role and task prompts to determine the data extraction prompt. The role-based task prompt integration includes at least material synthesis feature prompts, special data prompts, and return result prompts. Specifically, role and task prompts are defined, for example, using the prompt "You are a material synthesis expert and need to accurately extract key synthesis information from the following text" to clarify the model's role and task. Key feature words extracted in step S410 are then embedded sequentially, such as "Extract precursor dissolution conditions, including solvent type, dissolution temperature, and stirring rate." For specific expressions of material properties, prompts are added, such as "Identify and convert different representations of material purity (e.g., mass percentage, molar percentage)." Return result prompts are defined, for example, specifying the result format as JSON and defining fields such as "reaction temperature (unit: °C), reaction time (unit: hours), and product morphology description." All the above prompts are combined to form a logically coherent and complete prompt. The extraction accuracy and completeness are verified on a test dataset, and the prompt word expression and sequence are adjusted based on the results.

[0057] In one possible implementation, knowledge extraction processing is performed to determine the material synthesis extraction data. Step S400 further includes step S430, which merges the material synthesis knowledge text, the data extraction prompt, and the enhanced knowledge text to determine the fusion input data. Specifically, the material synthesis knowledge text, the data extraction prompt, and the enhanced knowledge text are read separately, and each text is cleaned and formatted, including removing redundant information and unifying the encoding format. A fusion strategy is determined based on task requirements. For example, the original knowledge text is provided first, followed by supplementary information from the enhanced text, and finally, a prompt is attached to clarify the extraction requirements. The text is then integrated according to a predetermined strategy. For example, the material synthesis knowledge text is used as a base, interspersed with background information and focus points from the enhanced knowledge text, and finally, the data extraction prompt is attached.

[0058] Step S440: Import the fused input data into the extraction model for knowledge extraction processing, and output the material synthesis extraction data. Specifically, load the extraction model and set appropriate parameters, such as setting the temperature value to 0.1, to reduce the randomness of the output. Input the fused input data into the model to start the knowledge extraction process. The model performs inference based on the fused input data and outputs preliminary extraction results. Verify the format and accuracy of the output results. If they do not meet the requirements, analyze the reasons and correct the fused input data or model parameters. Organize the verified output results into a structured format (such as JSON, CSV) as the final material synthesis extraction data.

[0059] In one possible implementation, the method further includes: associating with an external domain knowledge base, aligning the material synthesis extraction data with domain knowledge, compensating for missing fields through knowledge reasoning, and performing logical verification and correction through domain rules, which is then used as the material synthesis extraction result. The alignment method includes entity disambiguation and coreference resolution.

[0060] Specifically, entity information is extracted from a selected external domain knowledge base, and entities such as materials and process parameters are identified in the material synthesis extraction data. The identified entities are matched and linked with entities in the knowledge base. Entity disambiguation and coreference resolution techniques are used to analyze the context, eliminate entity ambiguity, resolve coreference issues, and determine unique entities.

[0061] A knowledge graph is constructed based on data extracted from material synthesis and a knowledge base. By using inference algorithms and contextual information, possible values ​​of missing fields are inferred and the inferred values ​​are filled into the data to enhance data integrity.

[0062] Based on knowledge in the field of materials synthesis, logical rules are defined. The extracted materials synthesis data is input into the rule engine for logical consistency verification. Based on the verification results, logical errors in the data are corrected to ensure that the data conforms to domain logic, ultimately yielding the extracted materials synthesis results.

[0063] This application employs data denoising and OCR text conversion on materials synthesis literature data to generate structured materials synthesis knowledge text. Pre-training is performed on a basic large-scale model, and guided learning is achieved through LoRA fine-tuning technology combined with domain instruction datasets and prompting engineering to obtain a dedicated extraction large-scale model. Semantic segmentation and vectorization are performed on the knowledge text, a multi-level retrieval architecture is constructed and integrated with a materials science knowledge base to generate enhanced retrieval knowledge text. A data extraction prompt template is designed, and the enhanced knowledge text and the original knowledge text are input together into the extraction large-scale model. Through knowledge extraction processing, structured materials synthesis extracted data is finally output. These techniques solve the technical problems of insufficient accuracy, completeness, and generalization ability in existing materials synthesis data extraction due to the difficulty in obtaining large amounts of labeled data, thus achieving the technical effect of improving the accuracy, completeness, and generalization ability of data extraction.

[0064] In the above text, refer to Figure 1 A method for extracting material synthesis data based on a knowledge-enhanced large model according to an embodiment of the present invention is described in detail. Next, reference will be made to... Figure 2 This invention describes a material synthesis data extraction system based on a knowledge-enhanced large model according to an embodiment of the present invention.

[0065] A material synthesis data extraction system based on a knowledge-enhanced large model, according to an embodiment of the present invention, addresses the technical problem of insufficient accuracy, completeness, and generalization ability in existing material synthesis data extraction methods due to the difficulty in obtaining large amounts of labeled data, thereby improving the accuracy, completeness, and generalization ability of data extraction. The system comprises: a material synthesis knowledge text construction module 10, a large model fine-tuning module 20, a knowledge enhancement module 30, and a knowledge extraction processing module 40.

[0066] The material synthesis knowledge text construction module 10 is used to construct material synthesis knowledge text by retrieving literature data on material synthesis and performing data denoising and OCR text conversion; the large model fine-tuning module 20 is used to fine-tune the basic large model using LoRA, introduce domain instruction datasets for fine-tuning learning, and determine the extraction large model; the knowledge enhancement module 30 is used to traverse the material synthesis knowledge text, perform semantic segmentation and quantization, construct a multi-level retrieval architecture, combine with the materials science knowledge base for retrieval enhancement, and determine the enhanced knowledge text; the knowledge extraction processing module 40 is used to construct a data extraction prompt, combine the enhanced knowledge text and the material synthesis knowledge text, input them into the extraction large model, perform knowledge extraction processing, and determine the material synthesis extracted data.

[0067] The specific configuration of the material synthesis knowledge text construction module 10 will be described in detail below. As mentioned above, the material synthesis knowledge text construction module 10 may further include: a targeted retrieval unit for constructing a retrieval formula and performing a targeted retrieval of the literature pool to determine the literature data; a text denoising preprocessing unit for parsing the literature data, and if the literature data is in text format, performing text denoising preprocessing to determine the material synthesis knowledge text, wherein the preprocessing method includes at least special element conversion and redundancy denoising; and a second-order processing unit for performing first-order processing based on OCR text conversion and second-order processing based on text denoising if the literature data is in graph format to determine the material synthesis knowledge text.

[0068] The specific configuration of the large model extraction and determination module 20 will be described in detail below. As mentioned above, a domain instruction dataset is introduced for fine-tuning learning to determine the large model extraction. The large model fine-tuning module 20 may further include: a prompt word design unit for designing prompt words based on material synthesis terminology; an instruction dataset construction unit for constructing an instruction dataset based on the prompt words, wherein the instruction dataset is in the format of a structured triple of instruction-input-output; and a fine-tuning learning unit for fine-tuning learning the basic large model based on the instruction dataset to determine the large model extraction.

[0069] The fine-tuning learning unit may further include: transferring the basic large model, inserting the LoRA module into the basic large model, and performing fine-tuning learning with the instruction dataset. The learning steps include: freezing the weights of the basic large model, using the LoRA module as the training target, using a preset convergence as the training iteration condition, introducing a learning rate, and performing iterative training learning steps based on the instruction dataset.

[0070] The specific configuration of the enhanced knowledge text determination module 30 will be described in detail below. As mentioned above, the enhanced knowledge text determination module 30 may further include: a text segmentation unit for segmenting the material synthesis knowledge text into blocks through contextual semantic perception, and determining a knowledge block set, wherein complete technical semantic units are used as segmentation constraints; a vectorization unit for traversing the knowledge block set, introducing embedded words and vectorizing them to determine a semantic vector library, wherein the embedded words are feature words for domain enhancement; a text semantic reordering unit for retrieving candidate sets from the semantic vector library using vector similarity retrieval, performing text semantic reordering, and determining the text ranking result, wherein the semantic vector library members that meet the similarity are grouped into a candidate set, and the ranking is constrained by the weight of technical keywords in the material synthesis domain; and a text enhancement processing unit for performing knowledge query and text enhancement processing based on the text ranking result to determine the enhanced knowledge text.

[0071] The text enhancement processing unit further includes: a query vector determination subunit for traversing the text sorting results, extracting technical keywords, and determining a query vector; a database distribution and retrieval subunit for distributing and retrieving data from a database based on the query vector, and determining highly relevant texts that meet a preset relevance level, wherein the database is the source database for knowledge enhancement; and an additional information generation subunit for semantically aligning the highly relevant texts with the knowledge block set and generating additional information to determine the enhanced knowledge text, wherein the additional information includes at least background information and a focus of attention.

[0072] The specific configuration of the knowledge extraction processing module 40 will be described in detail below. As mentioned above, to construct the data extraction prompt, the knowledge extraction processing module 40 may further include: a key feature word extraction unit for extracting key feature words of material synthesis based on the material synthesis process; and a content integration unit for integrating the key feature words with role and task prompt words to determine the data extraction prompt, wherein the role-based task prompt word integration includes at least material synthesis feature prompt words, special data prompt words, and return result prompt words.

[0073] The knowledge extraction processing module 40 may further include: a fusion input data determination unit for fusing the material synthesis knowledge text, the data extraction prompt, and the enhanced knowledge text to determine fusion input data; and a knowledge extraction processing unit for importing the fusion input data into the extraction model, performing knowledge extraction processing, and outputting the material synthesis extraction data.

[0074] The system may further include: a missing field compensation module for associating with an external domain knowledge base, aligning the material synthesis extraction data with domain knowledge, compensating for missing fields through knowledge reasoning, and performing logical verification and correction through domain rules, which serves as the material synthesis extraction result. The alignment method includes entity disambiguation and coreference resolution.

[0075] The material synthesis data extraction system based on a knowledge-enhanced large model provided in this invention can execute the material synthesis data extraction method based on a knowledge-enhanced large model provided in any embodiment of this invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0076] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0077] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A method for extracting material synthesis data based on a knowledge-enhanced large model, characterized in that, The method includes: By retrieving literature data on material synthesis and performing data denoising and OCR text conversion, a knowledge text on material synthesis is constructed. LoRA fine-tuning is performed on the basic large model, and the domain instruction dataset is introduced for fine-tuning learning to determine the extracted large model. The material synthesis knowledge text is traversed, semantic segmentation and quantization are performed, a multi-level retrieval architecture is constructed, and retrieval enhancement is performed in combination with the materials science knowledge base to determine the enhanced knowledge text. Construct a data extraction prompt, combine the enhanced knowledge text and the material synthesis knowledge text, input the extraction model, perform knowledge extraction processing, and determine the material synthesis extraction data; This includes introducing a domain instruction dataset for fine-tuning learning to determine the optimal large model, including: Design prompts based on materials synthesis terminology; Based on the prompt words, construct an instruction dataset, wherein the instruction dataset is in the format of a structured triple of instruction-input-output; Based on the instruction dataset, the basic large model is fine-tuned to determine the extracted large model; Specifically, the basic large model is called during migration, and a LoRA module is inserted into the basic large model for fine-tuning learning using the instruction dataset; The learning steps include: freezing the weights of the basic large model, using the LoRA module as the training target, using a preset convergence as the training iteration condition, introducing a learning rate, and executing iterative training learning steps according to the instruction dataset. Among them, the enhanced knowledge text is identified, including: By using contextual semantic awareness, the material-synthesized knowledge text is segmented into blocks to determine a knowledge block set, wherein complete technical semantic units are used as segmentation constraints. Traverse the knowledge block set, introduce embedded words and vectorize them to determine the semantic vector library, where the embedded words are domain-enhanced feature words; Using vector similarity retrieval, a candidate set is recalled from the semantic vector library, text semantic reordering is performed, and the text ranking result is determined. Among them, the semantic vectors that meet the vector similarity are taken as a candidate set, and the ranking is constrained by the weight of technical keywords in the field of materials synthesis. Based on the text sorting results, knowledge query and text enhancement processing are performed to determine the enhanced knowledge text; Specifically, based on the text sorting results, knowledge query and text enhancement processing are performed to determine the enhanced knowledge text, including: Traverse the sorted text results, extract technical keywords, and determine the query vector; Based on the query vector, database distribution and retrieval are performed to determine highly relevant texts that meet a preset relevance, wherein the database is a source database for knowledge enhancement; The highly relevant text and knowledge block set are semantically aligned and additional information is generated to determine the enhanced knowledge text, wherein the additional information includes at least background information and focus of attention.

2. The method for extracting material synthesis data based on a knowledge-enhanced large model as described in claim 1, characterized in that, Constructing a knowledge text on materials synthesis, including: By constructing search queries, targeted searches of the literature pool are performed to determine the literature data; The document data is parsed. If the document data is in text format, text denoising preprocessing is performed to determine the material synthesized knowledge text. The preprocessing method includes at least special element transformation and redundancy denoising. If the literature data is in graph format, perform first-order processing based on OCR text conversion and second-order processing based on text denoising to determine the material-synthesized knowledge text.

3. The method for extracting material synthesis data based on a knowledge-enhanced large model as described in claim 1, characterized in that, Construct a data extraction prompt, including: Based on the material synthesis process, extract key characteristic words for material synthesis; By integrating the key feature words with role and task prompts, the data extraction prompt is determined. Among them, the integration of role-based task prompts includes at least material synthesis feature prompts, special data prompts, and return result prompts.

4. The method for extracting material synthesis data based on a knowledge-enhanced large model as described in claim 3, characterized in that, Perform knowledge extraction processing to determine the material synthesis extraction data, including: By integrating the material synthesis knowledge text, the data extraction prompt, and the enhanced knowledge text, the fusion input data is determined. The fused input data is imported into the extraction model for knowledge extraction processing, and the extracted material synthesis data is output.

5. The method for extracting material synthesis data based on a knowledge-enhanced large model as described in claim 1, characterized in that, By associating with an external domain knowledge base, the material synthesis extraction data is aligned with domain knowledge, missing fields are compensated by knowledge reasoning, and logical verification and correction are performed by domain rules, which serve as the material synthesis extraction result. The alignment methods include entity disambiguation and coreference resolution.

6. A material synthesis data extraction system based on a knowledge-enhanced large model, characterized in that, The system is used to implement the material synthesis data extraction method based on a knowledge-enhanced large model as described in any one of claims 1-5, the system comprising: The Materials Synthesis Knowledge Text Construction Module is used to construct materials synthesis knowledge text by retrieving literature data on materials synthesis and performing data denoising and OCR text conversion. The large model fine-tuning module is used to fine-tune the basic large model using LoRA, introduces the domain instruction dataset for fine-tuning learning, and determines the extracted large model. The knowledge enhancement module is used to traverse the material synthesis knowledge text, perform semantic segmentation and quantization, construct a multi-level retrieval architecture, combine the material science knowledge base to perform retrieval enhancement, and determine the enhanced knowledge text. The knowledge extraction processing module is used to construct a data extraction prompt, combine the enhanced knowledge text and the material synthesis knowledge text, input the extraction model, perform knowledge extraction processing, and determine the material synthesis extraction data.

Images