Multi-Transformer collaborative chemical reaction mining methods, devices, processors, and storage media

Abstract

This invention relates to the field of natural language processing technology, specifically disclosing a method, apparatus, processor, and storage medium for chemical reaction mining using a multi-Transformer collaborative approach. The method includes: acquiring chemical reaction data and a pre-trained multi-Transformer collaborative model; performing component standardization on the chemical reaction data to generate standardized chemical reaction components; performing reaction rule analysis on the chemical reaction data based on atomic mapping rules, the chemical reaction components, and the multi-Transformer collaborative model to generate component reaction rules; extracting structured steps from target documents based on the multi-Transformer collaborative model to obtain structured reaction steps; and performing information mining on the structured reaction steps based on the component reaction rules to generate mined chemical reactions. By training the multi-Transformer collaborative model, chemical reactions in documents can be accurately and comprehensively identified and mined, meeting practical needs and improving work efficiency.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a method, apparatus, processor, and storage medium for chemical reaction mining using a multi-Transformer collaborative approach. Background Technology

[0002] In the fields of chemistry, pharmaceuticals, and materials science, efficiently and accurately extracting chemical reaction information from massive amounts of literature is a fundamental and crucial task. Technicians often need to refer to numerous documents to aid their research and experiments. Currently, this process is primarily manual, requiring technicians to read each document, identify key information such as reactants, products, catalysts, and solvents, and manually input this information into databases or analytical tools.

[0003] However, in practical applications, manual data mining has significant drawbacks: on the one hand, the sheer volume and rapid growth of literature leads to a heavy workload and low efficiency; on the other hand, the reaction descriptions in the literature often suffer from problems such as vague expressions, inconsistent terminology, or missing information (for example, the term "palladium on carbon catalyst" does not specify dry or wet states, loading, or particle size), which can easily lead to human misjudgment or omissions and affect the reliability of the data.

[0004] Furthermore, the standardization of chemical reaction data faces significant challenges. The same chemical entity may have multiple names or specifications due to differences in experimental conditions or literature descriptions (e.g., "palladium on carbon" includes different forms such as Pd / C (dry) and Pd / C (wet, containing 10% water)). The roles of reagents or solvents also lack a unified standard—some solvents (such as water and alcohols) may participate as reactants and provide atoms, while others may only act as inert media. This heterogeneity of data not only increases the difficulty of manual data aggregation and analysis but also hinders the automated mining of chemical reaction knowledge, database construction, and the training and application of machine learning models. Summary of the Invention

[0005] To overcome the aforementioned technical problems in the prior art, embodiments of the present invention provide a method, apparatus, processor, and storage medium for chemical reaction mining using a multi-Transformer collaborative model. By training a multi-Transformer collaborative model, chemical reactions in literature can be accurately and comprehensively identified and mined, meeting practical needs and improving work efficiency.

[0006] To achieve the above objectives, embodiments of the present invention provide a multi-Transformer collaborative chemical reaction mining method, the method comprising: acquiring chemical reaction data and a pre-trained multi-Transformer collaborative model; performing component standardization processing on the chemical reaction data to generate standardized chemical reaction components; performing reaction rule analysis on the chemical reaction data based on atomic mapping rules, the chemical reaction components, and the multi-Transformer collaborative model to generate component reaction rules; extracting structured steps from target literature based on the multi-Transformer collaborative model to obtain structured reaction steps; and performing information mining on the structured reaction steps based on the component reaction rules to generate mined chemical reactions.

[0007] Preferably, the method further includes: constructing a chemical entity recognition model based on SciBERT; constructing a chemical reaction recognition model based on a cross-module attention mechanism and a gated feature aggregation mechanism; determining the optimal reaction data based on historical reaction data; optimizing the chemical reaction recognition model based on the optimal reaction data to generate an optimized model; constructing a product prediction model; and generating a pre-trained multi-Transformer collaborative model based on the chemical entity recognition model, the optimized model, and the product prediction model.

[0008] Preferably, the construction of the chemical reaction recognition model based on the cross-module attention mechanism and the gated feature aggregation mechanism includes: constructing a relationship recognition Transformer based on the cross-attention mechanism; constructing a role classification Transformer based on the gated feature aggregation mechanism and the historical reaction data; obtaining a preset standard terminology library and constructing a standardized Transformer based on the preset standard terminology library; and generating the chemical reaction recognition model based on the relationship recognition Transformer, the role classification Transformer, and the standardized Transformer.

[0009] Preferably, the step of standardizing the chemical reaction data to generate standardized chemical reaction components includes: extracting chemical components from the chemical reaction data; standardizing the expression of the chemical components to generate standardized components; and labeling the standardized components to generate standardized chemical reaction components, wherein the standardized chemical reaction components include reactants, products, reagents, and reaction condition information.

[0010] Preferably, the step of performing reaction rule analysis on the chemical reaction data based on atomic mapping rules, the chemical reaction components, and the multi-Transformer collaborative model to generate component reaction rules includes: performing atomic mapping analysis on reactants and products in the chemical reaction data based on atomic mapping rules to generate atomic mapping information; performing atomic matching on the atomic mapping information based on the multi-Transformer collaborative model to generate atomic matching information; determining unreacted components that have not undergone atomic mapping and reacted components that have undergone atomic mapping from the chemical reaction components based on the atomic matching information; determining the corresponding first component reaction rule based on the physical morphology of the unreacted components; determining the second component reaction rule for the reacted components based on the atomic matching information; and generating component reaction rules based on the first component reaction rule, the second component reaction rule, and the reaction condition information.

[0011] Preferably, the step of extracting structured steps from the target document based on the multi-Transformer collaborative model to obtain structured reaction steps includes: extracting synthesis actions from the target document based on the multi-Transformer collaborative model; arranging the synthesis actions in a structured manner to generate structured synthesis information; and generating structured reaction steps based on the structured synthesis information.

[0012] Preferably, the step of mining information from the structured reaction step based on the component reaction rules to generate a post-mined chemical reaction includes: determining whether the structured reaction step has an expression defect; if so, obtaining the defect type; if the defect type is missing components, performing component completion on the structured reaction step based on the component reaction rules to generate a completed chemical reaction; if the defect type is a component mapping error, performing reaction correction on the structured reaction step based on the component reaction rules to generate a corrected chemical reaction; if the defect type is missing reaction conditions, supplementing the reaction conditions on the structured reaction step based on the component reaction rules to generate a supplemented chemical reaction; and generating a post-mined chemical reaction based on the completed chemical reaction and / or the corrected chemical reaction and / or the supplemented chemical reaction.

[0013] Accordingly, the present invention also provides a multi-Transformer collaborative chemical reaction mining device, the device comprising: an acquisition unit for acquiring chemical reaction data and a pre-trained multi-Transformer collaborative model; a processing unit for performing component standardization processing on the chemical reaction data to generate standardized chemical reaction components; an analysis unit for performing reaction rule analysis on the chemical reaction data based on atomic mapping rules, the chemical reaction components, and the multi-Transformer collaborative model to generate component reaction rules; an extraction unit for extracting structured steps from target literature based on the multi-Transformer collaborative model to obtain structured reaction steps; and a mining unit for performing information mining on the structured reaction steps based on the component reaction rules to generate mined chemical reactions.

[0014] On the other hand, the present invention also provides a processor for running a program, wherein the program is run to perform the method provided in the embodiments of the present invention.

[0015] On the other hand, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method provided in the embodiments of the present invention.

[0016] The present invention has at least the following technical effects through the technical solution provided by the present invention:

[0017] By improving existing literature reading methods, a multi-Transformer collaborative model is pre-trained and analyzed to generate standard component reaction rules from publicly available standard chemical reaction data. Based on this, the multi-Transformer collaborative model and component reaction rules are used to automatically read the literature and automatically identify and mine chemical reaction information, thereby providing technical personnel with fast, accurate and complete chemical reaction information, improving user experience and meeting the actual needs of enterprises.

[0018] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings:

[0020] Figure 1 This is a flowchart illustrating the specific implementation of the multi-Transformer collaborative chemical reaction mining method provided in this embodiment of the invention.

[0021] Figure 2This is a schematic diagram illustrating the principle matching of atomic mapping information provided in an embodiment of the present invention;

[0022] Figure 3 This is a schematic diagram illustrating the extraction of structured synthesis information from target documents based on a multi-Transformer collaborative model, provided by an embodiment of the present invention.

[0023] Figure 4 This is a schematic diagram illustrating the conversion of structured synthesis information into a chemical expression, provided by an embodiment of the present invention.

[0024] Figure 5 This is a schematic diagram illustrating the process of mining and improving defects in chemical expressions provided by an embodiment of the present invention;

[0025] Figure 6 This is a schematic diagram of the structure of the multi-Transformer collaborative chemical reaction mining device provided in an embodiment of the present invention. Detailed Implementation

[0026] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0027] In this invention, the terms "system" and "network" are used interchangeably. "Multiple" refers to two or more; therefore, in this invention, "multiple" can also be understood as "at least two." "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / ", unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship. Furthermore, it should be understood that in the description of this invention, terms such as "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or order.

[0028] Please see Figure 1 This invention provides a multi-Transformer collaborative chemical reaction mining method, the method comprising:

[0029] S10: Acquire chemical reaction data and pre-trained multi-Transformer collaborative models;

[0030] S20: Perform component standardization processing on the chemical reaction data to generate standardized chemical reaction components;

[0031] S30: Based on the atomic mapping rules, the chemical reaction components, and the multi-Transformer collaborative model, perform reaction rule analysis on the chemical reaction data to generate component reaction rules;

[0032] S40: Based on the aforementioned multi-Transformer collaborative model, structured steps are extracted from the target document to obtain structured reaction steps;

[0033] S50: Based on the component reaction rules, information mining is performed on the structured reaction steps to generate the mined chemical reaction.

[0034] In one possible implementation, chemical reaction data and a pre-trained multi-Transformer collaborative model are first acquired. For example, to ensure the comprehensiveness and accuracy of the acquired chemical reaction data, it can be obtained from publicly available compound databases such as SCIFINDER, ACS, and USPTO, as well as journal article databases. In recent years, Natural Language Processing (NLP) technologies, especially Transformer-based models (such as BERT and GPT), have demonstrated powerful capabilities in text information extraction tasks. However, single Transformer models still have limitations when processing chemical reaction text: chemical reaction information involves multiple sub-tasks such as entity recognition (e.g., reactants, products), relation extraction (e.g., catalysis, solvation), and role classification (e.g., solvents participating / not participating in the reaction). A single model cannot optimize all tasks simultaneously, leading to performance degradation. General pre-trained models (such as BERT) lack chemical domain knowledge and have insufficient semantic understanding of specialized terms (e.g., "palladium on carbon," "DMF"), making them prone to misjudgments. Chemical reaction descriptions often contain complex conditions (such as temperature and pressure) and nested entities (such as "10% Pd / C"), and a single model is difficult to effectively capture this kind of structured information.

[0035] To address the aforementioned technical problems, in this embodiment of the invention, the method further includes: constructing a chemical entity recognition model based on SciBERT; constructing a chemical reaction recognition model based on a cross-module attention mechanism and a gated feature aggregation mechanism; determining optimal reaction data based on historical reaction data; optimizing the chemical reaction recognition model based on the optimal reaction data to generate an optimized model; constructing a product prediction model; and generating a pre-trained multi-Transformer collaborative model based on the chemical entity recognition model, the optimized model, and the product prediction model.

[0036] In one possible implementation, a chemical entity recognition module (NER Transformer) is first constructed, such as a domain-pre-trained Transformer (e.g., SciBERT, ChemBERTa), specifically designed to identify key chemical entities such as reactants, products, catalysts, and solvents. During training, SMILES strings or reaction formulas (e.g., C=O+[H]→CO) are used as input, and reactants, reagents, and conditions are encoded into dense vectors. The model extracts atomic / bond-level change features (e.g., functional group migration) to output corresponding information. Then, a chemical reaction recognition model is constructed based on cross-module attention mechanisms and gated feature aggregation mechanisms.

[0037] In this embodiment of the invention, a multi-Transformer modular architecture for chemical reactions is constructed. The construction of a chemical reaction recognition model based on a cross-module attention mechanism and a gated feature aggregation mechanism includes: constructing a relationship recognition Transformer based on a cross-attention mechanism; constructing a role classification Transformer based on a gated feature aggregation mechanism and the historical reaction data; obtaining a preset standard terminology library; constructing a standardized Transformer based on the preset standard terminology library; and generating a chemical reaction recognition model based on the relationship recognition Transformer, the role classification Transformer, and the standardized Transformer.

[0038] Specifically, a cross-attention mechanism is employed to model the reaction relationships between entities (e.g., "A+B→C", "Pd / C catalysis"), i.e., constructing a relationship recognition Transformer. Then, a role classification Transformer is created based on a gating feature aggregation mechanism and historical reaction data. Specifically, the Transformer is trained using this mechanism to effectively distinguish the involvement of solvents / reagents (e.g., "water as reactant" vs. "DMF as solvent only"). Finally, a standardized Transformer is constructed based on a pre-defined standard terminology library. Specifically, it is trained using standardized chemical reaction terms from this library to map unstructured descriptions (e.g., "palladium on carbon") to standard terms (e.g., "Pd / C (5% loaded, dry)").

[0039] After constructing the chemical reaction recognition model, the optimal reaction data is further determined based on historical reaction data. For example, in this embodiment of the invention, a reaction condition optimizer can be further constructed. For instance, a reaction condition generator can be built using the chemical reaction recognition model. During training, it is trained by inputting reactants, target products, and historical condition data to output data such as the optimal reagent / catalyst (e.g., "Pd / C") and the numerical range of reaction conditions (temperature, pH, etc.). This output is used to optimize the chemical reaction recognition model to obtain an optimized model. Then, a product prediction model is further constructed. For example, a hybrid architecture combining graph neural networks (GNN) and Transformers can be used to construct the product prediction model. The GNN part is used to process molecular graph structures (e.g., MPNN), and the Transformer part is used to model long-range dependencies (e.g., interatomic interactions). An attention mechanism is used to dynamically adjust the importance of atoms (e.g., higher weight for the reaction center atom). During training, reactant SMILES and reaction rules are input to output the probability distribution of candidate products (e.g., Top-k SMILES). Thus, the construction of a multi-Transformer collaborative model is completed.

[0040] For example, in one embodiment, a document describes that "benzoic acid reacts with ethanol to produce ethyl benzoate under Pd / C catalysis". The chemical entity recognition model identifies the entities: Pd / C (catalyst), benzoic acid (reactant), ethanol (reactant), and ethyl benzoate (product). The relationship recognition Transformer extracts the relationship: benzoic acid + ethanol → ethyl benzoate, Pd / C catalysis. The role classification Transformer confirms that ethanol is a reactant (non-solvent). The normalization Transformer normalizes "Pd / C" to "Pd / C (10% loading, dry)".

[0041] In this embodiment of the invention, by training a multi-Transformer collaborative model, the limitations of existing models in text processing in the chemical field can be effectively addressed, and various defects in the process of chemical reaction recognition in the chemical field can be resolved, thereby effectively improving the accuracy of chemical reaction recognition and enhancing the comprehensiveness, accuracy, and reliability of chemical reaction mining.

[0042] After completing model training and data collection, a standard database needs to be prepared in advance to achieve standardization in the subsequent identification and response mining of various documents.

[0043] In this embodiment of the invention, the step of standardizing the chemical reaction data to generate standardized chemical reaction components includes: extracting chemical components from the chemical reaction data; standardizing the expression of the chemical components to generate standardized components; and labeling the standardized components to generate standardized chemical reaction components, wherein the standardized chemical reaction components include reactants, products, reagents, and reaction condition information.

[0044] In one possible implementation, chemical components are first extracted from chemical reaction data, such as the name, chemical meaning, and type of each chemical component. Then, the chemical components are standardized. Specifically, for chemical components that are not standardized using abbreviations or acronyms, they are standardized according to the IUPAC standard to generate standardized components. Then, the components are further labeled, such as reactants, products, reagents, catalysts, and reaction conditions, to generate standardized reaction components.

[0045] Then, the reaction rules in the above chemical reaction data are further analyzed to generate component reaction rules, and combined with the above standardized chemical reaction components to construct a standard reaction database for subsequent use in chemical reaction data mining.

[0046] In existing technologies, reaction rule extraction is often performed manually by technicians. However, traditional analysis methods are inefficient, highly subjective, and prone to significant bias, failing to meet practical needs. Conversely, directly employing natural language processing (NLP) for rule extraction may result in substantial recognition errors due to semantic differences.

[0047] To address the aforementioned technical problems, in this embodiment of the invention, the step of performing reaction rule analysis on the chemical reaction data based on atomic mapping rules, the chemical reaction components, and the multi-Transformer collaborative model to generate component reaction rules includes: performing atomic mapping analysis on reactants and products in the chemical reaction data based on atomic mapping rules to generate atomic mapping information; performing atomic matching on the atomic mapping information based on the multi-Transformer collaborative model to generate atomic matching information; determining unreacted components that have not undergone atomic mapping and reacted components that have undergone atomic mapping from the chemical reaction components based on the atomic matching information; determining the corresponding first component reaction rule based on the physical morphology of the unreacted components; determining the second component reaction rule for the reacted components based on the atomic matching information; and generating component reaction rules based on the first component reaction rule, the second component reaction rule, and the reaction condition information.

[0048] In one possible implementation, to achieve more accurate chemical reaction rule analysis, the principle of principle mapping is used to extract reaction rules. Specifically, firstly, based on the atomic mapping rules, atomic mapping analysis is performed on the reactants and products in the chemical reaction data to generate atomic mapping information. For example, each chemical reaction in the chemical reaction data is converted into a chemical reaction process expressed by a molecular formula, and the mapping of each atom in the chemical reaction process is analyzed through the molecular formula to generate atomic mapping information.

[0049] Then, based on the multi-Transformer collaborative model, atomic matching is performed on the atomic mapping information to generate atomic matching information. Specifically, atoms are matched by finding the atom with the maximum attention weight in the self-attention layer to generate atomic matching information, for example, please refer to [link to relevant documentation]. Figure 2 At this point, unreacted and reacted components are further identified from the chemical reaction components based on atomic matching information. For example, in this embodiment, compounds without atomic mapping indicate that they did not participate in the reaction, such as solvents, and can therefore be identified as unreacted components. Based on the physical morphology of the unreacted components, their first-component reaction rules in the chemical reaction process can be determined. For example, in one embodiment, water, as a solution, did not participate in the reaction, and was therefore identified as a solution based on its physical morphology. Simultaneously, based on the atomic matching information of the reacted components, their second-component reaction rules in the reaction process can be determined, finally generating the component reaction rules for the entire chemical reaction. Based on this, a large database of standard reaction rules for chemical reactions is constructed. In subsequent chemical reaction mining processes, the reaction processes can be standardized, comprehensive, and rapidly matched and improved based on this database of standard reaction rules.

[0050] In this embodiment of the invention, the step of extracting structured steps from the target document based on the multi-Transformer collaborative model to obtain structured reaction steps includes: extracting synthesis actions from the target document based on the multi-Transformer collaborative model; arranging the synthesis actions in a structured manner to generate structured synthesis information; and generating structured reaction steps based on the structured synthesis information.

[0051] In one possible implementation, a company needs to quickly search and learn about chemical reactions in a certain document. Manually reading through the document is inefficient. Therefore, the multi-Transformer collaborative model provided in this embodiment of the invention is first used to extract structured steps from the target document, thereby extracting structured reaction steps. Specifically, the multi-Transformer collaborative model identifies synthesis actions from the target document, ignoring contextual connection statements and statements about reaction effects, thus improving reading efficiency. Then, the synthesis actions are arranged in a structured manner, generating structured synthesis information. For example, please refer to [link to relevant documentation]. Figure 3 .

[0052] For example, a chemical reaction process may include the following:

[0053] Reactant preparation: Add 3.0 g of 4-iodobenzoic acid (12.1 mmol) and 100 mL of chloroform to a round-bottom flask.

[0054] Add reagents: Add thioyl chloride (5.0 mL) dissolved in 10 mL of chloroform and 2-3 drops of dimethylformamide (DMF) to the solution.

[0055] Reaction conditions: The mixture was heated under reflux for 2 hours, and the reaction was monitored by oil bubble.

[0056] Results: A clear solution of 4-iodobenzoyl chloride was obtained. After removing volatile substances, a colorless oily substance was obtained, which solidified upon cooling.

[0057] After extracting the structured reaction steps, the following information can be obtained:

[0058] 1. Preparation of solution: Dissolve 3.0 g (12.1 mmol) of 4-iodobenzoic acid in 100 mL of chloroform.

[0059] 2. Add thiocyanate chloride (SLN).

[0060] 3. Preparation of thioyl chloride solution: Dissolve thioyl chloride (5.0 mL) in 10 mL of chloroform.

[0061] 4. Add thiocyanate chloride (SLN) again.

[0062] 5. Adding excessive amounts of dimethylformamide.

[0063] 6. Recirculate for 2 hours.

[0064] 7. 4-Iodobenzoyl chloride was obtained.

[0065] 8. Concentrated.

[0066] 9. Obtain colorless oil.

[0067] Traditional methods stop at simply processing text into structured synthetic information and presenting it to technicians. This information may lack certain compounds that are difficult for technicians to identify intuitively from textual information alone. Therefore, to address this technical problem, after generating the structured synthetic information, it is further converted into standard chemical expressions to facilitate analysis and discovery. For example, please refer to [link to relevant documentation]. Figure 4 .

[0068] Finally, after generating the chemical expression, information mining is performed on the structured reaction steps using the generated component reaction rules. In this embodiment of the invention, the step of mining information on the structured reaction steps based on the component reaction rules to generate a mined chemical reaction includes: determining whether the structured reaction steps have expression defects; if so, obtaining the defect type; if the defect type is missing components, performing component completion on the structured reaction steps based on the component reaction rules to generate a completed chemical reaction; if the defect type is incorrect component mapping, performing reaction correction on the structured reaction steps based on the component reaction rules to generate a corrected chemical reaction; if the defect type is missing reaction conditions, supplementing the reaction conditions on the structured reaction steps based on the component reaction rules to generate a supplemented chemical reaction; and generating a mined chemical reaction based on the completed chemical reaction and / or the corrected chemical reaction and / or the supplemented chemical reaction.

[0069] In one possible implementation, when automatically identifying chemical reactions in literature, the model can first generate preliminary identification information. Specifically, the preliminary identification information can be a preliminarily generated chemical formula of the reaction or a structured reaction step. Then, it is determined whether the structured reaction step has expression defects. In the embodiments of the present invention, the expression defects include, but are not limited to, lack of reactants, lack of products, lack of reagents, inability to distinguish between reagents and solvents, lack of reaction conditions, etc.

[0070] If a chemical reaction described in the current literature is found to have an expression defect, the defect type is immediately determined, and chemical reaction mining is performed. For example, in the first embodiment, a defect in a chemical reaction with missing products was detected.

[0071] CC(C)S.CN(C)C=O.Fc1cccnc1F.O=C([O-])[O-].[K+].[K+]>>

[0072] Therefore, product completion is automatically performed, and a completed chemical reaction is generated. Based on the same mining logic, if other expression defects are found, mining and defect improvement operations are automatically performed according to the generated component reaction rules or relevant standard reaction databases, and a mined chemical reaction is generated. For example, please refer to [link to relevant documentation]. Figure 5 This is a schematic diagram of the completed chemical reaction.

[0073] In this embodiment of the invention, by employing a multi-Transformer collaborative model to accurately identify and perform detailed mining of chemical reactions in the literature, various defects in the chemical reactions recorded in the literature are effectively repaired, assisting technicians to obtain reference materials in the literature better and more accurately, improving user experience and work efficiency.

[0074] The multi-Transformer collaborative chemical reaction mining device provided in the embodiments of the present invention will be described below with reference to the accompanying drawings.

[0075] Please see Figure 6 Based on the same inventive concept, this invention provides a multi-Transformer collaborative chemical reaction mining device, comprising: an acquisition unit for acquiring chemical reaction data and a pre-trained multi-Transformer collaborative model; a processing unit for performing component standardization processing on the chemical reaction data to generate standardized chemical reaction components; an analysis unit for performing reaction rule analysis on the chemical reaction data based on atomic mapping rules, the chemical reaction components, and the multi-Transformer collaborative model to generate component reaction rules; an extraction unit for extracting structured steps from target literature based on the multi-Transformer collaborative model to obtain structured reaction steps; and a mining unit for performing information mining on the structured reaction steps based on the component reaction rules to generate mined chemical reactions.

[0076] Furthermore, embodiments of the present invention also provide a processor for running a program, wherein the program is run to execute the methods described in embodiments of the present invention.

[0077] Furthermore, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the methods described in the embodiments of the present invention.

[0078] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above embodiments. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention.

[0079] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not describe the various possible combinations separately.

[0080] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0081] Furthermore, various different implementations of the present invention can be combined arbitrarily, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed in the present invention.

Claims

1. A multi-Transformer collaborative chemical reaction mining method, characterized in that, The method includes: Acquire chemical reaction data and pre-trained multi-Transformer collaborative models; The chemical reaction data is subjected to component standardization processing to generate standardized chemical reaction components; Based on the atomic mapping rules, the chemical reaction components, and the multi-Transformer collaborative model, reaction rule analysis is performed on the chemical reaction data to generate component reaction rules; Based on the aforementioned multi-Transformer collaborative model, structured steps are extracted from the target documents to obtain structured reaction steps; Information mining is performed on the structured reaction steps based on the component reaction rules to generate the mined chemical reaction; The method further includes: A chemical entity recognition model based on SciBERT was constructed. A chemical reaction recognition model is constructed based on cross-module attention mechanism and gating feature aggregation mechanism; Determine the optimal reaction data based on historical reaction data; The chemical reaction identification model is optimized based on the optimal reaction data to generate an optimized model; Construct a product prediction model; A pre-trained multi-Transformer collaborative model is generated based on the chemical entity recognition model, the optimized model, and the product prediction model.

2. The method according to claim 1, characterized in that, The chemical reaction recognition model constructed based on cross-module attention mechanism and gating feature aggregation mechanism includes: A Transformer for relationship recognition is built based on a cross-attention mechanism; A role classification Transformer is constructed based on the gating feature aggregation mechanism and the historical reaction data. Obtain a preset standard terminology library, and construct a standardized Transformer based on the preset standard terminology library; A chemical reaction recognition model is generated based on the relationship recognition Transformer, the role classification Transformer, and the standardized Transformer.

3. The method according to claim 1, characterized in that, The step of standardizing the chemical reaction data to generate standardized chemical reaction components includes: Extract chemical components from the chemical reaction data; The chemical components are expressed in a standardized manner to generate standardized components; The standardized components are labeled to generate standardized chemical reaction components, which include reactants, products, reagents, and reaction condition information.

4. The method according to claim 3, characterized in that, The reaction rule analysis of the chemical reaction data based on the atomic mapping rules, the chemical reaction components, and the multi-Transformer collaborative model generates component reaction rules, including: Based on atomic mapping rules, atomic mapping analysis is performed on reactants and products in the chemical reaction data to generate atomic mapping information. Based on the multi-Transformer collaborative model, atomic matching is performed on the atomic mapping information to generate atomic matching information; Based on the atomic matching information, unreacted components that have not undergone atomic mapping and reacted components that have undergone atomic mapping are identified from the chemical reaction components. The corresponding first component reaction rule is determined based on the physical morphology of the unreacted component; The second component reaction rule of the reacted component is determined based on the atomic matching information; Component reaction rules are generated based on the first component reaction rule, the second component reaction rule, and the reaction condition information.

5. The method according to claim 1, characterized in that, The structured step extraction of the target document based on the multi-Transformer collaborative model to obtain structured reaction steps includes: Based on the multi-Transformer collaborative model, synthetic actions are extracted from the target documents; The synthesis actions are arranged in a structured manner to generate structured synthesis information; The structured reaction steps are generated based on the structured synthesis information.

6. The method according to claim 1, characterized in that, The process of mining information from the structured reaction steps based on the component reaction rules to generate post-mined chemical reactions includes: Determine whether the structured reaction step has expression defects; If so, obtain the defect type; When the defect type is a missing component, the structural reaction step is supplemented with components based on the component reaction rules to generate a supplemented chemical reaction. In the case where the defect type is component mapping error, the structured reaction step is corrected based on the component reaction rule to generate a corrected chemical reaction. When the defect type is missing reaction conditions, the reaction conditions of the structured reaction step are supplemented based on the component reaction rules to generate a supplemented chemical reaction. The post-mining chemical reaction is generated based on the completed chemical reaction and / or the corrected chemical reaction and / or the supplemented chemical reaction.

7. A multi-Transformer collaborative chemical reaction mining device, characterized in that, The device includes: The acquisition unit is used to acquire chemical reaction data and a pre-trained multi-Transformer collaborative model. The processing unit is used to perform component standardization processing on the chemical reaction data to generate standardized chemical reaction components; The analysis unit is used to perform reaction rule analysis on the chemical reaction data based on the atomic mapping rules, the chemical reaction components, and the multi-Transformer collaborative model, and generate component reaction rules. The extraction unit is used to extract structured steps from the target document based on the multi-Transformer collaborative model to obtain structured reaction steps; The mining unit is used to mine information about the structured reaction steps based on the component reaction rules and generate the mined chemical reaction. The method for generating the multi-Transformer collaborative model includes: A chemical entity recognition model based on SciBERT was constructed. A chemical reaction recognition model is constructed based on cross-module attention mechanism and gating feature aggregation mechanism; Determine the optimal reaction data based on historical reaction data; The chemical reaction identification model is optimized based on the optimal reaction data to generate an optimized model; Construct a product prediction model; A pre-trained multi-Transformer collaborative model is generated based on the chemical entity recognition model, the optimized model, and the product prediction model.

8. A processor, characterized in that, Used to run a program, wherein the program is run to perform the method of any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the method described in any one of claims 1-6.

Images