Method for extracting chemical reaction information from journal paper

Abstract

The invention provides a method for extracting chemical reaction information from journal papers, and relates to the technical field of chemical informatics. The method comprises the following steps: firstly, extracting a reaction information unit from a journal paper PDF (Portable Document Format) file; carrying out reaction information modal segmentation on the reaction information units; automatically segmenting the extracted reaction information unit into three types of picture information including a chemical reaction expression, a table and a conditional description text; respectively analyzing the segmented chemical reaction expression picture, the table picture and the conditional description text picture; and finally, integrating the analyzed information into complete chemical reaction information. Meanwhile, the extracted, segmented, analyzed and integrated reaction information is stored in an optimized mode, and subsequent use and updating are facilitated. The method can be used for efficiently and accurately extracting reactants, products, reaction conditions, yield, remarks and other information related to the chemical reaction in the journal paper.

Description

Technical Field

[0001] The present invention relates to the technical field of chemical informatics, and in particular to a method for extracting chemical reaction information from journal articles. Background Art

[0002] Chemical reactions and their related reactants, products, reaction conditions, product characteristics and other information are one of the most important basic knowledge in the fields of medicinal chemistry, material chemistry and biochemistry. By mastering this knowledge, relevant personnel can infer the reaction mechanism, design and optimize the synthesis route, predict the main product and possible by-products of the reaction, and thus efficiently synthesize target compounds such as drugs, natural products, new compounds and new materials.

[0003] The main source of chemical reactions and related information is journal literature, but this information is presented in various forms, which are multimodal and highly unstructured, including pictures, tables, texts, etc. Efficiently and accurately extracting chemical reactions and related information from journal literature can provide strong support for the development and application of related disciplines and fields.

[0004] At present, there are two methods for extracting chemical reaction information from journal literature: one is manual extraction based on expert knowledge; the other is automatic extraction based on artificial intelligence. The manual extraction method mainly relies on the knowledge and understanding of experts in the field. Its advantage is that it can ensure the high accuracy of information, but its disadvantage is that it is extremely inefficient and requires a lot of time and labor costs. Manual extraction is the main method currently used by mainstream chemical reaction information databases, including SciFinder, Reaxys, etc. Corresponding to manual extraction is the recently developed automatic extraction method based on artificial intelligence. This type of automatic extraction method is based on big data and artificial intelligence technology, and establishes an intelligent algorithm based on part of the annotated data to achieve automatic extraction of chemical reaction information. For example, patent CN115481627A discloses a chemical reaction information extraction method, which is based on semantic recognition of reaction information in chemical reactions, and compares the corresponding chemical formula structure through image recognition. This method makes full use of the full text information in the journal, but cannot parse the logical information in the chemical reaction schematic diagram. In fact, professionals are more accustomed to first looking at the schematic diagram of the chemical reaction and its attached charts, because the common information of the chemical reaction, including reactants, products, reaction conditions, etc., can be quickly and accurately identified from the chemical reaction schematic diagram and attached charts. In addition, the information density of chemical reaction schematics and their accompanying charts is much higher than that of text information. Therefore, it is more practical to develop an intelligent method for extracting chemical reaction information based on chemical reaction schematics and their accompanying charts. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a method for extracting chemical reaction information from journal articles based on chemical reaction schematic diagrams and accompanying charts in view of the deficiencies of the above-mentioned prior art.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for extracting chemical reaction information from journal articles, for the chemical reaction formula schematic diagram and its attached charts in the journal article, the extraction process is decomposed into process steps, specifically including reaction information unit extraction, reaction information modal segmentation, chemical reaction formula analysis, table analysis, text analysis, reaction information integration, data storage and multiple processes, including the following steps:

[0007] Step 1: extracting a reaction information unit from a journal document; the reaction information unit is a complete paper diagram containing at least one chemical reaction and its attached tables and condition description text in the journal document;

[0008] Step 2: segment the reaction information modality for the reaction information unit;

[0009] Automatically segment the reaction information unit extracted in step 1 into three types of image information: chemical reaction formula, table and condition description text;

[0010] Step 3: Analyze the chemical reaction formula image, table image and condition description text image segmented in step 2 respectively;

[0011] Step 3.1: Chemical reaction analysis;

[0012] Parse the chemical reaction formula image segmented in step 2, identify the reactants, products, arrows, reaction conditions, connectors, and molecular label information therein, and recognize the molecular structure diagram expressed in the form of an image as a machine-readable format;

[0013] Step 3.2: Table parsing;

[0014] Parse the table image segmented in step 2 and match the reaction conditions and remarks information according to the custom terms;

[0015] Step 3.3: Text parsing;

[0016] Parse the condition description text image segmented in step 2, and match the reaction conditions and remarks information according to the custom dictionary;

[0017] Step 4: reaction information integration;

[0018] Integrate the information parsed in step 3 into complete chemical reaction information; the integrated chemical reaction information includes: reactants, products, solvents, catalysts, reagents, molecular labels, reaction temperature, reaction time, yield and remarks;

[0019] Step 5: Data storage;

[0020] The reaction information extracted, segmented, parsed and integrated in steps 1 to 4 is stored in an optimized manner to facilitate subsequent use and updating.

[0021] Preferably, the specific method of extracting from the journal document in step 1 is:

[0022] Step S11: converting all pages of the journal document containing chemical reaction information into image format;

[0023] Step S12: manually marking the positions of the reaction information units in the image, thereby establishing an initial training set for reaction information unit extraction;

[0024] Step S13: using the annotated reaction information unit image data to establish a reaction information unit extraction machine learning model;

[0025] Step S14: using a reaction information unit extraction machine learning model to automatically extract reaction information units from new journal images;

[0026] Step S15: sampling the results extracted by the reaction information unit extraction machine learning model, manually re-labeling the pictures with extraction errors and adding them to the training set;

[0027] Step S16: Continuously update the iterative reaction information unit extraction machine learning model using the updated training data until the manual sampling error rate is lower than the set threshold.

[0028] Preferably, the specific method of segmenting the reaction information modality in step 2 is:

[0029] Step S21: manually marking the positions of chemical reaction formulas, tables, and texts in the reaction information unit images extracted in step 1, thereby establishing an initial training set for reaction information modality segmentation;

[0030] Step S22: establishing a reaction information modality machine learning model using the annotated reaction information modality image data;

[0031] Step S23: automatically segmenting the reaction information modality in the unlabeled reaction information unit image using the reaction information modality machine learning model;

[0032] Step S24: manually sampling the results of the segmentation of the reaction information modality machine learning model, manually re-labeling the images with segmentation errors and adding them to the training set of the reaction information modality segmentation;

[0033] Step S25: Continuously update and iterate the reaction information modality machine learning model using the updated reaction information modality segmentation training data until the manual sampling error rate is lower than the set threshold.

[0034] Preferably, the chemical reaction formula analysis in step 3.1 includes three sub-processes: chemical reaction sub-item identification, chemical reaction formula identification and molecular structure identification;

[0035] Among them, the purpose of chemical reaction sub-item recognition is to identify the locations of reactants, products, arrows, reaction conditions, connectors and molecular labels in the chemical reaction formula image;

[0036] The purpose of chemical reaction formula recognition is to determine the properties of molecules identified in chemical reaction sub-item recognition, namely: reactants, products, reaction conditions and molecular labels;

[0037] The purpose of the molecular structure recognition is to convert the molecular structure image identified in the chemical reaction sub-item recognition into a machine-readable format.

[0038] Preferably, the specific method for identifying the chemical reaction sub-item is:

[0039] (1) manually annotating the positions of molecules, arrows, reaction conditions, labels, and connectors in the chemical reaction formula image segmented in step 2, thereby establishing an initial training set for chemical reaction sub-item recognition;

[0040] (2) Using labeled chemical reaction sub-item image data to establish a reaction sub-item recognition machine learning model;

[0041] (3) Automatically identify the reaction sub-items in the segmented chemical reaction formula images using a chemical reaction sub-item recognition machine learning model;

[0042] (4) Manually proofread the results of the automatic recognition of the chemical reaction sub-item recognition machine learning model, and manually re-label the images that were misrecognized and add them to the training set of chemical reaction sub-item recognition;

[0043] (5) Using the updated training data of chemical reaction sub-item recognition, the reaction sub-item recognition machine learning model is continuously updated and iterated until the manual proofreading error rate is lower than the set threshold.

[0044] Preferably, the chemical reaction formula recognition process is performed according to the relative positions of the molecule and the arrow using the following rule-based method:

[0045] 1) Each arrow represents a chemical reaction;

[0046] 2) The molecules at the starting point of the arrow are reactants, the molecules at the end point of the arrow are products, and the other texts at the vertical direction of the arrow are reaction conditions;

[0047] 3) The molecules before and after or above and below the symbol “+” have the same properties;

[0048] 4) Based on the principle of closest distance, define the molecule to which each molecular tag belongs;

[0049] The reaction formulas identified according to the above rules are manually proofread, and the judgment rules are modified according to the proofreading results until the manual proofreading error rate is lower than the set threshold.

[0050] Preferably, the specific method of molecular structure recognition is:

[0051] (a) Artificially draw the molecular structure diagram identified in the chemical reaction sub-item recognition, so as to establish the initial training set for molecular structure recognition;

[0052] (b) Establishing a machine learning model for molecular structure recognition using artificially drawn molecular structure graph data;

[0053] (c) Automatically identify new molecular structure diagrams using molecular structure recognition machine learning models;

[0054] (d) manually proofreading the recognition results, and manually modifying or redrawing the corrected results and adding them to the molecular structure recognition training set;

[0055] (e) Using the updated molecular structure recognition training data, the molecular structure recognition machine learning model is continuously updated and iterated until the manual proofreading error rate is lower than the set threshold.

[0056] Preferably, the purpose of the table parsing in step 3.2 is to identify the table in the image format identified in step 2 as a machine-readable format, and the specific method of the identification process is:

[0057] Step S321) using an open source table parsing machine learning model to parse the table image;

[0058] Step S322) Manually perform random checks on the results of table analysis, manually re-mark the tables with random checks errors, and add them to the table analysis training set;

[0059] Step S323) Continuously update and iterate the table parsing machine learning model using the updated table parsing training data until the manual sampling error rate is lower than the set threshold.

[0060] Preferably, the purpose of the text parsing in step 3.3 is to identify the condition description text in the picture format identified in step 2 into a machine-readable format. The specific method of the identification process is:

[0061] Step S331: using an open source text parsing machine learning model to parse the condition description text image;

[0062] Step S332: Manually check the results of text analysis, manually re-label the erroneous condition description text images, and add them to the text analysis training set;

[0063] Step S333: Continuously update and iterate the text parsing machine learning model using the updated text parsing training data until the manual sampling error rate is lower than the set threshold.

[0064] Preferably, the purpose of the reaction information integration in step 4 is to integrate the chemical reaction information extracted and analyzed from all the above processes according to the set data categories, and the specific method is:

[0065] Step S41: Integrate the molecular structure diagram extracted in step 3.1 into the reaction information of reactants, products, solvents, catalysts and reagents according to the analyzed properties;

[0066] Step S42: The text information identified in step 3.1 and the text information parsed in step 3.3 are combined and processed, and the matching solvent, catalyst, reagent, molecular label, reaction temperature, reaction time and yield information are automatically extracted according to the custom dictionary using an open source semantic parsing model; the unmatched information is classified as the remark information;

[0067] Step S43: Integrate the table information parsed in step 3.3 into reaction condition information; create a table header dictionary, establish the correspondence between the table header and the reaction information attribute, and automatically extract the matching solvent, catalyst, reagent, molecular label, reaction temperature, reaction time and yield information; classify the unmatched information into the remark information;

[0068] Step S44: sampling the results of reaction information integration and correcting erroneous results;

[0069] Step S45: Integrate the data according to the modified reaction information and re-execute steps S41 to S44 until the sampling error rate is lower than the set threshold.

[0070] The beneficial effect of adopting the above technical solution is that: the method provided by the present invention for extracting chemical reaction information from journal articles can efficiently and accurately extract information such as reactants, products, reaction conditions, yields, notes, etc. related to chemical reactions in journal articles. The method decomposes the extraction process into process steps for chemical reaction formula schematics and their attached charts in journal articles, specifically including reaction information unit extraction, reaction information modal segmentation, chemical reaction formula analysis, table analysis, text analysis, reaction information integration, data storage and other processes. Each process includes two processing links: automatic preprocessing and human-computer interaction: automatic preprocessing uses machine learning-based or mechanism-based models to automatically process tasks, and human-computer interaction performs manual inspection and model update. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 A flow chart of a method for extracting chemical reaction information from a journal article provided by an embodiment of the present invention;

[0072] Figure 2 A schematic diagram of reaction information unit extraction provided by an embodiment of the present invention;

[0073] Figure 3 A schematic diagram of the modal segmentation of response information provided by an embodiment of the present invention;

[0074] Figure 4 A schematic diagram of chemical reaction sub-item identification provided by an embodiment of the present invention;

[0075] Figure 5 Schematic diagram of reactants and products of a chemical reaction provided in an embodiment of the present invention, wherein (a) is reactant 1, (b) is reactant 2, and (c) is the product. DETAILED DESCRIPTION

[0076] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0077] In this embodiment, a method for extracting chemical reaction information from journal articles is described as follows: Figure 1 As shown in the figure, for the chemical reaction diagrams and their attached charts in journal articles, the extraction process is decomposed into process steps, including reaction information unit extraction, reaction information modal segmentation, chemical reaction formula analysis, table analysis, text analysis, reaction information integration, data storage and other processes, which specifically include the following steps:

[0078] Step 1: extracting a reaction information unit from a journal document; the reaction information unit is a complete paper diagram containing at least one chemical reaction and its attached tables and condition description text in the journal document;

[0079] Step S11: converting all pages of the journal document containing chemical reaction information into image format;

[0080] Step S12: manually marking the positions of the reaction information units in the image, thereby establishing an initial training set for reaction information unit extraction;

[0081] Step S13: using the annotated reaction information unit image data to establish a reaction information unit extraction machine learning model;

[0082] Step S14: using a reaction information unit extraction machine learning model to automatically extract reaction information units from new journal images;

[0083] Step S15: sampling the results extracted by the reaction information unit extraction machine learning model, manually re-labeling the pictures with extraction errors and adding them to the training set;

[0084] Step S16: Continuously update the iterative reaction information unit extraction machine learning model using the updated training data until the manual sampling error rate is lower than the set threshold.

[0085] This example randomly selects 500 journal articles from the journal literature data, converts all pages into pictures, and manually annotates the units containing chemical reaction information, requiring that the reaction information unit contains at least one chemical reaction. This example manually annotates 3775 reaction information units from 500 journal articles, uses these data to train the machine learning model for target detection, thereby obtaining a reaction information unit extraction machine learning model, and realizing the extraction of pictures of chemical reaction information from journal article PDF files.

[0086] The established reaction information unit extraction machine learning model is used to extract reaction information units in subsequent journal articles. During the extraction process, about 10% of the extraction results are randomly selected for manual proofreading. If the extraction accuracy is greater than or equal to 95%, new journal articles are input for information extraction; if the accuracy is less than 95%, the corrected error results and all manually proofread data are added to the training set to retrain a new target detection machine learning model.

[0087] The extracted reaction information units are stored in a database and enter the next step of reaction information modal segmentation process.

[0088] In this embodiment, the chemical reaction information unit extracted from a PDF page of a journal file is as follows: Figure 2 As shown, the page contains two independent response information units.

[0089] Step 2: Perform reaction information modality segmentation on the reaction information unit; automatically segment the reaction information unit extracted in step 1 into three types of image information: chemical reaction formula, table, and condition description text. The specific method is as follows:

[0090] Step S21: manually marking the positions of chemical reaction formulas, tables, and texts in the reaction information unit images extracted in step 1, thereby establishing an initial training set for reaction information modality segmentation;

[0091] Step S22: establishing a reaction information modality machine learning model using the annotated reaction information modality image data;

[0092] Step S23: automatically segmenting the reaction information modality in the unlabeled reaction information unit image using the reaction information modality machine learning model;

[0093] Step S24: manually sampling the results of the segmentation of the reaction information modality machine learning model, manually re-labeling the images with segmentation errors and adding them to the training set of the reaction information modality segmentation;

[0094] Step S25: Continuously update and iterate the reaction information modality machine learning model using the updated reaction information modality segmentation training data until the manual sampling error rate is lower than the set threshold.

[0095] In this embodiment, 500 information units are randomly selected from the reaction information units extracted in step 1 for manual annotation. The positions of chemical reaction formulas, tables, and texts are required to be annotated. The machine learning model for target detection is trained using the 500 manually annotated reaction information modal data, thereby obtaining a reaction information modal machine learning model to achieve segmentation of the reaction information unit image.

[0096] The established reaction information modality machine learning model is used for the segmentation of reaction information modality in subsequent reaction information units. During the segmentation process, about 10% of the segmentation results are randomly selected for manual proofreading. If the extraction accuracy is greater than or equal to 95%, new units are continuously input for modality segmentation; if the accuracy is less than 95%, the corrected error results and all manually proofread data are added to the training set to retrain a new machine learning model.

[0097] The segmented reaction information modalities are stored in the database for the next step of analyzing chemical reaction formulas, tables and texts.

[0098] In this embodiment, the modal segmentation result of the second reaction information unit extracted in step 1 is as follows: Figure 3 As shown, the reaction information unit includes a chemical reaction formula image, a table image and two text images.

[0099] Step 3: Analyze the chemical reaction formula image, table image and condition description text image segmented in step 2 respectively;

[0100] Step 3.1: Chemical reaction formula analysis: analyze the chemical reaction formula image segmented in step 2, identify the reactants, products, arrows, reaction conditions, connectors, and molecular label information, and identify the molecular structure diagram expressed in the form of an image into a machine-readable format;

[0101] The chemical reaction formula analysis includes three sub-processes: chemical reaction sub-item identification, chemical reaction formula identification and molecular structure identification;

[0102] Among them, the purpose of chemical reaction sub-item recognition is to identify the locations of reactants, products, arrows, reaction conditions, connectors and molecular labels in the chemical reaction formula image;

[0103] The purpose of chemical reaction formula recognition is to determine the properties of molecules identified in chemical reaction sub-item recognition, namely: reactants, products, reaction conditions and molecular labels;

[0104] The purpose of the molecular structure recognition is to convert the molecular structure image identified in the chemical reaction sub-item recognition into a machine-readable format (mol code).

[0105] The specific method for identifying the chemical reaction sub-item is:

[0106] (1) manually marking the positions of molecules, arrows, reaction conditions, labels, and connectors (such as "+", etc.) in the chemical reaction formula image segmented in step 2, thereby establishing an initial training set for chemical reaction sub-item recognition;

[0107] (2) Using labeled chemical reaction sub-item image data to establish a reaction sub-item recognition machine learning model;

[0108] (3) Automatically identify the reaction sub-items in the segmented chemical reaction formula images using a chemical reaction sub-item recognition machine learning model;

[0109] (4) Manually proofread the results of the automatic recognition of the chemical reaction sub-item recognition machine learning model, and manually re-label the images that were misrecognized and add them to the training set of chemical reaction sub-item recognition;

[0110] (5) Using the updated training data of chemical reaction sub-item recognition, the reaction sub-item recognition machine learning model is continuously updated and iterated until the manual proofreading error rate is lower than the set threshold.

[0111] In this embodiment, 500 chemical reaction formula modes are randomly selected from the reaction modes segmented in step 2 for manual annotation. It is required to mark the location of information such as reactants, products, arrows, reaction conditions, connectors, and molecular labels in the picture. The machine learning model for target detection is trained using the data of the manually annotated reaction sub-items, thereby obtaining a machine learning model for chemical reaction sub-item recognition, and realizing the recognition of molecules, reaction conditions, arrows, molecular labels, etc. in the reaction formula picture.

[0112] The established chemical reaction sub-item recognition model is used to identify reaction sub-items in subsequent chemical reaction images. During the recognition process, all recognition results are manually proofread; after the error results are corrected and all manually proofread data are added to the training set, new machine learning models are retrained regularly.

[0113] In this embodiment, the result of identifying the chemical reaction sub-item in the chemical reaction modal image is as follows: Figure 4 As shown, the chemical reaction picture contains 3 molecules, 2 sets of conditions, 1 arrow, 1 connector and 3 molecular labels.

[0114] The chemical reaction formula recognition process is performed according to the relative positions of the molecules and the arrows using the following rule-based method:

[0115] 1) Each arrow represents a chemical reaction;

[0116] 2) The molecules at the starting point of the arrow are reactants, the molecules at the end point of the arrow are products, and the other texts at the vertical direction of the arrow are reaction conditions;

[0117] 3) The molecules before and after or above and below the symbol “+” have the same properties;

[0118] 4) Based on the principle of closest distance, define the molecule to which each molecular tag belongs;

[0119] The reaction formulas identified according to the above rules are manually proofread, and the judgment rules are modified according to the proofreading results until the manual proofreading error rate is lower than the set threshold.

[0120] In this embodiment, the reactants and products of the chemical reaction are determined based on the arrow directions and connectors identified in the chemical reaction sub-item identification process. During the identification process, all identification results are manually proofread, and the identification rules are regularly revised based on erroneous results.

[0121] In this embodiment, according to the above rules, Figure 4 The chemical reaction reactants and products obtained by extracting the chemical reaction shown are as follows Figure 5 As shown;

[0122] The specific method of molecular structure recognition is:

[0123] (a) Artificially draw the molecular structure diagram identified in the chemical reaction sub-item recognition, so as to establish the initial training set for molecular structure recognition;

[0124] (b) Establishing a machine learning model for molecular structure recognition using artificially drawn molecular structure graph data;

[0125] (c) Automatically identify new molecular structure diagrams using molecular structure recognition machine learning models;

[0126] (d) manually proofreading the recognition results, and manually modifying or redrawing the corrected results and adding them to the molecular structure recognition training set;

[0127] (e) Using the updated molecular structure recognition training data, the molecular structure recognition machine learning model is continuously updated and iterated until the manual proofreading error rate is lower than the set threshold.

[0128] In this embodiment, 500 molecular images are randomly selected and the molecular structure diagram is manually drawn. The target detection model is established using the data to detect the chemical bonds and atomic symbols in the image, and the type and direction of the chemical bonds are classified at the same time, thereby obtaining a molecular structure recognition machine learning model; then the mol code is generated using the nearest distance principle and the mol code encoding rules. The established molecular structure recognition machine learning model is used for the subsequent identification of chemical molecular structures. During the recognition process, all recognition results are manually proofread; after the error results are corrected and all manually proofread data are added to the training set, the new machine learning model is retrained regularly.

[0129] In this embodiment, the molecular structure recognition machine learning model identifies Figure 4 The molecular structure information is shown in Table 1.

[0130] Table 1 Molecular structure recognition results

[0131]

[0132] In this embodiment, the mol code generated corresponding to the first molecule image in the first column of Table 1 is as follows:

[0133]

[0134] The generated mol code corresponding to the second molecular image in the first column of Table 1 is as follows:

[0135]

[0136]

[0137] The generated mol code corresponding to the third molecular image in the first column of Table 1 is as follows:

[0138]

[0139]

[0140] Step 3.2: Table parsing: Parse the table image segmented in step 2, and match the reaction conditions and remarks information according to the custom terms;

[0141] The purpose of the table parsing is to identify the table in the image format identified in step 2 as a machine-readable format. The specific method of the recognition process is:

[0142] Step S321) using an open source table parsing machine learning model to parse the table image;

[0143] Step S322) Manually perform random checks on the results of table analysis, manually re-mark the tables with random checks errors, and add them to the table analysis training set;

[0144] Step S323) Continuously update and iterate the table parsing machine learning model using the updated table parsing training data until the manual sampling error rate is lower than the set threshold.

[0145] In this embodiment, the open source table parsing machine learning model PP-Structure is used to parse the table. During the parsing process, about 10% of the parsing results are manually sampled. If the accuracy rate is greater than or equal to 95%, the parsing continues; if the accuracy rate is less than 95%, the open source machine learning model is optimized with the erroneous results corrected and all manually proofread data until the error rate is lower than the threshold.

[0146] In this example, the analysis results of the segmented table are shown in Table 2.

[0147] Table 2 Table analysis results

[0148]

[0149]

[0150] Step 3.3: Text parsing: parse the condition description text image segmented in step 2, and match the reaction conditions and remarks information according to the custom dictionary;

[0151] The purpose of the text parsing is to identify the condition description text in the picture format identified in step 2 into a machine-readable format. The specific method of the identification process is:

[0152] Step S331: using an open source text parsing machine learning model to parse the condition description text image;

[0153] Step S332: Manually check the results of text analysis, manually re-label the erroneous condition description text images, and add them to the text analysis training set;

[0154] Step S333: Continuously update and iterate the text parsing machine learning model using the updated text parsing training data until the manual sampling error rate is lower than the set threshold.

[0155] In this embodiment, the open source text parsing machine learning model PP-OCRv4 is used to parse the text. During the parsing process, about 10% of the text parsing results are manually sampled. If the accuracy rate is greater than or equal to 95%, the parsing continues; if the accuracy rate is less than 95%, the open source machine learning model is optimized with the erroneous results corrected and all manually proofread data until the error rate is lower than the threshold.

[0156] In this embodiment, the parsing results of the segmented text are shown in Table 3.

[0157] Table 3 Text parsing example results

[0158]

[0159] Step 4: Reaction information integration: Integrate the information parsed in step 3 into complete chemical reaction information; the integrated chemical reaction information includes: reactants, products, solvents, catalysts, reagents, molecular labels, reaction temperature, reaction time, yield and remarks;

[0160] The purpose of reaction information integration is to integrate the chemical reaction information extracted and analyzed by all the above processes according to the set data categories. The specific method is as follows:

[0161] Step S41: Integrate the molecular structure diagram extracted in step 3.1 into the reaction information of reactants, products, solvents, catalysts and reagents according to the analyzed properties;

[0162] Step S42: The text information identified in step 3.1 and the text information parsed in step 3.3 are combined and processed, and the matching solvent, catalyst, reagent, molecular label, reaction temperature, reaction time and yield information are automatically extracted according to the custom dictionary using an open source semantic parsing model; the unmatched information is classified as the remark information;

[0163] Step S43: Integrate the table information parsed in step 3.3 into reaction condition information; create a table header dictionary, establish the correspondence between the table header and the reaction information attribute, and automatically extract the matching solvent, catalyst, reagent, molecular label, reaction temperature, reaction time and yield information; classify the unmatched information into the remark information;

[0164] Step S44: sampling the results of reaction information integration and correcting erroneous results;

[0165] Step S45: Integrate the data according to the modified reaction information and re-execute steps S41 to S44 until the sampling error rate is lower than the set threshold.

[0166] In this embodiment, the molecular structure diagram extracted in step 3.1 is integrated into reaction information such as reactants and products according to the parsed attributes. The text information uses the open source semantic parsing model ChemRxnExtractor to automatically extract information such as solvents, catalysts, reagents, molecular tags, reaction temperature, reaction time, and yield. The table information automatically extracts information such as solvents, catalysts, reagents, molecular tags, reaction temperature, reaction time, and yield by creating a header dictionary and establishing a correspondence between the header and the reaction information attributes. Information that is not automatically extracted is classified as remarks information.

[0167] In this embodiment, during the reaction information integration process, about 10% of the results are manually sampled. If the accuracy rate is greater than or equal to 95%, the integration will continue; if the accuracy rate is less than 95%, the integration rules will be optimized after the erroneous results are corrected and all manually proofread data are used.

[0168] In this embodiment, the chemical reaction information integration includes reaction formula information integration and reaction condition information integration. The integration results are shown in Tables 4 and 5.

[0169] Table 4 Reactive information integration results

[0170]

[0171]

[0172] Table 5 Reaction condition information integration results

[0173]

[0174] Step 5: Data storage;

[0175] The reaction information extracted, segmented, parsed and integrated in steps 1 to 4 is stored in an optimized manner to facilitate subsequent use and updating.

[0176] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A method for extracting chemical reaction information from journal articles, characterized in that: For the chemical reaction diagrams and their attached charts in journal articles, the extraction process is decomposed into multiple steps, including reaction information unit extraction, reaction information modal segmentation, chemical reaction formula analysis, table analysis, text analysis, reaction information integration, and data storage. The following steps are included: Step 1: extracting a reaction information unit from a journal document; the reaction information unit is a complete paper diagram containing at least one chemical reaction and its attached tables and condition description text in the journal document; Step 2: Perform reaction information modality segmentation on the reaction information unit; automatically segment the reaction information unit extracted in step 1 into three types of image information: chemical reaction formula, table, and condition description text; Step 3: Analyze the chemical reaction formula image, table image and condition description text image segmented in step 2 respectively; Step 3.1: Chemical reaction formula analysis: analyze the chemical reaction formula image segmented in step 2, identify the reactants, products, arrows, reaction conditions, connectors, and molecular label information, and identify the molecular structure diagram expressed in the form of an image into a machine-readable format; Step 3.2: Table parsing: Parse the table image segmented in step 2, and match the reaction conditions and remarks information according to the custom terms; Step 3.3: Text parsing: parse the condition description text image segmented in step 2, and match the reaction conditions and remarks information according to the custom dictionary; Step 4: Reaction information integration: Integrate the information parsed in step 3 into complete chemical reaction information; the integrated chemical reaction information includes: reactants, products, solvents, catalysts, reagents, molecular labels, reaction temperature, reaction time, yield and remarks; Step 5: Data storage: The reaction information extracted, segmented, parsed and integrated in steps 1 to 4 is stored in an optimized manner to facilitate subsequent use and updating.

2. A method for extracting chemical reaction information from journal articles according to claim 1, characterized in that: The specific method of extracting from the journal document in step 1 is: Step S11: converting all pages of the journal document containing chemical reaction information into image format; Step S12: manually marking the positions of the reaction information units in the image, thereby establishing an initial training set for reaction information unit extraction; Step S13: using the annotated reaction information unit image data to establish a reaction information unit extraction machine learning model; Step S14: using a reaction information unit extraction machine learning model to automatically extract reaction information units from new journal images; Step S15: sampling the results extracted by the reaction information unit extraction machine learning model, manually re-labeling the pictures with extraction errors and adding them to the training set; Step S16: Continuously update the iterative reaction information unit extraction machine learning model using the updated training data until the manual sampling error rate is lower than the set threshold.

3. A method for extracting chemical reaction information from journal articles according to claim 1, characterized in that: The specific method of step 2 for segmenting the reaction information modality is: Step S21: manually marking the positions of chemical reaction formulas, tables, and texts in the reaction information unit images extracted in step 1, thereby establishing an initial training set for reaction information modality segmentation; Step S22: establishing a reaction information modality machine learning model using the annotated reaction information modality image data; Step S23: automatically segmenting the reaction information modality in the unlabeled reaction information unit image using the reaction information modality machine learning model; Step S24: manually sampling the results of the segmentation of the reaction information modality machine learning model, manually re-labeling the images with segmentation errors and adding them to the training set of the reaction information modality segmentation; Step S25: Continuously update and iterate the reaction information modality machine learning model using the updated reaction information modality segmentation training data until the manual sampling error rate is lower than the set threshold.

4. A method for extracting chemical reaction information from journal articles according to claim 1, characterized in that: The chemical reaction formula analysis described in step 3.1 includes three sub-processes: chemical reaction sub-item identification, chemical reaction formula identification and molecular structure identification; Among them, the purpose of chemical reaction sub-item recognition is to identify the locations of reactants, products, arrows, reaction conditions, connectors and molecular labels in the chemical reaction formula image; The purpose of chemical reaction formula recognition is to determine the properties of molecules identified in chemical reaction sub-item recognition, namely: reactants, products, reaction conditions and molecular labels; The purpose of the molecular structure recognition is to convert the molecular structure image identified in the chemical reaction sub-item recognition into a machine-readable format.

5. A method for extracting chemical reaction information from journal articles according to claim 4, characterized in that: The specific method for identifying the chemical reaction sub-item is: (1) manually annotating the positions of molecules, arrows, reaction conditions, labels, and connectors in the chemical reaction formula image segmented in step 2, thereby establishing an initial training set for chemical reaction sub-item recognition; (2) Using labeled chemical reaction sub-item image data to establish a reaction sub-item recognition machine learning model; (3) Automatically identify the reaction sub-items in the segmented chemical reaction formula images using a chemical reaction sub-item recognition machine learning model; (4) Manually proofread the results of the automatic recognition of the chemical reaction sub-item recognition machine learning model, and manually re-label the images that were misrecognized and add them to the training set of chemical reaction sub-item recognition; (5) Using the updated training data of chemical reaction sub-item recognition, the reaction sub-item recognition machine learning model is continuously updated and iterated until the manual proofreading error rate is lower than the set threshold.

6. A method for extracting chemical reaction information from journal articles according to claim 4, characterized in that: The chemical reaction formula recognition process is performed according to the relative positions of the molecules and the arrows using the following rule-based method: 1) Each arrow represents a chemical reaction; 2) The molecules at the starting point of the arrow are reactants, the molecules at the end point of the arrow are products, and the other texts at the vertical direction of the arrow are reaction conditions; 3) The molecules before and after or above and below the "+" symbol have the same properties; 4) Based on the principle of closest distance, define the molecule to which each molecular tag belongs; The reaction formulas identified according to the above rules are manually proofread, and the judgment rules are modified according to the proofreading results until the manual proofreading error rate is lower than the set threshold.

7. A method for extracting chemical reaction information from journal articles according to claim 4, characterized in that: The specific method of molecular structure recognition is: (a) Artificially draw the molecular structure diagram identified in the chemical reaction sub-item recognition, so as to establish the initial training set for molecular structure recognition; (b) Establish a machine learning model for molecular structure recognition using artificially drawn molecular structure graph data; (c) Automatically identify new molecular structure diagrams using molecular structure recognition machine learning models; (d) manually proofreading the recognition results, and manually modifying or redrawing the corrected results and adding them to the molecular structure recognition training set; (e) Using the updated molecular structure recognition training data, the molecular structure recognition machine learning model is continuously updated and iterated until the manual proofreading error rate is lower than the set threshold.

8. A method for extracting chemical reaction information from a journal article according to claim 1, characterized in that: The purpose of the table parsing in step 3.2 is to identify the table in the image format identified in step 2 as a machine-readable format. The specific method of the identification process is: Step S321) using an open source table parsing machine learning model to parse the table image; Step S322) Manually perform random checks on the results of table analysis, manually re-mark the tables with random checks errors, and add them to the table analysis training set; Step S323) Continuously update and iterate the table parsing machine learning model using the updated table parsing training data until the manual sampling error rate is lower than the set threshold.

9. A method for extracting chemical reaction information from journal articles according to claim 1, characterized in that: The purpose of text parsing in step 3.3 is to identify the condition description text in the image format identified in step 2 into a machine-readable format. The specific method of the identification process is: Step S331: using an open source text parsing machine learning model to parse the condition description text image; Step S332: Manually check the results of text analysis, manually re-label the erroneous condition description text images, and add them to the text analysis training set; Step S333: Continuously update and iterate the text parsing machine learning model using the updated text parsing training data until the manual sampling error rate is lower than the set threshold.

10. The method for extracting chemical reaction information from journal articles according to claim 1, characterized in that: The purpose of step 4 reaction information integration is to integrate the chemical reaction information extracted and analyzed by all the above processes according to the set data categories. The specific method is: Step S41: Integrate the molecular structure diagram extracted in step 3.1 into the reaction information of reactants, products, solvents, catalysts and reagents according to the analyzed properties; Step S42: The text information identified in step 3.1 and the text information parsed in step 3.3 are combined and processed, and the matching solvent, catalyst, reagent, molecular label, reaction temperature, reaction time and yield information are automatically extracted according to the custom dictionary using an open source semantic parsing model; the unmatched information is classified as the remark information; Step S43: Integrate the table information parsed in step 3.3 into reaction condition information; create a table header dictionary, establish the correspondence between the table header and the reaction information attribute, and automatically extract the matching solvent, catalyst, reagent, molecular label, reaction temperature, reaction time and yield information; classify the unmatched information into the remark information; Step S44: sampling the results of reaction information integration and correcting erroneous results; Step S45: Integrate the data according to the modified reaction information and re-execute steps S41 to S44 until the sampling error rate is lower than the set threshold.

Images