Block chain-based medical department data management method and system

Abstract

The invention provides a block chain-based medical department data management method and system, and the method comprises the steps: obtaining K to-be-analyzed text paragraphs which are text paragraphs which are arranged in sequence, carrying out the text advanced transformation of each to-be-analyzed text paragraph in the K to-be-analyzed text paragraph, and obtaining the to-be-analyzed text paragraph in the K to-be-analyzed text paragraph; the method comprises the following steps: acquiring a to-be-analyzed text paragraph, acquiring advanced transformation texts corresponding to the to-be-analyzed text paragraph, performing text combination on K advanced transformation texts to acquire K combined texts, loading the K combined texts into a target learning network, and acquiring an information identification result of to-be-identified information in the to-be-analyzed text paragraph, the target learning network is a neural network obtained by performing multiple rounds of optimization on a transformation training text and a first text indication mark corresponding to the first training text, and a basic transformation training text and an advanced transformation training text corresponding to the second training text, and the second text indication mark is a network learning sample.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a blockchain-based medical science and technology data management method and system. Background Art

[0002] In the pharmaceutical field, with the rapid development of science and technology, the importance of pharmaceutical scientific and technological data management is increasing day by day. The pharmaceutical industry chain covers many complex links such as research and development, production, and sales, and each link will generate massive amounts of data. These data are of great significance to the management of many aspects of the pharmaceutical industry, such as quality monitoring, safety supervision, intellectual property protection, and market circulation control.

[0003] With the continuous development of the pharmaceutical industry and the deepening of research, the types and amount of information in the pharmaceutical industry chain have increased dramatically. For example, the emergence of new drug development, new production processes, and the expansion of the pharmaceutical market have made pharmaceutical data increasingly rich and diverse. When managing pharmaceutical scientific and technological data, especially when tracing information, it is necessary to accurately identify various types of information in the data. However, the workload required for accurate labeling of pharmaceutical data is huge, because pharmaceutical data contains complex R&D processes, production links, sales paths and other information. It is very difficult to fully label these data. Therefore, a method is needed to improve the performance of the model by using unlabeled data based on limited labeled data. For example, different transformations are performed on some labeled data and unlabeled data. First, a more reliable temporary label is generated on the data that has undergone a transformation using an existing model, and then the model for information recognition is trained by combining the data that has undergone another transformation and these temporary labels, and then the target information in the unlabeled data is recognized using this model. However, there are problems with this approach. In the management of medical science and technology data, existing methods require that the results of specific transformations of the data be consistent in length when processing medical data. This can lead to the loss of some key information in the information recognition process (such as entity location information in medical information traceability, etc.), which in turn leads to large recognition errors and the inability to accurately identify the target information in the medical data (such as traceability information), affecting the accuracy and effectiveness of medical science and technology data management. Summary of the invention

[0004] In view of this, the present application provides a blockchain-based pharmaceutical science and technology innovation data management method and system. The technical solution of the present application is implemented as follows: On the one hand, the present application provides a blockchain-based pharmaceutical science and technology data management method, including: in response to a pharmaceutical traceability request, obtaining a pharmaceutical industry chain record text in a block node matched by the pharmaceutical traceability request, wherein the pharmaceutical industry chain record text includes K text paragraphs to be analyzed, wherein the K text paragraphs to be analyzed are text paragraphs arranged in sequence, and K is a natural number greater than 1; performing a high-level text transformation on each of the K text paragraphs to be analyzed to obtain a high-level transformation text corresponding to the text paragraph to be analyzed; performing text combination on the K high-level transformation texts to obtain K combined texts; loading the K combined texts into a target learning network to obtain an information recognition result of the information to be recognized in the text paragraph to be analyzed, wherein the target learning network is a neural network obtained by performing multiple rounds of optimization on the transformation training text corresponding to the first training text and the first text indication mark, and the basic transformation training text corresponding to the second training text and the high-level transformation training text, and the second text indication mark as network learning samples, wherein the first training text is a text carrying the first text indication mark, and the second text indication mark is a temporary mark obtained after the target guidance network performs information recognition reasoning on the basic transformation training text corresponding to the second training text.

[0005] On the other hand, the present application provides a computer system, including a memory and a processor, wherein the memory stores a computer program executable on the processor, and the processor implements the steps in the above method when executing the program.

[0006] Beneficial effects of the present application: In the embodiment of the present application, the target learning network is a neural network obtained by performing multiple rounds of optimization on the transformation training text corresponding to the first training text and the first text indicator mark, and the basic transformation training text, advanced transformation training text, and second text indicator mark corresponding to the second training text as network learning samples. In addition, the first training text is a text carrying the first text indicator mark, and the second text indicator mark is a temporary mark obtained after the target guidance network performs information recognition reasoning on the basic transformation training text corresponding to the second training text. Based on this, after obtaining K sequentially arranged text paragraphs to be analyzed, each of the K text paragraphs to be analyzed is subjected to a text high-level transformation to obtain the high-level transformation text corresponding to each text paragraph to be analyzed, and thus K high-level transformation texts are obtained. Then, the K high-level transformation texts are combined to obtain K combined texts, and the K combined texts are loaded into the target learning network at the same time, so as to identify the target information of the K combined texts based on the target learning network, and obtain the information recognition result of the information to be recognized in the text paragraphs to be analyzed. The embodiment of the present application does not require that the text lengths of the above-mentioned K high-level transformation texts be consistent, but instead adopts text combination of the K high-level transformation texts, so that the combined K combined texts contain the information in each high-level transformation text, and retain the entity location information in the target information recognition to the maximum extent. When the target learning network performs information recognition on the K combined texts, the semantic errors caused by missing sentences can be reduced, and the accuracy of the information recognition results retrieved can be improved. In this way, when tracing the source of medical information, more accurate traceability information can be automatically provided. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 A schematic diagram of the implementation process of a blockchain-based pharmaceutical science and technology innovation data management method provided in an embodiment of the present application.

[0008] Figure 2 A hardware entity diagram of a computer system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0009] The present application embodiment provides a blockchain-based pharmaceutical science and technology data management method, which can be executed by a processor of a computer system. The computer system can refer to a device with data processing capabilities such as a server, a laptop, a tablet computer, a desktop computer, a mobile device, etc. Figure 1 A schematic diagram of the implementation process of a blockchain-based pharmaceutical science and technology innovation data management method provided in an embodiment of the present application is as follows: Figure 1 As shown, the method includes: Step S100: In response to a pharmaceutical traceability request, obtain a pharmaceutical industry chain record text in a block node that matches the pharmaceutical traceability request. The pharmaceutical industry chain record text includes K text paragraphs to be analyzed. The K text paragraphs to be analyzed are text paragraphs arranged in sequence, and K is a natural number greater than 1.

[0010] In the embodiment of the present application, the pharmaceutical traceability request is an externally initiated query demand, which aims to obtain the traceability information of pharmaceutical-related products or processes. For example, in the field of pharmaceutical production, it may involve a traceability query of information such as the source of raw materials of a particular drug, various links in the production process, quality test results, and sales channels. This request can be initiated by the drug regulatory department to ensure the quality and safety of the drug; it can also be initiated by the pharmaceutical company itself for internal management and quality control; it may even be initiated by consumers to understand the detailed information of the purchased drugs. After receiving the pharmaceutical traceability request, the computer system starts the corresponding processing mechanism, first parses the request, and identifies the key information in the request, such as the name of the drug to be traced, the production batch, the time range of the traceability, etc. Taking a batch of influenza vaccine as an example, the pharmaceutical traceability request may contain key information such as "Influenza vaccine-XYZ batch-2023 production-from raw material procurement to finished product sales full process traceability". After the computer system parses this information, it can determine the direction and goal of subsequent operations.

[0011] In blockchain technology, block nodes are the basic units that make up the blockchain network. Each block node stores specific data blocks, which contain various information related to pharmaceutical science and technology innovation. This information is stored on the node in an encrypted form, ensuring the security and non-tamperability of the data. For example, a block node may store the production data of a pharmaceutical company in a specific period of time, including raw material supplier information, operating parameters of production equipment, and operation records of production personnel.

[0012] Different block nodes may be responsible for storing different types or stages of medical data. For example, some nodes are dedicated to storing data in the research and development stage, such as the clinical trial results of new drugs; some nodes focus on storing data in the production stage, such as environmental monitoring data in the production workshop.

[0013] The pharmaceutical industry chain record text is a data record in text form that contains information from all links in the pharmaceutical industry chain. It covers relevant information from the entire industry chain, from pharmaceutical research and development, raw material procurement, manufacturing, quality testing to sales and after-sales. Taking a certain anti-cancer drug as an example, its pharmaceutical industry chain record text may include the molecular structure design of the drug during the research and development process, data in animal and human clinical trials (such as the effectiveness and side effects of the drug); in the raw material procurement link, the origin of the raw materials, the qualifications of the supplier and other information; in the production and manufacturing link, the production process, the sanitary standards of the production environment, etc.; in the quality inspection link, various test indicators and test results; in the sales link, sales channels, sales prices, etc.; in the after-sales link, patient feedback, adverse reaction reports and other information.

[0014] The pharmaceutical industry chain record text consists of K text paragraphs to be analyzed, where K is a natural number greater than 1, indicating the number of text paragraphs. These paragraphs are arranged in sequence. For example, in a pharmaceutical industry chain record text about the production of a certain antibiotic, the first paragraph may be the relevant paragraphs of the R&D stage, which describes the R&D process of the antibiotic, the research results of the R&D team, etc.; followed by the raw material procurement paragraph, which introduces the source of the antibiotic raw materials (such as the components of the culture medium for fermentation, etc.); then the production and manufacturing paragraph, which details the equipment, production process, quality control measures, etc. of the production workshop; there may be quality inspection paragraphs, sales paragraphs, and after-sales paragraphs in the follow-up. These paragraphs are arranged in the logical order of the pharmaceutical industry chain, forming a complete pharmaceutical industry chain record text.

[0015] In summary, in step S100, the computer system accurately locates the matching block node by responding to the medicine traceability request, and then obtains the medicine industry chain record text containing K sequentially arranged text paragraphs to be analyzed.

[0016] Step S200: performing high-level text transformation on each of the K to-be-analyzed text paragraphs to obtain high-level transformed text corresponding to the to-be-analyzed text paragraph.

[0017] In step S200 of the embodiment of the present application, the computer system performs high-level text transformation on each of the K text paragraphs to be analyzed obtained in step S100, and thereby obtains high-level transformed text corresponding to each text paragraph to be analyzed.

[0018] The K text paragraphs to be analyzed in the pharmaceutical industry chain record text are text parts arranged in sequence according to different links or different logical relationships of the pharmaceutical industry chain. For example, in the pharmaceutical industry chain record text about a new anti-cancer drug, one of the text paragraphs to be analyzed may be a detailed record of the drug's research and development stage, including the ideas of the research and development team for designing and optimizing the molecular structure of the drug, the experimental methods used, and the preliminary results in early cell experiments; another text paragraph to be analyzed may focus on the production and manufacturing of the drug, describing the layout of the production workshop, the types of production equipment used, and the key control points in the production process.

[0019] Advanced text transformation is a complex data processing method implemented by computer systems on the text paragraphs to be analyzed. It aims to transform the original text paragraphs through specific algorithms or models to extract or highlight the features that are more conducive to subsequent analysis. This transformation is different from simple text format adjustment or basic text editing operations.

[0020] Taking the aforementioned text paragraph to be analyzed in the R&D stage of anticancer drugs as an example, assume that the original text paragraph contains a large number of R&D details, such as "When determining the molecular structure of the anticancer drug, the R&D team referred to a variety of existing similar drug structures. After multiple computer simulation analyses, it was found that replacing some groups in the original structure with a certain chemical group can enhance the drug's targeting of cancer cells. Therefore, a series of in vitro cell experiments were conducted. In the experiment on the lung cancer cell line A549, different concentrations of candidate drugs were used for treatment, and cell viability was detected by MTT method. It was found that when the drug concentration reached xμmol / L, the cell viability decreased significantly, indicating that the drug had a certain inhibitory effect on the cell line. At the same time, it was also observed during the experiment that the drug can affect certain signaling pathways of the cell, such as the reduced activity of the PI3K-AKT signaling pathway." When a computer system performs advanced text transformation, it can reorganize, associate, or extract features from key information. For example, the relationship between drug molecular structure design and cell experiment results can be strengthened, and the content about molecular structure replacement and decreased cell viability, signal pathway changes, etc. can be presented in a more compact and relevant way. The text can be transformed into a form that emphasizes cause and effect, such as "By analyzing the structure of existing similar drugs, specific chemical groups in the molecular structure were replaced. When the lung cancer cell line A549 was tested by the MTT method, the cell viability was significantly decreased when the drug concentration reached xμmol / L, and the activity of the PI3K-AKT signaling pathway was reduced, indicating that this change in molecular structure has an inhibitory effect on cancer cells." After the computer system performs advanced text transformation, it can focus on the relationship between the parameters of the production equipment and the production process. For example, it can be transformed into "the high-precision reactor (volume z liters, can accurately control temperature and pressure) in the production workshop is coordinated with the automated filling production line. After the raw materials are put into the reactor in proportion and react at a specific temperature and pressure, they are filtered and purified and then packaged by the filling line. The functional area division of the workshop assists the entire production process." By performing such high-level text transformation on each text paragraph to be analyzed, the computer system can obtain the high-level transformation text corresponding to each text paragraph to be analyzed. Compared with the original text paragraph to be analyzed, these high-level transformation texts highlight the relationship between key information and remove some possible redundant information, providing a more valuable data foundation for the in-depth analysis, combination and identification of target information of the pharmaceutical industry chain in subsequent steps. This operation helps to improve the efficiency and accuracy of the entire pharmaceutical science and technology data management method, especially when dealing with complex pharmaceutical industry chain data, it can better mine the potential information behind the data to meet various needs such as pharmaceutical traceability.

[0021] Step S300: performing text combination on K high-level transformation texts to obtain K combined texts.

[0022] After performing high-level text transformation on each text paragraph to be analyzed and obtaining the high-level transformation text corresponding to each text paragraph to be analyzed, these K high-level transformation texts can also be combined to obtain K combined texts. Each combined text in the K combined texts can be understood as a fusion of part of the sub-text in each of the K high-level transformation texts. For example, if K=4, the 4 high-level transformation texts can include high-level transformation text 1 to high-level transformation text 4, and the 4 combined texts can include combined text 1 to combined text 4. At this time, combined text 1 can be considered to be a fusion of the sub-text in high-level transformation text 1, the sub-text in high-level transformation text 2, the sub-text in high-level transformation text 3, and the sub-text in high-level transformation text 4. Similarly, for the remaining combined texts 2 to combined text 4, refer to the introduction of combined text 1.

[0023] In one implementation, step S300 may also be implemented by the following steps: Step S310: decomposing each high-level transformation text into T clauses, where T is a natural number greater than 1; Step S320: perform clause integration on the T clauses in the K advanced transformation texts to obtain K first importance information, each of which includes T first attention scores, and the a-th first attention score among the T first attention scores indicates that the text position where the a-th first attention score is located corresponds to the a-th clause in the d-th advanced transformation text, d is equal to the a-th first attention score, d is a natural number greater than 0 and less than K, and a is a natural number greater than 0 and less than T; Step S330: Generate a corresponding combined text according to each first importance information in the K first importance information to obtain K combined texts.

[0024] In step S310, the computer system decomposes each high-level transformation text into T clauses, where T is a natural number greater than 1. The high-level transformation text is the result of high-level text transformation of the text paragraph to be analyzed in the medical industry chain record text after the previous steps. It has reorganized and highlighted the key information in the original text. For example, in a high-level transformation text about a certain drug production link, it may contain information on production process, quality control, equipment use, etc. The computer system decomposes this high-level transformation text into clauses, which are smaller semantic units in the high-level transformation text. For example, the description of "in terms of production process, first pre-process the raw materials to ensure that their purity meets the standard" can be used as a clause, and "in the quality control link, samples on the production line are sampled every hour" can also be a clause. This decomposition operation helps the computer system to analyze the information structure in the high-level transformation text more carefully, laying the foundation for subsequent combination operations.

[0025] In step S320, the computer system performs clause integration on the T clauses in the K high-level transformation texts, thereby obtaining K first importance information, each of which includes T first attention scores. The first attention score is an indicator used to measure the importance of each clause in the integration process. Taking a case containing three high-level transformation texts (K=3) as an example, assume that the first high-level transformation text is about drug research and development, the second is about production, and the third is about sales. Each high-level transformation text is decomposed into T=5 clauses. During the clause integration process, the computer system determines the first attention score of each clause based on the logical relationship of the pharmaceutical industry chain and the correlation between information. For example, for the clause "determine the effective dosage range of the drug through a large number of clinical trials" in the high-level transformation text of the R&D stage, if this clause is very critical to understanding the characteristics of the drug in the context of the entire traceability or information analysis, then it can be assigned a higher first attention score. For the clause "the sales team has developed a preliminary marketing plan" in the high-level transformation text of the sales stage, if it has a relatively weak relationship with the core information to be identified (such as the safety traceability of the drug), it can be assigned a lower first attention score. These first attention scores have clear meanings, where the ath first attention score indicates that the text position of the ath first attention score corresponds to the ath clause in the dth advanced transformation text, and d is equal to the ath first attention score. For example, the 3rd first attention score corresponds to the 3rd clause in the 3rd advanced transformation text. In this way, the computer system can accurately mark the importance position relationship of each clause in different advanced transformation texts. This relationship reflects the relative importance of the clause in the entire pharmaceutical industry chain information system, which helps to screen out the most valuable information for constructing a combined text.

[0026] Finally, step S330 is performed, in which the computer system generates a corresponding combined text according to each of the K first importance information to obtain K combined texts. When generating the combined text, the computer system will comprehensively consider each clause in each first importance information and their corresponding first attention scores. Continuing with the above example, if in the first importance information of the research and development stage, the clause "determine the target of the drug" has a higher first attention score, in the first importance information of the production stage, the clause "produce in strict accordance with GMP standards" has a higher first attention score, and in the first importance information of the sales stage, the clause "locate the target customer group according to the characteristics of the drug" has a higher first attention score, then when generating the combined text, the computer system will give priority to combining these clauses with higher importance scores in a reasonable logical order. For example, the combined text can be constructed in a logical order such as "after the target of the drug is determined, it is produced in strict accordance with GMP standards, and then the target customer group is located according to the characteristics of the produced drug". At the same time, for those clauses with low first attention scores but still potentially complementary to the overall information, they will also be appropriately incorporated into the combined text according to certain logic and importance relationships. Each of the K combined texts generated in this way can comprehensively and reasonably integrate key information from different high-level transformation texts, highlighting important information while taking into account the overall logic and integrity, providing a high-quality data foundation for subsequent steps (such as loading the combined text into the target learning network for identification of the information to be identified).

[0027] Step S400: K combined texts are loaded into a target learning network to obtain information recognition results of information to be recognized in the text paragraph to be analyzed. The target learning network is a neural network obtained by performing multiple rounds of optimization on the transformation training text and the first text indication mark corresponding to the first training text, and the basic transformation training text and the advanced transformation training text and the second text indication mark corresponding to the second training text as network learning samples. The first training text is a text carrying the first text indication mark, and the second text indication mark is a temporary mark obtained after the target guidance network performs information recognition reasoning on the basic transformation training text corresponding to the second training text.

[0028] In the embodiment of the present application, the target learning network is a neural network that has been optimized for multiple rounds. The construction of this network is based on a large number of training samples, and is intended to accurately identify the information of the input text data. The target learning network has a complex structure and algorithm mechanism. It can conduct in-depth analysis of the input combined text and explore the hidden information relationships therein.

[0029] In summary, the present application performs high-level text transformation on the acquired text paragraphs. The semantic information in medical data is complex and difficult to parse directly. This transformation helps to dig deeper semantic content and make text information easier to process. When processing K high-level transformation texts, the text length is not restricted to be consistent and the text is combined to obtain K combined texts. This operation mode ensures that the combined text can contain the information in each high-level transformation text, and the entity location information can be retained to the maximum extent. In the medical information traceability scenario, entity location information is a key element for accurate traceability. Traditional methods may cause sentence missing due to various restrictions (such as requiring a specific format or consistency), thereby causing semantic errors. This step of the present solution avoids this situation, ensures the integrity of the information, and provides a reliable basis for subsequent accurate recognition. Finally, the target learning network is carefully constructed and optimized, and it can be processed based on the complete information retained in the previous steps. Since the combined text in step S300 contains complete information, the target learning network can effectively avoid semantic errors caused by missing sentences when performing recognition, thereby accurately identifying the information to be recognized and improving the accuracy of the information recognition results. This overall process enables more accurate traceability information to be automatically provided when tracing the source of medical information, overcoming problems such as large identification errors and inaccurate information in medical data management.

[0030] As an implementation mode, step S400, loading K combined texts into the target learning network to obtain information recognition results of the information to be recognized in the text paragraph to be analyzed, may include: Step S410: performing text embedding coding on each of the K combined texts to obtain a text implicit representation of the corresponding combined text; Step S420: performing a text splitting operation on the K combined texts according to the text implicit representations in the K combined texts to obtain K split text sentences; Step S430: performing text embedding coding on each of the K split text sentences to obtain a text implicit representation of the corresponding split text sentence; Step S440: Decode and infer the text implicit representations of the K split text sentences to obtain information recognition results of the information to be recognized in the text paragraph to be analyzed.

[0031] In step S410, when processing K combined texts, the computer system first performs a text embedding coding operation. Text embedding coding is a technical means of converting text data into a low-dimensional vector representation. In the context of medical science and technology data management, the combined text contains processed information from different links in the pharmaceutical industry chain. For example, a combined text may describe the key technologies in the research and development process of a certain drug, the key points of quality control in the production process, and the market positioning in the sales process. The computer system will use a pre-trained word vector model (such as Word2Vec, GloVe, etc.) or a text embedding model based on deep learning (such as BERT, etc.) to process this combined text. Taking Word2Vec as an example, assuming that the vocabulary set in the combined text is , the Word2Vec model learns each word through training Vector representation of For the entire combined text, the computer system can obtain the implicit text representation of the combined text by combining the vocabulary vectors in the text (such as simple average, weighted average, or using a more complex neural network structure to combine) .

[0032] Assume that a simple average method is adopted, and the combined text T consists of vocabulary , then the text implicitly represents The calculation formula is: ; This text implicitly indicates It can capture the semantic information in the combined text in the form of a low-dimensional vector. It represents the lexical relationship and semantic content in the combined text in the form of a vector, providing an easy-to-process data structure for subsequent operations.

[0033] In step S420, after obtaining the implicit text representation of the combined text, the computer system performs a text splitting operation based on these representations. This operation is intended to further decompose the combined text into smaller semantic units, that is, split text sentences, so as to better understand different information parts in the text and perform decoupling processing.

[0034] For example, for the combined text about drugs mentioned above, its implicit representation may contain information about different links such as R&D, production and sales mixed together. Through text splitting operations, the computer system can separate the information from these different links.

[0035] The computer system can use the attention mechanism-based method to split the text. Assume that the text of the combined text is implicitly represented as , the computer system uses an attention function To determine the importance of information at different positions. This attention function can be a neural network model, such as a multi-layer perceptron (MLP).

[0036] set up , where W is the weight matrix, b is the bias vector, is an activation function (such as a sigmoid function). The obtained attention scores represent the importance of different parts in the combined text. Based on these attention scores, the computer system splits the combined text into K split text sentences. For example, if the combined text describes the technological innovation in the drug R&D process, the quality assurance measures in production, and the market strategy in sales, the computer system can split it into split text sentences about R&D technological innovation, split text sentences about production quality assurance, and split text sentences about sales market strategy, etc., by determining the importance of different parts through the attention mechanism.

[0037] In step S430, after obtaining K split text sentences, the computer system performs a text embedding coding operation on each split text sentence. This operation is similar to the text embedding coding in step S410, but is aimed at the smaller semantic unit after the split - the split text sentence.

[0038] For example, for a split text sentence about drug R&D technology innovation, which contains content such as "using new gene editing technology to improve drug efficacy", the computer system also uses a pre-trained word vector model or a deep learning text embedding model to process it.

[0039] Assuming that the same Word2Vec model as in step S410 is used, the vocabulary set in the split text sentence is , through the learned vocabulary vector , according to some combination method (such as the simple average mentioned above), the text implicit representation of the split text sentence is obtained .

[0040] This implicit textual representation of split text sentences can more accurately capture the semantic information in the split text sentences, because the split text sentences are more focused on a specific link in the pharmaceutical industry chain or a specific type of information than the combined text.

[0041] As an implementation mode, step S440, decoding and reasoning the text implicit representations of the K split text sentences to obtain the information recognition result of the information to be recognized in the text paragraph to be analyzed, includes: Step S441: performing target information retrieval on the text implicit representations of the K split text sentences to obtain entity location information of the information to be identified; Step S442: classify, decode and infer the text implicit representations of the K split text sentences and the entity position information of the information to be identified to obtain the information identification result of the information to be identified.

[0042] In step S441, the task of the computer system in this sub-step is to perform target information retrieval. Target information retrieval refers to the operation of searching for entities and their location information related to the information to be identified in the implicit text representation of K split text sentences. In the context of medical science and technology data management, the information to be identified is the specific medical-related content of concern, such as the name of a specific raw material in the production process of a certain drug, the name of a key technology in a certain research and development link, etc. The entity location information is used to accurately define the specific location range of these information to be identified in the text, which is crucial for subsequent accurate information identification and analysis.

[0043] The computer system can use technical means based on the vector space model. First, the text implicit representation of the split text sentence is a low-dimensional vector form, which represents the semantic information in the text sentence in a vector form. Assume that there are K split text sentences, and their text implicit representations are .

[0044] For the information to be identified, it is also converted into vector form and set as For example, if the information to be identified is "a certain drug ingredient X", the computer system converts "a certain drug ingredient X" into a vector using a pre-trained word vector model (such as Word2Vec, etc.) .

[0045] Then, the computer system calculates the similarity between the implicit text representation of each split text sentence and the information vector to be recognized. A commonly used similarity calculation method is cosine similarity, and its calculation formula is: ,in is the vector dot product, They are vector quantities Model.

[0046] By calculating this similarity, the computer system can determine which of the split text sentences are most relevant to the information to be recognized. The value of is the highest, then the split text sentence 3 is likely to contain entities related to the information to be identified.

[0047] Next, the computer system determines the entity location information of the information to be identified in the most relevant split text sentence. For example, if split text sentence 3 is "In the process of drug production, drug ingredient X and other auxiliary ingredients are added and mixed according to a specific process flow", the computer system can analyze the vector Zhongyu The relevant part is to determine the starting and ending positions of the entity "drug ingredient X" in the sentence, which is the entity position information.

[0048] Another technical approach is to use a retrieval model based on a neural network. A computer system can build a neural network specifically for information retrieval, such as a convolutional neural network (CNN) or a recurrent neural network (RNN).

[0049] Taking the convolutional neural network as an example, the text of the split text sentence is implicitly represented As the data of the input layer, the network extracts features and reduces the dimension of the input data through structures such as convolutional layers and pooling layers.

[0050] For the quasi-identification information vector , or it can be used as part of the network input or matched as a target feature. During the network training process, a large amount of labeled data (such as text sentence samples that are known to contain the information to be identified and its location) is used to learn how to accurately find the entities and their locations related to the information to be identified.

[0051] In step S442, after determining the entity location information of the information to be identified, the computer system performs classification decoding reasoning. The purpose of this operation is to finally classify and judge the information to be identified based on the implicit text representation of the K split text sentences and the entity location information of the information to be identified that has been obtained, so as to obtain the information recognition result. This result may be to determine whether a certain drug contains a specific ingredient, whether a certain research and development process uses a specific technology, etc., which is of great significance for the management, traceability, quality control and other aspects of medical science and technology data.

[0052] The computer system can use a logistic regression model for classification decoding reasoning. First, the text implicit representation of the K split text sentences and the entity position information of the information to be recognized are integrated. For example, assuming that the text implicit representation of the split text sentence is , the entity location information of the information to be identified is , the computer system combines them into a new vector , this vector contains the semantic information of the text sentence related to the information to be recognized and its position information in the text.

[0053] The basic form of the logistic regression model is ,in It is a parameter of the model. y=1 can represent a specific classification result, such as "the drug contains specific ingredients."

[0054] Suppose we want to determine whether a drug contains a specific drug ingredient X (this is the information to be identified). The computer system combines the split text sentences containing relevant information into vectors in the above way. , and then substitute it into the logistic regression model to calculate .if >0.5 (0.5 is a commonly used threshold value, which can also be adjusted according to actual conditions), the computer system determines that the drug contains a specific drug ingredient X; if , it is determined that the drug does not contain the ingredient.

[0055] For example, in a split text sentence about drug ingredients, "The main ingredients of this drug are A and B, and no specific drug ingredient X was found during the test", the computer system combines the implicit text representation of this sentence with the entity position information of "specific drug ingredient X" (the entity does not appear here, and the position information can be represented by a special tag) into a vector , and then substituted into the logistic regression model to obtain P(y=1| )=0.1, thus concluding that the drug does not contain specific drug ingredient X.

[0056] In addition to logistic regression models, computer systems can also use more complex neural network classifiers, such as multi-layer perceptrons (MLPs). As the input layer of the MLP.

[0057] MLP consists of an input layer, multiple hidden layers, and an output layer. In the hidden layer, neurons perform nonlinear transformations on the input through activation functions (such as the ReLU function f(x)=max(0,x)). For example, suppose the input layer has n neurons and the first hidden layer has m neurons. For the input vector Each element in , the output of neuron j in the first hidden layer is ,in is the weight connecting the input layer neuron i and the hidden layer neuron j, is bias.

[0058] Through the processing of multiple hidden layers, the classification result is finally obtained in the output layer. For example, if the output layer has two neurons, representing "contains specific components" and "does not contain specific components" respectively, the final classification result is determined by the activation value of the output layer neurons.

[0059] Taking the example of determining whether a drug uses a specific production technology (to be identified information), the computer system combines the information of the relevant split text sentences into a vector After input into the MLP and calculation by the network, if the neuron activation value indicating "includes specific production technology" is large, it is determined that the drug uses this production technology; otherwise, it is determined that it does not use this technology.

[0060] As an implementation mode, step S420, performing a text splitting operation on the K combined texts according to the text implicit representations in the K combined texts to obtain K split text sentences, includes: Step S421: dividing the text implicit representation of each of the K combined texts to obtain T text sub-implicit representations of the corresponding combined text; Step S422: performing a text splitting operation on the K combined texts according to the T text sub-implicit representations of the K combined texts to obtain K second importance information, wherein the K second importance information includes T second attention scores, wherein the b-th second attention score among the T second attention scores indicates that the text position where the b-th second attention score is located corresponds to the b-th text sub-implicit representation in the e-th combined text, where e is a natural number greater than 0 and less than K, and b is a natural number greater than 0 and less than T; Step S423: Generate a corresponding split text sentence according to each second importance information in the K second importance information to obtain K split text sentences.

[0061] In step S421 of the embodiment of the present application, the computer system divides the text implicit representation of each combined text in the K combined texts to obtain T text sub-implicit representations of the corresponding combined text.

[0062] The implicit text representation of the combined text is the result obtained by the computer system after the combined text is embedded in the previous step. It contains the semantic information of the combined text in the form of a low-dimensional vector. For example, a combined text about the entire process of drug production may have a multi-dimensional vector obtained by text embedding. The different dimensions in this vector represent the quantitative representation of different semantic features in the text.

[0063] When a computer system performs a partitioning operation, it can use a variety of technical means. One method is based on a clustering algorithm, such as the K-Means clustering algorithm. Assume that the text of the combined text is implicitly represented as a vector , with dimension n. The computer system will As a point in n-dimensional space. The K-Means algorithm first randomly initializes T cluster centers , these cluster centers are also n-dimensional vectors.

[0064] Then, for the vector For each element in (i.e., each dimension value in the implicit representation of the text), the computer system calculates its distance to each cluster center. The distance can be calculated using the Euclidean distance formula ,in are two n-dimensional vectors, are their values ​​in the i-th dimension respectively.

[0065] According to the calculated distance, each element is assigned to the class represented by the nearest cluster center. After multiple iterations (the position of the cluster center is continuously updated during the iteration process so that the sum of the distances from the elements in each cluster to the cluster center is minimized), the text implicit representation of the combined text is finally Divided into T different parts, each part corresponds to a cluster, which results in T implicit sub-representations of the text.

[0066] For example, for a text implicit representation of a combined text containing information on drug research and development, production, and sales, after being divided by the K-Means clustering algorithm, three text sub-implicit representations may be obtained, corresponding to the embodiment of the semantic information of the three links of research and development, production, and sales in the text implicit representation.

[0067] Next is step S422, in which the computer system determines the importance of each part in the combined text based on the previously obtained text sub-implicit representation, which is represented by a second attention score.

[0068] For example, the computer system can use an attention mechanism to calculate the second attention score. Assume that for the e-th combined text, its T text sub-sub-subs are implicitly represented as The computer system first defines a query vector ,This query vector can be set according to the specific task requirements. ,For example, if you want to determine the part that is most relevant to the ,quality traceability of drugs, the query vector can be a predefined vector related to ,quality traceability.

[0069] Then, the computer system calculates the relevance score of each text sub-implicit representation with the query vector. The relevance score can be calculated using a dot product operation, i.e. ,here It is the relevance score between the bth text sub-implicit representation in the eth combined text and the query vector.

[0070] In order to convert the relevance score into an attention score (i.e., the second attention score), the computer system can use the Softmax function for normalization. The formula of the Softmax function is ,here It is the bth second attention score in the eth combined text.

[0071] In this way, the computer system calculates the second attention score for each text sub-implicit representation in each combined text, thereby obtaining K second importance information. For example, in a combined text about pharmaceutical production, if one of the text sub-implicit representations is related to quality control, and the query vector is related to quality traceability, then this text sub-implicit representation can get a higher second attention score, indicating that it is more important in the task of quality traceability.

[0072] Finally, in step S423, the computer system generates split text sentences according to the second importance information (i.e., the second attention score) of each combined text. For example, for the e-th combined text, it has T text sub-implicit representations and corresponding second attention scores .

[0073] The computer system can sort the text sub-implicit representations according to the size of the second attention score. Assume that the sorted order is , the corresponding second attention score is .

[0074] Then, the computer system extracts the parts corresponding to these text sub-implicit representations from the original combined text to generate split text sentences. For example, if The corresponding text sub-implicit representation is related to the drug research and development stage, and the computer system extracts the sentence part related to the research and development stage from the original combined text, thus forming a split text sentence.

[0075] In this way, the computer system generates a split text sentence for each combined text, and finally obtains K split text sentences. These split text sentences can better reflect the importance of different parts of the combined text, and are helpful for the subsequent further processing of text information. For example, in the process of medicine traceability, information at different links can be analyzed more accurately, laying the foundation for accurate identification of the intended information.

[0076] In the actual application of medical science and technology data management, this text splitting operation based on implicit text representation is of great significance. For example, in the process of tracing the source of adverse drug reactions, this method can be used to split the information related to various links such as research and development, production, and sales in the combined text, and more clearly analyze the links that may cause adverse reactions. If the part related to the production link in a split text sentence shows that specific raw materials or processes are used in the production process, and these factors may be related to adverse reactions, then the problems in the production link can be further investigated.

[0077] In the decision-making process of pharmaceutical companies, this text splitting operation is also helpful for the analysis of pharmaceutical data. For example, if a company wants to optimize the sales strategy of a drug, by splitting the combined text containing R&D, production and sales information, the company can formulate a more reasonable sales strategy based on the information related to the sales link in the split text sentences and the impact of the R&D and production links on product characteristics.

[0078] In terms of knowledge mining of medical data, this text splitting operation can help researchers explore the potential knowledge in medical data more deeply. For example, researchers can find new research directions by analyzing the relationship between different links in the split text sentences, such as how new technologies in the R&D link affect the efficiency and quality of the production link, and then affect the market competitiveness of the sales link.

[0079] As an implementation manner, in step S400, before loading the K combined texts into the target learning network and obtaining the information recognition result of the information to be recognized in the text paragraph to be analyzed, the method further includes: Step S401: obtaining a plurality of sequentially arranged training text paragraphs, wherein the plurality of sequentially arranged training text paragraphs include a plurality of first training texts and a plurality of second training texts, each first training text is a text carrying a first text indicator mark, and each second training text is a text not carrying a text indicator mark, and the first text indicator mark indicates an information recognition result of the information to be recognized in the first training text; Step S402: performing a text transformation operation on each first training text to obtain a transformed training text of the corresponding first training text, and performing a basic text transformation and an advanced text transformation on each second training text to obtain a basic transformed training text and an advanced transformed training text of the corresponding second training text; Step S403: using the basic learning network to perform information recognition and reasoning processing on each transformed training text to obtain the corresponding reasoning mark of the first training text.

[0080] In step S401 of the embodiment of the present application, the computer system obtains a plurality of training text paragraphs arranged in sequence, and these training text paragraphs include a plurality of first training texts and a plurality of second training texts, each first training text is a text carrying a first text indicator mark, and each second training text is a text not carrying a text indicator mark, and the first text indicator mark represents the information recognition result of the intended information to be recognized in the first training text.

[0081] In the context of medical science and technology data management, training text paragraphs are the data foundation for training the target learning network. These paragraphs contain a wealth of medical-related information, such as descriptions of the drug development process, production process, quality inspection, sales channels, etc. Taking drug development as an example, a training text paragraph may elaborate on the entire process of a new drug from initial target discovery to drug molecule design to preclinical experiments.

[0082] The first training text is a text with a special mark, which is the first text indicator. Assuming that you want to conduct identification training on the traceability of raw materials in the drug production process, the first training text may be a detailed record of the production of a certain drug, which clearly points out that a specific raw material is used in the production process (this is the intended identification information), and the first text indicator clearly indicates that the raw material is indeed used in this production process, which is the information recognition result of the intended identification information. For example, the first training text describes "In the production process of ABC drug, after precise formula preparation, ingredient X is added as the main raw material, and then a series of processing steps are carried out..." Here, "ingredient X is added as the main raw material" is the intended identification information, and the first text indicator indicates that this information is correct, that is, ingredient X is indeed used.

[0083] The second training text does not have such clear indication marks. Continuing with the example of raw material traceability for drug production, the second training text may only describe the drug production process, such as "the production process of ABC drugs includes raw material preparation, mixing, processing, packaging and other links. In the raw material preparation link, various ingredients are mixed in a certain proportion...", but it does not directly indicate whether specific raw materials are used. The computer system needs to process through subsequent steps to determine the relevant circumstances of the information to be identified.

[0084] Next is step S402, which is a text transformation operation for the first training text. This is a purposeful conversion operation on the original text, the purpose of which is to increase the diversity of training data and improve the generalization ability of the target learning network. For example, when transforming the above-mentioned first training text about drug production and clarifying the use of raw materials, the technical means of vocabulary replacement can be used. Assuming that the original text contains "X ingredient is added as the main raw material", "add" can be replaced with "use", and "main raw material" can be replaced with "key ingredient" to obtain the transformed training text "In the production process of ABC drug, X ingredient is used as the key ingredient, followed by a series of processing steps...". This transformation increases the diversity of the text without changing the core semantics of the text (i.e., the use of raw materials).

[0085] For the transformation of the second training text, basic text transformation and advanced text transformation are transformation operations at different levels. Basic text transformation can be some simple grammatical structure adjustments, such as converting active voice to passive voice. Assuming that the second training text is "technicians conducted a drug synthesis experiment in the laboratory", after basic transformation, it is obtained as "drug synthesis experiments were conducted by technicians in the laboratory".

[0086] Advanced text transformation may involve adjustments at the semantic level, such as reorganizing the semantic relationship in a sentence. If the second training text is "The sales of drugs are affected by market demand and competitive products", it may become "Market demand and competitive products affect the sales of drugs" after advanced transformation. Through these transformations, the basic transformation training text and advanced transformation training text of the second training text are obtained, providing more diverse data for subsequent training steps.

[0087] Then comes step S403, the basic learning network is the initial state of the target learning network, which has a certain network structure and parameter initial values. Taking a simple neural network structure as an example, it may include an input layer, a hidden layer and an output layer. The input layer receives the transformed training text, for example, the transformed text about the use of raw materials for drug production is encoded and input into the network.

[0088] Assume that this neural network adopts a fully connected layer structure, and the neurons in the hidden layer perform nonlinear transformation on the input through the activation function (such as ReLU function f(x)=max(0,x)). For the encoding vector corresponding to each word in the input transformation training text, the weighted sum is performed in the hidden layer and processed by the activation function.

[0089] At the output layer, the inference result of the information to be recognized in the first training text, i.e., the inference mark, is obtained. For example, if the output layer adopts a binary classification method (such as whether a specific raw material is used or not), when the transformed training text using component X as a key component is input, the network may output an inference mark indicating that the component is used (such as a probability value close to 1, indicating that the component is highly likely to be used). This inference mark is the preliminary judgment result of the information to be recognized in the first training text obtained by the basic learning network after analyzing the transformed training text according to its current structure and parameters.

[0090] In the management of medical and scientific data, this process of acquiring inference marks for the first training text helps to adjust the parameters of the basic learning network, so that it can gradually optimize in the direction of accurately identifying medical-related information. For example, in the training related to drug quality testing, if the first training text describes the test results of a certain test method for a specific drug, the basic learning network can learn the relationship pattern between the test method and the test results by acquiring inference marks for the transformed training text, so that it can more accurately identify the information related to drug quality testing in subsequent training and practical applications.

[0091] Step S404: labeling and constructing each basic transformation training text according to the target guidance network to obtain a second text indication label corresponding to the second training text, where the second text indication label is a temporary label of the corresponding second training text; Step S405: performing information recognition and reasoning processing on each high-level transformation training text according to the basic learning network to obtain the corresponding reasoning mark of the second training text; Step S406: determining a target training error according to the inference mark of the first training text, the first text indication mark, the second text indication mark, and the inference mark of the second training text.

[0092] As an implementation manner, step S406, determining a target training error according to the reasoning mark of the first training text, the first text indication mark, the second text indication mark, and the reasoning mark of the second training text, includes: Step S4061: determining an error between the inference mark of the first training text and the first text indicator mark to obtain a first training error, and determining an error between the second text indicator mark and the inference mark of the second training text to obtain a second training error; Step S4062: Determine a target training error according to the first training error and the second training error.

[0093] Step S407: Correct the network weight variables of the basic learning network according to the target training error to obtain the target learning network.

[0094] In step S404 of the present embodiment, each basic transformation training text is labeled and constructed according to the target guidance network to obtain a second text indicator corresponding to the second training text, and the second text indicator is a temporary label of the corresponding second training text. In this step, the computer system processes the base transformation training text obtained after the second training text undergoes base transformation with the help of the target guidance network, thereby constructing the second text indicator mark. The target guidance network is a network that has been trained to a certain extent or has a specific structure and parameters, and is used to analyze the base transformation training text and generate temporary marks.

[0095] For example, in the scenario of medical science and technology data management, suppose the second training text is a description of drug sales, such as "the sales of this drug in various regions show different growth trends, with faster sales growth in economically developed regions and slower sales growth in some remote regions." After the text basic transformation, the basic transformation training text obtained may be "the sales of this drug in various regions have different growth trends, with faster sales growth in economically developed regions and slower sales growth in some remote regions."

[0096] The computer system inputs the base transformation training text into the target guidance network. The target guidance network may be a multi-layer neural network structure, such as an input layer, multiple hidden layers (each hidden layer has a number of neurons), and an output layer. The input layer receives a certain encoding representation of the base transformation training text (for example, the text can be converted into a vector form input using word vector encoding).

[0097] Inside the network, neurons in the hidden layer perform nonlinear transformations on the input through specific activation functions (such as the ReLU function, f(x)=max(0,x)). After being processed by multiple hidden layers, a second text indicator corresponding to the second training text is generated in the output layer. This mark is a temporary mark, also known as a pseudo-label, which is used to represent a certain inference result related to the information to be identified in the basic transformation training text, and serves as a supervisory signal for the training target learning network. For example, if the information to be identified is to determine whether drug sales are affected by the level of regional economic development, the output layer may output a binary classification mark (such as 0 for unaffected and 1 for affected). In this example, if the network outputs 1, then this 1 is the second text indicator mark of the second training text regarding this information to be identified, indicating that the computer system temporarily infers that drug sales are affected by the level of regional economic development based on the analysis of the target guidance network.

[0098] In step S405, the computer system uses the basic learning network to perform information recognition reasoning on the high-level transformed training text obtained by the high-level transformation of the second training text. The basic learning network has previously performed reasoning mark acquisition operations on the transformed training text of the first training text and has certain parameters and structures.

[0099] Taking data management in pharmaceutical R&D as an example, suppose the second training text is about the R&D process of a certain drug, such as "When studying the drug, the R&D team tried a variety of chemical structure modifications to improve the biological activity of the drug." After advanced transformation, an advanced transformation training text is obtained, such as "In order to improve the biological activity of the drug, the R&D team tried a variety of chemical structure modifications in the research of the drug."

[0100] This advanced transformation training text is input into the basic learning network. The basic learning network is also a multi-layer neural network structure. Assume that the input layer receives the advanced transformation training text in the form of word vectors and passes it to the hidden layer through the fully connected layer. The neurons in the hidden layer are interconnected, and the weights and biases are the parameters that need to be optimized during the network training process.

[0101] In the hidden layer, the calculation method of neurons is similar to the hidden layer calculation method in the target-guided network mentioned above. For example, for the kth neuron, its output is , where u_l is the element of the input vector, is the connection weight, is the bias and p is the dimension of the input vector.

[0102] After being processed by the hidden layer, the corresponding inference mark of the second training text is obtained in the output layer. This inference mark is an inference result about the information to be identified after the basic learning network analyzes the high-level transformation training text. For example, if the information to be identified is to determine whether the chemical structure modification in the research and development process effectively improves the biological activity of the drug, the output layer may output a probability value (such as 0.8, indicating that there is a high probability that it is effective), and this probability value is the inference mark.

[0103] In step S406, the reasoning mark of the first training text includes the inference results of different aspects related to the information to be identified. For example, in the first training text of the drug production link, if the information to be identified is whether the amount of a certain raw material used is within the specified range, the reasoning mark may include information such as the numerical estimate of the amount of raw material used and the confidence of this estimate. The first text indicator mark is a clear correct result about the information to be identified, such as the accurate value of the amount of raw material used and the accurate judgment of whether it is within the specified range.

[0104] Taking the usage of a raw material in drug production as an example, suppose the estimated usage of the raw material in the inference mark of the first training text is x_1 kg, and the accurate usage of the raw material in the first text indication mark is Kilograms. The error calculation here can be done using the mean square error (MSE) formula, which is (In this simple example n=1), this value is part of the first training error and represents the deviation of the numerical estimate of the raw material usage.

[0105] In addition, if the inference in the inference mark as to whether the amount of raw material used is within the prescribed range is wrong (for example, the inference mark indicates that it is within the range, while the first text indication mark indicates that it is not within the range), this also constitutes part of the first training error, indicating an error in determining whether the amount of raw material used is compliant.

[0106] The computer system comprehensively considers these different errors and obtains the first training error. For example, if the mean square error in numerical estimation is , errors in compliance judgment are recorded as (You can use a simple 0-1 error label, correct is 0, wrong is 1), then the first training error Can be ,in It is a weight coefficient used to balance the proportion of numerical estimation error and compliance judgment error in the overall first training error.

[0107] The second text indicator is a temporary mark obtained by the target guidance network from the basic transformation training text of the second training text, and the inference mark of the second training text is the inference result obtained by the basic learning network from the advanced transformation training text of the second training text. For example, in the example of drug sales, if the information to be identified is whether the sales channels are diversified, the second text indicator may be a temporary classification result (such as 0 for no, 1 for yes), and the inference mark of the second training text may be a probability value (such as 0.6, indicating that there is a certain possibility that the sales channels are diversified).

[0108] For the error calculation of the classification result, the cross-entropy formula can be used. Assume that the second text indication mark is , the inference markup of the second training text is (here and is a probability distribution, which can be regarded as a numerical value in the case of simple binary classification). The cross entropy formula is , this value is part of the second training error, indicating the difference between the classification results.

[0109] If there are other differences (such as differences in the accuracy of probability estimates), they also need to be taken into account in the second training error.

[0110] Similar to the determination of the first training error, the computer system comprehensively considers errors from various aspects to determine the second training error. Assume that the error calculated by cross entropy is The other errors are , the weight coefficient is , then the second training error .

[0111] The computer system calculates the first training error based on the obtained and the second training error To determine the target training error. A simple way is to directly add the two together, that is, the target training error E = + However, more complex weighted combination methods can also be used according to actual conditions, such as ,in is a weight coefficient used to adjust the relative importance of the first training error and the second training error in the target training error.

[0112] In step S407, the computer system corrects the network weight variables of the basic learning network based on the target training error. The network weight variables of the basic learning network determine how the network processes the input data and the output results. One correction method is based on the gradient descent algorithm. Assume that the target training error E is about the network weight variables (here represents all weights and bias variables), that is, E=E( ). The basic idea of ​​the gradient descent algorithm is to update the network weight variables along the negative gradient direction of the target training error function so that the target training error gradually decreases. The gradient can be calculated using the back propagation algorithm. For a simple neural network layer, such as , if the derivative of the target training error E with respect to the output y is known , then the target training error can be calculated with respect to the weight The derivative of ,in .

[0113] According to the calculated gradient, the update formula of the network weight variable is ,in is the current network weight variable, is the updated network weight variable, is the learning rate, a pre-set hyperparameter that controls the step size of the weight update.

[0114] Through multiple iterations of this updating process, the network weight variables of the basic learning network are continuously adjusted according to the target training error, and finally an optimized target learning network is obtained. This target learning network can more accurately identify the information to be identified in the medical science and technology data, laying a good foundation for loading K combined texts into the target learning network in the subsequent step S400 to obtain the information recognition result of the information to be identified.

[0115] As an implementation manner, the inference mark of the first training text includes a first entity position information inference mark and a first information content inference mark, and the first text indication mark includes a first entity position information indication mark and a first information content indication mark. Based on this, in step S4061, determining the error between the inference mark of the first training text and the first text indication mark to obtain the first training error includes: Step S40611: determining an error between the first entity position information inference mark and the first entity position information indication mark to obtain a first sub-training error, where the first sub-training error represents a prediction deviation of the entity position information of the to-be-recognized information in the first training text; Step S40612: determining an error between the first information content inference mark and the first information content indication mark to obtain a second sub-training error, where the second sub-training error represents an information content prediction deviation of the to-be-recognized information in the first training text; Step S40613: Determine a first training error according to the first sub-training error and the second sub-training error.

[0116] In step S40611 of the embodiment of the present application, the error between the first entity location information inference mark and the first entity location information indication mark is determined to obtain a first sub-training error, and the first sub-training error represents the prediction deviation of the entity location information of the to-be-recognized information in the first training text. In the management of medical and scientific data, entity location information is crucial for accurately understanding and processing medical-related texts. For example, in a first training text about the drug development process, the information to be identified may be the location of a specific drug ingredient in the development process (for example, at which experimental stage it was first mentioned or used).

[0117] The first entity location information inference mark is the inference result about the entity location obtained by the basic learning network after performing information recognition and inference processing on the first training text. Assume that in this example, the first entity location information inference mark indicates that a specific drug ingredient is inferred to be used for the first time in the third experimental stage. The first entity location information indication mark is the actual correct information in the first training text, that is, the drug ingredient was actually used for the first time in the second experimental stage.

[0118] In order to determine the error between the two, the computer system can use a variety of technical means. One method is to calculate the absolute difference. Suppose the first entity position information inference mark is , the first entity location information indication mark is , then the first sub-training error In the above example, , this error value of 1 indicates a deviation in the prediction of entity location information.

[0119] This error calculation of entity location information is of great significance in pharmaceutical data management. For example, in the process of drug traceability, if the introduction location of raw materials in the production process is predicted incorrectly, it can lead to a deviation in the understanding of the entire production process, thus affecting aspects of drug quality control and supervision.

[0120] Another way is to use weighted difference to calculate the error. For example, different weights are assigned according to the importance of different experimental stages. Suppose the weight of the second experimental stage is , the weight of the third experimental stage is (The weight can be determined based on factors such as the impact of the experimental stage on the drug development results.) Then the first sub-training error so This weighted difference calculation method takes into account the impact of differences between different positions on the overall error in more detail.

[0121] In step S40612, in the scenario of medical science and innovation data management, accurate identification of information content is equally important. Continuing with the example of drug development, the information to be identified in the first training text may be a description of the mechanism of action of a certain drug. The first information content inference mark is the inference content about the drug mechanism obtained by the basic learning network through inference on the first training text, and the first information content indicator mark is the actual correct description of the drug mechanism in the text.

[0122] For example, the first information content inference marker infers that a drug exerts its efficacy by inhibiting the activity of a specific enzyme (assuming mechanism A), while the first information content indication marker indicates that the drug actually exerts its efficacy by regulating the intracellular signaling pathway (assuming mechanism B).

[0123] In order to determine the error between the two, the computer system can use a text similarity calculation method. For example, cosine similarity calculation based on word vectors. In addition to cosine similarity, edit distance can also be used to measure the error of information content. Edit distance refers to the minimum number of editing operations required to convert one string to another string (editing operations include inserting, deleting, and replacing characters). Treat the description of the drug's mechanism of action as a string and calculate the edit distance d from the first information content inference marker to the first information content indication marker. For example, if the edit distance d=5, it can be converted into an error value based on a certain mapping relationship (For example, according to the statistical results of the training data, set ,in is the maximum possible edit distance among all training samples).

[0124] The calculation of this information content error is critical for the accurate inheritance and application of medical knowledge. In drug development, if the understanding of the drug's mechanism of action is wrong, it can lead to problems such as the wrong direction of subsequent research and development and unreasonable clinical trial design.

[0125] In step S40613, the computer system obtains the first sub-training error (representing the prediction deviation of entity location information) and the second sub-training error (represents the information content prediction deviation) After that, these two errors need to be combined to determine the first training error.

[0126] One way to determine the first training error is to directly add the two together, that is, the first training error For example, if =0.2 (prediction deviation of entity position information calculated previously), = 0.7 (the information content prediction bias calculated previously), then .

[0127] However, in practical applications, according to the needs of different medical data management tasks, a weighted summation method can be adopted. For example, different weights can be assigned according to the relative importance of entity location information and information content in the overall medical data understanding. Assume that the weight of entity location information is = 0.3, the weight of information content is =0.7 (these weights can be determined based on the experience of medical experts or analysis of large amounts of training data).

[0128] The first training error This weighted summation method can more flexibly determine the final first training error according to the importance of different error factors in medical data management.

[0129] Determining the first training error is of great significance in the entire blockchain-based pharmaceutical science and innovation data management method. It is one of the key indicators for optimizing the basic learning network. By accurately calculating the first training error, the network can adjust the weights more specifically in the subsequent training process to improve the recognition accuracy of the information to be recognized in the first training text, thereby improving the performance of the entire pharmaceutical data management system in terms of drug traceability, R&D process monitoring, quality control, etc.

[0130] As an implementation manner, the inference mark of the second training text includes a second entity location information inference mark and a second information content inference mark, and the second text indication mark includes an entity location information temporary mark and an information content temporary mark. Based on this, in step S4061, determining the error between the second text indication mark and the inference mark of the second training text to obtain the second training error includes: Step S40614: determining an error between the second entity location information inference mark and the entity location information temporary mark to obtain a third sub-training error, where the third sub-training error represents a prediction deviation of the entity location information of the to-be-recognized information in the second training text; Step S40615: determining an error between the second information content inference mark and the information content temporary mark to obtain a fourth sub-training error, where the fourth sub-training error represents an information content prediction deviation of the to-be-recognized information in the second training text; Step S40616: Determine a second training error according to the third sub-training error and the fourth sub-training error.

[0131] In step S40614 of the embodiment of the present application, the error between the second entity location information inference mark and the entity location information temporary mark is determined to obtain a third sub-training error, which represents the prediction deviation of the entity location information of the to-be-recognized information in the second training text. The second entity location information inference mark is the inference result about the entity location obtained after the basic learning network performs information recognition and inference processing on the second training text. Assume that in this example of drug production, the basic learning network infers that a specific quality inspection link is carried out in the fifth step of the production process (this is the second entity location information inference mark). The temporary mark of the entity location information is the temporary judgment result about the entity location obtained by the target guidance network by marking the basic transformation training text of the second training text. Assume that this temporary mark indicates that the specific quality inspection link should be carried out in the fourth step of the production process. In order to determine the error between the two, the computer system can use different technical means. One method is to calculate the absolute difference (see above).

[0132] The calculation of the prediction deviation of this entity position information is of great significance in pharmaceutical production management. For example, if the prediction of the location of the quality inspection link is wrong, it can affect the quality control and optimization strategy of the entire production process. Because the various links in the production process are interrelated, the deviation of the location of the quality inspection link may lead to an incorrect assessment of production efficiency, product quality, etc. In addition to the absolute difference, the computer system can also use the relative difference calculation method. The calculation of the relative difference can take into account the impact of the number of steps in the entire production process.

[0133] In step S40615, the second information content inference mark is the inference content of the main factors affecting drug sales obtained by the basic learning network through inference on the second training text, and the information content temporary mark is the temporary judgment result of the information content obtained by the target guidance network through marking and constructing the basic transformation training text of the second training text. For example, the second information content inference mark infers that the main factor affecting drug sales is price competition (assuming factor A), while the information content temporary mark indicates that the main factor affecting the drug sales is that market demand has not been fully explored (assuming factor B).

[0134] In order to determine the error between the two, the computer system can use a semantic distance calculation method. One way is to build a semantic vector space based on the semantic knowledge base. First, map the descriptions of factor A and factor B into this semantic vector space to obtain the corresponding semantic vectors. and .

[0135] Then, the semantic distance between the two can be calculated using the Euclidean distance formula , where n is the dimension of the semantic vector, They are and The value in the i-th dimension. This semantic distance can be used as the fourth sub-training error A measure of .

[0136] Another method is to use a topic model to measure the error of information content. For example, the Latent Dirichlet Allocation (LDA) model is used. The text containing the description of factor A and factor B is used as input respectively, and their topic distribution is obtained through the LDA model. Suppose the topic distribution of factor A is , the topic distribution of factor B is , where k is the number of topics.

[0137] Then, we can use some distance metric, such as KL divergence (Kullback-Leibler Divergence), to calculate the difference between the two. The formula for KL divergence is , this value can be used as a measure of the fourth sub-training error e4. This topic model-based method can measure the difference in information content from a more macroscopic topic level.

[0138] In step S40616, after obtaining the third sub-training error e3 (indicating the entity location information prediction deviation) and the fourth sub-training error e4 (indicating the information content prediction deviation), the computer system needs to combine these two errors to determine the second training error.

[0139] A simple way to determine the second training error is to directly add the two together, that is, the second training error E2 = e3 + e4. For example, if e3 = 0.1 (the entity position information prediction deviation calculated previously), e4 = 0.3 (the information content prediction deviation calculated previously), then E2 = 0.1 + 0.3 = 0.4.

[0140] As an implementation mode, in step S407, after the network weight variables of the basic learning network are modified according to the target training error and the target learning network is obtained, the method further includes: Step S408: modifying the weight of the initial guidance network according to the dynamic mean of the target learning network to obtain the target guidance network.

[0141] In the embodiment of the present application, the dynamic mean (EMA) is a method of calculating the weighted average during the optimization process, which gives higher weight to recent data. In this scenario, after multiple rounds of optimization, the target learning network's network weight variables have been able to better adapt to various patterns in medical science and technology data, and the dynamic mean of the target learning network integrates the weight information in these optimization processes.

[0142] For example, suppose the weight variable of the target learning network has a series of values ​​in different training rounds. If the simple average method is used to calculate the mean of the weights, the weights of all rounds will be treated equally. But when using the moving average (EMA), for example, for the weight of round t , dynamic mean The calculation can be done using the formula ,in is a coefficient between 0 and 1 that determines the relative importance of the new weight and the old moving average in calculating the new moving average. =0.1, which means that the new weight w_t accounts for 10% of the dynamic mean, while the old dynamic mean As time goes by (training rounds), this dynamic mean can reflect a comprehensive state of the target learning network weights that is biased towards recent optimization results.

[0143] The initial guidance network is a network used to perform operations such as labeling and constructing the base transformation training text during the process of building the target learning network. The weights of the initial guidance network may be randomly initialized or determined based on some predefined rules in the initial state. However, as the target learning network is optimized, the knowledge learned by the target learning network (reflected by its weight variables) can be used to improve the initial guidance network.

[0144] The computer system uses the dynamic mean of the target learning network to correct the weights of the initial guidance network. This correction process can be done in a variety of ways. One way is to directly adjust the weights of the initial guidance network in proportion to the dynamic mean of the target learning network. For example, if a weight of the initial guidance network The corresponding target learning network dynamic mean part is , can be obtained through the formula To calculate the new weight ,in is a basic dynamic mean part used for normalization.

[0145] As an implementation mode, step S404, labeling and constructing each basic transformation training text according to the target guidance network to obtain a second text indication label corresponding to the second training text, includes: Step S4041: performing text embedding coding on each basic transformation training text to obtain a training text implicit representation of the corresponding basic transformation training text; Step S4042: Decode and infer the training text implicit representation of each basic transformation training text to obtain a second text indicator corresponding to the second training text.

[0146] In step S4041, the computer system performs text embedding coding to convert the basic transformation training text into a low-dimensional vector representation that is convenient for subsequent processing, that is, an implicit representation of the training text. For example, in a basic transformation training text about drug production, information such as the type of production equipment, the requirements of the production environment, and the various steps in the production process may be described.

[0147] Computer systems can use pre-trained word vector models (such as Word2Vec or GloVe) or deep learning-based text embedding models (such as BERT) to perform text embedding encoding. Taking Word2Vec as an example, it is based on the idea of ​​distributed representation and learns a fixed-dimensional vector representation for each word by training on a large-scale corpus.

[0148] Assume that the basis transformation training text is ,in Represents the i-th word in the text. The Word2Vec model learns the co-occurrence relationship of words in the corpus and generates a Generate a vector representation For the entire basic transformation training text T, the computer system can obtain the implicit representation of the training text by combining the word vectors in the text (such as simple average or weighted average) .

[0149] If a simple average method is used, the implicit representation of the training text The calculation formula is: This text implicitly indicates It can capture the semantic information in the basic transformation training text, represent the lexical relationship in the text in the form of vectors, and provide a suitable data structure for subsequent target-guided network operations.

[0150] In step S4042, after obtaining the implicit representation of the training text of the basic transformation training text, the computer system needs to decode and infer it to obtain a second text indication mark. This mark is a temporary mark corresponding to the second training text, which is used for subsequent operations such as determining the target training error in the process of building the target learning network.

[0151] For example, after a basic transformation training text about drug sales obtains an implicit representation of the training text through step S4041, the computer system needs to determine the information in the text that is related to the information to be identified (such as whether the sales channels are diversified) and generate a temporary tag to represent this relationship.

[0152] The computer system can use a neural network-based decoder to perform decoding inference. Assume that the decoder is a multi-layer perceptron (MLP) structure, which consists of an input layer, multiple hidden layers, and an output layer.

[0153] The input layer receives the implicit representation of the training text , and then pass it to the hidden layer. In the hidden layer, neurons are connected by weights, and an activation function (such as the ReLU function f(x)=max(0,x)) is used to perform a nonlinear transformation on the input. For the jth neuron in the hidden layer, its input is After the weight matrix W and the bias vector The result after transformation is , and then the output is obtained through the activation function .

[0154] After being processed by multiple hidden layers, the second text indicator corresponding to the second training text is obtained in the output layer. The structure of the output layer depends on the type of mark to be generated. If it is a binary classification problem (such as whether the sales channel is diversified or not), the output layer can be a neuron, and the Sigmoid function is used to convert the output into a probability value between 0 and 1. If the probability value is greater than a certain threshold (such as 0.5), it means that it is a diversified sales channel, and the corresponding second text indicator is 1; otherwise, it is 0, indicating that it is not a diversified sales channel.

[0155] This process of decoding and reasoning the implicit representation of the training text of the basic transformation training text to obtain the second text indicator mark is an important step in the process of building the target learning network. It helps the computer system to generate temporary marks based on the target guidance network in the absence of clear indicator marks of the second training text, providing necessary support for the subsequent accurate determination of the target training error, optimization of the basic learning network, and ultimately achieving accurate information recognition in the management of medical and scientific data.

[0156] As an implementation mode, step S4042, decoding and reasoning the training text implicit representation of each basic transformation training text to obtain a second text indicator corresponding to the second training text, includes: Step S40421: performing target information retrieval on the implicit representation of the training text of each basic transformation training text to obtain entity position information of the to-be-recognized information corresponding to the basic transformation training text; Step S40422: Process the training text implicit representation of each basic transformation training text and the entity position information of the to-be-recognized information of the basic transformation training text to obtain a second text indicator corresponding to the second training text.

[0157] In step S40421 of the present embodiment, target information retrieval refers to the operation of the computer system to find specific information related to the information to be identified from the implicit representation of the training text of the basic transformation training text. The entity position information of the information to be identified is used to determine the specific position range of the information to be identified in the basic transformation training text, which is crucial for accurately understanding and marking the basic transformation training text.

[0158] The computer system can use technical means based on the vector space model. The implicit representation of the training text of the basic transformation training text is a low-dimensional vector, which represents the semantic information of the text in the form of a vector. Assume that this vector is , with dimension n.

[0159] For the intended identification information, for example, in the drug production process, the intended identification information is the use of specific raw materials. The keywords related to this intended identification information (such as the name of the raw material, etc.) can also be converted into vector form and set as The computer system then calculates and similarity, such as cosine similarity.

[0160] Next, the computer system needs to determine the entity position information of the information to be identified. Assuming that the basic transformation training text is "In the process of drug production, raw materials A and B are first mixed, and then special ingredient X is added, and after a series of processing steps...", if the part related to special ingredient X (the information to be identified) is determined by vector calculation, the computer system can determine the starting and ending positions of special ingredient X based on the character positions in the text, which is the entity position information.

[0161] Computer systems can also build retrieval models based on neural networks, such as a convolutional neural network (CNN) or a recurrent neural network (RNN).

[0162] Taking convolutional neural network as an example, the training text is implicitly represented As the data of the input layer, the network extracts features and reduces the dimension of the input data through structures such as convolutional layers and pooling layers.

[0163] For the information to be identified, it is also converted into a target vector (which can be obtained by using the same encoding method as the implicit representation of the training text). During the network training process, a large amount of labeled data (such as basic transformation training text samples that are known to contain the information to be identified and its location) is used to learn how to accurately find the entities and their locations related to the information to be identified.

[0164] In step S40422, after determining the entity location information of the information to be recognized, the computer system needs to further process it in combination with the implicit representation of the training text of the basic transformation training text to obtain a second text indicator corresponding to the second training text. This second text indicator is a temporary mark used to indicate a certain inference result related to the information to be recognized in the basic transformation training text during the process of building the target learning network.

[0165] The computer system can process according to predefined rules. For example, in the pharmaceutical sales process, if the information to be identified is whether the sales channels are diversified, the training text implicit representation of the basic transformation training text determines the part related to the sales channel (entity location information) through the previous target information retrieval.

[0166] Assume that according to the rules, if at least three different types of sales channels (such as hospital pharmacies, retail pharmacies, online sales platforms, etc.) are mentioned in the text portion determined by the entity location information, the second text indication is marked as 1, indicating that the sales channels are diversified; if there are less than three sales channels mentioned, the second text indication is marked as 0, indicating that the sales channels are not diversified enough. For example, the basic transformation training text is "The sales channels of this drug mainly include hospital pharmacies and retail pharmacies, and the possibility of online sales platforms is currently being explored." According to the rules, two existing sales channels and one sales channel being explored are mentioned here, so the second text indication is marked as 1.

[0167] The computer system can be processed using a neural network (such as a multi-layer perceptron, MLP). The implicit representation of the training text of the basic transformation training text and the vector representing the entity position information (which can be obtained by encoding the entity position information) are used as inputs of the input layer. After being processed by multiple hidden layers, the second text indication mark is obtained in the output layer. For example, if the output layer is a neuron, the Sigmoid function is used to convert the output into a probability value between 0 and 1. If this probability value is greater than a certain threshold (such as 0.5), the second text indication mark is 1; otherwise, it is 0.

[0168] As an implementation mode, step S405, performing information recognition and reasoning processing on each high-level transformation training text according to the basic learning network to obtain the reasoning mark of the corresponding second training text, includes: Step S4051: performing text combination on a plurality of advanced transformation training texts to obtain a plurality of combined training texts; Step S4052: performing text embedding coding on each combined training text to obtain a text implicit representation of the corresponding combined training text; Step S4053: performing a text splitting operation on the plurality of combined training texts according to the text implicit representations of the plurality of combined training texts to obtain a plurality of split training text sentences; Step S4054: performing text embedding coding on each of the plurality of split training text sentences to obtain a text implicit representation of the corresponding split training text sentence; Step S4055: Process the text implicit representation of each split training text sentence to obtain the inference mark of each second training text.

[0169] In step S4051 of the embodiment of the present application, the computer system performs a text combination operation in order to integrate the information in these scattered high-level transformation training texts so as to perform more effective information recognition and reasoning processing later. For example, in the field of medicine, one high-level transformation training text may focus on the experimental design part of the drug development process, describing information such as sample selection, grouping, and control settings of the experiment; another high-level transformation training text may focus on the data analysis part of the development process, mentioning the statistical methods used and the data processing process.

[0170] Computer systems can use a variety of technical means to combine texts. A simple method is splicing based on semantic coherence. First, the computer system analyzes the semantic themes of each high-level transformation training text. For example, for the two high-level transformation training texts about drug development mentioned above, the computer system identifies the keywords (such as "experimental design", "sample selection", "data analysis", "statistical methods", etc.) in them to determine that they are all related to the drug development process and have semantic coherence.

[0171] Then, these advanced transformation training texts are spliced ​​together in a certain order to form a combined training text. This order can be determined according to the logic of the pharmaceutical business, for example, according to the order of the R&D process, the content related to experimental design is placed first, followed by the content related to data analysis.

[0172] Another technical approach is to use the attention mechanism for text combination. Assume that there are n high-level transformation training texts, whose text representations are The computer system defines a query vector q for each high-level transformation training text, and the query vector can be set according to the information to be recognized or the overall medical task requirements.

[0173] For each high-level transformation training text , the computer system calculates its relevance score s_i with the query vector q. The relevance score can be calculated using a dot product operation, i.e. Then, based on these relevance scores, the computer system assigns a weight to each advanced transformation training text. , the weight calculation can be normalized using the Softmax function, that is, .

[0174] Finally, combine the training text It can be obtained by weighted summation, that is, In this way, the combined training text can better highlight the parts related to the query vector (that is, the task), improving the information integration effect.

[0175] In step S4052, although the combined training text has integrated the high-level transformation training text, it is still in text form. The computer system needs to perform text embedding encoding on it and convert it into a low-dimensional vector representation that is more suitable for subsequent processing, namely, text implicit representation.

[0176] Computer systems can use pre-trained word vector models (such as Word2Vec, GloVe, etc.) or deep learning-based text embedding models (such as BERT, etc.) to perform text embedding encoding. Taking the BERT model as an example, BERT is a pre-trained language model based on the Transformer architecture, which can learn the semantic information in the text and encode it into a vector representation.

[0177] Assume that the combined training text is ,in Represents the i-th word in the combined training text. The BERT model takes this combined training text as input and, after processing it through a multi-layer Transformer structure, obtains a fixed-dimensional vector representation. , this vector It is the implicit text representation corresponding to the combined training text.

[0178] This implicit text representation can capture the semantic information in the combined training text, including the semantic relationship between words, the semantic structure of sentences, etc. For example, for a combined training text "In the process of drug research and development, rigorous experimental design and scientific data analysis jointly ensure the effectiveness of drugs, which has a positive impact on subsequent market positioning and sales channel expansion", after BERT's text embedding encoding, the obtained implicit text representation can reflect the semantic association between the research and development process, drug effectiveness, market positioning and sales channels.

[0179] In step S4053, after obtaining the text implicit representation of the combined training text, the computer system performs a text splitting operation, the purpose of which is to further decompose the combined training text into smaller semantic units, that is, split the training text sentences, so as to analyze the information in the text more carefully.

[0180] The computer system can adopt a splitting method based on semantic boundaries. In the text implicit representation of the combined training text, different semantic parts may have different characteristics in the vector space. For example, in the text implicit representation of a combined training text about drug production and sales, there may be some difference in the vector representation between the part related to the production link and the part related to the sales link.

[0181] The computer system determines the semantic boundaries by analyzing the vector features of the implicit representation of the text. For example, a clustering algorithm can be used to identify these semantic boundaries. Assume that the implicit representation of the text is , with dimension d. The computer system can The points are considered as points in d-dimensional space, and then the K-Means clustering algorithm is used to cluster these points into different groups.

[0182] The K-Means algorithm first randomly initializes k cluster centers , and then calculate the distance (such as Euclidean distance) from each point to the cluster center, and assign the point to the group to which the nearest cluster center belongs according to the distance. After multiple iterations, until the cluster center no longer changes significantly.

[0183] Through this clustering method, the computer system can determine the semantic boundaries in the implicit representation of the text, and then split the combined training text into several split training text sentences based on these boundaries.

[0184] In step S4054, the purpose of text embedding encoding for the split training text sentence is similar to that of text embedding encoding for the combined training text, which is to convert the text into a low-dimensional vector representation for subsequent information processing. A pre-trained word vector model or a text embedding model based on deep learning can also be used. For example, for the split training text sentence "In the process of drug research and development, the selection of experimental samples follows strict standards", the Word2Vec model is used for text embedding encoding. Assume that the word set in this split training text sentence is , Word2Vec for each word Generate a vector representation .

[0185] The implicit text representation of the split training text sentence can be obtained by simple averaging This implicit text representation can capture the semantic information in the split training text sentences. For example, in this example, it can reflect the relevant semantic features of the experimental sample selection in the drug development process.

[0186] In step S4055, after obtaining the text implicit representation of the split training text sentences, the computer system needs to process it to obtain the inference mark of each second training text. This inference mark is the result of the basic learning network performing information recognition inference processing on the high-level transformation training text, and is used to represent a certain inference related to the information to be recognized in the second training text.

[0187] The computer system can use a neural network-based classifier for processing. For example, a multi-layer perceptron (MLP) is constructed as a classifier. The text implicit representation of the split training text sentence is input as the input layer of the MLP. The MLP consists of an input layer, multiple hidden layers, and an output layer. In the hidden layer, the neurons perform nonlinear transformations on the input through activation functions (such as the ReLU function f(x)=max(0,x)). Assume that the input layer has n neurons and the first hidden layer has m neurons. For the input vector Each element in , the output of neuron j in the first hidden layer is ,in is the weight connecting the input layer neuron i and the hidden layer neuron j, is bias.

[0188] After being processed by multiple hidden layers, the inference mark is obtained in the output layer. For example, if the information to be identified is to determine whether the drug development process meets the specifications, the output layer can be a binary classification structure, outputting 0 or 1. If the output is 1, it means that the computer system infers that the development process meets the specifications based on the analysis of the split training text sentences; if the output is 0, it means that the inference of the development process does not meet the specifications.

[0189] Figure 2 A hardware entity diagram of a computer system provided in an embodiment of the present application is shown in FIG. Figure 2 As shown, the hardware entity of the computer system 1000 includes: a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program that can be run on the processor 1001, and the processor 1001 implements the steps in the method of any of the above embodiments when executing the program.

Claims

1. A blockchain-based medical science and technology data management method, characterized in that: include: In response to a medicine traceability request, a medicine industry chain record text is obtained in a block node matched by the medicine traceability request, wherein the medicine industry chain record text includes K text paragraphs to be analyzed, wherein the K text paragraphs to be analyzed are text paragraphs arranged in sequence, and K is a natural number greater than 1; Performing high-level text transformation on each of the K to-be-analyzed text paragraphs to obtain a high-level transformed text corresponding to the to-be-analyzed text paragraph; Performing text combination on the K advanced transformation texts to obtain K combined texts; The K combined texts are loaded into a target learning network to obtain information recognition results of the information to be recognized in the text paragraph to be analyzed. The target learning network is a neural network obtained by performing multiple rounds of optimization on the transformation training text and the first text indication mark corresponding to the first training text, and the basic transformation training text and the advanced transformation training text and the second text indication mark corresponding to the second training text as network learning samples. The first training text is a text carrying the first text indication mark, and the second text indication mark is a temporary mark obtained after information recognition reasoning is performed on the basic transformation training text corresponding to the second training text by the target guidance network.

2. The method according to claim 1, characterized in that The step of combining the K advanced transformation texts to obtain K combined texts includes: Decomposing each of the high-level transformation texts into T clauses, where T is a natural number greater than 1; Perform clause integration on the T clauses in the K advanced transformation texts to obtain K first importance information, each of which includes T first attention scores, and the a-th first attention score among the T first attention scores indicates that the text position where the a-th first attention score is located corresponds to the a-th clause in the d-th advanced transformation text, d is equal to the a-th first attention score, d is a natural number greater than 0 and less than K, and a is a natural number greater than 0 and less than T; A corresponding combined text is generated according to each of the K first importance information to obtain K combined texts.

3. The method according to claim 1 or 2, characterized in that: The K combined texts are loaded into the target learning network to obtain information recognition results of the information to be recognized in the text paragraph to be analyzed, including: Performing text embedding coding on each of the K combined texts to obtain a text implicit representation of the corresponding combined text; Performing a text splitting operation on the K combined texts according to the text implicit representations in the K combined texts to obtain K split text sentences; Performing text embedding coding on each of the K split text sentences to obtain a text implicit representation of the corresponding split text sentence; Decoding and reasoning are performed on the text implicit representations of the K split text sentences to obtain information recognition results of the information to be recognized in the text paragraph to be analyzed.

4. The method according to claim 3, characterized in that The decoding and reasoning of the implicit text representations of the K split text sentences to obtain the information recognition result of the to-be-recognized information in the to-be-analyzed text paragraphs includes: Performing target information retrieval on the implicit text representations of the K split text sentences to obtain entity location information of the information to be identified; Classification, decoding and reasoning are performed on the implicit text representations of the K split text sentences and the entity position information of the information to be identified to obtain an information identification result of the information to be identified.

5. The method according to claim 3, characterized in that: Performing a text splitting operation on the K combined texts according to the text implicit representations in the K combined texts to obtain K split text sentences, including: Dividing the text implicit representation of each of the K combined texts to obtain T text sub-implicit representations corresponding to the combined text; Performing a text splitting operation on the K combined texts according to the T text sub-implicit representations of the K combined texts to obtain K second importance information, wherein the K second importance information includes T second attention scores, and the b-th second attention score among the T second attention scores indicates that the text position where the b-th second attention score is located corresponds to the b-th text sub-implicit representation in the e-th combined text, where e is a natural number greater than 0 and less than K, and b is a natural number greater than 0 and less than T; A corresponding split text sentence is generated according to each of the K second importance information to obtain K split text sentences.

6. The method according to claim 1 or 2, characterized in that: Before loading the K combined texts into the target learning network to obtain the information recognition result of the to-be-recognized information in the to-be-analyzed text paragraph, the method further includes: Acquire a plurality of training text paragraphs arranged in sequence, wherein the plurality of training text paragraphs arranged in sequence include a plurality of first training texts and a plurality of second training texts, each of the first training texts is a text carrying a first text indicator mark, each of the second training texts is a text not carrying a text indicator mark, and the first text indicator mark indicates an information recognition result of the information to be recognized in the first training text; Performing a text transformation operation on each of the first training texts to obtain a transformed training text of the corresponding first training text, and performing a basic text transformation and an advanced text transformation on each of the second training texts to obtain a basic transformed training text and an advanced transformed training text of the corresponding second training text; Using a basic learning network to perform information recognition and reasoning processing on each of the transformed training texts to obtain a corresponding reasoning tag of the first training text; Marking and constructing each of the basic transformation training texts according to the target guidance network to obtain a second text indication mark corresponding to the second training text, where the second text indication mark is a temporary mark of the corresponding second training text; Performing information recognition and reasoning processing on each of the high-level transformation training texts according to the basic learning network to obtain a corresponding reasoning tag of the second training text; determining a target training error according to the inference mark of the first training text, the first text indication mark, the second text indication mark, and the inference mark of the second training text; The network weight variables of the basic learning network are modified according to the target training error to obtain the target learning network.

7. The method according to claim 6, characterized in that Determining a target training error according to the inference mark of the first training text, the first text indication mark, the second text indication mark, and the inference mark of the second training text includes: Determining an error between the inference mark of the first training text and the first text indicator mark to obtain a first training error, and determining an error between the second text indicator mark and the inference mark of the second training text to obtain a second training error; The target training error is determined according to the first training error and the second training error.

8. The method according to claim 7, characterized in that The inference mark of the first training text includes a first entity position information inference mark and a first information content inference mark, the first text indication mark includes a first entity position information indication mark and a first information content indication mark, and determining the error between the inference mark of the first training text and the first text indication mark to obtain a first training error includes: Determine an error between the first entity position information inference mark and the first entity position information indication mark to obtain a first sub-training error, where the first sub-training error represents a prediction deviation of the entity position information of the to-be-recognized information in the first training text; Determine an error between the first information content inference mark and the first information content indication mark to obtain a second sub-training error, wherein the second sub-training error represents an information content prediction deviation of the to-be-recognized information in the first training text; determining the first training error according to the first sub-training error and the second sub-training error; The inference mark of the second training text includes a second entity location information inference mark and a second information content inference mark, the second text indication mark includes an entity location information temporary mark and an information content temporary mark, and determining the error between the second text indication mark and the inference mark of the second training text to obtain a second training error includes: Determine an error between the second entity position information inference mark and the entity position information temporary mark to obtain a third sub-training error, wherein the third sub-training error represents a prediction deviation of the entity position information of the to-be-recognized information in the second training text; Determine an error between the second information content inference mark and the information content temporary mark to obtain a fourth sub-training error, wherein the fourth sub-training error represents an information content prediction deviation of the to-be-recognized information in the second training text; The second training error is determined according to the third sub-training error and the fourth sub-training error.

9. The method according to claim 6, characterized in that After the network weight variables of the basic learning network are modified according to the target training error to obtain the target learning network, the method further includes: Modifying the weight of the initial guidance network according to the dynamic mean of the target learning network to obtain the target guidance network; The step of constructing a label for each of the basic transformed training texts according to the target guiding network to obtain a second text indication label corresponding to the second training text includes: Performing text embedding coding on each of the basic transformation training texts to obtain a training text implicit representation of the corresponding basic transformation training text; Decoding and reasoning the implicit representation of the training text of each of the basic transformation training texts to obtain a second text indicator corresponding to the second training text; The step of decoding and reasoning the implicit representation of the training text of each of the basic transformed training texts to obtain a second text indicator corresponding to the second training text includes: Performing target information retrieval on the implicit representation of the training text of each of the basic transformation training texts to obtain entity position information of the to-be-recognized information corresponding to the basic transformation training text; Processing the training text implicit representation of each of the basic transformed training texts and the entity position information of the to-be-recognized information of the basic transformed training texts to obtain a second text indicator corresponding to the second training text; The step of performing information recognition and reasoning processing on each of the high-level transformation training texts according to the basic learning network to obtain the corresponding reasoning mark of the second training text includes: Combining the plurality of advanced transformation training texts to obtain a plurality of combined training texts; Performing text embedding coding on each of the combined training texts to obtain a text implicit representation corresponding to the combined training text; Performing a text splitting operation on the plurality of combined training texts according to the text implicit representations of the plurality of combined training texts to obtain a plurality of split training text sentences; Performing text embedding coding on each of the plurality of split training text sentences to obtain a text implicit representation corresponding to the split training text sentence; The text implicit representation of each of the split training text sentences is processed to obtain an inference tag of each of the second training texts.

10. A computer system comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, wherein: When the processor executes the program, the steps in the method according to any one of claims 1 to 9 are implemented.

Images