Scientific literature information extraction method and device, computer device and storage medium

Abstract

The application relates to a scientific literature information extraction method and device, computer equipment, a storage medium and a computer program product. The method comprises the following steps: obtaining reference scientific literature, determining an information extraction mode based on the attributes of the reference scientific literature, constructing an information extraction instruction set based on the reference scientific literature and the information extraction mode, the information extraction instruction set comprising information extraction instructions corresponding to different information, labeling the reference scientific literature based on the information extraction instruction set to obtain an information extraction data set, the information extraction data set comprising labeled information, training an initial information extraction model based on the reference scientific literature and the information extraction data set to obtain a target information extraction model, inputting a scientific literature to be processed and a user information extraction instruction into the target information extraction model to obtain a target information extraction result. Information can be extracted from multiple heterogeneous scientific literatures, and the accuracy and scalability of information extraction are improved.

Description

Technical Field

[0001] This application relates to the field of information extraction technology, and in particular to a method, apparatus, computer device, and storage medium for extracting scientific literature information. Background Technology

[0002] Scientific literature, as documents recording the results of scientific research, typically records and describes research questions, methods, results, and conclusions, playing a vital role in advancing academic research, disseminating research findings, and promoting knowledge sharing. Through scientific literature, researchers can understand previous research findings, the current state of research, and development trends, thereby better enabling them to conduct their own research. However, with the development of information technology, the scale and amount of information contained in scientific literature are increasing dramatically. Relying solely on manual reading to extract useful information requires a significant amount of time and effort. Therefore, information extraction technology, as a tool that combines natural language processing and database technology, is increasingly being widely used by researchers.

[0003] Traditional information extraction techniques typically employ conventional machine learning methods, based on rules or classifiers such as support vector machines and decision trees, to extract information from scientific literature. However, this approach only performs well on specific domains or datasets and is not applicable to all scientific literature. Furthermore, the diverse and heterogeneous nature of scientific literature, with its varied sources and formats, significantly impacts the accuracy and robustness of information extraction. Therefore, there is an urgent need for a method that can extract information from diverse and heterogeneous scientific literature while simultaneously improving its accuracy and scalability. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, computer equipment, and computer-readable storage medium for extracting scientific literature information that can extract diverse and heterogeneous scientific literature while improving the accuracy and scalability of the extraction, in order to address the above-mentioned technical problems.

[0005] Firstly, this application provides a method for extracting information from scientific literature. The method includes:

[0006] Obtain reference scientific literature, and determine the information extraction method based on the attributes of the reference scientific literature;

[0007] An information extraction instruction set is constructed based on the aforementioned reference scientific literature and information extraction methods. The information extraction instruction set includes information extraction instructions corresponding to different types of information.

[0008] The referenced scientific literature is labeled based on the information extraction instruction set to obtain an information extraction dataset, which includes annotation information.

[0009] The initial information extraction model is trained based on the aforementioned reference scientific literature and information extraction dataset to obtain the target information extraction model;

[0010] Input the scientific literature to be processed and the user information extraction instructions into the target information extraction model to obtain the target information extraction results.

[0011] Optionally, in one embodiment of this application, the step of obtaining reference scientific literature and determining the information extraction method based on the attributes of the reference scientific literature includes:

[0012] Identify the source and format of the referenced scientific literature;

[0013] The information extraction method is determined based on the source and carrier format.

[0014] Optionally, in one embodiment of this application, constructing the information extraction instruction set based on the referenced scientific literature and information extraction method includes:

[0015] Multiple information extraction targets are identified based on the content and metadata of the referenced scientific literature;

[0016] Multiple information extraction instructions are generated based on the multiple information extraction objects and information extraction methods, forming an information extraction instruction set.

[0017] Optionally, in one embodiment of this application, the step of annotating the reference scientific literature based on the information extraction instruction set to obtain the information extraction dataset includes:

[0018] The extraction accuracy and completeness of the information extraction instruction are determined based on the referenced scientific literature and annotation information.

[0019] Optionally, in one embodiment of this application, training an initial information extraction model based on the reference scientific literature and the information extraction dataset to obtain a target information extraction model includes:

[0020] The referenced scientific literature is input into the initial information extraction model to obtain initial information;

[0021] The extraction loss function is determined based on the initial information and the annotation information, and the parameters of the initial information extraction model are adjusted based on the extraction loss function.

[0022] The referenced scientific literature is input into the adjusted initial information extraction model to obtain the extraction result, and the extraction result is input into the reward scoring model to obtain the scoring result, wherein the reward scoring model is trained based on the adjusted initial information extraction model;

[0023] Based on the extraction and scoring results, a reinforcement loss function is determined, and the parameters of the initial information extraction model are adjusted based on the reinforcement loss function to obtain the target information extraction model.

[0024] Optionally, in one embodiment of this application, the reward scoring model is trained based on the adjusted initial information extraction model in the following manner:

[0025] The initial information is input into the adjusted initial information extraction model to obtain the cumulative reward;

[0026] Based on the cumulative reward and standard score, a reward loss function is determined, and the parameters of the initial information extraction model are adjusted based on the cumulative reward and policy loss function to obtain the reward scoring model.

[0027] Optionally, in one embodiment of this application, the step of inputting the scientific literature to be processed into the target information extraction model and obtaining the target information extraction result includes:

[0028] The extracted target information is stored in a scientific literature extraction database.

[0029] Secondly, this application also provides a scientific literature information extraction device. The device includes:

[0030] The reference scientific literature acquisition module is used to acquire reference scientific literature and determine the information extraction method based on the attributes of the reference scientific literature.

[0031] An information extraction instruction set construction module is used to construct an information extraction instruction set based on the referenced scientific literature and information extraction methods. The information extraction instruction set includes information extraction instructions corresponding to different types of information.

[0032] An information extraction dataset construction module is used to annotate the reference scientific literature based on the information extraction instruction set to obtain an information extraction dataset, wherein the information extraction dataset includes annotation information;

[0033] The information extraction model training module is used to train an initial information extraction model based on the reference scientific literature and the information extraction dataset to obtain the target information extraction model.

[0034] The information extraction module is used to input the scientific literature to be processed and the user information extraction instructions into the target information extraction model to obtain the target information extraction results.

[0035] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the steps of the methods described in the various embodiments above.

[0036] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps of the methods described in the various embodiments above.

[0037] The aforementioned scientific literature information extraction method, apparatus, computer equipment, and storage medium first acquire reference scientific literature, determine an information extraction method based on the attributes of the reference scientific literature, then construct an information extraction instruction set based on the reference scientific literature and the information extraction method, the information extraction instruction set including information extraction instructions corresponding to different information, then annotate the reference scientific literature based on the information extraction instruction set to obtain an information extraction dataset, the information extraction dataset including annotation information, then train an initial information extraction model based on the reference scientific literature and the information extraction dataset to obtain a target information extraction model, and finally input the scientific literature to be processed and the user's information extraction instructions into the target information extraction model to obtain the target information extraction result. In other words, by leveraging the characteristics of scientific literature, employing different information extraction methods, generating information extraction instructions based on different information extraction methods, constructing a training dataset, and training the extraction model, information extraction from diverse and heterogeneous scientific literature is achieved. Furthermore, by using a generative model, it can automatically adapt to scientific literature from different fields and formats, improving the accuracy and scalability of information extraction. Attached Figure Description

[0038] Figure 1 This is a diagram illustrating the application environment of a scientific literature information extraction method in one embodiment.

[0039] Figure 2 This is a flowchart illustrating a scientific literature information extraction method in one embodiment;

[0040] Figure 3 This is a schematic diagram of the system interface for implementing a scientific literature information extraction method in one embodiment.

[0041] Figure 4 This is a structural block diagram of a scientific literature information extraction device in one embodiment;

[0042] Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0044] The scientific literature information extraction method provided in this application embodiment can be applied to, for example, Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or located in the cloud or on other network servers. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. Portable wearable devices can include smartwatches, smart bracelets, head-mounted devices, etc. Server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0045] In one embodiment, such as Figure 2 As shown, a method for extracting information from scientific literature is provided, and this method is applied to... Figure 1 Taking the server in the example, the following steps are included:

[0046] S201: Obtain reference scientific literature and determine the information extraction method based on the attributes of the reference scientific literature.

[0047] In this embodiment, firstly, reference scientific literature is obtained. Optionally, the reference scientific literature is diverse and heterogeneous, and may come from different sources, such as academic journals, academic conferences, and academic papers. It may also have different structures and can usually be obtained through publicly available scientific literature websites such as arXiv. Then, the information extraction method is determined based on the attributes of the obtained reference scientific literature. Different reference scientific literature has different attributes due to different sources, such as different layouts and different carrier formats. In addition, different reference scientific literature also has a variety of presentation forms. Based on different attributes, there are different information extraction methods, all of which can be used to extract the required information from the reference scientific literature.

[0048] S203: Construct an information extraction instruction set based on the aforementioned reference scientific literature and information extraction method, wherein the information extraction instruction set includes information extraction instructions corresponding to different types of information.

[0049] In this embodiment, after determining the information extraction method, multiple information extraction instructions are generated based on different information in the referenced scientific literature and the information extraction methods corresponding to different attributes. Each information extraction instruction can extract one piece of information from a referenced scientific literature, and multiple information extraction instructions corresponding to different information constitute an information extraction instruction set. Specifically, the information in the referenced scientific literature includes title, abstract, author, main text, references, publication year, etc.

[0050] S205: Based on the information extraction instruction set, the reference scientific literature is labeled to obtain an information extraction dataset, which includes the labeling information.

[0051] In this embodiment, after constructing the information extraction instruction set, different information from the referenced scientific literature is extracted based on different information extraction instructions in the instruction set to obtain extracted information. This extracted information is then labeled to obtain labeled information. All labeled information constitutes the information extraction dataset. In specific applications, labeling is typically done through machine learning or manual annotation. Manual annotation refers to annotation performed by experts or professionals, adding corresponding labels to the extracted information.

[0052] S207: Train an initial information extraction model based on the aforementioned reference scientific literature and information extraction dataset to obtain the target information extraction model.

[0053] In this embodiment, after obtaining the information extraction dataset, the initial information extraction model is trained based on referenced scientific literature and the information extraction dataset. By continuously adjusting the model parameters, the target information extraction model is finally obtained. Specifically, the initial information extraction model is a generative model, optionally, such as a Large Language Model (LLM), specifically represented by the following formula.

[0054] P(x)=\sum_z P(x|z)P(z)

[0055] Where x represents the generated data, z represents the latent variable that generated the data, P(x|z) represents the probability of generating data x under the latent variable z, and P(z) represents the prior probability of the latent variable.

[0056] In LLM, the latent variable z is typically represented as contextual information from text generation, such as scientific literature text and instructions. LLM estimates P(x|z) and P(z) by learning from a large amount of text data.

[0057] Optionally, the model can also be trained by maximizing the document generation likelihood, as shown in the following formula.

[0058] θ * =argmax θ E~d ata [pθ()]

[0059] Here, x represents the document, and pθ() is the probability that the model generates x. By maximizing the generation likelihood of the document, the model learns to predict the next word in the document, thereby generating coherent and grammatically correct text.

[0060] The model can also be trained by minimizing the sampling error, as shown in the following formula.

[0061]

[0062] Where x is the document, y is the actual annotation of the extracted information in the document, and f θ () represents the model's extraction result for x, and L is the loss function, commonly MSE, Cross-Entropy, etc. By minimizing the error between the extracted result and the true annotation, the model learns to accurately extract information from the document.

[0063] Alternatively, by taking the combination of maximizing the likelihood of document generation and minimizing the sampling error as joint training objectives, a model that performs well on both tasks can be obtained, as shown in the following formula, where λ is used to adjust the weights of the two tasks.

[0064]

[0065] S209: Input the scientific literature to be processed and the user information extraction instructions into the target information extraction model to obtain the target information extraction results.

[0066] In this embodiment of the application, after obtaining the target information extraction model, the scientific literature to be processed and the received user information extraction instructions are input into the trained target information extraction model to obtain the target information extraction result, and the target information extraction result is displayed to the user.

[0067] In the aforementioned scientific literature information extraction method, firstly, reference scientific literature is acquired, and an information extraction method is determined based on the attributes of the reference scientific literature. Then, an information extraction instruction set is constructed based on the reference scientific literature and the information extraction method. This instruction set includes information extraction instructions corresponding to different types of information. Next, the reference scientific literature is labeled based on the information extraction instruction set to obtain an information extraction dataset, which includes labeled information. Then, an initial information extraction model is trained based on the reference scientific literature and the information extraction dataset to obtain a target information extraction model. Finally, the scientific literature to be processed and the user's information extraction instructions are input into the target information extraction model to obtain the target information extraction result. In other words, by leveraging the characteristics of scientific literature, employing different information extraction methods, generating information extraction instructions based on these methods, constructing a training dataset, and training the extraction model, information extraction from diverse and heterogeneous scientific literature is achieved. Furthermore, by using a generative model, it can automatically adapt to scientific literature from different fields and formats, improving the accuracy and scalability of information extraction.

[0068] In one embodiment of this application, the step of obtaining reference scientific literature and determining the information extraction method based on the attributes of the reference scientific literature includes:

[0069] S301: Identify the source and format of the referenced scientific literature.

[0070] S303: Determine the information extraction method based on the source and carrier format.

[0071] In one embodiment of this application, firstly, the source and carrier format of the referenced scientific literature are identified. The source refers to the origin of the referenced scientific literature, such as a journal, conference proceedings, or collection of papers. The carrier format refers to the format of the medium that carries and displays the referenced scientific literature, including PDF, DOC, and HTML. Then, the information extraction method is determined based on the source and carrier format. Since different sources and carrier formats result in different representations of the referenced scientific literature, and different referenced scientific literatures have different expression formats, different information extraction methods can be determined according to their source and carrier format. In specific applications, processing modules such as PDF processing, DOC processing, and HTML processing can be used to extract information from different referenced scientific literatures. It should be noted that the PDF, DOC, and HTML processing modules include natural language processing and image recognition technologies, which can be used to process content in different formats such as text, images, and tables in the referenced scientific literature.

[0072] In this embodiment, by identifying the source and carrier format of the referenced scientific literature, and determining the information extraction method based on the source and carrier format, it is possible to perform preliminary processing on diverse and heterogeneous scientific literature, extract structured information, and meet different needs.

[0073] In one embodiment of this application, constructing the information extraction instruction set based on the referenced scientific literature and information extraction method includes:

[0074] S401: Based on the content and metadata of the referenced scientific literature, identify multiple information extraction targets.

[0075] S403: Generate multiple information extraction instructions based on the multiple information extraction objects and information extraction methods to form an information extraction instruction set.

[0076] In one embodiment of this application, the referenced scientific literature contains various kinds of information, including content such as title, abstract, main text, references, etc., and metadata such as author, affiliation, publication year, etc. These contents and metadata are multiple information extraction objects. Based on these multiple information extraction objects and the previously determined information extraction method, with the assistance of experts, multiple information extraction instructions are generated to form an information extraction instruction set.

[0077] In this embodiment, multiple information extraction objects are determined based on the content and metadata of referenced scientific literature, and multiple information extraction instructions are generated based on the multiple information extraction objects and information extraction methods to form an information extraction instruction set, which can simplify the construction and use of the information extraction model.

[0078] In one embodiment of this application, the step of annotating the reference scientific literature based on the information extraction instruction set to obtain the information extraction dataset includes:

[0079] The extraction accuracy and completeness of the information extraction instruction are determined based on the referenced scientific literature and annotation information.

[0080] In one embodiment of this application, after annotating reference scientific literature based on the information extraction instruction set to obtain the information extraction dataset, the information extraction instructions also need to be evaluated. Specifically, the performance of the information extraction instructions is evaluated by determining the extraction precision and extraction completeness. In specific applications, the extraction precision is determined by calculating the proportion of truly relevant results among the extracted results, i.e., extraction precision = number of correctly extracted results / total number of extracted results. The extraction completeness is determined by the F1 score, i.e., the extraction precision and extraction recall are used to comprehensively judge and evaluate the extraction completeness. The extraction recall is calculated as the proportion of relevant results extracted out of all relevant results, i.e., extraction recall = number of correctly extracted results / total number of relevant results in the reference standard. The calculation method of the F1 score is shown in the following formula.

[0081] F1=2*Precision*Recall / (Precision+Recall)

[0082] Optionally, in specific applications, the performance of information extraction instructions can also be evaluated using other metrics, such as coverage, which is calculated as the number of extracted information components divided by the total number of information components in the document. Consistency and scalability can also be assessed. Consistency refers to evaluating the consistency of repeated extractions from the same document using the same set of extraction instructions, while scalability refers to evaluating the ease with which the extraction instruction set can be extended to new domains and information types.

[0083] In this embodiment, by determining the extraction accuracy and completeness of the information extraction instruction based on referenced scientific literature and annotation information, the performance of the information extraction instruction can be comprehensively evaluated, making the information extraction results more accurate and complete.

[0084] In one embodiment of this application, training an initial information extraction model based on the reference scientific literature and the information extraction dataset to obtain a target information extraction model includes:

[0085] S501: Input the referenced scientific literature into the initial information extraction model to obtain initial information.

[0086] S503: Determine the extraction loss function based on the initial information and the annotation information, and adjust the parameters of the initial information extraction model based on the extraction loss function.

[0087] S505: Input the referenced scientific literature into the adjusted initial information extraction model to obtain the extraction result, and input the extraction result into the reward scoring model to obtain the scoring result, wherein the reward scoring model is trained based on the adjusted initial information extraction model.

[0088] S507: Determine the enhancement loss function based on the extraction results and scoring results, and adjust the parameters of the initial information extraction model based on the enhancement loss function to obtain the target information extraction model.

[0089] In one embodiment of this application, firstly, referenced scientific literature is input into an initial information extraction model to obtain initial information. This initial information extraction model is a text-trained scientific literature base model capable of understanding basic paper structure and semantics. Its implementation logic involves first encoding the training scientific literature into a sequence, then decoding it back into the original text based on the sequence. Next, an extraction loss function is determined based on the initial information extracted by the initial information extraction model and the labeled information in the information extraction dataset. The parameters of the initial information extraction model are then adjusted based on this extraction loss function. The initial information extraction model training ends when the value of the extraction loss function gradually decreases and stabilizes. In practical applications, a supervised training method is typically used to train the initial information extraction model, and the loss is generally calculated using the following formula.

[0090]

[0091] Where, f(x) i ;) refers to initial information, y i The metrics are as follows: N represents the number of samples, and l represents the loss function. It should be noted that in practical applications, the information extraction dataset is typically divided into a training set and a validation set. The training set is used to train the initial information extraction model, while the validation set is used to validate the training results of the initial information extraction model. Validation accuracy is typically used to verify the performance of the trained initial information extraction model. Validation accuracy is calculated using the following formula.

[0092]

[0093] Where, f(x) i ;) refers to initial information, yi The indicator note indicates that N refers to the number of samples.

[0094] Next, the referenced scientific literature is input again into the adjusted initial information extraction model to obtain the extraction results. The obtained extraction results are then input into the pre-trained reward scoring model to obtain the scoring results. Specifically, the reward scoring model is trained based on the adjusted initial information extraction model.

[0095] Next, a reinforcement loss function is determined based on the extraction and scoring results. The parameters of the initial information extraction model are then adjusted based on this reinforcement loss function. When the reinforcement loss function reaches its minimum value, the initial information extraction model is successfully trained, yielding the target information extraction model. In practical applications, unsupervised reinforcement learning is typically used to further train the initial information extraction model. The training objective is to maximize the agent's cumulative reward value. Taking Proximal Policy Optimization (PPO) as an example, the loss function used in PPO is the same as that in Trust Region Policy Optimization (TRPO), and its specific calculation method is shown below.

[0096]

[0097] Where θ represents the policy parameters, θ′ represents the updated policy parameters, and a t Let KL(π) represent the dominance function at time t. θ′ |πθ) represents the KL divergence between two policies. The state space determines the agent's action space. In the loss function, the state space represents the state s at time t. t The state space contains all the information the agent has at the current moment. In Reinforcement Learning from Human Feedback (RLHF), the state space can include the current input text, the model's internal state, etc. The action space determines the number of actions the agent can take. In the loss function, the action space represents the action a at time t. t The action space is all the operations an agent can take at the current moment. In RLHF, the action space can include generating text, selecting the next word, etc. The reward function determines the agent's behavioral objective. In the loss function, the reward function represents the reward r obtained at time t. t The reward function is the reward an agent receives after taking a certain action. In RLHF, the reward function can be trained using human feedback data. The policy determines the agent's performance. In the loss function, the policy represents the performance of the agent in state s. t Take action a tThe probability π θ (a t |s t A policy is the probability that an agent will take a certain action in a given state. In RLHF, the policy can be trained using the PPO algorithm. The optimizer determines the efficiency of the training process. In the loss function, the optimizer is used to update the policy parameters θ. The optimizer is the algorithm used to update the policy parameters. In RLHF, PPO typically uses the Adam optimizer. Hyperparameters affect the effectiveness of the training process. In the loss function, hyperparameters are used to control the training process, such as the learning rate and the optimizer parameters. Hyperparameters are parameters that control the training process. In RLHF, the hyperparameters of PPO include the learning rate and the optimizer parameters.

[0098] In this embodiment, initial information is obtained by inputting reference scientific literature into the initial information extraction model. An extraction loss function is determined based on the initial information and the labeled information. The parameters of the initial information extraction model are adjusted based on the extraction loss function. The reference scientific literature is then input into the adjusted initial information extraction model to obtain the extraction result. The extraction result is then input into the reward scoring model to obtain the scoring result. The reward scoring model is trained based on the adjusted initial information extraction model. A reinforcement loss function is determined based on the extraction result and the scoring result. The parameters of the initial information extraction model are adjusted based on the reinforcement loss function to obtain the target information extraction model. By supervised fine-tuning training and introducing a reward scoring mechanism, and unsupervised reinforcement training, the accuracy, robustness, and generalization of the model are improved.

[0099] In one embodiment of this application, the reward scoring model is trained based on the adjusted initial information extraction model in the following manner:

[0100] S601: Input the initial information into the adjusted initial information extraction model to obtain the cumulative reward.

[0101] S603: Determine the reward loss function based on the cumulative reward and standard score, and adjust the parameters of the initial information extraction model based on the cumulative reward and policy loss function to obtain the reward scoring model.

[0102] In one embodiment of this application, firstly, initial information is input into an adjusted initial information extraction model to obtain a cumulative reward. Specifically, in the model, the cumulative reward is calculated using the following formula.

[0103]

[0104] Among them, R t The cumulative reward at time t, r i Refers to an immediate reward.

[0105] Next, a reward loss function is determined based on the cumulative reward and the standard score. The parameters of the initial information extraction model are then adjusted based on the cumulative reward and the reward loss function. The reward scoring model is considered complete when the cumulative reward is higher and the value of the reward loss function is smaller and gradually stabilizes. Specifically, the reward loss function in the model can be expressed by the following formula.

[0106] L(θ)=(R t -(s t ;)) 2

[0107] Wherein, V(s) t ;) refers to state s t The value estimate, i.e., the standard score.

[0108] In this embodiment, by inputting initial information into an adjusted initial information extraction model, a cumulative reward is obtained. A reward loss function is determined based on the cumulative reward and the standard score. The parameters of the initial information extraction model are adjusted based on the cumulative reward and the policy loss function to obtain a reward scoring model. By introducing a reward mechanism, the robustness of the information extraction model can be further improved.

[0109] In one embodiment of this application, the step of inputting the scientific literature to be processed into the target information extraction model and obtaining the target information extraction result includes:

[0110] The extracted target information is stored in a scientific literature extraction database.

[0111] In one embodiment of this application, after obtaining the target information extraction result, the target information extraction result is stored in a scientific literature extraction database. When a user initiates information extraction again from scientific literature that has already undergone information extraction, the information extraction result can be directly matched from the scientific literature extraction database and returned to the user for viewing.

[0112] In this embodiment, by storing the target information extraction results in a scientific literature extraction database, structured representation and management are achieved, which facilitates user retrieval and utilization.

[0113] The following specific embodiment illustrates the detailed implementation steps of the scientific literature information extraction method of this application. For example... Figure 3 The image shows the interface of the scientific literature information extraction method system. Users can upload their personal literature, which is received and stored in the system's database. Users can also select existing scientific literature and define information extraction instructions on the interface. The system receives the information and extracts information based on the selected scientific literature and the information extraction instructions.

[0114] First, step S701 involves acquiring reference scientific literature and determining the information extraction method based on the attributes of the reference scientific literature. Specifically, steps S703-S705 involve identifying the source and carrier format of the reference scientific literature and determining the information extraction method based on the source and carrier format. In the system, this is implemented by a multimodal parsing module, which can parse scientific literature in various formats and parse multiple modalities within a unified format, such as text, images, and tables.

[0115] Next, in step S707, an information extraction instruction set is constructed based on the referenced scientific literature and the information extraction method. The information extraction instruction set includes information extraction instructions corresponding to different types of information. Specifically, in steps S709-711, multiple information extraction objects are determined based on the content and metadata of the referenced scientific literature, and multiple information extraction instructions are generated based on the multiple information extraction objects and the information extraction method to form an information extraction instruction set.

[0116] Next, in step S713, the reference scientific literature is annotated based on the information extraction instruction set to obtain an information extraction dataset, which includes annotation information. Then, in step S715, the extraction accuracy and extraction completeness of the information extraction instructions are determined based on the reference scientific literature and the annotation information.

[0117] Next, in step S717, an initial information extraction model is trained based on the reference scientific literature and the information extraction dataset to obtain the target information extraction model. Specifically, in steps S719-725, the reference scientific literature is input into the initial information extraction model to obtain initial information; an extraction loss function is determined based on the initial information and the annotation information, and the parameters of the initial information extraction model are adjusted based on the extraction loss function; the reference scientific literature is input into the adjusted initial information extraction model to obtain extraction results, and the extraction results are input into a reward scoring model to obtain scoring results, wherein the reward scoring model is trained based on the adjusted initial information extraction model; a reinforcement loss function is determined based on the extraction results and the scoring results, and the parameters of the initial information extraction model are adjusted based on the reinforcement loss function to obtain the target information extraction model. The reward model can be trained in the following way: S727-729, the initial information is input into the adjusted initial information extraction model to obtain the cumulative reward; the reward loss function is determined based on the cumulative reward and the standard score; the parameters of the initial information extraction model are adjusted based on the cumulative reward and the policy loss function to obtain the reward scoring model.

[0118] Next, in step S731, the scientific literature to be processed and the user information extraction instruction are input into the target information extraction model to obtain the target information extraction result. Finally, in step S733, the target information extraction result is stored in the scientific literature extraction database. When the information extraction is completed or the information required by the user is matched from the scientific literature extraction database, the information is displayed to the user in a structured form.

[0119] In one embodiment of this application, the initial information extraction model is selected as either the generative model Chat-GLM-6B or Llama-2-7B. Both Chat-GLM-6B and Llama-2-7B are Transformer-based generative models that have achieved excellent results on various generative tasks. Specifically, Chat-GLM-6B can directly encode the scientific document to be extracted and the extraction instructions into a sequence, and then decode it based on that sequence. Llama-2-7B can also directly encode the scientific document into a sequence and then decode it based on that sequence. During training, Chat-GLM-6B and Llama-2-7B learn the semantics and structure of the scientific document, thereby improving the accuracy and robustness of extraction. Furthermore, Chat-GLM-6B and Llama-2-7B can also improve the coverage and richness of extraction by fusing information from multiple modalities and types. For example, for scientific documents containing both text and images, Chat-GLM-6B and Llama-2-7B can fuse information from the text and images, thereby improving the accuracy of extraction.

[0120] Specifically, first, a training dataset containing a scientific literature base model is collected. This training dataset can include information from various modalities such as text, images, and tables. Next, text training is performed on the scientific literature base model dataset to obtain the PaperExtractGPT_base model, enabling the model to understand basic paper structure and semantics. Encoding involves encoding scientific literature into a sequence, and decoding involves decoding the scientific literature back into its original text based on the sequence. Then, the PaperExtractGPT_base model is trained in a supervised manner using a labeled scientific literature extraction training dataset to obtain PaperExtractGPT_1. Finally, a reward model, PaperExtractGPT_reword, is trained based on the scientific literature extraction model PaperExtractGPT_1 trained by the supervised fine-tuning module. This model scores the results extracted by the scientific literature extraction model (0-5 points), with 5 points for qualified extractions and low points for unqualified extractions. Subsequently, the model was fine-tuned in an unsupervised manner using an unlabeled scientific literature extraction training dataset, and PaperExtractGPT_reword was used as a supervised model to improve the model's generalization, resulting in the final target information extraction model PaperExtractGPT.

[0121] Finally, the target information extraction model is used to extract scientific literature information. The scientific literature to be processed and the user information extraction instructions are input into the target information extraction model. The target information extraction model will extract information from the scientific literature according to the user information extraction instructions and generate JSON format according to the instructions. The model generation and processing module will process the JSON text generated by the model into MySQL fields, that is, the target information extraction structure, and store it in the specified scientific literature extraction database for researchers to conduct research and use.

[0122] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0123] Based on the same inventive concept, this application also provides a scientific literature information extraction device for implementing the scientific literature information extraction method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the scientific literature information extraction device provided below can be found in the limitations of the scientific literature information extraction method described above, and will not be repeated here.

[0124] In one embodiment, such as Figure 4 As shown, a scientific literature information extraction device 400 is provided, including: a reference scientific literature acquisition module 401, an information extraction instruction set construction module 403, an information extraction dataset construction module 405, an information extraction model training module 407, and an information extraction module 409, wherein:

[0125] The reference scientific literature acquisition module 401 is used to acquire reference scientific literature and determine the information extraction method based on the attributes of the reference scientific literature.

[0126] The information extraction instruction set construction module 403 is used to construct an information extraction instruction set based on the referenced scientific literature and information extraction method. The information extraction instruction set includes information extraction instructions corresponding to different types of information.

[0127] The information extraction dataset construction module 405 is used to annotate the reference scientific literature based on the information extraction instruction set to obtain an information extraction dataset, wherein the information extraction dataset includes annotation information.

[0128] The information extraction model training module 407 is used to train an initial information extraction model based on the reference scientific literature and the information extraction dataset to obtain a target information extraction model.

[0129] The information extraction module 409 is used to input the scientific literature to be processed and the user information extraction instructions into the target information extraction model to obtain the target information extraction result.

[0130] In one embodiment of this application, the reference scientific literature acquisition module is further configured to:

[0131] Identify the source and format of the referenced scientific literature;

[0132] The information extraction method is determined based on the source and carrier format.

[0133] In one embodiment of this application, the information extraction instruction set construction module is further configured to:

[0134] Multiple information extraction targets are identified based on the content and metadata of the referenced scientific literature;

[0135] Multiple information extraction instructions are generated based on the multiple information extraction objects and information extraction methods, forming an information extraction instruction set.

[0136] The scientific literature information extraction device further includes an extraction instruction evaluation module. In one embodiment of this application, the extraction instruction evaluation module is used for:

[0137] The extraction accuracy and completeness of the information extraction instruction are determined based on the referenced scientific literature and annotation information.

[0138] In one embodiment of this application, the information extraction model training module is further configured to:

[0139] The referenced scientific literature is input into the initial information extraction model to obtain initial information;

[0140] The extraction loss function is determined based on the initial information and the annotation information, and the parameters of the initial information extraction model are adjusted based on the extraction loss function.

[0141] The referenced scientific literature is input into the adjusted initial information extraction model to obtain the extraction result, and the extraction result is input into the reward scoring model to obtain the scoring result, wherein the reward scoring model is trained based on the adjusted initial information extraction model;

[0142] Based on the extraction and scoring results, a reinforcement loss function is determined, and the parameters of the initial information extraction model are adjusted based on the reinforcement loss function to obtain the target information extraction model.

[0143] In one embodiment of this application, the information extraction model training module is further configured to:

[0144] The initial information is input into the adjusted initial information extraction model to obtain the cumulative reward;

[0145] Based on the cumulative reward and standard score, a reward loss function is determined, and the parameters of the initial information extraction model are adjusted based on the cumulative reward and policy loss function to obtain the reward scoring model.

[0146] The scientific literature information extraction device further includes an extraction result management module. In one embodiment of this application, the extraction result management module is used for:

[0147] The extracted target information is stored in a scientific literature extraction database.

[0148] Each module in the aforementioned scientific literature information extraction device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0149] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a method for extracting scientific literature information. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0150] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0151] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0152] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0153] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0154] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0155] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0156] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0157] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for extracting information from scientific literature, characterized in that, The method includes: Obtain reference scientific literature, and determine the information extraction method based on the attributes of the reference scientific literature; An information extraction instruction set is constructed based on the aforementioned reference scientific literature and information extraction methods. The information extraction instruction set includes information extraction instructions corresponding to different types of information. The referenced scientific literature is labeled based on the information extraction instruction set to obtain an information extraction dataset, which includes annotation information. The initial information extraction model is trained based on the aforementioned reference scientific literature and information extraction dataset to obtain the target information extraction model; Input the scientific literature to be processed and the user information extraction instructions into the target information extraction model to obtain the target information extraction results; The process of training an initial information extraction model based on the reference scientific literature and information extraction dataset to obtain a target information extraction model includes: The referenced scientific literature is input into the initial information extraction model to obtain initial information; The extraction loss function is determined based on the initial information and the annotation information, and the parameters of the initial information extraction model are adjusted based on the extraction loss function. The referenced scientific literature is input into the adjusted initial information extraction model to obtain the extraction result, and the extraction result is input into the reward scoring model to obtain the scoring result, wherein the reward scoring model is trained based on the adjusted initial information extraction model; Based on the extraction results and scoring results, a reinforcement loss function is determined, and the parameters of the initial information extraction model are adjusted based on the reinforcement loss function to obtain the target information extraction model. The reward scoring model is trained based on the adjusted initial information extraction model in the following way: The initial information is input into the adjusted initial information extraction model to obtain the cumulative reward; Based on the cumulative reward and standard score, a reward loss function is determined, and the parameters of the initial information extraction model are adjusted based on the cumulative reward and policy loss function to obtain the reward scoring model.

2. The method according to claim 1, characterized in that, The method of obtaining reference scientific literature, and determining the information extraction method based on the attributes of the reference scientific literature, includes: Identify the source and format of the referenced scientific literature; The information extraction method is determined based on the source and carrier format.

3. The method according to claim 1, characterized in that, The construction of the information extraction instruction set based on the aforementioned scientific literature and information extraction method includes: Multiple information extraction targets are identified based on the content and metadata of the referenced scientific literature; Multiple information extraction instructions are generated based on the multiple information extraction objects and information extraction methods, forming an information extraction instruction set.

4. The method according to claim 1, characterized in that, After annotating the referenced scientific literature based on the information extraction instruction set to obtain the information extraction dataset, the process includes: The extraction accuracy and completeness of the information extraction instruction are determined based on the referenced scientific literature and annotation information.

5. The method according to claim 1, characterized in that, The step of inputting the scientific literature to be processed into the target information extraction model and obtaining the target information extraction result includes: The extracted target information is stored in a scientific literature extraction database.

6. A scientific literature information extraction device, characterized in that, The device includes: The reference scientific literature acquisition module is used to acquire reference scientific literature and determine the information extraction method based on the attributes of the reference scientific literature. An information extraction instruction set construction module is used to construct an information extraction instruction set based on the referenced scientific literature and information extraction methods. The information extraction instruction set includes information extraction instructions corresponding to different types of information. An information extraction dataset construction module is used to annotate the reference scientific literature based on the information extraction instruction set to obtain an information extraction dataset, wherein the information extraction dataset includes annotation information; The information extraction model training module is used to train an initial information extraction model based on the reference scientific literature and the information extraction dataset to obtain the target information extraction model. The information extraction module is used to input the scientific literature to be processed and the user information extraction instructions into the target information extraction model to obtain the target information extraction result; The information extraction model training module is further configured to: input the reference scientific literature into the initial information extraction model to obtain initial information; determine an extraction loss function based on the initial information and annotation information; adjust the parameters of the initial information extraction model based on the extraction loss function; input the reference scientific literature into the adjusted initial information extraction model to obtain extraction results; input the extraction results into a reward scoring model to obtain scoring results, wherein the reward scoring model is trained based on the adjusted initial information extraction model; determine a reinforcement loss function based on the extraction results and scoring results; adjust the parameters of the initial information extraction model based on the reinforcement loss function to obtain the target information extraction model. The reward scoring model is trained based on the adjusted initial information extraction model in the following manner: the initial information is input into the adjusted initial information extraction model to obtain a cumulative reward; a reward loss function is determined based on the cumulative reward and the standard score; and the parameters of the initial information extraction model are adjusted based on the cumulative reward and the policy loss function to obtain the reward scoring model.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images