Information extraction method and device, storage medium and program product

Abstract

The invention discloses an information extraction method and device, a storage medium and a program product, and relates to the technical field of information extraction, and the method comprises the steps: obtaining an information extraction task and a to-be-extracted document; based on the information extraction task, constructing an information extraction directed graph corresponding to the to-be-extracted document; wherein the information extraction directed graph comprises at least one main node and auxiliary nodes associated with the main node, each main node corresponds to one type of main information to be extracted, and each auxiliary node corresponds to one type of auxiliary information to be extracted; based on the information extraction directed graph, calling a large language model to execute information extraction operation on the to-be-extracted document to obtain an information extraction result; wherein the information extraction result comprises a trunk information extraction result and an affiliated information extraction result. The technical problem of low information extraction accuracy in related technologies can be solved.

Description

Technical Field

[0001] The present application relates to the technical field of information extraction, and in particular to an information extraction method, device, storage medium and program product. Background Art

[0002] In related technologies, the extraction task and the original document can be directly input into the large language model as input information. The large language model uses its powerful parameter basis to capture and understand the input information, and then obtains the corresponding information extraction results.

[0003] However, in practical applications, for more complex extraction tasks, large language models may experience "hallucination", that is, they may create content that seems reasonable but is actually wrong, thereby reducing the accuracy of information retrieval. Summary of the invention

[0004] The main purpose of this application is to provide an information extraction method, device, storage medium and program product, aiming to solve the technical problem of low information extraction accuracy in related technologies.

[0005] To achieve the above purpose, the present application proposes an information extraction method, which includes: Obtain information extraction tasks and documents to be extracted; Based on the information extraction task, an information extraction directed graph corresponding to the document to be extracted is constructed; wherein the information extraction directed graph includes at least one main node and subsidiary nodes associated with the main node, each main node corresponds to a type of main information to be extracted, and each subsidiary node corresponds to a type of subsidiary information to be extracted; Based on the information extraction directed graph, the large language model is called to perform information extraction operations on the document to be extracted to obtain information extraction results; wherein the information extraction results include trunk information extraction results and subsidiary information extraction results.

[0006] In one embodiment, based on the information extraction directed graph, calling the large language model to perform information extraction operation on the document to be extracted, the step of obtaining the information extraction result includes: Based on the information extraction directed graph, the large language model is called to perform a backbone information extraction operation on the document to be extracted to obtain a backbone information extraction result; wherein the backbone information extraction result includes at least one backbone information set, and each backbone information set includes multiple backbone information; Based on the main information extraction result and the information extraction directed graph, the large language model is called to extract the subsidiary information associated with the main information from the document to be extracted, and the subsidiary information extraction result is obtained; Based on the main information extraction result and the subsidiary information extraction result, an information extraction result is obtained.

[0007] In one embodiment, based on the information extraction directed graph, calling the large language model to perform a trunk information extraction operation on the document to be extracted, before obtaining the trunk information extraction result, the step further includes: Based on the information extraction directed graph, determine the prior relationship combination between multiple trunk information to be extracted; Based on the information extraction directed graph, the steps of calling the large language model to perform the main information extraction operation on the extracted document and obtaining the main information extraction result include: For each type of backbone information to be extracted, the large language model is called to perform zero-prior information extraction on the document to be extracted to obtain a zero-prior backbone information set; wherein each zero-prior backbone information set corresponds to one type of backbone information to be extracted, and each zero-prior backbone information set includes at least one backbone information; Based on the combination of prior relations, a conditional backbone information set as a prior condition is determined from all zero-prior backbone information sets; Each backbone information in the conditional backbone information set is used as prior information, and the large language model is called to extract prior information of the document to be extracted, so as to obtain the prior backbone information set corresponding to the conditional backbone information set; Based on the conditional backbone information set and the prior backbone information set, the backbone information extraction result is determined.

[0008] In one embodiment, based on the trunk information extraction result and the information extraction directed graph, the large language model is called to extract the subsidiary information associated with the trunk information from the document to be extracted, and the step of obtaining the subsidiary information extraction result includes: In the case where the auxiliary information to be extracted is preset structural information, for each trunk information in the trunk information extraction result, the trunk information is used as a priori condition, and the large language model is called to extract the priori information of the document to be extracted to obtain the first auxiliary information; In the case where the auxiliary information to be extracted is non-preset structural information, initial trunk information is determined from the trunk information extraction result; The initial backbone information is used as a priori condition, and the large language model is called to extract the priori information of the document to be extracted to obtain the second subsidiary information; The prior condition is updated based on the second subsidiary information, and the step of calling the large language model to extract the prior information from the document to be extracted to obtain the second subsidiary information is returned to be executed until a preset stop condition of the information extraction operation is met; Based on all the first attached information and all the second attached information, an attached information extraction result is obtained.

[0009] In one embodiment, based on the information extraction directed graph, calling the large language model to perform information extraction operation on the document to be extracted, the step of obtaining the information extraction result includes: The document to be extracted is divided according to a preset zoom factor to obtain multiple sub-documents; Based on the information extraction directed graph, perform information extraction operations on each sub-document respectively to obtain a first information extraction result corresponding to each sub-document; For each sub-document to which each first information extraction result belongs, segmenting the target document according to a preset scaling factor to obtain sub-target documents; The sub-document is updated based on the sub-target document, and the step of returning to execute based on the information extraction directed graph, performing information extraction operations on each sub-document respectively, and obtaining a first information extraction result corresponding to each sub-document, until the byte length corresponding to the sub-target document is less than or equal to the preset byte length; Based on all the first information extraction results, an information extraction result is obtained.

[0010] In one embodiment, the step of obtaining the information extraction result based on all the first information extraction results includes: Performing average weighting on all first information extraction results to obtain weighted scores corresponding to each extracted information in all first information extraction results; Based on the weighted scores, all extracted information is screened and filtered to obtain the information extraction results.

[0011] In one embodiment, the step of obtaining the document to be extracted includes: Obtaining an initial document; wherein the initial document is the original text from which information extraction is required; When the number of bytes of the initial document exceeds the preset number of bytes, the initial document is segmented to obtain a plurality of documents to be extracted; wherein the number of bytes of each document to be extracted does not exceed the preset number of bytes; When the number of bytes of the initial document does not exceed the preset number of bytes, the initial document is determined as the document to be extracted.

[0012] In addition, to achieve the above objectives, the present application also proposes an information extraction device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the information extraction method as described above.

[0013] In addition, to achieve the above-mentioned purpose, the present application also proposes a storage medium, which is a computer-readable storage medium. A computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the information extraction method as described above are implemented.

[0014] In addition, to achieve the above-mentioned purpose, the present application also proposes a computer program product, which includes a computer program. When the computer program is executed by a processor, the steps of the above-mentioned information extraction method are implemented.

[0015] One or more technical solutions proposed in this application have at least the following technical effects: In the information extraction method proposed in the present application, an information extraction task and a document to be extracted can be obtained; and based on the information extraction task, an information extraction directed graph corresponding to the document to be extracted is constructed; wherein the information extraction directed graph includes at least one main node and subsidiary nodes associated with the main node, each main node corresponds to a main information to be extracted, and each subsidiary node corresponds to a subsidiary information to be extracted; and then according to the information extraction directed graph, a large language model is called to extract information, and a final information extraction result including a main information extraction result and a subsidiary information extraction result is obtained. The present application decomposes a complex information extraction task into an information extraction directed graph including a main node and subsidiary nodes, thereby clarifying the association between each main information to be extracted and each subsidiary information to be extracted according to the directed graph, so that when calling the large model for information extraction, the large model can consider the logical relationship between the information as a whole, constrain and guide the extraction process of the large model, reduce the degree of freedom of the large language model, and avoid generating irrelevant or fictitious inaccurate content, thereby effectively improving the accuracy of the information extraction result. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related technologies, the drawings required for use in the embodiments or the related technical descriptions are briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 A schematic diagram of a process flow provided for the first embodiment of the information extraction method of this application; Figure 2 Extracting a directed graph for information of an example; Figure 3 It is a detailed flow chart of step S300; Figure 4 It is a detailed flow chart of step S310; Figure 5 It is a detailed flow chart of step S320; Figure 6 A schematic diagram of the process flow provided for the second embodiment of the information extraction method of this application; Figure 7 Schematic diagram of the device structure of the hardware operating environment involved in the information extraction method in the embodiment of the present application.

[0019] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0020] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and are not used to limit the present application.

[0021] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0022] The main solution of the embodiment of the present application is: obtain information extraction tasks and documents to be extracted; based on the information extraction tasks, construct an information extraction directed graph corresponding to the documents to be extracted; wherein the information extraction directed graph includes at least one main node and subsidiary nodes associated with the main node, each main node corresponds to a type of main information to be extracted, and each subsidiary node corresponds to a type of subsidiary information to be extracted; based on the information extraction directed graph, call a large language model to perform information extraction operations on the documents to be extracted to obtain information extraction results; wherein the information extraction results include main information extraction results and subsidiary information extraction results.

[0023] The field of information extraction, especially information extraction based on large models, is in a stage of rapid technological development. The goal of information extraction (IE: Information Extraction) is to extract structured knowledge such as entities, relations, events, etc. from natural language text. This field faces the challenges of specific task patterns and complex text expressions. With the emergence of large language models (LLM: Large Language Model), we see a new possibility that the natural language understanding capabilities of these models can be used to handle information extraction tasks. Large language models, through their powerful parameter base, are able to capture and understand complex patterns and relationships in text, thus showing great potential in information extraction tasks.

[0024] In related technologies, information extraction methods mainly extract information from identified target paragraphs, and the information to be extracted must be concentrated in the paragraph, which not only increases the screening work, but also limits the application scenarios of the method, and often lacks sufficient flexibility and generalization capabilities. The use of a large language model can solve the problems of flexibility and generalization capabilities, and directly input the extraction task and the original document as input information into the large language model. The large language model can capture and understand the input information through its powerful parameter foundation, and obtain the corresponding information extraction results.

[0025] In practical applications, although large language models have relatively better flexibility and generalization capabilities, for more complex extraction tasks, large language models will still have the "hallucination" phenomenon, that is, they may create content that seems reasonable but is actually wrong, thereby reducing the accuracy of information retrieval.

[0026] The present application provides a solution that can decompose complex information extraction tasks into an information extraction directed graph containing main nodes and subsidiary nodes, thereby clarifying the association relationship between each main information to be extracted and each subsidiary information to be extracted based on the directed graph, so that when calling a large model for information extraction, the large model can consider the logical relationship between the information as a whole, constrain and guide the extraction process of the large model, reduce the degree of freedom of the large language model, avoid generating irrelevant or fictitious inaccurate content, and thus effectively improve the accuracy of the information extraction results.

[0027] It should be noted that the execution subject of this embodiment can be a computing service device with data processing, network communication and program running functions, such as a tablet computer, a personal computer, a mobile phone, etc., or an electronic device capable of realizing the above functions, etc. The following takes the information extraction device as an example to illustrate this embodiment and the following embodiments.

[0028] Based on this, the present application embodiment provides an information extraction method, referring to Figure 1 , Figure 1 This is a flow chart of the first embodiment of the information extraction method of this application.

[0029] In this embodiment, the information extraction method includes steps S100 to S300: Step S100, obtaining information extraction tasks and documents to be extracted.

[0030] Step S200: constructing an information extraction directed graph corresponding to the document to be extracted based on the information extraction task.

[0031] The information extraction directed graph includes at least one main node and subsidiary nodes associated with the main node, each main node corresponds to a type of main information to be extracted, and each subsidiary node corresponds to a type of subsidiary information to be extracted.

[0032] Step S300: Based on the information extraction directed graph, the large language model is called to perform information extraction operations on the document to be extracted to obtain an information extraction result.

[0033] The information extraction results include main information extraction results and auxiliary information extraction results.

[0034] Specifically, when there is a need for information extraction, the user can input the information extraction task and the document to be extracted into the information extraction device; the document to be extracted can be a text file such as a corporate report, an academic paper, or social media; the information extraction task indicates the specific content that needs to be obtained from the document to be extracted. For example, the document to be extracted can be an academic paper based on the surface reaction of organic molecules, and the information extraction task can be to extract the relevant content of the surface reaction of the experiment in the academic paper.

[0035] According to the information extraction task, the main information to be extracted and the subsidiary information to be extracted that the information extraction task may correspond to can be determined, and an information extraction directed graph including at least one main node and subsidiary nodes can be constructed according to the relationship between each main information to be extracted and the subsidiary information to be extracted; in the information extraction directed graph, a main node corresponds to a type of main information to be extracted, and a subsidiary node corresponds to a type of subsidiary information to be extracted, and the subsidiary information to be extracted generally depends on the main information to be extracted; the main node and the subsidiary node can be organized into a directed graph according to the logical relationship of the corresponding information, and the edge of the directed graph can represent the dependency or association relationship between the nodes. For example, when the information extraction task is a surface reaction, it can be determined as follows Figure 2 The information extraction directed graph shown in the figure has reactants, products and substrates as three main information to be extracted, and the reaction functional groups of reactants, reaction sites of products, intermediate products (intermediate product 1, intermediate product 2, etc.) and reaction conditions (reaction condition 1, reaction condition 2, etc.) as auxiliary information to be extracted. It should be noted that in practical applications, the information extraction device can use the existing large model to analyze and process the information extraction task to obtain a possible information extraction directed graph; or the user can analyze the information extraction task in advance based on manual experience to obtain the information extraction directed graph, and then input the information extraction directed graph into the information extraction device when performing information extraction, so that the information extraction device can extract relevant information based on the information extraction directed graph and the document to be extracted.

[0036] After determining the information extraction directed graph, you can directly call a large language model that has the ability to process text and / or images and can perform information extraction, such as GPT, Wenxinyiyan, etc. These models have powerful language understanding and generation capabilities and can handle complex natural language tasks. The document to be extracted, the information extraction directed graph, and the corresponding task instructions are passed to the large language model, so that the large language model can perform information extraction based on the logical relationship between the nodes in the information extraction directed graph, and obtain information extraction results including the main information extraction results and the auxiliary information extraction results. The directed graph can clarify the goals and logical associations during extraction, constrain and guide the information extraction process of the large language model, reduce the degree of freedom of the large model, and avoid generating irrelevant or fictitious inaccurate content.

[0037] In a feasible implementation manner, step S300 may specifically include steps S310 to S330. Figure 3 As shown, Figure 3 It is a detailed flowchart of step S300.

[0038] Step S310: Based on the information extraction directed graph, the large language model is called to perform a main information extraction operation on the document to be extracted to obtain a main information extraction result.

[0039] The backbone information extraction result includes at least one backbone information set, and each backbone information set includes a plurality of backbone information.

[0040] Step S320: Based on the trunk information extraction result and the information extraction directed graph, the large language model is called to extract the subsidiary information associated with the trunk information from the document to be extracted, and the subsidiary information extraction result is obtained.

[0041] Step S330, obtaining an information extraction result based on the main information extraction result and the auxiliary information extraction result.

[0042] Specifically, the information extraction directed graph defines the backbone information to be extracted. Based on the backbone information to be extracted, the large language model can be called to process the document to be extracted, and possible backbone information can be extracted from the document to be extracted. For example, the backbone information to be extracted includes three types: reactants, substrates, and products. The large language model can extract all specific reactants, substrates, and products from the document to be extracted input by the user. The same type of information can be regarded as the same set of backbone information sets. For example, all extracted reactants constitute a set of backbone information sets, and the multiple backbone information contained in the backbone information set are multiple specific reactants. In a feasible implementation, the large language model can be called to obtain the backbone information extraction result by means of zero prior extraction and prior information extraction. Therefore, before step S310, the prior relationship combination between multiple backbone information to be extracted can also be determined based on the information extraction directed graph. Under this implementation, step S310 can specifically include steps S311~S314. Figure 4 As shown, Figure 4 It is a detailed flowchart of step S310.

[0043] Step S311 , for each type of backbone information to be extracted, calling a large language model to perform zero prior information extraction on the document to be extracted, and obtaining a zero prior backbone information set.

[0044] Each zero a priori backbone information set corresponds to a type of backbone information to be extracted, and each zero a priori backbone information set includes at least one backbone information.

[0045] Step S312: Based on the prior relationship combination, a conditional backbone information set serving as a priori condition is determined from all zero priori backbone information sets.

[0046] Step S313, taking each piece of backbone information in the conditional backbone information set as prior information, calling the large language model to extract prior information from the document to be extracted, and obtaining a prior backbone information set corresponding to the conditional backbone information set.

[0047] Step S314, determining a backbone information extraction result based on the conditional backbone information set and the prior backbone information set.

[0048] It should be noted that the information extraction directed graph can reflect the logical relationship between multiple trunk information to be extracted. Based on the logical relationship, multiple prior relationship combinations can be determined for subsequent prior information extraction; for example, in surface reactions, both reactants and substrates have a clear causal relationship with products, which can be expressed as: reactants → products ← substrate; at this time, the following three prior relationship combinations can be determined: 1) reactants and substrates are inferred to products as priors; 2) reactants and products are inferred to substrates as priors; 3) products and substrates are inferred to reactants as priors.

[0049] For each type of backbone information to be extracted in the information extraction directed graph, the large language model can be called to perform zero prior information extraction on the document to be extracted, extract all possible backbone information in the document to be extracted, and obtain a zero prior backbone information set; zero prior information extraction means directly calling the large model to extract the required information to be extracted from the document to be extracted without prior knowledge; the backbone information of the same type extracted can be attributed to the same zero prior backbone information set. Further, according to the combination of prior relationships between the backbone information to be extracted, a conditional backbone information set as a prior condition is determined from all zero prior backbone information sets, and each backbone information in the conditional backbone information set is used as a prior condition, and the backbone information related to these prior conditions is further extracted. The result of the prior information extraction is called a prior backbone information set. The conditional backbone information set and its corresponding prior backbone information set are integrated to form the final backbone information extraction result.

[0050] To facilitate understanding of the above content, T is used here to represent the document to be extracted, and T contains multiple sets of target information d corresponding to the information extraction task (for example, the information extraction task is to extract the surface reaction in the paper, and the paper contains multiple sets of surface reactions d), that is, T={d 1 ,…,d k}; where k represents the number of target information d; in the subsequent representation, superscripts are used to distinguish different groups of information in a document to be extracted, and subscripts are used to distinguish different information in a group (for example, reactants, substrates, and products are three different information, and the subscripts can be represented by 1, 2, and 3 respectively). Sometimes, for the sake of simplicity, superscripts are ignored. When superscripts are ignored, the default description is for information in the same group. The set of backbone information included in the i-th group of target information can be represented as M i ;M i ={m1 i ,..,m n i} (m is the main information, n is the number of main information in the group), the corresponding set of auxiliary information can be expressed as S i ; S i ={s1 i ,..,s l i} (s is the additional information, l is the number of additional information); d 1 =[M 1 ,S 1 ],d 2 =[M 2 ,S 2 ],d 3 =[M 3 ,S 3 ]…. It should be noted that M i ={m1,..,m n The structure of} is fixed, and S i It is possible to change. For example, in a chemical reaction, there may be multiple intermediate reaction steps. Therefore, we can cluster M1={m1 1 , m1 2 , m1 3 ,m1 4 …}, so as to further extract information from the dimension of a single backbone information.

[0051] When performing the backbone information extraction operation, we can define the function f for extracting target information based on the large language model. Then, for the backbone information set M1, it can be expressed as f(M1|T). This method does not require prior knowledge, that is, zero prior knowledge. For prior information extraction, prior knowledge needs to be provided as a priori condition for information extraction; for example, given m2 i , we need to extract the corresponding possible M1, which can be expressed as (M1|T, m2 i Here, we take M1, M2, and M3 as examples to illustrate the three types of backbone information to be extracted, where M1 and M2 are causes and M3 is the result, i.e., M1→M3←M2. For each type of backbone information to be extracted, zero prior information extraction is performed on T to obtain M1，zero = f(M1|T); ​​M 2，zero = f(M2|T); M 3，zero = f(M3|T); where M 1，zero 、M 2，zero and M 3，zero They represent three zero-prior backbone information sets obtained based on M1, M2 and M3 respectively.

[0052] On this basis, we extract prior information, that is, we use the result of zero prior extraction as a prior condition to constrain the output space of the large language model. For ease of understanding, we take M1→M2 as an example, extract M2 given M1 (that is, extract M1 as the conditional backbone information set to obtain the corresponding M2), and then extract M1 given M2 (that is, extract M2 as the conditional backbone information set to obtain the corresponding M1). 1，zero = Each backbone information m1 in f(M1|T) i , M 2，prior = f(M2|T,m1 i ); for M 2，zero = f(M2|T) for each m2 i , M 1，prior = f(M1|T, m2 i );M 1，prior and M 2，prior are the two prior backbone information sets obtained; that is, M 1，zero Each backbone information in m1 i As a priori condition, extract the prior information of M2 and obtain each backbone information m1 i The corresponding prior backbone information, the set of these prior backbone information is M 2，prior ;M 1，prior The acquisition of is similar, and will not be repeated here. Combining the conditional backbone information set and the corresponding prior backbone information set, the backbone information extraction result can be obtained. In the above example, the backbone information extraction result is two sets of { m1 i , m2 i |M 1，zero} (representing the zero prior backbone information of M1 and the set of prior backbone information M2 obtained based on M1) and { m1 i , m2 i |M 2，zero} (representing the set of zero prior backbone information of M2 and the prior backbone information M1 obtained based on M2).

[0053] Based on this, for the above-mentioned M1→M3←M2 situation, there are three prior relationship combinations: 1) M1 and M2 are used as priors to be inferred to M3; 2) M1 and M3 are used as priors to be inferred to M2; 3) M2 and M3 are used as priors to be inferred to M1. The following is an explanation using the combination of M1 and M2 as priors to be inferred to M3 (using other combinations or using all combinations is also feasible). Specifically, using M 1，zero A certain m1 in i As a priori condition, to extract M2, M3: for M 1，zero = each m1 in f(M1|T) i , M 2，prior = f(M2|T,m1 i ) ; then use M 2，prior The main information in m2 i Combined with the corresponding m1 i To extract M3, that is, for M 1，zero = each m1 in f(M1|T) i and M 2，prior = f(M2|T) for each m2 i , M 3，prior = f(M3|T, m1 i , m2 i ); thus we can get a set of extraction results {m1 i ,m2 i ,m3 i |M 1，zero , M 2，prior}; Similarly, M 2，zero As a priori conditions, the above operation can obtain another set of {m1 i ,m2 i ,m3 i |M 2，zero , M1，prior}; Finally, two groups of results can be obtained. In order to obtain the most accurate backbone information extraction results, the multiple groups of results can be weighted, intersected, unioned or processed in other ways. By selecting different merging methods, the coverage (i.e., obtaining as many results as possible) and accuracy (i.e., obtaining the most accurate results possible) can be balanced. i ,m2 i ,m3 i |M 1，zero , M 2，prior} and {m1 i ,m2 i ,m3 i |M 2，zero , M 1，prior} two sets of results as an example, we can select the intersection of the two sets or merge them based on other confidence indicators. This operation balances the final results of the extracted information in terms of recall and accuracy, and can increase the accuracy of the extracted information while ensuring the coverage of the information extraction. For example, we can select the intersection of the two sets to maximize the accuracy, and get the final backbone information extraction result as {m1 i ,m2 i ,m3 i |M 1，zero , M 2，prior}∩{m1 i ,m2 i ,m3 i |M 2，zero , M 1，prior In addition, for the case of multiple steps of backbone information to be extracted, such as M1→M2→M3→M4→M5, the aforementioned steps can be executed step by step, such as M1→M2 first and then M2→M3, until the entire backbone information extraction operation is completed to obtain the final backbone information extraction result.

[0054] Taking the zero prior backbone information set as the basis, calling the large language model to extract prior backbone information can constrain the output space of the large language model and reduce the difficulty of task implementation, thereby reducing the "fantasy" possibility of the large language model (that is, the results output by the large language model are not based on any factual data and contain information that does not conform to the facts) and increasing the accuracy of information extraction.

[0055] After obtaining the main information extraction result, the information extraction directed graph and the extracted main information extraction result can be further used as the basis to call the large language model to extract the auxiliary information associated with the main information from the document to be extracted to obtain the auxiliary information extraction result; the auxiliary information extraction result is a supplement and detailed description of the main information, which can help enhance the integrity of the entire information extraction. In a feasible implementation, step S320 can specifically include steps S321 to S325, such as Figure 5 As shown, Figure 5 It is a detailed flowchart of step S320.

[0056] Step S321, when the subsidiary information to be extracted is preset structural information, for each trunk information in the trunk information extraction result, the trunk information is used as a priori condition, and the large language model is called to perform priori information extraction on the document to be extracted to obtain the first subsidiary information.

[0057] Step S322: when the auxiliary information to be extracted is non-preset structural information, initial trunk information is determined from the trunk information extraction result.

[0058] Step S323, taking the initial trunk information as a priori condition, calling the large language model to extract priori information from the document to be extracted, and obtaining the second supplementary information.

[0059] Step S324, updating the prior condition based on the second supplementary information, and returning to execute the step of calling the large language model to extract prior information from the document to be extracted to obtain the second supplementary information, until the preset stop condition of the information extraction operation is met.

[0060] Step S325: obtaining an additional information extraction result based on all the first additional information and all the second additional information.

[0061] Specifically, after extracting the backbone information, the subsidiary information can be extracted according to the type of subsidiary information to be extracted. The subsidiary information to be extracted can be preset structural information, that is, subsidiary information with fixed structures such as molecular properties and experimental conditions; when the subsidiary information to be extracted is preset structural information, the aforementioned prior information extraction method can be adopted, and each backbone information in the backbone information extraction result is used as a prior condition, and the large language model is called to extract information in combination with the preset structural information to obtain the first subsidiary information. For example, when the subsidiary information s1 to be extracted with a fixed structure needs to be extracted, that is, for each backbone information m1 in the backbone information extraction result, s1=f(s1|T,m1).

[0062] The auxiliary information to be extracted may also be non-preset structural information, that is, the structure of the auxiliary information to be extracted is not fixed, and the subsequent auxiliary information to be extracted may depend on the result of the previous auxiliary information to be extracted. For example, the reaction process and reaction conditions in the surface reaction are usually chain structures of indefinite length; when the auxiliary information to be extracted is non-preset structural information, the initial trunk information can be determined from the trunk information extraction result, and used as a priori condition, and the large language model is called to extract prior information to obtain the second auxiliary information; since there may be a front-to-back dependency relationship between the auxiliary information to be extracted, the obtained second auxiliary information can be used as a new priori condition, and then the step of calling the large language model to extract prior information from the document to be extracted to obtain the second auxiliary information is performed until the preset stop condition is met before stopping the information extraction operation. For example, in addition to constructing the information extraction function f, a preset stop condition function g can also be constructed. After each information extraction, g is used to determine whether the stop condition is met. If not, the current extracted information (i.e., the extracted second auxiliary information) is used as a new priori condition to extract information again, and the above process is continuously repeated until g meets the preset condition. The following uses the surface reaction as an example: the initial state (i.e., the initial backbone information) can be determined as t initial ;t initial = (reactant m1 i , substrate m2 i The final state is tfinal ;t final =(Product m3 i , substrate m2 i ); function f is based on the current state t i Extract the next intermediate state t i+1 =f(T,t i , t initial , t final ); function g is based on the current state t i Determine whether the final state is reached if_end=g(T,t i , t initial , t final ) (i.e. the preset stop condition is that the current state = the final state). It can be expressed in pseudo code as: "Reaction process attached information r = [] t i <- t initial While not g(if_end|T,t i , t initial , t final ): t i+1 =f(t i+1 |T,t i , t initial , t final ) r.append(t i ) t i <- t i+1 ”; The above steps can be used to complete the extraction of all non-fixed structured ancillary information (i.e., second ancillary information); then, by integrating all first ancillary information and all second ancillary information, the ancillary information extraction results of all ancillary information can be obtained. Finally, by combining the main information extraction results and the ancillary information extraction results, a complete and structured information extraction result can be obtained.

[0063] In addition, since this application needs to call a large language model, and the current large language model usually supports the processing of input text with a length of 16k to 128k, corresponding to 50k to 300k characters; for most academic articles (except reviews), these large language models can directly process them. However, for articles that exceed the length that the large language model can process at one time, they need to be processed in advance. Therefore, in a feasible implementation, the step of obtaining the document to be extracted can be specifically as follows: obtaining the initial document; wherein the initial document is the original text for information extraction; when the number of bytes of the initial document exceeds the preset number of bytes, the initial document is segmented to obtain multiple documents to be extracted; wherein the number of bytes of each document to be extracted does not exceed the preset number of bytes; when the number of bytes of the initial document does not exceed the preset number of bytes, the initial document is determined as the document to be extracted. That is, when the number of bytes of the initial document meets the processing requirements of the large language model, the large language model is directly called to perform the aforementioned information extraction operation on the initial document; when the number of bytes of the initial document exceeds the processing range of the large language model, it is necessary to divide the initial document into multiple documents to be extracted whose byte numbers do not exceed the processing range of the large language model, so as to call the large language model to perform the aforementioned information extraction operation. For example, the map-reduce approach can be used to divide the original article into multiple parts, process each part independently, and then merge the results. For example, for the extraction of the main information M1, after extracting M1 from each part separately, the results are taken as the union to obtain a complete result.

[0064] It is not difficult to understand that the information extraction method provided by the present application can be directly input with original data (such as the original text of the article, without the need to extract the target paragraph) and output the final extraction result, so as to realize end-to-end information extraction without manual intervention. And by constructing a directed graph for the information extraction task, it is divided into the main information to be extracted and the subsidiary information to be extracted, and the extraction of complex information is converted into a sub-problem that is easier to solve. Therefore, the association relationship between each main information to be extracted and each subsidiary information to be extracted can be clarified according to the directed graph, so that when the large model is called for information extraction, the large model can consider the logical relationship between the information as a whole. Not only can the association relationship between each main information to be extracted and each subsidiary information to be extracted be clarified according to the directed graph, so that when the large model is called for information extraction, the large model can consider the logical relationship between the information as a whole, but also the extraction process of the large model can be constrained and guided, reducing the degree of freedom of the large language model, avoiding the generation of irrelevant or fictitious inaccurate content, and avoiding the omission and incompleteness of information caused by local information extraction, thereby effectively improving the accuracy of the information extraction results.

[0065] Due to the limitation of large model capabilities in some more complex task extraction, the output results of the model still have a certain degree of "illusion" problem, and the extracted information is still not accurate enough. Therefore, based on the first embodiment of this application, the second embodiment of this application is proposed. In the second embodiment of this application, the same or similar content as the above-mentioned first embodiment can be referred to the above introduction, and will not be repeated later. On this basis, please refer to Figure 6 , step S300 may include steps A100 to A500 to improve the accuracy of the information extraction process: Step A100, segmenting the document to be extracted according to a preset zoom factor to obtain a plurality of sub-documents.

[0066] Step A200: Based on the information extraction directed graph, perform information extraction operations on each sub-document respectively to obtain a first information extraction result corresponding to each sub-document.

[0067] Step A300 , for the sub-documents to which each first information extraction result belongs, the target document is segmented according to a preset scaling factor to obtain sub-target documents.

[0068] Step A400, updates the sub-document based on the sub-target document, returns to the step of executing based on the information extraction directed graph, performs information extraction operations on each sub-document respectively, and obtains the first information extraction result corresponding to each sub-document, until the byte length corresponding to the sub-target document is less than or equal to the preset byte length.

[0069] Step A500, obtaining an information extraction result based on all first information extraction results.

[0070] Specifically, when performing information extraction, the document to be extracted can be segmented according to a preset scaling ratio. For example, if the document to be extracted is 50k bytes and the preset scaling ratio is 0.2, then it can be segmented according to the order of the text in the document to obtain 5 10k-byte sub-documents; then, according to the information extraction directed graph, the information extraction operation is performed on each sub-document to obtain the first information extraction result of each sub-document (for details, refer to the information extraction process in the aforementioned embodiment 1). Further narrow the information extraction scope, that is, for the sub-documents to which each first information extraction result belongs, the target document is segmented again according to the preset scaling factor to obtain a sub-target document; for example, for a 10k-byte sub-document, the sub-document is further segmented according to the preset scaling ratio of 0.2 to obtain 5 2k-byte sub-target documents; the sub-target document is used as a new sub-document, and then based on the information extraction directed graph, the aforementioned information extraction operation is performed on each sub-document to obtain the first information extraction result corresponding to the current sub-document; the aforementioned process of segmenting the document according to the preset scaling ratio is continuously repeated until the byte length corresponding to the sub-target document is less than or equal to the preset byte length; all the first information extraction results are integrated to obtain the information extraction result. The above-mentioned division of the documents to be extracted according to the preset scaling ratio can realize the gradual contraction of the information extraction scope, that is, it can realize the segmentation from coarse to fine. By limiting the scope of fine-grained segmentation, it is guaranteed to the greatest extent that the information corresponding to the information extraction task will not be removed by mistake.

[0071] In practical applications, the above-mentioned division of documents to be retrieved can be implemented based on the adaptive retrieval enhancement method (RAG: Retrieval Augmented Generation). Its basic principle is explained in the following three steps: 1) Define the task: For an information extraction task with a prior condition p, the conventional extraction method is generally o= f(o|t,p), where t is the input document, p is the prior condition, and o is the extraction result corresponding to the information extraction task.

[0072] 2) Construct a retrieval question based on p: Based on the prior condition p, construct a retrieval question for RAG, query(p), and then use RAG to obtain relevant content from the original text based on the retrieval question. As for the retrieval method, considering that the prior condition may usually include specific terms, a hybrid retrieval method is adopted, that is, combining the results of sparse and dense retrieval; among them, sparse retrieval is usually based on information such as keywords and word frequency (such as BM25 algorithm), and dense retrieval is based on the similarity of feature vectors.

[0073] 3) Adaptive segmentation to construct the data to be retrieved: Usually in RAG, a fixed segmentation length is used for the original text, or multi-level segmentation is used for joint retrieval. However, considering the knowledge-intensive nature of the task, this application adopts an adaptive method of gradually narrowing the retrieval range. For example, a segmentation from coarse to fine is adopted. The first step of segmentation uses 1200 words, the second step is 600 words, and the third step is 300 words until the preset fine-grained range is reached, so as to ensure to the greatest extent that the target information is not removed by mistake. The following is an example pseudocode for illustration: The minimum segmentation length (that is, the preset byte length) is l min , the scaling factor (i.e., the preset scaling ratio) α of each segmentation length. The corresponding pseudo code can be expressed as: “The extraction results of the target extraction task of RAG rag =[] RAG retrieved content rag =[] l i <- len(T)* α # Split length t i <- T #Documents to be extracted# Input the content of the large language model o = f(o|t i ,p) #Based on the results of the extraction of the document to be extracted o rag .append(o) while l i > min : t i+1 =retrieval(t i ,l i ,query(p)) l i <- len(t i+1 )*α o = f(o|t i+1 ,p) o rag .append(o) t rag .append(t i+1 ) t i <- t i+1 ”; Finally, we can get the information extraction results based on different article paragraphs according to the information extraction task. rag and the corresponding article paragraphs rag Assume that the extraction is divided into four stages (for example, 1200 bytes, 600 bytes, 300 bytes and 150 bytes). ragThere are 4 stages of results, o rag = [[a1,a2,a3],[a1,a3],[a2,a3],[a1,a3]]; wherein a1, a2 and a3 are the extracted information in the extraction result; [a1,a2,a3] is the first information extraction result corresponding to the 1200-byte division, [a1,a3] is the first information extraction result corresponding to the 600-byte division, [a2,a3] is the first information extraction result corresponding to the 300-byte division, and [a1,a3] is the first information extraction result corresponding to the 150-byte division. It should be noted that the above-mentioned adaptive method of gradually shrinking the information extraction range can be used in the main information extraction process as well as in the subsidiary information extraction process. By integrating all the first information extraction results, the final information extraction result can be obtained.

[0074] In order to further improve the accuracy of the information extraction result, obtain a more effective information extraction result, and avoid redundant extracted information in the extraction result, in a feasible implementation, step A500 may specifically include steps A510 to A520: Step A510, weighting all first information extraction results on an average basis to obtain weighted scores corresponding to each extracted information in all first information extraction results.

[0075] Step A520, based on the weighted scores, all the extracted information is screened and filtered to obtain information extraction results.

[0076] Specifically, the importance or credibility of each extracted information in the first information extraction result obtained above can be evaluated by the average weighted method to provide a basis for subsequent screening; for example, a corresponding weight can be assigned to each extraction result according to the type of information and the correlation between the information, and then a corresponding weighted score can be calculated for each extracted information; information screening can be performed based on this, such as pre-setting a screening threshold, and for all extracted information, screening is performed according to its weighted score to filter out information with lower scores; for example, assuming that the preset screening threshold is 0.75, only the extracted information with a weighted score higher than 0.75 is retained, and the extracted information finally retained is integrated to obtain the information extraction result; alternatively, the weighted score can reflect the confidence of the corresponding extracted information, or the corresponding weighted score can be calculated based on the extracted information to calculate the confidence of each information extraction result, and information screening can be performed based on the confidence to obtain the final information extraction result.

[0077] Taking the above content as an example, we have obtained o rag= [[a1,a2,a3],[a1,a3],[a2,a3],[a1,a3]] This extraction result adopts the weighted average method, that is, the weighted score of each step is the same, and the weighted result can be obtained as follows: a1 weighted score is 0.75, a2 weighted score is 0.5, and a3 weighted score is 1.0. Or for some extraction tasks, the first information extraction result obtained by setting the finest-grained division step to correspond to a higher weighted score according to actual needs. According to the weighted scores of each extracted information, it can be calculated that the confidence of [a1,a2,a3] is (0.75+0.5+1) / 3=0.75; the confidence of [a1,a3] is (0.75+1) / 2=0.875; the confidence of [a2,a3] is (0.5+1) / 2=0.75; then for o rag =[[a1,a2,a3],[a1,a3],[a2,a3],[a1,a3]] The confidence sum and average of all the extraction results can be obtained. rag The mean confidence value is (0.75+0.875+0.75+0.875) / 4=0.815, which reflects the reliability of the extraction result and can provide a reference for subsequent processing.

[0078] It can be understood that the information extraction method provided in this embodiment can gradually narrow the extraction scope of the documents to be extracted. For various information to be extracted corresponding to the information extraction task, weighted statistics can be performed through the corresponding retrieval enhancement results (i.e., the first information extraction results) to obtain the confidence of the corresponding extraction results, thereby further improving the accuracy of the results.

[0079] The present application provides an information extraction device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the information extraction method in the above-mentioned embodiment one.

[0080] Reference below Figure 7 , which shows a schematic diagram of the structure of an information extraction device suitable for implementing an embodiment of the present application. The information extraction device in the embodiment of the present application may include but is not limited to mobile terminals such as mobile phones, laptop computers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions: tablet computers), etc., and fixed terminals such as desktop computers, etc. Figure 7 The information extraction device shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0081] like Figure 7 As shown, the information extraction device may include a processing device 1001 (such as a central processing unit, a graphics processor, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM: Read Only Memory) 1002 or a program loaded from a storage device 1003 to a random access memory (RAM: Random Access Memory) 1004. In RAM1004, various programs and data required for the operation of the information extraction device are also stored. The processing device 1001, ROM1002 and RAM1004 are connected to each other through a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Generally, the following systems can be connected to the I / O interface 1006: an input device 1007 including, for example, a touch screen, a touch pad, a keyboard, a mouse, an image sensor, a microphone, etc.; an output device 1008 including, for example, a liquid crystal display (LCD: Liquid Crystal Display), a speaker, a vibrator, etc.; a storage device 1003 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1009. The communication device 1009 can allow the information extraction device to communicate wirelessly or wired with other devices to exchange data. Although the figure shows an information extraction device with various systems, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems can be implemented or have alternatively.

[0082] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are executed.

[0083] The information extraction device provided by the present application adopts the information extraction method in the above embodiment, which can solve the technical problem of low accuracy of information extraction in the related art. Compared with the related art, the beneficial effects of the information extraction device provided by the present application are the same as the beneficial effects of the information extraction method provided by the above embodiment, and the other technical features in the information extraction device are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.

[0084] It should be understood that the various parts disclosed in this application can be implemented by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0085] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0086] The present application provides a computer-readable storage medium having computer-readable program instructions (ie, computer programs) stored thereon, wherein the computer-readable program instructions are used to execute the information extraction method in the above-mentioned embodiment.

[0087] The computer-readable storage medium provided in the present application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems or devices, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM: Random Access Memory), a read-only memory (ROM: Read Only Memory), an erasable programmable read-only memory (EPROM: Erasable Programmable Read Only Memory or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM: CD-Read Only Memory), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program, which may be used by or in combination with an instruction execution system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency: Radio Frequency), etc., or any suitable combination of the above.

[0088] The computer-readable storage medium may be included in the information extraction device; or may exist independently without being assembled into the information extraction device.

[0089] The above-mentioned computer-readable storage medium carries one or more programs. When the above-mentioned one or more programs are executed by the information extraction device, the information extraction device: obtains an information extraction task and a document to be extracted; based on the information extraction task, constructs an information extraction directed graph corresponding to the document to be extracted; wherein the information extraction directed graph includes at least one main node and subsidiary nodes associated with the main node, each main node corresponds to a type of main information to be extracted, and each subsidiary node corresponds to a type of subsidiary information to be extracted; based on the information extraction directed graph, calls a large language model to perform information extraction operations on the document to be extracted to obtain an information extraction result; wherein the information extraction result includes a main information extraction result and a subsidiary information extraction result.

[0090] Computer program code for performing the operations of the present application may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0091] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0092] The modules involved in the embodiments described in this application may be implemented by software or hardware, wherein the name of the module does not constitute a limitation on the unit itself in some cases.

[0093] The readable storage medium provided in the present application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-mentioned information extraction method, and can solve the technical problem of low accuracy of information extraction in the related art. Compared with the related art, the beneficial effects of the computer-readable storage medium provided in the present application are the same as the beneficial effects of the information extraction method provided in the above-mentioned embodiment, and will not be described in detail here.

[0094] The present application also provides a computer program product, including a computer program, which implements the steps of the above-mentioned information extraction method when executed by a processor.

[0095] The computer program product provided by this application can solve the technical problem of low accuracy of information extraction in related technologies. Compared with related technologies, the beneficial effects of the computer program product provided by this application are the same as those of the information extraction method provided by the above embodiment, which will not be described in detail here.

[0096] The above descriptions are only some embodiments of the present application, and are not intended to limit the patent scope of the present application. All equivalent structural changes made using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect applications in other related technical fields are included in the patent protection scope of the present application.

Claims

1. An information extraction method, characterized in that: The information extraction method comprises: Obtain information extraction tasks and documents to be extracted; Based on the information extraction task, construct an information extraction directed graph corresponding to the document to be extracted; wherein the information extraction directed graph includes at least one main node and subsidiary nodes associated with the main node, each of the main nodes corresponds to a type of main information to be extracted, and each of the subsidiary nodes corresponds to a type of subsidiary information to be extracted; Based on the information extraction directed graph, a large language model is called to perform an information extraction operation on the document to be extracted to obtain an information extraction result; wherein the information extraction result includes a main information extraction result and an auxiliary information extraction result.

2. The information extraction method according to claim 1, characterized in that: The step of calling a large language model to perform an information extraction operation on the document to be extracted based on the information extraction directed graph to obtain an information extraction result comprises: Based on the information extraction directed graph, calling the large language model to perform a trunk information extraction operation on the document to be extracted, and obtaining the trunk information extraction result; wherein the trunk information extraction result includes at least one group of trunk information sets, and each group of the trunk information sets includes multiple trunk information; Based on the trunk information extraction result and the information extraction directed graph, calling a large language model to extract the subsidiary information associated with the trunk information from the document to be extracted, to obtain the subsidiary information extraction result; The information extraction result is obtained based on the main information extraction result and the auxiliary information extraction result.

3. The information extraction method according to claim 2, characterized in that: Before the step of calling a large language model to perform a main information extraction operation on the document to be extracted based on the information extraction directed graph and obtaining the main information extraction result, the step further includes: Determine a priori relationship combinations between a plurality of trunk information to be extracted based on the information extraction directed graph; The step of calling a large language model to perform a main information extraction operation on the document to be extracted based on the information extraction directed graph to obtain a main information extraction result comprises: For each type of backbone information to be extracted, a large language model is called to perform zero-prior information extraction on the document to be extracted to obtain a zero-prior backbone information set; wherein each zero-prior backbone information set corresponds to one type of backbone information to be extracted, and each zero-prior backbone information set includes at least one backbone information; Based on the prior relationship combination, determining a conditional backbone information set as a priori condition from all the zero prior backbone information sets; Each of the backbone information in the conditional backbone information set is used as prior information, and a large language model is called to extract prior information from the document to be extracted, so as to obtain a prior backbone information set corresponding to the conditional backbone information set; Based on the conditional backbone information set and the priori backbone information set, the backbone information extraction result is determined.

4. The information extraction method according to claim 2, characterized in that: The step of extracting the ancillary information associated with the main information from the document to be extracted by calling a large language model based on the main information extraction result and the information extraction directed graph to obtain the ancillary information extraction result comprises: In the case where the to-be-extracted subsidiary information is preset structural information, for each of the trunk information in the trunk information extraction result, the trunk information is used as a priori condition, and a large language model is called to perform priori information extraction on the to-be-extracted document to obtain first subsidiary information; In the case where the auxiliary information to be extracted is non-preset structural information, determining initial trunk information from the trunk information extraction result; Taking the initial trunk information as a priori condition, calling the large language model to extract priori information from the document to be extracted, and obtaining the second subsidiary information; The prior condition is updated based on the second subsidiary information, and the step of calling the large language model to extract the prior information of the document to be extracted to obtain the second subsidiary information is returned to be executed until a preset stop condition of the information extraction operation is met; The auxiliary information extraction result is obtained based on all the first auxiliary information and all the second auxiliary information.

5. The information extraction method according to claim 1, characterized in that: The step of calling a large language model to perform an information extraction operation on the document to be extracted based on the information extraction directed graph to obtain an information extraction result comprises: Segmenting the document to be extracted according to a preset zoom factor to obtain a plurality of sub-documents; Based on the information extraction directed graph, performing information extraction operations on each of the sub-documents respectively to obtain a first information extraction result corresponding to each of the sub-documents; For each sub-document to which the first information extraction result belongs, segmenting the target document according to the preset scaling factor to obtain sub-target documents; The sub-document is updated based on the sub-target document, and the step of returning to the step of executing the information extraction directed graph based on the information extraction, performing information extraction operations on each of the sub-documents respectively, and obtaining a first information extraction result corresponding to each of the sub-documents, until the byte length corresponding to the sub-target document is less than or equal to a preset byte length; Based on all the first information extraction results, the information extraction result is obtained.

6. The information extraction method according to claim 5, characterized in that: The step of obtaining the information extraction result based on all the first information extraction results comprises: Performing average weighting on all the first information extraction results to obtain weighted scores corresponding to each extracted information in all the first information extraction results; Based on the weighted scores, all the extracted information is screened and filtered to obtain the information extraction result.

7. The information extraction method according to any one of claims 1 to 6, characterized in that: The step of obtaining the document to be extracted comprises: Obtaining an initial document; wherein the initial document is an original text from which information extraction is required; When the number of bytes of the initial document exceeds the preset number of bytes, the initial document is segmented to obtain a plurality of the documents to be extracted; wherein the number of bytes of each of the documents to be extracted does not exceed the preset number of bytes; When the number of bytes of the initial document does not exceed a preset number of bytes, the initial document is determined as the document to be extracted.

8. An information extraction device, characterized in that: The device comprises: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the information extraction method according to any one of claims 1 to 7.

9. A storage medium, characterized in that: The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, the steps of the information extraction method according to any one of claims 1 to 7 are implemented.

10. A computer program product, characterized in that The computer program product comprises a computer program, and when the computer program is executed by a processor, the steps of the information extraction method according to any one of claims 1 to 7 are implemented.

Images