An entity pair guided scientific literature document level relation extraction method and system

Abstract

The application provides a kind of entity pair guided scientific and technical literature document level relationship extraction method and system, its method includes: the entity identification of input scientific and technical literature document is obtained in scientific and technical literature document entity set;Based on the entity pair pre-screening mechanism of multiple sampling and similarity verification, from all possible entity pairs of entity set, candidate entity pair set is screened out;Then generate enhanced relationship description that integrates corresponding entity type information and relationship semantics between entity pairs;Based on the pre-constructed relationship semantic knowledge base, a double-layer filtering mechanism is used to retrieve the corresponding fine-screening candidate relationship set for each enhanced relationship description;Then use the detailed semantic description of candidate relationship to guide the large language model to make triple fact judgment, and get the output result.The application significantly reduces the computing overhead of long text processing while ensuring high-precision relationship extraction, making scientific and technical literature document level relationship extraction more accurate and efficient.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to a method and system for extracting document-level relationships from scientific and technological literature guided by entity pairs. Background Technology

[0002] Entity relation extraction from scientific and technological literature is a crucial task in natural language processing and knowledge graph construction, with wide applications in fields such as scientific and technological intelligence analysis, literature knowledge mining, and academic recommendation systems. With the rapid development of scientific research, the number of scientific and technological documents is growing exponentially. How to automatically extract structured entity relation information from massive amounts of literature has become a current research hotspot and challenge.

[0003] Traditional relation extraction methods primarily target short texts at the sentence level, employing rule-based, statistical, or deep learning-based approaches. However, scientific and technical documents typically contain multiple paragraphs, and identifying relationships between entities often requires reasoning across sentences or even paragraphs, posing a significant challenge to traditional methods.

[0004] In recent years, document-level relation extraction has gradually become a research focus. Early document-level methods were mainly based on graph neural networks (GNNs), which modeled the dependencies between entities and sentences by constructing the graph structure of the document. For example, the EoG method in 2019 adopted an edge-guided graph structure, the GAIN method in 2020 designed a dual-graph reasoning mechanism, and the SIRE method in 2021 separated the reasoning processes within and between sentences. These methods have improved the performance of document-level relation extraction to some extent, but they still have problems such as difficulty in predefining graph structures and insufficient capture of long-distance dependencies when dealing with complex scientific and technical documents.

[0005] Subsequently, methods based on pre-trained language models gradually emerged. For example, the ATLOP method in 2021 employed adaptive thresholding and local context pooling techniques, the EIDER method in 2022 introduced an efficient mechanism for fusing evidence extraction and inference stages, and the DREEAM method in 2023 utilized an evidence-guided attention mechanism to improve relation extraction. These methods leveraged the powerful semantic understanding capabilities of the Transformer architecture and achieved significant progress in document-level relation extraction tasks, but they are still limited by the parameter size of the pre-trained model and the amount of training data.

[0006] Recently, Large Language Models (LLMs) have demonstrated groundbreaking capabilities in the field of natural language processing, bringing new research directions to document-level relation extraction. Existing relation extraction methods based on Large Language Models mainly fall into two categories:

[0007] (1) Non-fine-tuning methods. For example, the PromptRE method in 2023 combines prompting techniques and data programming to enhance relation extraction capabilities through multiple weakly supervised sources; the DocGNRE method in 2023 integrates a large language model with a natural language inference module to enhance document-level relation extraction datasets. These non-fine-tuning methods do not require task-specific fine-tuning of the large language model, but their performance is often limited by the prompting design and the quality of the weakly supervised data.

[0008] (2) Fine-tuning methods. The LMRC method in 2024 incorporates relation set and entity pair information into the prompts, and improves relation extraction performance through fine-tuning. The AutoRE method in 2024 proposes a three-stage processing paradigm: first, it filters possible relation types from the entire document as candidate relations; then, it identifies the head entities in the document; and finally, it extracts triple facts from the head entities and candidate relations.

[0009] The main features of the AutoRE method include: First, it employs a document-level candidate relation filtering strategy, filtering candidate relations based on the overall document content and outputting a list of possible relation types after analyzing the document using a large language model. The advantage of this approach is that it can obtain all potential relations in the document at once. Second, it introduces a head entity recognition step, identifying head entities that may serve as relation subjects after obtaining candidate relations, which helps narrow down the search scope. Third, in the final stage, it performs relation judgment on head entities and candidate relations, including relation descriptions in the prompts to guide the large language model in its judgment. Although AutoRE achieves state-of-the-art performance among large language model-based methods, its F1 score on the DocRED dataset is approximately 47.15, and on the Re-DocRED dataset, it is approximately 60.17, still lagging behind the best performance of traditional small models.

[0010] However, existing technologies still have the following problems:

[0011] First, there is insufficient relevance between candidate relations and entity pairs.

[0012] Second, the problem of unrelated data imbalance has not been effectively solved. In scientific literature datasets, the proportion of entity pairs that have no relationship is extremely high, and existing methods have failed to effectively alleviate the false positive problem caused by this data imbalance.

[0013] Third, the efficiency of long text processing needs to be improved. Scientific and technological documents often contain lengthy technical descriptions and background information. Existing methods take the entire document as input, which is computationally expensive and makes it difficult to locate key information related to specific entity pairs.

[0014] Fourth, existing methods mainly rely on the pre-trained knowledge of large language models and the contextual information of the current document, failing to fully utilize the entity relationship pattern knowledge accumulated in the training data.

[0015] In summary, existing methods for extracting document-level relationships in scientific and technological literature still suffer from low processing efficiency and low accuracy. Summary of the Invention

[0016] To address the aforementioned problems in the prior art, this invention provides a method and system for extracting document-level relationships in scientific and technological literature guided by entity pairs, making the extraction of document-level relationships in scientific and technological literature more accurate and efficient.

[0017] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0018] In a first aspect, the present invention provides a method for extracting document-level relationships in scientific and technological literature guided by entity pairs, comprising:

[0019] Step S1: Perform entity recognition on the input scientific and technological documents to obtain the entity set in the scientific and technological documents;

[0020] Step S2: Based on the entity pair pre-screening mechanism of multiple sampling and similarity verification, a candidate entity pair set is selected from all possible entity pairs in the entity set.

[0021] Step S3: For each entity pair in the candidate entity pair set, generate an enhanced relation description that integrates the corresponding entity type information and the semantics of the relationship between entity pairs;

[0022] Step S4: Based on the pre-built relational semantic knowledge base, a two-layer filtering mechanism including coarse screening and fine screening is adopted to retrieve the corresponding fine screening candidate relation set for each enhanced relation description;

[0023] Step S5: For each candidate relation in the finely filtered candidate relation set, use the detailed semantic description of the candidate relation to guide the large language model to perform triple fact judgment. If the judgment is true, output a triple containing the head entity, tail entity and relation type.

[0024] The beneficial effects of this invention are as follows: First, it significantly reduces the scale of ineffective computation by pre-screening entity pairs, alleviating data imbalance; second, it generates a refined representation rich in context for each entity pair through semantic enhancement, improving information density and relevance; third, it achieves accurate and efficient matching of candidate relations through a two-layer retrieval based on a historical knowledge base; and finally, it ensures the accuracy of triple extraction through relational semantic guidance. The entire scheme significantly reduces the computational overhead of long text processing while ensuring high-precision relation extraction, making document-level relation extraction in scientific and technological literature more accurate and efficient.

[0025] Optionally, in step S4, a two-layer filtering mechanism including coarse screening and fine screening is adopted to retrieve a corresponding set of finely screened candidate relations for each enhanced relation description, including:

[0026] Using the same sentence converter model Encoder as that used to build the relational semantic knowledge base, the enhanced relational description of the entity pairs to be queried is transformed. Encoded as query vector :

[0027] ;

[0028] Calculate query vector With each vector in the relational semantic knowledge base K cosine similarity According to the cosine similarity Sort the samples in descending order, select the samples corresponding to the k most similar vectors, extract the relation labels corresponding to the k samples, and after deduplication, form a coarsely selected candidate relation set R. coarse ;

[0029] Based on the predefined relation transformation template T rel and entity prior knowledge P know The coarsely screened candidate relation set R coarse Each relation label in the table is converted into a label about the entity pair. The natural language statements form the initial option set Q;

[0030] The scientific and technological document D, the initial option set Q, and the entity pairs are used. and fine screening instructions Input a large language model and add irrelevant options as independent choices to obtain the answer. :

[0031] ;

[0032] The answer Transform into the final candidate relation set R fine .

[0033] Optionally, step S4, based on a pre-built relational semantic knowledge base, includes:

[0034] Obtain each labeled entity pair from the training set T. and its relational tags Generate an enhanced relation description for each labeled entity pair. ;

[0035] Each augmented relation description is encoded into a dense vector of fixed dimensions using a pre-trained sentence converter model Encoder. :

[0036] ;

[0037] Establish a set of mapping relationships to constitute the relational semantic knowledge base K:

[0038] .

[0039] Optionally, step S2 includes:

[0040] Step S21: Define instruction template The large language model is guided to identify possible entity relationship pairs from the scientific and technological document D and the entity set E;

[0041] Step S22: Using the large language model, based on the instruction template The document D and the entity set E are sampled K times independently, and each sampling generates a subset of entity pairs that may contain relationships. :

[0042] ;

[0043] Step S23: For the subset of entity pairs obtained from each sampling... Each entity pair in The sentence embedding model is used to calculate the pairs of real entities in the entity set E. cosine similarity ;

[0044] Step S24: Set a similarity threshold for filtering to obtain a subset of valid entity pairs;

[0045] Step S25: Merge all valid entity pair subsets obtained after K samplings to obtain the final candidate entity pair set.

[0046] Optionally, step S3 includes:

[0047] Step S31: For any entity pair in the candidate entity pair set The large language model is invoked to extract head entities from the aforementioned scientific and technological documents. and entity Entity description information including type and key attributes:

[0048] ;

[0049] ;

[0050] Step S32: Invoke the large language model and summarize the entity pairs based on the scientific and technological document D. Generate relation description text based on the latent relation semantics between them. :

[0051] ;

[0052] Step S33: Concatenate the entity description information with the relationship description text to generate an enhanced relationship description. :

[0053] ;

[0054] In the formula This indicates a text concatenation operation.

[0055] Optionally, step S5 includes:

[0056] Step S51: Construct judgment instructions for each candidate relation r. It explicitly requires large language judgment triples Is it valid?

[0057] Step S52: Introduce a structured relation semantic description for each candidate relation r. The At least include the definition of relation r, entity Type constraints and entities Type constraints;

[0058] Step S53: Set the judgment instruction The relational semantic description And including the scientific and technological document D and entity pairs The test input x is submitted to the large language model for binary judgment, and the output y is either yes or no.

[0059] ;

[0060] Step S54: If the output y is yes, then output the relation triplet. .

[0061] Optionally, step S1 includes:

[0062] Entity recognition is performed on the input scientific and technological document D to obtain the entity set E in the document:

[0063] ;

[0064] In the formula, N is the number of entities contained in the scientific and technological document D. These are the i-th and j-th entities, respectively.

[0065] Optionally, the large language model is optimized using a phased supervised fine-tuning strategy, including:

[0066] For the three tasks of entity pair pre-screening, multiple-choice question-and-answer fine-tuning, and triplet fact judgment, low-rank adaptation technique is used to efficiently fine-tune the parameters.

[0067] Optionally, it also includes:

[0068] During the inference phase, different temperature parameters are used for different tasks.

[0069] In a second aspect, the present invention provides an entity-pair-guided document-level relationship extraction system for scientific and technological documents, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the entity-pair-guided document-level relationship extraction method for scientific and technological documents of the first aspect.

[0070] The technical effects of the entity-pair-guided scientific and technological document-level relationship extraction system provided in the second aspect are described in the relevant description of the entity-pair-guided scientific and technological document-level relationship extraction method provided in the first aspect. Attached Figure Description

[0071] Figure 1 This is a schematic diagram of the main process of an entity pair-guided method for extracting document-level relationships in scientific and technological literature according to an embodiment of the present invention;

[0072] Figure 2 This is a schematic diagram of the structure of a document-level relationship extraction system for scientific and technological literature guided by entity pairs, according to an embodiment of the present invention.

[0073] Explanation of reference numerals in the attached figures:

[0074] 1. A document-level relationship extraction system for scientific and technological documents guided by entity pairs;

[0075] 2. Processor;

[0076] 3. Memory. Detailed Implementation

[0077] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art. Example 1

[0078] Existing technologies for document-level relation extraction based on scientific and technological literature still suffer from low processing efficiency and low accuracy. This embodiment provides a precise and efficient solution for automated knowledge extraction from scientific and technological literature through an innovative entity pair guidance mechanism and a multi-level relation retrieval strategy, effectively solving several key technical challenges faced by traditional methods when processing document-level scientific and technological literature.

[0079] Please refer to the discussion below for details.

[0080] Please refer to Figure 1 A method for extracting document-level relationships in scientific and technological literature guided by entity pairs, comprising:

[0081] Step S1: Perform entity recognition on the input scientific and technological documents to obtain the entity set in the scientific and technological documents.

[0082] In this embodiment, step S1 includes:

[0083] Entity recognition is performed on the input scientific and technological document D to obtain the entity set E in the scientific and technological document:

[0084] ;

[0085] In the formula, N represents the number of entities contained in the scientific and technological document D. These are the i-th and j-th entities, respectively.

[0086] Among them, the scientific and technological document D contains sentences, and each sentence may contain L word units.

[0087] The following explanation uses the first example as an example:

[0088] Consider a scientific document about semiconductor manufacturing processes. The excerpt reads, "Company X has achieved mass production of 7-nanometer technology using extreme ultraviolet (EUV) light sources, based on the FinFET transistor architecture. Advances in photolithography have allowed Moore's Law to continue, but quantum tunneling has become a major physical obstacle to further miniaturizing transistors." This document contains the entity set E = {photolithography, EUV, 7-nanometer process, Company X, FinFET transistor, Moore's Law, and quantum tunneling}, totaling N = 7 entities, which can form C(7,2) = 21 entity pairs.

[0089] Step S2: Based on the entity pair pre-screening mechanism of multiple sampling and similarity verification, candidate entity pair sets are selected from all possible entity pairs in the entity set.

[0090] In this embodiment, to alleviate the data imbalance problem caused by a large number of unrelated entity pairs in scientific and technological literature, the present invention adopts an entity pair pre-screening mechanism based on multiple sampling. Specifically, step S2 includes:

[0091] Step S21: Define instruction template Ψ to guide the large language model to identify possible entity relationship pairs from scientific and technological documents D and entity set E.

[0092] Step S22: Use a large language model based on instruction templates. Perform K independent sampling operations on document D and entity set E, with each sampling generating a subset of entity pairs that may contain relationships. :

[0093] .

[0094] Step S23: For the subset of entity pairs obtained from each sampling... Each entity pair in The sentence embedding model is used to calculate the pairs of real entities in the entity set E. cosine similarity .

[0095] Among them, cosine similarity The function is located as follows:

[0096] ;

[0097] In the formula, u and v are subsets of entity pairs. The entity pairs and their corresponding real entities.

[0098] Therefore, for each sampled entity pair Similarity verification can improve the quality of screening.

[0099] Step S24: Set a similarity threshold for filtering to obtain a subset of valid entity pairs.

[0100] Step S25: Merge all valid entity pair subsets obtained after K samplings to obtain the final candidate entity pair set.

[0101] Therefore, by performing multiple sampling and similarity verifications in step S2, the number of entity pairs that need to be processed is significantly reduced while ensuring recall, effectively alleviating the data imbalance problem.

[0102] The following explanation uses the first example as an example:

[0103] By sampling K=5 times and using a similarity threshold θ=0.88, a set of highly relevant candidate entity pairs was identified. ={(Company X, extreme ultraviolet light source), (Company X, 7nm process), (extreme ultraviolet light source, 7nm process), (7nm process, FinFET transistor), (lithography technology, Moore's Law)}, a total of 6 entity pairs, reducing the subsequent processing scale from 21 to 6, with a compression rate of 71.4%, effectively filtering out low-relevance entity pairs such as (Company X, quantum tunneling effect).

[0104] Step S3: For each entity pair in the candidate entity pair set, generate an enhanced relation description that integrates the corresponding entity type information and the semantics of the relationship between entity pairs.

[0105] In this embodiment, step S3 includes:

[0106] Step S31: For any entity pair in the candidate entity pair set The large language model is invoked to extract head entities from scientific and technological literature documents. and entity Entity description information including type and key attributes:

[0107] ;

[0108] .

[0109] Step S32: Call the large language model and summarize entity pairs based on the scientific and technological document D. Generate relation description text based on the latent relation semantics between them. :

[0110] .

[0111] Step S33: Concatenate the entity description information with the relation description text to generate an enhanced relation description. :

[0112] ;

[0113] In the formula, This indicates a text concatenation operation.

[0114] Therefore, enhanced relation descriptions not only include the semantics of the relationship between entity pairs, but also incorporate entity type information and attribute features, providing richer semantic basis for subsequent candidate relation retrieval. Considering that relation descriptions usually have implicit constraints on entity types, the addition of entity information significantly improves the accuracy of relation matching.

[0115] The following explanation uses the first example as an example:

[0116] For the entity pair (extreme ultraviolet light source, 7nm process), the system first extracts the entity information:

[0117] (Extreme Ultraviolet Light Source) "Extreme ultraviolet light source is a photolithography light source with a wavelength of 13.5 nanometers, and it is a key piece of equipment in advanced semiconductor manufacturing."

[0118] (7nm process) = "7nm process refers to the integrated circuit manufacturing process node with a transistor gate length of approximately 7 nanometers".

[0119] Then extract the relation description:

[0120] (Extreme ultraviolet light source, 7-nanometer process) "Extreme ultraviolet light source is the core technology for achieving mass production of 7-nanometer process. Its short wavelength characteristics enable the lithography resolution to meet the process requirements."

[0121] The final fusion generates an enhanced relation description: "Extreme ultraviolet (EUV) light sources are lithography light sources with a wavelength of 13.5 nanometers, and are key equipment in advanced semiconductor manufacturing. 7-nanometer process refers to the integrated circuit manufacturing process node where the transistor gate length is approximately 7 nanometers. EUV light sources are the core technology for achieving mass production of 7-nanometer processes; their short wavelength characteristics enable lithography resolution to meet process requirements." This description clearly defines entity type constraints and technological dependencies, providing a sufficient semantic foundation for accurately matching relation types such as "key technology" or "enabling technology."

[0122] Step S4: Based on the pre-built relational semantic knowledge base, a two-layer filtering mechanism including coarse screening and fine screening is adopted to retrieve the corresponding fine screening candidate relation set for each enhanced relation description.

[0123] In this embodiment, to fully utilize the relational pattern knowledge accumulated in the training data, the present invention constructs a relational semantic knowledge base based on training samples and employs an innovative two-layer filtering mechanism to retrieve highly relevant candidate relations for each entity pair. Therefore, step S4 includes:

[0124] Step S41: Obtain the entity pair of each labeled relation from the training set T. and its relational tags Generate an enhanced relation description for each labeled entity pair. .

[0125] Step S42: Encode each augmented relation description into a fixed-dimensional dense vector using a pre-trained sentence converter model Encoder. :

[0126] .

[0127] Step S43: Establish a set of mapping relationships to form a relational semantic knowledge base K:

[0128] .

[0129] Step S44: Using the same sentence converter model Encoder as that used to build the relational semantic knowledge base, the enhanced relation description of the entity pair to be queried is transformed. Encoded as query vector :

[0130] .

[0131] Step S45: Calculate the query vector With each vector in the relational semantic knowledge base K cosine similarity According to cosine similarity Sort the samples in descending order, select the samples corresponding to the k most similar vectors, extract the relation labels corresponding to the k samples, and after deduplication, form a coarsely selected candidate relation set R. coarse .

[0132] Therefore, the above steps are arranged in descending order of similarity, and the samples corresponding to the top k most similar vectors are selected. Their corresponding relation labels are then extracted to form a coarsely screened candidate relation set R. coarse This layer of retrieval quickly locates historical cases with similar relationship patterns to the current entity pair, initially filtering out possible relationship type options and improving the relevance of candidate relationships to entity pairs.

[0133] Step S46: Based on the predefined relation transformation template T rel and entity prior knowledge P know The candidate relation set R is coarsely screened. coarse Each relation label in the table is converted into a label about the entity pair. The natural language statements form the initial option set Q.

[0134] In this step, the candidate relations R obtained from the coarse screening are... coarse Convert to natural language option set Q

[0135] Step S47: Combine the scientific and technological literature document D, the initial option set Q, and the entity pairs. and fine screening instructions Input a large language model and add irrelevant options as independent choices to obtain the answer. :

[0136] .

[0137] Step S48: Submit the answer Transform into the final candidate relation set Rfine .

[0138] In cases where the answers are related and there are multiple-choice options, then the answer will be... Transform into the final candidate relation set R fine If the answer is unrelated or the option set Q is empty, then the entity pair is considered to have no relation. .

[0139] Therefore, in this step, the two-layer filtering mechanism combines the efficiency of vector retrieval with the accuracy of multiple-choice question answering, ensuring both retrieval speed and improving the accuracy of candidate relationships.

[0140] The following explanation uses the first example as an example:

[0141] Continuing with the semiconductor manufacturing process example above, for the physical pair (extreme ultraviolet light source, 7nm process), its enhancement relationship is described as follows: Encoded as a 768-dimensional vector Assuming the knowledge base contains vector representations of 15,000 training samples, then the search is performed within the knowledge base K constructed from the training set, and the Top samples are selected after calculating similarity. k =8 most similar cases, and the relationship labels of these cases are as follows: =Key technologies, =Enabling technologies, =Applied to, =depends on, =Implementation method, =Technical foundation, =Manufacturing tools, =Core equipment. After deduplication, a coarsely selected candidate relation set is obtained:

[0142] R coarse ={Key technologies, enabling technologies, applied to, dependent on, implementation methods, technological foundations, manufacturing tools, core equipment}.

[0143] During the fine screening phase, the system converts these 8 candidate relationships into structured options:

[0144] A. Extreme ultraviolet light source is a key technology for 7-nanometer process;

[0145] B. Extreme ultraviolet light source is an enabling technology for 7-nanometer process;

[0146] C. Extreme ultraviolet light source is used in 7-nanometer process;

[0147] The D.7 nanometer process relies on extreme ultraviolet light sources;

[0148] E. Extreme ultraviolet light source is a method for realizing 7-nanometer process;

[0149] F. Extreme ultraviolet light sources are the technological basis for 7-nanometer processes;

[0150] G. Extreme ultraviolet light sources are manufacturing tools for 7-nanometer processes;

[0151] H. Extreme ultraviolet light source is a core device in the 7-nanometer process;

[0152] I. Extreme ultraviolet light sources are unrelated to 7-nanometer processes;

[0153] Combining entity prior knowledge P know "Extreme ultraviolet (EUV) light sources are the core components of lithography equipment; 7-nanometer processes require light sources of specific wavelengths." The large language model analyzes the document evidence "Extreme ultraviolet light sources are the core technology for achieving mass production of 7-nanometer processes" using a multiple-choice question-and-answer mechanism, selecting answer A={A,B,D,F}, and finally obtaining the following refined results:

[0154] R fine ={Key technologies, enabling technologies, dependent on, technological foundation}.

[0155] This fine screening process effectively eliminated irrelevant candidate relationships such as manufacturing tools and implementation methods, reducing the number of candidates from 8 to 4. The semantic relevance between candidate relationships and entity pairs increased from an average similarity of 0.76 in the coarse screening to 0.89 in the fine screening.

[0156] Step S5: For each candidate relation in the finely filtered candidate relation set, use the detailed semantic description of the candidate relation to guide the large language model to make a triplet fact judgment. If the judgment is true, output a triplet containing the head entity, the tail entity, and the relation type.

[0157] In this embodiment, for each candidate relation r∈R obtained through fine screening fine The final triple extraction is performed using a judgment mechanism guided by relational semantic description. A specific judgment instruction is constructed for each candidate relation r. Clearly define the judgment task and output requirements. Introduce a detailed relational semantic description. This helps large language models understand the precise semantics of relations. Relation descriptions include the relation definition, subject-object type constraints, typical examples, and other information. Therefore, step S5 includes:

[0158] Step S51: Construct judgment instructions for each candidate relation r. It explicitly requires large language judgment triples Whether it is valid or not.

[0159] Step S52: Introduce a structured relation semantic description for each candidate relation r. , At least include the definition of relation r, entity Type constraints and entities Type constraints.

[0160] Step S53: Determine the instruction Relational semantic description Δ r And including scientific and technological documents D and entity pairs The test input x is submitted to the large language model for binary judgment, and the output y is either yes or no.

[0161] .

[0162] Step S54: If the output y is yes, then output the relation triplet. .

[0163] If the output y is negative, then no output is made.

[0164] Therefore, in this step, guided by the semantic description of the relation, the model can accurately understand the semantic constraints and applicable conditions of the relation, which significantly improves the accuracy of relation judgment.

[0165] The following explanation uses the first example as an example:

[0166] For the "key technologies" of candidate relationships, the system constructs judgment instructions:

[0167] 关键技术 ="Based on the text content and domain knowledge, determine whether extreme ultraviolet light source is a key technology for 7-nanometer process".

[0168] Introducing a detailed semantic description of relations:

[0169] "The 'key technology' relationship indicates that the subject technology or equipment is an indispensable technical element for the objective process or system to realize its core functions. Constraints: (1) The subject must be a technology or equipment entity; (2) The objective must be a process or system entity; (3) The subject has an irreplaceable supporting role in realizing the core functions of the objective."

[0170] The test input x is constructed by including:

[0171] Document excerpt: "Company X uses extreme ultraviolet light to achieve mass production of 7-nanometer process... Advances in photolithography technology have enabled Moore's Law to continue."

[0172] Entity pair: (Extreme ultraviolet light source, 7nm process);

[0173] Large language models reason based on three aspects of evidence:

[0174] 1. The document explicitly states that "7-nanometer process is achieved using extreme ultraviolet light source," indicating a causal relationship;

[0175] 2. The entity type meets the constraints (extreme ultraviolet light source is the equipment, 7nm process is the process node).

[0176] 3. Domain knowledge indicates that the wavelength of traditional light sources cannot meet the lithography resolution requirements of the 7-nanometer process, making extreme ultraviolet light sources essential.

[0177] If the judgment result y is yes, output the triplet: (Extreme ultraviolet light source, 7 nanometer process, key technology).

[0178] The same process applies to other candidate relationships:

[0179] For the candidate relationship "Enabling Technology", if it is determined to be yes, output the triplet: (Extreme Ultraviolet Light Source, 7nm Process, Enabling Technology);

[0180] For the candidate relation "depends on", since this relation usually describes a reverse dependency, although the semantics are correct, the direction does not match the order of the entity pairs, so it is judged as negative.

[0181] The candidate relationship "Technical Foundation" does not reflect the core dependency relationship, so it is judged as no.

[0182] Ultimately, the entity pair outputs two high-confidence triples.

[0183] In this embodiment, the large language model is optimized using a phased supervised fine-tuning strategy, including:

[0184] For the three tasks of entity pair pre-screening, multiple-choice question-and-answer fine-tuning, and triplet fact judgment, low-rank adaptation technique is used to efficiently fine-tune the parameters.

[0185] Specifically, to improve the performance of large language models in the task of extracting relationships from scientific and technological documents, this invention adopts a phased fine-tuning strategy, performing supervised fine-tuning on three stages: entity pair pre-screening, multiple-choice question answering, and triplet judgment, and using low-rank adaptation techniques for efficient parameter fine-tuning. The key parameters are set as follows: low-rank parameter r... lora =64, scaling factor α lora =128; Dropout rate p drop =0.05; Maximum input length L context =2048; Learning rate The number of training rounds is adjusted to n based on the task complexity. epoch ∈[1,5].

[0186] Temperature parameters during the inference phase are differentiated for different tasks:

[0187] Entity pair pre-screening and relation description generation stages: Increase diversity;

[0188] Multiple choice question and answer stage: Seeking definitive answers;

[0189] Triplet judgment phase: To balance accuracy and flexibility.

[0190] Uniformly set the Top-p sampling parameter p nucleus =0.9, which maintains a certain degree of randomness while ensuring output quality.

[0191] Therefore, the complete method operation flow of this embodiment is as follows:

[0192] Input: Scientific and technical document D and a predefined set of relation types R all ;

[0193] Output: The set T of all relation triples in the scientific and technical document D. output .

[0194] Processing procedure:

[0195] 1. Document preprocessing: Identifying the entity set E in the document;

[0196] 2. Entity Pair Pre-screening: A candidate entity pair set is obtained through K samplings and similarity filtering. ;

[0197] 3. Enhanced relation description generation: For each entity pair ,generate ;

[0198] 4. Two-layer retrieval of candidate relations: Through coarse screening: ... Encoded as The knowledge base K is searched to obtain the Top-k similarity relations, which constitute R. coarse ; Refining R through carefully selected multiple-choice questions and answers coarse , to obtain R fine ;

[0199] 5. Triple fact judgment: For each r∈R fine Perform a relational semantic-guided judgment; if the result is yes, then... Join T output ;

[0200] 6. Output the final set of triples T output .

[0201] Therefore, the complete example operation flow of this embodiment is as follows:

[0202] Given semiconductor manufacturing process literature, the system identifies 7 entities, theoretically forming 21 entity pairs. After pre-screening of entity pairs, 6 highly relevant candidate entity pairs are obtained. Perform the above process for each entity pair, for example:

[0203] Solid pair (extreme ultraviolet light source, 7nm process): outputs 2 triplet sets (key technology, enabling technology);

[0204] Physical pair (7nm process, FinFET transistor): Output 1 triplet (7nm process, FinFET transistor, using architecture);

[0205] Entity pair (photolithography, Moore's Law): Outputs one triple (photolithography, Moore's Law, continuation);

[0206] Entity pair (quantum tunneling effect, FinFET transistor): Outputs a triplet (quantum tunneling effect, FinFET transistor, physical confinement).

[0207] The final system outputs a set of triples:

[0208] T output ={(Extreme Ultraviolet Light Source, 7nm Process, Key Technologies),(Extreme Ultraviolet Light Source, 7nm Process, Enabling Technologies),(7nm Process, FinFET Transistor, Architecture Adopted),(Photolithography Technology, Moore's Law, Continuation),(Quantum Tunneling Effect, FinFET Transistor, Physical Limitations)}.

[0209] In the example above, a hierarchical processing flow guided by entity pairs accurately extracts high-quality relation triples from 21 theoretical entity pairs. Compared to traditional document-level methods, which require processing all 21 entity pairs, retrieving 96 relation types per pair, and making a total of 2016 judgments, this invention only needs to process 6 entity pairs and an average of 4 candidate relations per pair, making a total of 24 judgments, reducing the computational load by 98.8%. At the same time, the accuracy, as shown in the F1 score on the DocRED dataset, is improved from 47.2% for traditional methods to 54.6%. Example 2

[0210] Please refer to Figure 2 A document-level relationship extraction system 1 for entity pairs in scientific and technological literature includes a memory 3, a processor 2, and a computer program stored on the memory 3 and executable on the processor 2. When the processor 2 executes the computer program, it implements the steps in Embodiment 1 above.

[0211] In summary, the present invention has the following advantages:

[0212] First, it significantly improves the accuracy of candidate relationship screening.

[0213] To address the problem of weak correlation between candidate relations and specific entity pairs caused by existing methods that filter all relation types at the document level, this invention proposes an entity pair-guided relation semantic enhancement and hierarchical retrieval mechanism. This mechanism achieves precise matching between candidate relations and specific entity pairs, thereby significantly improving the accuracy of relation extraction.

[0214] Second, it effectively alleviates the problem of data imbalance.

[0215] To address the false positive errors in scientific literature caused by unrelated entity pairs exceeding 95%, this invention introduces a pre-screening mechanism for entity pairs based on multiple sampling. This mechanism pre-identifies entity pairs that may contain relationships before relation extraction, effectively reducing interference from unrelated data and improving the model's discriminative ability.

[0216] Third, it significantly improves the efficiency of long text processing and the ability to locate key information.

[0217] To address the challenges of high computational costs and difficulty in locating key information caused by the length and complexity of scientific and technological literature, this invention compresses long texts into refined representations highly relevant to the entity pairs by generating enhanced descriptive texts that integrate entity type and relational semantics for each entity pair. This method significantly reduces computational overhead while preserving key information.

[0218] Fourth, achieve efficient use of historical knowledge.

[0219] To address the issue of insufficient utilization of relational patterns accumulated in training data by existing methods, this invention constructs a relational semantic knowledge base based on training samples and employs a two-layer filtering mechanism combining vector retrieval and question-answering refinement for candidate relation matching. This design enables efficient reuse and transfer of historical knowledge, improving the model's ability to analyze new documents.

[0220] Fifth, enhance the accuracy of relationship judgment and the depth of semantic understanding.

[0221] To address the misjudgments caused by the diverse types and subtle semantics of relationships in scientific and technological literature, this invention guides a large language model to perform fine-grained relationship judgments by combining detailed semantic descriptions of relationships with prior entity knowledge. This not only improves the accuracy of relationship judgments but also enhances the system's ability to understand complex relational semantics.

[0222] Overall, this invention provides a precise and efficient solution for automated knowledge extraction from scientific and technological documents through entity pair-guided mechanisms and multi-level relationship retrieval strategies, effectively overcoming several key technical bottlenecks in document-level relationship extraction.

[0223] Since the systems / devices described in the above embodiments of the present invention are systems / devices used to implement the methods of the above embodiments of the present invention, those skilled in the art can understand the specific structure and modifications of the systems / devices based on the methods described in the above embodiments of the present invention, and therefore will not be repeated here. All systems / devices used in the methods of the above embodiments of the present invention fall within the scope of protection of the present invention.

[0224] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0225] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions.

[0226] It should be noted that any reference numerals placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In claims that enumerate several means, several of these means may be embodied by the same hardware. The use of the terms first, second, third, etc., is merely for convenience of expression and does not indicate any order. These terms can be understood as part of the component names.

[0227] Furthermore, it should be noted that in the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0228] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the claims should be interpreted to include both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0229] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, then this invention should also include these modifications and variations.

Claims

1. A method for extracting document-level relationships in scientific and technological literature guided by entity pairs, characterized in that, include: Step S1: Perform entity recognition on the input scientific and technological documents to obtain the entity set in the scientific and technological documents; Step S2: Based on the entity pair pre-screening mechanism of multiple sampling and similarity verification, a candidate entity pair set is selected from all possible entity pairs in the entity set. Step S3: For each entity pair in the candidate entity pair set, generate an enhanced relation description that integrates the corresponding entity type information and the semantics of the relationship between entity pairs; Step S4: Based on the pre-built relational semantic knowledge base, a two-layer filtering mechanism including coarse screening and fine screening is adopted to retrieve the corresponding fine screening candidate relation set for each enhanced relation description; Step S5: For each candidate relation in the finely filtered candidate relation set, use the detailed semantic description of the candidate relation to guide the large language model to make a triplet fact judgment. If the judgment is true, output a triplet containing the head entity, the tail entity and the relation type. Step S2 includes: Step S21: Define instruction template Ψ to guide the large language model to identify possible entity relationship pairs from the scientific and technological document D and the entity set E; Step S22: Using the large language model, based on the instruction template Ψ, perform K independent sampling operations on the document D and the entity set E. Each sampling operation generates a subset Ω of entity pairs that may contain relationships. i : Oh i =LLM(Ψ,D,E); Step S23: For the entity pair subset Ω obtained from each sampling... i Each entity pair (E) p E q Using a sentence embedding model, it calculates the corresponding real entity pair (E) in the entity set E. p ',E q The cosine similarity σ(u,v) of '). Step S24: Set a similarity threshold for filtering to obtain a subset of valid entity pairs; Step S25: Merge all valid entity pair subsets obtained after K samplings to obtain the final candidate entity pair set; In step S4, a two-layer filtering mechanism including coarse screening and fine screening is used to retrieve a corresponding set of finely screened candidate relations for each enhanced relation description, including: Using the same sentence converter model Encoder as that used to construct the relational semantic knowledge base, the enhanced relational description Γ*(E) of the entity pair to be queried is transformed. p E q Encode as query vector V q : V q =Encoder(Γ*(E p ,E q )); Calculate query vector V q Each vector V in the relational semantic knowledge base K T cosine similarity σ(V) q V T According to the cosine similarity σ(V) q V T Sort the samples in descending order, select the samples corresponding to the k most similar vectors, extract the relation labels corresponding to the k samples, and after deduplication, form a coarsely screened candidate relation set R. coarse ; Based on the predefined relation transformation template T rel and entity prior knowledge P know The coarsely screened candidate relation set R coarse Each relation label in the table is converted to a relation about the entity pair (E). p E q The natural language statements are used to form the initial option set Q; The scientific and technological document D, the initial option set Q, and the entity pair (E) are used. p E q ) and fine screening instructions Ψ refine Inputting a large language model and adding unrelated options as independent choices yields answer A(E). p E q ): A(E p ,E q )=LLM(Ψ refine ,D,Q,E p ,E q ); The answer A(E) p E q Transform into the final candidate relation set R fine .

2. The method for extracting document-level relationships in scientific and technological literature guided by entity pairs according to claim 1, characterized in that, In step S4, based on a pre-built relational semantic knowledge base, the following is included: Obtain each labeled entity pair from the training set T. ) and its relational label Λ( ), generating an enhanced relation description Γ*( for each labeled entity pair. ); Each augmented relation description is encoded into a dense vector V of fixed dimensions using a pre-trained sentence converter model Encoder. ): V( )=Encoder(Γ*( )); Establish a set of mapping relationships to constitute the relational semantic knowledge base K: K={V( ):Λ( )}。 3. The method for extracting document-level relationships in scientific and technological literature guided by entity pairs according to claim 1, characterized in that, Step S3 includes: Step S31: For any entity pair (E) in the candidate entity pair set p E q The large language model is invoked to extract the head entity E from the aforementioned scientific and technological documents. p and entity E q Entity description information including type and key attributes: Φ(E p )=LLM(E p ,D); Φ(E q )=LLM(E q ,D); Step S32: Invoke the large language model and, based on the scientific and technological document D, summarize the entity pairs (E). p E q The latent relational semantics between Γ and Γ generate relational description text. (p,q) : Γ (p,q) =LLM(E p ,E q ,D); Step S33: Concatenate the entity description information with the relation description text to generate an enhanced relation description Γ*(E p E q ): Γ*(E p ,AND q )=Φ(E p )⊕Φ(E q )⊕Γ (p,q) ; In the formula, ⊕ represents the text concatenation operation.

4. The method for extracting document-level relationships in scientific and technological literature guided by entity pairs according to claim 1, characterized in that, Step S5 includes: Step S51: Construct a judgment instruction Ψ for each candidate relation r. r It explicitly requires large language decision triples (E p ,r,E q Is this true? Step S52: Introduce a structured relation semantic description Δ for each candidate relation r. r The Δ r At least include the definition of relation r, entity E p Type constraints and entity E q Type constraints; Step S53: Set the judgment instruction Ψ r The relational semantic description Δ r And including the scientific and technological document D and entity pairs (E) p E q The test input x is submitted to the large language model for binary judgment, and the output y is either yes or no. y=LLM(Ψ r ,D r ,x); Step S54: If the output y is yes, then output the relation triple (E). p ,r,E q ).

5. The method for extracting document-level relationships in scientific and technological literature guided by entity pairs according to claim 1, characterized in that, Step S1 includes: Entity recognition is performed on the input scientific and technological document D to obtain the entity set E in the document: Ω={(E i ,AND j )|i≠j,i,j∈{1,...,N}}; In the formula, N is the number of entities contained in the scientific and technological document D, and E i E j These are the i-th and j-th entities, respectively.

6. The method for extracting document-level relationships in scientific and technological literature guided by entity pairs according to claim 1, characterized in that, The large language model is optimized using a phased supervised fine-tuning strategy, including: For the three tasks of entity pair pre-screening, multiple-choice question-and-answer fine-tuning, and triplet fact judgment, low-rank adaptation technique is used to efficiently fine-tune the parameters.

7. The method for extracting document-level relationships in scientific and technological literature guided by entity pairs according to claim 6, characterized in that, The large language model is optimized using a phased supervised fine-tuning strategy, and also includes: During the inference phase, different temperature parameters are used for different tasks.

8. A system for extracting document-level relationships from scientific and technological documents guided by entity pairs, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the entity pair-guided method for extracting document-level relationships in scientific and technological literature as described in any one of claims 1 to 7.

Images