Entity linking method based on bidirectional evidence checking and lateral tuning

Abstract

The present application relates to a kind of entity linking method based on two-way evidence verification and lateral optimization, specific steps are as follows: step 1, candidate entity retrieval is carried out to input mention and its original context, step 2, enhanced mention context feature and candidate entity description feature are extracted using frozen backbone large language model respectively.Step 3, enhanced mention context feature is used as active detection end, step 4, candidate entity description feature is used as active detection end, enhanced mention context feature is used as verified end, entity side evidence alignment is executed, and entity side evidence matrix is obtained.Step 5, mention side evidence matrix and entity side evidence matrix are respectively pooled, in training stage, based on the correlation score corresponding to candidate entity, multi-classification cross-entropy loss function is constructed, and only subject transformation kernel, reference transformation kernel and evidence reconstruction kernel are updated.In inference stage, according to step 1 to step 5, the correlation score of the highest candidate entity is calculated and output as entity linking result.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing technology, specifically relating to an entity linking method based on bidirectional evidence verification and lateral optimization. Background Technology

[0002] Entity linking (EL) is a fundamental task in natural language processing and knowledge acquisition. Its goal is to map mentions in unstructured text to canonical entities in a knowledge base, thereby establishing a connection between textual information and structured knowledge. Entity linking plays a crucial role in downstream tasks such as information retrieval, question answering systems, and knowledge graph construction. Therefore, establishing accurate and reliable matching relationships between mention contexts and candidate entity descriptions is a key challenge in entity linking.

[0003] Existing entity linking methods can be broadly categorized into traditional feature engineering methods, fine-tuning methods based on pre-trained language models, and prompting or generative methods based on large language models. Early methods primarily relied on manually designed features and heuristic rules, such as surface form, lexical similarity, prior statistical information, and document-level entity consistency, to perform local or global disambiguation of mentions and candidate entities. While these methods have some effectiveness in specific scenarios, they generally suffer from strong reliance on manually designed features, insufficient cross-domain generalization ability, and limited system robustness.

[0004] With the development of pre-trained language models, entity linking is increasingly being modeled as a neural ranking problem. Existing representative methods typically employ a two-stage framework: first, dual encoders are used to recall candidate entities, and then a re-ranking module performs finer-grained matching between the mention context and the candidate entity description. While these methods can improve entity linking accuracy using deep semantic representations, the re-ranking stage usually involves directly concatenating the mention context and candidate entity description for joint encoding, essentially representing a holistic, implicit fusion matching approach. Although this method can obtain a global matching score, it struggles to explicitly distinguish which words, sentences, or descriptive attributes truly support the final linking result. Therefore, when faced with redundant information, local similarity, and noise interference, it is prone to unclear evidence attribution, scoring bias, and incorrect candidate ranking. Furthermore, these methods typically require full parameter fine-tuning of the pre-trained model, resulting in high training and deployment costs and potentially reducing the model's original general capabilities.

[0005] In recent years, large language models (LLMs) have also been applied to entity linking tasks. Related methods typically leverage the zero-shot or few-shot reasoning capabilities of LLMs to directly make candidate judgments based on prompts, or use LLMs to enhance the original context to supplement implicit information and resolve coreference in short texts. These methods offer high flexibility, but also suffer from high inference latency, limited context windows, and sensitivity to prompt design. Furthermore, while enhanced context may provide richer discriminative information, it may also introduce irrelevant details or redundant content, further exacerbating noise propagation. Especially in complex contexts or ambiguous mentions, relying solely on overall generative judgments or one-way similarity matching often lacks fine-grained verification mechanisms to ensure candidate entities truly receive bidirectional supporting evidence, thus easily leading to mismatches and unstable results.

[0006] Therefore, based on existing entity linking methods, this invention proposes an entity linking method based on bidirectional evidence verification and lateral tuning, formulating the entity linking problem as a bidirectional consensus verification process between mentions and candidate entities. This invention, by attaching a lateral network to a frozen backbone LLMs, performs evidence extraction and alignment on both the mention and entity sides. This not only retains the advantages of LLMs in context understanding and semantic representation but also achieves explicit, fine-grained verification of supporting evidence between mentions and candidate entities, thereby reducing the risk of noise-driven mismatches and enhancing the interpretability and stability of entity linking results. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the present invention aims to design an entity linking method based on bidirectional evidence verification and lateral optimization. This method addresses the problem of mismatches that easily occur in existing entity linking methods under conditions of short context, semantic ambiguity, and noise interference, while also ensuring accuracy, stability, and efficient parameter adaptation in entity linking tasks.

[0008] To achieve the above objectives, the present invention provides the following solution: an entity linking method based on bidirectional evidence verification and lateral optimization, the method comprising the following steps:

[0009] Step 1, Candidate Entity Retrieval and Context Enhancement. For input mentions and its original context First, the candidate retrieval module BLINK is used to generate candidates, and the top 10 entities are retained as the candidate entity set. And retrieve each candidate entity from the knowledge base. Corresponding text description Furthermore, a large language model is used to resolve referential inconsistencies and complete implicit ambiguities in the original context, resulting in an enhanced context. This process replaces pronoun references with explicit names and fills in missing contextual cues. The resulting enhanced input pairs are then used in subsequent steps such as bidirectional evidence alignment. .

[0010] Step 2, Context Feature and Entity Description Feature Extraction. This involves extracting the enhanced context features. With candidate entity description The frozen large language model is used for encoding. Specifically, the candidate entity descriptions are... The data is divided into sentence sequences, and each sentence is input into the frozen large language model. The hidden state of the last word in the last layer is taken as the sentence-level representation; the enhanced mention context is then used. By inputting the same large language model, we obtain contextual feature representations. This yields the enhanced contextual feature matrix and the candidate entity description feature matrix:

[0011]

[0012]

[0013] Where D represents the hidden layer size of the large open-source model, Indicates the number of lexical units that refer to the context. Or the number of sentences in the entity description . The context of mentions, including explanatory references and potential amplification noise, was encoded. Then the candidate entities were captured. These matrices extract sentence-level information. They abstract natural language into feature representations while maintaining the decoupling between the context of mentions and entity descriptions during the feature extraction stage.

[0014] Step 3, mention-side evidence alignment. The enhanced mention context features are used as the active probe, and the candidate entity description features are used as the verification target. This is achieved through a subject transformation kernel. and reference transformation kernel The features from both sides are mapped into a common low-dimensional verification subspace to construct the mention-side detection signal. The mention detection signal can be represented as:

[0015]

[0016] in, This represents the enhanced mention context feature matrix. This represents the feature matrix describing the candidate entities. Represents the main transformation kernel. Represents the reference transformation kernel; both belong to ; Represents the real number field. Let represent a matrix space consisting of D rows and K columns of real numbers, where D represents the dimension of the output features of the large language model and K represents the dimension of the low-dimensional validation subspace. express for OK, A matrix of real numbers in columns; represents the length of the feature sequence, and T represents the matrix transpose.

[0017] Subsequently, the evidence-bearing attributes in the candidate entity description are weighted according to the detection signal, and further processed through evidence reconstruction kernel. The supporting content in the candidate entity descriptions is aggregated to obtain the mention-side evidence matrix. This can be represented as:

[0018]

[0019] in, This represents the normalization function, used to generate weights for each feature position in the candidate entity description based on the mention-side detection signal. This represents the scaling factor used to scale the probe signal matrix. Scaling is performed.

[0020] Step 4, entity-side evidence alignment. The roles of the active probe and the verified entity are swapped. The candidate entity description features are used as the active probe, and the enhanced mention context features are used as the verified entity. The same subject transformation kernel as in Step 3 is applied. Reference transformation kernel and evidence reconstruction core Perform reverse evidence retrieval and aggregation to obtain the entity-side evidence matrix. The entire process is illustrated in the attached instruction manual. Figure 2 As shown. The entity-side evidence matrix is ​​used to characterize the degree of support of the mention context for the candidate entity, thereby relating it to the mention-side evidence matrix. Together they constitute the result of two-way evidence verification.

[0021] Step 5: Relevance scoring and result determination. This involves analyzing the evidence matrix related to the mention side. and entity-side evidence matrix Perform pooling operations separately to obtain the corresponding evidence vectors. and Then, candidate entities are calculated based on the similarity between the evidence vectors. With mention Correlation score between A candidate entity's relevance score is only high when both bidirectional evidence retrievals extract strong and mutually compatible supporting information. Finally, the candidate entity set is evaluated based on these relevance scores. The entities are sorted, and the candidate entities with the highest relevance scores in the sorting results are determined as the final entity linking results.

[0022] During the training phase, based on the candidate entity set containing real entities, the relevance score corresponding to each candidate entity is first calculated according to steps 1 to 5; then, a multi-class cross-entropy loss function is constructed based on the relevance scores, and the lateral network is trained. The loss function can be expressed as:

[0023]

[0024] in, Represents the set of candidate entities The real entities in the middle, Indicates candidate entities With mention The correlation score between them. During training, the parameters of the large language model remain frozen, and only the subject transformation kernel is updated. Reference transformation kernel and evidence reconstruction core This allows for efficient parameter adaptation for entity linking tasks while maintaining the general language knowledge of the large model.

[0025] During the inference phase, for the mentions to be linked, the relevance score of each candidate entity is calculated according to steps 1 to 5. That is, candidate entity retrieval and context enhancement are performed first, then feature extraction, bidirectional evidence alignment, and relevance scoring are performed on each candidate entity, and the candidate entity with the highest relevance score is determined as the final entity linking result.

[0026] This invention achieves mutual verification between mentions and candidate entities through a two-way evidence verification mechanism, and combined with an efficient parameter tuning method, improves the reliability and applicability of entity linking tasks while ensuring the generality of large language models.

[0027] Compared with the prior art, the advantages of the present invention are as follows:

[0028] 1. This solution proposes an explicit alignment method based on bidirectional evidence verification for the matching process between mention context and candidate entity descriptions in entity linking tasks. For entity linking tasks, identifying the key information that truly supports the linking result from mention context and candidate entity descriptions is a crucial step affecting entity linking performance. Traditional methods typically concatenate mention context and candidate entity descriptions directly and then perform overall matching or reordering. This approach is essentially implicit fusion, making it difficult to explicitly highlight the words, sentences, and descriptive attributes that truly determine the linking result. Furthermore, it is susceptible to interference from redundant information and noisy content, thus affecting the interpretability and stability of the entity linking result. In contrast, this invention models the enhanced mention context and candidate entity descriptions separately and performs evidence extraction and alignment on both the mention and entity sides. This allows for the explicit identification and highlighting of key information that supports the entity linking result, thereby improving the interpretability of the entity linking process and enhancing its discriminative ability under complex contexts and noisy conditions.

[0029] 2. This invention introduces the general language understanding capabilities of large language models into entity linking tasks by attaching a lateral network to a frozen backbone large language model. Traditional methods typically rely on specially trained matching models or direct sorting of the entire input, making it difficult to fully utilize the advantages of large language models in context understanding and semantic representation, and often requiring updates to numerous parameters during task adaptation. In contrast, this invention, by attaching a lightweight lateral network to a frozen backbone large language model, fully leverages the capabilities of large language models in context modeling and semantic representation, while training only the subject transformation kernel, reference transformation kernel, and evidence reconstruction kernel, achieving efficient parameter adaptation for entity linking tasks. This reduces the number of model parameters that need to be updated, lowers the computational resource consumption and memory usage during training, and improves the applicability of the model when transferring to different knowledge bases or different text domains.

[0030] 3. This invention improves the accuracy of entity linking in scenarios involving short texts, ambiguous referents, and similar candidate entity descriptions. In actual text processing, the context in which a link is mentioned may contain ambiguous referents, omitted words, abbreviations, or missing background information. Multiple candidate entities may also have similar names or descriptions. By enhancing the implicit semantics through contextual augmentation and combining a two-way evidence verification mechanism to finely differentiate candidate entities, this invention can more accurately identify the contextual evidence and entity description evidence that truly support the link relationship, thereby reducing erroneous links caused by polysemy, similar names, and similar descriptions.

[0031] 4. In the inference stage, this invention calculates and sorts relevance scores for candidate entities, enabling the entity linking results to be output in score form. These scores can be used not only to determine the final linked entities but also to assess the credibility of the linking results in subsequent system iterations. This facilitates quality control before knowledge graph entry, question-and-answer generation, or retrieval result display, reducing the propagation of errors caused by low-confidence linking results to downstream applications. Attached Figure Description

[0032] Figure 1 This is a flowchart of the entity linking method based on two-way evidence verification and lateral optimization of the present invention.

[0033] Figure 2 This invention presents an entity linking framework based on bidirectional evidence verification and lateral optimization. Detailed Implementation

[0034] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention, and should not be construed as limiting the scope of protection of this application.

[0035] Example 1: This invention proposes an entity linking method based on bidirectional evidence alignment and lateral tuning. This method is geared towards entity linking tasks, used to map mentions in text context to canonical entities in a knowledge base. For example... Figure 1 As shown, the main components include candidate entity retrieval and context enhancement, context and entity description feature extraction, bidirectional evidence alignment, relevance scoring, training, and inference. First, for the input mention and its original context, a candidate retrieval module obtains a predetermined number of candidate entities, and a large language model is used to enhance the original context to resolve coreference and supplement implicit background information. Then, a frozen backbone large language model is used to extract features from the enhanced context and candidate entity descriptions, obtaining context feature representations and entity description feature representations. Next, a lateral network attached to the frozen backbone large language model performs bidirectional evidence alignment on both the mention and entity sides, extracting bidirectional supporting evidence, and calculating the relevance score between the candidate entity and the mention based on the bidirectional supporting evidence. Finally, during the training phase, the lateral network is optimized using the relevance score; during the inference phase, candidate entities are ranked according to the relevance score, and the highest-ranked candidate entity is output as the entity link result. The process includes the following steps:

[0036] Step 1, Candidate Entity Retrieval and Context Enhancement. For input mentions and its original context First, the candidate retrieval module BLINK is used to generate candidates, and the top 10 entities are retained as the candidate entity set. And retrieve each candidate entity from the knowledge base. Corresponding text description Furthermore, a large language model is used to resolve referential inconsistencies and complete implicit ambiguities in the original context, resulting in an enhanced context. This process replaces pronoun references with explicit names and fills in missing contextual cues. The resulting enhanced input pairs are then used in subsequent steps such as bidirectional evidence alignment. .

[0037] Step 2, Context Feature and Entity Description Feature Extraction. This involves extracting the enhanced context features. With candidate entity description The frozen large language model is used for encoding. Specifically, the candidate entity descriptions are... The data is divided into sentence sequences, and each sentence is input into the frozen large language model. The hidden state of the last word in the last layer is taken as the sentence-level representation; the enhanced mention context is then used. By inputting the same large language model, we obtain contextual feature representations. This yields the enhanced contextual feature matrix and the candidate entity description feature matrix:

[0038]

[0039]

[0040] Where D represents the hidden layer size of the large open-source model, Indicates the number of lexical units that refer to the context. Or the number of sentences in the entity description . The context of mentions, including explanatory references and potential amplification noise, was encoded. Then the candidate entities were captured. These matrices extract sentence-level information. They abstract natural language into feature representations while maintaining the decoupling between the context of mentions and entity descriptions during the feature extraction stage.

[0041] Step 3, mention-side evidence alignment. The enhanced mention context features are used as the active probe, and the candidate entity description features are used as the verification target. This is achieved through a subject transformation kernel. and reference transformation kernel The features from both sides are mapped into a common low-dimensional verification subspace to construct the mention-side detection signal. The mention detection signal can be represented as:

[0042]

[0043] in, This represents the enhanced mention context feature matrix. This represents the feature matrix describing the candidate entities. Represents the main transformation kernel. Represents the reference transformation kernel; both belong to ; Represents the real number field. Let represent a matrix space consisting of D rows and K columns of real numbers, where D represents the dimension of the output features of the large language model and K represents the dimension of the low-dimensional validation subspace. express for OK, A matrix of real numbers in columns; represents the length of the feature sequence, and T represents the matrix transpose.

[0044] Subsequently, the evidence-bearing attributes in the candidate entity description are weighted according to the detection signal, and further processed through evidence reconstruction kernel. The supporting content in the candidate entity descriptions is aggregated to obtain the mention-side evidence matrix. This can be represented as:

[0045]

[0046] in, This represents the normalization function, used to generate weights for each feature position in the candidate entity description based on the mention-side detection signal. This represents the scaling factor used to scale the probe signal matrix. Scaling is performed.

[0047] Step 4, entity-side evidence alignment. The roles of the active probe and the verified entity are swapped. The candidate entity description features are used as the active probe, and the enhanced mention context features are used as the verified entity. The same subject transformation kernel as in Step 3 is applied. Reference transformation kernel and evidence reconstruction core Perform reverse evidence retrieval and aggregation to obtain the entity-side evidence matrix. The entire process is illustrated in the attached instruction manual. Figure 2 As shown. The entity-side evidence matrix is ​​used to characterize the degree of support of the mention context for the candidate entity, thereby relating it to the mention-side evidence matrix. Together they constitute the result of two-way evidence verification.

[0048] Step 5: Relevance scoring and result determination. This involves analyzing the evidence matrix related to the mention side. and entity-side evidence matrix Perform pooling operations separately to obtain the corresponding evidence vectors. and Then, candidate entities are calculated based on the similarity between the evidence vectors. With mention Correlation score between A candidate entity's relevance score is only high when both bidirectional evidence retrievals extract strong and mutually compatible supporting information. Finally, the candidate entity set is evaluated based on these relevance scores. The entities are sorted, and the candidate entities with the highest relevance scores in the sorting results are determined as the final entity linking results.

[0049] During the training phase, based on the candidate entity set containing real entities, the relevance score corresponding to each candidate entity is first calculated according to steps 1 to 5; then, a multi-class cross-entropy loss function is constructed based on the relevance scores, and the lateral network is trained. The loss function can be expressed as:

[0050]

[0051] in, Represents the set of candidate entities The real entities in the middle, Indicates candidate entities With mention The correlation score between them. During training, the parameters of the large language model remain frozen, and only the subject transformation kernel is updated. Reference transformation kernel and evidence reconstruction core This allows for efficient parameter adaptation for entity linking tasks while maintaining the general language knowledge of the large model.

[0052] During the inference phase, for the mentions to be linked, the relevance score of each candidate entity is calculated according to steps 1 to 5. That is, candidate entity retrieval and context enhancement are performed first, then feature extraction, bidirectional evidence alignment, and relevance scoring are performed on each candidate entity, and the candidate entity with the highest relevance score is determined as the final entity linking result.

[0053] This invention achieves mutual verification between mentions and candidate entities through a two-way evidence verification mechanism, and combined with an efficient parameter tuning method, improves the reliability and applicability of entity linking tasks while ensuring the generality of large language models.

[0054] Example 2: Method Implementation under Different Backbone Models. In another embodiment, the candidate entity retrieval, context enhancement, mention-side evidence alignment, entity-side evidence alignment, and relevance scoring processes remain unchanged. Only the frozen backbone large language model in step 2 is replaced with a large language model fine-tuned by other instructions. The replacement does not change the structure and training method of the lateral network. During training, the backbone model parameters are kept frozen, and only the subject transformation kernel is updated. Reference transformation kernel and evidence reconstruction core To complete the entity linking task.

[0055] This invention, while maintaining the candidate entity retrieval, context enhancement, bidirectional evidence alignment, and relevance scoring processes unchanged, can adapt to different types of instruction-based fine-tuning of large language models. Furthermore, by training only the subject transformation kernel, reference transformation kernel, and evidence reconstruction kernel in the lateral network, it achieves unified adaptation to different backbone models. This demonstrates that this invention does not rely on a single backbone model, possesses excellent model compatibility and transferability, and is of significant importance for deploying large language models in different entity linking scenarios.

[0056] Example 3: Method Implementation in Noisy Context. When the input enhanced context contains irrelevant information or redundant supplementary content, steps 1 to 5 are still performed to perform candidate entity retrieval, feature extraction, mention-side evidence alignment, entity-side evidence alignment, and relevance scoring. Among them, through the bidirectional verification mechanism of mention-side and entity-side, noisy content that cannot obtain bidirectional support is suppressed, and entity link results are output based on the candidate entity with the highest relevance score.

[0057] Even when the enhanced context contains irrelevant information or redundant supplementary content, this invention can still suppress noisy content that cannot obtain reciprocal support through a two-way evidence verification mechanism between the mention side and the entity side. This reduces the risk of noise-driven mismatches and improves the stability and reliability of entity linking results. Therefore, this invention is of great significance in entity linking tasks under complex contexts, information redundancy, or noise interference.

[0058] The applicant of this invention has provided a detailed description of the embodiments of the invention in conjunction with the accompanying drawings. However, those skilled in the art should understand that the above embodiments are merely preferred embodiments of the invention. The detailed description is only intended to help readers better understand the spirit of the invention and is not intended to limit the scope of protection of the invention. On the contrary, any improvements or modifications made based on the inventive spirit of the invention should fall within the scope of protection of the invention.

Claims

1. An entity linking method based on two-way evidence verification and lateral optimization, characterized in that, The method includes the following steps: Step 1: Perform candidate entity retrieval on the input mention and its original context, and enhance the original context to obtain the enhanced mention context and candidate entity set. Step 2: Using the frozen backbone large language model, extract the enhanced mention context features and candidate entity description features respectively. Step 3: Using the enhanced mention context features as the active probe and the candidate entity description features as the verified entity, perform mention-side evidence alignment to obtain the mention-side evidence matrix. Step 4: Using the candidate entity description features as the active probe and the enhanced mention context features as the verified entity, perform entity-side evidence alignment to obtain the entity-side evidence matrix. Step 5: Pool the mention-side evidence matrix and the entity-side evidence matrix to obtain corresponding evidence vectors. Calculate the relevance score based on the evidence vectors and select the candidate entity with the highest score as the final entity linking result. During the training phase, a multi-class cross-entropy loss function is constructed based on the relevance scores corresponding to candidate entities, and only the subject transformation kernel, reference transformation kernel, and evidence reconstruction kernel are updated. During the inference phase, the linked mentions are calculated according to steps 1 to 5, and the candidate entity with the highest relevance score is output as the entity link result.

2. The entity linking method based on bidirectional evidence verification and lateral optimization according to claim 1, characterized in that, Step 1 is as follows: For input mention and its original context First, the candidate retrieval module BLINK is used to generate candidates, and the top 10 entities are retained as the candidate entity set. And retrieve each candidate entity from the knowledge base. Corresponding text description By using a large language model to resolve referential inconsistencies and complete implicit ambiguities in the original context, an enhanced context is obtained. This process replaces pronoun references with explicit names and fills in missing contextual cues, resulting in enhanced input pairs used in subsequent steps such as bidirectional evidence alignment. .

3. The entity linking method based on bidirectional evidence verification and lateral optimization according to claim 1, characterized in that, Step 2 is as follows: For the enhanced context With candidate entity description The frozen large language model is used for encoding, where candidate entity descriptions are... The data is divided into sentence sequences, and each sentence is input into the frozen large language model. The hidden state of the last word (token) in the last layer is taken as the sentence-level representation; the enhanced mention context is then used. By inputting the same large language model, we obtain contextual feature representations, thereby deriving the enhanced contextual feature matrix and the candidate entity description feature matrix: Where D represents the size of the hidden layers in the backbone model. Indicates the number of context terms mentioned. Or the number of sentences in the entity description , The context of mentions, including explanatory references and potential amplification noise, was encoded. Then the candidate entities were captured. Sentence-level information.

4. The entity linking method based on bidirectional evidence verification and lateral optimization according to claim 1, characterized in that, Step 3 is as follows: The enhanced mention context features are used as the active probe, and the candidate entity description features are used as the verification probe. This is achieved through a subject transformation kernel. and reference transformation kernel The features from both sides are mapped into a common low-dimensional verification subspace to construct the mention-side detection signal, and the mention detection signal matrix. Represented as: in, This represents the enhanced mention context feature matrix. This represents the feature matrix describing the candidate entities. Represents the main transformation kernel. Represents the reference transformation kernel; both belong to ; Represents the real number field. Let represent a matrix space consisting of D rows and K columns of real numbers, where D represents the dimension of the output features of the large language model and K represents the dimension of the low-dimensional validation subspace. express for OK, A matrix of real numbers in columns; The length of the feature sequence is represented by T, and T represents the matrix transpose. Subsequently, the evidence-bearing attributes in the candidate entity descriptions are weighted based on the detection signals, and further evidence reconstruction is performed. Description of candidate entities The supporting content in the data is aggregated to obtain the mention-side evidence matrix. , represented as: in, This represents the normalization function, used to generate weights for each feature position in the candidate entity description based on the mention-side detection signal. This represents the scaling factor used to scale the probe signal matrix. Scaling is performed.

5. The entity linking method based on bidirectional evidence verification and lateral optimization according to claim 1, characterized in that, Step 4 is as follows: The roles of the active probe and the verified entity are swapped. The candidate entity description features are used as the active probe, and the enhanced mention context features are used as the verified entity. The same subject transformation kernel as in step 3 is used. Reference transformation kernel and evidence reconstruction core Perform reverse evidence retrieval and aggregation to obtain the entity-side evidence matrix. The entity-side evidence matrix is ​​used to characterize the degree of support that the mention context provides for the candidate entity, thus relating it to the mention-side evidence matrix. Together they constitute the result of two-way evidence verification.

6. The entity linking method based on bidirectional evidence verification and lateral optimization according to claim 1, characterized in that, Step 5 is as follows: Evidence matrix for mention and entity-side evidence matrix Perform pooling operations separately to obtain the corresponding evidence vectors. and Then, candidate entities are calculated based on the similarity between the evidence vectors. With mention Correlation score between Only when both bidirectional evidence retrievals extract strong and mutually compatible supporting information will the candidate entity's relevance score be high. Finally, the candidate entity set is ranked based on its relevance score. The entities are sorted, and the candidate entities with the highest relevance scores in the sorting results are determined as the final entity linking results.

7. The entity linking method based on bidirectional evidence verification and lateral optimization according to claim 1, characterized in that, The training phase is as follows: Based on a candidate entity set containing real entities, the relevance score for each candidate entity is first calculated according to steps 1 to 5. Then, a multi-class cross-entropy loss function is constructed based on the relevance scores, and the lateral network is trained. The loss function is expressed as: in, Represents the set of candidate entities The real entities in the middle, Indicates candidate entities With mention The correlation scores between them are maintained during training, with the parameters of the large language model frozen and only the subject transformation kernel updated. Reference transformation kernel and evidence reconstruction core This allows for efficient parameter adaptation for entity linking tasks while maintaining the general language knowledge of the large model.

8. The entity linking method based on bidirectional evidence verification and lateral optimization according to claim 1, characterized in that, The reasoning stage is as follows: The process involves obtaining the mention to be linked and its original context, generating a set of candidate entities corresponding to the mention based on the candidate entity retrieval model, enhancing the original context to obtain the enhanced mention context, and extracting the enhanced mention context feature matrix and the candidate entity description feature matrix of each candidate entity using the frozen backbone large language model. The process also involves calling the subject transformation kernel, reference transformation kernel, and evidence reconstruction kernel obtained through lateral tuning training, and performing mention-side evidence verification and entity-side evidence verification on each candidate entity without updating the backbone large language model parameters or the transformation kernel parameters. This yields a relevance score between each candidate entity and the mention to be linked, and the candidate entities are sorted according to their relevance scores. The candidate entity with the highest relevance score is then determined as the final entity linking result.

9. An electronic device 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 program, it implements the entity linking method based on bidirectional evidence verification and lateral tuning as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, the computer instructions implement the entity linking method based on bidirectional evidence verification and lateral tuning as described in any one of claims 1-8.

Images