Task requirement parsing method, system and storage medium based on entity linking

Through multi-stage entity linking and unlinkable prediction model, combined with BERT and GRU modules, the problem of insufficient accuracy in the traditional entity linking method in short text is solved, and more efficient entity linking and context understanding is achieved.

CN118468854BActive Publication Date: 2024-11-08NANJING UNIV OF INFORMATION SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410914696.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-09
Publication Date
2024-11-08
Estimated Expiration
2044-07-09

AI Technical Summary

Technical Problem

Traditional entity linking methods lack the interactive information between entity mentions and candidate entities when processing short texts, making lexical, semantic and syntactic relationships difficult to capture and affect accuracy.

Method used

Using multi-stage entity linking and unlinkable prediction model, the BERT model encodes feature representation, combines multi-layer perceptron and full-connection layer for preliminary and accurate prediction, and introduces the description information of linked entities in the knowledge base for re-prediction and global disambiguation. , update query statements using the GRU module to enhance the context.

Benefits of technology

Improves the accuracy and robustness of entity links, is suitable for the analysis of complex task requirements, and enhances the context semantic understanding of entity mentions in short texts.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118468854B_ABST
    Figure CN118468854B_ABST
Patent Text Reader

Abstract

The present invention discloses a task requirement parsing method, system and storage medium based on entity linking. The method comprises generating candidate entities according to a task requirement text and obtaining description information from a knowledge base; constructing a query statement according to the task requirement text, splicing the query statement with the candidate entity and the description information, obtaining a splicing result, performing unlinkable prediction on the splicing result, obtaining an unlinkable preliminary prediction result, and obtaining the matching probability between different linkable entities and their mentions; re-predicting the unlinkable entity mentions in the preliminary prediction result, adding the unlinkable entity mentions in the re-prediction result to the knowledge base, performing global disambiguation on the linkable entity mentions in the preliminary prediction result and the re-prediction result, and then sorting, encoding and globally scoring the linkable entity mentions according to the matching probability between the linkable entity and its mentions to obtain the optimal task requirement and improve the accuracy of entity linking.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a task requirement parsing method, system and storage medium based on entity linking, and belongs to the technical field of artificial intelligence. Background Art

[0002] With the explosive growth of mission elements, the method of relying on traditional commanders' human brain decision-making and template matching is not conducive to the rapid generation of mission requirements. For the task, most of the input text consists of dozens of terms, which makes the surrounding context of entity mentions scarce and concise, without rich contextual semantics. Therefore, the present invention mainly focuses on the problem of short text entity linking. Traditional entity linking methods mainly rely on the encoding vector of the encoder output layer for semantic matching, but lack the interactive information between entity mentions and their candidate entities, which makes it difficult for the model to capture lexical, semantic and syntactic relationships, thereby affecting the accuracy of entity linking. Summary of the invention

[0003] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a task requirement parsing method, system and storage medium based on entity linking to improve the accuracy of entity linking.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is:

[0005] In a first aspect, the present invention provides a task requirement parsing method based on entity linking, comprising:

[0006] Generate candidate entities based on the task requirement text and obtain description information from the knowledge base;

[0007] Construct a query statement according to the task requirement text, and concatenate the query statement with the candidate entity and description information to obtain a concatenation result;

[0008] Performing unlinkable prediction on the concatenated results to obtain unlinkable preliminary prediction results, matching probabilities between different linkable entities and their mentions;

[0009] The unlinkable entity mentions in the preliminary prediction results are re-predicted and added to the knowledge base. After global disambiguation of the linkable entity mentions in the preliminary prediction results and the re-prediction results, the linkable entity mentions are sorted according to the matching probability between the linkable entities and their mentions, the sorted results are encoded to obtain a global score, and the optimal task requirements are obtained based on the global score.

[0010] In combination with the first aspect, further, constructing a query statement according to the task requirement text includes:

[0011] Use reading comprehension to construct query statements based on the task requirements.

[0012] Further, the concatenation result is predicted to be unlinkable, and the unlinkable preliminary prediction result, the matching probability between different linkable entities and their mentions are obtained, including:

[0013] The concatenated result is encoded by the BERT model to obtain a feature representation, wherein the feature representation includes question features, candidate entities and their description features;

[0014] The problem features are used through a multi-layer perceptron to make a preliminary unlinkable prediction and generate a preliminary unlinkable judgment result;

[0015] Question features, candidate entities and their description features are passed through a fully connected layer for accurate unlinkable prediction, and the matching probabilities between different linkable entities and their mentions are obtained.

[0016] Furthermore, the method comprises:

[0017] Treat unlinkable entities and an option marked as NIL as candidate entities;

[0018] Question features, candidate entities and their description features are passed through a fully connected layer for accurate unlinkable prediction, and the matching probabilities between different linkable entities and their mentions are obtained.

[0019] Furthermore, the linkable entity mentions are ranked according to the matching probability between the linkable entities and their mentions, including:

[0020] Predict matching scores for candidate entity mentions based on the matching probabilities between linkable entities and their mentions;

[0021] Calculate the range of matching scores for each candidate entity mention;

[0022] Sort the candidate entity mentions according to their range size.

[0023] Furthermore, the sorted results are encoded to obtain a global score, and the optimal task requirements are obtained based on the global score, including:

[0024] Start disambiguation from the candidate entity mention with the largest range, reduce the interference of noise data, and obtain the disambiguation result encoding;

[0025] The disambiguation result encoding and the BERT encoding obtained after query sentence update are globally scored through the GRU module;

[0026] Get the optimal task requirements based on the global score.

[0027] Furthermore, the unlinkable entity mentions in the preliminary prediction results are re-predicted, including:

[0028] Incorporate the description information of linkable entities into the knowledge base to reconstruct the query statement, and encode the unlinkable entities in the task requirement text through the BERT model to obtain feature representation;

[0029] Regularize the feature representation through the Dropout layer;

[0030] The regularized feature representation is passed through the Dense layer to predict the unlinkable entities and obtain the linkable entities.

[0031] Furthermore, the mission requirements include: target type, mission means, military service, mission domain and mission purpose information.

[0032] In a second aspect, the present invention provides a task requirement parsing system based on entity linking, characterized in that it includes:

[0033] The candidate entity generation module is used to generate candidate entities based on the task requirement text and obtain description information from the knowledge base;

[0034] The query statement construction module is used to construct a query statement according to the task requirement text, splice the query statement with the candidate entity and description information, and obtain the splicing result.

[0035] An unlinkable prediction module performs unlinkable prediction on the concatenation result to obtain a preliminary unlinkable judgment result and matching probabilities between different candidate entities and their mentions;

[0036] A link order prediction module is used to sort the linkable entity mentions according to the matching probability between the linkable entities and their mentions after globally disambiguating the linkable entity mentions in the preliminary prediction results and the re-prediction results;

[0037] The multi-stage linking module encodes the sorted results to obtain a global score, and obtains the optimal task requirements based on the global score;

[0038] The unlinkable re-prediction module is used to re-predict the unlinkable entity mentions in the preliminary prediction results to obtain linkable entities.

[0039] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, it implements any of the entity linking-based task requirement parsing methods described above.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] (1) The present invention draws on the method of reading comprehension to construct query statements for the target type, task means, military service, task domain and task purpose of the current task respectively, and then splices the query statements with the candidate entities and entity descriptions in the knowledge graph to provide additional semantic information.

[0042] (2) The present invention updates the current query statement with the linked entity, adds the context description of the entity mention, and enriches the context semantics.

[0043] (3) The present invention designs a re-prediction, which is specifically a three-stage unlinkable verification strategy. The first stage roughly makes a preliminary judgment through the classification layer; the second stage verifies the specific options and returns the prediction results through intensive reading; the third stage makes a final judgment on the unlinkable entities by introducing the description information of the linked entities in the knowledge base. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A schematic diagram of the structure of a task requirement parsing system based on entity linking provided by an embodiment of the present invention;

[0045] Figure 2 A flowchart of the unlinkable prediction module provided by an embodiment of the present invention;

[0046] Figure 3 A flowchart of the link order prediction module provided by an embodiment of the present invention;

[0047] Figure 4 A workflow diagram of a multi-stage linking module provided by an embodiment of the present invention;

[0048] Figure 5 A flowchart of the unlinkable re-prediction module provided by an embodiment of the present invention;

[0049] Figure 6 A comparison chart of accuracy rates under different numbers of mentions provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0051] Embodiment 1:

[0052] The present invention provides a task requirement analysis method based on entity linking. On the basis of analyzing the task requirement analysis, key elements such as target type, task means, military service, task domain and task purpose are defined, and the requirement analysis is performed based on these key elements. The process includes:

[0053] By extracting knowledge from the current task text, we can extract the target type, task means, military service, task domain and task purpose information. We use the extracted task information as constraints and match it in the graph library through entity linking technology. Finally, we output the task actions and task forces that can complete the current task as the results of task requirement analysis, realize the mapping of tasks to actions and forces, and complete the task requirement analysis.

[0054] The present invention relates to a task requirement parsing method, system and storage medium based on entity linking. Aiming at the entity linking problem in short texts, the accuracy of entity linking is improved through multi-stage entity linking and unlinkable prediction models, which is suitable for parsing complex task requirements.

[0055] like Figure 1 As shown, it is a structural diagram of a task requirement parsing system based on entity linking provided by an embodiment of the present invention, including:

[0056] The candidate entity generation module generates candidate entities and extracts description information. The query statement construction module concatenates this information with the generated query statement. The unlinkable prediction module makes a preliminary judgment on the concatenation results and calculates the matching probability. The link order prediction module sorts the mentions according to the matching probability. The multi-stage link module performs global disambiguation in a dynamic multi-round manner. The unlinkable re-prediction module re-predicts the unlinkable entities after the text link is completed.

[0057] Figure 1 The multi-stage interaction-based entity linking is demonstrated, and efficient entity linking is achieved through multi-stage interaction and iterative processing.

[0058] Specifically, the query statement construction module can be implemented with the input short text as follows: According to the advance reconnaissance intelligence, there may be minefields in the sea areas near South Beach and North Beach. The minefields on the south and north sides should be cleared to open up a road passage.

[0059] The query statement is constructed as question (Q1): Based on the prior reconnaissance intelligence, where may the minefields exist? Clear the minefields on the south and north sides to open up a road passage.

[0060] The generated options (O1) are: sea area, air area, and land area, and the corresponding descriptions (D1) are: area located on the sea surface, area located in the air, and area located on land.

[0061] from Figure 1 From the figure, we can see the overall process of entity linking based on multi-stage interaction. Through close interaction and multi-stage processing between modules, the accuracy of entity linking in short texts is improved, which can be applied to task requirement parsing.

[0062] Given a task requirement text, the present invention first generates potential candidate entities through matching rules such as homonyms and aliases, and obtains a description from the knowledge base for each candidate entity. Drawing on the reading comprehension method, query statements are constructed for the target type, task means, military service, task domain, and task purpose of the current task, and the query statements are spliced ​​with the candidate entities and description information in the knowledge graph as the input of the BERT model, and then unlinkable prediction is performed.

[0063] Unlinkable prediction is divided into two stages. The first stage is to perform binary classification on the feature representation of the query sentence through a multilayer perceptron (MLP) to obtain the probability of whether it is an unlinkable (NIL) entity. The second stage is to obtain the matching probability between different entity candidates and mentions through a fully connected layer (FC) after the feature representation of the query sentence, candidate entities and description information is concatenated.

[0064] Traditional methods have difficulty handling the problem of calculating multiple mentions simultaneously during the cold start phase and lack an effective sorting mechanism. In order to solve the problem of entity linking cold start, the present invention simultaneously calculates the candidate probabilities of FC for multiple mentions. After the calculation, a sort is made according to the extreme values ​​of the probability differences between the candidates. The mentions and candidate entities with the largest extreme values ​​are the first disambiguated objects and disambiguation results. In the absence of initial data, the priority disambiguation objects are determined by comparing the probability differences of the candidate entities, which enhances the disambiguation effect during the cold start phase.

[0065] The disambiguation result encoding obtained in the previous module and the BERT encoding obtained after the query statement update are globally scored through the GRU module. Using the disambiguation results of the previous stage and the dynamically updated query information, richer context and global information can be captured, improving the effect of global disambiguation.

[0066] Subsequently, the scores obtained during sequential prediction are weighted and summed to complete disambiguation, which comprehensively considers local and global information and improves the accuracy and robustness of candidate entity ranking.

[0067] Finally, the unlinkable entity mentions are judged again using machine reading comprehension methods to improve the accuracy of unlinkable entity prediction.

[0068] The specific working process of each module is as follows:

[0069] 1. Candidate entity generation module:

[0070] Given a short text , which contains a set of identified mentions The goal of an entity linking-based task parsing system is to find a mapping that maps each entity mention to Link to a target entity , the target entity is an entity existing in the knowledge base, or is an unlinkable entity, i.e. .

[0071] Before entity disambiguation, for each entity mention Generate potential candidate entities using matching rules such as same name and alias ,in R is a predefined parameter used to prune the candidate set.

[0072] 2. Query statement construction module:

[0073] The local entity disambiguation is formalized into a multi-choice machine reading comprehension method consisting of query statements and modules. In addition, a NIL validator is designed to facilitate the judgment of whether the task requirement text is NIL.

[0074] In building i Entity mention query When the entity mentions Replaced with a single Token, for each candidate entity mentioned by the entity and its description ,use Indicates the start of a sequence and Indicates a sequence separator, two special tokens are concatenated as the input sequence .

[0075] (1)

[0076] The goal of each module is to make full use of the semantic information of the options and select the correct link entity from the available options. k For a query with answer options, first get k Input query: Then, each query is fed into the BERT encoder and the final prediction for each option is obtained using a softmax normalization function after BERT encoding.

[0077] (2)

[0078] (3)

[0079] In the formula, For the i The entity mentioned jThe representation of candidate entities, is the input sequence, and are the weight matrix and bias respectively, For the i The entity mentioned j The predicted probability of each candidate entity is given by the cross entropy, and the cross entropy is defined as the training loss function for linking entity prediction.

[0080] 3. Unlinkable prediction module:

[0081] like Figure 2 As shown, it is a workflow diagram of the unlinkable prediction module, which uses a multi-layer perceptron (MLP) and a fully connected layer (FC) to perform two-stage unlinkable prediction.

[0082] BERT encoding: First, the task requirement text is encoded through the BERT model to obtain feature representations, namely question features (Q1) and candidate entities and their description features (D1, Q1, O1).

[0083] Multilayer Perceptron (MLP): The multilayer perceptron is used to make a preliminary unlinkable prediction of the question feature (Q1) and generate a preliminary unlinkable judgment result.

[0084] Fully connected layer (FC): The candidate entities and their description features (D1, Q1, O1) are passed through the fully connected layer for accurate unlinkable prediction to obtain the matching probability between different candidate entities and mentions.

[0085] In order to solve the problem of unlinkability that often occurs in short texts, a two-stage verification mechanism is adopted. In the first stage, the query statement is roughly Put it into the multi-layer perceptron for classification and preliminarily determine whether it is an unlinkable entity.

[0086] (4)

[0087] In the formula, is the sigmoid function, MLP is a multi-layer perceptron, and the classifier is trained using binary cross entropy loss.

[0088] (5)

[0089] In the formula, represents the prediction loss function for unlinkable entities, Represents the prediction result, is the sample index, Indicates the sample size.

[0090] Then, in the second stage, the output of the multi-layer perceptron (MLP) is taken as input and an option marked as NIL is added, which can participate in the comparison between candidate entities. Finally, its joint loss is the weighted sum of the candidate entity prediction loss and the NIL loss.

[0091] (6)

[0092] In the formula, represents the total loss function, represents the loss function for link entity prediction, and Are two hyperparameters used to balance the weights of the two losses.

[0093] The unlinkable prediction module includes a local module, which is mainly responsible for the preliminary processing of the input short text, extracting candidate entities, and performing preliminary entity disambiguation. After local disambiguation, all candidate entity representations of each entity mentioned in the short text are obtained, which will be used for subsequent global disambiguation.

[0094] 4. Link order prediction module:

[0095] like Figure 3 Figure 2 shows the workflow of the link order prediction module, which ranks the mentions according to the prediction scores of the local modules, starting with the entity mentions that are easier to disambiguate.

[0096] Candidate entities and their scores: The figure lists the matching scores of different candidate entities, such as the candidate entity O 1 , O 2 , O 3 In the entity m 1 , m 2 , m 3 The score on .

[0097] Range calculation: Calculate the range R of the candidate entity matching scores for each entity mention to determine the difficulty of disambiguation.

[0098] Link order: Sort entity mentions according to the size of the range R, and start disambiguation from the entity mention with the largest range to reduce the interference of noisy data.

[0099] Before global disambiguation, starting with entity mentions that are easier to disambiguate will effectively reduce the interference of noisy data. Mentions are ranked based on the prediction scores generated by local modules, and the rules are as follows:

[0100] (7)

[0101] Where Rank{} represents the ranking of entity mentions, represents the set of entity mentions sorted by disambiguation difficulty, K represents the candidate entity set, L 1 is the norm representing the Manhattan distance between vectors, Indicates the predicted i The entity mentioned k The probability of a candidate entity, The predicted i The entity mentioned j The probability of a candidate entity.

[0102] 5. Multi-stage linking module:

[0103] like Figure 4 As shown, it is a workflow diagram of the multi-stage linking module provided in an embodiment of the present invention. In the process of disambiguating a task, the system first inputs information about the sea minefield and a description of the task to be performed, extracts features through the BERT model, and uses the GRU module to dynamically update the query information in combination with the disambiguation results at the previous moment, gradually accumulates context information, and finally achieves accurate disambiguation of candidate entities.

[0104] Figure 4 The arrow in the middle indicates the direction of information flow. The BERT model is used to obtain the feature representation of the input, and the GRU module is used to combine the contextual information of the previous and next moments for global scoring and disambiguation. By processing moment by moment, contextual information is accumulated to improve the accuracy of disambiguation. This method combines local and global information, which helps to match entities more accurately and improve the effect of entity linking.

[0105] For example, the questions in the multi-stage link module set in this application are: According to the advance reconnaissance intelligence, what may exist in the sea area, and clear the minefields on the south and north sides to open a road passage. This question asks about possible threats (for example, minefields) in the sea area and the actions to clear the minefields on the south and north sides to open a passage. The system needs to identify and link the entity "Possible threats in the sea area". According to the advance reconnaissance intelligence, there may be minefields in the sea area, and what tasks are performed to clear the minefields on the south and north sides. This question clearly states that there may be minefields in the sea area, and asks about the specific tasks after clearing the minefields, such as opening a passage. The system needs to identify and link the entity "Specific tasks after clearing the minefields".

[0106] The goal of the multi-stage linking module is to capture topical coherence and interdependence in a global disambiguation of all mentions. To achieve this goal, previously linked entities are leveraged to enhance the linking of later entities in a dynamic multi-round manner. After ranking the mentions, the entity mentions are The representation of the candidate entity with the highest score is set to , forming the initial history hint. In order to obtain the global entity score for subsequent mentions, the current query is first updated with the linked entity to accumulate previous context information. Then the updated query and each candidate entity mentioned currently are linked with a special word unit. and Concatenate them as the input sequence. Similar to the local module, the model uses the BERT encoder to obtain the representation of each candidate mentioned currently:

[0107] (8)

[0108] In the formula, is the updated candidate entity representation, It is i The entity mentioned j description of candidate entities, It is i The entity mentioned j candidate entities.

[0109] Subsequently, a GRU module is used to introduce historical cues to disambiguate the current mention.

[0110] (9)

[0111] (10)

[0112] (11)

[0113] In the formula, For the i The entity mentioned j The reset gating signal of the candidate entity. is the candidate entity with the highest score at the previous moment, is the update gating signal; Represents the trainable weight matrix of the current moment features in the reset gate; Represents the candidate hidden state at the current moment; Represents the trainable weight matrix of the hidden state at the previous moment in the reset gate; Represents the trainable weight matrix of the current moment features in the update gate; A trainable weight matrix representing the previous hidden state in the update gate; Represents the trainable weight matrix of the current feature; Represents the trainable weight matrix of the hidden state at the previous moment after being reset by the reset gate. An entity representing the output, Indicates the reset gate signal.

[0114] Finally, using the output to calculate the global score of the candidate entity mentioned currently. After prediction, the candidate entity representation with the highest global score in the current entity mention flows to the next step of global disambiguation as the updated history representation. The cross entropy function is used as the training loss function of the global module.

[0115] (12)

[0116] In the formula, is a global loss function used to evaluate the overall prediction accuracy of the model. is the global prediction of entity mentions, is the number of samples.

[0117] The prediction scores of the local and global modules are combined to determine the final disambiguated entity for each entity mention in the short text.

[0118] (13)

[0119] In the formula, Score is the final score, The weight to balance the local and global scores.

[0120] 6. Unlinkable re-prediction module:

[0121] like Figure 5 , which is a workflow diagram of the unlinkable re-prediction module. After the text link is completed, this module re-predicts the unlinkable entity mentions to improve the prediction accuracy.

[0122] The issues and options in the unlinkable prediction module are handled as follows:

[0123] Question (Q1): Based on the advance reconnaissance intelligence, we need to know what might exist in the sea area and clear the minefields on the south and north sides to open up a passage on the road.

[0124] Options (O1): 1. Minefields are an enemy threat; 2. Opening a passage on the road is a mission.

[0125] The annotation text is based on the existing fast and slow intelligence. There may be torpedoes in the nearby waters. Please report to the patrol section of the sea surface detection and combat readiness. The system needs to determine whether there are minefields in the sea area and the entities performing the mission. If the entity in the question cannot be directly linked to an entity in the knowledge base, the system will use a two-stage unlinkable prediction mechanism to first make a preliminary judgment and then further verify it through option comparison. Finally, the accuracy of unlinkable entity prediction is improved through the combination of BERT model, Dropout layer and Dense layer.

[0126] BERT encoding: The task requirement text is encoded through the BERT model to obtain feature representation, including question features (Q1), unlinkable question features (N1), and candidate entity features (O1).

[0127] Dropout layer: The feature representation after BERT encoding is regularized through the Dropout layer to prevent overfitting.

[0128] Dense layer: predicts unlinkable entities through the fully connected layer to determine whether they are unlinkable entities.

[0129] The unlinkable re-prediction module reconstructs the query statement using the additional description of the linked entities and makes judgments through BERT encoding, further improving the prediction accuracy of unlinkable entities.

[0130] After linking the entire text, the unlinkable entity mentions are re-predicted, and the query statement is reconstructed using the additional descriptions of the linked entities in the knowledge base, which are BERT-encoded and judged through the fully connected layer.

[0131] The cross entropy function is used as the training loss function of the non-linkable re-prediction module.

[0132] (14)

[0133] In the formula, is the training loss function of the non-linkable re-prediction module.

[0134] In the dynamic multi-round disambiguation process, the present invention uses previously linked entities to update the current query in each round. This method enhances the accuracy and consistency of subsequent entity links by continuously accumulating context information; and continuously updates and improves query statements to ensure that more accurate and context-rich queries can be provided in each round, thereby improving the accuracy of disambiguation.

[0135] like Figure 6 As shown, it is a comparison chart of accuracy rates under different numbers of mentions provided by an embodiment of the present invention. It can be seen from the chart that under different numbers of mentions, the accuracy rate of query updates is higher.

[0136] The disambiguation result encoding obtained in the previous step is combined with the BERT encoding of the updated query and globally scored through the GRU module. This combination method effectively utilizes the contextual information and global semantic relationships in the disambiguation process and improves the model's disambiguation ability.

[0137] Embodiment 2:

[0138] The present invention also provides a task requirement parsing system based on entity linking, comprising:

[0139] The candidate entity generation module is used to generate candidate entities based on the task requirement text and obtain description information from the knowledge base;

[0140] The query statement construction module is used to construct a query statement according to the task requirement text, splice the query statement with the candidate entity and description information, and obtain the splicing result.

[0141] An unlinkable prediction module performs unlinkable prediction on the concatenation result to obtain a preliminary unlinkable judgment result and matching probabilities between different candidate entities and their mentions;

[0142] A link order prediction module is used to sort the linkable entity mentions according to the matching probability between the linkable entities and their mentions after globally disambiguating the linkable entity mentions in the preliminary prediction results and the re-prediction results;

[0143] The multi-stage linking module encodes the sorted results to obtain a global score, and obtains the optimal task requirements based on the global score;

[0144] The unlinkable re-prediction module is used to re-predict the unlinkable entity mentions in the preliminary prediction results to obtain linkable entities.

[0145] Embodiment three:

[0146] The present invention provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, a task requirement parsing method based on entity linking is implemented. The method comprises:

[0147] Generate candidate entities based on the task requirement text and obtain description information from the knowledge base;

[0148] Construct a query statement according to the task requirement text, and concatenate the query statement with the candidate entity and description information to obtain a concatenation result;

[0149] Performing unlinkable prediction on the concatenated results to obtain unlinkable preliminary prediction results, matching probabilities between different linkable entities and their mentions;

[0150] The unlinkable entity mentions in the preliminary prediction results are re-predicted and added to the knowledge base. After global disambiguation of the linkable entity mentions in the preliminary prediction results and the re-prediction results, the linkable entity mentions are sorted according to the matching probability between the linkable entities and their mentions, the sorted results are encoded to obtain a global score, and the optimal task requirements are obtained based on the global score.

[0151] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0152] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0153] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0154] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0155] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A task requirement parsing method based on entity linking, characterized in that: include: Generate candidate entities based on the task requirement text and obtain description information from the knowledge base; Construct a query statement according to the task requirement text, and concatenate the query statement with the candidate entity and description information to obtain a concatenation result; Performing unlinkable prediction on the concatenated results to obtain unlinkable preliminary prediction results, matching probabilities between different linkable entities and their mentions; Re-predict the unlinkable entity mentions in the preliminary prediction results, add the unlinkable entity mentions in the re-prediction results to the knowledge base, perform global disambiguation on the linkable entity mentions in the preliminary prediction results and the re-prediction results, sort the linkable entity mentions according to the matching probability between the linkable entities and their mentions, encode the sorted results to obtain a global score, and obtain the optimal task requirements based on the global score; Re-predict the unlinkable entity mentions in the initial prediction results, including: Incorporate the description information of linkable entities into the knowledge base to reconstruct the query statement, and encode the unlinkable entities in the task requirement text through the BERT model to obtain feature representation; Regularize the feature representation through the Dropout layer; The regularized feature representation is passed through the Dense layer to predict the unlinkable entities and obtain the linkable entities. Linkable entity mentions are ranked based on the matching probability between linkable entities and their mentions, including: Predict matching scores for candidate entity mentions based on the matching probabilities between linkable entities and their mentions; Calculate the range of matching scores for each candidate entity mention; Sort candidate entity mentions by range size; The sorted results are encoded to obtain a global score, and the optimal task requirements are obtained based on the global score, including: Start disambiguation from the candidate entity mention with the largest range, reduce the interference of noise data, and obtain the disambiguation result encoding; The disambiguation result encoding and the BERT encoding obtained after query sentence update are globally scored through the GRU module; Get the optimal task requirements based on the global score.

2. The method according to claim 1, characterized in that The step of constructing a query statement according to the task requirement text includes: Use reading comprehension to construct query statements based on the task requirements.

3. The method according to claim 1, characterized in that: Perform unlinkable prediction on the concatenation result to obtain unlinkable preliminary prediction results, matching probabilities between different linkable entities and their mentions, including: The concatenated result is encoded by the BERT model to obtain a feature representation, wherein the feature representation includes question features, candidate entities and their description features; The problem features are used through a multi-layer perceptron to make a preliminary unlinkable prediction and generate a preliminary unlinkable judgment result; Question features, candidate entities and their description features are passed through a fully connected layer for accurate unlinkable prediction, and the matching probabilities between different linkable entities and their mentions are obtained.

4. The method according to claim 3, characterized in that: The method includes: Treat unlinkable entities and an option marked as NIL as candidate entities; Question features, candidate entities and their description features are passed through a fully connected layer for accurate unlinkable prediction, and the matching probabilities between different linkable entities and their mentions are obtained.

5. The method according to claim 1, characterized in that: The mission requirements include: target type, mission means, military service, mission domain and mission purpose information.

6. A task requirement parsing system based on entity linking, characterized in that: include: The candidate entity generation module is used to generate candidate entities based on the task requirement text and obtain description information from the knowledge base; The query statement construction module is used to construct a query statement according to the task requirement text, splice the query statement with the candidate entity and description information, and obtain the splicing result. An unlinkable prediction module performs unlinkable prediction on the concatenation result to obtain a preliminary unlinkable judgment result and matching probabilities between different candidate entities and their mentions; The link order prediction module is used to sort the linkable entity mentions according to the matching probability between the linkable entities and their mentions after globally disambiguating the linkable entity mentions in the preliminary prediction results and the re-prediction results, and to perform the following steps: Predict matching scores for candidate entity mentions based on the matching probabilities between linkable entities and their mentions; Calculate the range of matching scores for each candidate entity mention; Sort candidate entity mentions by range size; The multi-stage linking module encodes the sorted results to obtain a global score, and obtains the optimal task requirements based on the global score to perform the following steps: Start disambiguation from the candidate entity mention with the largest range, reduce the interference of noise data, and obtain the disambiguation result encoding; The disambiguation result encoding and the BERT encoding obtained after query sentence update are globally scored through the GRU module; Get the optimal task requirements based on the global score; The unlinkable re-prediction module is used to re-predict the unlinkable entity mentions in the preliminary prediction results to obtain linkable entities for performing the following steps: Incorporate the description information of linkable entities into the knowledge base to reconstruct the query statement, and encode the unlinkable entities in the task requirement text through the BERT model to obtain feature representation; Regularize the feature representation through the Dropout layer; The regularized feature representation is passed through the Dense layer to predict the unlinkable entities and obtain the linkable entities.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the task requirement parsing method based on entity linking as described in any one of claims 1 to 5 is implemented.