Marine exploration task decomposition method based on relation extraction and hierarchical task network
By using a relation extraction and hierarchical task network approach, the problem of time-consuming and labor-intensive task decomposition in traditional marine exploration is solved, achieving efficient and accurate task decomposition that adapts to complex and ever-changing marine environments.
Patent Information
- Application Number
- CN202511830773.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-06
- Publication Date
- 2026-03-27
AI Technical Summary
Traditional methods for decomposing marine exploration missions are cumbersome, time-consuming, and labor-intensive, making them difficult to adapt to the rapidly changing marine environment. Existing automated planning technologies cannot fully understand the logic and constraints of unstructured text, hindering scientific decision-making and rapid response.
We employ a relation extraction and hierarchical task network approach. By constructing and preprocessing a small sample dataset, we utilize the pre-trained BERT model for entity relation extraction and combine it with a deep neural network and a hierarchical task network (HTN) to achieve efficient decomposition of marine exploration tasks.
It significantly improves the efficiency of ocean exploration mission decomposition, shortens the decomposition time, and ensures high accuracy and adaptability.
Smart Images
Figure CN121743864A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of ocean exploration, in particular to a method for decomposing an exploration task. BACKGROUND
[0002] In modern ocean exploration tasks, the ocean environment is complex and variable, multi-task concurrency, decision-making command window compression, and exploration resource high coordination have become the new normal. In the face of this situation, the management and decision-making of ocean exploration tasks are facing unprecedented challenges. The traditional method of decomposing ocean exploration tasks relying on artificial experience is cumbersome and time-consuming, and it is difficult to adapt to the changing ocean environment. The existing automatic planning technology cannot fully understand the logic and constraint relationship contained in the unstructured text in the ocean exploration task, hindering scientific decision-making and rapid response capability.
[0003] To overcome the above bottleneck problems, the present application provides a method for decomposing an ocean exploration task based on relation extraction and hierarchical task network, aiming to promote the intelligent transformation of ocean exploration task decomposition. The core of the method is to rely on deep learning and natural language processing technology. First, the entity labeling method is used to construct a small sample training dataset, and the semantic relationship between different entities in complex text is modeled. Then, the model is trained to extract entity relationships from semi-structured or unstructured input data. Its core advantage is that it can automatically convert implicit text expressions into structured knowledge that computers can understand and process through relation extraction. Finally, the hierarchical task network is used to recursively decompose complex tasks into subtasks, greatly improving the efficiency of ocean exploration task decomposition. SUMMARY
[0004] In order to overcome the shortcomings of the prior art, the present application provides a method for decomposing an ocean exploration task based on relation extraction and hierarchical task network.
[0005] The technical solution adopted by the present application to solve its technical problems is as follows:
[0006] Step 1. Small sample dataset construction and preprocessing;
[0007] Given an ocean exploration task to be decomposed, the ocean exploration task is regarded as a text containing multiple subtasks. From the ocean exploration task to be decomposed, select the triplets of the form "entity1-relation-entity2" to form the training set. From an ocean exploration task to be decomposed, divide the triplets of the form "entity1-relation-entity2" into subtasks, and add the subtasks to the small sample dataset.
[0008] Step 2. Relation extraction;
[0009] For a given task instance , use the pre-trained model BERT to convert the task instance The semantic information of the task instance is encoded as a low-dimensional vector to represent the task instance Each word in the text is encoded to correspond to a vector.
[0010] (4)
[0011] In formula (4), BERT( ) represents that the input word is represented in low dimension by using a pre-trained model to obtain a vector represents the semantic information of the word, and all word vectors in the encoded task instance are input into a pooling feature extraction layer to obtain an embedding vector E with a fixed size:
[0012] (5)
[0013] In formula (5), Pool( ) represents the pooling feature extraction, which is used to extract word features and output a feature vector with a fixed dimension of to represent the semantic information of the entire instance;
[0014] After feature extraction, a deep neural network (DNN) is used to map different sentences to a vector, and a prototype vector of the task instance vector is obtained by calculating the mean value of the same class sample vector:
[0015] (6)
[0016] In formula (6), represents the prototype obtained by calculation, is the vector representation of a relationship in the classified set A; in classification, the distance between the unclassified set B and the vector is calculated by using the softmax function, so as to obtain the probability representation of the task instance e in the relationship set R:
[0017] (7)
[0018] In formula (7), d( ) represents the Euclidean distance between two vectors, f( represents the encoding of the task instance , and represents the probability that the task instance is a relationship type , and finally, the loss function of the entire model is defined as:
[0019] (8)
[0020] In formula (8), All parameters of the model are represented, and in the model training process, the model parameters are adjusted by back propagation to minimize the loss function to complete the classification training.
[0021] Step 3. Task decomposition based on HTN;
[0022] After completing the relation extraction, the subtasks obtained by the relation extraction are decomposed according to the complexity of the task by using HTN. First, the composite task is decomposed into different subtasks, and then the subtasks are further decomposed into indivisible meta tasks.
[0023] In step 1, after constructing the small sample dataset, first divide the small sample dataset into training set and test set, divide the training set into classified set A and unclassified set B, and the relation extraction task is to predict the entity pair (h, t) in the unclassified set B belongs to the relationship r in the classified set A. The relation extraction task forms a given relation set R, and the definitions of the classified set A and the unclassified set B are as follows:
[0024] (1)
[0025] (2)
[0026] (3)
[0027] Wherein, The jth specific instance is represented, i=1,2, …, N, wherein The ith relation is represented, the text in the ith relation is The head entity in the text in the ith relation is The tail entity in the text in the ith relation is N represents that there are N relations in the classified set A, M represents that there are M different instances for each relation type, and it is called N way M Shot classification task to judge which kind of the classified set A the unclassified instance belongs to. Through repeated N way M Shot classification training, the discrimination performance of the model for different relation categories can be enhanced.
[0028] In step 3, in the process of ocean exploration task decomposition, the decomposition is constrained according to the following four principles:
[0029] 1) Task decomposition should have a hierarchical nature from complex to simple, decompose a task into multiple subtasks, and further decompose the subtasks until they become meta tasks.
[0030] 2) The decomposed subtasks and metatasks have independence, reducing the conflicts between tasks;
[0031] 3) The decomposed subtasks have combinability, i.e. multiple metatasks can complete a complex task by combination;
[0032] 4) The task decomposition has appropriateness, and the total number of tasks, subtasks and metatasks meets the actual marine exploration task requirements.
[0033] Based on the four criteria, the correlation coefficient , the cohesion coefficient and the task granularity control the task decomposition.
[0034] The specific steps of task decomposition are as follows:
[0035] First, input the initial task W0, task database and domain knowledge; then generate the initial task network W1, select task T from the initial task network W1 and judge whether the task T is decomposable. If the task T is not decomposable, the task T is a metatask, which is directly removed from W1 and taken as the final task network W; if the task T is decomposable, calculate the correlation coefficient Q, cohesion coefficient C and task granularity G of the task T, and whether the distance D(Q), D(C) and D(G) between the target solution meet the user-defined conditions, if not, the task decomposition fails; if it meets, remove the task T from W1, and take T as the final task network W, judge whether the initial task network W1 is empty set after removing the task T, if the initial task network W1 is not empty set, then loop the step of judging whether the task T is decomposable until W1 is empty set, if the initial task network W1 is empty set, stop the loop, and get the final task network W.
[0036] The correlation coefficient represents the correlation between subtasks, first study the correlation between two tasks, which depends on time dependence, resource dependence, information dependence and other dependence; for time dependence, when the execution time of a task is completely determined by another task, the two tasks can be considered as time fully dependent, and its time dependence is denoted as: The time dependence between the two tasks is denoted as:
[0037] (9)
[0038] For material dependence Information dependence Other dependencies The scores were given by experts within the range of [0,1] and evaluated in five levels, based on time dependence, resource dependence, information dependence, and other dependencies. The correlation coefficient between the two tasks was determined. Represented as:
[0039] (10)
[0040] Correlation coefficient The correlation coefficient represents the correlation between two tasks. To ensure the independence between the two tasks, the correlation coefficient is used. The value should be as small as possible;
[0041] Cohesion coefficient This indicates the degree of connection between the subtasks that can be decomposed from a given task, reflecting the composability between tasks. If the cohesion coefficient of two tasks is greater than a threshold, the two tasks are considered related; the threshold is determined based on domain experience. If a task is the smallest unit of decomposition, then it is denoted as a task. Cohesion coefficient If a certain task Decomposed into For each subtask of task y, The cohesion coefficient is expressed as:
[0042] (12)
[0043] in, For all subtasks in the subtask The number of related subtasks: If the correlation coefficient Q between two tasks is greater than a threshold, the two tasks are considered related. The cohesion coefficient is expressed as:
[0044] (13)
[0045] For the cohesion coefficient of the meta-task, formula (12) is used for recursive calculation layer by layer; for the task decomposition scheme, the closer the connection between the lower-level subtasks of the subtask, the larger the cohesion coefficient. Therefore, the cohesion coefficient It should be as large as possible, but it should be noted that an increase in the cohesion coefficient may lead to a simultaneous increase in the correlation coefficient. This issue should be controlled during the task decomposition process.
[0046] Task granularity The decomposition degree of a task decomposition scheme, the task granularity is directly defined as the inverse of the number of tasks, which makes the task granularity insensitive to the number of tasks when the number of layers remains unchanged. Therefore, it is defined as the inverse of the square root of the number of tasks. On this basis, in order to make the task granularity of the un-decomposed state 1, a coefficient is added, so the task granularity is defined as:
[0047] (14)
[0048] When performing a decomposition, the task granularity is reduced;
[0049] To control the task decomposition process, three constraint conditions are defined as follows:
[0050] (15)
[0051] (16)
[0052] (17)
[0053] wherein, is the maximum allowed correlation coefficient, is the minimum allowed cohesion coefficient, and are the maximum and minimum values of the task granularity allowed, respectively, and in the process of task decomposition, the decomposition process is optimized by adjusting the distance.
[0054] The weight parameters , , and are calculated according to the self-defined importance , , , The importance value is between (0, 1), and the closer to 1, the more important the item is. The calculation method of the weight parameter is:
[0055] (11)
[0056] The weight parameters , , and are calculated according to formula (11).
[0057] The threshold value is in the range of 0.2-0.8.
[0058] An electronic device comprising: one or more processors; memory; one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs configured to perform the method as described above.
[0059] A computer-readable storage medium storing program code, the program code executable by a processor to perform the method as described above.
[0060] The present application has the beneficial effect of providing a small sample dataset construction and preprocessing method for ocean exploration task decomposition training, solving the current problem of data scarcity and difficulty in obtaining ocean exploration tasks. A method for decomposing ocean exploration tasks based on relationship extraction and hierarchical task networks is proposed, which greatly shortens the time consumption of ocean exploration task decomposition while ensuring high accuracy. BRIEF DESCRIPTION OF DRAWINGS
[0061] Figure 1 is a dataset construction schematic diagram of the present application.
[0062] Figure 2 is a feature extraction schematic diagram of the present application.
[0063] Figure 3 is the ocean exploration task decomposition process of the present application.
[0064] Figure 4 is the ocean exploration task decomposition flowchart of the present application. DETAILED DESCRIPTION
[0065] The present application will be further described below in conjunction with the drawings and examples.
[0066] Relationship extraction
[0067] Relationship extraction is one of the core sub-tasks of information extraction (IE) in natural language processing, aiming to identify specific semantic relationships between different entities (such as people, objects, places, organizations, etc.) in a text. It is widely used in the fields of knowledge graph construction, event extraction, sentiment analysis, and intelligent question answering. Existing relationship extraction methods mainly fall into the following three categories:
[0068] (1) Rule and template-based relationship extraction method
[0069] Early relationship extraction methods usually manually write rules or templates by experts in a certain field to identify and represent language patterns between specific relationships. These methods have strong explainability for specific domain tasks or structured text, and can achieve high accuracy results. However, these methods have the significant disadvantage of poor generalization ability, as they rely on expert knowledge and specific language patterns, making it difficult to handle complex and diverse natural language expressions and migrate to new domain applications. In addition, these methods are time-consuming and labor-intensive in constructing rules and templates, with high construction costs.
[0070] (2) Relationship extraction method based on supervised learning
[0071] Supervised learning-based relationship extraction methods usually require large-scale and high-quality training datasets, which are generally annotated according to the triplets of entity1-entity2-relation. These methods usually treat relationship extraction as a classification problem: given a sentence and a pair of candidate entities, a classifier is used to determine whether there is a specific relationship in the pre-set relationship set between the two entities. Currently, most supervised learning-based relationship extraction methods are based on deep learning techniques, using CNN, RNN, LSTM, etc. to extract local features or sequence features, and introducing attention mechanisms to make the model focus on more important entities for relationship judgment. These methods have achieved good results. However, these methods usually rely heavily on large amounts of high-quality data, and the cost of data annotation is extremely high.
[0072] (3) Relationship extraction method based on pre-trained model
[0073] In recent years, with the continuous development of large language models, fine-tuning of deep neural network models pre-trained on large-scale general training sets for relationship extraction has become a new research direction. These methods use pre-trained models to learn rich language knowledge, such as using BERT models to learn context-related word vectors, which can significantly improve relationship extraction performance.
[0074] Hierarchical task network
[0075] Hierarchical task network (HTN) is an artificial intelligence planning method that recursively decomposes complex tasks into simpler subtasks and performs threat checks and solutions during decomposition until the decomposition is a directly executable original task to solve problems and generate action plans. Its core idea is "task decomposition" and relies on a rich domain knowledge base for guidance. Currently, HTN has many applications in task decomposition and planning, such as completing multi-robot complex task decomposition through HTN and dynamic path planning methods based on HTN.
[0076] The existing marine exploration task decomposition method relies on manual experts to manually decompose the task, when facing complex tasks, the manual decomposition step is tedious and time-consuming, and it is difficult to adapt to the changing marine environment. In view of the above shortcomings, a marine exploration task decomposition method based on relation extraction and hierarchical task network is proposed, and the following problems are solved:
[0077] The marine exploration task decomposition data set is difficult to obtain. Since model training usually requires a large-scale labeled data set, the marine exploration task decomposition data set is difficult to obtain due to its particularity, and there is no standard data labeling rule at present. In view of this, the present application proposes a small sample data set construction and preprocessing method for marine exploration task decomposition;
[0078] The marine exploration task decomposition step is tedious and slow. When facing complex tasks, the existing method relies on manual step-by-step decomposition, which consumes a lot of manpower and material resources. In view of this, the present application applies relation extraction and hierarchical task network to marine exploration task decomposition, and quickly and intelligently completes marine exploration task decomposition.
[0079] The steps of the marine exploration task decomposition based on relation extraction and hierarchical task network of the present application are as follows:
[0080] Step 1. Small sample data set construction and preprocessing;
[0081] Given a marine exploration task to be decomposed, the marine exploration task is regarded as a text containing multiple subtasks, and the triplets of the form "entity1-relation-entity2" are selected from the marine exploration task to be decomposed to form a training set, and the construction process is as shown in Figure 1 The triplet subtask of the form "entity1-relation-entity2 (such as Figure 1 detector-deploy-underwater)" is divided from a marine exploration task to be decomposed, and the subtask is added to the small sample data set;
[0082] After constructing the small sample data set, first divide the small sample data set into a training set and a test set, divide the training set into a classified set A and an unclassified set B, and the relation extraction task is to predict the entity pair (h, t) in the unclassified set B belongs to the relation r in the classified set A. The relation extraction task forms a given relation set R, and the definitions of the classified set A and the unclassified set B are as follows:
[0083] (1)
[0084] (2)
[0085] (3)
[0086] Among them, represents the jth specific instance, i = 1, 2, …, N, wherein represents the ith relationship, the text in the ith relationship is , the head entity in the text in the ith relationship is , the tail entity in the text in the ith relationship is , N represents that there are N relationships in the classified set A, M represents that there are M different instances for each relationship type, and determining which one of the classified set A the unclassified instance belongs to is called an N way M Shot classification task, and through repeated N way M Shot classification training, the discrimination performance of the model for different relationship categories can be enhanced;
[0087] Step 2. Relationship extraction;
[0088] For a given task instance , a pre-trained model BERT is used to encode the task instance into a low-dimensional vector to represent the semantic information of the task instance, and for the task instance , each word in the text corresponds to a vector after encoding;
[0089] (4)
[0090] In formula (4), BERT( ) represents that the low-dimensional representation of the input word is obtained by using the pre-trained model, and the vector represents the semantic information of the word, and all word vectors in the encoded task instance are input into a pooling feature extraction layer to obtain an embedding vector E with a fixed size, as shown in Figure 2
[0091] (5)
[0092] In formula (5), Pool( ) represents a pooling feature extraction, which is used to extract word features and output a feature vector with a fixed dimension of to represent the semantic information of the entire instance;
[0093] After feature extraction, a deep neural network (DNN) is used to map different sentences into a vector, and a prototype vector of the task instance vector is obtained by calculating the mean value of the same class sample vector:
[0094] (6)
[0095] In formula (6), represent each category The prototype obtained by calculation, is a vector representation of a relationship in the classified set A ; In classification, the distance between the unclassified set B and the vector is calculated using the softmax function, so that the probability representation of the task instance e in the relationship set R is obtained:
[0096] (7)
[0097] In formula (7), d( ) represents the Euclidean distance between two vectors, f( represents the encoding of the task instance , represents the probability that the task instance is a relationship type , and finally, the loss function of the entire model is defined as:
[0098] (8)
[0099] In formula (8), represents all the parameters of the model, and in the model training process, the model parameters are adjusted through back propagation to minimize the loss function to complete the classification training.
[0100] Step 3. Task decomposition based on HTN;
[0101] After completing the relationship extraction, the subtasks obtained by relationship extraction are decomposed according to the complexity of the task using HTN, and the decomposition process is shown in Figure 3 , which first decomposes the composite task into different subtasks, and then further decomposes the subtasks into indivisible meta-tasks.
[0102] In the process of ocean exploration task decomposition, the following four principles are followed to constrain the decomposition:
[0103] 1) The task decomposition should have a hierarchical nature from complex to simple, decomposing a task into multiple subtasks, and further decomposing the subtasks until they become meta-tasks.
[0104] 2) The subtasks after decomposition are independent of each other, reducing conflicts between tasks.
[0105] 3) The subtasks obtained by decomposition have combinability, that is, through combination, multiple meta-tasks can complete a complex task.
[0106] 4) The task decomposition is appropriate, and the total number of tasks, subtasks and metatasks meets the actual marine exploration task requirements.
[0107] Based on the four criteria, the correlation coefficient , the cohesion coefficient and the task granularity control the task decomposition, and the specific steps are as follows:
[0108] The task decomposition process is shown in Figure 4 . First, input the initial task W0, the task database and the domain knowledge; then generate the initial task network W1, select the task T from the initial task network W1 and determine whether the task T is decomposable. If the task T is not decomposable, the task T is a metatask, and the task T is directly removed from W1 and taken as the final task network W; if the task T is decomposable, calculate the correlation coefficient Q of the task T, the cohesion coefficient C and the distance D(Q), D(C) and D(G) between the task granularity G and the target solution, and determine whether they meet the user-defined conditions. If not, the task decomposition fails; if so, remove the task T from W1 and take T as the final task network W. Then determine whether the initial task network W1 is empty after removing the task T. If the initial task network W1 is not empty, repeat the step of determining whether the task T is decomposable until W1 is empty. If the initial task network W1 is empty, stop the loop and obtain the final task network W.
[0109] The correlation coefficient represents the correlation between subtasks and reflects the independence of subtasks. In order to represent the overall correlation of the task decomposition scheme, first, the correlation between two tasks is studied, which depends on time dependence, resource dependence, information dependence and other dependence. For time dependence, when the execution time of a task is completely determined by another task, the two tasks are considered to be fully time-dependent, and the time dependence is denoted as . If the execution times of two tasks are and , under the specified time constraint , the free time of is denoted as:
[0110] (9);
[0111] For material dependence , information dependence and other dependence , experts score in the range of [0, 1] and judge in five grades. Based on time dependence, resource dependence, information dependence and other dependence, the correlation coefficient between two tasks is is denoted as:
[0112] (10);
[0113] wherein the weight parameter , , , is calculated according to the self-defined importance , , , , the importance value being between (0, 1), the closer to 1 indicating that the item is more important, and the calculation method of the weight parameter being:
[0114] (11);
[0115] the correlation coefficient represents the correlation between two tasks, and in order to ensure the independence between two tasks, the value of the correlation coefficient should be as small as possible;
[0116] the cohesion coefficient represents the close degree of the connection between the subtasks obtained by decomposing a certain task, and can reflect the combination between tasks; if the cohesion coefficients of two tasks are greater than a threshold value, it is considered that the two tasks are related, and the threshold value is determined according to the domain experience; if the task is the smallest unit decomposed for a certain task, the cohesion coefficient of the task is denoted as , if a certain task is decomposed into subtasks, for each subtask of the task y, the cohesion coefficient is denoted as:
[0117] (12);
[0118] wherein is the number of all subtasks having a connection with the subtask , if the correlation coefficient Q of two tasks is greater than a threshold value, it is considered that the two tasks have a connection, and the threshold value is in the range of 0.2-0.8, and the cohesion coefficient of the task is denoted as:
[0119] (13);
[0120] for the cohesion coefficient of the meta-task, formula (12) is adopted for recursive calculation; for the task decomposition scheme, the closer the connection between the lower subtasks of the subtasks, the greater the cohesion coefficient, and therefore, the cohesion coefficient It should be as large as possible, but it should be noted that an increase in the cohesion coefficient may lead to a simultaneous increase in the correlation coefficient. This issue should be controlled during the task decomposition process.
[0121] Task granularity This represents the degree of decomposition in a task breakdown scheme. It reflects both the hierarchical structure and scale of the task, directly impacting the efficiency and quality of task execution. It reflects the hierarchy and balance of subtasks. When task granularity is too large, the decomposition is insufficient, leading to high internal complexity and failing to achieve the purpose of task decomposition, potentially affecting efficiency. Conversely, when task granularity is too small, the decomposition level is too deep, resulting in complex relationships and excessive interactions between tasks, which may affect the independence of subtasks and increase resource interaction costs. The measurement of task granularity is related to task hierarchy. Number of subtasks Related to, and inversely proportional to, both. As the number of task layers increases, the number of tasks increases exponentially compared to the number of task layers. Defining task granularity directly as the reciprocal of the number of tasks would make the granularity insensitive to changes in the number of tasks when the number of layers remains constant. Therefore, it is defined as the reciprocal of the square root of the number of tasks. Based on this, to ensure that the granularity of undecomposed tasks is 1, a coefficient is added. Therefore, task granularity is defined as:
[0122] (14);
[0123] After one decomposition, the task granularity is... Decrease.
[0124] To control the task decomposition process, three constraints are defined to determine the distance between the constraint and the objective solution. as follows:
[0125] (15)
[0126] (16)
[0127] (17)
[0128] in, The maximum allowable correlation coefficient, The minimum allowable cohesion coefficient, and These represent the maximum and minimum allowable values for task granularity, respectively. During task decomposition, the decomposition process is optimized by adjusting the distance.
[0129] HTN is based on domain knowledge to decompose the complex task recursively to get the hierarchical network of ocean exploration task, which effectively solves the problem of complicated and time-consuming in current ocean exploration task decomposition method.
Claims
1. A method for decomposing ocean exploration tasks based on relation extraction and hierarchical task networks, characterized in that... Includes the following steps: Step 1. Construction and preprocessing of small sample datasets; Given a marine exploration task to be decomposed, the marine exploration task is regarded as a text containing multiple subtasks. The training set is formed by selecting triples in the form of "entity 1-relation-entity 2" from the marine exploration task to be decomposed. Subtasks in the form of "entity 1-relation-entity 2" are divided from the marine exploration task to be decomposed, and the subtasks are added to the small sample dataset. Step 2. Relation extraction; For a given task instance Using the pre-trained model BERT to process task instances Encode it as a low-dimensional vector to represent the semantic information of the task instance, for each task instance. Every word in the text After encoding, it corresponds to a vector; （4）； In formula (4), BERT ( This indicates that a pre-trained model is used to represent the input words in low dimension, resulting in a vector of vectors. To represent the semantic information of the word, all word vectors in the encoded task instance are input into a pooling feature extraction layer to obtain a fixed-size embedding vector E: （5）； In formula (5), Pool( () indicates pooling feature extraction, used to extract word features and output with a fixed dimension. The feature vectors are used to represent the semantic information of the entire instance; After feature extraction, a deep neural network is used to map different sentences into a vector. The prototype vector of the task instance vector is obtained by calculating the mean of the vectors of samples of the same class. （6）； In formula (6), Represent each category The prototype obtained through calculation For a relation in a classified set A The vector representation; during classification, the softmax function is used to calculate the relationship between the unclassified set B and the vector. The distance is used to obtain the probability representation of task instance e in relation set R: （7）； In formula (7), d( ) represents the Euclidean distance between two vectors, f( Indicates a task instance Encode, Represents a task instance It is a relation type Finally, the loss function of the entire model is defined as: （8）； In formula (8), This represents all the parameters of the model. During model training, the model parameters are adjusted through backpropagation, and the loss function is minimized to complete the classification training. Step 3. Task decomposition based on HTN; After relation extraction is completed, the subtasks obtained from relation extraction are decomposed according to the complexity of the task using HTN. First, the composite task is decomposed into different subtasks, and then the subtasks are further decomposed into indivisible metatasks.
2. The ocean exploration task decomposition method based on relation extraction and hierarchical task network according to claim 1, characterized in that: In step 1, after constructing the small sample dataset, it is first divided into a training set and a test set. The training set is further divided into a classified set A and an unclassified set B. The relation extraction task is to predict whether the entity pair (h, t) in the unclassified set B belongs to relation r in the classified set A. The relation extraction task constitutes a given relation set R. The classified set A and the unclassified set B are defined as follows: （1）； （2）； （3）； in, Let j represent the j-th specific instance, i = 1, 2, ..., N, where This represents the i-th type of relation, where the text in the i-th type of relation is... The head entity in the text of the i-th relation is The tail entity in the text of the i-th relation is N represents the number of relations in the classified set A, and M represents the number of distinct instances of each relation type. Determining which type of the classified set A an unclassified instance belongs to is called an N-way M-shot classification task. By repeating the N-way M-shot classification training multiple times, the model's ability to distinguish different relation categories can be enhanced.
3. The ocean exploration task decomposition method based on relation extraction and hierarchical task network according to claim 1, characterized in that: In step 3, the decomposition of the marine exploration mission is constrained by the following four principles: 1) Task decomposition should be hierarchical, from complex to simple, breaking down a task into multiple subtasks, and then further decomposing the subtasks until they become metatasks. 2) The decomposed subtasks are independent of each other and the metatasks are independent of each other, reducing conflicts between tasks; 3) The subtasks obtained from the decomposition are composable, that is, multiple metatasks can be combined to complete a complex task. 4) The task decomposition is appropriate, and the total number of tasks, sub-tasks and meta-tasks meets the requirements of actual marine exploration tasks; Based on four criteria, a correlation coefficient is proposed. Cohesion coefficient and task granularity Control task decomposition.
4. The ocean exploration task decomposition method based on relation extraction and hierarchical task network according to claim 3, characterized in that: The specific steps for task decomposition are as follows: First, input the initial task W0, the task database, and domain knowledge. Then, generate the initial task network W1. Select task T from the initial task network W1 and determine whether task T is decomposable. If task T is not decomposable, task T is a meta-task, and task T is directly removed from W1 and used as the final task network W. If task T is decomposable, calculate whether the correlation coefficient Q, cohesion coefficient C, and distances D(Q), D(C), and D(G) between task T's granularity G and the target solution satisfy the user-defined conditions. If not, the task decomposition fails. If they are satisfied, task T is removed from W1 and used as the final task network W. After removing task T, determine whether the initial task network W1 is an empty set. If the initial task network W1 is not an empty set, repeat the step of determining whether task T is decomposable until W1 is an empty set. If the initial task network W1 is an empty set, stop the loop and obtain the final task network W.
5. The ocean exploration task decomposition method based on relation extraction and hierarchical task network according to claim 4, characterized in that: Correlation coefficient To characterize the relationships between subtasks, we first examine the relationships between two tasks. These relationships depend on time dependency, resource dependency, information dependency, and other dependencies. Regarding time dependency, when the execution time of one task is entirely determined by the execution time of another task, the two tasks can be considered fully time-dependent, and their time dependency is denoted as... If the execution times of the two tasks are respectively and , exist Under the specified time constraints Free time in time refers to the time dependency between two tasks. Represented as: (9)； For material dependence Information dependence Other dependencies The scores were given by experts within the range of [0,1] and evaluated in five levels, based on time dependence, resource dependence, information dependence, and other dependencies. The correlation coefficient between the two tasks was determined. Represented as: (10)； Correlation coefficient The correlation coefficient represents the correlation between two tasks. To ensure the independence between the two tasks, the correlation coefficient is used. The value should be as small as possible; Cohesion coefficient This indicates the degree of connection between the subtasks that can be decomposed from a given task, reflecting the composability between tasks. If the cohesion coefficient of two tasks is greater than a threshold, the two tasks are considered related; the threshold is determined based on domain experience. If a task is the smallest unit of decomposition, then it is denoted as a task. Cohesion coefficient If a certain task Decomposed into For each subtask of task y, The cohesion coefficient is expressed as: (12)； in, For all subtasks in the subtask The number of related subtasks: If the correlation coefficient Q between two tasks is greater than a threshold, the two tasks are considered related. The cohesion coefficient is expressed as: (13)； For the cohesion coefficient of the meta-task, formula (12) is used for recursive calculation layer by layer; for the task decomposition scheme, the closer the connection between the lower-level subtasks of the subtask, the larger the cohesion coefficient. Therefore, the cohesion coefficient It should be as large as possible, but it should be noted that an increase in the cohesion coefficient may lead to a simultaneous increase in the correlation coefficient. This issue should be controlled during the task decomposition process. Task granularity Representing the degree of decomposition in a task breakdown scheme, directly defining task granularity as the reciprocal of the number of tasks would make the granularity insensitive to changes in the number of tasks when the number of levels remains constant. Therefore, it is defined as the reciprocal of the square root of the number of tasks. Furthermore, to ensure that the granularity of undecomposed tasks is 1, a coefficient is added. Therefore, task granularity is defined as: (14)； After one decomposition, the task granularity is... Decrease; To control the task decomposition process, three constraints are defined to determine the distance between the constraint and the objective solution. as follows: (15)； (16)； (17)； in, The maximum allowable correlation coefficient, The minimum allowable cohesion coefficient, and These represent the maximum and minimum allowable values for task granularity, respectively. During task decomposition, the decomposition process is optimized by adjusting the distance.
6. The ocean exploration task decomposition method based on relation extraction and hierarchical task network according to claim 5, characterized in that: Weight parameters , , and Based on the importance of the customization , , , The calculation is performed, with importance values ranging from (0, 1). The closer the value is to 1, the more important the item is. The weight parameter is calculated as follows: (11) ； The weight parameters are calculated according to formula (11). , , and .
7. The ocean exploration task decomposition method based on relation extraction and hierarchical task network according to claim 5, characterized in that: The threshold value ranges from 0.2 to 0.
8.
8. An electronic device, characterized in that, include: One or more processors; Memory; One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs being configured to perform the method as described in any one of claims 1-7.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program code that can be invoked by a processor to execute the method as described in any one of claims 1-7.