Hydrothermal zone habitat event chain prediction method based on deep sea domain knowledge enhancement

Abstract

The invention discloses a hydrothermal region habitat event chain prediction method based on deep sea domain knowledge enhancement, and the method comprises the steps: carrying out the entity recognition and disambiguation of an input event sequence and a candidate event set, constructing an event composition diagram, and calculating a matching score based on the entity type and description information to complete the entity link; using graph representation to learn and train an event component graph structure, extracting semantic features of an event chain and candidate events through a self-attention mechanism, and generating knowledge-enhanced event state embedding; fusing the knowledge enhancement feature with the chain-level event state representation to form a comprehensive event state representation; a knowledge-level reward mechanism is introduced into the reinforcement learning framework, and the knowledge-level reward mechanism and the chain reward jointly optimize an event prediction strategy. According to the method, external common knowledge knowledge is introduced, knowledge-enhanced event representation is constructed, and the accuracy and robustness of hydrothermal region habitat event chain prediction are effectively improved in combination with a self-attention and reinforcement learning mechanism.

Description

Technical Field

[0001] This invention relates to the field of deep-sea exploration and marine artificial intelligence, and in particular to a method for predicting the event chain of hydrothermal vent habitats based on knowledge enhancement in the deep-sea domain. This method can be used in applications such as analyzing biological activity patterns in hydrothermal vent areas, predicting the evolution of the deep-sea ecological environment, and providing decision support for seabed resource exploration. Background Technology

[0002] With advancements in deep-sea exploration technology and the accumulation of vast amounts of marine environmental data, the automated analysis and prediction of habitat evolution in hydrothermal vent areas has become a key requirement for deep-sea ecological research and resource exploration. Event chain prediction technology can infer subsequent potential events based on existing event sequences, and has significant application value in scenarios such as identifying biological activity patterns in hydrothermal vent areas, early warning of hydrothermal vent activity, and extrapolation of seabed ecological environment evolution.

[0003] Currently, most common event chain prediction methods are based on sequence modeling techniques, such as recurrent neural networks, graph neural networks, or pre-trained language models, primarily relying on event context information to learn event transition patterns. However, in complex habitats such as deep-sea hydrothermal vents, existing methods have the following significant limitations: Insufficient semantic understanding of events: Models often only perform pattern matching on the surface of the text, making it difficult to distinguish between events with similar semantics but different ecological meanings (such as "sulfide deposition" and "sulfur-oxidizing bacteria aggregation"). Errors tend to accumulate in multi-step predictions, and there is a lack of explicit constraints on the logical rationality of events, causing the prediction path to deviate from the actual ecological process.

[0004] Weak knowledge integration capability: Hydrothermal vents involve knowledge from multiple disciplines such as geology, chemistry, and biology, and take various forms (such as structured maps, literature descriptions, and observation reports). Existing methods lack a unified representation and fusion mechanism for multi-source heterogeneous knowledge, making it difficult to effectively introduce domain knowledge to support event reasoning.

[0005] Missing event-knowledge associations: The deep semantic associations between external knowledge and event chains are not fully captured, especially the spatiotemporal specificity and environmental dependence of hydrothermal events. Existing models struggle to organically combine domain knowledge with event sequence and habitat context, which limits the reliability and interpretability of prediction results.

[0006] Therefore, when faced with complex scenarios such as hydrothermal vent habitats where multiple factors intertwine, existing event chain prediction methods have significant shortcomings in semantic understanding, knowledge fusion, and event-knowledge interaction. There is an urgent need for a new method that can embed domain knowledge, enhance event semantic modeling, and improve the reasoning ability for complex habitat events. Summary of the Invention

[0007] This invention addresses the shortcomings of existing event chain prediction methods in semantic understanding and knowledge fusion by providing a hydrothermal vent habitat event chain prediction method based on deep-sea domain knowledge enhancement. This method introduces a deep-sea domain knowledge base to construct knowledge-enhanced event representations and combines self-attention mechanisms and reinforcement learning to effectively improve the accuracy and interpretability of hydrothermal vent habitat event chain prediction.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A method for predicting hydrothermal vent habitat event chains based on deep-sea domain knowledge enhancement includes the following steps: Step S1: Perform entity recognition and disambiguation on the input event chain and candidate event set, and obtain relevant descriptive information from the deep-sea domain knowledge base through entity links; Step S2: Construct a component graph of the event, with entities as nodes and predicate relations as edges, and integrate entity type matching degree and description semantic similarity to complete entity linking; Step S3: Train the event composition component graph based on graph representation learning methods to obtain entity and relation vector representations that integrate domain knowledge; Step S4: Semantically enhance the event chain and candidate event set through a self-attention mechanism to generate a knowledge-level event state representation; Step S5: Combine the knowledge-level event state representation with the chain-level event state representation based on time-series modeling to form a comprehensive event state representation; Step S6: Introduce a knowledge-level reward function into the reinforcement learning framework and combine it with chained rewards to optimize the event prediction process.

[0009] As an optional implementation, step S1 includes: identifying and disambiguating entities in the event text, generating contextual lexical embeddings using a pre-trained language model, and labeling entity boundaries using the BIO tagging method.

[0010] As an optional implementation, step S2 includes: performing average pooling on the entity lexical embedding to obtain an entity vector representation; calculating a type matching score based on the Euclidean distance between the entity vector and the candidate type vector; and combining the descriptive semantic score and the global entity priority to generate the final entity link score.

[0011] According to one aspect of the present invention, the formula for calculating the entity type score is as follows: The formula for calculating the entity type score is as follows: The model uses a linear layer and an activation function. To calculate the score. The input includes: entity embedding vectors. , feature vectors of candidate types ,and This refers to the set of candidate entities belonging to the corresponding type in the database.

[0012] Entity description score is obtained through a formula Calculation, where linear layer parameters and They are independent of each other.

[0013] The final entity link score is composed of the type score. Description score and global priority via linear layer It is obtained by weighted combination, and the calculation formula is: .

[0014] According to one aspect of the present invention, step S3 includes: Align and bind the identified event entities with the corresponding entities in the external knowledge base; Obtain the description text of the entity, encode it, and generate the initialization vector of the entity node; Construct an event component graph based on entity nodes and predicate relation edges; We use a graph representation learning model to embed the graph into the graph and optimize the vector representation of nodes and edges to enhance the semantic association of knowledge.

[0015] According to one aspect of the invention, the graph representation learning model employs the TransE model; during vector optimization, the TransE model is trained based on the following loss function: ,in,( , , ) represents a correct triple in the graph, consisting of the head entity, the relation, and the tail entity. , , () represents negative sample triples generated by replacing the head or tail entity. A component map of the event, This is a marginal parameter.

[0016] According to one aspect of the invention, step S4 includes: The vector representations of the subject, predicate, and object of the event are concatenated to form the initial knowledge-level vector representation of the event; The semantic associations between events in the event chain and between the event chain and the candidate event set are modeled using the self-attention mechanism to obtain semantically enhanced event representations. Average pooling is performed on the semantically enhanced event representation to generate a global knowledge-level event state representation.

[0017] According to one aspect of the invention, the initial knowledge-level vector representation of an event is generated by the following formula: ,in, A vector representation of the event subject. Vector representation of the object of an event. A vector representation of an event predicate; The model calculation involves the following two core formulas: Self-attention mechanism: Where Q (query), K (key), and V (value) are input vectors. It is the dot product normalization parameter used for scaling; Average pooling operation: Where m and n are the sequence lengths of the event chain and the candidate event set, respectively. and These are the enhanced event representations in the event chain and the candidate event set, respectively.

[0018] According to one aspect of the present invention, step S6 includes: Calculate the cosine similarity between the predicted event sequence and the target event sequence in the vector space, and use it as the event-level reward; The event-level rewards are weighted and fused with the chain rewards that reflect the coherence of the event chain structure to form a comprehensive reward signal; Based on the comprehensive reward signal, the parameters of the event prediction model are updated using a policy gradient algorithm to optimize the decision-making strategy for event selection.

[0019] According to one aspect of the present invention, the event-level reward is calculated using the following formula: ,in, The mean of a knowledge representation vector representing a sequence of predicted events. The mean of the knowledge representation vector of the target sequence. and They are transposes of each other; The final reward is obtained using the following weighted formula: The total reward consists of chain rewards. and event-level rewards We get the weighted sum. These are preset hyperparameters used to balance the contributions of both.

[0020] The policy gradient method aims to maximize the expected return, and its parameter update direction is as follows: ,in, As a discount factor, Rewards for each step, It is the strategy probability. gradient, It is the predicted probability value, which ultimately leads to the state. , This represents the amount of time the model parameters are updated.

[0021] According to one aspect of the present invention, in step S5, the chain-level event occurrence state representation is obtained by performing sequence modeling of the event chain using a time-series model; the chain-level event occurrence state representation is then... Representation of the occurrence state of knowledge-level events Vector concatenation is performed to generate the final representation of the event occurrence state. The concatenation formula is as follows: ,in, This indicates a vector concatenation operation.

[0022] The advantages of this invention are mainly reflected in the following aspects: First, by introducing deep-sea domain knowledge and constructing knowledge-enhanced event state representations, it enhances the deep understanding of the semantics of hydrothermal vent habitat events, effectively distinguishing events with similar semantics but different ecological meanings (such as "sulfide deposition" and "hydrothermal plume diffusion"), and significantly improving the accuracy of habitat event prediction. Second, by employing a TransE-based graph representation learning method, it unifies multi-source heterogeneous deep-sea knowledge (such as geological maps, species distribution, chemical parameters, etc.) into low-dimensional vectors, effectively solving the problem of knowledge heterogeneity. The method achieves deep fusion and structured modeling of multi-source knowledge. Furthermore, by combining chain-based and knowledge-level reward reinforcement learning mechanisms, it simultaneously considers the temporal coherence and semantic rationality of event chains during prediction, effectively suppressing error accumulation in long-sequence predictions and improving the stability and interpretability of predictions of complex ecological processes in hydrothermal vent areas. Finally, based on a general knowledge enhancement mechanism, this method does not rely on a large amount of labeled data, possesses good generalization ability, and can be transferred to various deep-sea habitat event reasoning scenarios such as hydrothermal vent activity prediction and deep-sea biological community evolution analysis. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating the hydrothermal vent habitat event chain prediction method based on deep-sea domain knowledge enhancement of the present invention. Figure 2 This is a schematic diagram of the framework of the hydrothermal vent habitat event chain prediction method based on deep-sea domain knowledge enhancement of the present invention; Figure 3 A schematic diagram illustrating the construction and training process of the event component component graph in this invention is shown; Figure 4 The implementation process of extracting knowledge-level event state representations through a self-attention mechanism is described; Figure 5 A schematic diagram illustrating an example of applying this invention to the prediction of hydrothermal vent activity is provided; Figure 6 The illustration shows a case study of using this invention to extrapolate events in the evolution of deep-sea biological communities. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Figure 1 and Figure 2 The overall process and model architecture of the hydrothermal vent habitat event chain prediction method based on deep-sea domain knowledge enhancement described in this invention are presented. The method mainly includes the following steps: Step S1: Entity Links and Terminology Matching This step aims to identify entities in the event text and accurately associate them with corresponding entities in the deep-sea domain knowledge base, providing a foundation for subsequent knowledge fusion. Specifically, it includes the following sub-steps: Step S1.1: Entity disambiguation and description information retrieval First, entities appearing in the input event chain and candidate event set are identified and disambiguated. In the context of hydrothermal vent habitats, entities may involve geological structures, biological species, chemical substances, etc., requiring clear identification of entity references within the context, such as distinguishing between similar entities like "hydrothermal vent" and "sulfide chimney." After disambiguation, structured or unstructured descriptive information of the corresponding entities, such as entity attributes and ecological roles, is retrieved from the deep-sea domain knowledge base to support subsequent knowledge embedding.

[0027] Step S1.2: Lexical Encoding and Context Embedding Generation A pre-trained language model is used to encode lexical units in event text, generating lexical embedding vectors that contain contextual information. This model can capture the semantic relationships of lexical units in event descriptions, improving the ability to represent hydrothermal vent area terminology.

[0028] Step S1.3: Entity boundary recognition based on BIO tags The BIO (Begin-Origin) tagging method is used to indicate whether a term is an entity component: B (Begin) indicates the starting term of an entity, I (Inside) indicates a term inside the entity, and O (Outside) indicates a non-entity term. This tagging clarifies entity boundaries and provides a basis for subsequent entity-level vector aggregation.

[0029] Step S2: Constructing the event composition component graph.

[0030] This step aims to transform entity semantic information into structured graph data, providing input for subsequent graph representation learning. Specifically, it includes the following sub-steps: Step S2.1: Generation of entity representation Based on the BIO tagging results in step S1.3, entity tokens tagged with "B" and "I" are selected, and their token embedding vectors are subjected to average pooling to generate entity vector representations of uniform dimension, denoted as e.

[0031] Step S2.2: Calculate Entity Type Score Calculate the entity vector e and the candidate entity type feature vector c in the deep-sea domain knowledge base. i The Euclidean distance between the entities is used as the type matching score. The smaller the distance, the higher the semantic similarity between the entities and that type.

[0032] Step S2.3: Calculate Entity Description Score The entity description text is encoded into a fixed-dimensional description embedding vector through an attention mechanism. The entity vector and the description embedding are mapped to the same vector space using a linear transformation. The dot product of the two is calculated as the description semantic association score.

[0033] Step S2.4: Final score fusion of entity links By integrating entity type scores, description scores, and global entity priorities based on statistical information, the final score of entity links is calculated through a linear layer. The candidate entity with the highest score is selected to complete the link, providing accurate entity correspondences for the construction of the event component graph.

[0034] Step S3: Knowledge Vector Acquisition Based on Graph Representation Learning This step aims to transform deep-sea domain knowledge into a structured vector representation, providing support for knowledge enhancement of event states. The specific process is as follows: Step S3.1: Associating and binding event entities with the domain knowledge base Based on the entity linking results of step S2.4, the entities identified in the event (such as "hydrothermal vent" and "chemosynthetic bacteria") are precisely bound to the corresponding entities in the deep-sea domain knowledge base to ensure the accuracy of entity description information extraction.

[0035] Step S3.2: Entity description information retrieval and vectorization Descriptive information (such as geological attributes and ecological functions) of bound entities is retrieved from the knowledge base, and the text description is encoded into a fixed-dimensional vector using a pre-trained language model to complete the numerical representation of the entity's semantic information.

[0036] Step S3.3: Construction of event component graph and vector initialization Using subject and object entities as nodes and predicate relations as edges, an event composition component graph is constructed. For example, for the event "hydrothermal vent releases hydrogen sulfide", "hydrothermal vent" and "hydrogen sulfide" are used as nodes, "release" is used as an edge, and the description vector generated in step S3.2 is assigned to the corresponding nodes and edges to form a structured graph with semantic embedding.

[0037] Step S3.4: TransE-based graph representation learning and vector optimization.

[0038] The goal of this step is to optimize the vector representations of all components in the graph using the TransE model. The core idea of ​​the TransE algorithm is to view the relation vector r as a translation from the head entity vector h to the tail entity vector t. The training objective of the model is defined as minimizing the following marginal loss function: ,in( , , )and( , , The vectors are the entity and relation vectors before and after training, respectively. To construct the event component map, is the marginal parameter. The optimization of this loss function aims to drive the model to adjust all vectors in the graph, so that the semantic relationships defined by them (such as h + r ≈t) are more accurately represented in the vector space.

[0039] Step S4: Fuse the representation of the event occurrence state using a self-attention mechanism. By fusing the knowledge vectors corresponding to the various components of an event, a holistic knowledge representation of the event is formed. The specific method is as follows: First, clarify the basic components of an event: each event Depend on( , , It consists of three parts, all of which have obtained the corresponding knowledge vector representation through step S3.

[0040] Subsequently, the component vectors of the event are merged using a concatenation method: the subject vector is... Object vector and predicate vector By concatenating these elements sequentially, a knowledge level representation vector for the event is formed, the expression of which is:

[0041] For example, if , , If all dimensions are 384, then the concatenated result is... The dimension is 1152. This method can preserve the complete semantic information of each component of the event, avoid information loss caused by improper fusion, and thus provide sufficient event knowledge input for subsequent self-attention mechanisms.

[0042] Step S4.2: Generate attention-enhanced event representations like Figure 4 As shown, a self-attention mechanism is used to capture the semantic association between the event chain and the candidate events, generating an attention-enhanced event representation.

[0043] Figure 4 As shown, this step utilizes a self-attention mechanism to capture the semantic relationships between events, thereby generating attention-enhanced event representations.

[0044] The specific steps are as follows: First, construct the input matrix of the self-attention layer. Then, represent the knowledge of the n events contained in the event chain. The knowledge representation of the m events contained in the candidate event set. The components are concatenated to form a matrix of dimension (n+m)×d, where d is the dimension representing the event.

[0045] Then, the self-attention transformation formula is executed as follows: Where Q is the query vector, K is the key vector, and V is the value vector. These are normalization parameters; The average pooling formula is: Where m and n represent the lengths of the event chain and the candidate event set, respectively. For attention enhancement representation of events in an event chain, This is an attention-enhanced representation of candidate events. For example, the candidate events "active hydrothermal vents" and "mineral deposition forming chimneys" are closely related due to common sense, so their attention weights are higher. and It will strengthen the semantic features that are related to each other.

[0046] Step S4.3: Generate a knowledge-level event occurrence status representation.

[0047] To obtain a compact representation containing global event information, this step applies average pooling to all event representations after attention enhancement, aiming to aggregate key features to generate the final knowledge-level state representation.

[0048] First, summarize the event chain obtained from step S4.2. With candidate event set All vectors are used to comprehensively cover relevant semantic information. These features are then aggregated using average pooling. / Generate a vector with invariant dimensions. The formula is It serves as a comprehensive knowledge-level state representation.

[0049] Step S5: Integrate knowledge-level and chain-level event state representations.

[0050] This step aims to integrate event state representations from different dimensions to form a final representation that simultaneously incorporates external common-sense semantics and event chain temporal features. Specifically, the input includes two parts: first, the knowledge-level event state representation obtained in step S4. The first is generated by concatenating vectors of event components, interacting through a self-attention mechanism, and performing average pooling, thus implying common-sense semantic relationships between events; the second is a chain-level event state representation obtained based on existing temporal modeling methods (such as temporal neural networks). The focus is on characterizing the temporal dependencies and evolution patterns within the event chain. Subsequently, a vector concatenation operation (denoted as ⊕) is used to directly connect the two, thereby generating an event state representation that incorporates external knowledge. Its calculation expression is: .

[0051] Step S6: Introduce knowledge-level constraints into the reinforcement learning reward function.

[0052] This step aims to optimize the model's prediction strategy through a composite reward mechanism, combining the temporal logic of the event chain with external common-sense semantics, thereby guiding the model to generate more reasonable candidate events. The specific steps are as follows: Step S6.1: Calculate the event-level reward.

[0053] The reward This is used to evaluate the semantic consistency between the predicted sequence and the true target sequence from a common-sense knowledge perspective. First, the two sequences are processed separately: for the predicted event sequence output by the model, the knowledge representation vector of each event is extracted and its average value is calculated, denoted as . For manually labeled real target sequences, the knowledge representation vectors of their events are also extracted and their average values ​​are calculated to obtain... and transpose it to .

[0054] Event-level rewards Cosine similarity is defined as the cosine similarity between the final state representations of the predicted sequence and the target sequence. The formula is: This reward directly reflects the quality of the model's predictions at the common-sense level: a logical prediction will receive a higher reward, while a prediction that violates common sense will receive a lower reward, thus providing clear semantic rationality guidance for model optimization.

[0055] Step S6.2: Calculate the final reward value.

[0056] This step will use chain rewards (reflecting temporal continuity). Event-level rewards (reflecting common sense and rationality) By combining these factors, a composite reward function that takes into account both temporal and semantic aspects can be constructed. The specific integration method is as follows: Set weight coefficients (This can be adjusted according to task requirements), calculate the weighted sum of the two types of rewards using the following formula:

[0057] For example, if Set to 0.5, for chain rewards It is 0.8, event-level reward. When it is 0.9, the final reward =0.8 + 0.5 × 0.9 = 1.25; when chain rewards It's 0.8, but the event-level reward... When it is 0.3, then =0.8 + 0.5 × 0.3 = 0.95. This mechanism can effectively balance the influence of the two types of rewards, and strengthen the guiding role of semantic rationality in model training while ensuring temporal continuity.

[0058] Step S6.3: Model loss calculation and parameter update.

[0059] This step is based on the final reward signal. The policy gradient algorithm of reinforcement learning is used to update the model parameters, aiming to optimize the prediction strategy for events.

[0060] The update objective is set as follows: improve the model's performance on high-reward events in the current state. This predictive tendency enhances the consistency of time sequence and the rationality of common sense.

[0061] The specific update method is as follows: For each prediction step from t+1 to t+k, consider its corresponding reward. And introduce discount factors (0 < γ < 1) to adjust the weights of long-term rewards. Then, the log probability gradient of the predicted events at each step is calculated. Parameter update volume It can be expressed by the following formula: Through iterative optimization, the model will gradually adjust its parameters. This makes the generated event sequences more consistent and reasonable in terms of timing and common sense, thereby improving the overall prediction performance and stability.

[0062] Figure 5 This paper demonstrates the predictive performance of the event chain prediction method proposed in this application in a case study of habitat evolution in hydrothermal vent areas, comparing it with other methods. The case study is based on the event chain: "active hydrothermal vents → release of hydrogen sulfide and metal sulfides → aggregation of chemosynthetic microbial communities → formation of deep-sea mussel and tube worm communities → increased biodiversity in hydrothermal vent areas → stabilization of the vent ecosystem." The candidate event set includes "weakened hydrothermal vent activity," "mineral deposition forming chimneys," "changes in the bottom water's chemical environment," "expansion of thermophilic microbial populations," "migration of macrobenthic animals," and "intensified competition for ecological niches in the vent area."

[0063] The figure illustrates the complete reasoning path of this method, which involves a series of steps including entity linking, event component component subgraph construction, TransE graph representation learning, self-attention semantic fusion, and reinforcement learning multi-reward constraints, ultimately selecting "mineral deposition forming chimney structures" and "thermophilic microbial population expansion" as the optimal subsequent events. This process intuitively demonstrates the crucial role of external deep-sea domain knowledge (such as "mineral deposition promotes vent structure stability" and "thermophilic microorganisms have a competitive advantage in high-temperature hydrothermal environments") in eliminating unreasonable candidate events (such as "weakened hydrothermal vent activity") and selecting events that conform to the logic of ecological evolution.

[0064] The comparative results show that our method outperforms other comparative methods (FEEL, MRSL and IE-RoBERTa) in terms of predictive reasonableness and semantic consistency, verifying its effectiveness and practicality in reasoning about complex habitat events.

[0065] Figure 6 This figure compares the effectiveness of the hydrothermal vent habitat event chain prediction method based on deep-sea domain knowledge enhancement proposed in this application with other methods in a case study of hydrothermal vent ecosystem evolution. The figure comprises three core parts: the narrative event chain, the candidate event set, and a comparison of the prediction results of different methods. The narrative event chain consists of "increased hydrothermal vent activity → abnormal aggregation of chemosynthetic bacteria → increased attachment density of deep-sea mussel larvae → intensified competition among tube worm populations → peak biodiversity in the vent area → sudden decrease in hydrothermal fluid flux → vent ecosystem entering a period of decline." The candidate event set consists of "continuously rising vent temperature, accelerated sulfide deposition rate, abrupt changes in bottom water chemical composition, significantly reduced bacterial metabolic activity, the migration of new predator populations, and further enhancement of biodiversity in the vent area."

[0066] This figure clearly demonstrates the differences in prediction results between the proposed method and three mainstream methods: FEEL, MRSL, and IE-RoBERTa. The proposed method predicts "significantly reduced bacterial metabolic activity and accelerated sulfide deposition rate," while FEEL, MRSL, and IE-RoBERTa incorrectly select events such as "further enhanced biodiversity in the vent area," "immigration of new predator populations," and "mutation of bottom water chemical composition," which contradict common sense in deep-sea ecology or are inconsistent with the development trend of the event chain. The comparison clearly highlights the semantic rationality and predictive accuracy advantages of the proposed method in assessing the evolution of hydrothermal vent habitat events after incorporating deep-sea knowledge.

[0067] Meanwhile, the hydrothermal vent habitat event chain prediction method proposed in this application, based on deep-sea domain knowledge enhancement, outperforms mainstream methods in both single-step prediction and event chain prediction tasks. As shown in Table 1, in the single-step prediction task, IE-RoBERTa, a leading mainstream method, achieves accuracies of 60.59%, 55.11%, 51.32%, and 50.19% on the NYT, Finance, ICEWS14, and ICEWS15 datasets, respectively. In contrast, our method (Ours) achieves accuracies of 76.23%, 74.39%, 72.11%, and 70.09% on the same datasets, representing an improvement of 15.64 to 20.9 percentage points over IE-RoBERTa. Even with a simplified version (Ours without external knowledge), its accuracy is still higher than all mainstream methods, reaching 68.74% on the NYT dataset, 8.15 percentage points higher than IE-RoBERTa, demonstrating the comprehensive advantages of our method in terms of infrastructure and common-sense knowledge enhancement.

[0068] In the event chain prediction task, the mainstream method FEEL achieved accuracies of 64.98%, 60.54%, 67.83%, and 65.51% on four datasets, respectively, while MRSL yielded corresponding results of 63.26%, 59.76%, 58.06%, and 60.66%. Our proposed method (Ours) achieved accuracies of 71.76%, 70.39%, 68.97%, and 69.35%, representing an improvement of 1.14 to 9.85 percentage points compared to FEEL. Furthermore, our method exhibited less performance fluctuation across different datasets; for example, the difference between ICEWS14 and ICEWS15 was only 0.38 percentage points, indicating superior stability.

[0069]

[0070] Table 1 The advantages of this invention are mainly reflected in the following aspects: First, by introducing deep-sea domain knowledge and constructing knowledge-enhanced event state representations, it enhances the deep understanding of the semantics of hydrothermal vent habitat events, effectively distinguishing events with similar semantics but different ecological meanings (such as "sulfide deposition" and "hydrothermal plume diffusion"), significantly improving the accuracy of habitat event prediction; Second, by employing a TransE-based graph representation learning method, it unifies multi-source heterogeneous deep-sea knowledge (such as geological maps, species distribution, chemical parameters, etc.) into low-dimensional vectors, effectively solving the problem of knowledge heterogeneity. First, it addresses the issue of sexuality, achieving deep fusion and structured modeling of multi-source knowledge. Second, by combining chain-based and knowledge-level reward reinforcement learning mechanisms, it simultaneously considers the temporal coherence and semantic rationality of event chains during the prediction process, effectively suppressing error accumulation in long-sequence predictions and improving the stability and interpretability of predictions of complex ecological processes in hydrothermal vent areas. Finally, based on a general knowledge enhancement mechanism, this method does not rely on a large amount of labeled data, has good generalization ability, and can be applied to various deep-sea habitat event reasoning scenarios such as hydrothermal vent activity prediction and deep-sea biological community dynamic analysis.

[0071] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting hydrothermal vent habitat event chains based on deep-sea domain knowledge enhancement, characterized in that, Includes the following steps: Step S1: Perform entity recognition and disambiguation on the input event chain and candidate event set, and obtain the description information of the corresponding entity from the preset knowledge base based on the entity link; Step S2: Construct a component graph of the event, where nodes represent entities, edges represent predicate relationships between entities, and calculate entity link scores based on entity type matching degree and description semantic similarity. Step S3: The event components are trained using a graph representation learning method to obtain entity and relation embedding vectors that incorporate external knowledge. Step S4: Semantic enhancement is performed on the event chain and candidate event set through a self-attention mechanism to generate knowledge-enhanced event state representations. Step S5: Concatenate the knowledge-enhanced event state representation with the chain-level event state representation to form a comprehensive event state representation; Step S6: Introduce a knowledge-level reward function into the reinforcement learning framework and combine it with a chain-structured reward to jointly optimize the event prediction strategy.

2. The method for predicting hydrothermal vent habitat event chains based on deep-sea domain knowledge enhancement according to claim 1, characterized in that, Step S1 includes: Identify and disambiguate entities appearing in the event text; The input lexical units are encoded using a pre-trained language model to generate context-aware lexical embedding vectors. The BIO annotation system is used to annotate the entity boundaries of lexical units.

3. The method for predicting hydrothermal vent habitat event chains based on deep-sea domain knowledge enhancement according to claim 1, characterized in that, Step S2 includes: Average pooling is performed on the embedding vectors of the words contained in the entity to generate a fixed-dimensional entity representation; Calculate the entity type matching score based on the Euclidean distance between the entity representation and the candidate type vector; Based on the dot product operation of entity representation and description text embedding, the semantic similarity score of entity description is calculated; By combining entity type score, description score, and global entity priority, a final discrimination score for entity links is generated.

4. The method for predicting hydrothermal vent habitat event chains based on deep-sea domain knowledge enhancement according to claim 3, characterized in that, The formula for calculating the entity type is: The model uses a linear layer and an activation function. To calculate the score. The input includes: entity embedding vectors. , feature vectors of candidate types ,and This refers to the set of candidate entities belonging to the corresponding type in the database. Entity description score is obtained through a formula Calculation, where linear layer parameters and They are independent of each other. The final entity link score is composed of the type score. Description score and global priority via linear layer It is obtained by weighted combination, and the calculation formula is: .

5. The method for predicting hydrothermal vent habitat event chains based on deep-sea domain knowledge enhancement according to claim 1, characterized in that, Step S3 includes: Align and bind the identified event entities with the corresponding entities in the external knowledge base; Obtain the description text of the entity, encode it, and generate the initialization vector of the entity node; Construct an event component graph based on entity nodes and predicate relation edges; We use a graph representation learning model to embed the graph into the graph and optimize the vector representation of nodes and edges to enhance the semantic association of knowledge.

6. The method for predicting hydrothermal vent habitat event chains based on deep-sea domain knowledge enhancement according to claim 5, characterized in that, The graph representation learning model employs the TransE model; during vector optimization, the TransE model is trained based on the following loss function: ,in,( , , ) represents a correct triple in the graph, consisting of the head entity, the relation, and the tail entity. , , () represents negative sample triples generated by replacing the head or tail entity. A component map of the event, This is a marginal parameter.

7. The method for predicting hydrothermal vent habitat event chains based on deep-sea domain knowledge enhancement according to claim 1, characterized in that, Step S4 includes: The vector representations of the subject, predicate, and object of the event are concatenated to form the initial knowledge-level vector representation of the event; The semantic associations between events in the event chain and between the event chain and the candidate event set are modeled using the self-attention mechanism to obtain semantically enhanced event representations. Average pooling is performed on the semantically enhanced event representation to generate a global knowledge-level event state representation.

8. The method for predicting hydrothermal vent habitat event chains based on deep-sea domain knowledge enhancement according to claim 7, characterized in that, The initial knowledge-level vector representation of the event is generated using the following formula: ,in, A vector representation of the event subject. Vector representation of the object of an event. A vector representation of an event predicate; The model calculation involves the following two core formulas: Self-attention mechanism: Where Q (query), K (key), and V (value) are input vectors. It is the dot product normalization parameter used for scaling; Average pooling operation: Where m and n are the sequence lengths of the event chain and the candidate event set, respectively. and These are the enhanced event representations in the event chain and the candidate event set, respectively.

9. The method for predicting hydrothermal vent habitat event chains based on deep-sea domain knowledge enhancement according to claim 1, characterized in that, Step S6 includes: Calculate the cosine similarity between the predicted event sequence and the target event sequence in the vector space, and use it as the event-level reward; The event-level rewards are weighted and fused with the chain rewards that reflect the coherence of the event chain structure to form a comprehensive reward signal; Based on the comprehensive reward signal, the parameters of the event prediction model are updated using a policy gradient algorithm to optimize the decision-making strategy for event selection.

10. The method for predicting hydrothermal vent habitat event chains based on deep-sea domain knowledge enhancement according to claim 9, characterized in that: The event-level reward is calculated using the following formula: ,in, The mean of a knowledge representation vector representing a sequence of predicted events. The mean of the knowledge representation vector of the target sequence. and They are transposes of each other; The final reward is obtained using the following weighted formula: The total reward consists of chain rewards. and event-level rewards We get the weighted sum. These are preset hyperparameters used to balance the contributions of both. The policy gradient method aims to maximize the expected return, and its parameter update direction is as follows: ,in, As a discount factor, Rewards for each step, It is the strategy probability. gradient, It is the predicted probability value, which ultimately leads to the state. , This represents the amount of time the model parameters are updated.

11. The method for predicting hydrothermal vent habitat event chains based on deep-sea domain knowledge enhancement according to any one of claims 1 to 10, characterized in that, In step S5: the chain-level event occurrence state representation is the result obtained by performing sequence modeling on the event chain using a time-series model; the chain-level event occurrence state representation is... Representation of the occurrence state of knowledge-level events Vector concatenation is performed to generate the final representation of the event occurrence state. The concatenation formula is as follows: ,in, This indicates a vector concatenation operation.

Images