A marine acoustic knowledge graph construction method and a large language model intelligent question answering system

Abstract

The application provides a marine acoustic knowledge graph construction method and a large language model intelligent question and answer system, and belongs to the technical field of marine engineering and underwater acoustic engineering; the marine acoustic intelligent knowledge system is constructed through the process of "data preprocessing-content extraction-graph structured expression-large language model knowledge calling", the method can systematically integrate and standardize the expression of marine acoustic knowledge, and support the large language model to realize the intelligent question and answer based on the marine acoustic knowledge through the knowledge graph, and is suitable for the scenes of marine acoustic intelligent question and answer, professional knowledge retrieval and auxiliary decision making, etc. The application can significantly reduce the labor cost, realize the structured management of marine acoustic knowledge, thereby improving the knowledge collation and utilization efficiency, and effectively reducing the error explanation, so that the analysis result is more accurate.

Description

Technical Field

[0001] This invention belongs to the field of marine engineering and underwater acoustic engineering technology, and mainly relates to a method for constructing a marine acoustic knowledge graph and an intelligent question-answering system based on a large language model. Background Technology

[0002] Marine acoustics research encompasses a wide range of complex concepts, including marine environmental parameters, sound field propagation characteristics, array signal processing methods, and underwater target detection mechanisms. It is characterized by its high level of specialization, numerous terms, diverse entity types, and complex physical relationships. Traditional underwater acoustic knowledge mining relies heavily on manual literature review and conceptualization, which is not only inefficient and costly but also struggles to maintain consistency, systematicity, and reusability as data scales up. Furthermore, while large language models possess strong language understanding and generation capabilities, they lack verifiability and structured expression mechanisms in specialized fields. Relying solely on their generated results can easily lead to unstable or even erroneous inferences; therefore, supplementation and enhancement with a structured knowledge system are necessary. Existing technologies primarily suffer from the following problems: (1) Knowledge graphs (KGs) play a prominent role in semantic standardization, prior constraints, and interpretable reasoning. By structurally representing environmental factors, acoustic mechanisms, propagation characteristics, simulation parameters, and evidence sources, knowledge graphs can provide clear, reliable, and traceable domain knowledge support for large language models, thereby significantly improving the stability and interpretability of the reasoning process. However, there is currently a lack of dedicated knowledge graphs for marine acoustics, making it difficult for large language models to obtain systematic professional knowledge support in this field.

[0003] (2) While large language models (LLMs) possess strong language understanding and generation capabilities, they lack verifiability and constraint mechanisms for professional knowledge. In highly specialized fields such as marine acoustics, where terminology accuracy is crucial, relying solely on LLMs to generate answers can easily lead to conceptual confusion or fabricated content. Therefore, structured knowledge graphs are needed to provide factual evidence and professional knowledge, thereby enabling the question-and-answer process to have an interpretable reasoning basis and traceability.

[0004] (3) The pain points of difficulty in interpreting acoustic simulation results and reliance on experience. Marine acoustic simulation (such as sound field simulation, coherence simulation, noise simulation, and optimization of mooring deployment locations) usually involves a large number of environmental parameters, terrain conditions, boundary conditions, and operational configurations. The calculation process is complex and the parameter coupling relationship is strong. The simulation output results are diverse, mainly including spatial and physical characteristics such as sound pressure field distribution, interference fringes, propagation path, energy convergence zone, and sound shadow zone. The analysis and interpretation of the above simulation results usually rely heavily on domain experience. There are problems such as complex influencing factors and large differences in assumptions between different simulation models, making it difficult to describe and understand the simulation results in a unified and standardized way. At the same time, the simulation results lack a structured connection with relevant acoustic theories, propagation mechanisms, and model assumptions, making it difficult for large language models to reliably analyze and interpret them without structured knowledge constraints.

[0005] Currently, there is a lack of a system that can structurally link simulation input parameters, simulation result characteristics, and marine acoustic theory knowledge. As a result, the analysis and interpretation of marine acoustic simulation results lack reliable basis, making it difficult to achieve systematic and reasonable intelligent analysis. Summary of the Invention

[0006] To overcome the problems of knowledge dispersion, retrieval difficulties, and lack of basis for large language model answers in existing methods of acquiring marine acoustic knowledge, this invention proposes a method for constructing a marine acoustic knowledge graph and an intelligent question-answering system based on a large language model. This invention constructs a marine acoustic intelligent knowledge system through a process of "data preprocessing—content extraction—graph-structured representation—large language model knowledge invocation." This method can systematically integrate and standardize marine acoustic knowledge, and supports intelligent question answering based on marine acoustic knowledge through a knowledge graph and a large language model. It is applicable to scenarios such as intelligent question answering in marine acoustics, professional knowledge retrieval, and decision support.

[0007] The technical solution adopted by this invention to solve its technical problem is a method for constructing a marine acoustic knowledge graph, which includes the following steps: Step S1: Collect marine acoustic text, perform unified preprocessing on the collected marine acoustic text, and obtain raw marine acoustic corpus; Step S2: Construct a sentence segmentation and annotation dataset based on the original marine acoustic corpus. Divide the sentence segmentation and annotation dataset into a component training set, a component validation set, and a component test set. Use the component training set to train the marine acoustic Chinese word segmentation model CWS to obtain the optimal marine acoustic Chinese word segmentation model. The input of the marine acoustic Chinese word segmentation model is the component training set; the output of the marine acoustic Chinese word segmentation model is a text character sequence; the text character sequence corresponds to the BMES tag sequence, where B, M, E, and S represent the beginning, middle, end, and single-character word formation positions of words, respectively. Step S3: Construct a term dataset based on the original marine acoustic corpus, and divide the term dataset into a term training set, a term validation set, and a term test set; use the term training set to train the marine acoustic entity recognition model NER to obtain the optimal marine acoustic entity recognition model NER; The input to the marine acoustic entity recognition model (NER) is a terminology dataset; the output of the NER is a terminology character sequence. The terminology character sequence includes entity, entity category, and location information. Entity terminology categories include E-type, R-type, P-type, D-type, L-type, and T-type; location information represents the position of the entity terminology category in the original marine acoustic corpus. The E-type represents entities related to marine environment and propagation structures; the R-type represents entities related to experiments and engineering; the P-type represents entities related to physical parameters; the D-type represents entities related to equipment systems; the L-type represents entities related to location, sea area, and orientation; and the T-type represents entities related to theoretical models and computational methods. The terminology character sequence is optimized, and a terminology thesaurus is established. Step S4: Construct a relation analysis dataset based on the original marine acoustic corpus. Divide the relation analysis dataset into a relation training set, a relation validation set, and a relation test set. Use the relation training set to train the marine acoustic relation extraction model to obtain the optimal marine acoustic relation extraction model. The input of the marine acoustic relation extraction model is the relation analysis dataset. The output of the marine acoustic relation extraction model is a relation triple, which includes a subject entity, a relation predicate, and an object entity. Step S5: Construct a marine acoustics knowledge graph and graph database; A marine acoustic knowledge graph is constructed using entities in all terminology character sequences as graph nodes, relation predicates in relation triples as graph relation edges, and location information and text content as graph attributes. The text content is the input corresponding to the relation triples in the marine acoustic relation extraction model (RE). After the marine acoustic knowledge graph is constructed, all graph nodes and graph relation edges are written into the graph database, and unique identifiers and indexes are set for the graph nodes. Furthermore, the preprocessing process involves removing garbled characters, full-width spaces, invisible control characters, spaces, and line breaks from the marine acoustic text, retaining only the text portion; and separately splitting numerical sequences, English strings, or long technical terms in the marine acoustic text at the character level.

[0008] Furthermore, the step of constructing the sentence segmentation and annotation dataset involves performing sentence component analysis on the original marine acoustic corpus, using the BMES labeling system to annotate the character sequences, and obtaining the sentence segmentation and annotation dataset.

[0009] Furthermore, the marine acoustic Chinese word segmentation model is based on the BERT-LSTM-CRF sequence annotation model structure.

[0010] Furthermore, the steps for obtaining the optimal marine acoustic Chinese word segmentation model are as follows: training the marine acoustic Chinese word segmentation model using a component training set, validating the training process using a component validation set, and finally evaluating the performance of the optimal model using a component test set. When the word segmentation model evaluation index is greater than 70%, the optimal marine acoustic Chinese word segmentation model is obtained. The word segmentation model evaluation index is classification precision, classification recall, and F1 score. Furthermore, the process of constructing the term dataset is as follows: perform term analysis on the original marine acoustic corpus, label the terms and term categories in the original marine acoustic corpus, and obtain the term dataset. Furthermore, the marine acoustic entity recognition model NER is based on the BERT-LSTM-CRF sequence labeling model structure.

[0011] Furthermore, the step of obtaining the optimal marine acoustic entity recognition model (NER) is as follows: the marine acoustic entity recognition model (NER) is trained using a terminology training set, and the optimal marine acoustic entity recognition model (NER) is obtained when the terminology model evaluation index is greater than 70%; the terminology model evaluation index is classification precision, classification recall, and F1 score.

[0012] Furthermore, the steps for constructing the relation analysis dataset are as follows: perform relation analysis on the original marine acoustic corpus, label the relationship types between professional terms in the original marine acoustic corpus, and obtain the relation analysis dataset; divide the relation analysis dataset into a relation training set, a relation validation set, and a relation test set; the relation types include causal relations, generation relations, influence relations, positive correlation relations, negative correlation relations, dependency relations, input relations, output relations, attribute relations, feature relations, co-occurrence evidence relations, simulated evidence relations, composition relations, and inclusion relations.

[0013] Furthermore, the marine acoustic relation extraction model RE is based on the BERT model.

[0014] Furthermore, the process of obtaining the optimal marine acoustic relation extraction model is as follows: the marine acoustic relation extraction model RE is trained using a relation training set, and the optimal marine acoustic relation extraction model RE is obtained when the cross-entropy loss function converges to a preset threshold; the loss function of the marine acoustic relation extraction model RE is the cross-entropy loss function.

[0015] Furthermore, the process of optimizing the terminology character sequence and obtaining a terminology thesaurus is as follows: Unify entities with the same concept in the terminology character sequence into the same entity name; and create a term thesaurus based on the different names of entities with the same concept.

[0016] A large language model intelligent question answering system that applies a marine acoustic knowledge graph construction method includes a marine acoustic knowledge graph, an optimal marine acoustic entity recognition model (NER), an optimal marine acoustic relation extraction model (RE), a large language model, and an intelligent question answering method. The intelligent question-answering method includes the following steps: Step Q1: The user enters a question text to ask a question; Step Q2: Process the received question text sequentially using the optimal marine acoustic entity recognition model (NER) and the optimal marine acoustic relation extraction model (RE). The optimal marine acoustic entity recognition model (NER) obtains the term character sequence in the question text; the optimal marine acoustic relation extraction model (RE) obtains the relation triples in the question text. Step Q3: Using the terminology character sequence of the question text, the relation triples of the question text, and the terminology thesaurus corresponding to the entities, construct a graph database retrieval statement and execute it to obtain the graph nodes that are the same as the entities in the question text and the graph relation edges corresponding to the graph nodes; Step Q4: Use the input large language model to reason about the text content in the graph nodes and graph relation edges to obtain the answer text to the question text.

[0017] The beneficial effects of this invention are: (1) In view of the problem of knowledge dispersion in the literature and technical data in the field of marine acoustics, this invention can identify entities, extract relationships, standardize and construct knowledge graphs from model assumptions, environmental parameters, propagation mechanisms and simulation results in a large number of literatures and data. Compared with the traditional method of relying on manual annotation and manual sorting, it can significantly reduce labor costs, realize the structured management of marine acoustics knowledge, and thus improve the efficiency of knowledge sorting and utilization.

[0018] (2) To address the difficulty of retrieving marine acoustic knowledge, this invention constructs a structured marine acoustic simulation knowledge graph to systematically organize the relationships between professional terms such as models, environmental parameters, and propagation mechanisms. This enables relevant knowledge to be retrieved through semantic query and associative reasoning, which significantly improves knowledge retrieval efficiency compared to traditional keyword-based retrieval methods.

[0019] (3) In response to the problem that the answers from large language models lack evidence, this invention uses marine acoustic knowledge graphs in conjunction with large language models to make the analysis of simulation results and the answers to related questions based on clear propagation mechanisms and theoretical basis. Compared with the method of simply relying on large language models to generate answers, this invention can effectively reduce misinterpretations and make the analysis results more accurate. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall process of a simulation analysis and intelligent question-answering system based on marine acoustic knowledge graphs. Figure 2 Flowchart for constructing a knowledge graph of marine acoustics; Figure 3 The figure shows the CWS, NER, and RE model structures used in this invention, as well as the collaborative question-answering process of the graph and the large language model; Figure 3 (a) is a schematic diagram of the CWS word segmentation model structure; Figure 3 (b) is a schematic diagram of the NER entity recognition model structure; Figure 3 (c) is a schematic diagram of the RE relation extraction model structure; Figure 3 (d) is a flowchart of the question-answering process combining knowledge graph and large language model; Figure 4 A partial example diagram of the marine acoustic knowledge graph constructed by the method of this invention; Figure 5 This is a schematic diagram of the graph relationship edges in the marine acoustic knowledge graph constructed by the method of this invention; Figure 6 This provides an example of question-answering using a large language model augmented with a knowledge graph. Figure 7 An example diagram illustrating the analysis of acoustic simulation results using a large language model; Figure 8 Example diagram of risk analysis of simulation scheme using a large language model; Figure 9 Example diagram showing optimization suggestions for large language models. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] A method for constructing a marine acoustic knowledge graph includes the following steps: Step 1: Collect marine acoustic text, perform unified preprocessing on the collected marine acoustic text to obtain raw marine acoustic corpus; The preprocessing process is as follows: remove garbled characters, full-width spaces, invisible control characters, spaces, and newlines from the marine acoustic text, and retain only the text portion of the marine acoustic text; separate the numerical sequences, English strings, or long technical terms in the marine acoustic text by character level. Step 2: Construct a sentence segmentation and annotation dataset. Divide the sentence segmentation and annotation dataset into a component training set, a component validation set, and a component test set. Use the component training set to train the marine acoustic Chinese word segmentation model CWS to obtain the optimal marine acoustic Chinese word segmentation model. Sentence component analysis was performed on the raw marine acoustic corpus, and the character sequences were labeled using the BMES labeling system to obtain a sentence segmentation and labeling dataset; The sentence segmentation and annotation dataset is divided into a component training set, a component validation set, and a component test set; The marine acoustic Chinese word segmentation model is based on the BERT-LSTM-CRF sequence annotation model structure; The input to the marine acoustic Chinese word segmentation model is the component training set; The output of the marine acoustic Chinese word segmentation model is a text character sequence; the text character sequence corresponds to the BMES tag sequence, where B, M, E, and S represent the beginning, middle, end, and single-character word positions of a word, respectively; The marine acoustic Chinese word segmentation model is trained using the component training set, the training process is validated using the component validation set and the optimal model is selected, and finally the performance of the optimal model is evaluated using the component test set. When the word segmentation model evaluation index is greater than 70%, the optimal marine acoustic Chinese word segmentation model is obtained. The evaluation metrics for word segmentation models are classification precision, classification recall, and F1 score. Step 3: Construct a term dataset, dividing it into a term training set, a term validation set, and a term test set; use the term training set to train the marine acoustic entity recognition model NER to obtain the optimal marine acoustic entity recognition model NER; Terminology analysis was performed on the raw marine acoustic corpus to identify the terms and term categories in the raw marine acoustic corpus, thus obtaining a term dataset; the term dataset was then divided into a term training set, a term validation set, and a term test set. The marine acoustic entity recognition model NER is based on the BERT-LSTM-CRF sequence annotation model structure; The input to the marine acoustic entity recognition model NER is a term dataset; The output of the marine acoustic entity recognition model NER is a term character sequence; the term character sequence includes entity, entity category and location information; entity term categories include E-type, R-type, P-type, D-type, L-type and T-type; the location information is the position of the entity term category in the original marine acoustic corpus; The E-type represents entities related to marine environment and propagation structures; the R-type represents entities related to experiments and engineering; the P-type represents entities related to physical parameters; the D-type represents entities related to equipment systems; the L-type represents entities related to location, sea area, and orientation; and the T-type represents entities related to theoretical models and calculation methods. The marine acoustic entity recognition model NER is trained using a terminology training set. The optimal marine acoustic entity recognition model NER is obtained when the terminology model evaluation index is greater than 70%. The evaluation metrics for the terminology model are classification precision, classification recall, and F1 score. Step 4: Construct a relation analysis dataset, dividing it into a relation training set, a relation validation set, and a relation test set. Use the relation training set to train the marine acoustic relation extraction model to obtain the optimal marine acoustic relation extraction model. The marine acoustic relation extraction model (RE) is based on the BERT model. Relation analysis was performed on the raw marine acoustic corpus to identify the relationship types between technical terms in the corpus, resulting in a relation analysis dataset. The dataset was divided into a training set, a validation set, and a test set. Relation types included causal relationships, generation relationships, influence relationships, positive correlation relationships, negative correlation relationships, dependency relationships, input relationships, output relationships, attribute relationships, feature relationships, co-occurrence evidence relationships, simulated evidence relationships, composition relationships, and inclusion relationships. The input to the marine acoustic relation extraction model is the relation analysis dataset; The output of the marine acoustic relation extraction model is a relation triple, which includes a subject entity, a relation predicate, and an object entity. The loss function of the marine acoustic relation extraction model (RE) is the cross-entropy loss function. The marine acoustic relation extraction model RE is trained using a relation training set. When the cross-entropy loss function converges to a preset threshold, the optimal marine acoustic relation extraction model RE is obtained. Step 5: Optimize the terminology character sequence to ensure the integrity and consistency of entities; The process of optimizing the term character sequence is as follows: Unify entities with the same concept in the terminology character sequence into the same entity name; and build a terminology thesaurus based on the different names of entities with the same concept. For terms lacking entity category labels, the category is inferred based on the text structure and semantic pattern. For example, “South China Sea, sea area, trench” is classified as sea area, “algorithm, model, equation” is classified as theoretical model, and “array, hydrophone, sonar” is classified as equipment system, so as to ensure the integrity and consistency of the entity category system. Step 6: Construction of a marine acoustics knowledge graph.

[0023] The entity in all terminology character sequences is used as a graph node, the relation predicate in the relation triple is used as a graph relation edge, and the location information and text content are used as graph attributes to construct a marine acoustic knowledge graph; the text content is the input corresponding to the relation triple in the marine acoustic relation extraction model RE. After the ocean acoustics knowledge graph is constructed, all graph nodes and graph relationship edges are written into the graph database, and unique identifiers and indexes are set for the graph nodes; A large language model intelligent question answering system that applies a marine acoustic knowledge graph construction method includes a marine acoustic knowledge graph, an optimal marine acoustic entity recognition model (NER), an optimal marine acoustic relation extraction model (RE), a large language model, and an intelligent question answering method. The intelligent question-answering method includes the following steps: Step Q1: The user enters a question text to ask a question; Step Q2: Process the received question text sequentially using the optimal marine acoustic entity recognition model (NER) and the optimal marine acoustic relation extraction model (RE). The optimal marine acoustic entity recognition model (NER) obtains the term character sequence in the question text; the optimal marine acoustic relation extraction model (RE) obtains the relation triples in the question text. Step Q3: Using the terminology character sequence of the question text, the relation triples of the question text, and the terminology thesaurus corresponding to the entities, construct a graph database retrieval statement and execute it to obtain the graph nodes that are the same as the entities in the question text and the graph relation edges corresponding to the graph nodes; Step Q4: Use the input large language model to reason about the text content in the graph nodes and graph relation edges to obtain the answer text to the question text.

[0024] The present invention will now be described using specific embodiments as examples; Step T1: Marine Acoustic Corpus Preprocessing. The collected marine acoustic text undergoes unified preprocessing. First, abnormal symbols, garbled characters, full-width spaces, and invisible control characters generated during text extraction are removed, and various whitespace characters are standardized to a uniform format. Simultaneously, different "character-tag" writing styles in the original corpus are automatically converted to the BMES annotation system, and characters or tags that do not conform to the annotation rules are corrected or removed. Furthermore, for continuous character content such as number sequences, English strings, or long technical terms, they are split at the character level to ensure that subsequent word segmentation, entity recognition, and relation extraction models can learn and process complex marine acoustic text at a consistent granularity.

[0025] Step T2: Chinese Word Segmentation Model Training (CWS). Based on the corpus preprocessing completed in Step T1, a Chinese word segmentation model suitable for the marine acoustics field is trained. The specific method includes: firstly, encoding the preprocessed acoustic text at the character level and adding start and end markers to the beginning and end of the sequence; using a fixed-length input format and introducing a mask vector to distinguish valid characters from padding positions. Subsequently, a word segmentation model structure containing a deep semantic coding layer, a bidirectional sequence modeling layer, and a conditional random field (CRF) global decoding layer is used to model the text sequence to learn the contextual continuity features and word boundary patterns of marine acoustics terminology. During model training, the BMES labeling system is used for sequence annotation, and global optimization is performed through a sequence loss function to obtain high-precision word segmentation results suitable for complex terms, technical vocabulary, and domain-specific long words.

[0026] The Chinese word segmentation model used in this invention is a deep learning-based sequence labeling model, whose basic structure adopts the mature BERT-LSTM-CRF sequence labeling model structure. This invention does not modify the core network structure of this model, but only constructs training data for marine acoustics corpora and performs domain-adaptive training on the model.

[0027] The input to this word segmentation model is a preprocessed sequence of marine acoustic text characters, with each character corresponding to a contextual semantic vector. The model output is a sequence of BMES tags that correspond one-to-one with the input character sequence, where B, M, E, and S represent the beginning, middle, end, and single-character word positions, respectively. The output of the word segmentation model can be decoded from the BMES tags to restore the word sequence, which can then be used as the basic or auxiliary input for subsequent entity recognition and relation extraction processes.

[0028] Step T3: Marine Acoustic Entity Recognition Model Training (NER). Based on the text that has undergone basic processing such as corpus preprocessing and word segmentation, an entity recognition model suitable for the marine acoustics field is constructed. A deep learning-based sequence annotation method is used to automatically identify multiple types of marine acoustic entities from professional text. The main entity categories involved in the marine acoustics field include, but are not limited to, the six categories shown in the table below:

[0029] To adapt to the structural characteristics of multi-category and multi-word technical terms in marine acoustic texts, a labeling system covering multiple entity types was first designed. This system employs a binary labeling structure of "entity category + location information," such as PB / M / E and EB / M / E, to accurately identify the entity category and its boundary position within the text. Based on this, an entity recognition model was constructed. The model consists of a pre-trained language model, a bidirectional sequence modeling layer, and a Conditional Random Field (CRF) global decoding layer. This model can simultaneously model the semantic continuity within entities and the dependencies between entities, effectively handling structurally complex entities commonly found in marine acoustic texts, such as multi-word combined terms, special symbol expressions, and deep-sea region names.

[0030] To adapt to specialized texts of varying lengths, the model employs a fixed-length input and masking mechanism. Sentences are converted into character sequences of uniform length, and positional masks distinguish between valid characters and filled regions. This enables the model to stably process marine acoustic corpora containing long term chains, multi-line descriptions, and irregularly structured text. After entity boundary parsing, the model's output label sequence automatically extracts entity text, entity categories, and their positions within the original sentence, achieving accurate recognition of continuous multi-character specialized terms, compound words with special formats, and entities with long-distance semantic dependencies.

[0031] The named entity recognition model in this invention adopts a sequence labeling structure based on a pre-trained language model. Its basic model is BERT + bidirectional sequence modeling layer + conditional random field (CRF) decoding layer. This invention does not modify the network topology of the model, but only constructs a dedicated entity labeling system for the marine acoustics domain and uses domain-labeled data to train the model.

[0032] The NER model takes a preprocessed sequence of marine acoustic text characters and their corresponding contextual semantic representation as input. The model outputs a sequence of entity labels of the same length as the input character sequence. Each label consists of "entity category + location information," used to identify the entity type and its boundary position. By parsing the entity boundaries of the output label sequence, the resulting entity text, entity category, and its location information within the original sentence are used as input to the relation extraction model for subsequent relation extraction and knowledge graph construction.

[0033] Step T4: Relation Extraction Model Training (RE). Based on the entity recognition results obtained in Step T3, a relation extraction model suitable for the marine acoustics field is constructed to identify semantic relationships between different entities in professional texts. The relation extraction model employs a triple extraction structure based on deep semantic modeling. It performs overall semantic encoding on acoustic sentences containing multiple entities and, under the constraints of entity recognition results, automatically identifies the semantic segments corresponding to the subject, predicate, and object, thereby generating candidate relation triples (S–P–O). The relation extraction model can identify multiple types of key logical relationships from marine acoustics descriptions, including but not limited to the examples shown in the table below:

[0034] To adapt to the diverse forms of relational expressions and inconsistent natural language representations in the field of marine acoustics, after obtaining candidate triples, a marine acoustic predicate library covering multiple types of relations, including influence, causality, dependence, composition, attribute, location, range, and application, is further constructed. By mapping different natural language expressions such as "determined by," "influencing factors," "includes parameters," and "distributed in," which appear in the text, to a unified semantic category, the extracted original predicates are semantically clustered and standardized. This automatically categorizes a large number of natural language predicates into approximately twenty standard relational types, forming a relational labeling system with a unified structure that facilitates reasoning.

[0035] After predicate standardization is completed, the identified entities and standardized relation types are combined to form structured triples (entity A, standard relation type, entity B). At the same time, the original sentence that generated the relation, the original predicate expression, and the context window information of the entity are retained for subsequent interpretable display of the knowledge graph, relation verification, and source analysis, ensuring the reliability of the relation extraction results in the process of automatic reasoning and calling large language models.

[0036] The relation extraction model in this invention is implemented based on a mature deep learning relation extraction framework. Its core network structure remains unchanged; only relation predicate standardization rules are constructed specifically for the marine acoustics domain, and the model is trained to adapt to the domain corpus. The input to the relation extraction model includes the original marine acoustic text sentences, as well as the entity text, entity category, and their position information within the sentence, output by the entity recognition model. The model output is a structured relation triplet result, where each triplet includes at least a subject entity, a relation predicate, and an object entity, and is associated with a corresponding evidence sentence for subsequent knowledge graph construction and interpretable display.

[0037] In the implementation of this invention, independent validation and test datasets were set up for the Chinese word segmentation model, the marine acoustic entity recognition model, and the relation extraction model to objectively evaluate the model training process. Precision, recall, and F1 score were used as evaluation metrics to judge the model training effect. The F1 score is a comprehensive evaluation metric for precision and recall in the model recognition results. Precision reflects the accuracy of the model's prediction results, while recall reflects the model's coverage of the true target. The F1 score, taking both into account, reflects the overall performance level of the model and is therefore suitable for evaluating the performance of these models.

[0038] Step T5: Entity Standardization and Synonym Fusion. Addressing issues such as mixed use of Chinese and English terms, diverse abbreviations, and inconsistent spellings in marine acoustics terminology, this invention unifies entity representations within the knowledge graph by employing an entity standardization and synonym fusion mechanism. The key features include: firstly, constructing a Chinese-English synonym dictionary for marine acoustics terminology, unifying various spellings with the same semantic meaning, such as "sound speed," "sound wave speed," and "sound speed," into standardized names; automatically deduplicating the same entity identified repeatedly in different sentences, assigning a unique identifier to each entity, and maintaining consistent referencing across all relationships in the knowledge graph; and for terms lacking entity category labeling, automatically inferring categories based on their textual structure features and semantic patterns, such as classifying "South China Sea, sea area, trench" as sea area, "algorithm, model, equation" as theoretical model, and "array, hydrophone, sonar" as equipment system, ensuring the integrity and consistency of the entity category system.

[0039] Step T6: Construction of the Marine Acoustics Knowledge Graph. Based on the processing results of steps T2 to T5, a marine acoustics knowledge graph is constructed. Its key features include: creating graph nodes from normalized entities and configuring attributes such as entity name, entity category, and entity number for each node; constructing graph relation edges from normalized triples, where each relation edge contains attributes such as relation type, relation number, source sentence, source document, subject entity type, and object entity type, represented in the following structural form: (Entity A) — [Relation Type / Relation Number] → (Entity B) After the knowledge graph is constructed, all entity nodes and relation edges are written into the graph database, and uniqueness constraints and efficient indexes are set for the nodes to ensure that the knowledge graph has characteristics such as clear structure, efficient query, semantic interpretability, scalability and reasonability, thereby supporting subsequent application scenarios such as intelligent question answering, large language model invocation and graph-based semantic retrieval.

[0040] Based on the aforementioned marine acoustics knowledge graph, the input marine acoustics domain question is analyzed, retrieved from the knowledge graph, and augmented with a large language model. This process includes the following steps: Step P1: Question Analysis and Entity Recognition. Semantic analysis is performed on the natural language question input by the user to identify marine acoustic entities, technical terms, and physical quantities. An entity recognition model is then used for entity extraction, and a marine acoustic thesaurus is used to expand the question terms with synonyms to improve the recall rate of subsequent map retrieval.

[0041] Step P2: Entity Expansion and Retrieval Based on Knowledge Graph. Using the terms obtained in Step P1 as search keywords, entity nodes whose names or aliases contain the terms are retrieved from the marine acoustics knowledge graph. Their adjacent entities and related relationships are obtained, and the corresponding evidence sentences and source documents are extracted, thus realizing the mapping from the problem terms to the relevant graph subgraphs.

[0042] Step P3: Query generation and graph retrieval execution. Based on the parsed entity and relation requirements, a graph database query statement is constructed. Without relying on a large language model to generate the query statement, a deterministic graph retrieval is performed in the graph database to obtain stable and controllable knowledge retrieval results.

[0043] Step P4: Collaborative Response of Graph Knowledge and Large Language Model. The graph retrieval results obtained in Step P3 (including entities, relations, and evidence sentences) are input together with the original question into the large language model, which is constrained to generate answers based on the true facts of the graph. When necessary, the large language model can supplement background explanations by combining its own existing knowledge, but it must not violate the facts of the graph and should reduce fictitious content.

[0044] Step P5: Answer Output and Interpretability Presentation. Output the answer generated by the large language model, and optionally display the graph evidence sentences supporting the answer, as well as related entities and relationships, to achieve interpretability presentation.

[0045] Furthermore, marine acoustic numerical simulation results are introduced as constraint information. By comprehensively utilizing semantic information such as simulation input parameters and numerical simulation results, image information such as propagation loss maps, and marine acoustic knowledge graphs, the simulation results are analyzed for interpretability and intelligently evaluated. The process includes the following steps: Step R1: Structured Extraction and Saving of Simulation Parameters. After completing the marine acoustic numerical simulation, the key parameters used in the simulation are structured and recorded. These simulation parameters include at least marine environmental parameters, acoustic parameters, deployment and geometric parameters, and performance evaluation threshold parameters. The simulation parameters are saved as a parameter file in structured data format for use in simulation result analysis and subsequent intelligent recommendations.

[0046] Step R2: Simulation Result Analysis and Numerical Index Extraction. The output file of the marine acoustic numerical simulation is analyzed to extract key numerical indicators related to deployment performance and propagation effect, including but not limited to optimal deployment depth combination, coverage ratio, effective detection distance, and performance statistics under threshold constraints. These numerical results are then uniformly formatted to serve as a quantitative basis for subsequent intelligent analysis.

[0047] Step R3: Propagation Loss Image Generation and Image Semantic Description. Based on the propagation loss field data output from the simulation, images such as propagation loss maps are generated to visualize the distribution characteristics of sound energy in the distance-depth space. Furthermore, a multimodal large language model is invoked to perform image semantic understanding on the propagation loss images, automatically generating image semantic description text containing sound energy attenuation features, spatial structure distribution, and explanations of potential physical mechanisms, to supplement spatial information that is difficult to express in numerical results.

[0048] Step R4: Key Simulation Parameter Extraction and Structured Summary Construction. Based on the complete set of simulation parameters obtained in Step 1, the simulation parameters are automatically screened and classified to extract key parameters that have a decisive impact on the acoustic propagation mechanism and deployment effect, and parameter summary information for intelligent analysis is constructed; wherein, the parameter summary includes at least: Environmental parameters characterizing the marine environment include seawater depth, topographic type, and seabed sediment type; Acoustic parameters characterizing acoustic conditions include the center frequency of the sound source and the type of sound velocity profile. Deployment parameters characterizing geometric and deployment constraints include the number of moorings and deployment depth; Evaluation parameters characterizing performance criteria include propagation loss threshold, coverage, and maximum detection distance.

[0049] The parameter summary is organized in a structured data format and, together with the numerical simulation results obtained in step R2 and the semantic description of the propagation loss image obtained in step R3, constitutes the unified input content for simulation result analysis, which is used for subsequent collaborative reasoning analysis based on knowledge graphs and large language models.

[0050] Step R5: Knowledge Retrieval Based on the Marine Acoustics Knowledge Graph. The simulation analysis problem proposed by the user is analyzed, and the technical terms in the problem are identified by combining the acoustic entity recognition model and the marine acoustic thesaurus. Entity nodes with related names or semantics are retrieved from the marine acoustics knowledge graph to obtain physical laws, empirical conclusions, and literature evidence related to the simulation scenario, thus realizing the association mapping from simulation parameters and results to domain knowledge.

[0051] Step R6: Collaborative Reasoning Analysis of Knowledge Graph and Large Language Model. The simulation parameter summary obtained in Step R4, the numerical results obtained in Step R2, the image semantic description obtained in Step R3, and the knowledge graph information retrieved in Step R5 are jointly input into the large language model. Under the constraint that the large language model relies on knowledge graph facts and simulation parameters, it is guided to comprehensively analyze the physical rationality, engineering effectiveness, and potential risks of the simulation results, and generate natural language analysis results with causal explanation capabilities.

[0052] Step R7: Output Analysis Results. Output the simulation results analysis conclusions generated by the large language model. The conclusions include at least an assessment of the rationality of the deployment scheme, an analysis of the impact of parameter settings, and suggestions for optimization and adjustment. Optionally, knowledge graph evidence related to the analysis conclusions can be displayed to ensure the interpretability and traceability of the analysis process.

[0053] Figure 1 A schematic diagram of the overall process of a simulation analysis and intelligent question-answering system based on a marine acoustic knowledge graph is presented. In application, user questions, simulation results, and knowledge graph query results are input into a large language model to achieve intelligent analysis and interpretation of marine acoustic simulation results, and output result analysis and auxiliary decision-making information.

[0054] Figure 2 A flowchart for constructing a marine acoustics knowledge graph is provided. It demonstrates the process of processing and constructing a knowledge graph from raw marine acoustics text, as described in this invention.

[0055] Figure 3 The CWS, NER, and RE model structures used in this invention, as well as the collaborative question-answering process of the graph and the large language model, are presented.

[0056] Figure 4 and Figure 5 An example of a marine acoustic knowledge graph constructed based on the method of this invention is given. When entity nodes are written to the graph database, node attributes include entity name, entity category, and unique number. When relation edges are written to the graph database, relation edge attributes include evidence sentence, supporting literature, relation name, subject node type, and object node type information.

[0057] Figure 6This paper presents an example of question-answering using a large language model enhanced with a knowledge graph. First, a natural language question is input, such as "What are the basic assumptions of the normal mode model?" The system identifies the key entity "normal mode model" in the question using an entity recognition model and expands the query using a thesaurus to improve the recall rate of knowledge retrieval. Then, nodes related to this entity and their relationships, such as model assumptions, applicable environments, and typical application scenarios, are retrieved from the knowledge graph. Finally, the retrieved graph entities and relationships, along with the user's question, are input into the large language model for reasoning to generate an answer. This ensures that the answer content is generated based on graph facts, thereby improving the accuracy and reliability of professional question answering.

[0058] Figures 7 to 9 An example of analyzing marine acoustic simulation results based on a marine acoustic knowledge graph is given.

[0059] Users request interpretation and analysis of simulation results. Under the knowledge constraints provided by the knowledge graph, the large language model analyzes the rationality of the simulation results and explains the performance of the simulation results in conjunction with the characteristics of the sound propagation path. Simultaneously, the system can analyze the potential environmental sensitivity and parameter change risks of the current simulation scheme based on the knowledge graph and provide corresponding suggestions for adjusting and optimizing simulation parameters.

Claims

1. A method for constructing a marine acoustic knowledge graph, characterized in that, Includes the following steps: Step S1: Collect marine acoustic text, preprocess the marine acoustic text to obtain raw marine acoustic corpus; Step S2: Construct a sentence segmentation and annotation dataset based on the original marine acoustic corpus. Divide the sentence segmentation and annotation dataset into a component training set, a component validation set, and a component test set. Use the component training set to train the marine acoustic Chinese word segmentation model CWS to obtain the optimal marine acoustic Chinese word segmentation model. The input of the marine acoustic Chinese word segmentation model is the component training set; the output of the marine acoustic Chinese word segmentation model is a text character sequence; the text character sequence corresponds to the BMES tag sequence, where B, M, E, and S represent the beginning, middle, end, and single-character word formation positions of words, respectively. Step S3: Construct a term dataset based on the original marine acoustic corpus, and divide the term dataset into a term training set, a term validation set, and a term test set; use the term training set to train the marine acoustic entity recognition model NER to obtain the optimal marine acoustic entity recognition model NER; The input of the marine acoustic entity recognition model (NER) is a term dataset; the output of the marine acoustic entity recognition model (NER) is a term character sequence; the term character sequence includes entity, entity category, and location information; entity term categories include E-type, R-type, P-type, D-type, L-type, and T-type; Location information refers to the position of entity term categories in the original marine acoustic corpus; E-type represents entities related to marine environment and propagation structures; R-type represents entities related to experiments and engineering; P-type represents entities related to physical parameters; D-type represents entities related to equipment systems; L-type represents entities related to location, sea area, and orientation; T-type represents entities related to theoretical models and calculation methods; the term character sequences are optimized, and a term thesaurus is established. Step S4: Construct a relation analysis dataset based on the original marine acoustic corpus, divide the relation analysis dataset into a relation training set, a relation validation set, and a relation test set, and use the relation training set to train the marine acoustic relation extraction model to obtain the optimal marine acoustic relation extraction model; The input to the marine acoustic relation extraction model is a relation analysis dataset; the output of the marine acoustic relation extraction model is a relation triple, which includes a subject entity, a relation predicate, and an object entity. Step S5: Construct a marine acoustics knowledge graph and graph database; A marine acoustics knowledge graph is constructed by using entities in all terminology character sequences as graph nodes, relation predicates in relation triples as graph relation edges, and location information and text content as graph attributes. The text content is the input corresponding to the relation triples in the marine acoustic relation extraction model (RE). After the marine acoustic knowledge graph is constructed, all graph nodes and graph relation edges are written into the graph database, and unique identifiers and indexes are set for the graph nodes.

2. The method for constructing a marine acoustic knowledge graph according to claim 1, characterized in that, The preprocessing process is as follows: remove garbled characters, full-width spaces, invisible control characters, spaces and newlines from the marine acoustic text, and retain only the text portion of the marine acoustic text; separate the numerical sequences, English strings or long technical terms in the marine acoustic text by character level.

3. The method for constructing a marine acoustic knowledge graph according to claim 1, characterized in that, The steps for constructing the sentence segmentation and annotation dataset are as follows: performing sentence component analysis on the original marine acoustic corpus, using the BMES tagging system to annotate character sequences, and obtaining the sentence segmentation and annotation dataset; the process for constructing the terminology dataset is as follows: performing terminology analysis on the original marine acoustic corpus, annotating the terms and term categories in the original marine acoustic corpus, and obtaining the terminology dataset; The steps for constructing the relation analysis dataset are as follows: perform relation analysis on the original marine acoustic corpus, label the relationship types between professional terms in the original marine acoustic corpus, and obtain the relation analysis dataset; The relation analysis dataset is divided into a relation training set, a relation validation set, and a relation test set; Relationship types include causal relationships, generation relationships, influence relationships, positive correlation relationships, negative correlation relationships, dependency relationships, input relationships, output relationships, attribute relationships, feature relationships, co-occurrence evidence relationships, simulated evidence relationships, composition relationships, and inclusion relationships.

4. The method for constructing a marine acoustic knowledge graph according to claim 1, characterized in that, The marine acoustic Chinese word segmentation model is based on the BERT-LSTM-CRF sequence annotation model structure.

5. The method for constructing a marine acoustic knowledge graph according to claim 1, characterized in that, The steps to obtain the optimal Chinese word segmentation model for marine acoustics are as follows: train the Chinese word segmentation model for marine acoustics using a component training set, verify the training process using a component validation set, and finally evaluate the performance of the optimal model using a component test set. When the word segmentation model evaluation index is greater than 70%, the optimal Chinese word segmentation model for marine acoustics is obtained. The word segmentation model evaluation index is classification precision, classification recall, and F1 score.

6. The method for constructing a marine acoustic knowledge graph according to claim 1, characterized in that, The process of optimizing the term character sequence and obtaining a term thesaurus is as follows: Unify entities with the same concept in the terminology character sequence into the same entity name; and create a term thesaurus based on the different names of entities with the same concept.

7. The method for constructing a marine acoustic knowledge graph according to claim 1, characterized in that, The marine acoustic entity recognition model NER is based on the BERT-LSTM-CRF sequence labeling model structure.

8. The method for constructing a marine acoustic knowledge graph according to claim 1, characterized in that, The steps to obtain the optimal marine acoustic entity recognition model (NER) are as follows: train the marine acoustic entity recognition model (NER) using a terminology training set, and obtain the optimal marine acoustic entity recognition model (NER) when the terminology model evaluation index is greater than 70%; the terminology model evaluation index is classification precision, classification recall, and F1 score.

9. The method for constructing a marine acoustic knowledge graph according to claim 1, characterized in that, The process of obtaining the optimal marine acoustic relation extraction model is as follows: the marine acoustic relation extraction model RE is trained using a relation training set, and the optimal marine acoustic relation extraction model RE is obtained when the cross-entropy loss function converges to a preset threshold.

10. A large language model intelligent question answering system that applies a marine acoustic knowledge graph construction method, including a marine acoustic knowledge graph, an optimal marine acoustic entity recognition model (NER), an optimal marine acoustic relation extraction model (RE), a large language model, and an intelligent question answering method; The intelligent question-answering method includes the following steps: Step Q1: The user enters a question text to ask a question; Step Q2: Process the received question text sequentially using the optimal marine acoustic entity recognition model (NER) and the optimal marine acoustic relation extraction model (RE). The optimal marine acoustic entity recognition model (NER) obtains the term character sequence in the question text; the optimal marine acoustic relation extraction model (RE) obtains the relation triples in the question text. Step Q3: Using the terminology character sequence of the question text, the relation triples of the question text, and the terminology thesaurus corresponding to the entities, construct a graph database retrieval statement and execute it to obtain the graph nodes that are the same as the entities in the question text and the graph relation edges corresponding to the graph nodes; Step Q4: Use the input large language model to reason about the text content in the graph nodes and graph relation edges to obtain the answer text to the question text.

Images