Marine ranch equipment intelligent question and answer architecture establishment method and system

Abstract

The invention discloses a marine ranch equipment intelligent question and answer architecture establishment method and system, and belongs to the technical field of knowledge question and answer. According to the method, a marine ranching equipment knowledge graph and an intention insight engine are constructed, and the intention insight engine comprises questions possibly proposed by a user about the field of marine ranching equipment and answers matched with the questions; then identifying the intention category of the original query statement of the user so as to grasp the query intention of the user; based on the intention category, the user complete query statement is determined, and the determined user complete query statement is beneficial to answer matching, so that the output result is more accurate; and finally, inputting the complete query statement of the user into the marine ranch equipment knowledge graph, and outputting a query result corresponding to the original query statement of the user. According to the method, the intelligent question and answer architecture of the marine ranching equipment can be established, the output query result is accurate, and the knowledge requirement of a user for the marine ranching equipment can be met.

Description

Technical Field

[0001] The present application relates to the field of knowledge question-answering technology, and in particular to a method and system for establishing an intelligent question-answering architecture for marine ranch equipment. Background Art

[0002] Knowledge graphs are graphical data structures that represent and store entities in the real world and their relationships through nodes and edges. The Big Language Model is an artificial intelligence system based on deep learning technology that can understand and generate natural language and is widely used in translation, conversation, writing assistance and other fields.

[0003] At present, there is a lack of research on knowledge graph question and answer in the field of marine ranch equipment, which cannot better meet actual needs. There is an urgent need for a method and system for establishing an intelligent question and answer architecture for marine ranch equipment. Summary of the invention

[0004] In view of this, the present application provides a method and system for establishing an intelligent question and answer architecture for marine ranch equipment, which can establish an intelligent question and answer architecture for marine ranch equipment. The output query results are relatively accurate and can meet users' knowledge needs for marine ranch equipment.

[0005] Specifically, the following technical solutions are included:

[0006] In a first aspect, the present application provides a method for establishing an intelligent question-answering architecture for marine ranching equipment, the method comprising:

[0007] According to the marine ranch equipment database, production data and production experience of marine ranch equipment related enterprises, a marine ranch equipment field data set is constructed, and based on the marine ranch equipment field data set, a marine ranch equipment knowledge graph is constructed, wherein the knowledge involved in the marine ranch equipment knowledge graph includes marine ranch equipment basic data, marine ranch equipment usage methods, marine ranch equipment fault diagnosis knowledge and marine ranch equipment maintenance data;

[0008] Building an intention insight engine based on the marine ranch equipment field data set and the usage demand information of marine ranch equipment practitioners;

[0009] According to the BERTGATE-CRF entity extraction model, extract entities and attributes of the user's original query sentence, wherein the BERTGATE-CRF entity extraction model includes a BERT network, a BiGRU network and a CRF network;

[0010] According to the preset natural language understanding model of the TMN temporal memory network and the entities and attributes of the user's original query sentence, the intent category of the user's original query sentence is identified;

[0011] Based on the intent category, determine the user's complete query statement;

[0012] The user's complete query statement is input into the marine ranch equipment knowledge graph, and the query result corresponding to the user's original query statement is output.

[0013] In some embodiments, the marine ranch equipment field data set is constructed based on the marine ranch equipment database, production data and production experience of marine ranch equipment related enterprises, and the marine ranch equipment knowledge graph is constructed based on the marine ranch equipment field data set, including:

[0014] Extract and summarize marine ranch equipment related content from the marine ranch equipment database, production materials and production experience of marine ranch equipment related enterprises, and obtain a marine ranch equipment field data set;

[0015] Randomly sampling the marine ranch equipment field dataset to obtain a sample dataset, and annotating the sample dataset by entity categories to obtain a first annotated dataset, wherein the entity categories include construction company, main function, maintenance measures, precautions, installation method, definition, and facilities;

[0016] The first labeled data set is cleaned by a regular expression to obtain a cleaned data set, wherein each piece of data in the marine ranch equipment field data set includes a marine ranch of an entity category, relevant attributes corresponding to the marine ranch of the entity category, and attribute values;

[0017] Performing sequence annotation on the cleaned data set by using a CERSA annotation method to obtain a second annotated data set;

[0018] Extracting entity relationships of the second labeled data set by using the BERTGATE-CRF entity extraction model, and verifying the entity relationships of the second labeled data set to obtain relationships between entities of the second labeled data set;

[0019] Based on the entities and relationships of Neo4j and the second annotated dataset, a knowledge graph of marine ranch equipment is constructed.

[0020] In some embodiments, the intention insight engine is constructed based on the marine ranch equipment field data set and the usage demand information of marine ranch equipment practitioners, including:

[0021] Based on the marine ranch equipment field data set and the usage demand information of marine ranch equipment practitioners, a marine ranch equipment question and answer corpus database is constructed;

[0022] Annotate the entities and inquiry intentions of the data in the ocean ranch equipment question-and-answer corpus, and build a question intention annotation library;

[0023] Based on the question intent annotation library and the preset matching mechanism, the questions that users may search for are determined and an intent insight engine is established.

[0024] In some embodiments, extracting entities and attributes of the user's original query statement according to the BERTGATE-CRF entity extraction model includes:

[0025] According to the BERTGATE-CRF entity extraction model, extract the entities and attributes of the user's original query statement;

[0026] Annotate the entities and attributes of the user's original query statement by means of CERSA;

[0027] The entities and attributes of the user's original query statement are standardized.

[0028] In some embodiments, obtaining the TMN temporal memory network preset natural language understanding model includes the following training steps:

[0029] The intent insight engine is divided into a training set and a validation set in a ratio of 7:3, and the questions in the training set and the validation set are preprocessed, wherein the preprocessing operation includes removing stop words, removing repeated words, and removing question segmentation words;

[0030] For the preprocessed question sentences in the training set, feature extraction is performed using a TF-IDF model to obtain a first feature vector, wherein each dimension of the first feature vector represents a word, and a numerical value represents the importance of the word in the preprocessed question sentences in the training set;

[0031] The pre-processed question sentences in the training set are encoded by the pre-trained BERT model to obtain a second feature vector;

[0032] The first eigenvector and the second eigenvector are fused by weighted average to obtain a mixed eigenvector;

[0033] Input the mixed feature vector into the classification model to obtain an intermediate intent classification result and an intermediate classification model, wherein the intermediate intent classification result includes the predicted classification label of the question in the training set, and the classification model includes a bidirectional TMN temporal memory network with a self-attention mechanism, a BPNN network connected to the bidirectional TMN temporal memory network, a fully connected layer connected to the BPNN network, a Softmax activation function, and a Multi-class Hinge Loss function;

[0034] In response to the intermediate intent classification result of the classification model and the intermediate classification model satisfying the training completion condition, the training is stopped.

[0035] In some embodiments, obtaining the TMN temporal memory network preset natural language understanding model includes the following verification steps:

[0036] Based on the validation set, a classification performance evaluation is performed on the trained intermediate classification model to obtain a classification performance evaluation result, wherein the classification performance evaluation result includes accuracy, precision, recall rate, and F1 score;

[0037] In response to the classification performance evaluation result satisfying the evaluation conditions, the trained intermediate classification model is used as the TMN temporal memory network preset natural language understanding model; in response to the classification performance evaluation result not satisfying the evaluation conditions, the parameters of the trained intermediate classification model are adjusted based on the back propagation algorithm and the Adam optimizer until the classification performance evaluation result meets the evaluation conditions.

[0038] In some embodiments, the pre-trained BERT model includes an MLM network and an NSP network connected to the MLM network, the MLM network is used to predict the gap between masked words and real words, and the NSP network is used to determine whether two questions are adjacent and coherent.

[0039] In some embodiments, determining the user's complete query statement based on the intent category includes:

[0040] Based on the intention insight engine, a query statement template for the marine ranch equipment knowledge graph is constructed;

[0041] Based on the query statement template and the intent category, the user's original query statement is completed to generate a user's complete query statement.

[0042] In some embodiments, the training completion condition is that the accuracy of three randomly selected intermediate intent classification results from all training rounds is not less than 90%, and the evaluation condition is that the accuracy, precision, recall rate and F1 score of the intermediate classification model do not decrease for three consecutive rounds.

[0043] In a second aspect, the present application provides a system for establishing an intelligent question-answering architecture for marine ranching equipment, the system comprising:

[0044] A graph construction module is used to construct a marine ranch equipment field data set based on the marine ranch equipment database, production materials and production experience of marine ranch equipment-related enterprises, and to construct a marine ranch equipment knowledge graph based on the marine ranch equipment field data set, wherein the knowledge involved in the marine ranch equipment knowledge graph includes marine ranch equipment basic data, marine ranch equipment usage methods, marine ranch equipment fault diagnosis knowledge and marine ranch equipment maintenance data;

[0045] An engine construction module, used to construct an intention insight engine based on the marine ranch equipment field data set and the usage demand information of marine ranch equipment practitioners;

[0046] An extraction module, used to extract entities and attributes of the user's original query sentence according to a BERTGATE-CRF entity extraction model, wherein the BERTGATE-CRF entity extraction model includes a BERT network, a BiGRU network, and a CRF network;

[0047] An identification module, used to identify the intent category of the user's original query statement according to the TMN temporal memory network preset natural language understanding model and the entities and attributes of the user's original query statement;

[0048] A determination module, used to determine a complete query statement of a user based on the intent category;

[0049] The output module is used to input the user's complete query statement into the marine ranch equipment knowledge graph and output the query result corresponding to the user's original query statement.

[0050] The beneficial effects of the technical solution provided by the embodiments of the present application include at least:

[0051] The embodiment of the present application provides a method and system for establishing an intelligent question-answering architecture for marine ranch equipment. The method first constructs a marine ranch equipment field data set based on a marine ranch equipment database, production materials and production experience of marine ranch equipment-related enterprises, and constructs a marine ranch equipment knowledge graph based on the marine ranch equipment field data set. The constructed marine ranch equipment field data set contains rich and accurate marine ranch equipment field-related knowledge, which is the basis for ensuring that users can obtain answers; then, based on the marine ranch equipment field data set and the usage demand information of marine ranch equipment practitioners, an intention insight engine is constructed, and the intention insight engine contains questions that users may ask about the field of marine ranch equipment and answers that match the questions; then, based on the BERTGATE-CRF entity extraction model, entities and attributes of the user's original query statement are extracted; then, based on the TMN temporal memory network preset natural language understanding model and the entities and attributes of the user's original query statement, the intent category of the user's original query statement is identified, thereby grasping the user's query intent; based on the intent category, the user's complete query statement is determined, and the determined user's complete query statement is conducive to matching the answer, so that the output result is more accurate; finally, the user's complete query statement is input into the marine ranch equipment knowledge graph, and the query result corresponding to the user's original query statement is output. This method can establish an intelligent question-answering architecture for marine ranch equipment, and the output query results are relatively accurate, which can meet users' knowledge needs for marine ranch equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0053] Figure 1 A method flow chart of a method for establishing an intelligent question-answering architecture for marine ranching equipment provided in an embodiment of the present application;

[0054] Figure 2 A schematic diagram of the effect of the method for establishing an intelligent question-answering architecture for marine ranching equipment provided in an embodiment of the present application;

[0055] Figure 3 A structural diagram of a system for establishing an intelligent question-and-answer architecture for marine ranching equipment provided in an embodiment of the present application. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0057] In order to make the technical solutions and advantages of the present application more clear, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0058] like Figure 1 As shown, in the first aspect, the present application provides a method for establishing an intelligent question-answering architecture for marine ranching equipment, the method comprising the following steps:

[0059] Step 101, construct a marine ranch equipment field data set based on the marine ranch equipment database, production materials and production experience of marine ranch equipment-related enterprises, and construct a marine ranch equipment knowledge graph based on the marine ranch equipment field data set, wherein the knowledge involved in the marine ranch equipment knowledge graph includes marine ranch equipment basic data, marine ranch equipment usage methods, marine ranch equipment fault diagnosis knowledge and marine ranch equipment maintenance data.

[0060] The constructed marine ranch equipment dataset contains rich and accurate knowledge related to the field of marine ranch equipment, which is the basis for ensuring that users can obtain answers.

[0061] In some embodiments, step 101 includes:

[0062] Step 1011, extract and summarize the marine ranch equipment database, the marine ranch equipment related content in the production materials and production experience of marine ranch equipment related enterprises, and obtain the marine ranch equipment field data set.

[0063] The marine ranch equipment dataset is based on the marine ranch equipment database, production materials and production experience of marine ranch equipment-related enterprises. It is rich in content and accurate in information.

[0064] In some embodiments, certain marine ranch equipment-related content is randomly sampled, the samples are labeled with entity categories, and then experts in the field of marine ranch equipment are combined to determine entity categories and entity relationships to ensure that marine ranch equipment materials are fully covered.

[0065] Step 1012, randomly sampling the marine ranch equipment field data set to obtain a sample data set, and annotating the sample data set with entity categories to obtain a first annotated data set, wherein the entity categories include construction company, main function, maintenance measures, precautions, installation method, definition and facilities.

[0066] Entity category labeling is performed so that the data in the first labeled data set has its corresponding category, which is convenient for subsequent model training.

[0067] In some embodiments, the sample data set can be labeled with entity categories based on the knowledge information of marine ranch equipment practitioners. After the sample data set is labeled with entity categories, the labeled entity categories can be verified based on the knowledge information of experts in the field of marine ranch equipment, and the possible relationships between entities can be determined to ensure the accuracy of the labeled entity categories.

[0068] Step 1013, cleaning the first labeled data set by regular expressions to obtain a cleaned data set, wherein each data in the cleaned data set includes an entity category of ocean ranch, and relevant attributes and attribute values ​​corresponding to the entity category of ocean ranch.

[0069] By cleaning, spaces, repeated contents, etc. in the first labeled data set can be removed.

[0070] In some embodiments, each data in the cleaned data set includes relevant attributes and attribute values ​​corresponding to a marine ranch of an entity category, and the format of the cleaned data can be expressed as a key-value pair in the form of {"key":"value","key":["value","value"]}, where the key "key" is an identifier used to identify or describe the corresponding value "value". There is a corresponding relationship between the key and the value, and the corresponding value can be obtained through the key.

[0071] In some embodiments, when a key-value pair contains one value, it is stored in the form of a string "value"; when it contains multiple values, it is stored in the form of an array [value].

[0072] For example, the format of the actual cleaning data can be expressed as {"type_name":"reef-type marine ranch","behalf_land":["northern waters of Chucun","Shandong Bay"]}.

[0073] Step 1014: sequence annotate the cleaned data set using the CERSA annotation method to obtain a second annotated data set.

[0074] Perform sequence annotation on the cleaned data set so that the cleaned data can obtain applicable sequences.

[0075] It should be noted that the Comprehensive Entity-Relation Structure Annotation (CERSA) is a complex annotation method that combines entity type labels, entity structure labels, entity relationship labels and BMESO sequence annotation methods. It is designed for the marine ranch equipment dataset; TE represents the equipment module entity, that is, the subject entity, SE represents the sub-entity, Ri represents the type of the i-th sub-entity SEi, that is, the relationship type between the subject entity and SEi, SE_Ri represents other entities that have a relationship with the marine ranch equipment module entity, B-TE represents the first letter of the subject entity, M-TE represents the middle character of the subject entity, E-TE represents the last character of the subject entity, S-TE represents the subject entity of a single character, B-SE_Ri represents the first character of the sub-entity SEi, M-SE_Ri represents the middle character of the sub-entity SEi, E-SE_Ri represents the last character of the sub-entity SEi, S-SE_Ri represents the single-character sub-entity SEi, and O represents other non-named entity characters.

[0076] Table 1 shows the sequence annotation results of “Jinghai No. 1 is equipped with a net washing machine” in the cleaning dataset.

[0077] Table 1 Sequence annotation table of “Jinghai No. 1 is equipped with a net washing machine” in the cleaning data set

[0078]

[0079] Step 1015: extract entity relationships of the second annotated data set through the BERTGATE-CRF entity extraction model, and verify the entity relationships of the second annotated data set to obtain relationships between entities in the second annotated data set.

[0080] Obtain the entity relationships of the second labeled dataset, thereby obtaining the connections or relationships between the entities in the second labeled dataset, and verify the entity relationships of the second labeled dataset based on the knowledge information of experts in the field of ocean ranching equipment to ensure the accuracy and integrity of the extracted entities and the relationships between entities.

[0081] In some embodiments, the BERTGATE-CRF entity extraction model includes a BERT network, a BiGRU network, and a CRF network.

[0082] In some embodiments, step 1015 includes the following steps:

[0083] Step 10151, the BERT network processes the input of the second labeled dataset as the sum of word vectors, segment vectors, and position vectors, which is used as the output of the BERT network.

[0084] In some embodiments, "Jing", "Hai", "Yi", "Hao", "Pei", "Bei", "Xi", "Wang", "Ji" are used as input characters, that is, the data in the second labeled dataset. The BERT network processes it as the sum of word vectors, segment vectors, and position vectors. See Table 2.

[0085] Table 2 Input Representation of the BERT Network

[0086]

[0087] Step 10152, the BiGRU network performs sequence modeling on the output of the BERT network to obtain a context-related feature representation, which is used as the output of the BiGRU network.

[0088] It should be noted that the bidirectional gated recurrent neural network (BiGRU, Bidirectional Gated Recurrent Unit) is used to perform sequence modeling on the input text. It consists of a forward GRU (Forward GRU) and a backward GRU (Backward GRU), and can effectively capture long-distance dependencies in the text sequence. GRU is a variant of the recurrent neural network (RNN) and contains a reset gate and an update gate. After the input sequence is encoded by the BiGRU, a context-related feature representation is obtained.

[0089] In some embodiments, the output of the BiGRU network can be expressed as BiGRU out = BiGRU(BERT out )

[0090] Step 10152, output the prediction label sequence with the highest probability through the CRF network for the output of the BiGRU network, and at the same time consider the dependencies between the labels, thereby reducing the probability of illegal labels and improving the accuracy of the prediction results.

[0091] It should be noted that the probabilistic graphical model (CRF, Conditional Random Field) is used to model the sequence labeling task. It utilizes the dependencies between the labeled sequences and globally optimizes the labeled sequences by maximizing the conditional probability, thereby improving the accuracy of sequence labeling.

[0092] The BERTGATE-CRF entity extraction model can make full use of the semantic representation ability of the BERT network, the sequence modeling ability of the BiGRU network, and the label dependency modeling ability of the CRF network, thereby achieving better performance in the entity relationship extraction task.

[0093] Marine ranch equipment entities refer to specific objects, concepts or entities in this field. Entities can be marine ranch types including fishing marine ranches, bottom seeding marine ranches, equipment marine ranches, reef-casting marine ranches, pastoral marine ranches, etc.; entities can be deep-water smart cages including Jinghai No. 1, Guobao No. 1, Changjing No. 1, etc.; entities can be environmental monitoring equipment including salinity sensors, flow sensors, remote-controlled smart winches, etc.; entities can be energy equipment including diesel generators, UPS uninterruptible power supplies, emergency generators, etc.; entities can be communication and navigation equipment including radars, manual whistles, radar transponders, AIS, etc.; entities can also be safety equipment including life jackets, life rafts, CO2 fire extinguishers, etc.

[0094] The marine ranch equipment entity attributes refer to the characteristics, properties or descriptions of the marine ranch equipment entity, which describe certain specific aspects or characteristics of the entity. The marine ranch equipment entity can be the introduction attribute when it is a multi-functional platform, large-scale aquaculture ship or other equipment module; the construction company attribute when the marine ranch equipment entity can be a deep-water intelligent cage, sea fishing boat or other equipment module; the main function attribute when the marine ranch equipment entity can be a multi-functional platform, deep-water intelligent cage or other equipment module; the maintenance measures attribute when the marine ranch equipment entity can be aquaculture equipment, energy equipment or other equipment; the precautions attribute when the marine ranch equipment entity can be a net equipment, water treatment equipment or other equipment; the installation method attribute when the marine ranch equipment entity can be aquaculture equipment, safety equipment or other equipment; the definition and facility attributes when the marine ranch equipment entity can be an equipment-type marine ranch or other marine ranch type.

[0095] The relationships between marine ranching equipment entities refer to the connections or relationships between different entities, describing the interactions, dependencies, and influences between the entities.

[0096] Step 1016, constructing a marine ranch equipment knowledge graph based on the entities and relationships of Neo4j and the second annotated dataset.

[0097] It should be noted that Neo4j is a graph database used to store and process graph-structured data. It is suitable for application scenarios that process complex relational data. Through its efficient graph traversal and real-time data processing capabilities, it provides users with excellent performance and intuitive data representation. It uses nodes, relationships, and properties to represent and store data. Nodes represent entities or objects in the graph, relationships represent connections or associations between nodes, and properties are key-value pairs on nodes or relationships for storing specific information.

[0098] Step 102, construct an intent insight engine based on the marine ranch equipment field data set and the usage demand information of marine ranch equipment practitioners.

[0099] The intent insight engine contains questions that users may ask about the field of marine ranch equipment (i.e., usage demand information of marine ranch equipment practitioners) and answers that match the questions, which can be used for subsequent model training.

[0100] In some embodiments, step 102 includes:

[0101] Step 1021, constructing a marine ranch equipment question and answer corpus database based on the marine ranch equipment field data set and the usage demand information of marine ranch equipment practitioners.

[0102] The marine ranch equipment question and answer corpus database includes the usage demand information of marine ranch equipment practitioners and related question and answer corpus.

[0103] Step 1022, annotate the entities and query intent of the data in the ocean ranch equipment question and answer corpus database, and build a question intent annotation library.

[0104] The entities and inquiry intentions of the data in the marine ranch equipment question and answer corpus database are labeled, so as to accurately lock in the entities and inquiry intentions in the usage demand information of marine ranch equipment practitioners, that is, which entity the marine ranch equipment practitioners have questions about and what aspects of the entity they have questions about.

[0105] In some embodiments, the entities and inquiry intent of the data in the labeled marine ranch equipment question and answer corpus database can be verified based on the knowledge information of experts in the marine ranch equipment field, so that the constructed question intent annotation library is more accurate.

[0106] In some embodiments, a question intention annotation library is constructed for the field of marine ranch equipment based on predefined relationships and attribute categories. The embodiment provided in this application divides question intentions into 18 categories, including 10 equipment module-relationship categories, 3 equipment module-attribute categories, and 5 equipment-attribute categories, see Table 3.

[0107] Table 3 Category table

[0108]

[0109]

[0110] Step 1023, based on the question intent annotation library and the preset matching mechanism, determine the questions that the user may search for and establish an intent insight engine.

[0111] Establishing an intent insight engine facilitates subsequent model training.

[0112] Table 4 gives some examples of the intent insight engine. For the question "What are the design sea area conditions for Genghai No. 1?", Genghai No. 1 is an equipment module, so the intent category of this sentence is "equipment module-design sea area conditions" in the relationship category, and the label is "8"; for the question "What are the safety equipment of Genghai No. 1?", Genghai No. 1 is an equipment module, so the intent type of this sentence is "equipment module-safety equipment" in the relationship category, and the label is "0".

[0113] Table 4 Some examples of intent insight engine

[0114]

[0115] Step 103: extract entities and attributes of the user's original query sentence according to the BERTGATE-CRF entity extraction model, wherein the BERTGATE-CRF entity extraction model includes a BERT network, a BiGRU network, and a CRF network.

[0116] Extract the entities and attributes of the user's original query statement to understand the user's query intention.

[0117] In step 103, the entities and attributes of the user's original query sentence are extracted, and the first two elements of the ocean ranch equipment triplet have been obtained, that is, the entity relationship or entity attribute combination has been partially obtained. Among them, the ocean ranch equipment triplet refers to (entity, relationship, entity) and (entity, attribute, attribute value) in the ocean ranch equipment knowledge graph.

[0118] Step 1031, extracting entities and attributes of the user's original query statement according to the BERTGATE-CRF entity extraction model for easy labeling.

[0119] Step 1032: annotate the entities and attributes of the user's original query statement through the CERSA annotation method to clarify the entity type, structure, and relationship.

[0120] Step 1033, standardize the entities and attributes of the user's original query statement to make them meet the requirements of professional vocabulary.

[0121] Step 104, based on the preset natural language understanding model of the TMN temporal memory network and the entities and attributes of the user's original query sentence, the intent category of the user's original query sentence is identified to facilitate giving a corresponding correct answer.

[0122] In some embodiments, obtaining the TMN temporal memory network preset natural language understanding model includes the following training steps:

[0123] (1) The intent insight engine is divided into a training set and a validation set in a ratio of 7:3, and the questions in the training set and the validation set are preprocessed. The preprocessing operations include removing stop words, removing repeated words, and removing question segmentation words.

[0124] (2) After preprocessing, the questions in the training set are subjected to feature extraction using the TF-IDF model to obtain the first feature vector V tfidf =[v1,v2,...,v n ], wherein each dimension of the first feature vector represents a word, and the numerical value represents the importance of the word in the question in the training set after preprocessing.

[0125] It should be noted that the TF-IDF model is a commonly used weighting method, in which TF (Term Frequency) represents the term frequency, which specifies the number of times word t appears in question d (hereinafter referred to as question d) in the training set after preprocessing; IDF (Inverse Document Frequency) represents the inverse document frequency, which is used to measure the importance of word t. The TF-IDF model is used to evaluate the importance of a word to a set of documents, and its importance is proportional to the number of times it appears in the document and inversely proportional to the frequency of its appearance in the corpus.

[0126] (3) The questions in the preprocessed training set are encoded using the pre-trained BERT model to obtain the second feature vector.

[0127] The pre-trained BERT model is the Chinese version, and the model parameters are 12-layer architecture, 768 hidden layers, and 12 attention heads.

[0128] It should be noted that the pre-trained BERT model is a pre-trained language model that learns the semantic representation of text data through large-scale unsupervised pre-training. It is based on the Transformer structure mainly composed of the self-attention mechanism, which is bidirectional, allowing the model to take into account the contextual information of each position in the sequence when encoding the input sequence, and automatically learn the rules and features of the corpus from massive unlabeled data, thereby obtaining feature representations containing richer semantic information, and then optimize specific tasks through fine-tuning and other methods to achieve excellent performance on specific tasks. Therefore, it performs well in a variety of natural language processing tasks such as sentence classification and intelligent question answering.

[0129] In some embodiments, the pre-trained BERT model includes an MLM network and an NSP network connected to the MLM network. The MLM network is used to predict the difference between masked words and real words, and the NSP network is used to determine whether two questions are adjacent and coherent. It can better extract features at the text word level and sentence level, and obtain word vectors containing more semantic information.

[0130] In the MLM network, given a part of the words in the input sequence are masked, the model needs to predict the labels of these masked words, and the cross-entropy loss function is used to evaluate the gap between the masked words predicted by the model and the real words; in the NSP network, given a pair of sentences, the model needs to determine whether the two sentences are adjacent and coherent, and the binary cross-entropy loss function is used to evaluate the model's judgment on the continuity of the sentences.

[0131] Since the input of the pre-trained BERT model is a set of text sequences, in some embodiments, it is first necessary to map each word into a word embedding vector using a pre-trained word vector model such as Word2Vec or GloVe, and then, position encoding is introduced before the pre-trained BERT model to capture the position information of the word in the sentence.

[0132] (4) The first eigenvector and the second eigenvector are fused by weighted averaging to obtain a mixed eigenvector.

[0133] The hybrid feature vector obtained by the fusion processing combines word frequency statistics and deep semantic information, so that the classification model can more accurately distinguish texts of different categories, which can facilitate the subsequent training of the classification model and make the trained component model more accurate.

[0134] In some embodiments, the first eigenvector is represented as V tfidf =[v1,v2,...,v n ], the second eigenvector is represented by V BERT =[V′1,V′2,...,V′ m ], assigning weights α and 1-1-α respectively. Then the fused feature vector is expressed as V fusion=α×V tfidf +(1-α)×V BERT . Among them, α is a weight value in the range of [0,1], which is used to balance the relative importance of the first eigenvector and the second eigenvector in the fusion; the obtained V fusion That is, it is a new eigenvector that combines the first eigenvector and the second eigenvector.

[0135] (5) The mixed feature vector is input into the classification model to obtain the intermediate intent classification result and the intermediate classification model. The intermediate intent classification result includes the predicted classification label of the question in the training set. The classification model includes a bidirectional TMN temporal memory network with a self-attention mechanism, a BPNN network connected to the bidirectional TMN temporal memory network, a fully connected layer connected to the BPNN network, a Softmax activation function and a Multi-class Hinge Loss function.

[0136] It should be noted that the internal structure of the bidirectional TMN temporal memory network with self-attention mechanism includes a unit, which contains three parts: information controller, memory regulator and result regulator. Each part controls the flow of information and when to remember or forget certain information through sigmoid function and dot product operation. The self-attention mechanism is applied to the output of the bidirectional TMN temporal memory network, so that the classification model can dynamically assign different weights to the hidden state of each time step, so that the classification model pays more attention to the important parts, thereby improving the classification performance.

[0137] In some embodiments, the BPNN network generally includes an input layer, a hidden layer, and an output layer, and a back propagation algorithm is used to optimize parameters between the hidden layer and the output layer. The BPNN network can learn more complex nonlinear functional relationships and enhance the representation ability of the model.

[0138] In some embodiments, the BPNN network is given an input feature vector A i , output y i The calculation process is as follows:

[0139] z i =ReLU(W z ·A i +b z )

[0140] y i =Softmax(W y ·z i +b y )

[0141] Among them, W z and W y is the weight matrix, b z and by is the bias vector, ReLU represents the rectified linear unit activation function, Z i It represents the hidden layer output after ReLU, and Softmax is the activation function.

[0142] In some embodiments, the output of the BPNN network is used as the input of the fully connected layer, and the Softmax activation function is used to output the intermediate intent classification result. The denominator of the Softmax function is the sum of exponential terms, so it will amplify the gap for larger original scores, allowing the model to more accurately classify the highest scoring category.

[0143] In some embodiments, the difference between the predicted classification label and the true classification label of the question in the model training set is determined by the Multi-class Hinge Loss loss function, which can reflect the accuracy of the intermediate intent classification result. The Multi-class Hinge Loss loss function not only focuses on the correct classification, but also focuses on the confidence and boundary interval of the classification, which helps to improve the generalization ability of the model.

[0144] (6) In response to the intermediate intent classification result of the classification model and the intermediate classification model satisfying the training completion conditions, stop training.

[0145] In some embodiments, obtaining the TMN temporal memory network preset natural language understanding model includes the following verification steps:

[0146] (1) Based on the validation set, the classification performance of the trained intermediate classification model is evaluated to obtain the classification performance evaluation results, where the classification performance evaluation results include accuracy, precision, recall rate, and F1 score.

[0147] (2) In response to the classification performance evaluation result satisfying the evaluation conditions, the trained intermediate classification model is used as the TMN temporal memory network preset natural language understanding model; in response to the classification performance evaluation result not satisfying the evaluation conditions, the parameters of the trained intermediate classification model are adjusted based on the back propagation algorithm and the Adam optimizer until the classification performance evaluation result satisfies the evaluation conditions.

[0148] The Adam optimizer combines the inertia concept of the momentum method with the characteristics of the adaptive learning rate. It can converge faster and has different learning rates for different parameters, making it more adaptable.

[0149] In some embodiments, the training completion condition is that the accuracy of three randomly selected intermediate intent classification results from all training rounds is not less than 90%, and the evaluation condition is that the accuracy, precision, recall rate and F1 score of the intermediate classification model do not decrease for three consecutive rounds.

[0150] Step 105: Determine the user's complete query statement based on the intent category.

[0151] The determined complete query statement of the user is conducive to matching the answer, making the query result more accurate.

[0152] In some embodiments, step 105 includes the following steps:

[0153] Step 1051, based on the intent insight engine, construct a query statement template for the marine ranch equipment knowledge graph.

[0154] The query statement template can make the user's original query statement more conducive to matching the corresponding answer.

[0155] In some embodiments, the specific implementation method of step 1051 is: according to the labels of different question intent categories in the intent insight engine, construct their respective Cypher query templates for the marine ranch equipment knowledge graph based on the Cypher language, thereby obtaining a query statement template for the marine ranch equipment knowledge graph.

[0156] Step 1052, based on the query statement template and the intent category, the user's original query statement is completed to generate a user's complete query statement.

[0157] The user's complete query statement helps match the answer, making the output result more accurate.

[0158] In some embodiments, step 1052 is specifically implemented by filling the intent categories of entities and attributes of the user's original query statement extracted in step 103 into the query statement template obtained in step 1051 to generate a complete query statement of the user.

[0159] Step 106, input the user's complete query statement into the marine ranch equipment knowledge graph, and output the query result corresponding to the user's original query statement.

[0160] In step 103, the entities and attributes of the user's original query are extracted, and the first two elements of the ocean ranch equipment triple are obtained, that is, the entity relationship or entity attribute combination is obtained. The last element of the ocean ranch equipment triple is obtained through the user's complete query statement in step 106, and the query result, that is, the answer to the user's question, is obtained.

[0161] For example, the user asks "What are the design sea area conditions for Genghai No. 1?", and the entities "Genghai No. 1" and "Design sea area conditions" are extracted in step 103, that is, a partial triple is obtained: (Genghai No. 1, Design sea area conditions,?), and the actual content represented by "?" is the answer to the question. For example, the Cypher query command for the user's question "What are the design sea area conditions for Genghai No. 1?" with the label "8" is: "match(entity)-[r]->(p)where entity.name='Genghai No. 1'and r.name='Design sea area conditions'return p" to perform a relational query, and all matching nodes can be obtained.

[0162] Figure 2 A schematic diagram of the effect of the method for establishing the intelligent question-answering architecture for marine ranching equipment provided in the embodiment of the present application, from Figure 2 It can be seen that the answer to "What are the design sea conditions of Genghai No. 1?" is "Design sea conditions of Genghai No. 1; ambient temperature (-10℃~36℃); wave period (8s); maximum storm wind speed (33.0m / s); maximum storm wave height (4.6m); maximum operating wind speed (17.1m / s); maximum operating wave height (3.95m); design water depth (10m); design flow rate (1.0 / 0.5lm / s)". The query results are accurate and can meet users' knowledge needs for marine ranch equipment.

[0163] In summary, the method for establishing an intelligent question and answer architecture for marine ranch equipment provided in the embodiment of the present application can establish an intelligent question and answer architecture for marine ranch equipment, and the output query results are relatively accurate, which can meet the user's knowledge needs for marine ranch equipment.

[0164] Second, this application provides a system for establishing an intelligent question-answering architecture for marine ranching equipment, see Figure 3 , the system comprises:

[0165] A graph construction module 301 is used to construct a marine ranch equipment field data set based on a marine ranch equipment database, production data and production experience of marine ranch equipment-related enterprises, and to construct a marine ranch equipment knowledge graph based on the marine ranch equipment field data set, wherein the knowledge involved in the marine ranch equipment knowledge graph includes marine ranch equipment basic data, marine ranch equipment usage methods, marine ranch equipment fault diagnosis knowledge and marine ranch equipment maintenance data;

[0166] An engine construction module 302 is used to construct an intention insight engine based on the marine ranch equipment field data set and the usage demand information of marine ranch equipment practitioners;

[0167] An extraction module 303 is used to extract entities and attributes of the user's original query sentence according to a BERTGATE-CRF entity extraction model, wherein the BERTGATE-CRF entity extraction model includes a BERT network, a BiGRU network, and a CRF network;

[0168] The identification module 304 is used to identify the intent category of the user's original query statement according to the TMN temporal memory network preset natural language understanding model and the entities and attributes of the user's original query statement;

[0169] A determination module 305 is used to determine a complete query statement of a user based on the intent category;

[0170] The output module 306 is used to input the user's complete query statement into the ocean ranch equipment knowledge graph and output the query result corresponding to the user's original query statement.

[0171] In some embodiments, the graph construction module 301 includes:

[0172] The data set extraction unit is used to extract and summarize the marine ranch equipment database, the production materials and production experience of marine ranch equipment related enterprises, and obtain the marine ranch equipment field data set;

[0173] The first labeled data set unit is used to randomly sample the marine ranch equipment field data set to obtain a sample data set, and to label the sample data set with entity categories to obtain a first labeled data set, wherein the entity categories include construction company, main function, maintenance measures, precautions, installation method, definition, and facilities;

[0174] A data set cleaning unit, used to clean the first labeled data set by regular expression to obtain a cleaned data set, wherein each data in the marine ranch equipment field data set includes a marine ranch of an entity category, and relevant attributes and attribute values ​​corresponding to the marine ranch of the entity category;

[0175] A second annotated data set unit is used to perform sequence annotation on the cleaned data set by using a CERSA annotation method to obtain a second annotated data set;

[0176] A verification unit, used to extract entity relationships of the second labeled data set through a BERTGATE-CRF entity extraction model, and verify the entity relationships of the second labeled data set to obtain relationships between entities of the second labeled data set;

[0177] Build a graph unit to construct a knowledge graph of marine ranch equipment based on the entities and relationships of Neo4j and the second annotated dataset.

[0178] In some embodiments, the engine building module 302 includes:

[0179] A database construction unit, used to construct a marine ranch equipment question-and-answer corpus database based on a marine ranch equipment field data set and usage demand information of marine ranch equipment practitioners;

[0180] The intention annotation library unit is used to annotate the entities and inquiry intentions of the data in the marine ranch equipment question-answering corpus and build a question intention annotation library;

[0181] The intent insight engine unit is used to determine the questions that users may search for based on the question intent annotation library and the preset matching mechanism, and to establish the intent insight engine.

[0182] In some embodiments, the extraction module 303 includes:

[0183] The entity and attribute extraction unit is used to extract entities and attributes of the user's original query statement according to the BERTGATE-CRF,entity extraction model;

[0184] A CERSA annotation unit is used to annotate entities and attributes of the user's original query statement by using a CERSA annotation method;

[0185] The standardization unit is used to standardize the entities and attributes of the user's original query statement.

[0186] In some embodiments, obtaining the TMN temporal memory network preset natural language understanding model includes the following training steps:

[0187] The intent insight engine is divided into a training set and a validation set in a ratio of 7:3, and the questions in the training set and the validation set are preprocessed. The preprocessing operation includes removing stop words, removing repeated words, and removing question segmentation words;

[0188] For the questions in the preprocessed training set, feature extraction is performed using a TF-IDF model to obtain a first feature vector, wherein each dimension of the first feature vector represents a word, and a numerical value represents the importance of the word in the questions in the preprocessed training set;

[0189] The questions in the preprocessed training set are encoded using the pretrained BERT model to obtain the second feature vector;

[0190] The first eigenvector and the second eigenvector are fused by weighted average to obtain a mixed eigenvector;

[0191] Input the mixed feature vector into the classification model to obtain the intermediate intent classification result and the intermediate classification model. The intermediate intent classification result includes the predicted classification label of the question in the training set. The classification model includes a bidirectional TMN temporal memory network with a self-attention mechanism, a BPNN network connected to the bidirectional TMN temporal memory network, a fully connected layer connected to the BPNN network, a Softmax activation function, and a Multi-class Hinge Loss function.

[0192] In response to the intermediate intent classification result of the classification model and the intermediate classification model satisfying the training completion condition, the training is stopped.

[0193] In some embodiments, obtaining the TMN temporal memory network preset natural language understanding model includes the following verification steps:

[0194] Based on the validation set, the classification performance of the trained intermediate classification model is evaluated to obtain the classification performance evaluation results, where the classification performance evaluation results include accuracy, precision, recall, and F1 score;

[0195] In response to the classification performance evaluation result satisfying the evaluation conditions, the trained intermediate classification model is used as the TMN temporal memory network preset natural language understanding model; in response to the classification performance evaluation result not satisfying the evaluation conditions, the parameters of the trained intermediate classification model are adjusted based on the back propagation algorithm and the Adam optimizer until the classification performance evaluation result meets the evaluation conditions.

[0196] In some embodiments, the pre-trained BERT model includes an MLM network and an NSP network connected to the MLM network, the MLM network is used to predict the gap between masked words and real words, and the NSP network is used to determine whether two questions are adjacent and coherent.

[0197] In some embodiments, the determination module 305 includes:

[0198] The template unit is used to build a query statement template for the marine ranch equipment knowledge graph based on the intent insight engine;

[0199] The generation unit is used to complete the user's original query statement based on the query statement template and the intent category to generate a complete query statement for the user.

[0200] In some embodiments, the training completion condition is that the accuracy of three randomly selected intermediate intent classification results from all training rounds is not less than 90%, and the evaluation condition is that the accuracy, precision, recall rate and F1 score of the intermediate classification model do not decrease for three consecutive rounds.

[0201] In summary, the system for establishing an intelligent question and answer architecture for marine ranch equipment provided in the embodiment of the present application can establish an intelligent question and answer architecture for marine ranch equipment, and the output query results are relatively accurate, which can meet the user's knowledge needs for marine ranch equipment.

[0202] In the present application, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance. The term "plurality" refers to two or more than two, unless otherwise clearly defined.

[0203] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the present application disclosed herein. The present application is intended to cover any variations, uses or adaptations of the present application, which follow the general principles of the present application and include common knowledge or customary techniques in the art that are not disclosed in the present application. The specification and examples are intended to be exemplary only.

[0204] It should be understood that the present application is not limited to the precise structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims

1. A method for establishing an intelligent question-answering framework for marine ranching equipment, characterized in that: The method comprises: According to the marine ranch equipment database, production data and production experience of marine ranch equipment related enterprises, a marine ranch equipment field data set is constructed, and based on the marine ranch equipment field data set, a marine ranch equipment knowledge graph is constructed, wherein the knowledge involved in the marine ranch equipment knowledge graph includes marine ranch equipment basic data, marine ranch equipment usage methods, marine ranch equipment fault diagnosis knowledge and marine ranch equipment maintenance data; Building an intention insight engine based on the marine ranch equipment field data set and the usage demand information of marine ranch equipment practitioners; According to the BERTGATE-CRF entity extraction model, extracting entities and attributes of the user's original query sentence, wherein the BERTGATE-CRF entity extraction model includes a BERT network, a BiGRU network and a CRF network; According to the preset natural language understanding model of the TMN temporal memory network and the entities and attributes of the original query sentence of the user, the intent category of the original query sentence of the user is identified; Based on the intent category, determine the user's complete query statement; The user's complete query statement is input into the marine ranch equipment knowledge graph, and the query result corresponding to the user's original query statement is output.

2. The method for establishing an intelligent question-answering framework for marine ranching equipment according to claim 1, characterized in that: The marine ranch equipment field data set is constructed based on the marine ranch equipment database, production data and production experience of marine ranch equipment related enterprises, and the marine ranch equipment knowledge graph is constructed based on the marine ranch equipment field data set, including: Extract and summarize marine ranch equipment related content from the marine ranch equipment database, production materials and production experience of marine ranch equipment related enterprises, and obtain a marine ranch equipment field data set; Randomly sampling the marine ranch equipment field dataset to obtain a sample dataset, and annotating the sample dataset by entity categories to obtain a first annotated dataset, wherein the entity categories include construction company, main function, maintenance measures, precautions, installation method, definition, and facilities; The first labeled data set is cleaned by using a regular expression to obtain a cleaned data set, wherein each data in the cleaned data set includes an entity category of ocean ranch, and relevant attributes and attribute values ​​corresponding to the entity category of ocean ranch; Performing sequence annotation on the cleaned data set by using a CERSA annotation method to obtain a second annotated data set; Extracting entity relationships of the second labeled data set by using the BERTGATE-CRF entity extraction model, and verifying the entity relationships of the second labeled data set to obtain relationships between entities of the second labeled data set; Based on the entities and relationships of Neo4j and the second annotated dataset, a knowledge graph of marine ranch equipment is constructed.

3. The method for establishing an intelligent question-answering framework for marine ranching equipment according to claim 1 or 2, characterized in that: The intention insight engine is constructed based on the marine ranch equipment field data set and the usage demand information of marine ranch equipment practitioners, including: Based on the marine ranch equipment field data set and the usage demand information of marine ranch equipment practitioners, a marine ranch equipment question and answer corpus database is constructed; Annotate the entities and inquiry intentions of the data in the ocean ranch equipment question-and-answer corpus, and build a question intention annotation library; Based on the question intent annotation library and the preset matching mechanism, the questions that users may search for are determined and an intent insight engine is established.

4. The method for establishing an intelligent question-answering framework for marine ranching equipment according to claim 1, characterized in that: The entities and attributes of the user's original query statement are extracted according to the BERTGATE-CRF entity extraction model, including: According to the BERTGATE-CRF entity extraction model, extract the entities and attributes of the user's original query statement; Annotate the entities and attributes of the user's original query statement by means of CERSA; The entities and attributes of the user's original query statement are standardized.

5. The method for establishing an intelligent question-answering framework for marine ranching equipment according to claim 1, characterized in that: Obtaining the TMN temporal memory network preset natural language understanding model includes the following training steps: The intent insight engine is divided into a training set and a validation set in a ratio of 7:3, and the questions in the training set and the validation set are preprocessed, wherein the preprocessing operation includes removing stop words, removing repeated words, and removing question segmentation words; For the preprocessed question sentences in the training set, feature extraction is performed using a TF-IDF model to obtain a first feature vector, wherein each dimension of the first feature vector represents a word, and a numerical value represents the importance of the word in the preprocessed question sentences in the training set; The pre-processed question sentences in the training set are encoded by the pre-trained BERT model to obtain a second feature vector; The first eigenvector and the second eigenvector are fused by weighted average to obtain a mixed eigenvector; Input the mixed feature vector into the classification model to obtain an intermediate intent classification result and an intermediate classification model, wherein the intermediate intent classification result includes the predicted classification label of the question in the training set, and the classification model includes a bidirectional TMN temporal memory network with a self-attention mechanism, a BPNN network connected to the bidirectional TMN temporal memory network, a fully connected layer connected to the BPNN network, a Softmax activation function, and a Multi-class Hinge Loss function; In response to the intermediate intent classification result of the classification model and the intermediate classification model satisfying the training completion condition, the training is stopped.

6. The method for establishing an intelligent question-answering framework for marine ranching equipment according to claim 5, characterized in that: Obtaining the TMN temporal memory network preset natural language understanding model includes the following verification steps: Based on the validation set, a classification performance evaluation is performed on the trained intermediate classification model to obtain a classification performance evaluation result, wherein the classification performance evaluation result includes accuracy, precision, recall rate, and F1 score; In response to the classification performance evaluation result satisfying the evaluation conditions, the trained intermediate classification model is used as the TMN temporal memory network preset natural language understanding model; in response to the classification performance evaluation result not satisfying the evaluation conditions, the parameters of the trained intermediate classification model are adjusted based on the back propagation algorithm and the Adam optimizer until the classification performance evaluation result meets the evaluation conditions.

7. The method for establishing an intelligent question-answering framework for marine ranching equipment according to claim 5, characterized in that: The pre-trained BERT model includes an MLM network and an NSP network connected to the MLM network, the MLM network is used to predict the gap between masked words and real words, and the NSP network is used to determine whether two questions are adjacent and coherent.

8. The method for establishing an intelligent question-answering framework for marine ranching equipment according to claim 4, characterized in that: The determining of the user's complete query statement based on the intent category includes: Based on the intention insight engine, a query statement template for the marine ranch equipment knowledge graph is constructed; Based on the query statement template and the intent category, the user's original query statement is completed to generate a user's complete query statement.

9. The method for establishing an intelligent question-answering framework for marine ranching equipment according to claim 6, characterized in that: The training completion condition is that the accuracy of three randomly selected intermediate intent classification results from all training rounds is not less than 90%, and the evaluation condition is that the accuracy, precision, recall and F1 score of the intermediate classification model do not decrease for three consecutive rounds.

10. The system for establishing the intelligent question-answering architecture of marine ranch equipment is characterized by: The system comprises: A graph construction module is used to construct a marine ranch equipment field data set based on the marine ranch equipment database, production materials and production experience of marine ranch equipment-related enterprises, and to construct a marine ranch equipment knowledge graph based on the marine ranch equipment field data set, wherein the knowledge involved in the marine ranch equipment knowledge graph includes marine ranch equipment basic data, marine ranch equipment usage methods, marine ranch equipment fault diagnosis knowledge and marine ranch equipment maintenance data; An engine construction module, used to construct an intention insight engine based on the marine ranch equipment field data set and the usage demand information of marine ranch equipment practitioners; An extraction module, used to extract entities and attributes of the user's original query sentence according to a BERTGATE-CRF entity extraction model, wherein the BERTGATE-CRF entity extraction model includes a BERT network, a BiGRU network, and a CRF network; An identification module, used to identify the intent category of the user's original query statement according to the TMN temporal memory network preset natural language understanding model and the entities and attributes of the user's original query statement; A determination module, used to determine a complete query statement of a user based on the intent category; The output module is used to input the user's complete query statement into the marine ranch equipment knowledge graph and output the query result corresponding to the user's original query statement.

Images