A method, device and computer equipment for decomposing complex questions in vertical fields

Abstract

The present application discloses a method, device and computer equipment for decomposing complex question sentences in a vertical field. The method comprises: determining the vertical field to which the complex sentence to be decomposed belongs, inputting the complex sentence into a recognition model corresponding to the vertical field, extracting the subject entity and the auxiliary verb type; combining the subject entity and the auxiliary verb type to form a feature expression, and matching it with a syntactic template library to generate a knowledge graph query statement; inputting the query statement into the knowledge graph corresponding to the vertical field, retrieving the associated subgraph corresponding to the subject entity; screening out several entities with the highest degree of association with the subject entity from the associated subgraph, and generating multiple sub-questions after the complex sentence is decomposed according to the associated semantic information corresponding to each entity; the present invention utilizes the knowledge graph in combination with the syntactic template to realize the engineered generation of the decomposed sub-questions, overcomes the problems of poor flexibility and limited vision existing in the decomposition method relying solely on the rule template, and has certain theoretical reference and practical reference value.

Description

Technical Field

[0001] The present application relates to the technical field of natural language processing, and more specifically, to a method, device and computer equipment for decomposing complex questions in a vertical field. Background Art

[0002] With the development and application of natural language processing (NLG) technology, the decomposition of complex questions has become an important branch task in natural language processing. Its basic connotation is to decompose a specific complex question into several sub-questions that can encompass its complete semantics through natural language understanding and natural language generation technology. Its main significance lies in: on the one hand, decomposing vague and general complex questions into several simple sub-questions, answering each sub-question in turn and finally forming the answer to the complex question, it is easier to achieve the correct processing of the complex question answer than directly performing intent recognition on the complex question and then matching the answer; on the other hand, decomposing complex questions into simple sub-questions that are easier to understand can also more fully explore the implicit needs behind the complex questions, thereby inspiring and guiding users to ask more valuable questions.

[0003] At present, the decomposition methods of complex questions are mainly divided into decomposition methods based on rule templates and decomposition methods based on machine learning. They each have the following shortcomings:

[0004] (1) The decomposition method based on rule templates has a limited vision. The decomposition method based on rule templates mainly matches the complex question to be decomposed with the pre-established rule template, extracts the syntactic features of the complex question, and fills in the corresponding syntactic template to obtain the corresponding sub-questions. This method requires domain experts to formulate a large number of decomposition rules. However, things in the objective world are intricately related, and the formulated decomposition rules are limited by factors such as the expert's knowledge structure, subjective consciousness, and personal preferences, resulting in a limited decomposition vision and a decomposition direction that is greatly affected by expert human factors.

[0005] (2) The "black box effect" of decomposition methods based on machine learning actually exists. Decomposition methods based on machine learning mainly use a large number of training question pair corpora (i.e., the mapping relationship between complex questions and sub-questions) to achieve end-to-end decomposition of complex questions. However, with the increase of information elements in complex questions, the unexplainability of complex question decomposition methods based on machine learning becomes more and more obvious, resulting in poor decomposition results and the "black box effect" actually exists. Summary of the invention

[0006] In response to at least one defect or improvement need in the prior art, the present invention provides a vertical field complex question decomposition method based on knowledge graph and syntactic template. The knowledge graph is used to discover semantic association information that is objectively and closely related to complex questions, and then combined with the syntactic template to generate decomposed sub-questions in natural language form, ultimately achieving a more scientific, sufficient and effective complex question decomposition effect.

[0007] To achieve the above object, according to a first aspect of the present invention, a method for decomposing complex questions in a vertical domain is provided, which comprises:

[0008] Obtaining a complex sentence to be decomposed, and determining the vertical field to which the complex sentence belongs;

[0009] Inputting the complex sentence into the recognition model corresponding to the vertical field, extracting the subject entity and auxiliary verb type in the complex sentence;

[0010] The subject entity and the auxiliary verb type are combined to form a feature expression, the feature expression is matched with a pre-configured syntactic template library corresponding to the vertical field, a syntactic template with the highest matching degree is obtained, and a corresponding knowledge graph query statement is generated according to the syntactic template;

[0011] Input the query statement into the knowledge graph corresponding to the vertical field, and retrieve the associated subgraph corresponding to the subject entity;

[0012] A number of entities with the highest degree of association with the subject entity are extracted from the association subgraph, and a plurality of sub-questions after the complex sentence is decomposed are generated according to the association semantic information corresponding to each of the entities.

[0013] Preferably, in the above vertical field complex question decomposition method, the step of determining the vertical field to which the complex sentence belongs includes:

[0014] Inputting the complex sentence into a pre-trained domain recognition model, and determining the vertical domain to which the complex sentence belongs through the domain recognition model;

[0015] The domain recognition model is trained using multiple samples with vertical domain labels, and each sample is processed into a question-domain mapping pair and then input into the domain recognition model.

[0016] Preferably, the above vertical field complex question decomposition method, wherein the complex sentence is input into the recognition model corresponding to the vertical field, and the subject entity and auxiliary verb type in the complex sentence are extracted, comprises:

[0017] Perform word segmentation and vectorization on the complex sentence to obtain a word vector matrix;

[0018] Performing feature extraction and normalization processing on the word vector matrix to obtain predicted label scores of different subject entities corresponding to the complex sentence;

[0019] The predicted label scores are subjected to context constraints to obtain the subject entity with the highest predicted label score and its corresponding auxiliary verb type.

[0020] Preferably, the above vertical field complex question decomposition method, wherein the extracting of several entities with the highest correlation with the subject entity from the correlation subgraph, and generating multiple sub-questions after the complex sentence is decomposed according to the correlation semantic information corresponding to each of the entities, comprises:

[0021] Calculate the importance of all entities associated with the subject entity in the associated subgraph and sort them, and extract the top N entities as associated entities;

[0022] The associated entity is used as a root node to perform adjacent entity traversal to obtain multiple associated semantic information in triple format, and each of the associated semantic information and auxiliary verb types are combined to obtain multiple sub-questions after the complex sentence is decomposed.

[0023] Preferably, in the above vertical field complex question decomposition method, the step of calculating the importance of all entities associated with the subject entity in the associated subgraph and ranking them comprises:

[0024] The PageRank algorithm is used to calculate the importance of all entities in the associated subgraph, which is defined as follows:

[0025]

[0026] Among them, PR(A) represents the importance score of entity A associated with the subject entity, PR(T1)...PR(T n ) represent entities A and T1...T n The correlation between them, C(T1)...C(T n ) represent entities T1...T n The sum of the associations with all entities, where d is the damping coefficient in the range 0-1.

[0027] According to a second aspect of the present invention, there is also provided a vertical domain complex question decomposition device, which comprises:

[0028] A determination module, configured to obtain a complex sentence to be decomposed, and determine the vertical field to which the complex sentence belongs;

[0029] An extraction module, configured to input a complex sentence into a recognition model corresponding to the vertical field, and extract a subject entity and an auxiliary verb type in the complex sentence;

[0030] A matching module, which is configured to combine the subject entity and the auxiliary verb type to form a feature expression, match the feature expression with a pre-configured syntactic template library corresponding to the vertical field, obtain the syntactic template with the highest matching degree, and generate a corresponding knowledge graph query statement according to the syntactic template;

[0031] A retrieval module is configured to input the query statement into the knowledge graph library corresponding to the vertical field, and retrieve the associated subgraph corresponding to the subject entity;

[0032] The output module is configured to extract several entities with the highest correlation with the subject entity from the correlation subgraph, and generate multiple sub-questions after the complex sentence is decomposed according to the correlation semantic information corresponding to each of the entities.

[0033] According to the third aspect of the present invention, there is also provided an electronic device, which comprises at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the computer program is executed by the processing unit, the processing unit executes the steps of any one of the above-mentioned methods.

[0034] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0035] The vertical domain complex question decomposition method provided by the present invention first determines the vertical domain to which the complex sentence to be decomposed belongs, inputs the complex sentence into the recognition model corresponding to the vertical domain, and extracts the subject entity and the auxiliary verb type in the complex sentence; combines the subject entity and the auxiliary verb type to form a feature expression, matches it with a pre-configured syntactic template library corresponding to the vertical domain, obtains the syntactic template with the highest matching degree, and generates a corresponding query statement according to the syntactic template; inputs the query statement into the knowledge graph library corresponding to the vertical domain, retrieves the associated subgraph containing the subject entity; extracts several entities with the highest correlation with the subject entity from the associated subgraph, and generates multiple sub-questions after the complex sentence is decomposed according to the associated semantic information corresponding to each of the entities; the present invention utilizes the advantages of knowledge graph related technologies in semantic correlation, and combines syntactic templates to realize the engineered generation of decomposed sub-questions, overcomes the problems of poor flexibility and limited vision existing in the decomposition method that relies solely on rule templates for matching, and has certain theoretical reference and practical reference value. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in 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 paying creative work.

[0037] Figure 1 A flowchart of a vertical domain complex question decomposition method provided in this embodiment;

[0038] Figure 2 A schematic diagram of the topological structure of the recognition model provided in this embodiment;

[0039] Figure 3 This is a schematic diagram of the framework and construction process of the knowledge graph library;

[0040] Figure 4 It is a schematic diagram of the structure of the sub-network in the knowledge graph library;

[0041] Figure 5 A logic block diagram of a vertical domain complex question decomposition device provided in this embodiment;

[0042] Figure 6 A logical block diagram of a computer device provided in this embodiment. DETAILED DESCRIPTION

[0043] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0044] The terms "first", "second", "third", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "including", "having", "corresponding to" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally includes steps or units that are not listed, or optionally includes other steps or units inherent to these processes, methods, products or devices.

[0045] In recent years, the knowledge graph (KG) technology has been widely used in the industry. Its unique semantic network structure can effectively reflect the objective existence of massive data in the real world and fully retain the original semantic information between the data. It has been applied in vertical fields such as business, finance, and medical care, providing strong support for solving high-dimensional and complex nonlinear problems in related fields. Therefore, the present invention proposes to apply the knowledge graph to the complex question decomposition task in the vertical field, which can provide path guidance for the decomposition of complex questions and provide filling basis for the generation of sub-questions after decomposition.

[0046] Syntactic templates refer to summarizing the syntactic rules of human natural language into various fixed paradigms and serving as templates for natural language processing. By summarizing the syntactic characteristics of complex questions in specific vertical fields, corresponding vertical field syntactic templates (including complex questions and sub-question templates) are designed, which can cooperate with the constructed vertical field knowledge graph to achieve the decomposition task of complex questions in vertical fields.

[0047] Based on the above considerations, the present invention adopts knowledge graph and syntactic template technology, and uses vertical domain knowledge graph to mine semantic information closely related to complex questions, which can provide decomposition path and filling information for complex question decomposition tasks, and realize a vertical domain complex question decomposition method.

[0048] Figure 1 FIG. 1 is a flow chart of a method for decomposing complex questions in a vertical field provided in this embodiment. Figure 1 As shown, the method mainly includes the following steps:

[0049] S1 obtains a complex sentence to be decomposed, and determines the vertical field to which the complex sentence belongs;

[0050] In this embodiment, the vertical field refers to the specific field to which the complex sentence belongs, such as medical, financial, educational and other fields; for the complex sentence to be decomposed, it is first necessary to distinguish the vertical field to which it belongs.

[0051] In an optional implementation, the vertical field to which the complex sentence belongs is determined by a deep learning model, specifically including:

[0052] The complex sentence is input into a pre-trained domain recognition model, and the vertical domain to which the complex sentence belongs is obtained through the domain recognition model; wherein the domain recognition model is trained using multiple samples with vertical domain labels, and each sample is processed into a question-domain mapping pair and then input into the domain recognition model.

[0053] In this embodiment, a large number of complex sentence samples are first obtained and domain labels are added to each sample to form a sentence-domain mapping pair input into the domain recognition model, and the model learns the mapping relationship between complex sentences and domains. The complex sentence to be decomposed is input into the trained domain recognition model, and the domain recognition model predicts and outputs the domain to which the complex sentence to be decomposed belongs. In a specific example, the domain recognition model can adopt the GPT-2 model.

[0054] S2 inputs the complex sentence into the recognition model corresponding to the vertical field, and extracts the subject entity and auxiliary verb type in the complex sentence;

[0055] In this embodiment, corresponding recognition models, knowledge graphs and syntactic templates are pre-set for each vertical field. After determining the field to which the complex question to be decomposed belongs, the complex sentence is first input into the recognition model corresponding to the vertical field, and the subject entity and auxiliary verb type in the complex sentence are extracted through the recognition model.

[0056] Before inputting a complex sentence into the recognition model, the complex sentence needs to be preprocessed. Specifically, the complex sentence is decomposed word by word, and each decomposed word is vectorized to obtain a word vector matrix, which is then input into the recognition model.

[0057] Figure 2 is a schematic diagram of the topological structure of the recognition model provided in this embodiment, see Figure 2 The recognition model includes a bidirectional long short-term memory (Bi-LSTM) layer and a conditional random field (CRF) layer; Figure 2 , the extraction process of subject entities and auxiliary verb types is explained;

[0058] The Bi-LSTM layer performs feature extraction and normalization on the input word vector matrix to obtain the predicted label scores of different topic entities corresponding to the complex sentence;

[0059] The CRF layer contextualizes the predicted label scores and obtains the topic entity with the highest predicted label score and its corresponding auxiliary verb type.

[0060] For example, the subject entity recognition of the complex question in the consulting field "What are the key points of the consulting report?" The recognition results are: subject entity - consulting report, auxiliary verb type - action type;

[0061] S3 combines the subject entity and the auxiliary verb type to form a feature expression, matches the feature expression with a pre-configured syntactic template library corresponding to the vertical field, obtains the syntactic template with the highest matching degree, and generates a corresponding knowledge graph query statement according to the syntactic template;

[0062] In this embodiment, according to the syntactic rules of vertical fields and in accordance with the "5W1H" combination principle (time When, place Where, subject Which, attribute What, reason Why, method How), a syntactic template for decomposing complex questions in vertical fields is designed.

[0063] The subject entity and auxiliary verb type output by the recognition model are combined to form a feature expression, and the feature expression of the vertical field complex question is matched with the pre-established vertical field complex question syntactic template library to obtain the matched syntactic template and the corresponding Cypher query statement. Taking the above vertical field complex question as an example, the replaced feature expression is "consulting report + action class", and the matched syntactic template is the prediction method class How, then the corresponding Cypher query statement is "MATCH (m: {name: 'consulting report'}) where (m)-[r]-(n), return n".

[0064] S4 inputs the query statement into the knowledge graph library corresponding to the vertical field, and retrieves the associated subgraph corresponding to the subject entity;

[0065] The knowledge graph library is designed to describe the objective connections between concepts, entities, and events in specific fields (such as medical, financial, and educational). In vertical domain knowledge graphs, triples are the most basic form of expression, mapping the entities and concepts that exist objectively in the real world and the relationships between them into the form of <entity-relationship-entity> or <entity-attribute-attribute value>, and through the interaction of a large number of relationships (attributes) between entities, the value of data (information) is sublimated from quantity to quality. Its construction and function framework are as follows: Figure 3 From a structural point of view, the knowledge graph is essentially a large-scale semantic network, in which the sub-network composed of an entity and its associated entities (attributes) is called the semantic association information sub-graph, as shown in Figure 4 shown.

[0066] Input the Cypher query statement "MATCH(m:{name:'consulting report'})where(m)-[r]-(n),returnn" into the knowledge graph, and retrieve the associated subgraph corresponding to the subject entity "consulting report" from the knowledge graph.

[0067] S5 extracts several entities with the highest correlation with the subject entity from the correlation subgraph, and generates multiple sub-questions after the complex sentence is decomposed according to the correlation semantic information corresponding to each of the entities.

[0068] In the associated subgraph extracted in step S4, there are several entities associated with the subject entity "consulting report". It is necessary to screen out several entities with the highest correlation with "consulting report" from the associated entities, and then generate multiple sub-questions after the complex sentence is decomposed according to the associated semantic information corresponding to each entity.

[0069] In an optional implementation, extracting several entities with the highest degree of association with the subject entity from the association subgraph, and generating multiple sub-questions after decomposing the complex sentence according to the association semantic information corresponding to each of the entities, includes:

[0070] First, the importance of all entities associated with the subject entity in the association subgraph is calculated and sorted, and the top N (for example, top 3) entities are extracted as associated entities;

[0071] In this embodiment, the PageRank algorithm is used to calculate the importance of all entities in the associated subgraph, which is defined as follows:

[0072]

[0073] Among them, PR(A) represents the importance score of entity A associated with the subject entity, PR(T1)...PR(T n ) represent entities A and T1...T n The correlation between them, C(T1)...C(T n ) represent entities T1...T n The sum of the associations with all entities, where d is the damping coefficient in the range 0-1.

[0074] Then, the associated entity is used as the root node to traverse the adjacent entities to obtain multiple associated semantic information in triple format, and each of the associated semantic information and the auxiliary verb type is combined to obtain multiple sub-questions after the complex sentence is decomposed.

[0075] For example, the two entities "including" and "analysis" with the highest correlation with the subject entity "consulting report" are extracted from the associated subgraph, and the adjacent entities are traversed with the "including" and "analysis" entities as root nodes to obtain multiple associated semantic information in triple format <consulting report-including-current situation analysis>, <consulting report-analysis-existing problems and their causes>..., and by matching the syntactic template library, they are identified as: <consulting report-including-current situation analysis>-attribute class What, <consulting report-analysis-existing problems and their causes>-method class How. Finally, the associated semantic information and auxiliary verb types are combined to obtain multiple sub-questions after the decomposition of the complex sentence, such as "What content should be considered in the current situation analysis of the consulting report?" "How to analyze the problems existing in the consulting report and their causes?".

[0076] It should be noted that although the operations of the methods of the embodiments of this specification are described in a specific order in the above-mentioned embodiments, this does not require or imply that the operations must be performed in this specific order, or that all the operations shown must be performed to achieve the desired results. On the contrary, the steps depicted in the flowchart can change the order of execution. Additionally or alternatively, some steps can be omitted, multiple steps can be combined into one step for execution, and / or one step can be decomposed into multiple steps for execution.

[0077] This embodiment provides a device for decomposing complex questions in a vertical field, such as Figure 5 As shown, the device includes a determination module, an extraction module, an extraction module, a matching module, a retrieval module and an output module;

[0078] The determination module is configured to obtain a complex sentence to be decomposed and determine the vertical field to which the complex sentence belongs;

[0079] In an optional embodiment, the determination module inputs the complex sentence into a pre-trained domain recognition model, and obtains the vertical domain to which the complex sentence belongs through the domain recognition model; wherein the domain recognition model is trained using multiple samples with vertical domain labels, and each sample is processed into a question-domain mapping pair and then input into the domain recognition model.

[0080] The extraction module is configured to input the complex sentence into the recognition model corresponding to the vertical field, and extract the subject entity and auxiliary verb type in the complex sentence; the extraction module includes a word segmentation unit;

[0081] The word segmentation unit is used to perform word segmentation and vectorization processing on the complex sentence to obtain a word vector matrix and input it into the recognition model;

[0082] The recognition model performs feature extraction and normalization processing on the word vector matrix to obtain predicted label scores of different subject entities corresponding to the complex sentence; and, contextually constrains the predicted label scores to obtain the subject entity with the highest predicted label score and its corresponding auxiliary verb type.

[0083] The matching module is configured to combine the subject entity and the auxiliary verb type to form a feature expression, match the feature expression with a pre-configured syntactic template library corresponding to the vertical field, obtain the syntactic template with the highest matching degree, and generate a corresponding knowledge graph query statement according to the syntactic template;

[0084] The retrieval module is configured to input the query statement into the knowledge graph library corresponding to the vertical field, and retrieve the associated subgraph corresponding to the subject entity;

[0085] The output module is configured to extract several entities with the highest degree of association with the subject entity from the association subgraph, and generate multiple sub-questions after the complex sentence is decomposed according to the association semantic information corresponding to each of the entities; the output module includes:

[0086] A sorting unit, used to calculate the importance of all entities associated with the subject entity in the associated subgraph and sort them, and extract the top N entities as associated entities;

[0087] The traversal unit is used to traverse adjacent entities with the associated entity as a root node to obtain multiple associated semantic information in triple format, and to combine each of the associated semantic information and auxiliary verb types to obtain multiple sub-questions after the complex sentence is decomposed.

[0088] For the specific definition of the vertical field complex question decomposition device, please refer to the definition of the vertical field complex question decomposition method above, which will not be repeated here. Each module in the above-mentioned vertical field complex question decomposition device can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0089] This embodiment also provides an electronic device, such as Figure 6 As shown, it includes at least one processor and at least one memory, wherein a computer program is stored in the memory. When the computer program is executed by the processor, the processor executes the steps of the above-mentioned vertical field complex question decomposition method; in this embodiment, the types of processor and memory are not specifically limited, for example: the processor can be a microprocessor, a digital information processor, an on-chip programmable logic system, etc.; the memory can be a volatile memory, a non-volatile memory, or a combination thereof, etc.

[0090] The electronic device may also communicate with one or more external devices (such as a keyboard, a pointing terminal, a display, etc.), may also communicate with one or more terminals that enable a user to interact with the electronic device, and / or communicate with any terminal (such as a network card, a modem, etc.) that enables the electronic device to communicate with one or more other computing terminals. Such communication may be performed through an input / output (I / O) interface. Furthermore, the electronic device may also communicate with one or more networks (such as a local area network (LAN), a wide area network (WAN) and / or a public network, such as the Internet) through a network adapter.

[0091] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0092] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0093] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A vertical domain complex question decomposition method, characterized by: include: Obtaining a complex sentence to be decomposed, and determining the vertical field to which the complex sentence belongs; Inputting the complex sentence into the recognition model corresponding to the vertical field, extracting the subject entity and auxiliary verb type in the complex sentence; The subject entity and the auxiliary verb type are combined to form a feature expression, the feature expression is matched with a pre-configured syntactic template library corresponding to the vertical field, a syntactic template with the highest matching degree is obtained, and a corresponding knowledge graph query statement is generated according to the syntactic template; Input the query statement into the knowledge graph library corresponding to the vertical field, and retrieve the associated subgraph corresponding to the subject entity; A number of entities with the highest degree of association with the subject entity are extracted from the association subgraph, and a plurality of sub-questions after the complex sentence is decomposed are generated according to the association semantic information corresponding to each of the entities.

2. The vertical domain complex question decomposition method according to claim 1, characterized in that: Determining the vertical field to which the complex sentence belongs includes: Inputting the complex sentence into a pre-trained domain recognition model, and determining the vertical domain to which the complex sentence belongs through the domain recognition model; The domain recognition model is trained using multiple samples with vertical domain labels, and each sample is processed into a question-domain mapping pair and then input into the domain recognition model.

3. The vertical domain complex question decomposition method according to claim 1 or 2, characterized in that: The step of inputting the complex sentence into the recognition model corresponding to the vertical field and extracting the subject entity and auxiliary verb type in the complex sentence includes: Perform word segmentation and vectorization on the complex sentence to obtain a word vector matrix; Performing feature extraction and normalization processing on the word vector matrix to obtain predicted label scores of different subject entities corresponding to the complex sentence; The predicted label scores are subjected to context constraints to obtain the subject entity with the highest predicted label score and its corresponding auxiliary verb type.

4. The vertical domain complex question decomposition method according to claim 1 or 2, characterized in that: The extracting of several entities with the highest degree of association with the subject entity from the association subgraph, and generating multiple sub-questions after decomposing the complex sentence according to the association semantic information corresponding to each of the entities, includes: Calculate the importance of all entities associated with the subject entity in the associated subgraph and sort them, and extract the top N entities as associated entities; The associated entity is used as a root node to perform adjacent entity traversal to obtain multiple associated semantic information in triple format, and each of the associated semantic information and auxiliary verb types are combined to obtain multiple sub-questions after the complex sentence is decomposed.

5. The vertical domain complex question decomposition method according to claim 4, characterized in that: The calculation of the importance of all entities associated with the subject entity in the associated subgraph and the ranking thereof include: The PageRank algorithm is used to calculate the importance of all entities in the associated subgraph, which is defined as follows: in, Represents an entity associated with the subject entity The importance score of Representing entities and The correlation between Representing entities The sum of the associations with all entities, is the damping factor in the range 0-1.

6. A vertical domain complex question decomposition device, characterized in that: include: A determination module, configured to obtain a complex sentence to be decomposed, and determine the vertical field to which the complex sentence belongs; An extraction module, configured to input a complex sentence into a recognition model corresponding to the vertical field, and extract a subject entity and an auxiliary verb type in the complex sentence; A matching module, which is configured to combine the subject entity and the auxiliary verb type to form a feature expression, match the feature expression with a pre-configured syntactic template library corresponding to the vertical field, obtain the syntactic template with the highest matching degree, and generate a corresponding query statement according to the syntactic template; A retrieval module is configured to input the query statement into the knowledge graph corresponding to the vertical field, and retrieve the associated subgraph corresponding to the subject entity; The output module is configured to extract several entities with the highest correlation with the subject entity from the correlation subgraph, and generate multiple sub-questions after the complex sentence is decomposed according to the correlation semantic information corresponding to each of the entities.

7. The vertical domain complex question decomposition device according to claim 6, characterized in that: The determination module inputs the complex sentence into a pre-trained domain recognition model, and determines the vertical domain to which the complex sentence belongs through the domain recognition model; The domain recognition model is trained using multiple samples with vertical domain labels, and each sample is processed into a question-domain mapping pair and then input into the domain recognition model.

8. The vertical domain complex question decomposition device according to claim 6 or 7, characterized in that: The extraction module comprises: A word segmentation unit, used to perform word segmentation and vectorization processing on the complex sentence, obtain a word vector matrix and input it into the recognition model; The recognition model performs feature extraction and normalization processing on the word vector matrix to obtain predicted label scores of different subject entities corresponding to the complex sentence; and, contextually constrains the predicted label scores to obtain the subject entity with the highest predicted label score and its corresponding auxiliary verb type.

9. The vertical domain complex question decomposition device according to claim 6 or 7, characterized in that: The output module comprises: A sorting unit, used to calculate the importance of all entities associated with the subject entity in the associated subgraph and sort them, and extract the top N entities as associated entities; The traversal unit is used to traverse adjacent entities with the associated entity as a root node to obtain multiple associated semantic information in triple format, and to combine each of the associated semantic information and auxiliary verb types to obtain multiple sub-questions after the complex sentence is decomposed.

10. An electronic device, characterized in that: The method comprises at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the computer program is executed by the processing unit, the processing unit executes the steps of the method according to any one of claims 1 to 5.

Images