A small sample field knowledge text expansion method and system

Abstract

The application discloses a small sample field knowledge text expansion method and system, which is applied to the technical field of text processing and comprises the following steps: performing semantic recognition on all to-be-classified knowledge texts in a target field corresponding to a small sample field to obtain a relationship chain of each knowledge element; classifying all chains based on text attribute data of the target field to obtain a chain set corresponding to each type of text attribute; analyzing existing knowledge elements in the small sample field to obtain missing knowledge chains; performing multi-dimensional search attribute data based on the small sample field; screening the chains in the knowledge chain set to obtain a target knowledge chain matched with the search attribute data; when the similarity between the target knowledge chain and the existing knowledge elements meets a condition, filling an initial knowledge chain based on the target knowledge chain to obtain an updated knowledge chain; and generating a new knowledge text based on the updated knowledge chain when the updated knowledge chain meets a verification condition. The application realizes automatic and intelligent expansion of small sample field knowledge texts.

Description

Technical Field

[0001] This invention relates to the field of text processing technology, and in particular to a method and system for expanding domain knowledge text in small samples. Background Technology

[0002] With the popularization of text data processing technology, knowledge augmentation has become a popular choice for many industries to process domain-specific knowledge. Existing knowledge text augmentation methods can only be used for knowledge text processing within the same domain, such as transferring related knowledge within the same domain to expand domain knowledge.

[0003] As data services continue to grow and expand, it often requires the introduction of cross-domain knowledge content. However, existing technologies first need to collect massive amounts of knowledge content from practitioners in different fields, then filter it based on the human experience of reviewers, and finally construct the expanded knowledge text content. The entire process is extremely cumbersome and complex, which is not conducive to promoting the automation and intelligentization of text data processing. Summary of the Invention

[0004] This invention provides a method and system for expanding small-sample domain knowledge texts, so as to realize the automated and intelligent expansion of cross-domain knowledge texts.

[0005] To address the aforementioned technical problems, embodiments of the present invention provide a method for augmenting small-sample domain knowledge text, comprising: Determine the target domain corresponding to the small sample domain, perform semantic recognition on all unclassified knowledge texts under the target domain, and obtain the knowledge element relationship chain of each unclassified knowledge text. Based on the text attribute data of all the knowledge texts to be classified, the relationship chains of all the knowledge elements are classified to obtain the knowledge chain set corresponding to each type of text attribute. The existing knowledge elements in the small sample domain are analyzed to obtain the initial knowledge chain of the small sample domain; Search attribute data is obtained based on the multi-dimensional constraints of the small sample domain; Filter the knowledge element relationship chains in all the knowledge chain sets to obtain the target knowledge chain that matches the search attribute data; When the similarity between the target knowledge chain and the existing knowledge elements satisfies the similarity condition of the small sample domain, the initial knowledge chain is filled based on the target knowledge chain to obtain the updated knowledge chain of the small sample domain. When the updated knowledge chain satisfies the verification conditions of the small sample domain, new knowledge text of the small sample domain is obtained based on the updated knowledge chain, so as to expand the knowledge text of the small sample domain.

[0006] As one preferred embodiment, determining the target domain corresponding to the small sample domain includes: Based on the knowledge data to be expanded in the small sample domain, determine the technology categories to be expanded in the small sample domain; Determine the target technology categories in fields other than the small sample domain; Based on the cross-analysis results of the technology category to be expanded and all the target technology categories, the target domain corresponding to the small sample domain is selected from the other domains.

[0007] As one preferred embodiment, the step of performing semantic recognition on all unclassified knowledge texts in the target domain to obtain the knowledge element relationship chain for each unclassified knowledge text includes: Obtain all knowledge elements to be extracted corresponding to the small sample domain, and construct a keyword set for each knowledge element to be extracted; Based on each set of keywords, semantic recognition is performed on all knowledge texts to be classified in the target domain to obtain a set of knowledge entities for each set of knowledge texts to be classified. Based on each knowledge element to be extracted, entity relationship parsing and associated entity combination are performed sequentially on all knowledge entities in each knowledge entity set to obtain the knowledge element relationship chain of each knowledge text to be classified.

[0008] As one preferred embodiment, the step of sequentially performing entity relationship parsing and associated entity combination on all knowledge entities in each knowledge entity set based on each knowledge element to be extracted, to obtain the knowledge element relationship chain of each knowledge text to be classified, includes: Based on a pre-built knowledge entity relationship model, entity relationship parsing and associated entity combination are performed sequentially on all knowledge entities corresponding to each knowledge element to be extracted in each knowledge entity set to obtain all knowledge elements of each knowledge text to be classified. Based on a pre-built knowledge element relationship model, all knowledge elements of each text to be classified are logically connected to obtain a knowledge element relationship chain for each text to be classified.

[0009] As a preferred embodiment, based on a pre-constructed knowledge entity relationship model, entity relationship parsing and associated entity combination are performed sequentially on all knowledge entities corresponding to each knowledge element to be extracted in each knowledge entity set, resulting in all knowledge elements of each knowledge text to be classified, including: Filter all structured texts from the knowledge texts to be classified, and use the grammar model as the knowledge entity relation model for processing the structured texts; Define all knowledge entities corresponding to each knowledge element to be extracted in the knowledge entity set of the structured text as knowledge entities to be processed, and convert all knowledge entities to be processed into structured word vectors; Determine the part-of-speech tag of each structured word vector and the semantic similarity between any two structured word vectors; Through the entity integration operation of the grammar model, multiple structured word vectors with semantic similarity greater than a preset threshold are integrated into a single structured word vector. By using the part-of-speech relation connection operation of the grammatical model, multiple structured word vectors with part-of-speech dependency relations after entity integration operation are connected to obtain a structured sentence vector. The structured statement vector is transformed into a knowledge statement through the vector decoding operation of the grammar model. The knowledge statement is the knowledge element of the structured text.

[0010] As one preferred embodiment, the logical connection of all knowledge elements in each of the knowledge texts to be classified, based on a pre-constructed knowledge element relationship model, to obtain the knowledge element relationship chain for each of the knowledge texts to be classified, includes: Target knowledge texts within the target domain are filtered, and a graph neural network model is used as the knowledge element relationship model to process all knowledge elements of the target knowledge texts; the knowledge elements of the target knowledge texts meet a preset diversity requirement. Each knowledge element of the target knowledge text is used as a graph node of the graph neural network model, and the initial node features of each graph node are obtained based on the text encoding operation of the graph neural network model. Through various node connection operations of the graph neural network model, a target graph containing the graph nodes, the initial node features, and the initial edges is obtained; The initial node features are updated through the forward propagation operation of the graph neural network model; By combining the inference operations of the graph neural network model, any two nodes with updated features are combined to obtain the knowledge element relationship chain of the target knowledge text.

[0011] As one preferred embodiment, the step of parsing existing knowledge elements in the small sample domain to obtain the initial knowledge chain of the small sample domain includes: Based on the correspondence between existing knowledge elements and the relationship chain of the knowledge elements, update or transform the existing knowledge elements to obtain new knowledge elements in the small sample domain. Based on the element connection relationships of the knowledge element relationship chain, an initial knowledge chain is constructed that includes the new knowledge element and the missing knowledge element.

[0012] As one preferred embodiment, when the similarity between the target knowledge chain and the existing knowledge elements satisfies the similarity condition of the small sample domain, filling the initial knowledge chain based on the target knowledge chain to obtain the updated knowledge chain of the small sample domain includes: Based on the element positional relationship between the initial knowledge chain and all the target knowledge chains, the target knowledge element corresponding to the position of the existing knowledge element in each target knowledge chain is determined; Based on the semantic matching results between the existing knowledge elements and each of the target knowledge elements, a first knowledge chain is determined among all the target knowledge chains; the semantic similarity between the target knowledge elements of the first knowledge chain and the existing knowledge elements satisfies the similarity condition of the small sample domain. Determine the first knowledge element in the first knowledge chain that corresponds to the position of the missing knowledge element; The first knowledge element is filled into the initial knowledge chain to obtain the updated knowledge chain for the small sample domain.

[0013] As one preferred embodiment, the step of filling the initial knowledge chain with the target knowledge chain to obtain the updated knowledge chain for the small sample domain when the similarity between the target knowledge chain and the existing knowledge elements satisfies the similarity condition of the small sample domain includes: The relation classification model is used as the model to validate the updated knowledge chain; Through the encoding operation of the relation classification model, the existing knowledge elements and the first knowledge element are constructed into a digital tensor. The numerical tensor is converted into a summary vector containing element relationships through the semantic fusion operation of the relation classification model. The confidence level of the summary vector belonging to each preset relation category is obtained through the classification decision operation of the relation classification model. Based on the confidence level of the summary vector belonging to each preset relation category, it is determined whether the updated knowledge chain satisfies the verification conditions of the few-sample domain.

[0014] Another embodiment of the present invention provides a small-sample domain knowledge text augmentation system, comprising: The knowledge element relationship chain determination module is used to determine the target domain corresponding to the small sample domain, perform semantic recognition on all unclassified knowledge texts under the target domain, and obtain the knowledge element relationship chain of each unclassified knowledge text. The knowledge chain set determination module is used to classify all the knowledge element relationship chains based on the text attribute data of all the knowledge texts to be classified, and obtain the knowledge chain set corresponding to each type of text attribute. The initial knowledge chain determination module is used to parse the existing knowledge elements in the small sample domain to obtain the initial knowledge chain of the small sample domain. The search attribute data determination module is used to obtain search attribute data based on the multi-dimensional constraints of the small sample domain. The target knowledge chain determination module is used to filter the knowledge element relationship chains in all the knowledge chain sets to obtain the target knowledge chain that matches the search attribute data; The knowledge chain update module is used to fill the initial knowledge chain based on the target knowledge chain when the similarity between the target knowledge chain and the existing knowledge element satisfies the similarity condition of the small sample domain, so as to obtain the updated knowledge chain of the small sample domain. The knowledge text expansion module is used to obtain new knowledge text for the small sample domain based on the updated knowledge chain when the updated knowledge chain satisfies the verification conditions of the small sample domain, so as to expand the knowledge text of the small sample domain. Attached Figure Description

[0015] Figure 1 This is one of the flowcharts illustrating the small-sample domain knowledge text augmentation method provided by this invention; Figure 2 This is the second flowchart illustrating the small-sample domain knowledge text augmentation method provided by this invention; Figure 3 This is a schematic diagram of the structure of the small-sample domain knowledge text expansion system provided by the present invention.

[0016] Figure label: Among them, 301, Knowledge Element Relationship Chain Determination Module; 302, Knowledge Chain Set Determination Module; 303, Initial Knowledge Chain Determination Module; 304, Search Attribute Data Determination Module; 305, Target Knowledge Chain Determination Module; 306, Knowledge Chain Update Module; 307, Knowledge Text Expansion Module. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0018] In the description of this application, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0019] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. The terms "vertical," "horizontal," "left," "right," "upper," "lower," and similar expressions used herein are for illustrative purposes only and do not indicate or imply that the device or component referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0020] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0021] See Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the small-sample domain knowledge text augmentation method provided by the present invention, as shown below. Figure 1 As shown, this embodiment includes steps 100 to 700, and the specific steps are as follows: Step 100: Determine the target domain corresponding to the small sample domain, perform semantic recognition on all the knowledge texts to be classified under the target domain, and obtain the knowledge element relationship chain of each knowledge text to be classified. This embodiment provides a method for determining the target domain corresponding to a small-sample domain, as follows: A global knowledge graph encompassing multiple technical fields is pre-constructed. Each node in the graph represents a technical concept or entity, and edges represent the relationships between nodes. For a small-sample domain (e.g., the drone domain), the set of core concept nodes for the drone domain is located in the graph. Then, a graph embedding algorithm is used to map each technical field or document cluster into a low-dimensional vector. By calculating the cosine similarity or topological similarity between the small-sample domain vector and other domain vectors, the top N domains with the highest similarity are automatically selected as the target domains in this embodiment.

[0022] For example, this method reveals that "attitude control" in the drone field is most relevant not only to aircraft but also to other fields sharing similar "disturbance-feedback-stabilization" control logic, such as active suspension in automobiles, roll reduction in ships, and end effectors in industrial robots. This allows for the construction of a diverse set of target domains, facilitating the selection of target domains corresponding to small sample domains. The advantage of this method lies in its ability to overcome the limitations of traditional classification systems and discover hidden cross-domain relationships.

[0023] This embodiment also provides a method for determining the target domain corresponding to a small sample domain, as follows: Taking the complex terrain adaptive control of humanoid robot legged walking as the application scenario of the small sample domain, there is not much relevant knowledge text in this domain, which belongs to the small sample domain; the target domain is determined from the upper domain (robot control domain) corresponding to the small sample domain. The corresponding target domain can be the autonomous navigation domain of industrial mobile robots, which has more relevant knowledge text than the small sample domain.

[0024] In this embodiment, the knowledge text to be classified in the target domain refers to the existing knowledge literature that has been published in the target domain; the knowledge element relationship chain refers to the smallest unit extracted from a single knowledge text that describes a complete knowledge logic, such as a knowledge chain of "problem-solution-effect".

[0025] In this embodiment, knowledge elements are illustrated using problems, solutions, and effects as examples. The method for semantically recognizing the knowledge texts to be classified within the target domain to obtain the knowledge element relationship chain for each text can be: using an event extraction paradigm based on a combination of a pre-trained language model and a graph neural network. Specifically, for each knowledge text to be classified (such as a document on electric vehicle battery thermal management), firstly, a pre-trained language model fine-tuned on large-scale text (such as SciBERT or PatentBERT) is used to encode the entire text, obtaining a context-aware text representation. Next, a sequence labeling model or span extraction model is used to simultaneously identify all problem entities, solution entities, and effect entities in the text. Then, without relying on manually defined syntactic rules, these entities are used as nodes, and the co-occurrence relationships, referential relationships, or conjunctions of the entities in the text are used as potential edges to construct a text-level entity co-occurrence graph. Finally, a trained graph neural network is used to classify the relationships in this graph, determining whether any "problem-solution-effect" triple constitutes a logically coherent knowledge element relationship chain. This method is more robust and can handle complex sentence structures and non-standard expressions.

[0026] Step 200: Based on the text attribute data of all the knowledge texts to be classified, classify all the knowledge element relationship chains to obtain the knowledge chain set corresponding to each type of text attribute; In this embodiment, text attributes refer to metadata information describing the knowledge text to be classified. By grouping all knowledge element relationship chains through these attributes, a large number of knowledge element relationship chains can be organized in a structured manner, providing a basis for subsequent multi-dimensional screening and avoiding the computational overhead caused by full traversal.

[0027] This embodiment provides a method for classifying knowledge element relationship chains, as follows: It employs multi-dimensional text attributes for combined classification, including document publication time, knowledge category, and document influence level. Specifically, firstly, the document publication time is divided into multiple time intervals, the knowledge category into multiple subcategories, and the influence level into high and low levels. Then, all knowledge element relationship chains are grouped according to the above three-dimensional attribute combinations to obtain a knowledge chain set corresponding to each attribute combination, for example, a knowledge chain set corresponding to a certain time interval, a certain knowledge subcategory, and a higher level.

[0028] This embodiment also provides a method for classifying knowledge element relationship chains, as follows: Knowledge element relationship chains are classified according to the knowledge categories of the knowledge texts to be classified. In addition to standard classification numbers, knowledge categories can also employ dynamic topic clustering based on text semantics. First, a topic model (such as LDA or its variants) is used to model each text segment of all the knowledge texts to be classified, automatically mining several text knowledge topics (e.g., knowledge in the electric vehicle field such as "battery thermal runaway protection," "motor control algorithm," and "lightweight vehicle body design"). Each knowledge element relationship chain is then associated with the knowledge topic to which its to-be-classified knowledge text belongs. The knowledge chain set obtained in this way is no longer a set based on fixed classification numbers, but a dynamically evolving knowledge cluster. For example, not only can a knowledge chain set under a certain knowledge category be obtained, but also a knowledge chain set under a specific knowledge topic can be obtained. These knowledge chain sets serve as the foundation pool for subsequent knowledge transfer and filtering.

[0029] Step 300: Analyze the existing knowledge elements in the small sample domain to obtain the initial knowledge chain of the small sample domain; Existing knowledge elements in a small-sample domain refer to some clearly defined, unresolved knowledge points or desired effects that already exist in the small-sample domain. These constitute the existing knowledge elements in the small-sample domain. The purpose of this embodiment is to organize these scattered existing knowledge elements into a structured skeleton to be completed, i.e., the initial knowledge chain in this embodiment. For example, in the field of drones, existing knowledge elements could be: the knowledge point "complex wind fields at low altitudes in cities cause drone attitude instability," and the effect "achieving highly robust hovering." Then, the corresponding initial knowledge chain could be: "complex wind fields at low altitudes in cities cause attitude instability - missing knowledge - achieving highly robust hovering."

[0030] The analysis of existing knowledge elements can be accomplished through interactive knowledge graph exploration. First, a small-scale knowledge graph is constructed from a limited amount of publicly available knowledge text within a small sample domain. Then, breakpoints—missing knowledge elements—are identified in this graph, represented by nodes having only outgoing or incoming edges. These breakpoints and their context (e.g., constraints surrounding the breakpoints) are extracted. Next, instead of directly generating a linear chain, a query graph pattern centered on the missing knowledge element nodes is generated as the initial knowledge chain. This query graph pattern includes not only the knowledge element nodes to be resolved but also constraint nodes extracted from the limited knowledge text, such as constraints like "weight limits," "power consumption limits," and "real-time requirements." This query graph pattern, rich in constraint information, improves the accuracy of subsequent searches for relevant knowledge in the target domain.

[0031] Step 400: Obtain search attribute data based on the multi-dimensional constraints of the small sample domain; This embodiment transforms analytical intentions or scenario constraints in a small sample domain into computer-executable data filtering conditions. Multi-dimensional constraints can be user-issued instructions to focus the analysis scope, while search attribute data is a machine-readable representation of these instructions. The purpose of this embodiment is to transform unstructured filtering requirements into machine-recognizable query parameters.

[0032] Multi-dimensional constraints can be literature from the last five years, belonging to a specific technical category. These constraints are transformed into key-value pairs for filtering the knowledge chain set, i.e., search attribute data. In one feasible implementation, multi-dimensional constraints can be unstructured or fuzzy technical descriptions. For example, "finding an application of a biological mechanism similar to chameleons adapting to their environment by changing their body surface structure to evade predators in drone stealth technology." This embodiment can utilize a cross-modal or cross-domain semantic understanding model. First, this textual description is transformed into a set of concept vectors and relational constraints, and these are used as part of the search attribute data, rather than extracting specific, superficial keywords. This allows subsequent searches to go beyond surface-level terms like "chameleon" and "stealth," directly searching for knowledge literature related to concepts such as "adaptive surface deformation" and "impedance modulation."

[0033] In this embodiment, the search attribute data can filter out knowledge element relationship chains that do not meet the requirements, thus avoiding irrelevant interference.

[0034] Step 500: Filter the knowledge element relationship chains in all the knowledge chain sets to obtain the target knowledge chain that matches the search attribute data; Based on the search attribute data described above, this embodiment filters out candidate knowledge chains that match the search attribute data from all knowledge chain sets, which are the target knowledge chains in this embodiment.

[0035] In one feasible implementation, knowledge chain filtering can be based not only on metadata matching but also on structured similarity pre-screening. For the query graph pattern constructed in the above embodiments, a lightweight algorithm based on graph matching or subgraph isomorphism is used to quickly scan and pre-screen the global knowledge graph composed of all knowledge chain sets. This algorithm aims to identify target knowledge chains that are similar to the query graph pattern in topology (e.g., both contain a "perturbation" problem node and a "stability" effect node, and both problem nodes are connected to an "energy consumption" constraint node). This structure-based pre-screening may have a faster retrieval speed than calculating semantic similarity and can ensure that the recalled target knowledge chains are logically related, providing a high-quality candidate set for the next step of fine-grained semantic comparison.

[0036] Step 600: When the similarity between the target knowledge chain and the existing knowledge element satisfies the similarity condition of the small sample domain, the initial knowledge chain is filled based on the target knowledge chain to obtain the updated knowledge chain of the small sample domain. This embodiment is used to match transferable target domain knowledge to fill knowledge gaps in a small sample domain. First, it determines whether the target knowledge chain selected from the target domain is compatible with existing knowledge in the small sample domain. If compatible, the information from the target knowledge chain is used to fill the knowledge gaps in the small sample domain. The similarity condition in the small sample domain can be a compatibility criterion, requiring that the corresponding part of the target knowledge chain has a certain degree of correlation or similarity with existing knowledge elements in the small sample domain. For example, the knowledge problem that the target knowledge chain attempts to solve needs to be essentially similar to the knowledge problem to be solved in the small sample domain.

[0037] This embodiment provides an initial knowledge chain filling method, as follows: A Sentence-BERT model fine-tuned with a knowledge text corpus is used to calculate semantic similarity. Specifically, firstly, the question text of the few-sample domain and the question text of each target knowledge chain are encoded into multi-dimensional semantic vectors; then, the cosine similarity between the two vectors is calculated. When the cosine similarity is greater than a preset similarity threshold, the target knowledge chain is determined to meet the similarity condition of the few-sample domain, and knowledge transfer can be performed.

[0038] This embodiment also provides an initial knowledge chain filling method, as follows: Evaluation is performed based on the equivalence of "functional roles" and "causal mechanisms." For example, functional role similarity and causal mechanism equivalence are judged for the existing knowledge element "urban wind disturbance leads to pose instability" from the drone domain and the knowledge element "random road surface excitation leads to vehicle body vibration" from the target knowledge chain of the electric vehicle domain. Functional role similarity analysis identifies that both knowledge elements play the role of "external disturbance source" leading to "platform stability degradation" in their respective knowledge texts. Although their physical domains differ (aerodynamics and structural mechanics), their functional roles are highly consistent. Causal mechanism equivalence analysis examines the underlying principles of the knowledge elements (such as solutions) derived from both. For example, an electric vehicle's "adaptive suspension" uses sensors to detect disturbances, the controller to calculate, and the actuator to apply a reaction force to counteract the disturbances, forming a logical closed loop of "perception-decision-execution." This causal mechanism is abstractly equivalent to the logical closed loop required for drones: "perceiving wind disturbances - calculating attitude - controlling the motor."

[0039] When the combined score of these two judgments exceeds a preset threshold, the similarity condition of the small sample domain is met. At this point, the following knowledge filling operation is performed: the "perception-decision-execution" closed-loop mechanism inherent in the knowledge element "adaptive suspension control logic" in the target knowledge chain is abstracted into a cross-domain transferable solution pattern, and the missing knowledge element in the initial knowledge chain "Urban low-altitude complex wind field leads to attitude instability - ? - Achieving highly robust aerial hovering" is filled in. After filling, the updated knowledge chain can be: "Urban low-altitude complex wind field leads to attitude instability - Adaptive disturbance cancellation mechanism based on multi-sensor fusion perception and model predictive control - Achieving highly robust aerial hovering." This process is not a simple word replacement, but a knowledge transfer based on analogical reasoning.

[0040] Step 700: When the updated knowledge chain satisfies the verification conditions of the small sample domain, new knowledge text of the small sample domain is obtained based on the updated knowledge chain to expand the knowledge text of the small sample domain.

[0041] This embodiment is used to verify the feasibility of knowledge transfer and complete the knowledge expansion in a small-sample domain. The updated knowledge chain is input into the verification model, which determines whether the updated knowledge chain is sufficiently feasible and whether there are any obvious logical contradictions. The verification model can also perform consistency checks on the generated new knowledge text, determining whether the elements in the new knowledge text correspond to the knowledge elements in the updated knowledge chain, thereby reducing the risk of arbitrary interpretations in the generation of new knowledge text. The verification conditions are a crucial step in ensuring the reliability of cross-domain knowledge transfer. They can filter out unreasonable erroneous transfers, avoid introducing incorrect knowledge into the knowledge base of the small-sample domain, and ensure the quality of the expanded knowledge. The verification model is fine-tuned based on labeled data of cross-domain technology transfer, and can accurately determine the feasibility of cross-domain knowledge. The new knowledge text generated after successful verification effectively fills the technological gaps in the small-sample domain, completing the cross-domain knowledge expansion.

[0042] In one feasible implementation, the verification condition is a multi-dimensional comprehensive scoring model that includes "internal logical consistency verification" and "cross-domain evidence closed-loop verification", rather than a single classification model.

[0043] Internal logical consistency verification: The causal inference model verifies whether the updated knowledge chain conflicts with specific constraints in the UAV domain. For example, the model evaluates whether the knowledge element "model predictive control," which requires strong computing power, contradicts the "lightweight and low-power" requirements of the UAV domain. If an irreconcilable logical conflict exists, the score of the updated knowledge chain will be significantly reduced.

[0044] Cross-domain evidence closed-loop verification: This involves tracing back to the target domain to find indirect evidence supporting the updated knowledge chain. This includes not only source knowledge texts but also other knowledge texts within the target domain, determining whether other knowledge texts have used similar knowledge logic to solve similar problems. For example, searching fields like electric vehicles and industrial robots reveals that the "disturbance resistance method based on model predictive control + sensor fusion" has been validated and used in "flexible control of robot joints," in addition to automotive suspension control. These different but related application examples constitute an "evidence cluster," collectively supporting the validity of this updated knowledge chain within the broad category of "mobile platform disturbance resistance."

[0045] Once the validation criteria for the small sample domain are met, the updated knowledge chain generates a new knowledge text for the small sample domain, such as a knowledge analysis report. This process not only expands the knowledge base of the small sample domain, but more importantly, it provides traceable source evidence and understandable transfer logic for each expanded piece of knowledge, achieving credible and interpretable knowledge discovery in the small sample domain.

[0046] In another embodiment of the small-sample domain knowledge text augmentation method provided by the present invention, step 100 specifically includes: Step 110: Based on the knowledge data to be expanded in the small sample domain, determine the technology category to be expanded in the small sample domain; Step 120: Determine the target technology categories in fields other than the small sample domain; Step 130: Based on the cross-analysis results of the technology category to be expanded and all the target technology categories, select the target domain corresponding to the small sample domain from the other domains.

[0047] This embodiment uses the unmanned aerial vehicle (UAV) domain as an example to illustrate the small sample domain. The knowledge data to be expanded in the UAV domain can be extracted from UAV-related text data. This expanded knowledge is used to characterize the current mainstream research directions in UAVs. The classification of the knowledge data to be expanded in the UAV domain includes, but is not limited to, flight control knowledge data, communication knowledge data, and power knowledge data. Specifically, the classification of the knowledge data to be expanded in the UAV domain is determined by semantic recognition of UAV-related text data; that is, identifying text data related to UAV flight control, UAV communication, and UAV power from the UAV-related text data.

[0048] The specific process for determining the categories of technologies to be expanded in the drone field based on the knowledge data to be expanded in the drone field is as follows: Step 1: Obtain publicly available standard technical classification data. The standard technical classification data includes multiple technical categories and the specific technical content under each technical category. For example, the specific technical content under technical category V is aerospace.

[0049] Step 2: Based on the knowledge data to be expanded in the field of drones mentioned above, determine the initial technology categories to be expanded in the field of drones from the standard technology classification data. For example, the initially determined technology categories to be expanded in the field of drones should at least include V (aerospace), J (mechanical) and L (electronic components and information technology).

[0050] Step 3: Screen the initial technology categories to be expanded as determined in Step 2. For example, the initial technology categories to be expanded may also include B (agriculture and forestry) and P (engineering construction). However, if the amount of publicly available knowledge data corresponding to technology category B is the largest compared to other technology categories and exceeds the corresponding threshold, then technology category B is determined to be a relatively mature technology category in the UAV field, indicating that the priority of supplementing technology category B is not as high as other technology categories. If the amount of publicly available knowledge data corresponding to technology category P is the smallest compared to other technology categories and is below the corresponding threshold, then technology category P has relatively low research value in the UAV field. The two initial technology categories to be expanded exemplified in Step 3 need to be eliminated. After the screening in Step 3, the technology categories to be expanded in the UAV field are obtained.

[0051] Based on the technology category determination method described above in this embodiment, the technology categories corresponding to other fields besides drones are determined, which are the target technology categories in this embodiment. For example, the technology categories corresponding to the electric vehicle field include at least J (machinery), L (electronic components and information technology), T (vehicles), and R (road and water transportation); the technology categories corresponding to the computer field include at least F (energy), N (instrumentation), and L (electronic components and information technology).

[0052] After identifying the target technology categories in other fields, based on the cross-analysis results of the technology categories to be expanded in the UAV field and the target technology categories in all other fields, the target fields corresponding to the UAV field are selected from the other fields. The specific process is as follows: Step 1: Perform cross-analysis on the technology categories to be expanded in the UAV field and the target technology categories in other fields to obtain the same technology categories in the UAV field with each other field, i.e., the cross-analysis results in this embodiment. For example, the cross-technology categories between the UAV field and the electric vehicle field include J, L, and F; the cross-technology categories between the UAV field and the computer field include L and F.

[0053] Step 2: Based on the cross-technology categories between the UAV field and each other field, filter the target fields corresponding to the UAV field from the other fields. Prioritize fields with more or greater weight in the cross-technology categories with the UAV field as the target fields corresponding to the UAV field. The weight of these technology categories can be a pre-defined order of importance, for example, L greater than F.

[0054] Step 3: Based on the screening method in Step 2 above, sort the selected target fields according to one or more of the following conditions: the number of cross-technology categories and the weight of cross-technology categories, to obtain the target field ranking result. After obtaining the target field ranking result, execute the technical solution of the present invention from Steps 100 to 700 above for each target field. The embodiments of the present invention are illustrated using the electric vehicle field as an example.

[0055] The scope of the small sample domain and target domain in this invention can be flexibly adjusted. For example, the higher-level domain concept of the drone domain can be the automation domain; the lower-level domain concept of the drone domain can be the multi-rotor drone domain. The higher-level domain concept of the electric vehicle domain can be the automotive domain; the lower-level domain concept of the electric vehicle domain can be the pure electric vehicle domain.

[0056] See Figure 2 , Figure 2 This is a flowchart illustrating another embodiment of the small-sample domain knowledge text augmentation method provided by the present invention, as shown below. Figure 2 As shown, this embodiment includes steps 140 to 160, and the specific steps are as follows: Step 140: Obtain all knowledge elements to be extracted corresponding to the small sample domain, and construct a keyword set for each knowledge element to be extracted; Step 150: Based on each set of keywords, perform semantic recognition on all knowledge texts to be classified in the target domain to obtain a set of knowledge entities for each set of knowledge texts to be classified. Step 160: Based on each knowledge element to be extracted, perform entity relationship parsing and associated entity combination on all knowledge entities in each knowledge entity set in sequence to obtain the knowledge element relationship chain of each knowledge text to be classified.

[0057] This embodiment is a further explanation of the above-mentioned "semantic recognition to obtain the knowledge element relationship chain".

[0058] First, obtain all the knowledge elements to be extracted from the small sample domain, such as the three knowledge elements of problem, solution, and effect. For each knowledge element, construct a keyword set corresponding to each knowledge element based on the terminology system of the small sample domain. For example, the keyword set for the problem knowledge element includes "excessive fluctuation," "walking instability," and "poor terrain adaptability"; the keyword set for the solution knowledge element includes "damping adjustment," "adaptive control," and "suspension adjustment"; and the keyword set for the effect knowledge element includes "reducing fluctuation," "improving stability," and "improving adaptability."

[0059] Then, based on these keyword sets, semantic recognition is performed on all unclassified knowledge texts in the target domain. Keywords contained in each unclassified knowledge text are matched to extract the corresponding knowledge entity set. For example, from a certain unclassified knowledge text, "excessive wheel-end force fluctuation" (problem entity), "adaptive suspension damping adjustment" (solution entity), and "reducing wheel-end force fluctuation" (effect entity) are extracted. Finally, these knowledge entities are parsed and combined to obtain the knowledge element relationship chain of the unclassified knowledge text. The keyword set for each knowledge element to be extracted is an extraction rule customized for a small sample domain. Specific terms in the small sample domain improve the accuracy of semantic recognition and avoid false positives or false negatives in general extraction models. In this embodiment, the keyword set is sourced from existing knowledge literature in the target domain. The content of the keyword set is the core knowledge terminology of the target domain. The role of the keyword set is to improve the accuracy of semantic recognition in a small sample domain and ensure that the extracted knowledge entities meet the needs of the small sample domain.

[0060] In addition to keyword matching, this embodiment also provides a knowledge entity extraction method, as follows: A trained sequence labeling model (such as BiLSTM-CRF or a BERT-based token classification model) is used. This model takes the word vectors of the knowledge text to be classified as input and outputs entity labels (such as questions and solutions) for each word or sentence, thereby extracting all possible knowledge entities sentence by sentence from the knowledge text to be classified, resulting in a knowledge entity set for each knowledge text to be classified.

[0061] Based on each knowledge element to be extracted, entity relationship parsing and associated entity combination are performed sequentially on all knowledge entities in each knowledge entity set to form a structured knowledge element relationship chain. Entity relationship parsing refers to determining whether there are logical knowledge relationships between multiple extracted knowledge entities. For example, there is a "solved" logical relationship between "vehicle vibration" (problem) and "adaptive suspension damping control" (solution). Associated entity combination refers to combining multiple logically related knowledge entities into a complete knowledge chain. For example, the three knowledge entities "vehicle vibration," "adaptive suspension damping control," and "improved ride comfort" are combined through the logical relationships of "solved" and "caused and realized" to form a complete "problem-solution-effect" knowledge element relationship chain. This process accurately extracts structured technological innovation logic from unstructured text.

[0062] In another embodiment of the small-sample domain knowledge text augmentation method provided by the present invention, step 160 specifically includes: Step 161: Based on the pre-built knowledge entity relationship model, perform entity relationship parsing and associated entity combination on all knowledge entities corresponding to each knowledge element to be extracted in each knowledge entity set in turn to obtain all knowledge elements of each knowledge text to be classified. Step 162: Based on the pre-built knowledge element relationship model, logically connect all knowledge elements of each knowledge text to be classified to obtain the knowledge element relationship chain of each knowledge text to be classified.

[0063] This embodiment further details how to construct a logically related knowledge element relationship chain from the identified knowledge entities. This embodiment can be divided into two core sub-steps: forming knowledge elements and connecting knowledge elements.

[0064] Based on a pre-built knowledge entity relationship model, this model parses and combines all knowledge entities belonging to the same knowledge element to be extracted (such as a question element) within each knowledge entity set, resulting in an independent and complete knowledge element. The core function of this model is to identify different expressions describing the same knowledge element in the text to be classified, and to integrate these different knowledge entities into a single, standardized expression. This model can utilize coreference resolution, semantic similarity, and structured dependency relationships between knowledge entities to combine knowledge entities scattered across different locations in the text to be classified into a unified whole.

[0065] After obtaining all knowledge elements for each text to be classified, logical connections are made between these independent knowledge elements based on a pre-built knowledge element relationship model, ultimately generating a knowledge element relationship chain. The core task of this model is to determine logical relationships such as "cause and effect," "purpose-means," or "problem-solution." For example, it determines whether the solution element "adaptive suspension damping control system" is used to solve the problem element "road excitation causing vehicle body vibration instability" or the problem element "excessive suspension system weight." By performing all-to-many relationship reasoning and connections on all knowledge elements, one or more knowledge element relationship chains composed of knowledge elements and logical relationships are ultimately constructed for each text to be classified. This approach decouples element generation from chain construction, improving the accuracy of the two core sub-steps mentioned above.

[0066] In another embodiment of the small-sample domain knowledge text augmentation method provided by the present invention, step 161 specifically includes: Step 161-1: Filter all structured texts in the knowledge texts to be classified, and use the grammar model as the knowledge entity relation model for processing the structured texts; Step 161-2: Define all knowledge entities corresponding to each knowledge element to be extracted in the knowledge entity set of the structured text as knowledge entities to be processed, and convert all knowledge entities to be processed into structured word vectors. Step 161-3: Determine the part-of-speech tag of each structured word vector and the semantic similarity between any two structured word vectors; Step 161-4: Through the entity integration operation of the grammar model, multiple structured word vectors with semantic similarity greater than a preset threshold are integrated into a single structured word vector. Step 161-5: By using the part-of-speech relation connection operation of the grammar model, connect multiple structured word vectors with part-of-speech dependency relations after entity integration operation to obtain a structured sentence vector; Step 161-6: Through the vector decoding operation of the grammar model, the structured statement vector is transformed into a knowledge statement, and the knowledge statement is the knowledge element of the structured text.

[0067] This embodiment provides a processing method for structured text in knowledge text to be classified. In this case, a grammatical model (such as a dependency parsing model) is determined as the knowledge entity relation model for processing structured text. For all knowledge entities belonging to a certain knowledge element in the knowledge entity set of the structured text to be processed, they are first converted into structured word vectors. Then, the part of speech (e.g., noun, verb, and adjective) of each word vector is determined, and the semantic similarity between any two word vectors is calculated.

[0068] The core operation of this embodiment lies in the entity integration operation of the grammatical model. For multiple structured word vectors with semantic similarity greater than a preset threshold (e.g., "small turning radius", "small turning radius", and "flexible steering"), the grammatical model determines that they point to the same knowledge element concept and integrates them into a single structured core word vector, forming a standardized expression. This essentially performs the step of synonym entity unification. Then, through the part-of-speech relation connection operation of the grammatical model, the integrated structured word vectors are combined. This operation, based on the dependency relationship between parts of speech (e.g., the adjective "adaptive" modifies the noun "suspended"; the verb "reduce" governs the noun "vibrate"), connects multiple word vectors with syntactic relations and semantic coherence to form a structured sentence vector expressing a complete knowledge meaning. Then, through the vector decoding operation of the grammatical model, this structured sentence vector is restored to readable natural language text, which is a clearly structured and standardized knowledge element. For example, the word vectors of "adaptive", "suspended", "damping", and "control" scattered in various places are integrated and decoded into the scheme element "adaptive suspension damping control system".

[0069] In another embodiment of the small-sample domain knowledge text augmentation method provided by the present invention, step 162 specifically includes: Step 162-1: Filter the target knowledge text in the target domain, and use the graph neural network model as the knowledge element relationship model for processing all knowledge elements of the target knowledge text; the knowledge elements of the target knowledge text meet the preset diversity requirements; Step 162-2: Take each knowledge element of the target knowledge text as a graph node of the graph neural network model, and obtain the initial node features of each graph node based on the text encoding operation of the graph neural network model; Step 162-3: Through various node connection operations of the graph neural network model, a target graph containing the graph nodes, the initial node features, and the initial edges is obtained; Step 162-4: Update the initial node features through the forward propagation operation of the graph neural network model; Step 162-5: Through the combined reasoning operation of the graph neural network model, combine any two nodes after updating features to obtain the knowledge element relationship chain of the target knowledge text.

[0070] After obtaining all knowledge elements of each knowledge text to be classified based on the pre-constructed knowledge entity relationship model in the above embodiments, the logical relationships of these knowledge elements are connected to obtain the knowledge element relationship chain of each knowledge text to be classified. The knowledge element relationship chain of each knowledge text to be classified can be one, multiple, or zero. Zero chains mean that the corresponding knowledge text to be classified does not have complete knowledge logic.

[0071] This embodiment logically connects all knowledge elements of each knowledge text to be classified based on a pre-built knowledge element relationship model. First, target knowledge texts within the target domain are selected, meaning those whose knowledge elements meet preset diversity requirements. Taking three knowledge elements—"problem," "measure," and "effect"—as an example, the preset diversity requirement can be that there are at least three types of knowledge elements and at least three total knowledge elements. A graph neural network model is used as an example to process the logical relationships between all knowledge elements of the target knowledge text.

[0072] First, each knowledge element of the target knowledge text (e.g., problem P1, solution M1, and effect E1) is used as a graph node in the graph neural network model. Through the text encoding operations built into the graph neural network model, combined with the contextual semantics of the knowledge element in the original text to be classified, initial node features are obtained for each graph node. Then, the graph neural network model performs various node connection operations with the goal of establishing potential relational edges for the graph. These node connection operations can include: fully connected (generating all possible relations) or attention-based connections (allowing the model to automatically learn which node pairs may have relationships). This results in a target graph containing graph nodes, initial node features, and initial edges for all potential connections.

[0073] Next, through the forward propagation operation of the graph neural network model, information is passed and aggregated along the edges of the target graph, thereby updating the feature vector of each graph node. After multiple rounds of propagation, the features of a solution node will not only contain its own semantics but also incorporate information from its potentially related problem and effect nodes. Then, the graph neural network model performs combinatorial reasoning. For any two nodes with updated features, the graph neural network model predicts the probability that a specific relationship exists between them. By judging the joint probability of all possible "problem-solution-effect" combinations, the graph neural network model ultimately decodes the most probable combination, forming one or more accurate knowledge element relationship chains in the target knowledge text. This method effectively prevents incorrect connections between irrelevant problems and solutions.

[0074] In another embodiment of the small-sample domain knowledge text augmentation method provided by the present invention, step 300 specifically includes: Step 310: Based on the correspondence between existing knowledge elements and the knowledge element relationship chain, update or transform the existing knowledge elements to obtain new knowledge elements in the small sample domain. Step 320: Based on the element connection relationship of the knowledge element relationship chain, construct an initial knowledge chain containing the new knowledge element and the missing knowledge element.

[0075] This embodiment provides a specific method for constructing a knowledge chain to be completed based on existing knowledge in a small-sample domain. First, based on the correspondence between existing knowledge elements in the small-sample domain and the knowledge element relationship chain in the target domain, existing knowledge elements are updated or transformed. If the existing knowledge element is effect description text, it can be transformed into a question knowledge element. If the existing knowledge element is question description text but contains logical errors, it is updated to normalize it into a knowledge element similar to the knowledge elements in the knowledge element relationship chain of the target domain. Through update or transformation operations, new knowledge elements in the small-sample domain are obtained.

[0076] Next, based on the logical connection rules of the knowledge element relationship chain (e.g., a complete knowledge element relationship chain should at least contain the logical order of "problem-solution-effect"), an initial knowledge chain is constructed that includes the aforementioned new knowledge elements and explicitly identifies the missing knowledge elements. This initial knowledge chain is a missing knowledge chain to be filled. This initial knowledge chain clearly indicates the search target and filling position for the knowledge elements, for example, the intermediate solution that connects the known problem with the desired effect.

[0077] In another embodiment of the small-sample domain knowledge text augmentation method provided by the present invention, step 600 specifically includes: Step 610: Based on the element positional relationship between the initial knowledge chain and all the target knowledge chains, determine the target knowledge element in each target knowledge chain that corresponds to the position of the existing knowledge element; Step 620: Based on the semantic matching results between the existing knowledge elements and each of the target knowledge elements, determine the first knowledge chain among all the target knowledge chains; the semantic similarity between the target knowledge elements of the first knowledge chain and the existing knowledge elements satisfies the similarity condition of the small sample domain. Step 630: Determine the first knowledge element in the first knowledge chain that corresponds to the position of the missing knowledge element; Step 640: Fill the first knowledge element into the initial knowledge chain to obtain the updated knowledge chain of the small sample domain.

[0078] This embodiment further explains how to determine the similarity between the initial knowledge chain and the target knowledge chain, and how to perform a fill operation on the initial knowledge chain. First, based on the element positional relationships between the initial knowledge chain and each selected target knowledge chain, an alignment relationship for knowledge elements is established. Once the logical position of an existing knowledge element P_target in the initial knowledge chain is identified as corresponding to the logical position of each target knowledge element P_car in the target knowledge chain, the semantic matching result between P_target and each P_car is calculated. Various semantic similarity calculation methods can be used here. For example, a cross-domain semantic matching model based on adversarial training can be used. This model not only calculates surface text similarity but also identifies the high similarity between "urban wind disturbance" and "road surface excitation" in their functional roles of interfering with the stability of dynamic systems.

[0079] The target knowledge chain in this embodiment is defined as the first knowledge chain for all semantic matching scores that satisfy a preset similarity threshold (i.e., the similarity condition in the few-sample domain). Then, the first knowledge element corresponding to the missing knowledge element position in the initial knowledge chain is determined. For example, the scheme element M_car in the first knowledge chain. Then, a knowledge element filling operation is performed. After adaptive transformation, the first knowledge element M_car is filled into the missing position in the initial knowledge chain, thus obtaining the updated knowledge chain for the few-sample domain. Adaptive transformation means not simply copying it verbatim, but extracting the core mechanism of M_car (e.g., "adaptive cancellation control based on real-time disturbance feedback") and filling M_car with more general terms, resulting in the following updated knowledge chain: "Urban wind disturbance causing attitude oscillation – adaptive cancellation control mechanism based on real-time disturbance feedback – all-weather high-reliability hovering". This process completes an evidence-based and traceable cross-domain knowledge transfer.

[0080] In another embodiment of the small-sample domain knowledge text augmentation method provided by the present invention, the step 600 above specifically includes: Step 10: Use the relation classification model as the model to verify the updated knowledge chain; Step 20: Construct a digital tensor by encoding the existing knowledge element and the first knowledge element through the encoding operation of the relation classification model; Step 30: Through the semantic fusion operation of the relation classification model, the digital tensor is converted into a summary vector containing element relations; Step 40: Through the classification decision operation of the relation classification model, obtain the confidence level of the summary vector belonging to each preset relation category; Step 50: Based on the confidence level of the summary vector belonging to each preset relation category, determine whether the updated knowledge chain satisfies the verification conditions of the few-sample domain.

[0081] In this embodiment, the model used to verify whether the updated knowledge chain conforms to the complete knowledge logic can be a relation classification model, including the BERT model, the RoBERTa model, and the ERNIE model. Taking the BERT model as an example, the specific process of verifying the updated knowledge chain will be further explained.

[0082] First, updating the knowledge chain includes the first knowledge element and existing knowledge elements, where the knowledge elements can be one or more pieces of text. For example, the first knowledge element is the text "adopting a read-write separation architecture"; the existing knowledge element is the text "improving database query performance". The task of the BERT model is to determine whether there is a pre-defined logical relationship between the text "adopting a read-write separation architecture" and the text "improving database query performance".

[0083] Second, during the model training phase, the BERT model has already learned a specific input format. To maintain consistency with the model training phase, special markers, such as [CLS] and [SEP], are inserted into the texts "Adopting a read-write separation architecture" and "Improving database query performance." This results in the following formatted input sequence enclosed in double quotes: "[CLS] Adopting a read-write separation architecture. [SEP] Improving database query performance. [SEP]". The [CLS] marker is located at the beginning of the sequence, and its final state is used to summarize the information of the entire input pair for classification; the [SEP] marker is used to separate the two sentences and mark the end of the sequence.

[0084] The formatted input sequence described above is split into sub-token tokens using a tokenizer (e.g., WordPiece) that accompanies the BERT model. For example, the formatted input sequence might be split into: [CLS], “adopt”, “read and write”, “separate”, “architecture”, “.”, [SEP], “database”, “query”, “performance”, “improvement”, “.”, [SEP]. The tokenizer converts each token into a unique ID (i.e., index) corresponding to a predefined vocabulary. Simultaneously, a paragraph ID vector is generated, which indicates whether each token belongs to the first sentence (marked as 0) or the second sentence (marked as 1). The two texts are then converted into three numeric tensors: a token index sequence, a paragraph ID sequence (used to distinguish the two texts), and an attention mask (used to mark which positions are real tokens and which are padding characters).

[0085] Third, in the embedding layer of the BERT model, each token ID in the token index sequence is transformed into a dense vector (word embedding) through the pre-trained embedding matrix. At the same time, paragraph embedding and position embedding are added to form an initial representation sequence for each token, which is then passed through BERT's multi-layer Transformer encoder.

[0086] In each layer of the multi-layer Transformer encoder, a self-attention mechanism allows each token to see and fuse the semantic information of all other tokens in the initial representation sequence, thereby generating a context-sensitive representation. For example, the word "performance" becomes very specific in its vector meaning after encountering "database," "query," and "improvement." After all 12 or 24 layers of deep computation (depending on the model), the output is a final feature vector corresponding to each input token in the initial representation sequence, deeply fused with global contextual information.

[0087] Regarding the [CLS] token, during model training, it has learned to condense the relational semantics of the entire sentence pair into the final output vector H_[CLS] of this token. Therefore, extracting this H_[CLS] vector (usually a 768-dimensional or 1024-dimensional vector), the H_[CLS] vector is a mathematical summary of the relationship between the two texts mentioned above, i.e., a summary vector.

[0088] Fourth, input H_[CLS] into a pre-trained fully connected neural network layer (classification head), which is usually trained together with BERT during fine-tuning. The calculation of this fully connected neural network layer can be simplified to the following formula: Logits = W × H_[CLS] + b, where W and b are the trained weights and bias parameters. The output Logits is a real number vector, and the length of Logits is equal to the preset number of logical relation categories. For example, logical relation categories include "irrelevant", "realization-cause", and "premise-condition".

[0089] The Logits are converted into a probability distribution using the Softmax function. The probability value at each position represents the confidence level of belonging to that category. The final model outputs the logical relation category with the highest probability and its confidence score. For example, the output relation category is "achieve-cause", and its confidence score is 0.95.

[0090] The second part mentioned above refers to the encoding operation of the relation classification model in this embodiment, which obtains multiple digital tensors of existing knowledge elements and the first knowledge element; the third part mentioned above refers to the specific content of the semantic fusion operation in this embodiment, which obtains the summary vector of existing knowledge elements and the first knowledge element; the fourth part mentioned above refers to the specific content of the classification decision operation in this embodiment, which obtains the confidence score of the summary vector belonging to each preset relation category.

[0091] Assuming that the relation category "realization-cause" is the verification condition for existing knowledge elements and the first knowledge element in a small sample domain, when the confidence score of the relation category "realization-cause" is greater than a preset threshold, it is determined that the updated knowledge chain satisfies the verification condition of the small sample domain.

[0092] The small-sample domain knowledge text expansion system provided by this invention is described below. The small-sample domain knowledge text expansion system described below can be referred to in correspondence with the small-sample domain knowledge text expansion method described above.

[0093] Please refer to Figure 3 The present invention also provides a small-sample domain knowledge text augmentation system, comprising: The knowledge element relationship chain determination module 301 is used to determine the target domain corresponding to the small sample domain, perform semantic recognition on all unclassified knowledge texts under the target domain, and obtain the knowledge element relationship chain of each unclassified knowledge text. The knowledge chain set determination module 302 is used to classify all the knowledge element relationship chains based on the text attribute data of all the knowledge texts to be classified, and obtain the knowledge chain set corresponding to each type of text attribute. The initial knowledge chain determination module 303 is used to parse the existing knowledge elements in the small sample domain to obtain the initial knowledge chain of the small sample domain. Search attribute data determination module 304 is used to obtain search attribute data based on the multi-dimensional constraints of the small sample domain; The target knowledge chain determination module 305 is used to filter the knowledge element relationship chains in all the knowledge chain sets to obtain the target knowledge chain that matches the search attribute data. The knowledge chain update module 306 is used to fill the initial knowledge chain based on the target knowledge chain when the similarity between the target knowledge chain and the existing knowledge element satisfies the similarity condition of the small sample domain, so as to obtain the updated knowledge chain of the small sample domain. The knowledge text expansion module 307 is used to obtain new knowledge text in the small sample domain based on the updated knowledge chain when the updated knowledge chain satisfies the verification conditions of the small sample domain, so as to expand the knowledge text in the small sample domain.

[0094] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for expanding domain knowledge text using a small sample size, characterized in that, include: Determine the target domain corresponding to the small sample domain, perform semantic recognition on all unclassified knowledge texts under the target domain, and obtain the knowledge element relationship chain of each unclassified knowledge text. Based on the text attribute data of all the knowledge texts to be classified, the relationship chains of all the knowledge elements are classified to obtain the knowledge chain set corresponding to each type of text attribute. The existing knowledge elements in the small sample domain are analyzed to obtain the initial knowledge chain of the small sample domain; Search attribute data is obtained based on the multi-dimensional constraints of the small sample domain; Filter the knowledge element relationship chains in all the knowledge chain sets to obtain the target knowledge chain that matches the search attribute data; When the similarity between the target knowledge chain and the existing knowledge elements satisfies the similarity condition of the small sample domain, the initial knowledge chain is filled based on the target knowledge chain to obtain the updated knowledge chain of the small sample domain. When the updated knowledge chain satisfies the verification conditions of the small sample domain, new knowledge text of the small sample domain is obtained based on the updated knowledge chain to expand the knowledge text of the small sample domain.

2. The method for expanding domain knowledge text using small samples as described in claim 1, characterized in that, The determination of the target domain corresponding to the small sample domain includes: Based on the knowledge data to be expanded in the small sample domain, determine the technology categories to be expanded in the small sample domain; Determine the target technology categories in fields other than the small sample domain; Based on the cross-analysis results of the technology category to be expanded and all the target technology categories, the target domain corresponding to the small sample domain is selected from the other domains.

3. The method for expanding domain knowledge text using small samples as described in claim 1, characterized in that, The step of performing semantic recognition on all unclassified knowledge texts in the target domain to obtain the knowledge element relationship chain for each unclassified knowledge text includes: Obtain all knowledge elements to be extracted corresponding to the small sample domain, and construct a keyword set for each knowledge element to be extracted; Based on each set of keywords, semantic recognition is performed on all knowledge texts to be classified in the target domain to obtain a set of knowledge entities for each set of knowledge texts to be classified. Based on each knowledge element to be extracted, entity relationship parsing and associated entity combination are performed sequentially on all knowledge entities in each knowledge entity set to obtain the knowledge element relationship chain of each knowledge text to be classified.

4. The method for expanding domain knowledge text using small samples as described in claim 3, characterized in that, The step of performing entity relationship parsing and associated entity combination on all knowledge entities in each knowledge entity set according to each knowledge element to be extracted, to obtain the knowledge element relationship chain of each knowledge text to be classified, includes: Based on a pre-built knowledge entity relationship model, entity relationship parsing and associated entity combination are performed sequentially on all knowledge entities corresponding to each knowledge element to be extracted in each knowledge entity set to obtain all knowledge elements of each knowledge text to be classified. Based on a pre-built knowledge element relationship model, all knowledge elements of each text to be classified are logically connected to obtain a knowledge element relationship chain for each text to be classified.

5. The method for expanding domain knowledge text using small samples as described in claim 4, characterized in that, The pre-built knowledge entity relationship model involves sequentially parsing and combining entity relationships for all knowledge entities corresponding to each knowledge element to be extracted in each knowledge entity set, resulting in all knowledge elements of each knowledge text to be classified, including: Filter all structured texts from the knowledge texts to be classified, and use the grammar model as the knowledge entity relation model for processing the structured texts; Define all knowledge entities corresponding to each knowledge element to be extracted in the knowledge entity set of the structured text as knowledge entities to be processed, and convert all knowledge entities to be processed into structured word vectors; Determine the part-of-speech tag of each structured word vector and the semantic similarity between any two structured word vectors; Through the entity integration operation of the grammar model, multiple structured word vectors with semantic similarity greater than a preset threshold are integrated into a single structured word vector. By using the part-of-speech relation connection operation of the grammatical model, multiple structured word vectors with part-of-speech dependency relations after entity integration operation are connected to obtain a structured sentence vector. The structured statement vector is transformed into a knowledge statement through the vector decoding operation of the grammar model. The knowledge statement is the knowledge element of the structured text.

6. The method for expanding domain knowledge text using small samples as described in claim 4, characterized in that, The pre-constructed knowledge element relationship model logically connects all knowledge elements of each knowledge text to be classified, resulting in a knowledge element relationship chain for each knowledge text to be classified, including: Target knowledge texts within the target domain are filtered, and a graph neural network model is used as the knowledge element relationship model to process all knowledge elements of the target knowledge texts; the knowledge elements of the target knowledge texts meet a preset diversity requirement. Each knowledge element of the target knowledge text is used as a graph node of the graph neural network model, and the initial node features of each graph node are obtained based on the text encoding operation of the graph neural network model. Through various node connection operations of the graph neural network model, a target graph containing the graph nodes, the initial node features, and the initial edges is obtained; The initial node features are updated through the forward propagation operation of the graph neural network model; By combining the inference operations of the graph neural network model, any two nodes with updated features are combined to obtain the knowledge element relationship chain of the target knowledge text.

7. The method for expanding domain knowledge text using small samples as described in claim 1, characterized in that, The step of parsing existing knowledge elements in the small sample domain to obtain the initial knowledge chain of the small sample domain includes: Based on the correspondence between existing knowledge elements and the relationship chain of the knowledge elements, update or transform the existing knowledge elements to obtain new knowledge elements in the small sample domain. Based on the element connection relationships of the knowledge element relationship chain, an initial knowledge chain is constructed that includes the new knowledge element and the missing knowledge element.

8. The method for expanding domain knowledge text using small samples as described in claim 7, characterized in that, The step of filling the initial knowledge chain with the target knowledge chain to obtain the updated knowledge chain for the small sample domain when the similarity between the target knowledge chain and the existing knowledge elements satisfies the similarity condition of the small sample domain includes: Based on the element positional relationship between the initial knowledge chain and all the target knowledge chains, the target knowledge element corresponding to the position of the existing knowledge element in each target knowledge chain is determined; Based on the semantic matching results between the existing knowledge elements and each of the target knowledge elements, a first knowledge chain is determined among all the target knowledge chains; the semantic similarity between the target knowledge elements of the first knowledge chain and the existing knowledge elements satisfies the similarity condition of the small sample domain. Determine the first knowledge element in the first knowledge chain that corresponds to the position of the missing knowledge element; The first knowledge element is filled into the initial knowledge chain to obtain the updated knowledge chain for the small sample domain.

9. The method for expanding domain knowledge text using small samples as described in claim 8, characterized in that, When the similarity between the target knowledge chain and the existing knowledge elements satisfies the similarity condition of the small sample domain, the process of filling the initial knowledge chain based on the target knowledge chain to obtain the updated knowledge chain of the small sample domain includes: The relation classification model is used as the model to validate the updated knowledge chain; Through the encoding operation of the relation classification model, the existing knowledge elements and the first knowledge element are constructed into a digital tensor. The numerical tensor is converted into a summary vector containing element relationships through the semantic fusion operation of the relation classification model. The confidence level of the summary vector belonging to each preset relation category is obtained through the classification decision operation of the relation classification model. Based on the confidence level of the summary vector belonging to each preset relation category, it is determined whether the updated knowledge chain satisfies the verification conditions of the few-sample domain.

10. A small-sample domain knowledge text augmentation system, characterized in that, include: The knowledge element relationship chain determination module is used to determine the target domain corresponding to the small sample domain, perform semantic recognition on all unclassified knowledge texts under the target domain, and obtain the knowledge element relationship chain of each unclassified knowledge text. The knowledge chain set determination module is used to classify all the knowledge element relationship chains based on the text attribute data of all the knowledge texts to be classified, and obtain the knowledge chain set corresponding to each type of text attribute. The initial knowledge chain determination module is used to parse the existing knowledge elements in the small sample domain to obtain the initial knowledge chain of the small sample domain. The search attribute data determination module is used to obtain search attribute data based on the multi-dimensional constraints of the small sample domain. The target knowledge chain determination module is used to filter the knowledge element relationship chains in all the knowledge chain sets to obtain the target knowledge chain that matches the search attribute data; The knowledge chain update module is used to fill the initial knowledge chain based on the target knowledge chain when the similarity between the target knowledge chain and the existing knowledge element satisfies the similarity condition of the small sample domain, so as to obtain the updated knowledge chain of the small sample domain. The knowledge text expansion module is used to obtain new knowledge text for the small sample domain based on the updated knowledge chain when the updated knowledge chain satisfies the verification conditions of the small sample domain, so as to expand the knowledge text of the small sample domain.

Images