A method and system for automatically constructing a knowledge base with intelligent segmentation and annotation

Abstract

This invention discloses an automated knowledge base construction method and system for intelligent segmentation and annotation, relating to the field of electronic digital data processing technology. The method includes the following steps: inputting a segmentation annotation model, verifying word segmentation annotation results, and constructing an entity recognition knowledge base. This invention segments the current text using a pre-trained language model to obtain segments to be annotated, which are then input into a word segmentation annotation model. The resulting word segmentation annotations are output and compared to determine if the word segmentation annotation model needs optimization. A second comparison and verification is then performed on the optimized word segmentation annotation results. Finally, entity recognition is performed based on the verified word segmentation annotation results to generate a knowledge graph, which is then stored in a knowledge base for automated construction. This improves the accuracy of unstructured document annotation and solves the problem of inaccurate annotation caused by inappropriate annotation model selection in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and in particular to an automated method and system for constructing a knowledge base with intelligent segmentation and annotation. Background Technology

[0002] In order to intelligently segment the original unstructured document content and combine it with annotation to form structured knowledge, thereby building a knowledge base, a method is designed to generate a corresponding knowledge graph after intelligent segmentation and annotation, and input the knowledge graph into the knowledge base to realize the automated construction of the knowledge base. The overall implementation process of existing technologies usually includes four stages, namely the input processing stage, the segmentation model stage, the annotation model stage, and the knowledge modeling stage.

[0003] The input processing stage uses word vector models (Word2Vec, GloVe, FastText) and context vector models (ELMo, BERT Embedding) to convert the original text into vector sequences, which serve as input for subsequent segmentation and labeling models. The segmentation model stage uses various models (such as BiLSTM-CRF or LayoutLM) to take the vectorized text sequence as input and output corresponding paragraph or sentence boundary label sequences. The labeling model stage first performs named entity recognition, which can be achieved using BiLSTM-CRF or BERT-NER, followed by relation extraction and event extraction. Finally, it outputs entity, relation, and event labels from the segmented text fragments. The knowledge modeling stage uses entity disambiguation models (EntityLinking) to unify the same entity across different documents, and generates vectors for entities and relations using knowledge embedding models (TransE, RotatE, etc.) to support knowledge reasoning. It outputs the input entity, relation, and event triples as structured knowledge (stored in a knowledge graph or knowledge base).

[0004] For example, Chinese invention patent CN114444512B discloses an automatic annotation method for natural language domain datasets based on ontology knowledge base, which includes: preparation and preprocessing of ontology knowledge base; extraction of ontology attributes and ontology relations from ontology knowledge base; modeling of natural language expression habits of domain users; construction of natural language template library; filling natural language templates with ontology knowledge base information and automatically annotating data according to task type.

[0005] For example, Chinese invention patent CN113158688B discloses a method, apparatus, device, and storage medium for constructing a domain knowledge base, including: obtaining a seed knowledge base corresponding to a specified domain, wherein the seed knowledge base contains seed sentence patterns labeled under knowledge points in the specified domain; based on the seed sentence patterns in the seed knowledge base, mining domain sentence patterns that can be labeled under knowledge points in the specified domain from a domain sentence pattern library corresponding to the specified domain, and labeling the mined domain sentence patterns to obtain mined sentence patterns labeled under knowledge points in the specified domain; and forming a domain knowledge base corresponding to the specified domain from the seed sentence patterns labeled under knowledge points in the specified domain and the mined sentence patterns.

[0006] The above-mentioned technology has at least the following technical problems:

[0007] Unstructured documents (such as text, reports, legal documents, etc.) lack explicit labels and structure and contain a large amount of natural language content, which makes it complex to extract valuable information and build knowledge bases from them. Moreover, the content of these documents may span multiple domains, topics, and formats. In addition, in sequence labeling tasks, there are strong dependencies between labels (such as named entity recognition and part-of-speech tagging). For example, some entities in the text may depend on the tags in the context (such as place names and organization names). Without effective label dependency modeling, labeling errors may occur. Summary of the Invention

[0008] To address the technical problem of inaccurate annotation caused by inappropriate annotation model selection in existing technologies, this invention provides a method and system for automatically constructing an intelligent segmented annotation knowledge base. The technical solution is as follows:

[0009] On the one hand, a method for automatically constructing a knowledge base with intelligent segmentation and annotation is provided. This method includes: segmenting the current text into segments to be annotated based on a pre-trained language model used for segmenting unstructured documents; inputting each segment to be annotated into a model for word segmentation and annotation, and outputting the corresponding word segmentation and annotation results; comparing and verifying the output word segmentation and annotation results; determining whether to optimize the word segmentation and annotation model to improve the accuracy of the word segmentation and annotation results based on the comparison and verification results; and continuing to perform a second comparison and verification on the optimized word segmentation and annotation results to ensure the reliability of the optimized word segmentation and annotation results; performing entity recognition based on the compared and verified word segmentation and annotation results to generate the corresponding knowledge graph; and storing the knowledge graph in the knowledge base for automated construction.

[0010] On the other hand, an automated knowledge base construction system for intelligent segmentation and annotation is provided. This system includes: a segmentation and annotation model input module, a word segmentation and annotation result verification module, and an entity recognition knowledge base construction module. The segmentation and annotation model input module segments the current text based on a pre-trained language model used for segmenting unstructured documents, obtaining segments to be annotated. Each segment is then input into a word segmentation and annotation model, outputting the corresponding word segmentation and annotation results. The word segmentation and annotation result verification module compares and verifies the output word segmentation and annotation results. Based on the comparison and verification results, it determines whether to optimize the word segmentation and annotation model to improve the accuracy of the results. It then performs a second comparison and verification on the optimized word segmentation and annotation results to ensure their reliability. The entity recognition knowledge base construction module performs entity recognition based on the verified word segmentation and annotation results to generate a corresponding knowledge graph, which is then stored in the knowledge base for automated construction.

[0011] The beneficial effects of the technical solutions provided by the embodiments of the present invention include at least the following:

[0012] 1. The unstructured document is segmented and labeled. First, the current text is segmented based on a pre-trained language model to obtain segments to be labeled, ensuring that the length of the text fragments input to the word segmentation model is moderate to avoid performance degradation due to excessively long inputs. Each segment to be labeled is then input into the word segmentation and labeling model, outputting the corresponding word segmentation and labeling results. This transforms the document into a structured word sequence, laying the foundation for subsequent entity recognition and relation extraction. Next, the output word segmentation and labeling results are compared and verified, and it is determined whether to optimize the word segmentation and labeling model. This forms a feedback mechanism to automatically detect word segmentation errors, thereby improving the accuracy of the word segmentation and labeling results. Then, the optimized word segmentation and labeling results are compared and verified a second time, which not only ensures the reliability of the optimized word segmentation and labeling results but also forms a quality closed loop, gradually improving the stability of the word segmentation module. Finally, entity recognition is performed based on the verified word segmentation and labeling results to generate the corresponding knowledge graph. The knowledge graph is stored in the knowledge base for automated construction, ensuring more accurate entity recognition and reducing the error rate of entity recognition.

[0013] 2. If the results of HMM and CRF word segmentation are completely consistent, the current word segmentation result is used for entity recognition, ensuring the accuracy of entity recognition. If the results of HMM and CRF word segmentation are inconsistent, the corresponding word segmentation probability data sets for the segment to be labeled in both the HMM and CRF labeling models are obtained and numbered. This allows for more precise location of the inconsistency, providing a clear direction for subsequent optimization. This helps to address the potential word segmentation errors that may occur in different contexts with a single word segmentation model (such as HMM or CRF) in existing technologies, and also allows for locating the corresponding numbered word segmentation probability at the point of inconsistency in the word segmentation results. This method uses probability difference judgment to determine whether to optimize the word segmentation and labeling model. It addresses the problem that traditional methods often rely directly on a single model or use manual rules to make decisions when model results are inconsistent, lacking objective quantitative evidence. By judging the inconsistency of model output through probability difference, it can automatically identify potential performance bottlenecks of the model. Furthermore, the optimization process can continuously improve the model's performance at specific labeling positions, thereby improving the overall word segmentation and labeling accuracy, reducing labeling errors, and effectively avoiding the problem in traditional HMM and CRF applications where the output results usually only take the word segmentation sequence with the highest probability, ignoring the complete probability distribution information, resulting in insufficient interpretability and optimizability of the model results.

[0014] 3. In traditional multi-model word segmentation systems, a unified consistency judgment mechanism is lacking, and conflicts between models cannot be automatically resolved. Therefore, a mechanism is introduced to compare and judge the results of BERT-CRF word segmentation annotation and HMM word segmentation annotation. If the results of BERT-CRF word segmentation annotation and HMM word segmentation annotation are completely consistent, the current word segmentation annotation result is used for entity recognition, which helps to perform entity recognition more stably. If the results of BERT-CRF word segmentation annotation and HMM word segmentation annotation are still inconsistent, the word segmentation probability of the inconsistencies between BERT-CRF word segmentation annotation and HMM word segmentation annotation is obtained and the difference is quantified to obtain the difference in adjustment probability. This helps to solve the shortcomings of existing technologies that cannot accurately identify "conflict points" and their confidence differences, resulting in a lack of targeted optimization. Through the optimization scheme, the system can quickly adapt and adjust, gradually improving the performance and robustness of the word segmentation model. If the difference in adjustment probability is less than the probability judgment value, the segment to be annotated is recorded as an abnormal annotation segment and the corresponding exception is executed. The optimized annotation scheme reduces the probability of the traditional problem of insufficient abnormal segmentation identification, which prevents the system from optimizing for specific abnormal scenarios. The system can quickly adapt and adjust, which helps to gradually improve the performance and robustness of the word segmentation model. If the difference in adjustment probabilities is not less than the probability judgment value, the HMM, CRF, LSTM-CRF, and BERT-CRF word segmentation annotations corresponding to the segment to be annotated are all fed back to the human terminal for verification. This effectively solves the problem of inconsistent presentation of results from multiple models and low efficiency of manual review. By obtaining word segmentation annotation results determined by the human terminal, the system can use the human confirmation results to correct the model's annotations, reducing errors and biases. If the word segmentation annotation results determined by the human terminal are not obtained, the system requests the human terminal to supplement the word segmentation annotations, preventing the backlog of abnormal tasks in the existing technology from affecting the overall processing efficiency and ensuring high-quality output of word segmentation annotations.

[0015] 4. By obtaining the text feature parameters reflecting the text characteristics of the corresponding segments to be labeled, the model's automatic selection quantity is obtained. This quantifies the fit between the segments to be labeled and the model, solving the problem of the lack of a quantitative segment feature extraction mechanism in current technologies, which leads to unfounded model selection. This helps enhance the system's adaptability to diverse inputs. Since traditional methods lack characterization of the model's applicable scope, resulting in unclear model application scenarios, the model's automatic selection quantity is then matched with pre-set model selection intervals. This ensures that different text features are accurately matched to the optimal model within appropriate intervals. If the model automatically selects a value within the BERT-CRF model selection range, then the BERT-CRF model is prioritized for word segmentation and labeling. By selecting the BERT-CRF model based on model fit, the overall efficiency of the word segmentation task is improved. If the model automatically selects a value within the LSTM-CRF model selection range, then the LSTM-CRF model is prioritized for word segmentation and labeling. Furthermore, by using LSTM for feature extraction and CRF for post-processing, errors in word segmentation and labeling can be reduced. This solves the problem in existing systems where models cannot flexibly switch based on input segmentation features, resulting in a trade-off between performance and efficiency. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating an automated knowledge base construction method with intelligent segmentation annotation provided in an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of the process for determining probability differences provided in an embodiment of the present invention;

[0019] Figure 3 This is a flowchart illustrating the implementation of the HMM annotation model provided in this embodiment of the invention;

[0020] Figure 4 This is a flowchart illustrating the implementation of the CRF annotation model provided in this embodiment of the invention;

[0021] Figure 5 This is a schematic diagram of the process for performing secondary comparison verification provided in an embodiment of the present invention;

[0022] Figure 6 This is a schematic diagram of the structure of an automated knowledge base construction system with intelligent segmentation and annotation provided in an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0024] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms “first,” “second,” and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an,” “a,” or “the,” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising,” “including,” or “including,” and similar terms mean that the element or object preceding the word encompasses the element or object listed following the word and its equivalents, without excluding other elements or objects. The terms “connected,” “linked,” or “connected,” and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect.

[0025] It should be noted that the terms "up", "down", "left", "right", "front", and "back" used in this invention are only used to indicate relative positional relationships. When the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0026] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0027] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0028] This invention provides a method and system for automatically constructing a knowledge base with intelligent segmentation and annotation. For example... Figure 1 The diagram illustrates a process flow for an automated knowledge base construction method with intelligent segmentation and annotation. This method's processing flow may include the following steps:

[0029] The segmentation annotation model input is as follows: the current text is segmented based on a pre-trained language model used for segmenting unstructured documents to obtain each segment to be annotated, and each segment to be annotated is input into the word segmentation annotation model to output the corresponding word segmentation annotation results.

[0030] Word segmentation and annotation result verification: The output word segmentation and annotation results are compared and verified. Based on the comparison and verification results, it is determined whether to optimize the word segmentation and annotation model to improve the accuracy of the word segmentation and annotation results. The optimized word segmentation and annotation results are then compared and verified again to ensure the reliability of the optimized word segmentation and annotation results.

[0031] Entity recognition knowledge base construction: Entity recognition is performed based on the results of word segmentation and annotation after comparison and verification to generate the corresponding knowledge graph, and the knowledge graph is stored in the knowledge base for automated construction.

[0032] In this embodiment, the semantic understanding capability of the pre-trained model is used to automatically complete semantically reasonable segmentation, rather than simply truncating by punctuation or word count. This helps to solve the problems of unclear sentence boundaries and chaotic context in unstructured documents, improving the accuracy of subsequent word segmentation and annotation. At the same time, it ensures that the length of the text fragments input to the word segmentation model is moderate, avoiding the performance degradation caused by excessively long inputs. Moreover, it transforms the document into a structured word sequence, laying the foundation for subsequent entity recognition and relation extraction. After optimizing the word segmentation model, it is compared with the validation data again to ensure the reliability of the optimized word segmentation, avoiding "overfitting" or "deterioration after improvement," forming a quality closed loop. This gradually improves the stability and generalization ability of the word segmentation module. Meanwhile, the generated entities, relations, and attributes are stored in the knowledge base, enabling the knowledge base to continuously accumulate and update, and constantly enhancing the intelligence level of the knowledge base.

[0033] Preferably, each segment to be labeled is input into the model used for word segmentation and labeling, and the corresponding word segmentation and labeling results are output. Specifically, each segment to be labeled is input into the HMM (Hidden Markov Model) labeling model and the CRF (Conditional Random Field) labeling model, and the corresponding word segmentation and labeling are output. The word segmentation and labeling output by the HMM labeling model is recorded as HMM word segmentation and labeling output by the CRF labeling model is recorded as CRF word segmentation and labeling.

[0034] In this embodiment, by using both HMM and CRF word segmentation models simultaneously, the advantages of each can be leveraged. HMM is more suitable for reasoning based on statistical features, while CRF can capture more contextual information through global dependency modeling. Obtaining word segmentation results from both models can provide multiple perspectives for subsequent annotation decisions, increasing the reliability of the annotation results. After generating HMM and CRF word segmentation annotations, the consistency and differences of the annotation results can be compared. If the annotation results of the two models are consistent, it means that the models perform similarly on that segment and the annotation quality is high. Otherwise, a preferred analysis is performed. By comparing the annotation results of HMM and CRF, it is helpful to more accurately identify which segments may have annotation errors or inconsistencies, thereby providing a basis for further model optimization. By comparing the annotation results of the two models, the adaptability of different models to different text features can be improved, and the most suitable model can be selected for word segmentation processing.

[0035] like Figure 2 The diagram shows a flowchart of the probability difference determination process provided in this embodiment of the invention. The specific logic is as follows: Matching HMM and CRF word segmentation annotations: If the results of HMM and CRF word segmentation annotations are completely consistent, then the current word segmentation annotation result is used for entity recognition; if the results of HMM and CRF word segmentation annotations are inconsistent, then the corresponding word segmentation annotation probability data groups for the segment to be annotated in the HMM and CRF annotation models are obtained and numbered. Simultaneously, the word segmentation annotation probabilities corresponding to the inconsistent results are located to quantify the differences between the word segmentation annotation probabilities corresponding to the inconsistent results in the HMM and CRF probability data groups, obtaining the corresponding probability differences; if the probability difference is less than the probability determination value, then the segment to be annotated is recorded as an abnormal annotation segment and the corresponding abnormal annotation optimization scheme is executed; if the probability difference is not less than the probability determination value, then the segment to be annotated is re-inputted into the LSTM (Long Short-Term Array). The LSTM (Long Short-Term Memory) network model is used to input the output of the LSTM model into the CRF model to obtain the LSTM-CRF word segmentation and labeling results. Through the above process, the accuracy of the word segmentation and labeling model can be improved, thereby improving the labeling quality of the segments to be labeled.

[0036] Preferably, the output word segmentation and annotation results are compared and verified, and the word segmentation and annotation model is optimized based on the comparison and verification results. The specific process is as follows: HMM word segmentation and annotation and CRF word segmentation and annotation are matched to obtain two matching results.

[0037] The first approach is to use the current word segmentation result for entity recognition if the results of HMM and CRF word segmentation are completely consistent.

[0038] The second approach involves obtaining and numbering the corresponding word segmentation probability data sets for the segment to be labeled in both the HMM and CRF labeling models. Simultaneously, the approach locates the inconsistencies in the word segmentation results and determines the probability difference. This probability difference determination is used to decide whether to optimize the word segmentation model. The word segmentation probability data sets include both HMM and CRF probability data sets. The HMM probability data set covers the probabilities output by the HMM labeling model when performing word segmentation on the segment to be labeled, while the CRF probability data set covers the probabilities output by the CRF labeling model when performing word segmentation on the segment to be labeled.

[0039] It should be noted that the Hidden Markov Model (HMM) is a classic method for sequence labeling, commonly used in tasks such as part-of-speech tagging and named entity recognition. Figure 3 The diagram shows the implementation flowchart of the HMM annotation model provided in this embodiment of the invention. The initial probability (π[S]) represents the initial probability in the Hidden Markov Model (HMM), π[S] represents the probability starting from state S, and the state transition probability is A, which is the probability of state transition marked on each arrow in the diagram. It represents the probability of transitioning to the next hidden state when the character is in a hidden state at a certain moment. In Chinese word segmentation, if the hidden state of the current character is B (beginning of the word), then the hidden state of the next character is likely to be M (middle of the word) or E (end of the word). Each arrow also marks the point of word segmentation annotation, and the corresponding state transition probability is recorded as the output probability when performing word segmentation annotation.

[0040] Specifically, A[S→S] represents the probability of transitioning from state S to S, A[S→B] represents the probability of transitioning from state S to B, and A[B→M] represents the probability of transitioning from state B to M. Additionally, the emission probability represents the probability of generating an observation value given a state, that is, the probability of observing a certain word (or character) in a certain hidden state S. In Chinese word segmentation, the hidden state B (beginning of a word) may tend to emit high-frequency beginning characters (such as "中", "大"), while E (end of a word) may be more likely to emit common ending characters such as "国", "学". B[S→"我"] represents the probability of generating the character "我" in state S, B[S→"爱"] represents the probability of generating the character "爱" in state S, B[B→"北"] represents the probability of generating the character "北" in state B, B[E→"京"] represents the probability of generating the character "京" in state E, and A[M→E] represents the probability of transitioning from state M to E. According to the word segmentation result, the label sequence in the figure is S / S / B-E / B-M-E, which means: "我" is a single-character word, labeled as S, "爱" is a single-character word, labeled as S, "北京" is labeled as B-E, indicating that it is a complete word, and "图书馆" is labeled as B-M-E, indicating that "图" is the beginning of the word, "书" is in the middle of the word, and "馆" is the end of the word. Among them, regarding the states, S is the beginning of a word (such as "我"), M is in the middle of a word (such as "书"), E is the end of a word (such as "京"), and B is also the beginning of a word (such as "北").

[0041] Figure 3 In the figure, the complete process of "Chinese character sequence input, hidden state transition, observation value generation, and final word segmentation result" is shown step by step according to "time steps", which more intuitively understands how the HMM achieves Chinese word segmentation through hidden state annotation.

[0042] It should also be supplemented that Figure 4 is the implementation flowchart of the CRF annotation model provided by the embodiments of the present invention. In the figure, the input is a featureized observation sequence (including characters and multi-dimensional features). Here, the features include the features of each character itself (such as "我", "爱", etc.), and the context information of adjacent characters (such as the previous character and the next character). The output is the globally optimal BMES label sequence. This model annotates Chinese word segmentation through a conditional random field (CRF). The input is a sequence containing characters and their related features, and the output is the optimal BMES (beginning of a word, middle of a word, end of a word, single-character word) label sequence.

[0043] Among them, the model structure includes hidden states (BMES), that is, the hidden states of the CRF model. In Figure 4Among them, S, B, M, and E represent different annotation types. Specifically, S represents Single, indicating that the word is a single word and is usually marked as the only character of an independent word. For example, if a word has only one character (such as "我"), then this character will be marked as S; B represents Beginning, indicating the start part of the word and is usually marked as the first character of a word. For example, in Chinese word segmentation, "自然" in "自然语言处理" will be marked as B ("自" is B, and the subsequent "然" will be marked as M); M represents Middle, indicating the middle part of the word, and the middle characters within a word are usually marked as M (for example, both "语" and "言" in "语言" are marked as M); E represents End, indicating the ending part of the word and is usually marked as the last character of the word. For example, "然" in "自然" will be marked as E (indicating the end of the word "自然").

[0044] Each character in the figure has corresponding features, such as the feature of the current character (such as "我"), the previous character (such as "前"), and the next character (such as "爱"); in the CRF model, the relationship between the observed features and the hidden states is modeled through the conditional probability P(Y|X). Among them, the conditional probability is the probability output during word segmentation annotation.

[0045] In this embodiment, when the word segmentation results of HMM and CRF are consistent, the annotation results can be directly used for entity recognition. And the consistent word segmentation results indicate that the model has obtained relatively stable and consistent annotation results, without further optimization, and can directly enter the downstream entity recognition stage, avoiding multiple model optimizations and comparisons, and also reducing unnecessary computational overhead. At the same time, the consistent annotation results show that the performance of the model is relatively reliable, ensuring the accuracy of entity recognition. By obtaining the probability data group, the specific annotation positions of the inconsistencies can be accurately located, providing a clear direction for subsequent optimization. And with the support of specific probability data, the optimization decision is more data-driven, reducing the situation of "blind optimization", enhancing the interpretability and reliability of the model, and the verification link can also provide feedback for subsequent model iterations, continuously improving the robustness and accuracy of the word segmentation annotation model.

[0046] Preferably, the specific process of probability difference determination is as follows:

[0047] Quantify the difference in the word segmentation annotation probabilities corresponding to the numbered results of the inconsistencies in the HMM probability data group and the CRF probability data group, that is, perform subtraction operations on the word segmentation annotation probabilities corresponding to the numbered results of the inconsistencies in the HMM probability data group and the CRF probability data group, respectively obtaining the corresponding probability differences.

[0048] If the difference in probabilities is less than the pre-set probability judgment value used to limit the maximum difference in probabilities between two word segmentation labels, then the segment to be labeled is recorded as an abnormal labeling segment and the corresponding abnormal labeling optimization scheme is executed to eliminate the abnormal situation of word segmentation labeling.

[0049] Specifically, the probability determination value is obtained directly from the preset database and used directly. Usually, it is preset by the staff based on experience rules and stored in the preset database in advance.

[0050] If the difference in probabilities is not less than the probability judgment value, the segment to be labeled is re-inputted into the LSTM model, and the output of the LSTM model is then input into the CRF model to obtain the LSTM-CRF word segmentation and labeling results.

[0051] In this embodiment, by calculating the probability difference between HMM and CRF, the difference between their word segmentation and annotation results can be judged more accurately. This also helps to filter out problematic segments and optimize them. If the probability difference is less than the judgment value, and the difference in annotation results is small, the segment is marked as an abnormal annotation segment and an optimization scheme is executed. This helps to eliminate anomalies in the annotation, prevent errors from spreading to subsequent tasks, and ensure the high quality and consistency of word segmentation. Furthermore, LSTM, with its powerful long short-term memory capability, can capture long-term dependencies in text, making it particularly suitable for handling long texts and complex contextual dependencies. This helps to better solve contextual ambiguity problems in word segmentation. By combining LSTM and CRF models, the long-term memory of LSTM and the global modeling capability of CRF can be utilized simultaneously, further enhancing the accuracy and consistency of word segmentation results. LSTM provides sequence context information, while CRF can optimize annotation boundaries based on global information. After abnormal annotation segments are identified and optimized, the quality of the annotation results is improved, thereby enhancing the performance of entity recognition.

[0052] like Figure 5The diagram illustrates the process of secondary comparison verification provided in this embodiment of the invention. The specific logic is as follows: If the results of LSTM-CRF word segmentation and HMM word segmentation are completely consistent, then the current word segmentation result is used for entity recognition; if the results of LSTM-CRF word segmentation and HMM word segmentation are still inconsistent, then the word segmentation probabilities of the inconsistencies between the LSTM-CRF word segmentation and HMM word segmentation results are obtained and the differences are quantified to obtain the corresponding difference in optimized probabilities; if the difference in optimized probabilities is less than the probability judgment value, then the segment to be labeled is recorded as an abnormal labeled segment and the corresponding abnormal labeling optimization scheme is executed; if the difference in optimized probabilities is not less than the probability judgment value, then the segment to be labeled is re-inputted into the BERT model, and the output of the BERT model is input into the CRF model to obtain BERT-CRF word segmentation, while continuing to perform comparison verification to determine whether manual feedback is required; through the above process, the accuracy and reliability of labeling the optimized segment to be labeled are ensured.

[0053] Preferably, the optimized word segmentation and tagging results are compared and verified a second time. The specific steps are as follows:

[0054] S1. If the results of LSTM-CRF word segmentation and HMM word segmentation are completely consistent, then the current word segmentation result is used for entity recognition.

[0055] S2, if the results of LSTM-CRF word segmentation and HMM word segmentation are still inconsistent, then obtain the word segmentation probabilities where the results of LSTM-CRF word segmentation and HMM word segmentation are still inconsistent and perform difference quantization respectively. This means that the word segmentation probabilities where the results of LSTM-CRF word segmentation and HMM word segmentation are still inconsistent are calculated by difference to obtain the difference of the corresponding optimized probabilities.

[0056] S3. If the difference in optimization probabilities is less than the probability judgment value, then the segment to be labeled is recorded as an abnormal labeling segment and the corresponding abnormal labeling optimization scheme is executed.

[0057] S4. If the difference in optimization probabilities is not less than the probability judgment value, the segment to be labeled is re-inputted into the BERT (Bidirectional Encoder Representations from Transformers) model, and the output of the BERT model is input into the CRF model to obtain BERT-CRF word segmentation and labeling. At the same time, comparison and verification are continued to determine whether manual feedback is required.

[0058] In this embodiment, consistency verification between LSTM-CRF and HMM ensures the stability of word segmentation results under different modeling mechanisms. If consistent, the system directly proceeds to entity recognition, which can significantly improve processing efficiency. For cases with low probability of discrepancies, the system automatically identifies them as anomalous segments and performs optimization, which helps to quickly correct minor differences and maintain overall consistency. This helps to avoid the interference of error accumulation on subsequent entity recognition. Furthermore, when the differences are significant, BERT-CRF is introduced, which fully utilizes its advantages in semantic understanding and contextual modeling. At the same time, BERT provides deep features, and CRF ensures label consistency. The combination of the two helps to improve the word segmentation quality of complex text. Even if there are still uncertainties at the BERT-CRF level, the system will trigger manual verification, which also ensures the reliability of the final result.

[0059] Preferably, the specific process for continuing comparative verification to determine whether manual feedback is required is as follows:

[0060] S41. If the results of BERT-CRF word segmentation and HMM word segmentation are completely consistent, then the current word segmentation result shall be used for entity recognition.

[0061] S42, if the results of BERT-CRF word segmentation annotation and HMM word segmentation annotation are still inconsistent, then obtain the word segmentation probability of the part where the results of BERT-CRF word segmentation annotation and HMM word segmentation annotation are still inconsistent and perform difference quantification, that is, subtract the word segmentation probability of the part where the results of BERT-CRF word segmentation annotation and HMM word segmentation annotation are still inconsistent to obtain the difference of adjustment probability.

[0062] S43. If the difference in adjustment probabilities is less than the probability judgment value, then the segment to be labeled is recorded as an abnormal labeling segment and the corresponding abnormal labeling optimization scheme is executed.

[0063] S44. If the difference in adjusted probabilities is not less than the probability judgment value, the HMM word segmentation annotation, CRF word segmentation annotation, LSTM-CRF word segmentation annotation and BERT-CRF word segmentation annotation corresponding to the segment to be annotated are all fed back to the human terminal for verification in order to obtain the word segmentation annotation result determined by the human terminal.

[0064] S45. If no word segmentation annotation result is received from the human terminal within the set time interval, then request the human terminal to supplement the word segmentation annotation.

[0065] In this embodiment, when the results of the two models are consistent, no further optimization or manual intervention is required, saving subsequent processing time and computing resources. Consistent results indicate that the two models have fully captured the document structure, which helps to perform entity recognition more stably. Furthermore, by quantifying the probability differences, the reasons for the inconsistencies can be clearly identified, providing more data support to determine the root cause of the inconsistencies. Through optimization schemes, the system can quickly adapt and adjust, gradually improving the performance and robustness of the word segmentation model. At the same time, through accurate annotation by human terminals, the system can use human confirmation results to correct the model's annotations, reducing errors and biases. Supplementary requests ensure that each inconsistent segment can obtain authoritative annotation results, reducing the uncertainty and risk brought by manual judgment. Optimization schemes for abnormal segments also help improve the stability and accuracy of the model when encountering complex scenarios.

[0066] Preferably, the specific details of the anomaly annotation optimization scheme are as follows:

[0067] First, the segments to be labeled corresponding to the anomaly labels are input into the HMM labeling model, CRF labeling model, LSTM-CRF labeling model, and BERT-CRF labeling model respectively, and the corresponding word segmentation and labeling results are output.

[0068] Next, if there are more than a set number of identical word segmentation annotation results, entity recognition is performed based on the corresponding word segmentation annotation results.

[0069] Alternatively, if there are no consistent word segmentation annotation results exceeding the preset number, the word segmentation annotation results output by the HMM annotation model, CRF annotation model, LSTM-CRF annotation model, and BERT-CRF annotation model will all be fed back to the human terminal for verification.

[0070] In this embodiment, statistical models (HMM), conditional random fields (CRF), deep neural networks (LSTM-CRF), and pre-trained language models (BERT-CRF) are all included in the comparison. This fully utilizes the complementarity of multi-source heterogeneous models, which not only improves the robustness of word segmentation and annotation but also reduces the risk caused by single-model bias. Through a consistency judgment mechanism, when more than a certain number of models give the same result, it indicates that the result is highly credible and can be directly used for entity recognition. This "voting" mechanism helps reduce the frequency of manual intervention and improves processing efficiency. Furthermore, when there are significant differences between models, the system automatically triggers manual verification, avoiding annotation distortion caused by automation errors. Through multi-model parallel verification, consensus majority decision-making, and manual verification, a three-level guarantee mechanism is formed to ensure that the entire word segmentation and annotation system can output reliable results under texts of different complexity.

[0071] Preferably, the word segmentation and tagging model is optimized, and then the process further includes:

[0072] The first step is to obtain the text feature parameters of the corresponding segments to be labeled, which reflect the text features of the segments, in order to obtain the model automatic selection quantity that quantifies the degree of fit between the segments to be labeled and the model.

[0073] The second step is to match the automatic model selection with the pre-set model selection intervals, which include the BERT-CRF model selection interval and the LSTM-CRF model selection interval, and the two are continuous.

[0074] It should be added that the technical staff has pre-defined the model selection range based on the characteristics of BERT and LSTM models and empirical rules, and stored it in the preset database. When needed, it can be directly retrieved from the preset database.

[0075] The third step is to select the BERT-CRF model for word segmentation and labeling if the automatically selected amount falls within the selection range of the BERT-CRF model.

[0076] Fourth step: If the number of words automatically selected by the model falls within the selection range of the LSTM-CRF model, then the LSTM-CRF model will be selected first for word segmentation and labeling.

[0077] In this embodiment, by quantifying the text features of the segments to be labeled, the system can select the most suitable model based on the specific content of the segments. By quantifying the fit between the text and the model, blind model selection can be avoided, ensuring that the selected model performs optimally for the current text task. As different text features change, the model selection can be dynamically adjusted, enhancing the system's adaptability to diverse inputs. Simultaneously, the preset model selection interval helps ensure that different text features are accurately matched to the optimal model within a suitable range. The continuity of the model selection interval ensures that the fit value can be smoothly mapped to BERT-CRF or LSTM-CRF, achieving automatic balancing of the two. BERT-CRF is selected based on model fit. The model can leverage its powerful semantic understanding capabilities in suitable scenarios to improve the efficiency and effectiveness of the overall word segmentation task. LSTM-CRF is suitable for handling long-short-term dependencies and complex sequence structures, especially when the text contains complex language structures or named entities. Furthermore, the LSTM-CRF model can improve the annotation effect while memorizing the sequence history information. By dynamically calculating the fitness of text features, it can automatically select the most suitable model for word segmentation and annotation based on the features of the text to be annotated, thereby maximizing the accuracy of word segmentation. Based on the model fitness after the text features are quantified, the most suitable model is selected first, thereby maximizing the accuracy of word segmentation and providing higher quality input for subsequent tasks (such as entity recognition).

[0078] Preferably, the specific method for obtaining the model's automatic selection quantity is as follows:

[0079] Step 1: Obtain the text feature parameters and corresponding text feature weights of the segments to be labeled. The text feature parameters include average text length, memory usage, training time, inference latency, number of training samples, and number of labels. The text feature weights include context dependency length weight, memory consumption weight, computation time weight, inference speed weight, dataset size weight, and task complexity weight.

[0080] It's worth noting that average text length typically refers to the length of contextual information a model can capture when processing text. Different models use different methods to obtain this information. For example, for traditional models (such as HMM or CRF), it can be estimated by examining the feature extraction mechanism. Memory usage refers to the amount of memory a model needs to use when running (including model parameters and intermediate computation results). This can be obtained using memory analysis tools (such as psutil) or by examining the model's memory usage. Training time is measured using Python's time or timeit libraries. Inference latency refers to the model's response time when processing a single input. Inference speed can also be measured using time or timeit. The number of training samples refers to the size of the dataset used for model training, usually determined by the number of samples. The number of labels typically reflects the difficulty of the problem the model solves, indicating the types of labels required for the task. This is usually the number of task categories or entity relations. For example, named entity recognition tasks may involve 5-10 entity types, while relation extraction tasks may involve dozens of relation types.

[0081] Specifically, the text feature weights are pre-set by technical personnel based on the degree of influence of each text feature parameter on model selection. Meanwhile, the sum of the context dependency length weight, memory consumption weight, computation time weight, inference speed weight, dataset size weight, and task complexity weight is 1.

[0082] Step 2: The text feature parameters are classified by model selection and assigned corresponding representative values ​​based on the classification results. The representative values ​​of the text feature parameters include the representative value of context dependency length, memory consumption, computation time, inference speed, dataset size, and task complexity.

[0083] It should be noted that the specific expression for the context-dependent length representing the value C is as follows:

[0084] ;

[0085] Where L represents the average text length, and Lth represents the text length threshold set by the technical staff in the preset database. If a piece of text is very short (such as a microblog or a news headline), LSTM is sufficient; if a piece of text is very long (such as an academic article or a legal provision), BERT is more suitable.

[0086] It should be noted that the specific expression for the memory consumption value M is as follows:

[0087] ;

[0088] Where Mem represents memory usage, which is the product of the number of model parameters and the parameter storage size, and Mth is the memory usage threshold, which is preset by technical personnel and stored in a preset database.

[0089] It should be noted that the specific expression for calculating the time representative value T is as follows:

[0090] ;

[0091] Where t represents the training time and Tth is the training time threshold, which is usually preset by technicians and stored in a preset database.

[0092] It should be noted that the specific expression for the inference speed representative value S is as follows:

[0093] ;

[0094] Where Y represents inference latency and Sth is the inference latency threshold, which is usually preset by technical personnel and stored in a preset database.

[0095] It should be noted that the specific expression for the value D, representing the dataset size, is as follows:

[0096] ;

[0097] Where d represents the number of training samples and Dth represents the dataset size threshold, which is usually preset by technical personnel and stored in a preset database.

[0098] It should be noted that the specific expression for the task complexity value P is as follows:

[0099] ;

[0100] Where B represents the number of tags and Pth represents the tag number threshold, which is usually preset by technical personnel and stored in a preset database.

[0101] Step 3: Based on the text feature weights, the representative values ​​of the text feature parameters of the segment to be labeled are weighted and then aggregated to obtain the model's automatic selection quantity.

[0102] Specifically, the specific constraint expression for the model's automatically selected Score is as follows:

[0103] ;

[0104] In the formula, ωC represents the context dependency length weight, ωM represents the memory consumption weight, ωT represents the computation time weight, ωS represents the inference speed weight, ωD represents the dataset size weight, and ωP represents the task complexity weight.

[0105] In this embodiment, by combining text features with weights, model fit can be automatically calculated, avoiding manual adjustments and intervention, thus improving the intelligence and automation of the model selection process. Simultaneously, by quantifying the fit between the segments to be labeled and different models, the most suitable model can be quickly selected, reducing unnecessary computation. For example, if a certain text has a stronger fit than the LSTM model, the system can directly select the LSTM model instead of the more complex BERT model, thereby improving inference speed and reducing memory consumption. Furthermore, by providing different text feature weights for each model, the system can automatically switch between multiple models according to the needs of different tasks, helping to improve the system's flexibility and adaptability. Through accurate fit calculation, the system can select the best-matching model, improving the accuracy of the corresponding labeling task. Even if the context, domain, or complexity of the task differs, a suitable model can ensure the corresponding labeling quality and improve the accuracy of the segmentation criteria to be labeled.

[0106] Figure 6 This is a schematic diagram of the structure of an automated knowledge base construction system with intelligent segmentation and annotation provided in an embodiment of the present invention. (Refer to...) Figure 6 The system includes a segmentation annotation model input module, a word segmentation annotation result verification module, and an entity recognition knowledge base construction module.

[0107] The segmentation annotation model input module is used to segment the current text based on a pre-trained language model for segmenting unstructured documents to obtain each segment to be annotated, and then input each segment to be annotated into the model for word segmentation annotation, and output the corresponding word segmentation annotation results.

[0108] The word segmentation and annotation result verification module is used to compare and verify the output word segmentation and annotation results. Based on the comparison and verification results, it determines whether to optimize the word segmentation and annotation model to improve the accuracy of the word segmentation and annotation results, and continues to perform a second comparison and verification on the optimized word segmentation and annotation results to ensure the reliability of the optimized word segmentation and annotation results.

[0109] The entity recognition knowledge base construction module is used to perform entity recognition based on the results of word segmentation and annotation after comparison and verification to generate the corresponding knowledge graph, and then store the knowledge graph in the knowledge base for automated construction.

[0110] The following points need to be explained:

[0111] (1) The accompanying drawings of the embodiments of the present invention only involve the structures involved in the embodiments of the present invention. Other structures can refer to the general design.

[0112] (2) For clarity, the thickness of layers or regions is enlarged or reduced in the drawings used to describe embodiments of the invention, i.e., these drawings are not drawn to scale. It is understood that when an element such as a layer, film, region or substrate is referred to as being “above” or “below” another element, the element may be “directly” located “above” or “below” the other element or there may be intermediate elements.

[0113] (3) Where there is no conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other to obtain new embodiments.

[0114] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for automatically constructing a knowledge base with intelligent segmentation and annotation, characterized in that, The method includes: The current text is segmented based on a pre-trained language model used for segmenting unstructured documents to obtain segments to be labeled. Each segment to be labeled is then input into a model used for word segmentation and labeling, and the corresponding word segmentation and labeling results are output. The output word segmentation and annotation results are compared and verified. Based on the comparison and verification results, it is determined whether to optimize the word segmentation and annotation model to improve the accuracy of the word segmentation and annotation results. The optimized word segmentation and annotation results are then compared and verified again to ensure the reliability of the optimized word segmentation and annotation results. Entity recognition is performed based on the results of word segmentation and annotation after comparison and verification to generate the corresponding knowledge graph, and the knowledge graph is stored in the knowledge base for automated construction. The output word segmentation and annotation results are compared and verified, and the word segmentation and annotation model is optimized based on the comparison and verification results. The specific process is as follows: Match HMM and CRF tokenization tags: If the results of HMM word segmentation and CRF word segmentation are completely consistent, then the current word segmentation result will be used for entity recognition. If the results of HMM word segmentation and CRF word segmentation are inconsistent, the corresponding word segmentation probability data group of the segment to be labeled in the HMM labeling model and the CRF labeling model is obtained and numbered. At the same time, the word segmentation probability corresponding to the inconsistent word segmentation results is located to determine the probability difference. The probability difference determination is used to determine whether to optimize the word segmentation labeling model. The word segmentation and annotation probability data set includes an HMM probability data set and a CRF probability data set. The HMM probability data set covers the probabilities output by the HMM annotation model when performing word segmentation annotation on the segment to be annotated, and the CRF probability data set covers the probabilities output by the CRF annotation model when performing word segmentation annotation on the segment to be annotated.

2. The method for automatically constructing a knowledge base with intelligent segmentation and annotation according to claim 1, characterized in that, The process involves inputting each segment to be labeled into a model for word segmentation and labeling, and outputting the corresponding word segmentation and labeling results. The specific details are as follows: Each segment to be labeled is input into the HMM labeling model and the CRF labeling model respectively, and the corresponding word segmentation labels are output. The word segmentation tags output by the HMM annotation model are denoted as HMM word segmentation tags, and the word segmentation tags output by the CRF annotation model are denoted as CRF word segmentation tags.

3. The method for automatically constructing a knowledge base with intelligent segmentation and annotation according to claim 1, characterized in that, The specific process for determining the probability difference is as follows: The differences in the word segmentation annotation probabilities corresponding to the inconsistencies in the HMM probability data set and CRF probability data set are quantified to obtain the corresponding probability differences. If the difference in probability is less than the pre-set probability judgment value used to limit the maximum difference in the probabilities of the two word segmentation labels, then the segment to be labeled is recorded as an abnormal labeling segment and the corresponding abnormal labeling optimization scheme is executed to eliminate the abnormal situation of word segmentation labeling. If the difference in probabilities is not less than the probability judgment value, the segment to be labeled is re-inputted into the LSTM model, and the output of the LSTM model is then input into the CRF model to obtain the LSTM-CRF word segmentation and labeling results.

4. The method for automatically constructing a knowledge base with intelligent segmentation and annotation according to claim 1, characterized in that, The optimized word segmentation and annotation results are then compared and verified a second time. The specific steps are as follows: If the results of LSTM-CRF word segmentation and HMM word segmentation are completely consistent, then the current word segmentation result will be used for entity recognition. If the results of LSTM-CRF word segmentation and HMM word segmentation are still inconsistent, then obtain the word segmentation probabilities of the inconsistencies between the LSTM-CRF word segmentation and HMM word segmentation results and perform difference quantification to obtain the difference of the corresponding optimized probabilities. If the difference in optimization probabilities is less than the probability judgment value, then the segment to be labeled is recorded as an abnormal labeling segment and the corresponding abnormal labeling optimization scheme is executed. If the difference in optimization probabilities is not less than the probability judgment value, the segment to be labeled is re-inputted into the BERT model, and the output of the BERT model is then input into the CRF model to obtain BERT-CRF word segmentation and labeling. At the same time, comparison and verification are continued to determine whether manual feedback is required.

5. The method for automatically constructing a knowledge base with intelligent segmentation and annotation according to claim 4, characterized in that, The specific process for continuing the comparative verification to determine whether manual feedback is needed is as follows: If the results of BERT-CRF word segmentation and HMM word segmentation are completely consistent, then the current word segmentation result will be used for entity recognition. If the results of BERT-CRF word segmentation and HMM word segmentation are still inconsistent, then obtain the word segmentation probability of the inconsistencies between the results of BERT-CRF word segmentation and HMM word segmentation and perform difference quantification to obtain the difference in adjusted probabilities. If the difference in adjustment probabilities is less than the probability judgment value, then the segment to be labeled is recorded as an abnormal labeling segment and the corresponding abnormal labeling optimization scheme is executed. If the difference in the adjusted probabilities is not less than the probability judgment value, then the HMM word segmentation annotation, CRF word segmentation annotation, LSTM-CRF word segmentation annotation and BERT-CRF word segmentation annotation corresponding to the segment to be annotated will all be fed back to the human terminal for verification in order to obtain the word segmentation annotation result determined by the human terminal. If no word segmentation and annotation results are received from the human terminal within the set time interval, then a request will be made to the human terminal to supplement the word segmentation and annotation.

6. The method for automatically constructing a knowledge base with intelligent segmentation and annotation according to claim 5, characterized in that, The specific details of the anomaly annotation optimization scheme are as follows: The segments to be labeled corresponding to the anomaly labels are input into the HMM labeling model, CRF labeling model, LSTM-CRF labeling model and BERT-CRF labeling model respectively, and the corresponding word segmentation and labeling results are output. If there are more than a set number of identical word segmentation annotation results, entity recognition will be performed based on the corresponding word segmentation annotation results. If there are no consistent word segmentation annotation results exceeding the preset number, the word segmentation annotation results output by the HMM annotation model, CRF annotation model, LSTM-CRF annotation model, and BERT-CRF annotation model will all be fed back to the human terminal for verification.

7. The method for automatically constructing a knowledge base with intelligent segmentation and annotation according to claim 1, characterized in that, The optimization of the word segmentation and annotation model is followed by: Obtain the text feature parameters that reflect the text features of the corresponding segment to be labeled, so as to obtain the model automatic selection quantity that quantifies the degree of fit between the segment to be labeled and the model; The model selection variables are matched with pre-set model selection intervals, which include the BERT-CRF model selection interval and the LSTM-CRF model selection interval, and the two are continuous. If the number of words automatically selected by the model falls within the selection range of the BERT-CRF model, then the BERT-CRF model will be selected first for word segmentation and labeling. If the number of words automatically selected by the model falls within the selection range of the LSTM-CRF model, then the LSTM-CRF model will be selected first for word segmentation and labeling.

8. The method for automatically constructing a knowledge base with intelligent segmentation and annotation according to claim 7, characterized in that, The specific method for obtaining the automatically selected quantities in the model is as follows: Obtain the text feature parameters of the segment to be labeled and the corresponding text feature weights. The text feature weights include context dependency length weight, memory consumption weight, computation time weight, inference speed weight, dataset size weight, and task complexity weight. The text feature parameters are classified by model selection, and corresponding representative values ​​are assigned to the text feature parameters according to the classification results. The model automatically selects values ​​by weighting the representative values ​​of the text feature parameters of the segment to be labeled and aggregating them.

9. An automated knowledge base construction system with intelligent segmentation annotation, employing the automated knowledge base construction method with intelligent segmentation annotation as described in any one of claims 1-8, characterized in that, It includes a segmentation annotation model input module, a word segmentation annotation result verification module, and an entity recognition knowledge base construction module; The segmentation annotation model input module is used to segment the current text based on a pre-trained language model for segmenting unstructured documents to obtain each segment to be annotated, and input each segment to be annotated into the model for word segmentation annotation, and output the corresponding word segmentation annotation result. The word segmentation annotation result verification module is used to compare and verify the output word segmentation annotation results, determine whether to optimize the word segmentation annotation model based on the comparison and verification results to improve the accuracy of the word segmentation annotation results, and continue to perform a second comparison and verification on the optimized word segmentation annotation results to ensure the reliability of the optimized word segmentation annotation results; The entity recognition knowledge base construction module is used to perform entity recognition based on the results of word segmentation and annotation after comparison and verification to generate the corresponding knowledge graph, and store the knowledge graph in the knowledge base for automated construction.

Images