Self-evolution knowledge base construction method and device based on multi-algorithm fusion and large model, processor and computer readable storage medium thereof

Abstract

The invention relates to a self-evolution knowledge base construction method based on multi-algorithm fusion and a large model, and the method comprises the following steps: (1) carrying out data cleaning and sorting on input known business text data, and obtaining clean text data; (2) performing topic classification on the cleaned text data by using a core topic classification unit based on LDAWord2vec and a sub-topic classification unit based on SVM to obtain core topics, sub-topics and keywords and word vectors corresponding to the core topics and the sub-topics; (3) carrying out unknown topic classification and element extraction on database data based on a large model, and fusing the unknown topic classification and element extraction with a topic classification result in the step (2); and (4) performing correlation analysis, data integration and version management on the initial knowledge base and the incremental dynamic knowledge, updating the main dynamic knowledge base, and finally forming the dynamic knowledge base. By adopting the self-evolution knowledge base construction method based on multi-algorithm fusion and the large model, the classification precision can be effectively improved.

Claims

1. A method for constructing a self-evolving knowledge base based on multi-algorithm fusion and large models, characterized by: The method comprises the following steps: (1) Clean and organize the input known business text data to obtain clean text data; (2) Use the core topic classification unit based on LDA_Word2vec and the sub-topic classification unit based on SVM to classify the cleaned text data, obtain the core topics, sub-topics and their corresponding keywords and word vectors, and obtain the initial knowledge base; (3) performing unknown topic classification and element extraction on the data in the initial knowledge base based on the large model, and fusing the unknown topic classification results with the topic classification results of step (2) to obtain an incremental dynamic knowledge base; (4) Performing correlation analysis, data integration, and version management on the initial knowledge base and the incremental dynamic knowledge base to form a final dynamic knowledge base.

2. The method for constructing a self-evolving knowledge base based on multi-algorithm fusion and large models according to claim 1 is characterized in that: The step (1) is specifically as follows: Input known business text data and remove useless characters and auxiliary words; integrate the stop word list of Harbin Institute of Technology and Baidu to form a stop word list for business text data processing, which is used to remove stop words; Data preprocessing is completed by unifying word case, avoiding filtering specific words within quotation marks, and preventing the loss of key information.

3. The method for constructing a self-evolving knowledge base based on multi-algorithm fusion and large models according to claim 1 is characterized in that: The core topic classification unit based on LDA_Word2vec is specifically processed as follows: (2.1.a) Use the Word2vec model to train word vectors on the preprocessed data: Use the CBOW continuous bag-of-words model method to predict the central word through the context window, set the window size and word vector dimension, and optimize the log-likelihood function through gradient descent to generate the word relationship matrix M. word =R V×d , used to capture the semantic association of words, where V represents the size of the vocabulary, d represents the dimension of the word vector, and R represents a real number matrix of size V×d; (2.2.a) Use the LDA topic model to process the dataset: set the initial number of topics, and then iterate through Gibbs sampling to calculate the "document-topic probability matrix θ" and the "topic-word probability matrix ”, calculated as: in, Represents the topic labels of words other than the i-th word in the known d-th document The probability that the i-th word belongs to topic k under the condition of the word set W of all documents, Z d,i is the topic label of the i-th word in the d-th document, k is the topic number, k' initial topics, is the topic label of the words in the d-th document except the i-th word, W is the word set of all documents, is the number of words belonging to topic k in the d-th document, is the total number of words in the dth document, Words in topic k d,i The number of occurrences, ω d,i is the i-th word in the d-th document, is the sum of the occurrences of all words in topic k, α is the Dirichlet prior parameter of topic distribution, and β is the Dirichlet prior parameter of word distribution; (2.3.a) By iteratively calculating document weights and word associations, for each topic k, according to the topic-word probability matrix Extract high-frequency keywords and use M word The weighted average of the word vectors in the topic-word vector matrix M is generated topic , the calculation formula is: Among them, u k is the word vector representation of topic k, TopWords k is the set of high-frequency keywords extracted from topic k, is the probability of word t in topic k, v t is the word vector of word t; (2.4.a) Determine the optimal number of core topics for the corresponding interval based on the perplexity calculated by the model. The perplexity (D) is calculated as follows: Where D is the set of all documents, n is the total number of documents in the dataset, p(d) is the model's predicted probability for the dth document, and N(d) is the number of words in the dth document.

4. The method for constructing a self-evolving knowledge base based on multi-algorithm fusion and large models according to claim 3 is characterized in that: The specific processing steps of the sub-topic classification unit based on SVM include: (2.1.b) For each core topic k, extract the word vector set of its keywords {V t |t∈TopWords k }, forming the feature matrix X k ; (2.2.b) Use a one-to-one strategy to implement multi-category partitioning through pairwise classification to build a multi-classifier, and expand the business analysis dimension from the number of core topics K to the number of sub-topics K' × M, where M is the average number of sub-topics per core topic; (2.3.b) The Softmax function is used to output the sub-topic classification probability. If the probability is greater than the preset threshold, it is considered a valid classification. Otherwise, it is marked as data for review.

5. The method for constructing a self-evolving knowledge base based on multi-algorithm fusion and large models according to claim 4 is characterized in that: The specific processing steps for classifying unknown topics in step (3) include: (3.1.a) Attention mechanism optimization: Add a topic mask matrix to the self-attention layer to force the model to focus on semantic segments related to known topics; (3.2.a) Incremental fine-tuning: For newly emerged unknown topic data, after collecting a preset number of labeled samples, set the corresponding learning rate and training epoch to perform incremental fine-tuning on the model; (3.3.a) Uncertainty Quantification: The uncertainty of the classification is assessed through the log probability variance of the model output. Data with variance greater than the preset threshold will automatically enter the review process.

6. The method for constructing a self-evolving knowledge base based on multi-algorithm fusion and large models according to claim 5 is characterized in that: The step (3) of extracting elements specifically includes: (3.1.b) Rule pre-screening: Quickly locate text segments that may contain elements through regular expression matching; (3.2.b) Deep parsing of large models: Utilize the named entity recognition capability of the Qwen-72B to identify entity types in the text and extract the relationships between entities through the relation extraction model; (3.3.b) Multi-round verification mechanism: Perform logical verification on the extracted elements and complete multiple rounds of verification.

7. The method for constructing a self-evolving knowledge base based on multi-algorithm fusion and large models according to claim 6 is characterized in that: The step (4) specifically includes the following steps: (4.1) Vector similarity calculation: The knowledge items in the initial knowledge base and the incremental dynamic knowledge base are vectorized using the Sentence-BERT model and the cosine similarity is calculated. Among them, a and b are the elements corresponding to the knowledge items to be associated in the knowledge base; Ua and Ub are semantic vectors; |Ua|.|Ub| is the vector dot product, which represents the degree of overlap of the same semantic features. The larger the value, the more common semantics; ||Ua||||Ub|| is the vector norm; the potential associated items are determined based on the calculated cosine similarity; (4.2) Knowledge Graph Reasoning: Build a knowledge graph containing topics, keywords, and elements, and use the TransE algorithm to train entity relationship embeddings to infer implicit associations. (4.3) Dynamic weight adjustment: Dynamically adjust the associated weight according to the popularity of the business scenario.

8. A self-evolving knowledge base construction device based on multi-algorithm fusion and large model, characterized in that: The device comprises: a processor configured to execute computer-executable instructions; A memory storing one or more computer-executable instructions, wherein when the computer-executable instructions are executed by the processor, the steps of the method for constructing a self-evolving knowledge base based on multi-algorithm fusion and large model according to any one of claims 1 to 7 are implemented.

9. A self-evolving knowledge base construction processor based on multi-algorithm fusion and large model, characterized in that: The processor is configured to execute computer-executable instructions. When the computer-executable instructions are executed by the processor, the steps of the method for constructing a self-evolving knowledge base based on multi-algorithm fusion and large model described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium, characterized in that A computer program is stored thereon, and the computer program can be executed by a processor to implement the steps of the method for constructing a self-evolving knowledge base based on multi-algorithm fusion and a large model as described in any one of claims 1 to 7.

Images