Intelligent question answering system for online art education

Abstract

The invention discloses an intelligent question-answering system for online art education, and belongs to the technical field of artificial intelligence and art education, and the system comprises a data set, a named entity recognition model, a knowledge graph, a question-answering subsystem and an evaluation subsystem. The data set is used for collecting data in the art education field; the named entity recognition model is used for recognizing entity information in the data set, constructing an entity relation graph and constructing a knowledge graph according to the entity relation graph; the question and answer subsystem retrieves and generates answers according to the knowledge graph; and the evaluation subsystem is used for evaluating and optimizing the system. The intelligent question and answer system constructed through the data set, the named entity recognition model, the knowledge graph, the question and answer subsystem and the evaluation subsystem can provide targeted accurate answers for online art education, and has important practical significance.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence and art education, and in particular relates to an intelligent question-answering system for online art education. Background Art

[0002] With the development of computer technology, more and more technologies are being applied in the field of education. Traditional education is gradually transforming into intelligent education. Higher requirements are being placed on intelligent education in all areas and directions of education, and higher requirements are also being placed on current online education. At present, existing intelligent question-answering systems have been put into practice in some fields, but they can often only answer some basic and encyclopedia questions based on text, and are unable to explore the deep meaning of keywords in user questions, especially in the field of art education. Since the field of art education itself is highly professional, personal art education is often affected by many factors. Many online education users will face difficulties and confusion in choosing art professional courses. Therefore, how to enable the majority of online art education users to obtain a more friendly and professional online education experience and develop an intelligent question-answering system for online art education has important practical significance for improving the quality of online art education. Summary of the Invention

[0003] In view of this, the present invention provides an intelligent question-answering system for online art education.

[0004] The present invention adopts the following technical solutions:

[0005] An intelligent question-answering system for online art education, used for full-scenario teaching, comprising a dataset, a named entity recognition model, a knowledge graph, a question-answering subsystem, and an evaluation subsystem.

[0006] The dataset is used to collect data in the field of art education; the named entity recognition model is used to identify entity information in the dataset and construct an entity relationship graph, and then construct a knowledge graph based on the entity relationship graph; the question-answering subsystem retrieves and generates answers based on the knowledge graph; and the evaluation subsystem is used to evaluate and optimize the system.

[0007] Furthermore, the data set includes: entity name, meaning, quantity, average length and maximum length.

[0008] Furthermore, the entity names and corresponding meanings include: NAME is the name of the musician; TYPE is the type of creation; WORK is the name of the work; FULLN is the full name of the musician; TIME is the time of birth and death; COUNTRY is the nationality; PERIOD is the period.

[0009] Furthermore, the named entity recognition model includes a BERT layer, a BILSTM layer and a CRF layer.

[0010] Furthermore, the BERT layer is used to: preprocess the data of the data set to obtain a labeled data set, and use the labeled data set to perform training on the BERT layer.

[0011] Furthermore, the BILSTM layer is used to: embed semantic representation information of word vectors in the BiLSTM layer for feature extraction; and calculate the maximum probability label corresponding to each word using the softmax function.

[0012] Furthermore, the CRF layer is used to: use the maximum probability label as the CRF layer input to perform label prediction, and obtain the optimal label sequence corresponding to the input sequence through the CRF layer decoding module.

[0013] Furthermore, the construction of the knowledge graph includes: crawling structured data and unstructured data through the Internet; the structured data includes: the musician's name, nationality, period or birth and death time, and the structured data is used to construct the knowledge graph through rule scripts; the unstructured data includes: musicians or works, and the unstructured data is extracted from the artist's profile text through an entity recognition model to form structured data and then construct the knowledge graph.

[0014] Furthermore, the process of retrieving and generating answers includes a RAG process;

[0015] The RAG process includes: an input prompt, which includes a natural language query or question input by the user; a document repository, which is an external database or document collection for retrieving the source of relevant knowledge; a retrieved document, which is a document or content related to the user input retrieved from the document repository, used to provide contextual information and improve the accuracy of the answer; a generator, which uses a pre-trained language model to generate an answer; and a response, which is the answer or text response ultimately returned to the user.

[0016] Furthermore, the process of retrieving and generating answers also includes a GraphRAG process;

[0017] The GraphRAG process includes:

[0018] User query input preparation stage: input natural language questions; named entity recognition, using the BERT model to identify named entities in the query; graph retrieval preparation, combining the identified named entities, associating the query with the knowledge graph, and extracting semantic query information;

[0019] Graph retrieval phase: Knowledge graph query: In the knowledge graph, starting from the identified entity, graph traversal or adjacent node retrieval is performed to extract subgraphs related to the entity; the subgraphs most relevant to the user query are selected from the knowledge graph;

[0020] Graph-enhanced information retrieval stage: Graph structure and text fusion: convert the most relevant subgraph into usable information blocks; text retrieval is used to supplement relevant historical background or detailed description;

[0021] Input construction in the answer generation phase: The natural language question, the most relevant subgraph, relevant document information, and graph structure relationships are spliced ​​into context input; the final answer is formed based on the context input.

[0022] Beneficial effects:

[0023] Through the above technical solution, the present invention realizes the intelligent dialogue interaction function of intelligent question and answer with high accuracy, high degree of automation and intelligence, which can greatly reduce the cost of manual service, improve the response speed of online art education, and enhance customer satisfaction.

[0024] This system uses the coordinated setting of multiple modules including data sets, named entity recognition models, knowledge graphs, question-answering subsystems, and evaluation subsystems to enable the question-answering system to communicate with online art education users more efficiently and provide users with more detailed and professional responses about online art education. This is of great significance for the development and improvement of the level of online art education in my country. DETAILED DESCRIPTION

[0025] It should be understood that the embodiments described are only a portion of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative work are within the scope of protection of the present invention.

[0026] Example 1

[0027] An intelligent question-answering system for online art education, used for full-scenario teaching, comprising a dataset, a named entity recognition model, a knowledge graph, a question-answering subsystem, and an evaluation subsystem.

[0028] The dataset is used to collect data in the field of art education; the named entity recognition model is used to identify entity information in the dataset and construct an entity relationship graph, and then construct a knowledge graph based on the entity relationship graph; the question-answering subsystem retrieves and generates answers based on the knowledge graph; and the evaluation subsystem is used to evaluate and optimize the system.

[0029] Furthermore, the data set includes: entity name, meaning, quantity, average length and maximum length.

[0030] Furthermore, the entity names and corresponding meanings include: NAME is the name of the musician; TYPE is the type of creation; WORK is the name of the work; FULLN is the full name of the musician; TIME is the time of birth and death; COUNTRY is the nationality; PERIOD is the period.

[0031] Furthermore, the named entity recognition model includes a BERT layer, a BILSTM layer and a CRF layer.

[0032] Furthermore, the BERT layer is used to preprocess the data in the dataset to obtain an annotated dataset, which is then used for training on the BERT layer. The learned semantic and contextual information for each word is the original word vector for each word in the input text, and the output is a vector representation of each word in the text that incorporates the semantic information of the entire text.

[0033] Preprocessing includes data cleaning, removing useless characters, removing or replacing special characters, and standardizing text writing; word segmentation, which breaks the text into a series of subword units. Hugging Face's transformers library can be used to automatically handle word segmentation; constructing the input format, including the input ID, attention mask, and segment mask; padding and truncating the input to ensure it meets the model's expected length (usually 512 tokens), which requires padding or truncating; and batch processing, which uses the tokenizer's batch processing feature for multiple samples.

[0034] Furthermore, the BILSTM layer is used to: embed semantic representation information of word vectors in the BiLSTM layer for feature extraction; and calculate the maximum probability label corresponding to each word using the softmax function.

[0035] Furthermore, the CRF layer is used to: use the maximum probability label as the CRF layer input for label prediction, and obtain the optimal label sequence corresponding to the input sequence through the CRF layer decoding module, which is the final effect of named entity recognition.

[0036] Furthermore, the construction of the knowledge graph includes: crawling structured data and unstructured data through the Internet; the structured data includes: the musician's name, nationality, period or birth and death time, and the structured data is used to construct the knowledge graph through rule scripts; the unstructured data includes: musicians or works, and the unstructured data is extracted from the artist's profile text through an entity recognition model to form structured data and then construct the knowledge graph.

[0037] Furthermore, the process of retrieving and generating answers includes a RAG process;

[0038] The RAG process includes: an input prompt, which includes a natural language query or question input by the user; a document repository, which is an external database or document collection for retrieving the source of relevant knowledge; a retrieved document, which is a document or content related to the user input retrieved from the document repository, used to provide contextual information and improve the accuracy of the answer; a generator, which uses a pre-trained language model to generate an answer; and a response, which is the answer or text response ultimately returned to the user.

[0039] Furthermore, the process of retrieving and generating answers also includes a GraphRAG process;

[0040] The GraphRAG process includes:

[0041] User query input preparation stage: input natural language questions; named entity recognition, using the BERT model to identify named entities in the query; graph retrieval preparation, combining the identified named entities, associating the query with the knowledge graph, and extracting semantic query information;

[0042] Graph retrieval phase: Knowledge graph query: In the knowledge graph, starting from the identified entity, graph traversal or adjacent node retrieval is performed to extract subgraphs related to the entity; the subgraphs most relevant to the user query are selected from the knowledge graph;

[0043] Graph-enhanced information retrieval stage: Graph structure and text fusion: convert the most relevant subgraph into usable information blocks; text retrieval is used to supplement relevant historical background or detailed description;

[0044] Input construction in the answer generation phase: The natural language question, the most relevant subgraph, relevant document information, and graph structure relationships are spliced ​​into context input; the final answer is formed based on the context input.

[0045] Example 2

[0046] 1. Dataset, crawling musician information, the data set is as follows:

[0047] Entity Name meaning quantity Average length Maximum length Minimum length Frequency NAME Musician's name 6066 3.29 9 2 31.14％ TYPE Creation Type 2867 2.86 6 2 14.72％ WORK Title of the work 4834 4.81 12 2 24.82％ FULLN Musician's full name 664 7.96 16 3 3.41％ TIME Date of birth and death 877 16.62 22 9 4.50％ COUNTRY Country of Citizenship 1232 2.67 7 2 6.33％ IDENTITY identity 2014 3.64 14 2 10.34％ PERIOD Period 924 3.90 6 2 4.74％

[0048] 2. BERT-BILSTM-CRF named entity model.

[0049] BERT layer: After data preprocessing, the BERT layer is trained using a labeled dataset to learn the semantics and context of words. This is the original word vector for each word in the input text. The output is a vector representation of each word in the text that incorporates the semantic information of the entire text.

[0050] BILSTM layer: The semantic representation of the sentence vector is embedded in the BiLSTM layer for feature extraction. The softmax function is used to calculate the maximum probability label corresponding to each word.

[0051] CRF layer: The above results are used as the input of the CRF layer for label prediction. Through the CRF layer decoding module, the optimal label sequence corresponding to the input sequence is obtained, which is the final effect of named entity recognition.

[0052] Experimental results:

[0053] Training loss: It decreases rapidly from the beginning of training and eventually approaches zero with little fluctuation, indicating that the model fits the training set well.

[0054] Validation loss: It initially dropped rapidly, but gradually stabilized during training, and then showed an upward trend, especially after 2000 steps, when the validation loss increased significantly.

[0055] The model performs well overall, with stable accuracy and F1 score on the validation set.

[0056] The recall rate is higher than the precision rate, which means that the model tends to identify entities more actively, which may lead to a small number of false positives.

[0057] The model has high prediction accuracy for most entity labels, but is still confused for a small number of labels.

[0058] 3. Knowledge graph construction.

[0059] The previous section detailed named entity recognition in text and developed a model through neural network training. This model was used to construct an encyclopedic knowledge graph for musicians. The figure shows the process for constructing the knowledge graph used in this study. The crawled data consists of both unstructured and structured data. Information about musicians' names, nationalities, time periods, and birth and death dates can be directly obtained from the Xinba website. This constitutes structured data and can be directly used to construct the graph using rule-based scripts. Information about musicians and their works is extracted from the artist profiles using a deep learning model. Once the structured data is generated, the graph is constructed.

[0060] 4. Question-answering system based on knowledge graph enhancement.

[0061] Retrieval-augmented generation (RAG) is a technique that improves the output quality of large language models (LLMs) by incorporating real-world information. RAG is a key component of most LLM-based tools. Most RAG methods use vector similarity as a retrieval technique, which we refer to as baseline RAG.

[0062] GraphRAG uses knowledge graphs to significantly improve question answering performance when reasoning about complex information. When reasoning about complex data, GraphRAG demonstrates superior performance to the baseline RAG, especially with the help of knowledge graphs.

[0063] (1)RAG process:

[0064] ①Prompt (input prompt)

[0065] A natural language query or question entered by the user. For example: "When was Beethoven's Ninth Symphony composed?"

[0066] ②Document Store

[0067] An external database or collection of documents containing large amounts of information, such as text paragraphs, articles, or structured data. A document repository is a source for retrieving relevant knowledge.

[0068] ③Retrieved Documents

[0069] Documents or content related to the user input are retrieved from the document repository. For example, when a user asks, "When was Beethoven's Ninth Symphony composed?" the system retrieves documents from the repository containing information about "Beethoven" and "Ninth Symphony." These documents provide contextual information to the generative model, improving the accuracy of the answer.

[0070] ④Generator (Language Model)

[0071] Use a pre-trained language model (such as GPT or ChatGLM) to generate responses. During this process, the language model combines the input prompt and retrieved documents to generate the final output. For example, the response "Beethoven's Ninth Symphony was composed in 1824" can be generated.

[0072] ⑤Response

[0073] The answer or text response that is ultimately returned to the user. For example: "Beethoven's Ninth Symphony was composed in 1824 and is one of his important late works."

[0074] (2) GraphRAG, GraphRAG operation process:

[0075] ① Input preparation stage: User query: input natural language question; Named Entity Recognition (NER): use the BERT model to identify named entities in the query; identify key entities, such as "Beethoven" (person entity); Graph retrieval preparation: combine the identified named entities, associate the query with the knowledge graph, and extract semantic query information.

[0076] ② Graph Retrieval Phase: Knowledge Graph Query: Starting from the identified entity in the knowledge graph, graph traversal or adjacent node search is performed. Subgraphs related to the entity are extracted, for example, nodes such as Beethoven, Symphony, and Piano Sonata. The subgraphs most relevant to the user's query are selected from the knowledge graph to form the information blocks that are input to the subsequent models.

[0077] 3. Graph-enhanced information retrieval: Graph structure and text fusion: This involves converting subgraph structures (entities and relationships) retrieved from the knowledge graph into usable information blocks. For example, consider the entity relationship: Beethoven → composed → Symphony No. 9. Text retrieval: This provides additional historical context or a detailed description of the Symphony No. 9.

[0078] ④ Answer Generation Phase Input Construction: The user question, graph search results, relevant document information, and graph structure relationships are combined into contextual input. ChatGLM2-6b generates answers: Based on the input contextual information, the ChatGLM2-6b model performs natural language generation to produce the final answer.

[0079] The embodiments of the present invention are described in detail above. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core ideas. At the same time, for those skilled in the art, according to the ideas of the present invention, there may be changes in the specific implementation methods and application scopes. In summary, the contents of this specification should not be understood as limiting the present invention.

Claims

1. An intelligent question-answering system for online art education, which is used for full-scenario teaching and is characterized by: The system includes a dataset, a named entity recognition model, a knowledge graph, a question-answering subsystem, and an evaluation subsystem; The dataset is used to collect data in the field of art education; the named entity recognition model is used to identify entity information in the dataset and construct an entity relationship graph, and then construct a knowledge graph based on the entity relationship graph; The question-answering subsystem retrieves and generates answers based on the knowledge graph; The evaluation subsystem is used for evaluating and optimizing the system.

2. The intelligent question-answering system according to claim 1, characterized in that: The data set includes: entity name, meaning, number, average length and maximum length.

3. The intelligent question-answering system according to claim 2, characterized in that: The entity names and their corresponding meanings include: NAME is the name of the musician; TYPE is the type of creation; WORK is the name of the work; FULLN is the full name of the musician; TIME is the date of birth and death; COUNTRY is the nationality; PERIOD is the period.

4. The intelligent question-answering system according to claim 3, characterized in that: The named entity recognition model includes a BERT layer, a BILSTM layer, and a CRF layer.

5. The intelligent question-answering system according to claim 4, characterized in that: The BERT layer is used to preprocess the data of the data set to obtain a labeled data set, and use the labeled data set to perform training on the BERT layer.

6. The intelligent question-answering system according to claim 5, characterized in that: The BILSTM layer is used to: embed the semantic representation information of the word vector in the BiLSTM layer for feature extraction; and calculate the maximum probability label corresponding to each word using the softmax function.

7. The intelligent question-answering system according to claim 6, characterized in that: The CRF layer is used to: use the maximum probability label as the CRF layer input to perform label prediction, and obtain the optimal label sequence corresponding to the input sequence through the CRF layer decoding module.

8. The intelligent question-answering system according to claim 7, characterized in that: The construction of the knowledge graph includes: crawling structured data and unstructured data through the Internet; the structured data includes: the musician's name, nationality, period or birth and death time, and the structured data is used to construct the knowledge graph through rule scripts; the unstructured data includes: musicians or works, and the unstructured data is extracted from the artist's profile text through an entity recognition model to form structured data and then construct the knowledge graph.

9. The intelligent question-answering system according to claim 8, characterized in that: The process of retrieving and generating answers includes a RAG process; The RAG process includes: an input prompt, which includes a natural language query or question input by the user; a document repository, which is an external database or document collection for retrieving the source of relevant knowledge; a retrieved document, which is a document or content related to the user input retrieved from the document repository, used to provide contextual information and improve the accuracy of the answer; a generator, which uses a pre-trained language model to generate an answer; and a response, which is the answer or text response ultimately returned to the user.

10. The intelligent question-answering system according to claim 9, characterized in that: The process of retrieving and generating answers also includes a GraphRAG process; The GraphRAG process includes: User query input preparation stage: input natural language questions; named entity recognition, using the BERT model to identify named entities in the query; graph retrieval preparation, combining the identified named entities, associating the query with the knowledge graph, and extracting semantic query information; Graph retrieval phase: Knowledge graph query: In the knowledge graph, starting from the identified entity, graph traversal or adjacent node retrieval is performed to extract subgraphs related to the entity; the subgraphs most relevant to the user query are selected from the knowledge graph; Graph-enhanced information retrieval stage: Graph structure and text fusion: convert the most relevant subgraph into usable information blocks; text retrieval is used to supplement relevant historical background or detailed description; Input construction in the answer generation phase: The natural language question, the most relevant subgraph, relevant document information, and graph structure relationships are spliced ​​into context input; the final answer is formed based on the context input.