Construction and Application of Music Education Knowledge System in Colleges and Universities under the Framework of Knowledge Graph Technology

Knowledge graph is a heterogeneous semantic network that describes the concepts, entities and their relationships in the objective world, and the use of visualization tools as a carrier can express the huge and redundant information in the Internet into a form that is closer to the human cognitive world. The knowledge graph can be divided into the data layer and the schema layer. The data layer is the underlying architecture, and the data is stored in the form of ternary groups, such as (head entity, relationship, tail entity) or (head entity, attribute, attribute value). Entity, as the most basic concept in the knowledge network, can be a specific person, place, organization, or abstract concept. Relationships can be either associations or attributes between entities. The schema layer is on top of the data layer to regulate and constrain it and is at the core of the knowledge graph, and only the knowledge constrained by certain rules and axioms can enter the schema layer. It can be said that the schema layer is the conceptual model and logical foundation of the whole knowledge graph, which is a definition of all the concepts in the domain, and the schema layer of the knowledge graph is usually managed by the ontology repository.

The construction of a knowledge graph starts with the processing of raw data, from which factual elements are condensed and stored in the data and schema layers of the knowledge base. There are two ways to construct a knowledge graph: (1) Top-down construction: the general encyclopedia website is used as a structured data source, and the knowledge elements are extracted from these valid data that have been summarized, inducted and stored into the knowledge base. (2) Bottom-up construction: entities, attributes and relational elements are automatically extracted from massive open-domain data, and then more advanced concepts and entities are gradually constructed.

The construction process of the knowledge graph is shown in Figure 1. This paper focuses on the bottom-up construction technique of a domain knowledge graph, which mainly includes six links: knowledge modeling, knowledge storage, knowledge extraction, knowledge fusion, knowledge computation, and knowledge application.

Figure 1.

The construction process of the knowledge map

Raw data can be divided into structured, semi-structured, and unstructured data depending on the degree of data structuring. According to the different structured forms of data, data are integrated and knowledge extracted, and knowledge fusion, i.e., entity disambiguation and co-reference disambiguation, is carried out after obtaining the initial representation of knowledge. Standardized knowledge is obtained through knowledge reasoning and knowledge discovery. Before entering the knowledge graph, it is generally necessary to update the knowledge representation according to the relevant data specifications and model revisions. In order to ensure the quality and security of the graph, it is also necessary to specifically assess the quality of this standard knowledge, and only those that have been assessed can be imported into the knowledge base.

The LDA model is an unsupervised machine learning algorithm, also known as a three-layer Bayesian probabilistic model. The model uncovers potential topic information from documents in a large-scale corpus and can be used in several domains for topic delineation, text clustering, and hot topic identification.

The latent Dirichlet distribution model is one of the text clustering methods that was proposed by BLEI and other scholars in 2003. The model is an unsupervised machine learning based on the bag-of-words model, according to the vocabulary in the document to find the topic to which the document belongs, is a generative probabilistic model of the corpus, is a three-layer Bayesian probabilistic model contains vocabulary, topic and document three-layer structure. Its main idea is that a document can be represented as a random mixture of implicit topics, and each topic is a probability distribution characterized by its vocabulary.The document picks a specific topic with a certain chance and picks a specific vocabulary from that same topic with a certain chance.By introducing the potential Dirichlet of topics and vocabularies, the overfitting problem in the traditional topic model can be effectively solved. The model can be used in the case of irregular textual expressions and uneven distribution of topic expressions, can handle a larger number of documents and polysemous words in the documents, reduce the influence of noise, does not need to determine the number of topics in advance, and can handle documents and vocabularies in multiple languages. For the lack of text features in the LDA model in the field of short text applications, the use of Word2vec word vectors trained based on a large number of corpus to expand the text data with synonymous features, which in turn can improve the number of text topic features, enhance the co-occurrence frequency of the feature vocabulary for the same topic, and ultimately solve the drawbacks of the LDA model in the field of topic extraction for short text.

The LDA model considers the document-topic distribution as well as the topic-word distribution as obeying a Dirichlet polynomial distribution with parameters. Thus, the probability of generating each word in a document and the joint probability of generating the document can be described as the following two equations: (1) P(w|dm,α,β)=∑z=1KP(w|z,α)P(z|dm,β)\[P(w|{{d}_{m}},\alpha ,\beta )=\sum\limits_{z=1}^{K}{P}(w|z,\alpha )P(z|{{d}_{m}},\beta )\] (2) P(w→|dm,α,β)=∏i=1N∑z=1KP(wi|z,α)P(z|dm,β)\[P(\overrightarrow{w}|{{d}_{m}},\alpha ,\beta )=\prod\limits_{i=1}^{N}{\sum\limits_{z=1}^{K}{P}}({{w}_{i}}|z,\alpha )P(z|{{d}_{m}},\beta )\]

Mathematical modeling of the above process. Assuming that the corpus has M documents and V words, each document contains N words and K topics, θ denotes the polynomial distribution relation between documents and topics, the polynomial distribution relation between topics and words is denoted by φ, and α and β are used to denote the hyper-parameters of θ and φ, respectively, and satisfy the Dirichlet distribution.

In the empirical session, not all cases are the topic distribution is known in advance so as to generate the document, most of the cases are given the document using the model to solve the hidden topic probability distribution. In the document generation process of the LDA model, the number of topics, as well as the two hyperparameters of Dirichlet distribution, need to be given in advance. For a document containing N vocabularies, the set z of topic compositions assigned to each vocabulary by the document, the polynomial distribution of topics θ, and the joint probability density function of the set w of vocabularies under α and β conditions are shown in Equation (3): (3) P(θ,z,w|α,β)=P(θ|α)∏n=1NP(zn|θ)P(wn|zn,β)\[P(\theta ,z,w|\alpha ,\beta )=P(\theta |\alpha )\prod\limits_{n=1}^{N}{P}({{z}_{n}}|\theta )P({{w}_{n}}|{{z}_{n}},\beta )\]

Summing over the θ integrals, all of z to obtain the boundary distribution of this document, yields Equation (4): (4) P(w|α,β)=∫P(θ|α)(∏n=1N∑znP(zn|θ)P(wn|zn,β))dθ\[P(w|\alpha ,\beta )=\int{P}(\theta |\alpha )(\prod\limits_{n=1}^{N}{\sum\limits_{{{z}_{n}}}{P}}({{z}_{n}}|\theta )P({{w}_{n}}|{{z}_{n}},\beta ))d\theta \]

The product of the boundary distributions of each document in the document collection yields the vocabulary generation probability of the full document collection as in Equation (5): (5) P(D|α,β)=∏m=1M∫P(θm|α)(∏n=1Nm∑zmnP(zmn|θm)P(wmn|zmn,β))dθ\[P(D|\alpha ,\beta )=\prod\limits_{m=1}^{M}{\int{P}}({{\theta }_{m}}|\alpha )\left( \prod\limits_{n=1}^{{{N}_{m}}}{\sum\limits_{{{z}_{mn}}}{P}}({{z}_{mn}}|{{\theta }_{m}})P({{w}_{mn}}|{{z}_{mn}},\beta ) \right)d\theta \]

Finally, the vocabulary is assigned to generate all topics according to the probability that the vocabulary is generated on the topic.

Due to the relatively high degree of coupling between parameter β and parameter θ when summing the hidden variables, this can cause the complexity of the a posteriori computation to become exceptionally difficult. Therefore, this paper adopts the Gibbs sampling approximation for the solution.

Gibbs sampling is an algorithm for continuously sampling the conditional distribution of variables whose state distributions converge to the true distribution in the long run. Markov chain Monte Carlo algorithms are commonly used for sampling and joint distribution of multivariate variables. The posterior distribution of the hidden variable z_ij in the LDA model can be decomposed from the probability distribution of the model based on Bayes’ law to obtain Equation (6): (6) P(zij=k|z¬ij,w)∝P(wij|zij=k,z¬ij,w¬ij)P(zij=k|z¬ij)\[P({{z}_{ij}}=k|{{z}_{\neg ij}},w)\propto P({{w}_{ij}}|{{z}_{ij}}=k,{{z}_{\neg ij}},{{w}_{\neg ij}})P({{z}_{ij}}=k|{{z}_{\neg ij}})\]

Where z¬_ij denotes the set of so topics other than the topics of the previous vocabulary, P(w_ij | z_ij = k, z_¬ij, w_¬ij) denotes the likelihood, and P(z_ij = k | z_¬ij) denotes the prior probability. And the likelihood can also be reduced to equation (7): (7) P(wij|zij=k,z¬ij,w¬ij)=∫P(wij|zij=k,φk)P(φk|z¬ij,w¬ij)d φk\[P({{w}_{ij}}|{{z}_{ij}}=k,{{z}_{\neg ij}},{{w}_{\neg ij}})=\int{P}({{w}_{ij}}|{{z}_{ij}}=k,{{\varphi }_{k}})P({{\varphi }_{k}}|{{z}_{\neg ij}},{{w}_{\neg ij}})d\text{ }{{\varphi }_{k}}\]

Again, because of equation (8): (8) P(φk|z¬ij,w¬ij)∝P(w¬ij|φk,z¬ij)P(φk)\[P({{\varphi }_{k}}|{{z}_{\neg ij}},{{w}_{\neg ij}})\propto P({{w}_{\neg ij}}|{{\varphi }_{k}},{{z}_{\neg ij}})P({{\varphi }_{k}})\]

Again, P(φ_k) = Dir(β) and P(w_¬ij | φ_k, z_¬ij) are conjugate to each other, so the posterior formula Dir(β+nk,¬wij(w))$Dir(\beta +n_{k,\neg {{w}_{ij}}}^{(w)})$ for P(w_¬ij | φ_k, z_¬ij) represents the total number of topic z_ij s assigned to vocabulary w_ij s other than the current vocabulary.

Then the above equation becomes equation (9): (9) P(wij|zij=k,z¬ij,w¬ij)=nk,¬wij(wij)+βnk,¬wij(⋅)+Vβ\[P({{w}_{ij}}|{{z}_{ij}}=k,{{z}_{\neg ij}},{{w}_{\neg ij}})=\frac{n_{k,\neg {{w}_{ij}}}^{({{w}_{ij}})}+\beta }{n_{k,\neg {{w}_{ij}}}^{(\cdot )}+V\beta }\]

Where nk,¬wij(⋅)$n_{k,\neg {{w}_{ij}}}^{(\cdot )}$ denotes the total count of words under each subject of k.

Integration of Θ yields Eq. (10): (10) P(zij=k|z¬ij)=∫P(zij=k|θi|z¬ij)dθi=nk,−Wij(di)+αn,−Wij(di)+Kα\[P({{z}_{ij}}=k|{{z}_{\neg ij}})=\int{P}({{z}_{ij}}=k|{{\theta }_{i}}|{{z}_{\neg ij}})d{{\theta }_{i}}=\frac{n_{k,-{{W}_{ij}}}^{({{d}_{i}})}+\alpha }{n_{,-{{W}_{ij}}}^{({{d}_{i}})}+K\alpha }\]

The following equation (11) can be obtained: (11) P(zij|z¬ij,w)∝nk,¬Wij(Wij)+βnk,¬Wij(⋅)+Vβnk,¬Wij(di)+αn,¬Wij(di)+Kα\[P({{z}_{ij}}|{{z}_{\neg ij}},w)\propto \frac{n_{k,\neg {{W}_{ij}}}^{({{W}_{ij}})}+\beta }{n_{k,\neg {{W}_{ij}}}^{(\cdot )}+V\beta }\frac{n_{k,\neg {{W}_{ij}}}^{({{d}_{i}})}+\alpha }{n_{,\neg {{W}_{ij}}}^{({{d}_{i}})}+K\alpha }\]

An intuitive representation of a Monte Carlo algorithm randomly selects a topic from among the K topics and assigns a value to z_ij as the initial state of the Markov chain. The z_ij is updated by calculating the value of P(z_ij | z_−ij, w), and this operation is performed cyclically for each component of the Markov chain, obtaining an update of the state of the entire Markov chain. Repeating the above process yields the target distribution, and the z_ij at this point is then used to estimate the hyperparameters of the LDA model.

According to the national standards, Art is a first-level discipline, and there are 12 second-level disciplines under Art Psychology, Music, Drama, Theater, Dance, Film, Radio and Television Literature, Fine Arts, Arts and Crafts, Calligraphy, Photography, and Other Disciplines of Art, and 4 third-level disciplines under the second-level discipline of Music, including Musicology. This paper uses knowledge mapping as the research method for bibliometric analysis and presents the development and evolution process of music disciplines in Chinese higher education from 2014 to 2023 through visualization and quantitative analysis. Figure 2 displays the breakdown of source journal publication data in the music discipline.

Figure 2.

Change the output at any time

Changes in the number and time distribution of Chinese music education research literature since the new century can be found that the trend between 2005-2023 can be roughly divided into three periods, respectively: 2005-2010, 2011-2018, and 2019-2023. The first period is a period of rapid rise, in which the research on music education in China has gradually received attention from the core journals in the field, showing a steady growth, from 94 articles in 2005 to 221 articles in 2011, and the core journals in the field of music education, People’s Music, Music in China, and Chinese Musicology, have issued a total of 106 articles in this period, accounting for 48% of the total number of articles issued in this period. In the second period, the number of articles gradually stabilized at an annual average of 290 and reached a peak of 367 in 2017, with the publication Grand Stage surging in 2017-2019 and then returning to zero. The third stage is the platform period. The number of articles issued declined, falling to 214 in 2020. The interest in journals in music education is decreasing, and the improvement in 2021 and 2022 is related to the number of articles issued by Fujian Music Leaf in these two years. Combined with the CNKI core journals in the field of music education in 2005-2023, it can be found that the overall number of articles in the field of music education is still steady and rising, and the research in this field still has a large potential for development in the next 2-3 years.

Through further sorting, merging and optimization adjustment, the author’s co-citation visualization map is finally obtained, as shown in Figure 3. By analyzing and examining the co-citation relationship between experts and scholars in the field of musicology research in China, experts and scholars who have had a significant impact on the development of the music discipline in China can be uncovered. Therefore, based on the citation data of source journals in the music discipline, the co-citation mapping of authors in the music discipline from 2014 to 2023 was drawn. Through statistics, there are 91 expert scholars with 10 or more citations, 62 with 20 or more citations, 39 with 50 or more citations, and 7 authors with 100 or more citations who can be regarded as leaders in Chinese higher education musicology. In this section, the expert scholars in the field of musicology with the top 7 citation frequencies are filtered out from them. This section takes two years as a time slice and selects a suitable threshold value. Yang Yinliu, Yu Runyang, Huang Xiangpeng, Wang Yaohua, Qiao Jianzhong, Xiang Yang, and Cai Zhongde are located in the center of the whole map. Their intermediary centrality degrees are all higher, which fully demonstrates that the above mentioned experts and scholars have a pivotal and important position in the field of academic research in the music discipline in China.They have an extremely important role to play in the flow and control of disciplinary knowledge. This fully indicates that the above experts and scholars have a significant position in the academic research field of the Chinese music discipline and that they play an extremely important role in the flow and control of disciplinary knowledge.

Figure 3.

The music field experts are used to map and analyze

Keywords can be said to be the basic language unit for researchers to carry out academic communication, but also a high degree of summary and generalization of the article. In general, it can clearly reflect the main object of research in an academic journal, so this paper, within a specific timeframe, conducted a statistical analysis of the high-frequency keywords appearing in the literature so as to find out the hotspots and research directions in the field. In order to more accurately reflect the development of music education in China in the past 20 years, this paper analyzes the main research content of music education in China by screening the relevant literature with the keyword “music education” from CNKI. Therefore, through the visual analysis based on the keyword “music education”, 257 keywords were found, of which the top 10 high-frequency important keywords and knowledge map were selected for analysis, and their distribution is shown in Figure 4. Keyword centrality represents the distribution of academic journals on music education. The size of the centrality responds to the frequency of keywords cited in the field, where the larger centrality indicates a higher frequency of occurrence. The connecting line between the centrality degrees indicates a common attribute, “music education” (0.15), “art education” (0.50), “quality education” (0.08), “musical art” (0.04) “ethnic music” (0.08) “multiculturalism” (0.11) “aesthetic education” (0.20), etc.

Figure 4.

High frequency keyword list

Mutant words refer to the sudden increase of professional terms in the literature published in certain years, which are words analyzed mainly based on the statistics of the rate of change of the frequency of citation or the number of occurrences of keywords in the literature and are suitable for characterizing the development trend of the research frontiers.

The identification and tracking of research frontiers can provide researchers with the latest research dynamics in the field of music education, which can enable researchers in the field of music education to timely and accurately grasp the frontiers and evolutionary dynamics of research in the field. Compared to the traditional way of analyzing high-frequency topic words, the mutant word analysis method is more suitable for detecting emerging trends in disciplinary development. The rate of change in the number of occurrences of keywords in music education research is shown in Table 1, “The types of theme mutations based on mutation detection algorithms are ascending, descending, ascending and then descending, emergent, and stabilizing theme mutations”, in which ascending, emergent and other mutated words can better reflect the cutting-edge issues in music education. By analyzing the types of the top 10 mutated words, it is found that the mutation types of the mutated words with high mutation degrees in the past five years are all decreasing, which indicates that the attention of the research on the theme of music education of “senior teachers” and “international music” is diminishing. Therefore, more attention should be paid to the construction of “senior teacher” and “international music” in future music education work.

In this paper, the above 10 themes are categorized into three parts according to quadrants, with themes 3, 4, 7, 5, 8, and 6 in quadrant IV as the first part, themes 1, 2, and 10 in quadrant III as the second part, and theme 9 in quadrant II as the third part. In Part I, Theme 6 and Theme 8, Theme 4 and Theme 7 have a high degree of overlap, indicating that in the field of music education, these two themes often co-occur in the same literature. In the third part, there is only theme 9, and this theme is the furthest away from the other nine themes, which indicates that the scope of the study is more distant from the other themes. Further analysis of the list of themes, based on the results of the LDA model to summarize each theme, the detailed analysis of the research themes is shown in Table 2. Theme 1 is “Multiculturalism and Chinese national cultural heritage” in the field of music education, including the dissemination and exchange of the traditional Chinese national culture and the world’s cultures and the impact of the development of multiculturalism on human beings in the fields of art, economy and other areas, Theme 1 is “Multiculturalism and Chinese National Cultural Heritage” in the field of music education, including the spread and exchange of Chinese traditional culture and world culture, and the influence of multicultural development on human beings in the fields of art and economy. Theme 2 is “International Music”, including music education thought, the influence of Western music education trends on Chinese music education philosophy, and the comparison of different stages of music education philosophy. Theme 3 is “Quality Education”, including teacher training in colleges and universities, curriculum reform of music majors in colleges and universities, and training of music teachers.