Exploration of Multicultural Teaching and Students’ Musical Literacy Enhancement in Music Education under the Construction of Knowledge Mapping

The architecture of knowledge graph contains logical architecture and technical architecture for building knowledge graph. 1)

Logical Architecture

The logical architecture of knowledge graph contains two layers: schema layer and data layer. Among them, the schema layer is the core of knowledge graph, which is the conceptual model of knowledge graph, and generally adopts ontology as the schema layer, which is the conceptual template of structured knowledge base, and constrains the data layer of knowledge graph with the help of the rules defined by the template. In the schema layer of the knowledge graph, nodes represent ontology concepts and edges represent relationships between concepts. According to the above hierarchical structure of knowledge graph, there are two main construction methods of knowledge graph, one is bottom-up construction method, and the other is top-down construction method, which is opposite to the first method and refers to defining the ontology and data schema for knowledge graph first, and then populating the entities, relations and attributes into the data layer of knowledge graph. At present, most domain knowledge graphs are constructed using the top-down approach, and since the concepts involved in domain knowledge graphs are relatively fixed, the knowledge accuracy of the constructed knowledge graphs is usually higher.

Figure 1.

Construct the technical framework of knowledge graph

For data collection, a crawler script was utilized to collect music education pedagogical data from music education websites. Data acquisition is the first step in building a knowledge graph. The data sources of knowledge graph can be divided into two categories according to the channel, one category is the open source data on the Internet, which are generally in the form of web pages and are unstructured data. The other category is business data, which is generally stored in the internal data table of the business and structured in a structured form, and is non-public or semi-public data. According to the different data structures, which can be divided into three kinds: structured data, semi-structured data, and unstructured data, different methods can be chosen to deal with different data types.

Knowledge extraction is the process of extracting knowledge from various channels and types of data, and then storing it in a knowledge graph. Based on the knowledge graph of music education, a knowledge extraction model is proposed, which consists of three parts: a text representation method based on fused word vectors, an entity recognition model and a relationship extraction model. 1)

Text representation method based on fused word vectors

The Skip-gram structure is shown in Fig. 2. Firstly, a context window (CW) is set up and it is assumed that other words in the window except the target word are independent of each other, and then the target word is used to predict other words in the window. Take context window CW = 2 as an example, the one-hot encoding of the center word is w_i, and the set of its predicted word encodings is {wi−2,wi−1,wi+1,wi+2} . The model parameters contain the hidden layer matrix H ∈ R^n×d and the output layer matrix O ∈ R^d×n. Where: n is the size of the word list, and d is the dimension of the word vector.

Figure 2.

Skip-gram model structure

The specific training process of word vectors is as follows: firstly, input the one-hot encoding w_i of the center word, look up the table to get the corresponding word vector s of the center word in the hidden layer, and then convert it to the d-dimensional vector o through the output layer matrix, and the computation process is as shown in Eqs. (1) and (2): (1)s=wi⋅H (2)o=s⋅O

Further processed by Softmax function to get the predicted probability of the context word as shown in equation (3), where o_i is the output value of the ind node. The calculation of the loss value loss is carried out using the cross-entropy loss function as shown in equation (4): (3)pk=Softmax(ok)=eok∑i=1neoi (4)loss=−∑i=1nqilgpi

Where: p_i is the ind element of the output probability distribution and q_i is the actual label corresponding to the ith element. The hidden layer matrix H obtained after convergence of the loss function is the required word vector matrix. Its word vector training process is: (5)G=∑i,jnf(Xi,j)(viTvj+bi+bj−ln(Xi,j))2

Where: G is the loss value of the GloVe model; v_i and v_j are the word vectors of the ith and jth words; X is the co-occurrence matrix, whose element X_i,j represents the number of times words v_i and v_j co-occur in a window, and the size of the window is a customized parameter, which is usually taken as 7~10; b_i and b_j are the bias terms; and f(⋅) is the weight function. Namely: (6)f(x)={ (xxmax)t x<xmax 1 x≥xmax

Where: x_max and t are hyperparameters to adjust the loss weights for different word frequencies, in this paper, we set x_max = 100, t = 0.75. The final fused word vector can be obtained by weighting and summing the two word vectors {wF,i}i=1s∈Rs×d . That is: (7)αi=Softmax(cat(wH,i,wG,i)⋅A+b) (8)wF,i=αi,1wH,i+αi,2wG,i

Where: cat(⋅) is the line splicing operation, A ∈ R^2d×2 and b ∈ R² are the learnable parameters.

The high quality text representation matrix obtained by establishing the word vector fusion mechanism will be further used for entity recognition and relationship extraction. 2)

Entity Recognition

In this paper, we adopt a model combining BiLSTM and CRF for entity recognition, and the main body of the model is the BiLSTM layer and the CRF layer. The BiLSTM layer can model the dependency effects between the front and back of the knowledge text and solve the problem of vanishing gradient, which speeds up the convergence speed of the model, and it can effectively solve the problem of long-distance dependency of entities in the knowledge text of the autonomous industrial software [22-23]. The CRF layer is employed to address the issue of serialized annotation. Layer is used to solve the serialization annotation problem, which can derive the conditional probability distribution of another set of output sequences under the condition of one set of input sequences, and add constraints to the final generated predictive labels. For example, if the input sequence is a piece of text, the output sequence will be the lexical properties of the words in the text.

Figure 3.

CRF-BiLSTM model structure

where: v_ntt ∈ R^2u, W ∈ R^2u, b ∈ R^c are the trainable parameters of the self-attention layer; c is the number of relation types to be output, i.e., c = 7; tanh(⋅) is the hyperbolic tangent activation function, the (11)tanh(x)=ex−e−xex+e−x

After the above knowledge extraction process, some logically confusing, redundant or even erroneous knowledge will be obtained, and then different knowledge about the same entity or concept will be integrated through knowledge fusion to form a knowledge base.

Knowledge fusion usually includes two ways, one way is entity linking, i.e., the entity extracted from knowledge and the corresponding entity in the knowledge base will be measured for similarity, and entity integration will be carried out according to the result of the measurement; the other way is knowledge merging, if the source of the knowledge is a third-party knowledge base or structured data, then it can be merged. In the process of knowledge fusion, there are mainly techniques such as denotation disambiguation, entity disambiguation, entity linking, and knowledge merging.

Massive data after knowledge extraction, knowledge fusion to get a series of factual expressions can not be equated with the knowledge of the standard, but also need to go through the process of knowledge processing, in order to add the quality of qualified knowledge to the knowledge base. Knowledge processing mainly includes three aspects: ontology construction, knowledge reasoning, and quality assessment. 1)

Ontology Construction

Ontology construction on the knowledge graph is mainly to summarize the knowledge of the relevant domain, identify the terms that have gained consensus in the domain, and then describe these contents with the definition of format specification. Normatively, the ontology defines the basic terms and their relationships that make up the vocabulary of the domain, as well as the rules for defining the extents of the vocabulary in conjunction with these terms and relationships. In fact, the entities and their attribute relationships extracted in the previous stages are specific representations of the ontology, which summarize them.

Generally speaking, data sources can be divided into structured, semi-structured and unstructured data, usually there are two ways to store these three types of data, the first way is RDF recommended by the W3C standard.RDF is the Resource Description Framework, which is essentially a data model, and it adopts XML to write metadata, using a unified standard to describe resources. The second way uses graph databases, such as the most popular Neo4j, Twitter’s FlockDB, as well as AllegroGrap and GraphDB, etc., which are able to store structured data on the network structure, and the knowledge graph will be more intuitive to store using graph databases. Combined with the multicultural teaching characteristics of music education, this paper uses a graph database form to store the knowledge graph of music teaching resources.

Determination of teaching objectives (enhancement of students’ musical literacy) is the first stage of music teaching based on knowledge mapping. Firstly, according to the knowledge map in the knowledge mapping learning platform, we check the knowledge points included in the teaching content and the relationship between the knowledge points and the knowledge points, and make clear the antecedents of the knowledge points learned in this lesson as well as the important and difficult points; then, we analyze the learning situation of the students, and determine the learning level of the students, precisely locate the nearest development zone of the students, and precisely determine the knowledge objectives and ability objectives to be achieved by students based on the group learning analysis and the individual learning analysis of students. Then, we analyze the students’ learning situation, determine their learning level, pinpoint their nearest development zone, and determine the knowledge and ability goals that they will achieve. The learning situation analysis of Knowledge Mapping Learning Platform is divided into group learning situation analysis and individual learning situation analysis, which includes class knowledge mapping and class overall statistical analysis in group learning situation analysis, and individual knowledge mapping and individual statistical analysis in individual learning situation analysis.

Precise implementation of the teaching process is the second stage of music teaching based on a knowledge map, and it is also the core of the precise teaching model of a knowledge map-based learning platform. Teaching implementation is divided according to the teaching process before, during and after class, and there are corresponding teacher activities and student activities in the three segments before, during and after class respectively. 1)

Before class

Before class, teachers release learning tasks and pre-test questions corresponding to the pre-study knowledge points through the “pre-course guide” in the platform, and the release of learning tasks includes uploading learning task lists and teaching resources involved in the learning tasks, such as microclasses, flash courseware, etc. Students accept the learning task lists and carry out independent learning. Students take on the list of learning tasks and engage in independent learning. After students complete the learning tasks, teachers view the knowledge map of the students’ test questions and analyze the learning statistics to identify the root knowledge points of the students’ problems, as well as determine the teaching activities based on the score rate of the knowledge points. Combining the characteristics of visual knowledge mapping, tracing the root causes of students’ problems, determining the focus and difficulty of teaching, and summarizing the scores of students’ knowledge points involved in the testing process to determine the teaching activities. For the knowledge points where the percentage of students’ scores in the class reaches 75%, the teaching activities of independent learning + individual counseling are adopted, because the knowledge points where the percentage of students’ scores in the class reaches 75% means that most students have mastered the relevant knowledge, and only a small number of students have not mastered it. Independent learning is to formulate the progressive learning tasks related to the knowledge points for those students who have already mastered the relevant knowledge, and the teacher Issues higher-order tasks, in the classroom for them to carry out independent learning, for a small number of students who have not mastered the knowledge point, individual counseling, for them to explain the problems encountered.

Summarization and reflection is the final stage, which is divided into teachers’ and students’ reflections. Teachers’ reflections are focused on the teaching process, while students’ reflections are focused on the learning process. Teachers reflect on the whole teaching process according to the students’ class learning situation analysis report and individual learning situation analysis report, checking the degree of achievement of the teaching objectives, reflecting on the problems in the design and implementation of teaching activities, and then improving and avoiding the next time in the determination of teaching objectives and the design of teaching activities. According to the analysis of individual learning reports and teacher feedback, students analyze the causes of their own problems, trace the root causes, and establish their own error books. At the same time, reflecting on whether their problems are a lack of understanding of knowledge or problems with their own learning methods, and making adjustments to their learning strategies, the reflection of teachers and students will make the next precise teaching more optimized.

In order to ensure the accuracy of the evaluation, the crawled music education teaching data will be partitioned into a training set and a testing set in the ratio of 8:2. Specifically, 80% of the data was used to train the model, while the remaining 20% was used to test and evaluate the performance of the model. This split allows for training the model on a large amount of data while still having a sufficient amount of data for evaluation. The statistical information of the dataset is shown in Table 1, which contains eight data types of music teaching resources (instruments, sheet music, audio, video, electronic textbooks, school-based materials, interdisciplinary materials, online course materials, etc.). It can be clearly seen that among the music teaching data crawled, school-based materials (training set: 451, test set: rounded to 90), musical instruments (training set: 396, test set: 79), and videos (training set: 392, test set: 78) were mainly dominated.

The effectiveness of knowledge extraction was evaluated using precision rate, recall rate and F1 score, whose corresponding formulas are as follows, respectively: (12)Precision=TPTP+FP (13)Recall=TPTP+FN (14)F1=2×precision×Recallprecision+Recall

Where TP stands for true example, TN stands for true negative example, FP stands for false positive example, FN stands for false negative example, and the total number of samples is the sum of the number of TP, TN, FP, and FN. For the entity recognition task, an entity is considered to be correctly recognized only when the target entity correctly matches the entity span and entity type. For the relationship classification task, a relationship is considered to be correctly extracted only when two entities are correctly identified and the relationship type is correct.

The environment used for the experiments was the subject’s server, OS: Ubuntu16.7, CPU: 2*Intel(R) Xeon(R) Gold 6148 CPU @ 4.00GHz, GPU: 4*NVIDIA RTX4090, Hard disk: 32T, Memory: 256G DDR4, Python version: Python 4.6, Cuda version: Cuda11.7, Pytorch version: Pytorch1.17.0.

The model parameter settings are shown in Table 2, which describes the specific detailed parameters of the model implementation, including the hidden layer size, the number of training rounds, the learning rate, and the optimizer of each module, which facilitates the subsequent research work.

In this section of the experiments, the models will be evaluated in two parts, the comparison experiment and the ablation experiment, using the dataset crawled above.

Three joint extraction models and one pipeline model are set up for the comparison experiments, and the reference models are described below: 1)

SpERT model: a joint entity-relationship extraction model based on spanning and pre-trained models.

The results of the experiments comparing the knowledge extraction effect of different models are shown in Fig. 4, where (a)~(b) are entity recognition and relationship extraction, respectively. From the figure, it can be observed that the model in this paper obtains the best performance performance, and the F1 score of the proposed method in the entity recognition task is improved by 4.98% compared with the traditional pipelined method Vulcan, which is ahead of the pipelined method. Meanwhile the F1 score of the relationship extraction task is improved by 9.79%, which indicates that the use of the BILSTM-CRF model reduces the impact of error propagation. Comparing the GraphRel method and the SpERT method, the F1 score of the relationship extraction task is improved by 14.69% and 4.02%, respectively, which is due to the fact that both methods take the approach of predicting the entity pairs first and then performing the relationship extraction, which pays less attention to the connection between the two sub-tasks, whereas the BILSTM-CRF model is able to better capture the bi-directional interactions between the entities and the relationships. Comparing with the PFN model, the performance of relationship extraction has improved by 1.31%, which validates that using the BILSTM-CRF model for feature extraction can better capture the connections between words.

Figure 4.

Comparison of knowledge extraction effect of different models.

After completing the model knowledge extraction effect comparison experiment, the knowledge extraction effect of this paper’s model is further evaluated by ablation experiment next, and the analysis results of the ablation experiment are shown in Figure 5. It can be seen that the evaluation indexes (precision rate, recall rate and F1 score) of the benchmark model LSTM are in the range of 0.6~0.65, and with the addition of the two-way propagation mechanism on the basis of the benchmark model LSTM, the evaluation indexes are improved to 0.80~0.85. Finally, with the introduction of the two-way propagation mechanism and CRF algorithm at the same time on the basis of the benchmark model LSTM, the evaluation indexes reach 0.90 or more, the knowledge extraction effect fully meets the requirements of being able to knowledge map, providing data support for the construction of knowledge map construction work. The extracted knowledge feature data are fused and processed to finalize the design of the knowledge graph.

Figure 5.

The results of ablation experiment were analyzed

In order to test whether the music teaching mode that integrates knowledge mapping can effectively improve students’ music literacy, 18 students majoring in musicology in a university were taken as the research samples and divided into an experimental group (9) and a control group (9), with the control group adopting the traditional music teaching mode and the experimental group adopting the music teaching mode that integrates knowledge mapping. The content of the music literacy test was as follows: 1)

The content of the test of correctness of entering the rhythm: play the music that conforms to the music rhythm of 132 beats/min, let the students listen to the music to find the point of correctly entering the music rhythm, and record the time used to correctly enter the music rhythm.

After going through 15 weeks of instruction, the study sample was chosen to take a post-test of music literacy skills, indicators in the sixteenth week. After the test was completed, the data from the test indicators were organized. Finally, the changes in the indicators before and after the experiment were compared between the two classes.

Music is usually 132 beats per minute, and it is extremely important for music literacy to accurately get the rhythm of the music correct, to perceive the time value of the beat, and to find the entry point of the beat for practicing or completing musical movements. Therefore, in the test of “correctness of entering the rhythm”, a piece of complete music was selected for the test, and the students listened to the music and counted out the correct rhythm of the music, and the shorter the time taken to enter the rhythm, the better the sense of rhythm, and vice versa. The teacher records the final data by timing. The test of “Music Application Ability” includes playing different music clips, students’ ability to recognize and answer questions for selection, and teachers are responsible for recording students’ scores, the test consists of 30 questions and is worth 100 points. “Rhythmic Strength Recognition Ability” is an indicator of the teacher’s ability to play the rhythmic strength of the 2/4, 4/4 beat music clips, usually 10s, the students can completely count the strong and weak beats of the music, and the teacher records the student’s test data through the final count. The results of the comparison between the experimental group and the control group before the intervention are shown in Figure 6, where (a) ~ (c) are the content of the test of entering rhythmic correctness, music use ability, and rhythmic strong and weak beat recognition ability, and EG and CG in the figure represent the experimental group and the control group, respectively. The results of entering rhythmic correctness test, the average time of entering rhythmic correctness in the experimental group was 17.78±3.26s, and the average time of entering rhythmic correctness in the control group was 18.54±2.89s. Independent samples T-test was conducted on the test scores of students in the experimental group and the control group before the experiment and the results showed that T (entering rhythmic correctness) = -1.471, P (entering rhythmic correctness) = 0.27 (p>0.05 ), so there is no significant difference between the two groups of students before the experiment in the index of entering the rhythm correctness, music use ability and rhythmic strong and weak beat recognition ability is the same, can be compared and referred to through the subsequent experiments.

Figure 6.

Comparison of experimental group and control group before intervention

The results of the comparison between the experimental groups before and after the intervention are shown in Figure 7. Before the experiment, the experimental group took about 17.78±3.26s in the index of “correctness in entering rhythm”, and after 16 weeks of teaching, the experimental group took about 14.71±2.07s in the index of “correctness in entering rhythm”, and analyzed the results by using the paired-sample t-test. Sample t-test results T (into the rhythm correctness) = -0.507, p into the rhythm correctness = 0.039 (p < 0.05) analyzed in the “into the rhythm correctness” of this indicator has a significant difference, and the other two also has a significant difference, indicating that in the fusion of knowledge mapping of music teaching, the student’s This shows that students’ music literacy skills (rhythmic correctness, music application skills, and ability to recognize strong and weak rhythmic beats) have improved to a great extent under the integration of knowledge mapping in music teaching.

Figure 7.

The experimental group compared the results before and after intervention

The results of the control group before and after the intervention are shown in Figure 8. After 16 weeks of teaching, the paired samples t-test showed that the control group did not show any significant difference in the indicator of “rhythmic correctness” with T (rhythmic correctness) = 1.028 and P (rhythmic correctness) = 0.074 (p > 0.05). The indicator “ability to use music” yielded a non-significant difference of T (ability to use music) = -0.322 and P (ability to use music) = 0.671 (p > 0.05). There is no significant difference in the indicator “ability to recognize strong and weak rhythmic beats” T (ability to recognize strong and weak rhythmic beats) = -2.018, P (ability to recognize strong and weak rhythmic beats) = 0.229 (p > 0.05). From the above three test indexes, the test indexes before and after the experiment have different degrees of improvement, but the magnitude of the improvement is small, and the p-value derived from the T-test is greater than 0.05, which indicates that there is no significant difference between the control group before and after the experiment in the ability of music literacy. Analyzing the reasons, the students in the control group in the 16 weeks of conventional teaching method teaching, just mechanical imitation of the teacher’s movements, and in the music listening, rhythm learning involves the type of music, rhythm type are relatively single, which leads to the students in the learning process of the rhythm of the rhythm of the grasp and the metrical rhythm of the sensitivity of the degree of low, is not conducive to the enhancement of music literacy.

Figure 8.

Comparative analysis of control group before and after intervention

Comparative analysis of the experimental group and the control group after the intervention is shown in Figure 9, through the 16-week teaching experiment, the experimental group of students in the music literacy ability is significantly better than the control group of students, T (into the rhythm correctness) = -3.268, P (into the rhythm correctness) = 0.014 (p < 0.05), the other two exist in the same situation, the analysis of the reasons for the integration of the knowledge map of the process of music teaching, emphasizing the The cooperation between movement and music rhythm, the experimental group has a special music rhythm training session in the classroom design, so that students can gradually improve the ability to perceive music. Teachers added a variety of rhythmic games to cultivate students’ sense of musical rhythm, and at the same time, the ability to use music was significantly improved. This teaching method not only helps students to deepen their memory of technical movements, but also enables them to grasp the rhythm of music more accurately, thus enhancing their core musical literacy.

Figure 9.

The results were compared after intervention