AIGC-enabled Education Information Technology Integration Application and Research--Taking Information Technology Teaching of Preschool Education Major as an Example

In order to quickly understand what specific knowledge points are contained in the teaching video and help users construct a knowledge framework based on the teaching video, this study proposes a mind map generation method based on the teaching video. To achieve this goal, this study integrates the results of previous research and aims to realize the generation of mind maps based on teaching videos through multiple processing stages. In the first step, the teaching video is pre-prepared to get the text data under different topics; in the second step, the data is pre-processed with the aim of getting the content summaries under the texts of different topics; in the third step, in order to establish the relationship between the keywords, the sub-texts are subjected to the textual syntactic dependency analysis and semantic dependency analysis, so as to get the hierarchical structure of the keywords; finally, the obtained topics, sub-topics, and semantically linked keywords are visualized and generated to get the mind map illustration. The idea of generating mind maps using teaching videos is shown in Figure 1.

Figure 1.

Mind Map Generation Method Based on Teaching Video

For the text data obtained from transcription, the next step is mainly to summarize the transcribed text information and obtain text summaries. Data preprocessing is a necessary operation before analyzing Chinese text. It refers to obtaining more reliable data through Chinese word splitting and other operations, correcting errors and abnormal data, and improving the accuracy of analysis. The preprocessing of text will directly affect the quality of the subsequent text analysis. Therefore, an efficient preprocessing operation in Chinese text is an important condition to ensure that an accurate mind map is obtained. The specific process of data preprocessing is depicted in Figure 2.

Figure 2.

Data preprocessing process

Text summary extraction refers to the technique of converting text content into a short summary that contains key information. At present, extractive and generative methods are the two main types of automatic text summary extraction.

The main idea of extractive text summarization is to rely on the original text, with a number of sentences before and after the original text to indicate the summary, the specific approach is to split the text into sentences, and according to certain rules to calculate the weight of each sentence, the weight of a number of sentences with the highest ranking as a text summary. The generative text summarization algorithm, on the other hand, needs to understand and summarize the text content. With the deep development of deep learning technology, the Seq2Seq model has excellent performance in the generative text summary extraction task. However, its generation results are not stable, and due to the complexity and non-interpretability of the model, it cannot meet the requirements of practical applications at present. Therefore, extractive-based text summarization methods continue to be widely popular.

This study mainly focuses on quickly obtaining the summaries in the subtext data obtained in the previous section through the TF-IDF algorithm and the TextRank algorithm, with the aim of eliminating disposable or poorly related statements and analyzing the main content specifically covered in each subtopic. By comparing the TF-IDF algorithm and the TextRank algorithm, it is found that the former is more dependent on the corpus environment, and therefore has more statistical advantages; while the latter further considers the semantic relationship between words within the document, and has the advantage of reflecting the co-occurrence information between words within the document. Therefore, the TF-IDF algorithm and TextRank have their own advantages and disadvantages, and the two algorithms will be used comprehensively in this study to compare each other and improve the accuracy of text summary extraction. 1)

TF-IDF algorithm

TF-IDF is an algorithm based on statistical analysis to indicate the importance of a word or sentence in a text. A simple idea is to find the word with the highest frequency of occurrence, and the more frequently it is used, the more important it is indicated. Thus, the concept of word frequency is introduced. However, the most frequent words in a text are likely to be “of”, ‘“in”, “is”, etc. This is a common category of meaningless words, so we need to filter out these words in the actual task. These are commonly used meaningless words, so we need to filter them out in the actual task. However, when only words with actual meaning are left, we find that the words “computer”, “hardware”, and “memory” appear as often as they do in a computer text. Obviously they are not of the same importance. This is because “computer” is a very common word in computer texts. Relatively speaking, “hardware” and “memory” are less common. This also means that “hardware” and “memory” are more important than “computer”, so in the keyword ranking, “hardware” and “memory” should be ranked before “computer”.

Therefore, the concept of reverse document frequency was introduced to measure how common a word is. More common words, such as “computer”, are given less weight, while less common words, such as “memory”, are given more weight.

TF-IDF value is the product of TF and IDF value, the larger the TF-IDF value, the more important the word, we simply call it “keyword”. The importance of a sentence can be measured by its “keywords”. If a sentence contains more “keywords”, the more important it is. The N sentences with the highest ranked importance level are selected to form a text summary. In conclusion, TF-IDF is a typical automatic summarization algorithm based on word frequency statistics. The specific formula is shown below. (1)

Word frequency: the frequency of keyword w in the text D_i. Where count(w) indicates the number of occurrences of keyword w, |Di|$$\left| {{D_i}} \right|$$ indicates the total number of words in the text. 1TFw,Di=count(w)|Di|$$T{F_{w,{D_i}}} = \frac{{count\left( w \right)}}{{\left| {{D_i}} \right|}}$$

By invoking the TF-IDF algorithm to extract text summaries for the text under each sub-topic separately, the main content under the topic can be obtained. For example, if we call the TF-IDF algorithm on the text under the topic of “Memory”, we can get the summary of the text as “So the second major component is the memory, which is also the part of the computer used to store the program and data. So there are two main parts of the memory, one is the internal memory and the other is the external memory.” 2)

TextRank Algorithm

TextRank algorithm was proposed by Mihalcea and Tarau in 2004 as an algorithm that can be used as keyword extraction as well as phrase extraction and summary extraction. The algorithm is inspired by the PageRank algorithm, which assumes that the Internet is a directed graph and the nodes are web pages, and determines the importance of a web page based on the probability of linking to it and being linked to it. Similarly, the TextRank algorithm sees the document as a network of constituent units (words or sentences), and the semantic relationships between words (sentences) and words (sentences) are used as links in the network, and the ranking of the words is obtained by constructing a candidate keyword graph, and the highly ranked words are used as text keywords, and the key phrases as well as the text summaries are obtained based on the number of keywords.

TextRank algorithm extracts the text summary process: first, according to the need to integrate the contents of multiple articles into the same text, and split them into sentences, these sentences are represented as vectors; after that, the similarity between two sentences is calculated, and the similarity score is used as the weight of the edges between the two sentences, and the scores of each sentence are obtained by iteratively calculating the weights of the edges, and finally the sentences are ranked in accordance with the Finally, the sentences are ranked according to their scores, and the top N sentence is selected as the summary sentence.

The similarity of two sentences is calculated by comparing the number of identical words between the two sentences. Where S_i and S_j denote two sentences respectively. The sentence similarity formula is shown below. 3Similarity(Si,Sj)=|WK|WK∈Si∩WK∈Sjlog(|Si|)+log(|Sj|)$$Similarity\left( {{S_i},{S_j}} \right) = \frac{{\left| {{W_K}} \right|{W_K} \in {S_i} \cap {W_K} \in {S_j}}}{{\log \left( {\left| {{S_i}} \right|} \right) + \log \left( {\left| {{S_j}} \right|} \right)}}$$

The score of each sentence was calculated iteratively according to the following formula. The scores of the sentences were sorted in descending order, and the first N sentences with a high degree of importance were taken as the summary. 4WS(Vi)=(1−d)+d*(∑Vi∈In(Vi)wji∑Vx∈out(vj)wjkWS(vj))$$W{S_{\left( {{V_i}} \right)}} = \left( {1 - d} \right) + d^*\left( {\sum\limits_{{V_i} \in In\left( {{V_i}} \right)} {\frac{{{w_{ji}}}}{{\sum\limits_{{V_x} \in out\left( {{v_j}} \right)} {{w_{jk}}} }}W{S_{\left( {vj} \right)}}} } \right)$$

By calling the TextRank algorithm to extract the text summary of each sub-topic, you can get the main content of the topic. For example, by calling TextRank algorithm on the text under the topic of “memory”, we can get the text summary as “So the second major component is the memory, which is also used to store programs and data in the computer. Then the memory also contains two major parts, one is the internal memory and the other is the external memory. U Disks are more compact and easy to carry around.”

Combining the TF-IDF algorithm and TextRank algorithm, we can simply take the overlapping part of the extracted summary content as the sub text summary. With the development of natural language understanding, text summary extraction algorithms are also constantly innovating and developing, and in the future about the text summary extraction part of the research task of generating mind maps, we can use more advanced algorithms to further improve the accuracy of the text summary extraction, and to provide a higher quality for the generation of mind maps.

In order to implement the requirements of the “Information Technology” course standard for preschool education majors, improve students’ learning efficiency, and cultivate students’ active inquiry ability in the digital learning environment. To design preschool classroom teaching based on the mind map generated by teaching video proposed above. Following the principles of visualization, inspiration, induction, subjectivity and differentiation in the process of instructional design, using instructional video as a medium, generating mind maps as a scaffold, and mind maps as a tool, instructional design is conceived based on these three perspectives, and the roles of the instructional video, the generating mind maps, and the mind map tool are fully brought into play. In the instructional design process, the preschool education program is designed according to the three stages of pre-course preparation, in-class learning and post-course consolidation.

In order to effectively utilize the knowledge graph of teaching resources to accomplish the recommendation task, this paper utilizes the context in the knowledge graph of teaching resources from two aspects: local context information as well as non-local context information. In the pedagogical resource knowledge graph context processing method, the pedagogical resource knowledge graph is utilized to extract features and embed information for all learner history interaction items. Each history interaction item is regarded as a center node (head entity), and the global context embedding of each head node is obtained after processing its respective context information.

In the local context processing method, the first-order neighborhood entities of all historical interaction items are processed. The feature information of the first-order neighborhood entities of the head node is extracted first, and the relationship between the head node and the first-order neighborhood entities is then embedded. Then a join operation is performed on the two, and a learner-specific attention coefficient is computed by combining the head node information and the learner’s own characteristics. Next, using this attention coefficient, the embedding of each first-order neighborhood entity is linearly combined to obtain the local neighborhood embedding of the head node. Finally, the local context embeddings of all historical interaction items are obtained by aggregating the embeddings of the head node entities and the local neighborhood embeddings.

In the non-local context processing method, the non-local context of the head node is first extracted using a biased random wandering sampling strategy, then the gated loop unit is input in reverse order to output the non-local neighborhood embeddings, and finally the embeddings of the head node entities are aggregated with the non-local neighborhood embeddings to obtain the non-local context embeddings of all the historical interaction items.

The local and non-local context embeddings are fed into a gating mechanism that adaptively aggregates these two embeddings to finally obtain the global context embedding of the head node.

An entity in the knowledge graph is always linked to numerous other entities that can enhance its information in the knowledge graph. Graph neural networks can be used to gather feature information from neighbouring items in the knowledge graph, but aggregating neighboring information directly cannot capture learners’ specific preferences for entities. In order to identify learners’ personalized preferences for local contexts, this paper proposes a graph attention mechanism for each learner to aggregate the neighbor information of entities in the knowledge graph of teaching resources. For each learner, different attention coefficients are assigned to the same neighborhood entity of an item, and then these attention coefficients are used to aggregate neighborhood entities.

In this paper, we use Chl={t|(h,r,t)∈D}$$C_h^l = \{ t|(h,r,t) \in D\}$$ to denote the local neighborhood of entity h and define Chl$$C_h^l$$ as the local graph context of h in the knowledge graph of teaching resources. This study argues that inter-entity relationships in the knowledge graph of teaching resources play an important role in understanding semantic information, and if neighboring entities are connected through different relationships, they may have different impacts. In order to incorporate the relations into the attention mechanism, the embedding of the neighboring entities t∈Chl$$t \in C_h^l$$ and the embedding of the corresponding relations r are first integrated by a linear transformation of Eq. (5): 5ert=(er‖et)W0$${e_{rt}} = \left( {{e_r}\left\| {{e_t}} \right.} \right){W_0}$$

where ∥ denotes the splicing operation and W0∈ℝ2d×d$${W_0} \in {\mathbb{R}^{2d \times d}}$$ is the weight matrix. For the target learner u, the learner-specific attention coefficient describing the importance of entity t to entity h is calculated as shown in Equation (6): 6αu(h,r,t)=exp[πu(h,r,t)]∑(h,r˜,x˜)=Dexp[πu(h,r˜,t˜)]$${\alpha _u}(h,r,t) = \frac{{\exp \left[ {{\pi _u}(h,r,t)} \right]}}{{\sum\limits_{(h,\tilde r,\tilde x) = \mathcal{D}} {\exp } \left[ {{\pi _u}(h,\tilde r,\tilde t)} \right]}}$$

Operation π_u(h, r, t) is performed by a single-layer feed-forward neural network, which is defined as shown in Equation (7): 7πu(h,r,t)=tanh[(eh‖ert)W1+b1]muT$${\pi _u}(h,r,t) = \tanh \left[ {\left( {{e_h}\left\| {{e_{rt}}} \right.} \right){W_1} + {b_1}} \right]m_u^T$$

where W1∈ℝ2d×d$${W_1} \in {\mathbb{R}^{2d \times d}}$$ and b1∈ℝ1×d$${b_1} \in {\mathbb{R}^{1 \times d}}$$ are the weight matrix and deviation vector, respectively. m_u is a nonlinear transformation of e_u which increases the flexibility of the model and it is defined as shown in Eq. (8): 8mu=ReLU(euW1˜+b˜1)$${m_u} = \operatorname{Re} LU\left( {{e_u}\widetilde {{W_1}} + {{\tilde b}_1}} \right)$$

where W1˜∈ℝ1×d$$\widetilde {{W_1}} \in {\mathbb{R}^{d \times d}}$$ and b˜1∈ℝ1×d$${\tilde b_1} \in {\mathbb{R}^{1 \times d}}$$ are the weight matrix and bias vector, respectively. According to Eqs. (5) to (8), the attention mechanism becomes learner-specific due to the use of e_u in the computation of the attention coefficients, which can reflect the learner’s personalized preference for neighbors. Given the normalization coefficients of each neighboring entity of h, linear aggregation of feature embeddings is performed on the entities in the local neighborhood to obtain the local neighborhood embedding of each head node as shown in Eq. (9): 9echl=∑t∈Chlαu(h,r,t)et$${e_{c_h^l}} = \sum\limits_{t \in C_h^l} {{\alpha _u}} (h,r,t){e_t}$$

Then, the embedding of the aggregated entity h and its local neighborhood embedding echl$${e_{c_h^l}}$$ are aggregated to form the local context embedding cht$$c_h^t$$ of h, as shown in Equation (10): 10chl=tanh[(eh‖echl)W2+b2]$$c_h^l = \tanh \left[ {\left( {{e_h}\left\| {{e_{c_h^l}}} \right.} \right){W_2} + {b_2}} \right]$$

where W2∈ℝ2d×d$${W_2} \in {\mathbb{R}^{2d \times d}}$$ and b2∈ℝ1×d$${b_2} \in {\mathbb{R}^{1 \times d}}$$ are the weight matrix and deviation vector, respectively.

The recommendation task is to predict whether learner u is potentially interested in item i, which he has not interacted with before, based on the learner’s historical session information and the knowledge graph of instructional resources. Specifically, the ultimate goal is to learn the scoring function y^i$${\hat y_i}$$, which represents the probability that u may interact with item i, and based on the learned scoring function, the prediction scores of the candidate items are computed and ranked, and the top-ranked items are selected to be recommended to the learner. For the teaching resource recommendation task, this section proposes the DB-CGAT model, which combines the proposed teaching resource knowledge graph contextual processing method with the dual behavioral aggregation method, which is jointly used for multidimensional preference personalized recommendation of teaching resources.The structure of the DB-CGAT model is shown in Fig. 3.

Figure 3.

The DB-CGAT model

DB-CGAT consists of two main modules: a context processing module using resource graphs and a dual behavior aggregation module using behavior graphs. Based on the item representations and session representations learned from these two modules, the probability of a learner interacting with a target item is predicted for personalized multidimensional preference recommendations.

In the context processing module, the teaching resource knowledge graph is first processed with the teaching resource knowledge graph context processing method, where all the historical interaction items in the session are treated as the center nodes respectively, the feature information of their first-order neighborhoods is aggregated to the center nodes, and the embedding of the center nodes themselves is aggregated with the feature embeddings of the local neighborhoods to obtain the local context embeddings of the nodes. Next, the non-local context of the center node is extracted by biased random wandering sampling, and the non-local context embedding of the center node is obtained by using the gated loop unit. Finally, the local and non-local context embeddings are adaptively aggregated to obtain the global context embedding for each center node.

In the dual-behavior aggregation module, the node embedding representations containing structural information are involved in the operation of the dual-behavior aggregation method as input data, and all historical interaction nodes are added to the target behavior sequence and auxiliary behavior sequence, respectively. Through graph attention network and graph neural network learning, the structural global context embeddings of nodes are aggregated with the neighbor embeddings in the interaction sequences, and the embedding representation of each node is updated to obtain the true global representation of the node. The node global representations are substituted back into the behavioral sequences, and the target and auxiliary behavioral sequence representations are aggregated again using a gating mechanism to obtain a session-based representation of the learner’s preferences.

When performing multidimensional preference personalized recommendation, the initial embedding of the learner and the session-based learner representation are aggregated to obtain the learner’s contextual representation, which is again aggregated with the contextual representation of the target item to obtain the learner’s preference score of the target item, and based on the score the item to be recommended for the learner is determined in the end.

To validate the effectiveness of the DB-CGAT model for recommendation tasks, the experiments in this chapter are conducted using the publicly available Yelp2018, Amazon-Book dataset and the self-built CoLR dataset. These datasets differ in terms of domain, data size, and sparsity. The following is a brief description of the datasets: 1)

Yelp2018 dataset for business location recommendations.Yelp2018 is a dataset provided by Yelp, the largest social consumer review site within the U.S. Yelp2018 features addresses, ratings, review content, and user information for registered businesses across the U.S. in 2018, and contains data from more than 6,000,000 user reviews covering all 50 U.S. states.

Each dataset consists of two parts: user-item interaction graph and knowledge graph, in addition to user-item interactions, a corresponding knowledge graph needs to be constructed for each dataset. Referring to a similar setup in KGAT, the knowledge graphs for the three datasets are constructed by mapping items to Freebase entities through title matching, extracting the triples that are directly related to the item-aligned entities and considering nodes in their two-hop range. Various types of information are sampled in the datasets as entities of the knowledge graph, such as category and location of business premises, author and publisher of books, author and type of instructional resources, etc., and connectivity relationships between entities are generated.

1)

Model recommendation performance experiment

The results of the tests on the accuracy of the recommendation tasks are shown in Table 1. It can be found that DB-CGAT outperforms other baselines and brings significant improvement in recommendation results on all three datasets, which validates the effectiveness of the knowledge graph-based DB-CGAT model for recommendation tasks. The datasets used for the evaluation produce different result preferences due to the variability of factors such as data sparsity, knowledge graph features, and different recommendation scenarios. The results in the table show that DB-CGAT can achieve better performance in most cases in the comparison of the different datasets listed and several mainstream baseline methods, which proves the versatility and flexibility of the proposed DB-CGAT model. Benefiting from the combination of the contextual processing method and the dual behavioral aggregation method, DB-CGAT can denoise entity-related relationships and capture more accurate semantic features.

From the results presented in Table 2, it is clear that the proposed DB-CGAT model still has better performance when competing with other knowledge-aware recommendation models. Specifically, DB-CGAT works best in mitigating knowledge graph noise, while still generating the best evaluation results on items with sparse knowledge entities. 3)

Experiments on the effect of the temperature parameter τ.

In the DB-CGAT model for contextual processing, due to the existence of certain negative samples that are very similar to the feature representation of the positive samples leading to the model is difficult to separate these difficult negative samples from the positive samples, so the identification of the difficult negative samples contributes a lot to the gradient descent of the loss function, and it can effectively speed up the convergence of the model results and generate a more robust parameter of the node feature representation τ Mining of difficult negative samples has an important significance for mining difficult negative samples. The performance of different τ-value models under different datasets is shown in Fig. 6.

Figure 6.

Performance of models with different tau values under different data sets

In both CoLR and Yelp2018 datasets, the DB-CGAT model has the best Recall@20 recommendation performance when τ = 0.3. And when the value of τ is too large, it will lead to poor performance and require more rounds of epoch training to converge, while too small a value of τ will harm the model performance due to convergence and other issues. Therefore, the choice of τ should not be too large and too small, and the τ value can be fine-tuned in the range of [0.1,1.0] to determine the best choice of τ value.