The global proliferation of COVID-19 has catalyzed a substantial transition from traditional educational settings to online learning environments. This shift has precipitated exponential growth in online educational content on video-sharing platforms like YouTube. However, this abundance of content often leaves learners navigating from the massive number of videos. Many learners struggle to identify learning videos that align with their learning objectives. To cope with the challenge, we present CONREC, a prototype application for online learners that recommends the next learning videos based on concept maps as a network of knowledge the user has learned. CONREC features an adaptive recommendation that re-ranks the candidates of learning videos based on the combined scores between inter-concepts learned and not learned by the user. We implemented a general-purpose interface that allows learners to continuously watch new learning videos and browser the concept maps of the current state in a visual form.
ucation space, ranging from higher education to personal guidance. This has opened new possibilities for Massive
tional Resources (OER)[
cause of the excellent advantages of online education such as flexibility and convenience.
Recently, a video-sharing platform like YouTube has
the potential to provide an elastic and liberal medium for
knowledge sharing and instruction[
nEvelop-O CEUR htp:/ceur-ws.org ISN1613-073
CEUR
videos, losing their way toward their learning goals.
The massive number of online learning videos can
become a hindrance to learning, with the challenge being
videos[
Our main contribution are as follows. To address these challenges aforementioned, we propose a CONREC system that recommends the next learning videos based on knowledge(a.k.a concept) learned from a number of ifrst to show that CONREC ranks recommended videos based on the user’ understood and not-yet-understood knowledge as a concept map, prioritizing videos that address the latter. Secondly, we present a general-purpose pipeline that extracts semantic-enriched, highly-relevant concepts from a learning video. Thirdly, we propose a recommendation method that computes the similarity between the user’ concept map and concepts in the learning video. Last, we implement an web-based learning application that supports an adaptive navigation of learning videos as well as a visual interface of concept maps.
Our code and demo of CONREC is available in
the sequential information in learning videos. From the learning videos returned by the search, we extract subtitles, and a boundary-based segmentation is performed on the subtitle text to extract a certain number of concepts. When video texts are provided in the learning video, a list of all texts present in the video is first retrieved. From this list, we determine the existence of video texts and
hosts complimentary lectures from globally renowned universities such as MIT, Caltech, and Harvard. FurtherThis section introduces an overview of CONREC. We first more, YouTube provides an API, which is instrumental in describe how to extract semantic and high-relevant con- extracting both learning videos and their corresponding cepts within content. The extracted concept can be added videos text. to the user’ concept map or used to model information of We extract a limited range of lecture videos to extract a learning video. Next, we present our recommendation videos related to the topic. When a user inputs a search method that computes the similarity between concept term and presses the search button, the number of videos vectors as a one-hot representations of concepts. It also returned from the YouTube API is retrieved. We selechelps to re-rank rankings of candidate videos by contin- tively filter out videos that are not relevant for learning uously applying to them. purposes, specifically those that are either too short (less than 10 minutes) or excessively long (more than 120 min2.1. Concept Extraction utes). Overly brief videos may not encompass suficient knowledge content, while exceedingly long videos could potentially impair the concentration of learners.
tent as concepts that subjects taught in the learning video. To extract concepts from the video, we introduce data sources used, video segmentation, and concept acquisition.
1–6 a concept vector in which concepts in the are understood explicitly by the user.
extract the corresponding texts. We have specifically set up fixed interval of 5 minutes, which typically corresponds to approximately 300-400 words. In addition, we have set the margin to include words within a range of 5 seconds in the existing segmentation, preventing them from being included in the following segment. We have made it possible to adjust the range of the margin based on input parameters. The adjustment of the margin can be seen in Figure 1(A).
used Wikification[
Maps
cepts related to the concept map-based recommendation • ⃖⃖⃖′⃖⃗ is the concept vector in which concepts in the method. learning video understood explicitly by the
A collection of learning videos is a finite set of user. videos related to a specific subject, such as the recom- • ⃖⃖⃖⃗′ is the concept vector in which concepts are mendation system, denoted as = { 1, 2, ..., }, where learned from a multiple number of learning represents a learning video. Each video is consid- videos the user has watched. Note that ⃖⃖⃖⃗′ is all ered a document containing its video texts, which are of the concept maps learned from the learning provided in the form of video subtitles or speech scripts. videos that the user watched related to a particu
Concept map represents the set of concepts that can lar subject. be learned or have been learned from the watched videos.
Let us denote the concept map as = ∪ ′, where Given , , and , our recommendation method com and ′ indicate sets of learnable or learned concepts, putes the similarity between concept vectors as the folrespectively. Both sets contain one or more concepts, lowing equation. referred to as = { 1, 2, ..., }, where each concept is mentioned in the video texts.
For eficient comparison between concept maps, we ( , , ) = ⋅ ( ⃖⃖⃖⃖⃗, ⃖⃖⃖⃖⃗) + (1 − ) ⋅ ( ⃖⃖⃖⃗′ − ⃖⃖⃖′⃖⃗, ⃖⃖⃖⃖⃗) vectorize the concept map using a one-hot representation (1) of concepts, which is referred to as the concept vector , where function denotes either Jaccard or Cosine and denoted as ⃖⃗. The size of ⃖⃗ is determined by the similarity computations and denotes a weight value nuFmobrecrlaorfitcyo,nwceepvtisewin le.arnable concepts as concepts betHweereen, 0( an⃖d⃖⃖⃖⃗1,.⃖⃖⃖⃖⃗) determines how many learnable that are not clearly understood by a user from he most- concepts are included in the new video. ( ⃖⃖⃖⃗′− ⃖⃖⃖′⃖⃗, ⃖⃖⃖⃖⃗) recent learning video the user has watched. In other reduces the importance of previously-learned concepts words, learnable concepts are concepts that the user while maintaining concepts learned from the most-recent wants to learn from another lecture video. We define learning video. learnable concepts from the most-recent learning video The weight values can determine the degree of importhey watched as follows. tance of the learning video containing learning concepts. In other words, we can restrict videos containing learnable concepts from being ranked high when recommending the next videos. It is worth noting that we allow the user to adjust according to the domain. • Given the most-recent learning video , ⃖⃖⃖⃖⃗ is a concept vector in which concepts in the are not yet understood by the user. Otherwise, ⃖⃖⃖′⃖⃗ is
1–6
3.1. Application Overview CONREC supports an adaptive navigation of video content for learners to explore videos based on the concept map of a user. It searches YouTube videos based on the concept map given by users and return re-rank videos based on the score calculated by our recommendation method. Figure 1 shows four panels of the main interface: (A) User Configuration, (B) Search based on concept map, (C) Feature tabs, and (D) Contents. 3.2. Main Features
uration, search module, and feature tabs. 3.2.1. User configuration The feature tabs consists of four sub-tabs: new learning, In the sidebar panel, these parameters are used to spec- history, concept map, and watching tabs. ify configurations according to the characteristics of the In the new learning tab, a user can see candidates learning domain. There are four options that users can of learning videos through the recommendation system determine (1) the number of videos to search, (2) how when searched in the top search bar. In the Figure 1 (D), many seconds to divide the video to extract key concepts, each video is composed of four components: the watch (3) how many concepts to extract from the sections di- button, the video, the recommendation score, the video vided by seconds in the video, and (4) how much weight title, and the video description. First, based on the results to give when measuring similarity. Especially, he weight of a keyword search in the search box, learning videos value indicates the learnable concepts in the video. It appear for the learners. When the ’watch button’ of a means value closer to 1 suggests that the video contains displayed video is clicked, the selected video is classified learnable concepts, while a value closer to 0 indicates that as a ’watched video’, and later, the concepts learned from the video primarily consists of learned concepts. With that video are used for recommendations. Furthermore, this parameter, users can tailor and receive video lecture the clicked video can be viewed in the ’watching video recommendations based on their preferences. tab’. The recommended score represents a value calculated by the recommendation algorithm. The higher the recommended score a video has, the higher it is placed. 3.2.2. Search module based on concept map In the history tab, it displays a list of videos that the When the user enters a keyword into the search bar and user has watched. The list includes the video, video title, clicks the search button, the system requests the search a re-watch button, and segment information about the keyword to the YouTube API. The videos are retrieved concepts contained in the video. The Re-Watch button according to the number of videos and stored as a video allows you to re-watch the video, enabling learners to object. Each video object contains the title, URL, video re-learn from sections they might not have fully grasped description, and video length. Next, it extracts video texts initially. Additionally, learners can check segment inforfrom the selected videos when they provides subtitles. mation for the videos they’ve watched, which displays Otherwise, we obtain it through automatic speech recog- information about learnable and learned concepts for nition (ASR) provided by the YouTube platform. We then each segment of the video. This segment information is divide the video texts into a fixed number of segments. provided in table, with each concept marked as either After the division, it apply the Wikification method to 0 or 1, depending on whether it was understood by the all segments in order to extract concepts depending on learner. the number of concepts. Here, smaller than the number In the concept map tab, it visualizes the concepts from of concepts are extracted when there are not enough the watched videos as a network graph. In Figure 2, number of keywords. The extracted concepts includes a Each sky blue node represents the videos that have been name, URL, and PageRank score. watched, each black node represents the concepts that were not understood from the video, and red nodes signify the concepts that were understood. The size of these concept nodes increases based on the PageRank value obtained through Wikifier, with higher values resulting in larger nodes. If diferent videos share the same concept and link to the same Wikipedia concept, they are interconnected.
In the watching tab, it allows users to view the video they selected from either the new learning tab or the history tab. In the watching tab, there is the selected video, segments related to the video, and the concepts contained within each segment. The number of segments and concepts displayed reflect the settings made in the options of Figure 1(A). Each segment contains concepts up to the set number, and they are displayed in order of importance. Initially, all concepts are in a ’learnable’ state and are shown in black. As users watch the video and understand a concept, they can click on it, turning it red, which signifies a ’learned concept’. The concepts learned in the Watching tab are reflected in the history tab and the concept map tab. These learned concepts are then used as a basis for recommending new learning videos
line videos based on concept maps. The following user scenario illustrates a continuous learning process.
In the video-sharing platform, a user searches for learning videos relating to keywords of the concept she wants to learn (e.g., recommender system). The video-sharing platform retrieves multiple videos. She can select the ifrst video from them to learn about the concept and watch the selected video. As shown in Figure 3, she can interactively mark either learnable or learned concepts in the middle or at the end of the video she is watching. Her concept map is incrementally created whenever she marks either learned or learnable concepts.
Regardless of whether she watches the video completely or not, she can select the next videos to understand learnable or learned concepts. At that time, the rank of the lecture videos was rearranged according to the current state of her concept map. Here, the user settings aforementioned afect the rank of the videos.
The visualized concept map allows her to check what she has learned. In addition, the co-occurrence of concepts in the watched videos is displayed. Moreover, she can watch the learning video again in her watched history if she either did not understand some concepts or re-learn concepts. She searches for keywords of either learnable or learned concepts to obtain additional candidates for lecture videos. 1–6
To summarize, we have proposed CONREC, a recommendation system tailored for exploring learning videos based on concept maps. First, CONREC has identified suitable learning videos based on knowledge the user has learned but also re-ranked them considering his understanding. Second, CONREC has navigated video content adaptively, dividing videos into segments and visualizing concepts as network graphs. Third, we have defined both learnable and learned concepts and represented the concept vector as a one-hot representation of concepts. Last, our graphical interface has ensured users can customize their learning experiences, enhancing content relevance, and engagement.
In the future work, we will improve CONREC that aligns video recommendations with potential career paths. Our future work provides learners a personalized learning path that not only enriches their knowledge but also enhances their career prospects.