Short, Long, and Segmented Learning Videos: From YouTube Practice to Enhanced Video Players

1.1 Short vs. Long Videos

There are numerous claims in the literature regarding the optimal length of learning videos, but there is a lack of valid reasons and empirical evidence. Short videos are generally preferred by users, but the definition of “short” is ambiguous. YouTube classifies videos as “short” if they are under 4 min and “long” if they are over 20 min. Students also prefer short videos for learning. The short length of the videos is intended to prevent the viewer from becoming too passive and to prevent thoughts from wandering (Brame, 2016; Risko et al., 2013). Guo et al. (2014) analysed the engagement time in multiple MOOCs by comparing videos of different types (e.g. lecture, tutorial, other), production style (e.g. slide presentations, classroom, Khan-style), and video length. At least for MOOC-like settings, Guo et al. (2014) recommended that individual videos should be kept shorter than 6 min, except for tutorials where the duration does not matter. However, these results can not be generalised for all types of video, production styles, and learning scenarios. For an in-depth analysis, a comprehensive classification of learning videos such as those presented by Seidel (2018) should be taken into account.

The process of shortening videos involves various design decisions. During production, content can be streamlined by removing side issues, pre-knowledge repetition (cf. pre-training principle (Mayer, 2009)), and redundant information (cf. redundancy principle (Mayer & Clark, 2012)). Speaking rates can be increased, though this demands more attention from viewers (Guo et al., 2014). Content can be outsourced to other videos. In post-production, pauses can be removed (Guo et al., 2014), playback speed increased, and visual content reduced to relevant still images. Such compression may necessitate multiple viewings, making the video less self-contained. Furthermore, densely packed information can overwhelm viewers, leaving little time for cognitive breaks.

1.2 Video Segmentation

As an alternative to shortening videos, the segmenting principle suggests dividing longer passages into smaller, separate sections. After watching each individual segment, the learner typically needs to actively proceed by clicking on the next video (Brame, 2016). This segmentation can be achieved either during post-production or through better planning of video lengths during pre-production. Segmentation offers two advantages that positively impact learning: reducing cognitive load and improving the structuring of learning material (Spanjers et al., 2010).

Mayer and Chandler investigated the effect of segmentation on learning by showing learners an animation of lightning formation. One group watched a continuous version of the animation, while another group viewed it in 16 individual segments (Mayer & Chandler, 2001). In a subsequent transfer test, the learners who interacted with the segmented version, clicking to advance to each new section, performed significantly better. Additionally, another experiment examined the optimal sequence for presenting holistic and segmented videos. The “whole-part” approach involves showing the entire video first, followed by the segmented sections, whereas the “part-whole” approach presents the segmented sections first, then the whole video. Consistent with cognitive load theory, learners in the “part-whole” group achieved better scores (Mayer & Chandler, 2001).

These advantages of segmentation demonstrated by Mayer and Chandler were confirmed a few years later by Moreno (2007). In Moreno’s study, prospective teachers were shown videos about teaching principles or techniques. A control group watched a continuous 20-minute video on teaching techniques, while a test group viewed the same content divided into seven segmented videos, each presenting one technique. The learners who watched the segmented videos scored better on transfer tests, which assessed their ability to implement the learned material.

Other studies have shown similar results. In a study by Mayer et al. (2003), learners were shown an animation explaining how an electric motor works. The segmented version included interactive elements where learners could click on a question and view the corresponding segment. The non-segmented version was a continuous video with identical content. Once again, the group with the segmented version performed better.

Boucheix and Guignard (2005) also demonstrated the effectiveness of segmentation in slideshows. Learners who received the segmented version achieved measurably better results. They could control the learning pace by clicking the Next button while two other groups watched either a fast or a slow continuous version of the same slideshow.

In the technical literature, the pacing principle is often used synonymously with the segmenting principle. There is no clear distinction between learner-controlled segmentation and program-controlled segmentation with pauses. Moreno and Mayer also use the term “pacing” as an alternative to segmentation (Moreno & Mayer, 2007). In Doolittle et al. (2015), segmentation is directly defined as control over pacing.

Segmentation without forced pauses by the program but with chapter subdivision and short transitions should enhance learning facilitation through better material structuring. This type of segmentation still requires the learning material to be logically divided into connected sections. The second benefit, reducing cognitive load, could be partially achieved if transitions allow enough time for learners to organise their thoughts before moving on to the next chapter. Unlike segmentation with forced pauses until learner interaction (e.g., via a Next button), pauses fixed by the video or learning program cannot adapt to the learner’s knowledge level or processing speed.

Spanjers et al. (2012) investigated segmentation without forced pauses by the program. Their study included four conditions: non-segmented, segmentation via brief screen blackening, segmentation with pauses, and segmentation without pauses.

A meta-analysis by Mayer and Pilegard (2014) demonstrated a segmentation effect in 10 out of 10 studies examined. The median effect size was 0.79, indicating a strong positive effect of segmentation on learning.

1.3 Video Analytics

In learning analytics research, videos are mostly studied in the context of certain learning modes (e.g. situated learning, collaborative learning) in different settings (e.g. MOOCs, higher education, K12). In this regard, video-based learning in online learning environments enables the recording of clickstream data and also a precise recording of the video segments viewed. The methods for using this data are well researched (Mongy et al., 2006; Seidel, 2014), but not yet widely used in research and practice (Poquet et al., 2018).

Apart from that, research on video analytics has mainly focused on learners’ behaviour but not on the context in which video-based learning processes occur. When investigating learning behaviour, context variables like the learning video and the video learning environment need to be considered.

The actual learning videos as an underlying learning resource have been considered as an independent variable in only a few studies (Guo et al., 2014; Maya et al., 2022). In contrast, it is assumed that learning processes and learning outcomes can be studied independently of the particular video. However, basal properties such as video length significantly impact the view duration (Park et al., 2021) and thus on learning time and, ultimately, learning success.

A comprehensive empirical study on the length of educational videos does not yet exist. In 2013, according to Che et al. (2015), 96 % of videos on YouTube were shorter than 600 s. This is mainly due to the limit of 10 min imposed by YouTube in March 2006, which was increased to 15 min in July 2010. The length of videos depends very much on the category. Park et al. (2021) examined 1,125 randomly selected YouTube videos. According to this study, the video length significantly impacts the view duration. Shorter videos were watched much longer, depending on the proportion of the video viewed. However, the studies by Che et al. (2015) and Park et al. (2021) refer to YouTube videos in general and not to learning videos.

From the perspective of video analytics research, the length of learning videos has a crucial influence on the choice of research methods for video-based learning processes. With long videos, learners interact more within a video (e.g. play, pause, seek), whereas when multiple short videos are offered, interactions between videos (e.g. search, browse) dominate. Schwan and Riempp (2004) therefore, distinguish between intra-video and inter-video interactions. While data of intra-video interactions can be analysed with time series analysis methods, the selection and sequence of the accessed videos and, thus, methods like frequent pattern mining and sequential pattern mining play a more important role in inter-video interactions.

With the ambition that learning analytics should bring insight into the design of learning opportunities, the study of videos as learning resources can potentially provide important insights to create compelling video-based learning opportunities. Guo et al. (2014) and Li et al. (2015), for instance, have developed pattern-like design recommendations based on studies of user behaviour in video-based MOOCs. Empirically well-founded principles for multimedia learning were proposed by Mayer (2009) and Clark et al. (2006). Although these recommendations, principles, and guidelines have practical use, they do not mention the design of adequate software applications. Furthermore, these principles lack concrete advice about implementing video content and video players. The design of video players for explicitly using longer learning videos has not yet been investigated.

1.4 Objective and Research Questions

This article aims to analyse common design practices regarding the length of educational videos and to explore design options for handling longer videos.

In pursuit of this aim, we investigate the typical lengths of learning videos available on YouTube, the world’s largest video platform. By doing so, we aim to fill a long-standing research gap and empirically determine whether the lengths of these learning videos in practice align with the recommendations from the literature.

In response to the generalised criticism of longer learning videos, we aim to explore design options to segment these videos using interactive video players. Regardless of a video’s length, the video player’s capabilities should support the learner in video-based learning. This raises the question of how longer learning videos can be segmented and thus made more conducive to learning by using interactive video players.

By combining these two directions, this research maps the current landscape of learning video lengths and provides actionable insights into improving the design and usability of longer educational videos through segmentation and interactive features. Concerning the general definition of learning analytics (Siemens, 2013), this research contributes to the field because it aims to collect and analyse data about the context in which learning takes place to understand and optimise learning environments.

To address the research questions, we present two studies. The first study, detailed in Sect. 2, examines the video lengths and segmentations of 895 German and 41,081 English channels from YouTube Education (a total of 85,303 and 3,923,938 videos, respectively). The second study, outlined in Sect. 3, explores design options for structuring longer videos into segments. We developed a video player capable of presenting a single (longer) video file under three different conditions: (i) as a non-segmented video, (ii) as a video with an additional chapter overview, visible chapter boundaries, and navigation options for the segments, and (iii) as a sequence of segmented videos of suitable length.

These three conditions were examined in an online user study. Two videos with varying amounts of information were used to compare learning effects and video interactions. Twenty-two participants were divided into groups for each condition.

2.1 Methods

2.1.1 Research Design

The first study investigates the current practice of producing and applying educational videos. For this purpose, a large number of relevant educational videos on YouTube are examined. As the largest video platform in the world, YouTube is particularly well suited to determine video trends and analyse viewer behaviour (Snelson, 2011; Che et al., 2015). Therefore, the trend towards greater acceptance and demand for shorter learning videos should also be reflected on YouTube. This analysis aims to examine educational YouTube channels in terms of the length of their videos. For this purpose, a full set of educational YouTube channels in German and a sample of English-speaking channels were examined for video length. We have chosen English as it is the most spoken language in YouTube videos. German was added as a contrasting language to observe language-specific or cultural differences.

A mixed research method was used, in which data were collected using information retrieval methods and then examined using data mining methods (Suganya & Vijayarani, 2022). This design is characterised by the collection and analysis of quantitative data about video-based learning resources in the first phase of research, followed by the analysis of video metadata in a second phase that builds on the results of the initial findings.

2.1.2 Dataset

YouTube does not provide a public list of channels or videos assigned to educational purposes. However, users must select one of eight content categories when creating a new channel. One of these content categories is “education”. A web service called channelcrawler.com constantly crawls YouTube channels, including information about the content category, language, country of origin, number of subscribers, number of videos, number of video views, and the creation date of the channel. We scrapped a complete list of all German-language channels within the education category from this web service. We scraped another list of all channels assigned to the education category that stated they were produced in English. Table 1 provides an overview of the resulting lists with the data from channelcrawler.com.

Each of these channels mentioned in Table 1 includes videos that are accessible from the channels’ landing page. We wrote a Python script to extract metadata of the videos filed under a YouTube channel. Among the available metadata, we could obtain information about the video length, title, the number of views, the average user rating, and the upload date. While we could extract metadata from all videos of the German-language channels, only a random sample of about 1 % of the English-language channels were analysed.

The datasets created, and the software used to scrape Channelcrawler.com and extract metadata from a list of YouTube channels are publicly available (Seidel, 2024).

2.1.3 Data Analysis

It was ensured that there were no duplicates in the data sets. Outliers regarding a video length of fewer than 10 s and a length of more than the average plus three times the standard deviation were removed. For the German-language videos, 1,749 and the English-language videos, 3,013 were classified as outliers.

We applied clustering to a single dimension to gain further insight into common video lengths used for educational purposes. The standard iterative k-means algorithm (Lloyd, 1982) was used for clustering. The algorithm calculates the within-cluster sum of squared distances iteratively, changes each point’s group membership to minimise the within-cluster sum of squared distances, and computes new cluster centres until local convergence is achieved. The number of clusters was determined by the distortion score and silhouette score (Rousseeuw, 1987). The distortion score is the mean sum of data points’ squared distances to the cluster centre.(Figs. 1 and 2)

The Regular Expressions in Listing 1 were used to identify videos from the sample whose titles contained indications of segmentation. Specifically, the keywords part, episode, volume, period, # – including plural forms and common abbreviations for German and English language – were searched for in conjunction with digits. The video titles were examined for strings that indicated segmentation to improve the regular expression. The video titles detected in this way with indications of segmentation were all manually checked for false positives.

Used Regular Expression to identify segmented videos

Full size image

2.2 Results

As shown in Table 2, metadata of 83,558 German and 154,370 English videos has been extracted. These videos have been uploaded between 2005 and 2022. In the considered sample, the yearly number of uploads increased to 25,000 and 37,000 in recent years. The German videos, with a mean length of 14.63 min, are only slightly shorter than the English videos, with a mean length of 15.07 min. The video length’s standard deviation (SD) for both languages exceeds the mean (M). The video length ranges from 10 to 4,973 s for German videos and 7,942 s for English videos. The overall shape of the distribution of the video length is skewed to the left (see Figs. 3 and 4). The histogram indicates a uni-modal distribution.

The degree of correlation between the video length and the view count is very low (Pearson, EN 0.074/DE \(-\)0.010). Figures 5 and 6 show the distribution of video length in relation to the view count. The median view count is 113 for German and 1389 for English videos, while only a few outliers of the English videos (about 1 %) achieved the attention of more than one million views.

Upload time for German videos

Full size image

Upload time for English videos

Full size image

Distribution of video length for German videos

Full size image

Distribution of video length for English videos

Full size image

Views per video length for German videos

Full size image

Views per video length for English videos

Full size image

We used K-Means as a clustering algorithm to identify and describe possible groups. The number of clusters was computed using the average silhouette of observations for k clusters with \(k=[1,\dots ,15]\). The maximum of the average silhouette over the range values for the k cluster was 3 (silhouette score.79 for EN and.66 for DE). Table 3 provides an overview of the clusters found in the data set. The smallest cluster contains 99 longer videos with a duration ranging from 19:25 to 56:55 min (M=27.70, SD=7.84). The cluster of medium-length videos contains 1422 videos with a total length between 7:32 and 19:15 min (M=10.85, SD=2.32). The biggest cluster is the one with short videos. The duration of 3102 videos is between 9 s and 7:31 min (M=4.20, SD=1.64). (Figs. 7 and 8)

Overall distribution of the video length clustered by video length - German videos

Full size image

Overall distribution of the video length clustered by video length - English videos

Full size image

The remarkably large total number of videos in clusters DE1 and EN1 (cf. Table 3) raises the question of whether these videos stand alone or are part of a series. A series can be understood as a continuation of a story or as a segmentation of a longer video. Using Regular Expression, segmented videos could be identified by their title. Figure 9 and 10 emphasize the share of segmented videos among the total number of videos concerning their length.

All clusters contain a considerable amount of segmented videos (12 % (DE) and 9 % (EN)). As shown in Table 3, there are 8.9 % segmented videos in cluster DE1 and 8.1 % in cluster EN1. For videos with a medium length, like in the clusters DE2 and EN2, the share increases to 21.2 % and 14.5 %, while the clusters DE3 and EN3 with longer videos also show a high share of 18.6 % and 13.9 %. Further analysis of the publication date of segmented videos within a YouTube channel did not reveal any insights.

Overall distribution of video length of segmented and unsegmented videos - German videos

Full size image

Overall distribution of video length of segmented and unsegmented videos - English videos

Full size image

2.3 Discussion

Channelcrawler selected the YouTube channels. However, the approaches they applied to collecting data from YouTube are unknown. Potential algorithmic biases from YouTube search and the crawler could not be excluded. The reliability of the identified YouTube channel was examined by random sampling. Based on the start pages of the channels as well as the associated videos, a relation to education and education videos could be determined in all samples.

While the German-speaking educational YouTube channels could be examined fully, the English channels could not. The amount of more than 42,000 channels could not be processed in a considerable time. The limitation factor is the response time for every metadata request. The sample of about 1 % of the overall number of English channels/videos may appear small, but the number of about 150,000 videos is comparable to the amount of examined German videos.

The metadata extraction for the individual videos was very reliable, as the information could be obtained directly from YouTube. The technical information about the video length can also be considered very reliable. On the other hand, the calculation of the view counts follows a metric defined by YouTube. Accordingly, a video is considered to have been viewed as soon as only a few seconds of the video have been played. It is impossible to conclude about the time spent viewing or the repeated viewing of individual sections. Referring to Hardman et al. (1999), the media element time needs to be distinguished from the time spent interacting with the video player and the video content.

For the individual videos, it could not be determined beyond doubt whether they were all learning videos in the strict sense. The content of the videos (e.g. topic, field), the production style (e.g. office video, presentation, khan-style), the type of learning video (e.g. instruction, explanation, observation), and the target audience (e.g. K12, university students) was not examined in detail. Possible differences in video duration concerning these dimensions were not investigated.

Each video was viewed individually. The context of videos that belong to a series or sequence of videos was determined by common keywords appearing in the video title. Semantic relations that were not expressed in the video title but in the video itself could not be considered.

A technical feature of the video was used to determine the video length. The duration of the didactic intervention conveyed in the video could not be recorded. One would have to subtract the time used for the opening credits, closing credits, or other interruptions from the video length.

The investigation of video length based on the length of the video in educational YouTube channels has shown that there are three groups of videos of different lengths. The largest group represents two-thirds of the videos, which are shorter than about 14 min (DE) and about 23 min (EN). This result shows that the production of short videos has proven itself within the examined sample of YouTube channels.

It is noticeable that other researchers consider videos of 4 min (YouTube) or 6 min (Guo et al., 2014; Geri et al., 2017) to be short, while here, videos almost six times as long are still considered short. Guo et al. (2014), however, related his recommendation to MOOCs, i.e., a particular formal learning setting. YouTube videos are mostly intended for informal learning settings and do not have to follow strict sequences of short videos and quizzes as is in typical video-based MOOCs.

One possible explanation for the different ranges of videos considered short between German and English educational YouTube channels is cultural differences between German and English-speaking audiences concerning preferences about video length.

Viewers’ preferred language is often linked to their geographical origin and cultural background (Taneja & Webster, 2016), which influences their media consumption behaviour. This behaviour is shaped by the relevance of the content to their cultural context, where relevance also stems from geographical reference points. Similar cultural proximity hypotheses have been confirmed for YouTube music videos (Baek, 2015) and cultural differences in play styles of video games (Bialas et al., 2014).

Producing videos with up to about 14 or 22 min of playing time can be considered common practice. However, videos with a median length between 14 and 39 min (DE), respectively 22 and 64 min (EN), also make up a significant proportion, while longer videos are the exception.

The channels considered here show a clear picture concerning the segmentation of educational videos. About 8 % of the videos with short length are part of a series of segmented videos. Very short videos with less than a minute duration were not segmented. A lower limit for the length of a segment is shown here. It is also clear that videos of medium length are part of a series in 14 % to 22 % of the cases. This result shows that more comprehensive topics can be conveyed in the form of a video if they are divided into smaller parts. In the YouTube channels considered here, it is common practice to divide videos into medium-and short-length segments.

However, the implications set out above apply only to YouTube’s possible uses. The videos of the channels examined here were produced exclusively for YouTube and the tools available there. Other video players may support other tools and enable other usage patterns. The following study examines how video segmentation can be supported by the design of video players.

3.1 Methods

3.1.1 Research Design

The analysis of educational YouTube channels has revealed a certain practice of segmenting videos for learning. However, YouTube only has a single design option for segmenting longer videos. Alternative design variants for segmenting videos have not been considered from a user’s perspective. The following user study aims to compare design variants for video players, each of which supports the segmentation of longer videos in different ways. Comparing different player variants should lead to findings for designing video players that are conducive to learning.

The study is an experimental study. Participants were randomly exposed to different treatments. The treatments consisted of two video types and three segmentation variants – two treatments and one control condition. The test persons received a short e-mail explaining the scope of the data collected and the content of the study. The attached link took them to the web application where they could participate in the study. All participating persons went through the same process steps. First, some personal data was collected in step 1. In step 2, previous knowledge of the topic dealt with in the video was asked for. In step 3, the participating persons were shown approximately 45 min of video material. In step 4 an evaluation of the shown learning material was asked for, and in step 5 the same knowledge query was made as in step 2. After submitting the knowledge query again, participation was finished. It took about 5 to 10 min each to answer the two questionnaires at the beginning and the end of the test participation. Together with a running time of the videos of about 45 min each, this results in a total time of about one hour for the participants. Finally, the learning success is calculated and evaluated based on the answers to the two knowledge questions.

The following hypotheses have been tested:

1. H1

   Segmented videos with sequences of up to 10 min in length (variant 2) achieve a higher learning effect than long, non-segmented videos (variant 1).

2. H2

   Segmented videos (variants 2 and 3) are perceived by viewers as better structured than non-segmented videos (variant 1).

3. H3

   Long videos are just as conducive to learning as segmented individual videos (variant 3), provided that the video learning experience is supported by a table of contents or chapter index, navigation options, short chapter transitions, and visible chapter boundaries (as markers on the Annotated Timeline) (variant 2).

3.1.2 Participants

Computer science students and colleagues were invited to participate in the study by e-mail. It is estimated that 100 to 150 people received the invitation since the invitation mail has been forwarded to instant messaging channels, among others. Of these, 38 people called up the web application and filled out the first questionnaire (20 for the lecture video and 18 for the documentary). During the second step of the study (pre-test before watching the videos, 4 persons ended the experiment prematurely (all of them were assigned to the lecture video). Of the remaining 34 persons, another 2 persons interrupted their participation in the study while watching the video (one person per video type). Of the remaining 32 persons, 6 briefly watched the video and then proceeded to the next step, only stopping the experiment without editing the two questionnaires afterwards (2 were assigned to the lecture video, 4 were assigned to the documentary).

Of the 26 persons who filled in all questionnaires, 4 were not considered because they had seen less than 10 % of the video material (all of them were of the lecture type). In total, 22 of the 38 persons who had started the study could still evaluate the data. This corresponds to a rate of 57.8 %. In the following, the number of participants will always refer to the 22 persons (14 male, 8 female). 17 participants, and thus the majority, were between 25 and 45 years old. Two participants were under 25 years of age, and 3 participants were over 45 years old. Table 3 shows how the participants are divided between the three treatments and the two video types. (Tables 4 and 5)

3.1.3 Material

The study was conducted on two different types of video. The first video is a lecture entitled “Mojib Latif: Herausforderung Klimawandel” (Engl. “Climate change challenges”, duration 1 hr 6 min 28 s),^{Footnote 1} which strongly resembles a lecture recording style in terms of structure and content. The second video is a TV documentary “Klimawandel konkret” (English: “Climate change concretely”, duration 50 min 18 s), which was a broadcast on the German state television channel ARD (Arbeitsgemeinschaft der öffentlich-rechtlichen Rundfunkanstalten der Bundesrepublik Deutschland).^{Footnote 2}

In the control condition (variant 1), participants can watch a non-segmented video. The video is not divided into subchapters and does not contain chapter titles. Participants can pause and resume the video anytime or jump forward or backwards as desired by clicking on the progress bar. The UI of the player as shown in Figure  11 is therefore hardly different from a standard HTML5 video player.

For variants 2 and 3, the video was segmented at previously selected points. Chapter transition animations were integrated at these points for both variants. The screen goes dark for 1–2 s, and the title of the current segment is displayed. However, both variants differ in the way the segmented content is presented. The learning environment of variant 2 (see Fig. 12) has been enhanced compared to variant 1 with a chapter overview, chapter boundaries, and navigation options. In variant 2, the video learning environment has been changed so that a chapter overview is permanently displayed on the left side of the video player. There, the participants can see the outline of the chapter structure and jump directly to a chapter by clicking on a chapter title. The chapters are also displayed on the video timeline. Red markers distinguish the different chapters from each other. Next to the timeline, another chapter navigation option that is also available in full-screen mode has been integrated.

In variant 3, the video was cut at the segment boundaries and represented as a series of individual videos, not played in a row. The segment boundaries were the same as for variants 1 and 2. As seen in Fig. 13, the individual videos are displayed as thumbnails above the video player. The corresponding video is loaded by clicking on the thumbnails or the titles. The next video is loaded by clicking on the button “Continue to chapter ...”. The participants must actively decide whether the next segment should follow directly, whether the person wants to reflect briefly on what he has learned or perhaps watch some content again. In addition, the title of the video that is currently shown is displayed directly above the player.

The video in variant 1 was not segmented. The video player provided only basic functionality

Full size image

Variant 2 with segmented chapters, which were labelled on the timeline and in a table of contents

Full size image

Variant 3 divided chapters into separate video vignettes. The chapter videos were provided as in variant 1 but with a thumbnail menu (see above) linked to the remaining chapter videos

Full size image

3.1.4 Procedure

Since the invitation e-mail contained a link to the web application, the recipients could start the web app immediately. When opening the web app for the first time, the participants were randomly assigned to a video type (lecture or documentary) and segmentation variant. During and after the study, the investigators did not communicate with the participants, and the conduct of the study was not actively monitored by the investigators. There were no problems during the implementation and no requests for assistance. The video usage data was collected by the web server. The survey results were stored in a Google spreadsheet.

3.1.5 Data Collection and Analysis

Since the study was conducted online and the investigators had no personal contact with the participants, socio-demographic data, knowledge acquisition, and feedback on the respective intervention were collected using four questionnaires and video analytics methods.

• Personal data: In detail, demographic data (age, gender, educational level), video consumption behaviour, interest in the video topic, and a self-assessment of learning and memory skills were collected.

• Pre-test: A test of the participants’ prior knowledge adapted to the video. The test consisted of 10 multiple-choice questions.

• Video analytics: During the video playback, clickstream data and time update events were collected in a log file.

• Feedback: Participants evaluate the learning environment regarding entertainment, structure, information content, and length. Additionally, questions about the learning content and climate change are asked again.

• Post-test: Repeat performance of the pre-test. The learning success results from the difference between the pre-test and post-test scores.

3.2 Results

The collected research data, including analysis scripts, is publicly available (Seidel, 2024).

Learning Gain Participants had minimal prior knowledge of all videos and across all groups (see Table 6). The topics covered in the documentary were intellectually less demanding, so the participants scored slightly better in the pre-test than in the lecture. The participants assigned to the three variants hardly differed from one another in terms of their prior knowledge. Only the participants in the control group (variant 1) recorded slightly better results (M=1.29, SD=1.11).

The post-test showed larger differences in performance. The differences between the two video types are only marginal, but the learners with variants 2 and 3 achieved significantly better results than the control group with variant 1. Compared to the control group, the learners with variant 2 achieved a good 22 % better test results (M=8.43, SD=0.98). With variant 3, the learners achieved an increase of 13 % (M=7.60, SD=1.85) compared to the control group (M=6.57, SD=1.62).

The differences in post-test are reflected in the learning gain. The participants scored almost 29 % more points with variant 2 than the control group. And also, with variant 3, the results were better by 22 %. Figure 14 and 15 show each group’s learning gains distribution.

To assess whether there were statistically significant differences in learning gains between variants 1 and 2, a Kruskal-Wallis rank sum test was performed. The results indicated a statistically significant difference in learning gains between the variants, \(\chi ^{2}(1) = 5.5844, p = 0.01812\). Given that the p-value is less than the alpha level of 0.05, we reject the null hypothesis. The mean learning gain for variant 2 (7.43) is higher than that of variant 1 (5.29) by approximately 2.14 points, based on the difference in the sample means, indicating a significant difference in performance.

The Kruskal-Wallis rank sum test for variants 1 and 3 resulted in no statistically significant differences, \(\chi ^{2}(1) = 1.81, p = 0.1785\). Since the p-value exceeds the alpha level of 0.05, we fail to reject the null hypothesis. This suggests that the observed differences in mean learning gains between variant 1 (5.29) and variant 3 (6.63) are not statistically significant and could be attributed to random chance. Although the mean learning gain for variant 3 is 1.34 higher than that of variant 1, this difference is not statistically meaningful. Further research with a larger sample size might be necessary to conclusively determine if there is a true difference between the variants.

Consequently, H1 can be accepted.

Box plot of learning gains by variant

Full size image

Distribution of learning gains by variant

Full size image

Perceived structuredness for each variant

Full size image

Video Interactions Based on the evaluations of other studies (cf. Biard et al. (2018); Schär and Zimmermann (2007)), it was to be expected that the pause function would be used only rarely. These expectations were confirmed. Only three participants used the pause function once each. Since there was no field study here, it can be assumed that the participants wanted to complete those presented in the study quickly and without interruptions. The mean paused periods were 27 s (type documentation variant 1), 89 s (type presentation variant 2) and 5 s (type presentation variant 1). In all cases, these are group participants who were shown a variant without automated breaks. Participants in group 3, who were shown the individual videos segmented with forced pauses, did not pause but, on average, allowed more time to pass before clicking on to the next video, usually 30 s to 2 min. One person waited over 3 h between two video segments. It is conceivable that another activity was started during this time, and the test participation continued afterwards.

Compared to the pauses, jumps in the video were used relatively often. 14 of 22 participants used the possibility to jump forward or backward on the progress bar in the video. The participants who used this possibility jumped on average 3 times while watching the video in one or the other direction. Of the 42 recorded jumps, 34 were forward, and 8 were backwards. When jumping backwards, it was mostly only a few seconds (approximately 7 to 15 s). When jumping forward, however, it was longer distances (approximately 30 s to 4 min). There was no repeated playing of a chapter or a longer passage; the few rebounds that were performed were usually only a few seconds. It is conceivable that something was overheard here which should be played again. It is also conceivable that the test persons remembered the questions of the prior knowledge test and recognised that what they had just heard or seen contained an answer to a question and should be looked at again. There were also strong outliers in the jumps. One participant jumped forward by 36 min directly after the second. One explanation would be that there was a technical problem (e.g., the browser crashed), and the participant had to reload the page.

Participants of group 1 were provided with the non-segmented version with the standard player, accordingly, they could not select chapters or segments. Participants of group 2 were able to orient themselves by the chapter markers on the progress bar and to jump to a chapter at will via the chapter menu. Out of 8 people in group 2, only 2 persons did not use the chapter menu and simply let the video scroll through from start to finish. Two people used the table of contents to jump to a desired location only once, three people three times, and one person four times. Group 3 participants had to click either the “Next” button or one of the chapter thumbnails to move on to the next video. All participants followed the chapters in the given order. There was no non-linear viewing of the videos. If chapters or sections were left out, they were not viewed later.

Perceived Structuredness The participants were asked to state how well-structured they perceived the video material. Both segmentation variants were rated on average 23 % as better structured than the non-segmented video (see Fig. 16). For both types of video, a clear picture emerges when evaluating the structuring. The advantages of segmentation on learning success identified in other studies are partly attributed to increased learning facilitation through better structuring of the learning material. The data collected here confirmed that the viewers also perceived and confirmed the improvement in structuring. It is striking that the evaluation of structuring, more than all other variables discussed here, hardly differs according to video type. The data suggest that there is a discernible tendency to perceive segmented videos as more structured if they have short chapter transitions and chapter titles. Consequently, H2 can be accepted.

Segmentation Support Variants 2 and 3 supported the segmentation of longer learning videos in different ways. This study should determine the extent to which both variants can be regarded as conducive to learning.

The comparison between variants 3 and 2 provides a mixed picture. Concerning the perceived structuring, both still achieved similarly good results. Variant 2, however, received an overall rating of 6.8 % better than variant 3 and a 10.5 % better rating of the information content, but was rated 5 % worse in the learners’ self-assessment, has a lower probability of being recommended to others and was also rated more often and more strongly than too long.

With regard to the measured learning success relevant to hypothesis H3, variant 2 (M = 7.25, SD = 1.16) scored 9.4 % better than variant 3 (M = 6.63, SD = 2.13). Thus, H3 can be confirmed. The chapter menu and the visible chapter boundaries raised this segmentation variant to the same high level as the variant segmented with individual videos concerning structuring. The chapter menu not only improved the structure and served as an identifier for the selected chapters but was also actively used as an interactive navigation element by the test persons. As the evaluation of the interactions with the video player showed, these possibilities were used by 8 of the 10 participants of group 2.

It is noticeable that all participants adhered to the given sequence of the segments, which were numbered with a chapter number. The navigation options were thus used to jump from the current segment to the beginning of the next segment or to skip a segment completely-but not to view segments in any order.

3.3 Discussion

The number of 22 participants was relatively small. Due to examining longer videos, the participants had to invest considerable time without receiving any financial or non-material incentives. An expense allowance might have increased the motivation to participate in the study and reduced the dropout.

Regarding video reception, it was impossible to determine how much attention the participants paid to the learning content. Additionally, it was impossible to control whether participants used external assistance when answering the test questions. Unfortunately, for technical reasons, it was impossible to record exact playback behaviour beyond clickstream data. Consequently, it was impossible to connect the learning effect and playback behaviour (e.g., rewatching certain sections).

The videos examined here both dealt with realistic video footage about climate change. Other domain-specific topics with the player variants examined here may lead to better or worse learning outcomes. The same applies to effects due to production styles other than lecture and documentation. Besides, it cannot be ruled out that certain learning scenarios may favour or restrict learning with one of the player variants.

The experimental groups were not pre-selected according to their level of knowledge. Although the level of prior knowledge was queried and can be compared among the test participants, the 10 multiple-choice questions are hardly suitable for grouping dedicated knowledge level differences. Disturbing variables, such as the handling of ignorance, can falsify the results. Some participants chose “Don’t know” for questions whose answers they did not know, and others chose any answer because they were ignorant. In addition to the “Don’t know” option, there were usually five other possible answers. With five answers, the strategy of randomly choosing between the possible answers would provide the correct answer on average in 20 % of the cases. With only 10 questions, this can skew the evaluation of the participants’ prior knowledge level. In addition, the learning content shown in both video types was aimed at people with no prior knowledge about climate change. Accordingly, the content is designed so that all learners are assumed to have no prior knowledge of the topic. Experts could probably only benefit from their previous knowledge to a very limited extent.

The multiple-choice questionnaires measured the participants’ memory and attention immediately after watching the video. Statements about the long-term learning effect require the implementation of follow-up tests.

Nevertheless, the study shows a correlation between the video player variants and the increase in knowledge. The segmentation effect known from previous studies (e.g. Spanjers et al. (2010); Mayer and Pilegard (2014)) could be confirmed. Both variant 2 and variant 3 showed better knowledge gains than variant 1.

By comparing variants 1 and 2, it could be shown that significantly better learning performances can be achieved with the same video length if the navigation of the video content per segment is supported. Compared to the work of Spanjers et al. (2010), the transitions between segments were not marked by a black screen or a pause but supported by an interactive table of contents and markers on the timeline. This form of segmentation variant could be called implied segmentation because the boundaries of the segments are only marked instead of breaking the video into independent video segments, as in variant 3. The implied segmentation realised by the video player supports the given content structure, making it more visible and easier to grasp cognitively.

In contrast to previous work in which mainly video length was considered as a dependent variable (Manasrah et al., 2021; Guo et al., 2014), this study was able to confirm results according to which the interactivity of the video players has an influence on learning performance (Geri et al., 2017). In the present study, however, it was shown that this interactivity can be designed to improve learning with longer-lasting videos. This is made possible by a table of contents, an annotated timeline with the segment boundaries, temporal pauses with identification of segment transitions, and segment titles. Thus, meaningful video segments become visible and entitled. This may thereby aid the chunking of memory components and aid the creation of a mental model about the sequence and relation of the segments’ contents (Gobet, 2005). The participants accordingly perceived the support of the segmentation of longer learning videos investigated in variant 2 as a more structured learning offer.

In the comparison of variant 1 with the separated videos in variant 3, an increase in knowledge could be determined, although not significantly and less than in the comparison of variants 1 and 2. The form of segmentation was also perceived by the participants as structured. However, the given sequencing and segmentation led to rigid processing of the video segments, i.e. without breaks within a segment and without jumping back between the segments. The participants seem to have subordinated or at least adapted to the pre-organised segments. Participants in variant 3 could probably maintain attention over the length of the segments. It should be noted that variant 3 is very similar in design to the segmentation on YouTube but does not intermix with other potentially distracting video series.

In contrast to Moreno and Mayer (2007) and Doolittle et al. (2015), in all three variants, the learners could control the learning pace and the sequence of the segments themselves and thus regulate their learning. Results of another study with longer-lasting videos confirmed that external control of forced pauses at segment boundaries and alternative locations is not conducive to learning (Merkt et al., 2018).

The player variants were firmly defined. In variants 2 and 3, several design elements were combined. Compared to the more isolated examination of particular features supporting segmentation (Merkt et al., 2018; Merkt & Schwan, 2014; Moreno & Mayer, 2007; Doolittle et al., 2015; Manasrah et al., 2021; Guo et al., 2014) the combination of several features may have complemented each other.

In this article, we presented two studies. In the first study, we examined the video length and segmentations of 895 German and 524 English channels from YouTube Education (83,558 of 154,370 videos). We clustered the videos by their length into three groups. The largest group represents two-thirds of the videos, which are shorter than about 14 min (DE)/23 min (EN). This result shows that the production of short videos has proven itself within the examined sample of YouTube channels. Producing videos with up to about 14 or 22 min of playing time can be considered common practice. However, videos with a medium length between 14 and 39 min (DE), respectively 22 and 64 min (EN), also make up a significant proportion, while longer videos are the exception.

The channels considered here show a clear picture concerning the segmentation of educational videos. About 8 % of the videos with short lengths were part of a series of segmented videos. Very short videos with less than a minute duration were not segmented. This seems to be a lower limit for the length of a segment. Videos of medium length were part of a series in 14 % to 22 % of cases. This result shows that more comprehensive topics can be transported in the form of a video if it is divided into smaller segments. In the YouTube channels considered here, dividing comprehensive video-based learning resources into multiple segments is a common practice. Longer videos have their place in informal learning settings but are also important in formal university teaching (e.g. lecture recordings).

However, the implications set out above apply only to YouTube’s possible uses. The videos of the channels examined here were almost exclusively produced for YouTube and the tools available. Other video players may support other educational practices and enable other video usage patterns.

We conducted a second study to investigate design options for structuring longer videos into segments. This study aimed to gain insight into enhancing video players beyond the standard design provided by YouTube. We developed a video player capable of presenting a single (long) video file in three different conditions: (i) as a non-segmented video, (ii) as a video with an additional chapter overview, visible chapter boundaries, and navigation options for the segments, and (iii) as a sequence of segmented videos of suitable length. These three conditions were examined in an online user study. Two videos with varying amounts of information were used to compare learning effects and video interactions. Twenty-two participants were divided into groups corresponding to each condition.

Results showed that segmented videos, with sequences of up to 10 min, resulted in higher learning effects than the non-segmented version of the same video. Participants perceived the segmented videos in conditions (ii) and (iii) better structured. The provided controls in the segmented conditions were utilised to switch between the segments.

The findings from both studies highlight several practical implications for instructional designers, educators, and developers of educational technology: First, video players should support segmentation by incorporating features such as a table of contents, annotated timelines with segment boundaries, temporal pauses at segment transitions, and titles for each segment. This structure helps learners navigate and comprehend the content more effectively while maintaining control over playback. Second, dividing long and medium-long videos into shorter, sequentially accessible segments enhances learning outcomes. This method allows learners to focus on specific content parts and revisit them as needed. As seen in YouTube EDU channels, platforms like YouTube can facilitate this by listing related videos within playlists. Third, the advanced design features demonstrated in variants 2 and 3 can be implemented using browser-side web technologies such as JavaScript. This approach does not require splitting the underlying video file, making it a practical and efficient solution for enhancing video-based learning. Fourth, the overall length of the video may not be as critical as the ability to divide it into meaningful segments. Instructional designers should prioritise creating videos that can be easily segmented and navigated, ensuring that each segment is coherent and focused on specific learning objectives.

For future research, we propose to investigate what variants of video segmentation support learning in formal educational settings over a longer duration. Additionally, it should be explored how learners would individually adjust video segmentation according to their learning progress and current learning context to meet their specific needs in the current situation. From a learning analytics perspective, we suggest analysing video viewing behaviour using different segmentation support methods.