TikTok is a video-sharing social networking service that is rapidly growing in popularity. It was the second most downloaded app in the app world in 2020. While the platform is known for having users post videos of themselves dancing, lip-syncing, or showing off other talents, videos of users sharing specific knowledge have increased because of initiatives such as #learnontiktok. This study aims to assess the types of knowledge and learnings shared on TikTok and the profile of its users. We collect a set of videos. Trough and innovative framework implemented with computer vision, natural language processing, and machine learning techniques, we show the main teaching topics published in the #learnontiktok campaign and the disciplines with the highest audience engagement.
Micro-Scholarship, What it is, How can it help me.pdf
Tiktok and Education-ICEDEG 2021.pptx
1. Tiktok and Education: Discovering
Knowledge through Learning Videos.
Carlos Fiallos
Angel Fiallos
Stalin Figueroa
July 2021
2. Agenda
• Introduction
• Objective
• Methodology
• Data Collection
• Video Indexer Process.
• Exploratory and Demographic Analysis
• Multi-Label Categorization
• Results
• Conclusions
2
3. 3
Introduction
TikTok is considered a social media
platform because, like Twitter and
Instagram. However, TikTok offers users a
unique way to share creative videos of
themselves, their surroundings, or a
collection of external audiovisual content.
4. Introduction
TikTok launched LearnOnTikTok
program, which consists of
educational videos to facilitate
learning during COVID-19
lockdowns. These videos are
authored by professionals, students,
and other users, who have shared
their knowledge to this social
network's audiences.
4
5. Objective
This study aims to discover the types of knowledge and learnings
shared on #learnontiktok campaign, using a framework that
integrates computer vision, natural language processing, and
machine learning techniques.
5
6. 6
Methodology
The pipeline starts with data
collection from the TikTok
platform. Then it continues
with Computer Vision
processes and finishes with
multi-classification text
models available to predict
the knowledge areas of the
educational videos.
7. 7
Data Collection
We used scraping algorithms
developed for this objective. Next, a
sample of 1495 TikTok posts using
the hashtag #learnontiktok was
selected to obtain metadata from
them, such as video file, post
description, counts of likes, date,
number of views, author information,
and profile picture.
8. 8
Video Indexer Process.
Azure Video Analyzer for Media is a cloud
application, part of Azure Applied AI
Services. The API extracts the insights
from your videos using Video Analyzer for
Media video and audio models.
Also, we use an Azure Optical Character
Recognition (OCR) service for text
extraction on specific images.
10. 10
Multi-Label Categorization
• We designed a simple CNN network composed for an input layer and a convolution
layer of word vectors obtained from Word2Vec model. The training dataset was
composed of text sentences tagged in 20 specific areas from Wikipedia
Excerpt from Wikipedia sentences dataset
11. 11
Results
The Face API process
was applied to the
TikTok profile's photos
for the recognition of
facial properties. The
following results show
the percentages
belonging to gender
and user groups by age
range.
Percentages of detected genre.
Percentages of detected age ranges.
12. 12
Results
Figure shows a word cloud with the
most relevant terms related to video
descriptions registered by authors
Figure shows tags belonging to the
elements identified by the video
indexer process were selected for
each of the videos
13. 13
Results
Histogram with the most relevant labels
from videos
Percentages of knowledge areas with the
highest engagement
14. 14
Conclusions
The proposed framework allows us to identify the main areas of
knowledge associated with educational videos on the TikTok
platform. This information would allow us to add efforts in
important knowledge areas, but which are not widely accepted or
have few content creators.
We find a more extensive collection of Health Sciences videos and
related to STEM areas, even higher than social sciences such as
law and education.
TikTok is considered a social media platform like Twitter and Instagram. It has more than 800 million monthly active users
its users have a social group of followers and other users they follow . However, TikTok offers to users a unique way to share creative videos of themselves (either dancing, lip-syncing), their surroundings, or a collection of external audiovisual content.
The most straightforward videos consist only of text superimposed on a colored background. Videos can be more complex by including images, video clips, and sounds.
TikTok launched LearnOnTikTok program, which consists of educational videos to facilitate learning during COVID-19 lockdowns. These videos are authored by professionals from different areas, students, and other users, who have shared their knowledge to this social network's audiences. The videos linked to the hashtag #learnontiktok.
The videos linked to the hashtag #learnontiktok, have varied topics: from chemistry experiments, cooking recipes, health tips, learning other languages, to creating origami figures, all created by its users.
This study aims to discover the types of knowledge and learnings shared on #learnontiktok campaign, using a framework that integrates computer vision and audio recognition approaches for processing text and metadata information that is part of videos. It also contemplates processes for classifying the information collected into science categories using natural language processing and ML techniques.
The pipeline starts with data collection from the TikTok platform. Then it continues with the use Computer vision and audio recognition models to obtain text metadata from the video files. Finally, a trained multi-label classification text model is available to use text metadata to predict the knowledge areas of the educational videos.
I will now explain each step
We used scraping algorithms developed for this purpose. First, we searched with the hashtag #learnontiktok, which yielded a limited number of videos. From that result, we selected a sample of 1,495 TikTok posts
Next, we collected posts metadata, such as video file, post description, counts of likes, date, number of views, author information, and profile picture. The period of the posts was from June 2020 to January 2021.
Azure Video Analyzer for Media is a cloud application, part of Azure Applied AI Services. The API extracts the insights from your videos using Computer vision models and also audio models for transcriptions. Also, we use an Azure Optical Character Recognition (OCR) service for text extraction on specific images. Azure uses an internal algorithm to infer the correct string to be presented by correcting the mistakes introduced by individual OCR detections. The audio transcriptions and text from video snippets were acquired for further processing.
Optical character recognition or optical character reader (OCR)
We used the images from the original videos that included the user’s face profiles and processed them via Microsoft’s Azure Face API, which allows us to infer gender and age in json format. Once the process is finished, we selected the photos in which the exposure value was greater than 0.5 and the gender and age properties could be detected.
Following the Kim approach [14], we designed a simple CNN network composed for an input layer with five different ngrams window sizes and one layer of convolution on top of word vectors obtained from Word2Vec unsupervised neural language model [15]. The training dataset was composed of text sentences tagged in 20 specific areas from Wikipedia, such as medicine, food and drink, legal, physics, chemistry, among others.
The networks try to predict 0 or 1 values on every label, and the model uses the confidence values to produce a ranking.
The Face API process was applied to the TikTok profile's photos for the recognition of facial properties. The following results show the percentages belonging to gender and user groups by age range: the male gender and the 18-34 age group had the highest percentage
The rest of the photos of user profiles, among other reasons, did not show the user's face or belonged to business profiles, could not identify gender and age properties.
First Figure shows a word cloud with the most relevant terms related to video descriptions registered by authors. Only terms identified as nouns, through Part of Speech Tagging libraries, were selected for analysis. Some words such as “psychology”, “food”, “life”, “amazon”, “fun”, could be identified, which give a weak idea about the topics related to educational videos.
Then, the tags belonging to the elements identified by the video indexer process were selected for each of the videos. Terms such as "person," "text," "indoor," "clothing", "hair" can be identified, which relate to characteristics of the authors and the video background but continue to give a weak idea about the topics and area of knowledge covered in the video
Next, we applied the multiclassification model to the keywords obtained from the identification of the text snippets and the text transcription of the audios to assign the knowledge areas to each video. The model returns a set of probabilities labels and the following knowledge areas were identified as having the highest counts. It can be seen in Figure , that Medicine, Food and Drink, Health, Cooking, Biology, Chemistry, among others, are the most relevant categories.
Finally, we associated the knowledge areas with the number of likes for each video in order to establish the categories that had the highest user engagement. The results are shown in Table . Medicine, Food and Drink, Health, Chemistry, and Technology are the areas with the best engagement by the TikTok audience.
The proposed framework allows us to identify in an automatic way the main areas of knowledge associated with educational videos on the TikTok platform and which areas are the most preferred by users. This information would allow us to add efforts in important knowledge areas, but which are not widely accepted or have few content creators.
In our sample, we find a more extensive collection of Health Sciences videos and related to STEM areas, even higher than social sciences such as law and education, which gives an impression of the potential of this type of videos for science learning. The study also confirmed that most authors are people under the age of 34, who also represent the largest audience on the social network.
This study also supports the idea that audio and text metadata information available in short TikTok videos contains concepts that give rise to a better understanding of the video learning topics than even the descriptions registered by the authors.