Metadata extraction and classification of YouTube videos using sentiment analysis

Full text

1
Metadata Extraction and Classification of YouTube
Videos Using Sentiment Analysis
Shanta Rangaswamy, Shubham Ghosh, Srishti Jha
Department of Computer Science and Engineering
R.V. College of Engineering, Bengaluru, India
Soodamani Ramalingam
School of Engineering and Technology
University of Hertfordshire, Hatfield, UK
Abstract: MPEG media have been widely adopted and is very
successful in promoting interoperable services that deliver
video to consumers on a range of devices. However, media
consumption is going beyond the mere playback of a media
asset and is geared towards a richer user experience that relies
on rich metadata and content description. This paper proposes
a technique for extracting and analysing metadata from a
video, followed by decision making related to the video content.
The system uses sentiment analysis for such a classification. It
is envisaged that the system when fully developed, is to be
applied to determine the existence of illicit multimedia content
on the web.
Keywords: MPEG, Metadata extraction, Video processing,
sentiment analysis, polarity.
I. I
NTRODUCTION
Recently, there has been a series of terrorist attacks such
as the July 2016 attack on Munich in Germany, Nice in
France and Dhaka in Bangladesh, to name just a few. The
gunman in the Nice attack is believed to have visited
websites that showed pictures of executions before making
his attack. Such terrorist activities call for monitoring of
web activity related to the gunmen involved. Governments,
in dealing with actions against such hate crime, have a need
to censor such websites. However, the growth of the
information uploads over the World Wide Web is increasing
on a very rapid scale that it is impossible to sieve through
them manually. Hence, there is a need for an automatic
content analysis that that can listen to, read and extract
relevant information that it is looking for which is termed as
‘metadata’. There are many web platforms that are used to
share non-textual content such as videos, images and
animations that allow users to add comments for each item
[1].
Sentiment analysis or opinion mining is one of the great
accomplishments of the last decade in the field of Language
Technologies. This field of study is related to the analysis of
opinions, sentiments, evaluations, attitudes, and emotions of
users which they express on social media and other online
resources. The revolution of social media sites has also
attracted the users towards video sharing sites. YouTube is
probably the most popular of them, with millions of videos
uploaded by its users and billions of comments for all of
these videos. Online users express their opinions or
sentiments on the video that they watch on such sites.
Classification of video is an increasingly prominent area of
research, rising with the quantity of videos shared online
through such sites [2-3]. In general, sentiment analysis
attempts to determine the attitude of material contributors
with respect to the topic of interest or the overall contextual
polarity of the content. That is, whether the expressed
opinion in the content may be classified for example as
positive, negative or neutral or equivalent.
Thus, we can perform a sentiment analysis related to
hate crime that would potentially extract related video
instigating such terrorist activities in real life.
Literature review reveals that sentiment analysis has
been typically been carried on textual data such as in [5].
Such systems involve carrying out a geographic analysis of
crime data in understanding high crime areas and hot spots
using Twitter data. Sentiment statistics enables the
categorization of tweets by type, occurrences and number of
associated tweets. A sentiment score indicating the central
idea of tweets is determined. Sophisticated machine learning
algorithms such as Deep Learning and Affective Computing
techniques have been used for real-time analysis.
However, carrying out a similar task on video poses
high volumes of data handling and extracting meaningful
information from the video content, which is a non-trivial
problem. This area of research is gaining importance due to
the advancements and availability of automation tools [6].
Sentiment analysis on generic video content consists
typically of the following stages [7]:
●Event classification- to classify an event with
importance levels such as critical, high, etc.
Automation tools such as Wordnet [8] may be used.
●Polarity detection-rating the opinion into positive,
negative, neutral descriptions. Again tools such as
Wordnet may be used. Furthermore, typical phrases
used in social media may be interpreted to
determine the polarity. For example, flagged, self-
promotion, propaganda, abusive, etc. are some
978-1-5090-1072-1/16/$31.00 ©2016 IEEE

2
typical phrase descriptions used in polarity
extraction.
● Polarity prediction-Analysing users’ comments for
predicting polarity. This is related to deciding
comments from viewers of videos and thereby
deducting the polarity. Clustering and aggregation
techniques may e used for this purpose.
● Evaluation of retrieved content based on metadata:
Precision and Recall are two objectives measures
commonly used in content based multimedia
retrieval systems that is applicable here.
To the best of our knowledge, there is a lack of
information on video metadata extraction on crime related
content. This paper therefore proposes an initial model
development that would effectively extract meta data from
YouTube social media that is related to crime.
The rest of the paper is organized as follows: Section II is
about the stages of the algorithmic approach of metadata
extraction. Section III talks about the experimental analysis
with different case studies, followed by conclusion and
further work in section IV.
II. M
ETADATA
E
XTRACTION
–
A
LGORITHMIC
A
PPROACH
The proposed system utilizes certain aspects of metadata
based retrieval system such as in [10] for extracting the
metadata within the video content. The general stages of
development are described in the following sub-section.
A. Metadata Extraction- Stages of Development
The key stage involves understanding of the relevant
database followed by aggregation of collected metadata
information from selected website. The system architecture
flowchart is illustrated in Fig.1.
1) Metadata categorization: The Dublin Network
Working Group separates metadata into three groups namely
content, intellectual property and instantiation and provides
detailed description of each of their elements [9]. This
classification is reproduced in Table I. Such a categorization
is a useful step during the development of a metadata
extraction process.
Table I Metadata classification [9]
Content
Intellectual Property
Instantiation
Title Creator Date
Subject Publisher Format
Description Contributor Identifier
Type Rights Language
Source
Relation
Coverage
2) Metadata extraction from a media file involves
retrieving information such as the name of the media file,
links to it, metatags, partner information supply the content,
etc. The result of extraction is a list comprising of media and
web page information such as URLs, titles, keywords,
author, genre, etc. A typical output may appear as shown in
Table II.
Table II Extracted Metadata
Field Contents
The referring URL http://www.youtube.com
Media URL https://youtu.be/h0SXO5KUZIo
Title "Cyber-security and Cyberwar: What
Everyone Needs to Know"
Channel Talks at Google
3) Metadata parsing: Some of the extracted data may
often contain outliers such as noise, white spaces or figure of
speech which is parsed and indexed, and has been cross
checked against a database with set fields. This provides an
opportunity to rectify and segregate the noisy fields, thereby
adding the relevant data to the extracted list as shown in
Table III.
4) Metadata Lookup: The meta data aggregated from
the previous step is matched against a known database. Let
us assume that a media file is found to have information
related to “cyber security and cyber war” in a google talk by
a media personality. The metadata is extracted and has the
following fields as shown in Table III. The fields are
compared against the known database. This process enables
to identify if the website is related to promoting crime.
Table III Metadata parsing
Field Contents
Published on Feb 10, 2014
Title “Cyber Terrorism and Warfare: The Emergent Threat”
Category News and Politics
Number of views Around 34k
Performer Peter Warren Singer
License YouTube Standard license
B. Metadata Extraction Principle and Procedure
The proposed system is developed in Python, version
2.7.9. Firstly, the source code of a YouTube webpage for a
specific video (given its URL) is obtained using Python
library urllib2 which uses request.get () function. For
metadata extraction, the Regular Expressions Python inbuilt
library is used as this makes the process of string matching
easier and efficient. Using string matching the Title,
Description, Number of Views and Category of the video are
extracted. Due to the flexibility in the metadata extraction
code, any other information can be extracted. Further, the
code uses Python's inbuilt Natural Language Processing
Toolkit (NLTK), which performs text parsing of the
metadata obtained above and it maps the required tags
depending on nature of the word to each token. The code
also considers a customized Python inbuilt Dictionary
Corpus which consists of all the possible English words with
their assigned ratings ranging from -15 to 15. This is then
mapped with the required tags to calculate the aggregated

3
rating of the video. The above setup is run on Ubuntu 16.04
platform, but it can be made to run on a windows
environment as well.
C. Metadata Extraction Implementation - Algorithm
In this Section, an algorithmic approach is outlined for
metadata extraction and analysis of the extracted data. It
works on the principle of seeking a URL, specifying specific
fields of interest that form part of the query, parsing
retrieved data, and classifying with reference to a predefined
dictionary.
Step 1: Extraction of source code from a YouTube video
Fetch the source code of a YouTube webpage of a video
given its URL using Python interface function urlopen:
usock = urllib2.urlopen(url)
The output is stored in usock and in the variable name is
“data”.
Step 2: Extraction of specific elements of the video
Using python string manipulation operations match the
string “<meta itemprop="genre" content="” from the inspect
element of the video URL to the obtained source code. After
the match occurs we copy the text following the matched
string into a text file called “Meta_data_outputs.txt”.
Pseudocode A provides the steps involved:
Pseudo Code A: Video information extraction
s2 = '<meta itemprop="genre" content="'
y = data. index(s2)
l = len(s2)
s3 = '<div id="watch-header" class="yt-card yt-
card-has-padding">'
y2 = data. index(s3)
category = data [y+l: y2]
category =re.sub ('">', '', category)
text_file = open ("Meta_data_outputs.txt", "a")
text_file.write("\r Category:%s\r" %category )
text_file.close()
The above procedure can be used to extract information
such as the video description, number of views, likes and
dislikes. Also, the comments can be extracted. Any other
information which may help in classifying the video may
also be extracted at this stage.
Step 3: Defining a dictionary of words
The next step is to recognize positive and negative
expressions. This is done by referencing a customized
dictionary file that is but with words and their assigned
rating. Such type of customized dictionary file is named
“Corpus.txt” and a sample from the same is shown below:
nice: [positive]
motivation: [positive]
inspirational: [positive]
bad: [negative]uninspired: [negative]
expensive: [negative]
This dictionary needs to be enhanced progressively for
better results by addition of more words with their respective
ratings.
Step 4: Parsing retrieved data
The retrieved data is in the form of a text as a series of
sentences. These are stored in a file Meta_data_outputs.txt.
During parsing of the output file, each token is assigned a
specific tag using the NLTK. These tags are labels
associated to each word depending on the type of the word,
such as a noun, verb, adjective etc.
The following structure will be used:
●Each text is a list of sentences.
●Each sentence is a list of tokens.
●Each token is a tuple of two elements: a word form
(the exact word that appeared in the text) and a list
of associated tags.
Pseudo Code B: Metadata Parsing
tokens=nltk.word_tokenize(temp) # variable temp holds
each sentence temporarily tagged=nltk.pos_tag(tokens)
Let us consider an example of retrieved text:
“All that is gold does not glitter”. This is tokenized as
follows: [[('All’, ['DT']), ('that’, ['DT']), ('is’, ['VBZ']),
('gold', ['NN']), ('does’, ['VBZ']), ('not’, ['RB']), ('glitter’,
['VB']), ('.', ['.'])],
Another example: “Not all those who wander are lost.”, this
is tokenized as per the following:
[('Not', ['RB']), ('all', ['DT']), ('those', ['DT']), ('who', ['WP']),
('wander’, ['NN']), ('are’, ['VBP']), ('lost', ['VBN'])]
S
tep 5: Mapping Metadata to the Dictionary:
In this step, the list of required tags is determined as
indicated in Pseudocode C.
Pseudo Code C: Tag Mapping with Rating
required tags= ['JJ', 'JJR', 'JJS', 'NN', 'NNS', 'NNP',
'NNPS', 'RB', 'RBR', 'RS', 'VBG', '-NONE-', 'VBZ']
The mapped metadata has a predefined rating for the
words present in the customized “Corpus.txt”
dictionary file.
Step 6: For Calculating rating of the video:
General formula:
●The positive and negative ratings are added separately.
Then final rating is calculated using an aggregation as
shown in Pseudocode D:

4
Pseudo Code D: Rating Aggregation
tsum=psum-nsum
if sum>0:
pol = (float) (tsum / psum)
if pol!= 1:
pol = pol * 0.5
tot += pol + 0.5
else:
pol=div(max(pword),4) *0.5
tot=tot+pol+0.5
else if sum< 0 :
pol = (float) (tsum / nsum)
if pol != 1 :
pol = pol * 0.5 * -1
tot += pol + 0.5
else:
pol=div(min(nword),-4)*(-0.5)
tot=tot+pol+0.5
Here, pol is a temporary variable, and tot is the
total rating of the document.
● Number of sentences are calculated. Average rating
of entire document is calculated using:
if sentence! =0: rating=(tot/sentence)
where “sentence” is number of sentences.
Fig.1: Diagrammatic representation of the entire
implemented system.
III. E
XPERIMENTAL
A
NALYSIS
In this Section, we consider a case study that demonstrates
the principle of metadata extraction from video and its
classification. Firstly, the datasets used for metadata
extraction is described followed by an analysis of the results
obtained for the datasets.
A.
Dataset
The dataset consists of a set of 15 YouTube videos. The
dataset categorization is as follows:
● Negative Sets: A set that is inappropriate for
viewership as they deal with criminal cases such as
theft, harassment and abuse.
● Positive Sets: A set of videos having high
viewership content and dealing with motivational
and inspirational values are considered. These
videos also talk about security and safety related
issues.
● Neutral Sets: A set of videos which are neutral and
don’t have any impact (positive or negative) on the
society. So each video is given a rating which
ranges between 0 to 1, based on the algorithm
designed above. Depending on the above rating
videos are classified into three class labels:
Positive, Negative and Neutral.
Case Study: Let’s consider the URL
“https://www.youtube.com/watch?v=Tdv3TIuFvMs “, The
metadata contained within the URL is as following:
“Title: One Minute Inspiration - Never Give Up! Be
Successful!
No of views: 25
Published on: Apr 18, 2014
Success is just around the corner for those who seize the day
and refuse to give up. Your goals are important and can be
met if you try hard and persevere.
You can do it!
Category: People & Blogs
License: Standard YouTube License”
This metadata is extracted and stored as a text file:
“Metadata_output.txt”.
This output file is then parsed to obtain the following tokens:
< ['\\Title’, ‘One', 'Minute', 'Inspiration', 'Never', 'Give', 'Up',
'Be', 'Successful’, ‘YouTube', 'Category’, ‘People’, ‘amp’,
‘Blogs', 'Find', 'out', 'why', 'Close']
['No', 'of', 'views’, ‘26', 'Description', 'Published', 'on', 'Apr',
'18', '2014', 'Success', 'is', 'just', 'around', 'the', 'corner', 'for',
'those', 'who', 'seize', 'the', 'day', 'and', 'refuse', 'to', 'give', 'up']
['Your', 'goals', 'are', 'important', 'and', 'can', 'be', 'met', 'if',
'you', 'try', 'hard', 'and', 'persevere']
['You', 'can', 'do', 'it', ‘!’] >
The next process is to obtain all possible tags from the
natural language processing toolkit (NLTK) which are
required for calculation of rating:
Required tags= ['JJ', 'JJR', 'JJS', 'NN', 'NNS', 'NNP', 'NNPS',
'RB', 'RBR', 'RS', 'VBG', '-NONE-', 'VBZ']
Tokens that are obtained from the tokenized metadata and
the tags obtained above are mapped together to associate a
predefined rating for the words present in the customized
dictionary.
Example illustrated below:
############################################

5
[('\\Title', 'NN'), ('One', 'CD'), ('Minute', 'NNP'),
('Inspiration', 'NNP'), ('Never', 'RB'), ('Give', 'VBP'), ('Up',
'RP'), ('Be', 'NNP'), ('Successful', 'JJ'), ('YouTube', 'JJ'),
('Category', 'NN'), ('People', 'NNP'), ('&', 'CC'), ('amp', 'NN'),
('Blogs', 'NNP'), ('>', 'NN'), ('Find', 'NNP'), ('out', 'RP'),
('why', 'WRB'), ('Close', 'JJ'), ('No', 'NN')]
('word = ', 'inspiration');
('rating = ', 10)
(‘word =’, ’Close’);
(‘rating = ‘, 2)
(‘word =’,’Successful’);
(‘rating = ‘, 8)
############################################
[('No', 'NN'), ('of', 'IN'), ('views', 'NNS'), ('26', 'CD'),
('Description', 'NN'), ('Published', 'VBN'), ('on', 'IN'), ('Apr',
'NNP'), ('18', 'CD'), ('2014', 'CD'), ('Success', 'NNP'), ('is',
'VBZ'), ('just', 'RB'), ('around', 'IN'), ('the', 'DT'), ('corner',
'NN'), ('for', 'IN'), ('those', 'DT'), ('who', 'WP'), ('seize',
'VBP'), ('the', 'DT'), ('day', 'NN'), ('and', 'CC'), ('refuse',
'NN'), ('to', 'TO'), ('give', 'VB'), ('up', 'RP')]
('word = ', 'just');
('rating = ', 3)
(‘word = ‘,’Success’);
(‘rating =’, 7)
############################################
[('Your', 'PRP$'), ('goals', 'NNS'), ('are', 'VBP'), ('important',
'JJ'), ('and', 'CC'), ('can', 'MD'), ('be', 'VB'), ('met', 'VBN'),
('if', 'IN'), ('you', 'PRP'), ('try', 'VBP'), ('hard', 'JJ'), ('and',
'CC'), ('persevere', 'JJ')]
('word = ', 'persevere')
('rating = ', 5)
(‘words =’,’goals’)
(‘rating =’, 8)
############################################
The aggregated rating of the entire set of extracted meta-data
is calculated using the algorithm explained in section II
(Algorithmic approach- Step 6) and will be categorized
based on the two threshold values defined in the next
section.
Result of the above sample set: Overall rating of the video
is: 0.9375. The ratings for the whole dataset is given in
Fig.2.
Category: Positive
B. Metadata Extraction: Sensitivity Analysis
The classification of the videos is done based on the final
aggregated rating obtained. A range of ratings have been
defined that is used to classify the videos as positive,
negative or neutral. Each of these set of ranges have a
defined threshold. We analyse the performance by using two
different thresholds:
Fig. 2. Graph Depicting Video Number Vs Rating
Table IV: Use of thresholds for determining polarity
Video
No.
Threshold 1 0 to 0.3
Negative
0.3 to 0.7
Neutral
0.7 to 1.0
Positive
Threshold 2 0 to 0.45
Negative
0.45 to 0.65
Neutral
0.65to1.0
Positive
Overall
Rating
V1 0.125
✓
✓
V2 0.40625
✓
✓
V3 0.1591
✓
✓
V4 0.6819
✓
✓
V5 0.9375
✓
✓
V6 0.075
✓
✓
V7 0.125
✓
✓
V8 0.75
✓
✓
V9 0.6629
✓
✓
V10 0.875
✓
✓
V11 0.5834
✓
✓
V12 0.512
✓
✓
V13 0.475
✓
✓
V14 0.4685
✓
✓
V15 0.5205
✓
✓

6
For example, Threshold 1 is set as: (0 to 0.3) - Negative; (0.3
to 0.7) - Neutral; (0.7 to 1.0) - Positive.
Similarly, Threshold 2 is set as: (0 to 0.45) - Negative; (0.45
to 0.65) - Neutral; (0.65 to 1.0) - Positive.
These thresholds then classify the videos into their
respective class labels. The resulting classification of video
dataset is provided in Table IV.:
From Table IV, it can be observed that though the
threshold values set for negative, neutral and positive
changes; the category into which a video is classified doesn’t
change by much. The dataset used in the experimental
analysis includes 15 videos each ranging from 30MB in size
to 400MB. Also, the metadata extracted such as description,
number of views, and category of a video is a text file and
has size in KB. But, when the size of the dataset increases
the classification will also start varying. The number of
variations withineach threshold will be large which is not
seen with the above sample range. It has also been observed
that threshold 2 classifies each of the videos accurately into
either positive negative and neutral. Hence, threshold 2 is
preferred over any other boundary conditions for each of the
class labels.
When Threshold 1 was used to classify the above sample
set, one video was classified inaccurately while the other 14
were accurate. On the other hand, with threshold 2 all of the
15 videos from the above sample set were classified
correctly. This variation in result is due to this particular
sample set size. It is believed that if the size of the dataset
increases the classification will also start varying. For a
larger dataset, different threshold ranges will need to be
established for much better classification.
IV. C
ONCLUSION AND
F
URTHER
W
ORK
In this paper, a technique for analyzing YouTube video
URLs is proposed. The technique follows a sentiment
analysis to determine the polarity for the video objects. The
system has been tested on a small dataset. The results are
promising.
The accuracy of the entire process depends on the list of
words with their respective ratings which are present in
Python inbuilt dictionary Corpus. For the purpose of testing
the enhancement of Corpus, by adding new words and their
respective ratings was done manually. This process needs to
be automated with growing size of the dataset. A manual
update of the Corpus dictionary would become time
consuming and impractical. To automate the process, we
propose using Machine Learning concepts such as Neural
Networks, Genetic Algorithms, SVMs and Bayesian
Learning. Further, when the size of the dataset increases the
classification of the videos may vary and in order to have a
robust classification, different threshold ranges will need to
be established.
V. R
EFERENCES
[1] Choudhury, Smitashree and Breslin, John G. (2010), “User sentiment
detection: a YouTube use case”, Proc. The 21st National Conference
on Artificial Intelligence and Cognitive Science, 30 August - 1
September 2010, Galway, Ireland, 2010.
[2] Kundi, F. M., Ahmad, S., Khan, A., &Asghar, M. Z. Detection and
Scoring of Internet Slangs for Sentiment Analysis Using
SentiWordNet. Life Science Journal, Vol. 11 No. 9, 2014.
[3] G. Vinodhini and RM.Chandrasekaran, “Sentiment analysis and
opinion mining: a survey”, International Journal of Advanced
Research in Computer Science and Software Engineering, Vol.2,
No.6, pp. 282-292, June 2002.
[4] Jain, R. and Fuller, C. and Gorkani, M.M. and Horowitz, B. and
Humphrey, R.D. and Portuesi, M.J. and Shu, C. and Hampapur, A.
and Gupta, A. and Bach, J.}, “Video cataloger system with
extensibility”, Google patents, US Patent 6463444, 2002.
[5] Raja Ashok Bolla,” Crime pattern detection using online social
media”, MSc Thesis, Missouri University of Science and Technology,
US, 2014.
[6] Wootton, Cliff, “Developing quality metadata: building innovative
tools and workflow solutions”, Focal Press, 2009, ISBN-13:978-0-
240-80869-7.
[7] Muhammad ZubairAsghar, Shakeel Ahmad, Afsana, Marwat,
FazalMasudKundi, “Sentiment Analysis on YouTube: A brief
survey”, MAGNT Research Report Vol.3, No.1, pp: 1250-1257, 2015,
ISSN. 1444-8939.
[8] Fellbaum, Christian, “WordNet and wordnets”. In: Brown, Keith et al.
(eds.), Encyclopedia of Language and Linguistics, Second Edition,
Oxford: Elsevier, 665-670, 2005.
[9] Stuart L. Weibel, John A. Kunze, Carl Lagoze and Misha Wolf,
“Dublin Core Metadata for Resource Discovery”, The Internet Society
1998, URL: http://www.ietf.org/rfc/rfc2413.txt, last accessed: 27 July
2016.
[10] Ken Alan Berkun, Austin David Dahl, Jennifer Lynn Kolar, Scott
Chao-Chueh Lee, Shannon E. McRae, Brad Steven, Miller, Mercer
Island, John Prince, Bellevue, Eric CarlRehm, Srinivasan
Sudanagunta, Seattle, Jonathan Robert Nowitz, Seattle, “Interpretive
Stream Metadata Extraction”, Patent No. US 2002/0103920 A1, 1
Aug.2002.
[11] Siersdorfer, S., Chelaru, S., Nejdl, W., & San Pedro, J. “How useful
are your comments? analyzing and predicting youtube comments
and comment ratings”, ACM. In Proceedings of the 19th international
conference on World Wide Web, pp. 891-900, 2010.
[12] Chetan Verma, Sujit Dey, Methods to Obtain Training Videos for
Fully Automated Application-Specific Classification, IEEE Access,
The Journal for rapid open access publishing, DOI: 10.1109/
ACCESS.2015.2461156, Aug 7 2015, pp1188- 1205
[13] Chng, X., Dale, C., & Liu, J., “Understanding the characteristics of
internet short video sharing: YouTube as a case study”, Technical
Report arXiv 0707.3670v1, Cornell University, July 2007.

2
APPENDIX A: List of Videos used as dataset:
●How to Make a Bomb Cracker (Home Made) - Easy
Tutorials - YouTube
(https://www.youtube.com/watch?v=o--87j1dysU)
●How to make Coloured smoke from Wax Crayons.
Smoke bomb/ grenade for paintball, airsoft..etc. -
YouTube(
https://www.youtube.com/watch?v=fdeXcGkqT_4 )
●Thieves Stealing ATM - YouTube
(https://www.youtube.com/watch?v=fGiCaheWGbI )
●Video: Mom Ditches Baby at Walmart After
Shoplifting - YouTube
(https://www.youtube.com/watch?v=pgn3j1FHOVs )
●Cyber Terrorism and Warfare: The Emergent Threat -
YouTube
(https://www.youtube.com/watch?v=CNE1tQoObbs&
feature=youtu.be )
●Video that will change your life. I have no words left.
- YouTube (https://www.youtube.com/watch?v=PT-
HBl2TVtI )
●STOP KILLING TIME ► Motivational Video -
YouTube
(https://www.youtube.com/watch?v=UX2tefQHNmk
)
●World Best Motivational Videos for Students -
YouTube
(https://www.youtube.com/watch?v=Tjnq5StX68g )
●INSPIRATIONAL - HOW GREAT I AM - YouTube
(https://www.youtube.com/watch?v=V6xLYt265ZM )
●One Minute Inspiration - Never Give Up! Be
Successful! - YouTube
(https://www.youtube.com/watch?v=Tdv3TIuFvMs )
●Factory made gasoline vapor carburetor from the past
– YouTube
(https://youtu.be/zaiygknSFHs)
●Mark Zuckerberg 2016 The Lifestyles of Young
Billionaire Entrepreneurs - YouTube
(https://youtu.be/gB9Tv7vtWVU)
●How to Fly a Drone
(https://youtu.be/OcxUCepBHkM)
●Playing to the Edge with General Michael Hayden -
YouTube
(https://youtu.be/etffkFDm2NQ)
●Jeremy Howard: The wonderful and terrifying
implications of computers that can learn - YouTube
(https://youtu.be/t4kyRyKyOpo)