�ݺ�ߣ

Searching and browsing through fragments of TED Talks
MARIELLA SABATINO – mariella.sabatino@eurecom.fr
GO!
25/09/2014
1

TED is a global set of conferences, held throughout North America, Europe and Asia.
TED Talks address a wide range of topics within the research and practice of science and culture.
The speakers are given a maximum of 18 minutes to present their ideas in the most innovative and engaging way they can, often through storytelling.
TED Talks
25/09/2014
2

Problem
Users are overwhelmed with audiovisual content
Users browse fast, looking for topic of interest
Which are the fragments potentially relevant without having to watch the entire video?
It is very difficult to find interesting documents
25/09/2014
3

Research questions
how to recommend related media fragments within the same video collection
1
2
3
detect segments of interest in a video?
recommend related media fragments within the same video collection?
design a web application that provides a rich environment for exploring a video collection?
HOW TO:
25/09/2014
4

Browsing and recommendation of Media Fragments of TED Talks based on entities extracted in the subtitles
Integration of the Media Fragments concept and the subtitles enrichment performed by NERD on a Node.js server
HyperTED
25/09/2014
5

Research question 1
1
2
3
HOW TO:
25/09/2014
6

2
3
What is a NER task?
1
Named Entity Recognition (NER) aims to locate and classify elements of textual document into pre-defined categories such as:
•People names;
•Organizations names;
•Places;
•Temporal and numerical expressions. These elements and the categories take respectively the name of entities and ontologies.
25/09/2014
7

2
3
For example…
1
“This is Nikita, a security guard from one of the bars in St. Petersburg.”
“This is Nikita, a security guard from one of the bars in St. Petersburg.”
NER
Example taken from the transcript of
https://www.ted.com/talks/2089
25/09/2014
8
PERSON
FUNCTION
LOCATION
Category: type in the NER task.
Natural Language Processing (NPL) Task  disambiguating URL in a knowledge base.
E.g. http://dbpedia.org/resource/Saint_Petersburg.

Web Tools that use NER algorithms.
Open APIs for research use.
2
3
NER extractors
1
25/09/2014
9

2
3
NERD
1
Compare performance of NER tools available on web.
Unify the results of NER extractors in a common output.
http://nerd.eurecom.fr/
25/09/2014
10

2
3
NER extractors evaluation
1
DOCUMENTS ANALYZED: 5 short TED Talks NUMBER OF EVALUATORS: 1 STEPS OF EVALUATION:
•Selection of the meaningful concepts on the subtitles;
•Run of each extractor;
•Comparison of the results.
25/09/2014
11
PRECISION: the fraction of retrieved documents that are relevant RECALL: is the fraction of relevant documents that are retrieved. F-MEASURE: is the level of accuracy considering both the Precision and the Recall

2
3
NER extractors evaluation
1
EXTRACTOR
PRECISION
RECALL
F-MEASURE
AlchemyAPI
0,15
0,03
0,05147488928
DataTXT
0,21
0,36
0,2652521588
DBpedia Spotlight
0,14
0,37
0,1994140988
Lupedia
0,18
0,02
0,04389924763
OpenCalais
0,27
0,09
0,1347540544
Saplo
0,00
0,00
0
Textrazor
0,17
0,40
0,2416065311
THD
0,12
0,05
0,07485426603
Wikimeta
0,13
0,08
0,09514781377
Yahoo! Content Analysis
0,52
0,13
0,202927267
Zemanta
0,44
0,18
0,2511994999
Combined
0,11
0,54
0,1859774587
25/09/2014
12

http://www.w3.org/TR/media-frags/
2
3
A Media Fragment is a part of a multimedia object.
Temporal Fragments
sections along the time dimension of the media resource with a start and an end point.
http://www.w3.org/TR/media-frags/
Media Fragments
1
25/09/2014
13

2
3
TED Talks have paragraphs:
a human-made subdivision of subtitles.
MF creation: chapters
1
25/09/2014
14

Extraction of topic from TextRazor and entities from NERD
Clustering of consecutive chapters which talks about similar topics
Filtering of those fragments based on annotation relevance
2
3
MF creation: hot spots
1
The Hot Spots are those fragments whose relative relevance falls under the first quarter of the final score distribution.
25/09/2014
15

Research question 2
1
2
3
HOW TO:
25/09/2014
16

1
3
A search engine is a system able to access to information previously stored and indexed.
The search engine indexing is the process of collecting, parsing and storing data to make searches faster.
We use it for indexing annotations in our database
Search Engine indexing
2
25/09/2014
17

1
3
Because they “contain” the meaning of the talk
Because they contain some very useful attributes:
•timing references (startNPT and endNPT);
•uuid;
•relevance references.
Annotation based index
2
WHY ANNOTATIONS?
25/09/2014
18
WHICH ANNOTATIONS? Entities and Topics

1
3
ElasticSearch is an open-source search engine.
It uses Apache Lucene™ for indexing.
It aims to make full text search easy by hiding the complexities of Lucene behind a simple RESTful API.
ElasticSearch
2
25/09/2014
19

1
3
ElasticSearch provides a full Query DSL based on JSON to define queries. In general, there are basic queries such as term or prefix.
HOW TO MAKE A QUERY
25/09/2014
20
ElasticSearch
2

1
3
Recommendation
2
Interlinking through chapters and topic
Interlinking to openCourseware and openUniversity
25/09/2014
21

Research question 3
1
2
3
HOW TO:
25/09/2014
22

1
2
Architecture
3
25/09/2014
23

1
2
DEMO
3
25/09/2014
24
http://linkedtv.eurecom.fr/mediafragmentplayer

Conclusions
25/09/2014
25
Evaluation of NER tools in the context of TED Talks
HotSpot detection based on topics and entities
Recommendation algorithm, hyperlinks between fragment of TED talks + external education resources
Nice and responsive UI

Publications
25/09/2014
26
HyperTED is one of the submitted app at the Challenge at LinkedUP - http://linkedup-challenge.org/
José Luis Redondo García, Mariella Sabatino, Pasquale Lisena and Raphaël Troncy.
Detecting Hot Spots in Web Videos. In International Semantic Web Conference (ISWC’14), Demo

�ݺ�ߣ

HyperTED - Searching and browsing through fragments of TED Talks

More Related Content

HyperTED - Searching and browsing through fragments of TED Talks