A framework for real time semantic social media analysis

A FRAMEWORK FOR REAL-TIME
SEMANTIC SOCIAL MEDIA
ANALYSIS
Diana Maynard, Ian Roberts,
Mark A. Greenwood, Kalina
Bontcheva
ZELIA BLAGA

DOMAIN
• collect and analyse large volume of social media content
• provides:
behavioral evidence
opinion mining
sentiment analysis
• case study: 2015 UK elections, Nesta-funded Political Futures Tracker
project [1]
2

CHALLENGES
• dynamic content
• reflection of author’s social and sentimental fluctuations
• nuances in communication such as sarcasm
• activity on social media sites is triggered by specific events (e.g.
sports events, celebrations, crises, news) and topics (e.g. global
warming, terrorism or immigration)
• Twitter: reactive medium, opinions tend to be event-driven rather
than topic-driven
3

RELATED WORK
• Twitris [2]
• sentiment analysis tool
• able to take a sample of social media chatter about a specific topic
and deduce real-time, large-scale, automated sentiment about the
specific topic
• cloud-based visualization
• static
• example: analyzed Twitter chatter leading up to the Great
Britain/European Union Membership Referendum (Brexit) on June
23rd; was able to predict six hours before the news broke that the
polls leaning toward the “remain” camp were incorrect
4

RELATED WORK
• TwitInfo [3]
• sentiment analysis tool
• used timeline to display tweet activity during a real-world event
• colour-coded for sentiment
• dynamic
• example: a football game
5

FRAMEWORK
• live processing system
1. Collector
2. Processor
3. Indexing and querying
6

FRAMEWORK
1. Collector
• receives tweets from Twitter via their streaming API and forwards
them to a reliable messaging queue; saves the raw JSON of the
tweets in backup files for later re-processing if necessary
• 2 constraints: track (textual filter that delivers all tweets that mention
specified keywords/ hashtags) and follow (a user ID filter that delivers
all tweets by specified Twitter users, as well as any tweet that is a
retweet of, or a reply to, a tweet by one of the specified users)
7

FRAMEWORK
2. Processor
• consumes tweets from the message queue, processes them with the
GATE analysis pipeline (Natural Language Processing) [4] and sends
the annotated documents to GATE Mimir [5] for indexing
• standalone Java application built using the Spring Boot framework
(message consumer application)
8

GATE
• GATE = Language Resources (ontologies, lexicons) + Processing
Resources (parsers, generators) + Visual Resources (GUI)
9

FRAMEWORK
3. GATE Mimir
• receives the annotated tweets and indexes their text and annotation
data, making it available for searching [5]
• not purely keyword based
• semantic-based search that can be performed over categories of
things
• can include synonyms, unknowns (e.g. amount of money), ranges
(e.g. values between 2 given numbers), restrictions to date periods,
domains etc.
10

MIMIR
• semantic search over text, document
structure, linguistic annotations, and formal
semantic knowledge
• indexes plain tweet text, structural metadata,
hashtags, mentions, semantic annotations,
and timestamps
• use Prospector to filter and visualize
correlations in large data sets (e.g. the most
frequent topics associated with positive
sentiment)
• temporal analytics – which topics are popular
over a time period
11

FRAMEWORK CHARACTERISTICS
• loosely coupled
• no direct dependency between the collector and processor
components
• communication is mediated through the message queue
• components can be distributed across different machines for higher
performance and/or robustness
• Mimir instance can easily sustain 10-15,000 tweets per minute
• multiple collectors can be run at the same time
12

ANALYSIS OF POLITICAL TWEETS
• long-term monitoring: list of all MPs, candidates and official party
accounts (560 MPs, 1811 candidates)
• collected 1.8 million tweets from 24 October 2014 until 13 February
2015, out of which 100k original, 700k replies, 1 million retweets
• for debates: track relevant hashtags (#leadersdebate, #BBCdebate),
and more general hashtags (#GE2015, #UKElection)
13

ANALYSIS OF POLITICAL TWEETS
• Semantic annotation done using TwitIE [6] - Named Entity
Recognition (identifying Persons, Places, Organisations etc.) and
Linking (mapping these to their respective URIs in Wikipedia or other
web-based knowledge sources)
• Sentiment analysis - detecting if a post conveys sentiment and if so,
whether it is positive or negative, the strength of this sentiment, and
whether the statement is sarcastic or not
• Topic detection - classifying terms according to the set of key
themes; performed by means of manually created gazetteer lists for
each topic
• MP and Candidate recognition
• Author recognition
14

TWITIE
• Information
Extraction
• GATE’s annotation +
ANNIE
• ANNIE = tokeniser,
sentence splitter, POS
tagger, gazetteer lists,
finite state transducer
• Gazetteer = lists such
as cities,
organisations, days of
the week; names of
useful indicators,
such as typical
company designators
(e.g. ‘Ltd.’), titles
15

SENTIMENT DETECTION EXAMPLE
17

SENTIMENT DETECTION EXAMPLE
18
• contains a wealth of useful information
on the domain of UK politics
• every UK MP is represented, along with
their constituency and the political party
to which they belong
• [7]

FUTURE THINKING
• identify phrases that may have a temporal orientation, and try to
guess what that orientation is
• temporal_thinking annotation [1]
temporal signals (e.g. simultaneously, as soon as, before, after)
soft signals (e.g. our children, planning, the long term, then-
current, the great war)
verb groups
• degree of future thinking = combined weights of relevant indicators
20

FUTURE THINKING
• Future tense: +0.65
• Past tense: -0.45
• Present perfect: -0.5
• Future timex: +1.0
• Past timex: -1.0
• Pre-2015 date: -1.0
• Pre-election date: -0.15
• Near future date: +0.5
• Far future date: +1.0
• Very far future date: +1.5
• Future temporal signal: +0.4
• Past temporal signal: -1.0
• Future soft indicator: +0.3
• Past soft indicator: +0.3
21

EXPLODING QUERIES
• Count all the sentiment expressions about the theme in tweets
written by party candidates for constituencies in region.
Count all the positive sentiment expressions about the “UK
economy” theme in tweets written by Labour candidates for
constituencies in Greater London.
Count all the negative sentiment expressions about the “UK
economy” theme in tweets written by Labour candidates for
constituencies in Greater London.
22
positive or negative
45 different political themes
7 main UK political parties
12 main
regions

VISUALISATIONS
• Top 10 topics mentioned by Green Party members
23

VISUALISATIONS
• Treemap showing the most frequent terms about climate change
mentioned by the Labour Party
24

VISUALISATIONS
• Dynamic Co-ccourence Matrix
25

VISUALISATIONS
• Choropleth depicting distribution of tweets about the economy
26

CLIMATE CHANGE ENGAGEMENT
• retweets, for self-gain and philanthropy, as a mean of disseminating
information to a wider audience and faster propagate a message
• 64.48% of climate change tweets were actually retweets
• high level of engagement (second in percentage of opinionated
tweets, after Europe)
• original tweets containing a URL are more likely to be retweeted
27

EVALUATION
• critical component: NLP processing
• must extract entities, topics and sentiments correctly
• used tools that had a high performance in previous analytics (like
TwitIE)
• Precision: 85.87%
• Recall: 53.05%
• can be adapted to different tasks (topics or groups of people)
28

CONCLUSIONS
• analysis is not limited to searching for relevant documents that
match a query
• questions can be expressed in a natural, broad way (e.g. „Which
political party talks the most about environmental topics?”, „Which
politician gets the most retweets when she/ he talks about climate
change?”)
• can lead to studies by social scientists, environmentalists and
politicians
• can be used for predictive analytics, "intent mining" - looking at
things like tendency to purchase, to prefer a brand, to switch etc.
• measure how the audience perceives the content and what kind of
emotional response it elicits
• in future, it could be combined with language generation for
creating content, e.g. writing news
29

BIBLIOGRAPHY
1. Diana Maynard, Ian Roberts, Mark A Greenwood, Leon Derczynski, Kalina Bontcheva. 2015.
Political Futures Tracker: Technical report. - Nesta working paper 15/12.
2. M. Nagarajan, K. Gomadam, A. Sheth, A. Ranabahu, R. Mutharaju, and A. Jadhav. Spatio-
temporal-thematic analysis of citizen sensor data: Challenges and experiences. In Web
Information Systems Engineering, pages 539–553, 2009.
3. A. Marcus, M. S. Bernstein, O. Badar, D. R. Karger, S. Madden, and R. C. Miller. TwitInfo:
Aggregating and visualizing microblogs for event exploration. In Proceedings of the 2011
Conference on Human Factors in Computing Systems (CHI), pages 227–236, 2011.
4. H. Cunningham, D. Maynard, K. Bontcheva, and V. Tablan. GATE: an Architecture for
Development of Robust HLT Applications. In Proceedings of the 40th Annual Meeting on
Association for Computational Linguistics, 7–12 July 2002, ACL ’02, pages 168–175, Stroudsburg,
PA, USA, 2002. Association for Computational Linguistics.
5. V. Tablan, K. Bontcheva, I. Roberts, and H. Cunningham. Mímir: an open-source semantic search
framework for interactive information seeking and discovery. Journal of Web Semantics, 2014.
6. K. Bontcheva, L. Derczynski, A. Funk, M. A. Greenwood, D. Maynard, and N. Aswani. TwitIE: An
Open-Source Information Extraction Pipeline for Microblog Text. In Proceedings of the
International Conference on Recent Advances in Natural Language Processing. Association for
Computational Linguistics, 2013.
7. C. Bizer, J. Lehmann, G. Kobilarov, S. Auer, C. Becker, R. Cyganiak, and S. Hellmann. DBpedia – a
crystallization point for the web of data. Journal of Web Semantics: Science, Services and Agents
on the World Wide Web, 7:154–165, 2009.
30

A framework for real time semantic social media analysis

More Related Content

What's hot (13)

Similar to A framework for real time semantic social media analysis (20)

Recently uploaded (20)

A framework for real time semantic social media analysis