Social media is a common place for people to post and share
digital reflections of their life events, including major events
such as getting married, having children, graduating, etc.
Although the creation of such posts is straightforward, the
identification of events on online media remains a challenge.
Much research in recent years focused on extracting major
events from Twitter, such as earthquakes, storms, and
floods. This paper however, targets the automatic detection
of personal life events, focusing on five events that psychologists
found to be the most prominent in people lives. We
define a variety of features (user, content, semantic and interaction)
to capture the characteristics of those life events
and present the results of several classification methods to
automatically identify these events in Twitter. Our proposed
classification methods obtain results between 0.84 and
0.92 F1-measure for the different types of life events. A novel
contribution of this work also lies in a new corpus of tweets,
which has been annotated by using crowdsourcing and that
constitutes, to the best of our knowledge, the first publicly
available dataset for the automatic identification of personal
life events from Twitter
Provide a quick overview of what I’m talking about There are three sections:
An overview of the background for why we’re doing this, and some terminology What we actually did with the experiment A discussion about our results and what we learnt from it.
Suggest dementia research maybe?
Anyone familiar with narratology? R.Q.1 is effectively the creation of a fabula R.Q.2 is effectively the story generation
Now before we go any further, it’d be useful to discuss what we mean by a life event
Chat about Autobiographical memory for a minute
As a way to kickstart our research, we focus on simple event classifiers These initially are to be used as a filter to kick start off our system, helping to filter posts that are more likely to be about life events We can then look at clustering similar posts together to help performance of our narrative generation component
We admit this isn’t the best, but it is definitely the easiest to obtain data from
Mention CrowdFlower quizzes and how they test users before they can do your job
Talk from Chris Welty earlier highlighted some of our issues with annotations from Twitter.
Most of our event specific classifiers had an F1 score of about 0.9 (falling in love was 0.841) Our binary classifier performed at 0.753 In almost all cases, content features (unigrams) ran far superior to other feature sets. This is likely due to bias of a keyword search, and the short nature of a tweet. Occasionally semantic features helped boost the performance slightly. Other feature sets like interaction and user did not perform so well. Semantic features
Talk by Chris earlier highlighted some of our issues with annotations from Twitter. We did sample some of our annotations, and found a number had been mislabled. For example news stories were included in out annotated dataset.
Identifying Prominent Life Events on Twitter - K-Cap 2015
Life Events on Twitter
SPEAKER – TOM DICKINSON
AUTHORS – TOM DICKINSON (OU - KMI), MIRIAM FERNANDEZ (OU - KMI), LISA
THOMAS (NORTHUMBRIA), PAUL MULHOLLAND (OU - KMI), PAM BRIGGS
(NORTHUMBRIA) AND HARITH ALANI (OU - KMI)
What we did
Why are we doing this?
As content creators, we post a lot of stuff on social media
This content can range from silly cats, to important life events that have happened to us
However, as users, we effectively lose access to this information about ourselves and forget
By being able to mine and present this data to users, we can look at giving users a tool to aid in
self reflection over their own online digital presence
So what are we doing?
As part of the Reellives project, we are looking at making short “Reels” from a users social media
These reels are intended as mini documentaries about a users life on social media
This presents two main problems for us to solve:
◦ R.Q. 1) How can we extract meaningful events about ourselves from our social media data?
◦ R.Q. 2) How can we present these events in a cohesive narrative?
R.Q. 1 is being tackled by KMI, where we are looking at event extraction.
R.Q. 2 is being tackled by Edinburgh, who are looking at taking our output, as their input, to
So what are we doing?
What is a life event?
There already exists a large body of research of event detection on social media.
However, not much has been done on focusing on life or personal events.
Semantically, they are no different:
◦ Both types will have a time and a location
◦ An action occurs
◦ The event is experienced by one or more agents
◦ With general events we care more about the broader social and political significance
◦ With life events we care more about the personal significance
What is a Life Event?
We can also get some intuition for life events from Autobiographical
Autobiographical memory is type of memory system that deals with
specific events that happened to us
◦ This is opposed to semantic memory which is our knowledge of things
It can be modelled with three separate layers
◦ Lifetime periods
◦ When I was at school I had my first kiss
◦ General Events
◦ I got married
◦ Event-Specific Knowledge
◦ My tie was red at the wedding
In our work, we can consider the event-specific knowledge to be
reflected in social media posts
Types of life events
To start off our research, we looked at identifying a finite number of life events.
The types of life events we chose are inspired by work done in Autobiographical Memory
◦ S. M. Janssen and D. C. Rubin. Age effects in cultural life scripts. Applied Cognitive Psychology
Their research showed a common consensus, amongst different age groups, of 48 life events that
would happen to a fictional child over the course of their life.
From this study, we selected 5 of the top events mentioned in a paper
◦ Getting Married
◦ Having Children
◦ Starting School
◦ A Parents Death
◦ First Love
We also look at combining all positive “about an event” into a training set to create a more general “Is
this about an event” classifier.
What we did – Data Collection
We chose Twitter due to ease of use for extracting large datasets.
Our selection methodology was based around a simple keyword search, where we considered
the root concepts for each of our events, and enhanced with synonyms from WordNet.
We extracted Tweets from Twitter’s front-end search, as opposed to their API
◦ This is due to their API having a 7 day limit
◦ Twitter now indexes every tweet, making it available to scrape from their front-end search application
Additional details were extracted for each Tweet, using their Lookup API with the extracted
What we did - Annotations
To annotate our dataset, we turned to CrowdFlower
To start with, we ran several small trials of annotation exercises on CrowdFlower to make sure
our questions were satisfactory
We initially had 7 questions:
◦ Is this tweet about Getting Married?
◦ Is this tweet about an event?
◦ Was the tweet before, during, or after the event?
◦ Is the author of the tweet experiencing the event?
◦ Is anyone else experiencing the event with the author?
◦ Is anyone else named in the tweet experiencing the event?
◦ Did the event happen where it was tweeted?
This did not prove too popular as we had large number of quiz failures
What we did - Annotations
Obvious failure for this initial test run were too many questions and possible subjectivity for our
given definition of an event.
After another trial, we finally settled on only asking two questions:
◦ Q1 - Is this tweet related to a particular topic theme? (Topic theme is the cluster we extracted from)
◦ Q2 - Is this tweet about an important life event?
We also provided users a list of the 46 life events that Jansen and Rubin identified, as a way to
get them to understand what we were after.
This ran much better, and our final agreement ratings were 89.5% and 87.17% respectively
What we did – Feature Sets
Our feature sets were divided into several groups:
◦ User features
◦ H1) Certain types of users may be more prone to share life events in Twitter
◦ Content Features
◦ H2) Posts written in a certain way may be related to life events
◦ Semantic Features
◦ H3) Posts about life events might be semantically associated with certain entities or concepts
◦ Interaction Features
◦ H4) Users who do not normally talk with the poster, might start interacting for certain types of life events
What we did - Classifiers
We ended up just testing two classifiers, as other work had already tested a number of different
classifiers on similar datasets:
◦ Naïve Bayes
We did try SVM’s as well, but due to poor performance, omitted it from our results.
To evaluate we used 10-fold cross validation, reporting standard classification performance
measures of Precision, Recall, and F1 scores.
Why the dominance of content features?
Unigrams outstripped performance of other feature sets.
This is similar to other similar papers, and slightly disappointing.
While the classifiers were biased towards the keywords chosen, it is disappointing other feature
sets did not perform well.
In the case of interaction features this might be because:
◦ We were limited in what types of interaction features we could obtain, due to the limits of Twitters API
◦ The dataset might have been annotated incorrectly
◦ For example, stories of other people are annotated, rather than people declaring an event about themselves
◦ Due to the nature of Twitter and it’s followers, interaction features might just not be a good
◦ Sites like Facebook though, which tends to be private, might have better performance in this area
Choice of targeting specific life events
Targeting only five specific life events, dilutes what we can actually extract from social media
Our binary classifier worked alright, but:
◦ Due to dependency on unigrams, it will probably not perform very well outside of these 5 events
This is no silver bullet for solving our research question
Collecting the dataset
The collected dataset was biased to certain words due to a keyword search
A better way to collect these datasets would be to randomly sample twitter profiles, and
annotate their timelines
However, it is likely that only a small number of tweets are actually about these types of events
in a users timeline
To achieve a decent training set, we would need to annotate lots of tweets which is very costly
Twitter and the annotation process
Using CrowdFlower is a great way to gain lots of annotations fast
However, with Twitter data we think the annotation is flawed for these types of questions
Lack of context
◦ Is a 140 character max text string enough context to annotate these types of events
◦ Example: Is “MadJacks Forever Memories” about getting married?
◦ Madjacks is a wedding venue in Las Vegas, so this might be?
First vs third party annotation
◦ While lack of context for a third party is an issue, if the owner of the tweet annotated it, would we get
Extracting useful interaction features is difficult
◦ There is no API to get conversations for tweets. Mining this manually is possible, but annoying.
◦ You can’t get access to which users have favourited a tweet
Facebook would be better…
…but it has heavy privacy controls to access user data
While this is great for users, it’s annoying for researchers
Retrieving content from Facebook all needs to be done within an application
◦ These days, a User ID is hashed with your application ID
◦ If you have a standard user ID, you can’t access the Facebook graph API to retrieve information about it
Asking people to just give us their Facebook data with a single sign on approach isn’t the best
◦ Users are reluctant to just give researchers their private data
◦ What do they get out of it? (besides the results of the research)
Is Instagram the middle ground?
Like Twitter, there are a lot of open Instagram accounts
◦ Sites like websta.me index large numbers of users and offer tag based search
Like Twitter, it is (currently) easy to extract Instagram data
◦ While the API, like Twitter, is limited, it is possible to extract full user profiles
◦ Instagram works with a REST based architecture, returning user posts in JSON feeds that can be
paginated allowing full extraction of posts
◦ Using the API each post can be augmented with additional information not available in the media
While we think of Instagram only being photos, most photos have short captions similar to
◦ Comments can also provide semantic context
We are currently looking at collecting Instagram and Facebook data for future experiments
◦ Facebook data is being collected with a trivial app that users can use
Unsupervised life event detection
◦ As opposed to targeting specific events, being able to extract any type would be of more value
◦ Currently we are looking at knowledge based approaches using ConceptNet to achieve this
Graph Classification of Posts
◦ So far we have employed fairly flat vectors when considering feature sets
◦ As opposed to this, an alternative is to treat posts as graphs, looking at relationships within semantic
(ConceptNet, DBpedia etc), interactions, and dependency parsing
◦ Graph frequent pattern mining might identify new feature sets that we can look at using