This is
the first year that this workshop runs and there are two possible
ways to participate:
Research Papers Submission
The workshop is open to the submission of papers
describing test collections or data sets suitable for early risk
prediction, early risk prediction challenges, tasks and
evaluation metrics or specific early risk detection solutions.
We understand that there are two main classes of early risk
prediction:
- With
multiple actors. We include in this category cases where there
is an external actor or intervening factor that explicitly
causes or stimulates the problem. For instance, sexual
offenders use deliberate tactics to contact vulnerable children
and engage them in sexual abuse. In such cases, early warning
systems need to analyse the interactions between the offender
and the victim and, in particular, the language of both.
Another example of risk provoked by external actions is
terrorist recruitment; there is currently massive online
activity aiming at recruiting young people (particularly,
teenagers) for joining criminal networks
- With
a single actors. We include in this second category cases
where there is no explicit external actor or intervening factor
that causes or stimulates the problem, but instead the risk
comes “exclusively” from the individual. For instance, a
teenager developing depression whose process is not caused or
stimulated by any intervention or action made by another
individual. Of course, there might be multiple personal or
contextual factors that affect (or even cause) this process
and, as a matter of fact, this is usually the case. However, as
it is notfeasible to have access to sources of data associated
to all these external conditions, the only element that can be
analysed is the language of the individual.
The
two classes of risks described above might interact one to each
other. For instance, individuals suffering from major depression
might be more inclined to fall prey to criminal networks. From a
technological perspective, different types of tools are likely
needed to develop early warning systems for these two types of
risks.
Essentially,
we look at early risk prediction as a process of sequential
evidence accumulation where alerts are made when there is enough
evidence about a certain type of risk. For the single actor type
of risk, the pieces of evidence could come from the
chronological sequence of entries written by a tormented subject
in the Social Media. For the multiple actor type of risk, the
pieces of evidence could come from a series of messages
interchanged by an offender and a victim in a chatroom or online
forum.
Notice
that we refer to early risk in a general way and the workshop is
open to contributions along these lines in any possible
application domain.
Important dates (Research Papers):
- Submission deadline: May 26, 2017
- Notification of acceptance:June 16, 2017
- Camera-ready: July 3, 2017
Submission instructions: We solicit papers up to 12 pages in length. The papers must be written in English and should follow the Springer-Verlag LNCS style.
For details see the Springer LNCS Author Instructions. Papers (PDF) must be submitted through EasyChair:(eRisk Workshop Papers).
The proceedings of this workshop will be published in the online CEUR-WS Proceedings and on the conference website.
Programme Committee:
- Leif Azzopardi, University of Glasgow, UK
- Alvaro Barreiro, University of A Coruña, Spain
- Meeyoung Cha, KAIST, South Korea
- Glen Coppersmith, Johns Hopkins University, USA
- Silvia Gabrielli, CREATE-NET, Italy
- Manuel Montes, National Institute of Astrophysics, Optics and Electronics, México
- Josianne Mothe, IRIT, France
- Raffaele Perego, CNR, Italy
- Philip Resnik, University of Maryland, USA
- Paolo Rosso, Universitat Politècnica de València, Spain
Important
Dates
Early Detection of Depression
The second way of participation consists in performing a pilot
task on early risk detection of depression. This is an
exploratory task on early risk detection of depression. The
challenge consists of sequentially processing pieces of evidence
and detect early traces of depression as soon as possible. The
task is mainly concerned about evaluating Text Mining solutions
and, thus, it concentrates on texts written in Social Media.
Texts should be processed in the order they were created. In
this way, systems that effectively perform this task could be
applied to sequentially monitor user interactions in blogs,
social networks, or other types of online media.
The test collection for this pilot task is the collection
described in [Losada
& Crestani 2016]. It is a collection of writings (posts or
comments) from a set of Social Media users. There are two
categories of users, depressed and non- depressed, and, for each
user, the collection contains a sequence of writings (in
chronological order). For each user, his collection of writings
has been divided into 10 chunks. The first chunk contains the
oldest 10% of the messages, the second chunk contains the second
oldest 10%, and so forth.
The
task is organized into two different stages:
- Training
stage. Initially, the teams that participate in this task will
have access to a training stage where we will release the whole
history of writings for a set of training users (we will
provide all chunks of all training users), and we will indicate
what users have explicitly mentioned that they have been
diagnosed with depression. The participants can therefore tune
their systems with the training data.
- Test
stage. The test stage will consist of 10 sequential releases
of data (i.e. done at different dates). The first release will
consist of the 1st chunk of data (oldest writings of all test
users), the second release will consist of the 2nd chunk of
data (second oldest writings of all test users), and so forth.
After each release, the participants will have a few days to
process the data and, before the next release, each
participating system has to choose between two options: a)
emitting a decision on the user (i.e. depressed or
non-depressed), or b) seeing more chunks.
This choice has to be made for each user in the collection. If the system emits a
decision, its decision will be final and it will be evaluated
based on the correctness of the system's decision and the
number of chunks required to make the decision (using a metric
for which the fewer writings required to make the alert, the
better). If the system does not emit a decision then it will
have access to the next chunk of data, that is it will have
access to more writings, but it will have a penalty for a
“later emission”).
Evaluation: The evaluation will take into account not only the
correctness of the system's output (i.e. whether or not the user
is depressed) but also the delay taken to emit its decision. To
meet this aim, we will consider the ERDE metric proposed in [Losada
& Crestani 2016].
Being
a pilot task, we expect it to be useful for instigating
discussion on how to create evaluation laboratories for early
risk prediction: proper size of the data, adequate early risk
evaluation metrics, alternative ways to formulate early
detection tasks, other possible application domains, etc.
Nov 30th, 2016: The training data has been released!. More info: here
