Social media enable fast and widespread dissemination of
information that can be exploited to efectively spread disinformation
by bad actors [
Several computational solutions have been proposed for the
identification of coordinated campaign on a single platform [
First, existing approaches focus on a single sources, such as Twitter or Reddit. Unfortunately, misinformation campaigns span multiple platforms and there is a recognized need to jointly analyze the difusion of content across diferent sources, such as social networks, online forums (e.g., Reddit), and traditional news outlets (e.g., comments in reputable sites).
Second, the content difusion graphs that are currently
generated from social network APIs are limited in quality. For example,
only the information about content re-posting (e.g., re-tweets) of
a user is directly provided. But information is disseminated also
by manipulating the original content to add bias, “evidence”, or
propaganda material. Moreover, fine granularity of the re-posting
is not available, with the recognized problem of star-efect for
re-tweets that can heavily degrade the quality of the network
model [
Consider the example in Figure 1 that shows the coordinated sharing of the same initial piece of content (say, a textual news) by three users over diferent platforms. With the current infrastructure and APIs, a journalist or a fact-checker willing to study the difusion network would look at each network in isolation. S/he would be able to follow content across users (nodes) only when they re-post explicitly (full edges across nodes).
In this example, the information would not be enough for the
early identification of the coordinated campaign started by the
three users. Looking at only one source with limited information
does not enable the analytics, neither in terms of scope nor
evidence, that we need to identify and understand how false and
biased content is used in online campaigns [
To overcome this limitation, recent approaches explore
evidence across users and platforms, such as coordinated link
sharing [
In fact, to tackle the first challenge, difusion networks should be heterogeneous, covering multiple platforms, with the ability to recognize the same content and the same users across services. In Figure 1, users that refer to the same real world person are annotated with nodes having the same color. Also, to handle the second challenge, the edges should be typed with fine-grained metadata that model diferent actions in the spread of the content.
We believe that data here plays a role as important as the
algorithms used for the analysis and therefore more attention is
deserved to the problem of creating such richer difusion networks.
Their creation can lead to better identification of coordinated
eforts [
The goal is therefor to develop methods for modeling and creating the rich difusion networks from the existing platform APIs. The resulting networks can be exploited to assist users, such as fact-checkers and journalists, in (1) monitoring sources at scale and recognize misleading information (in terms of false or biased textual content) on social networks and forum websites; (2) tracking the spread and difusion of the content in terms of time and actors; (3) generating visualizations that support the fight against misinformation and related literacy eforts.
This network generation is indeed challenging, as the desired metadata is not available and hard to profile automatically in an accurate way. We discuss next two research directions that we identify as critical to tackle these challenges.
The goal is to develop methods for the automatic modeling of content manipulation and difusion across time and diferent media sources, such as social networks, forums, and news outlets.
Not only we want the difusion graph for a given content to be across sources and very well described in terms of information, we also want it (i) to preserve precisely the provenance of the data (who created and shared, how and when) and (ii) to be as much as possible automatic in its creation, both to handle the Web scale and to not put additional burden on the users. There are therefore several challenges that we need to overcome • diferent sources do not contain any readily available information to connect users/content across networks and the automatic matching is a dificult task in both cases; • labeling the content in terms of being false, manipulated, and biased requires deep understanding of the language and of the reference and background information; • Web scale implies massive ingestion from heterogeneous data sources, but we would prefer tools that can be used by end users on their machines for confidentiality; • support for diferent languages as we aim at helping users across diferent countries.
Given the challenges above, a natural first line of work is to conduct data integration research to generate a unified representation from heterogeneous, non-aligned sources. A second line of work is to deploy natural language processing (NLP) techniques to profile the content and enrich the graph with typed nodes and edges.
Data integration. In the first line of research, the aim is the
online creation of a dissemination network for a given textual
content. Given a textual article, for example, the first task is the
identification of its citations and appearance across sources
(online articles, boards in forums, social posts) and time. This is not
trivial as one requirement is to go beyond the identification of
content by links, which act as unique identifiers. For this goal,
one promising direction is to exploit text-matching literature [
Example 1. Consider a textual article about a new vaccine
that circulates on social platform 1. Existing APIs allow the
tracing of the difusion of the specific content on platform 1 across
users 11, . . . , 1 , but the same content may be circulating in a
diferent form ( ′ ) and across a diferent platform 2 by users
12, . . . , 2 . We aim at identifying that the two posts refer to the
same content ( = ′ ) and at matching the subset of users that
are sharing the article across the two networks (e.g., 31 = 2).
6
Metadata from text. Existing NLP tools should be extended and
integrated to characterize the nature of the interactions across
actors w.r.t. the specific content. This can lead to labelling the
edges in the graph with information (metadata) about the
interaction between the nodes (actors). Possible metadata for such edges
include the nature of the connection between two users, if it is
based on friendship or topic afinity [
Example 2. Consider again the article example from Example 1. When it is shared across users, some of them introduce incorrect statistics about its impact (“it works only for young people”), or facts that are not supported by any source (“it will cost 100$ per dose”). We aim to enrich the network edges by recognizing how the content goes from its form to a new form ∗ when it is shared by a certain user .
We believe that an efective solution to the problem of
creating difusion networks for textual content across heterogeneous
sources would enable better disinformation campaigns detection.
The resulting graph with typed nodes (persons, organizations)
and typed relationships (copy or manipulation in terms of
content or form) can be then analyzed with existing methods such
as clustering and geometric deep learning [