The ODYCCEUS project is part of European Union's Horizon 2020 research and innovation programme. The project seeks conceptual breakthroughs in Global Systems Science, including a fine-grained representation of cultural conflicts based on conceptual spaces and sophisticated text analysis, extensions of game theory to handle games with both divergent interests and divergent mindsets, and new models of alignment and polarization dynamics.
The conference will cover all topics that we are working on in the project ranging from game theory, over opinion dynamics modeling up to natural language processing and opinion mapping. There will be sessions on projective game theory, frames alignment and polarization, conceptual spaces for political communication or precision language analysis for argument mining.
Financial support: The conference is part of a project that has received funding from the European Union's Horizon 2020 research and innovation programme under grant agreement No. 732942.
The Odycceus project tracks opinion dynamics and the emergence of conflict through social media and text. It is developing a platform for this purpose, Penelope, that allows a social science researcher to combine different textual resources, for example twitter data or newspaper commentaries, and apply various methods of analysis obtained through crowdsourcing. This talk first gives an overview of the platform and which components have been added so far. A crucial step in the pipeline is textual analysis. But out of necessity, most text analysis today uses quite shallow methods, ignoring most of the grammar and the semantics of what is written. Here I raise the issue in how far we can and should go deeper to capture the meaning of text and how this would improve social analysis.
Users to promote their favorite narratives and cluster themselves in groups where they accept information adhering their narrative (even if contains false claims) and resist dissenting information (i.e. the inefficacy of fact-checking). We extended the focus on the consumption of news on Facebook and we found evidence of the effect of disintermediation.
Indeed, according analyzing 376 million Facebook users’ interactions with more than 900 news outlets in six years we find that that selective exposure dominates news consumption on line creating a segregated non-communicating environment. The main problem behind fake news seems to be polarization of narratives.
This presentation offers an overview of how niche online communities use specialized fora to create and spread alternative styles of thought and networks of discursive references. We look at how their tactics, their mythos and their peculiar vernacular theories may be seen to have an influence on certain areas within mainstream social media, in particular through the spread of “alt-right” internet memes.
While new media “meme theorists” have in the past sought to conceptualize the use political memes in terms of a protest against liberal consensus politics — where consensus is seen as “the crystallization of relations of power” (Mouffe 1997, p.28) — we find repeated expressions of “absolute enmity” (Schmitt 1962) that would seem to precluded the possibility for such productive disagreement.
Emerging then from partially indexed regions of the web, that we call the “deep vernacular web”, our analysis furthermore considers aspects of the relationships between these online communities and the technical affordances of their platforms — where for example, in the case of the anonymous 4chan fora, memes can be understood as “credibility mechanisms” for demonstrating one’s in-group status in the absence of any persistent repetitional markers.
In a functioning democracy, it is crucial that the political sphere reacts to changes in societal prerequisites. Our analysis explores the impact of external events like the global financial crisis in 2008 on the Swiss parliamentary issue agenda. We analyze the verbatim transcripts of parliamentary debates and the issue orientation of parliamentary business from 1999 until 2018 by means of machine learning text analysis and other quantitative and qualitative methods to track the salient issues over time. The salience of the issues then gets compared to real world events to uncover correlations of the issue agenda and external events.
The iLCM project pursues the development of an integrated research environment for the analysis of structured and unstructured data ina “Software as a Service” architecture (SaaS). The research environment addresses requirements for the quantitative evaluation of large amounts of qualitative data with text mining methods as well as requirements for the reproducibility of data-driven research designsin the social sciences. For this, the iLCM research environment comprises two central components. First, the Leipzig Corpus Miner(LCM), a decentralized SaaS application for the analysis of large amounts of news texts developed in a previous Digital Humanities project. Second, the text mining tools implemented in the LCM are extended by an “Open Research Computing” (ORC) environmentfor executable script documents, so-called “notebooks”. This novel integration allows to combine generic, high-performance methods to process large amounts of unstructured text data and with individual program scripts to address specific research requirements in computational social science and digital humanities.
We discuss a novel method for an empirical analysis of political programs. Applying a set of binary text classifiers based on convolutional neural networks, we label statements in the political programs of the parties in Germany and USA. We discuss a long-term evolution that could be seen in the programs of major political parties and try to interpret the trends and relative positions of parties.
When considering the last five decades history of disciplinary relationships between social sciences and computation, two major steps can be identified. In a first stage, until the 1990s, the perceived necessity of quantifying for making analysis reproducible and formalizing for sharing knowledge was so imperative that the “computational turn” was unavoidable. In total submission most social scientists became dependent on computer scientists as users of predefined software for statistical analysis, automatic mapping or Geographical Information Systems. Very little of social science could be injected in computing, the dedication of some computer scientist for handling specific problems of social science was very rare. Even our first experiments of model building for simulating the evolution of social complex systems were made within an obvious asymmetry of information and power of decision. It is only in the recent years that a total change in the status of interdisciplinary interaction could occur, shifting towards real win-win interdisciplinary cooperation. In the academic world a large coordinated and well funded program coupled with an intelligent use of genetic algorithms and distributed computing are apparently necessary ingredients for success, while in the realm of business and public uses the injection of selected knowledge from social sciences in the algorithms becomes more and more a part of the intelligence injected in the algorithmic applications. The best practices in new data analytics will necessarily proceed from such voluntary collaboration.
The authors screen a set of French and Russian daily newspapers with different political and thematic orientations from 2005 to 2017 revealing the changing perceptions of the crisis arround Georgian breakaway regions (Abkhazia and South Ossetia) depending on political situation and the interests of the main actors. They apply a combination of qualitative and quantitaive methods ranging from discourse analysis to statistical modeling. First, the media attention to the crisis is shown. Second, they demonstrate the process of geopolitical scaling in their interpretative frames. Third, they focus on the evolution of narratives through time and the degree of national consensus in the vision of conflicts’ reasons and perspectives. In conclusion, they discuss the potential of their method for further studies of the dynamics of public opinion on geopolitical conflicts and its volatility.
The decline of Europe is a main topic in political and economic Geography since the beginning of the 20th century. But is Europe declining in minds and how fast is it declining? This communication will lay out a study of opinion dynamics through two analysis of two different objects depicted in the representations of Turkish undergraduate students: “Europe” as an identifiable and coherent entity and the European countries. The presentation will illustrate that these two ways of representing “Europe” are not changing in the same way. They are also not sensible to the same events or processes (global economic crisis, Arab Springs, current wars). These results could be an illustration of the difference between the branding processes of a global area and of a country in a specific population of study.
Path-constrained Random Walks to Reveal Specific Relations in GraphsHong-Lan BottermanLaboratoire d'Informatique de Paris 6, FrancePeople are subject to social influence as well as mass media influence. Individuals are therefore confronted to different sources of information which can have an impact on the topics they discuss. In this poster, such interactions are modelled by networks with different types of entities and links which are well adapted to reflect the different meanings of interactions. For instance, the question is whether the “individuals-topics” interactions in social media are random events or whether they are related to other existing interactions between the nodes. We propose a path-constrained random walk to perform statistical regression on networks and to thus unveil the full network topology.Cooperative vs. Competitive Spreading DynamicsFakhteh GhanbarnejadTU Berlin, GermanyWe show non-trivial dynamical effects of cooperation and competition in an ecological framework [1-5]; we study two (or more) spreading dynamics competing with each other for host resources but they can choose cooperation as successful strategy too. We treat dynamics in a homogeneously mixed population by means of mean-field theory and stability analysis and also on complex transmission networks. We study the impact of cooperation on the outcome of the global dynamics, which can be quantified in terms of dominance of one competing agent or the co-circulation of both of them. We show that the presence of cooperation can alter the outcome of competition as it may favor the more cooperative dynamics over the more infectious one.
L. Chen, F Ghanbarnejad, W. Cai, P. Grassberger, EPL 104, 5 (2013).
W. Cai, L. Chen, F. Ghanbarnejad, P. Grassberger, Nature Physics 11, 936 (2015).
P. Grassberger, L. Chen, F. Ghanbarnejad, W. Cai, Physical Review E 93, 042316 (2016).
J. P. Rodríguez, F. Ghanbarnejad, V. M. Eguíluz, Frontiers in Physics, V 5, P 46 (2017).
L. Chen, F. Ghanbarnejad, D. Brockmann, New J. Phys. 19, 103041(2017).
Local Discourse Norms and Dual Process Decisions in an Opinion Dynamics ModelMalte HeckelenUniversity of Stuttgart, GermanyDepending on paremeterization, basic Bounded Confidence models lead either to complete consensus or complete fragmentation of the opinion space. Noise terms can be introduced to achieve more realistic opinion dynamics. While this produces intuitively plausible results, it isn't easily mapped to current theories of social action and discourse. The detailed dual process models from Social Psychological Persuasion and Social Influence Studies in contrast use momentary energy and discourse investment to explain why an individual would evaluate messages systematically based on content in one situation and heuristically via meta-attributes (e.g. sender attributes) in another. This project integrates Social Psychological dual process models of information processing with the Model of Sociological Explanation from Sociology. The latter is a dual process model of social action, with the two modes corresponding to "slow", rational planning of actions and "fast", automatic selection of scripts based on familiarity with a situation. The situation classification uncertainty of inputs as well as momentary energy is taken into account to model the selection of one of the two modes in the sense of a latent utility function. This allows for a very general notion of norm generation. A layered dual process model of discourse decisions is constructed: In the inner layer, individuals are interested in maximizing opinion similarity in their neighborhood and minimizing energy loss, while also maximizing self-cohesion and self-neighborhood-cohesion. Decisions are made regarding the sending of "optimized" or "unoptimized" opinions as messages to maximize adoption probability as well as the reception mode (systematic or heuristc). The outer layer consists of the Model of Sociological Explanation: Individuals try to classify the discourse situation based on the available inputs and, if the uncertainty is too high and/or their energy is low, will choose to make all decisions in the rational mode as opposed to the automatic selection of situation-associated scripts concerning the inner layer. The project aims to implement this full model for the Hegselmann-Krause model in a step-wise fashion to ascertain minimal conditions for lasting, yet systematic dynamics. The goals for this project are then threefold: a) Determine the aggregation dynamics of dual process model assumptions from Persuasion Studies which were previously studied for small-groups, b) Link Opinion Dynamics models more closely with contemporary theory, while keeping complexity minimal and c) shift the focus of analysis toward actions, motivations, their local clustering and thus toward a more sociological research question. The account up to the inner layer is implemented for the Hegselmann-Krause model and shows plausible dynamics in exploratory runs. A first sensitivity analysis will be conducted and results will be presented on the poster.A simple adaptive network model for coup d'etatsLeonhard HorstmeyerMedical University Vienna, AustriaThe network of alliances and oppositions determine to a large extend the landscape of the political system. The instability of the system crucially hinges on the ties amongst the political actors and their temporal readjustment. Therefore these systems are amenable to a description by adaptive co-evolving networks, such as the autocatalytic network model by Jain and Krishna. In that model a network of species with catalytic ties is evolving via the deletion species that do not profit from the network and via the introduction of new species tying into the existing network. It exhibits the creation of stable autocatalytic subsystems as well as the growth of latent subsystems that can lead to partial or full collapse of catalytic subsystems. We propose an interpretation of this model in terms of political actors and their alliances and suggest that political systems may suffer similar collapse mechanisms during coup d'etats or revolutions. We investigate collapses and early-warning signs in this model and show how impending collapses can be detected.The Outlier Explorer: A Web Interface for the Exploration of Multidimensional DatasetsRobin Lamarche-PerrinCNRS / UPMC, FranceThis demonstration shows the current state of our Outlier Explorer, a Web service dedicated to the study of multidimensional datasets and for the detection of statistical outliers within. It is mainly a tool for data exploration allowing to have a first glance at it and to formulate research hypotheses to be later tested. This component takes as input a list of numeric observations described according to several categorical dimensions. For example, in the case of Twitter data, it can be the number of tweets (numeric observation) that have been published by a given user (first dimension) about a given topic (second dimension) at a given date (third dimension). Statistical outliers are then identified by first selecting some dimensions of interest, that is by subsetting or by aggregating the input dimensions. If needed, observations can also be normalised according to the marginal values along the selected dimensions, thus comparing the observed value to an expected value obtained by the uniform redistribution of the selected marginal values. Different statistical tests can then be chosen to measure the deviation between the observed and the expected values. The component finally returns a list of positive outliers, that is observations that are significantly higher than expected. This Web interface is hosted by the Huma-Num facility for digital humanities. Sources are open and available on GitHub.A Module in Penelope for Graph CompressionLéonard PanichiSorbonne Université, FranceIn this demonstration, we introduce a module for the Penelope API by the mean of a few simple examples. The module is designed to compress multidimensional graph-like data. The compression of data is, in our case, the construction of clusters of similar datapoints within a rectangular subspace of the dataset. We will use two datasets to explain these ideas and what can be expected from the data compression framework: the “Karate club” classical toy-graph and a corpus of geo-media data containing newspaper articles about international events. The demonstration will go from the interpretation of a single cluster to the multiscale compression of the datasets. We can also provide quick technical insights about how the algorithm works and is implemented.A Multidimensional Method to Find Irregular Patterns in Interaction DataAudrey WilmetLIP6 - Sorbonne Université, FranceTwitter is now an integral part of means of communication used by political leaders to disseminate information to the public. A politician may use it sporadically to merely broadcast to his followers or on the contrary employ it regularly and tweet at strategic moments. Likewise, their followers may be occasional spreaders or more involved online activists primarily retweeting a particular political figure. In this poster, we propose a multidimensional analysis method to describe structural and temporal relational patterns in a retweet network: the first step is to identify which type of irregular pattern one wants to detect, the second is to measure its activity, finally, the third step is to compare this activity to an expected activity so as to determine whether it is abnormal or not. We show that varying the scale of observation in each of these steps leads to interesting results such as the highlighting of political leaders or media events.
We will show how to qualify and quantify the activity of political communities in a multi-polar political environment as well as their temporal evolution through the study of their digital traces.
From the analysis of a corpora of 60 million Twitter exchanges between more than 2.4 million users who interacted with political figures during the 2017 French presidential elections, we characterize the socio-semantic networks of the French political environment, as well as their development over a period of eleven months preceding the election.
This reconstruction provides unprecedented insight into the opinion dynamics and the reconfigurations of political communities, giving access to an intermediate level of resolution, between traditional sociological surveys and large statistical studies (such as those conducted by national or international organizations).
We will show how this type of reconstruction, that are intended to constitute input for social systems modeling, can provide insights into some important societal issues.
Website : http://chavalarias.com & http://politoscope.org
We present a graph- and information-theoretical approach to the problem of the measurement of the diversity in the context of consumption, sharing, and recommendation of news media. We use these proposed metrics to analyze the diversity of news sources shared on Twitter during the 2017 French Presidential election campaign. We then explore the diversity of recommendations produced by different popular and ubiquitous recommendation algorithms when used to propose news sources to online users.
Social media platforms, comment boards, and online market places have created unprecedented potential for social influence and resulting opinion dynamics, which sparked a debate about the role of online media in the polarization of political opinions in many western countries. We propose an encompassing model that captures competing micro-level theories of social influence. Conducting an online lab-in-the-field experiment, we observe that individual opinions shift linearly towards the average of others' opinions. From this finding, we predict the macro-level opinion dynamics resulting from social influence. We test our predictions using data from another lab-in-the-field experiment and find that opinion polarization decreases in the presence of social influence. We corroborate these findings with large-scale field data.
In this talk, I will present two new models that have been developed in the ODYCCEUS project (www.odycceus.eu). The first model is based on the idea that agents evaluate alternative views on the basis of the social feedback obtained on expressing them. A high support of the favored and therefore expressed opinion in the social environment, is treated as a positive social feedback which reinforces the value associated to this opinion. In connected networks of sufficiently high modularity, different groups of agents can form strong convictions of competing opinions. Linking the social feedback process to standard equilibrium concepts allows to analytically characterize sufficient conditions for the stability of bi-polarization. The second model is based on a more complex representation of opinions and borrows ideas from psychological research on attitude structure and persuasive argument theory. In the model, political issues are represented as (partially overlapping) sets of arguments and these arguments are exchanged in the interaction process. If agents preferentially interact with other agents that hold similar attitudes, this gives rise to polarization and correlations of attitudes towards multiple political issues. Such correlations are usually exploited in the inference of political spaces.
Disentangling and understanding the main statements from an argumentative piece of text is an endeavour that is still far removed from the abilities of state-of-the-art AI systems. In fact, it even remains a hurdle for many undergraduate students, therefore resorting to reading comprehension courses. Today’s off-the-shelf NLP methods rely on small-window statistical co-occurrence counts or local information-theoretic measures, and remain therefore incapable of capturing an author’s claims and arguments. We refer to such existing techniques as distant reading tools, and contrast them to so-called close reading tools that go beyond statistical distributions in the training corpus but instead make use of the syntactic structure and the semantic frames underlying a specific sentence or paragraph.
In our talk, we propose a close reading method that makes use of Fluid Construction Grammar, a parsing and production engine that chunks together form-meaning mappings (i.e. constructions) into rich feature structures. It uses the exact same engine and constructions for building up an interpretation of a sentence as well as for composing an utterance based on a conceptualized meaning. We specifically show how such an engine can be used with a set of English constructions that are designed to extract argumentation frames from sentences (or paragraphs), through the theory of Frame Semantics. For instance, the Causation frame, which has frame elements such as an Actor and an Affected, could be detected by our system in sentences such as “Oxygen levels in oceans have fallen 2% in 50 years due to climate change.” (The Guardian, 20 Feb. 2017). Once these frames have been retrieved, and their frame elements have been filled in by pieces of the input text, there are two further steps that can be taken. In the first one, which we are currently exploring, the populated frames are returned to the human reader and can be used as a reading aid to grasp the main arguments and their relationship from a text. In a future second step, the AI system itself should be able to interpret the information present in the frames, based on their instantiation in Linked Data.
The aim of the ADD-up project is to build a computerized support system for face-to-face public deliberations. The system is intended to run in parallel to a debate, taking as input a live stenographic feed and processing it utterance by utterance based on a statistical model of deliberative communication. It then provides participants with an augmented view of the ongoing debate via large analytical displays. To this end, the ADD-up project combines recent advances in computational linguistics and visual analytics. In this talk, we focus on the statistical approach for the automatic detection of argument structures, and on the perspectives for delivering effective visual interventions based on those structures.
Our opinions, which things we like or dislike, depend on the opinions of those around us. Nowadays, we are influenced by the opinions of online strangers, expressed in comments and ratings on online platforms. Here, we perform novel "academic A/B testing" experiments with over 2,500 participants to measure the extent of that influence. In our experiments, the participants watch and evaluate videos on mirror proxies of YouTube and Vimeo. We control the comments and ratings that are shown underneath each of these videos. Our study shows that from 5% up to 40% of subjects adopt the majority opinion of strangers expressed in the comments. Using Bayes' theorem, we derive a flexible and interpretable family of models of social influence, in which each individual forms opinions stochastically following a logit model. The variants of our mixture model that maximize Akaike information criterion represent two sub-populations, i.e., non-influenceable and influenceable individuals. The prior opinions of the non-influenceable individuals are strongly correlated with the external opinions and have low standard error, whereas the prior opinions of influenceable individuals have high standard error and become correlated with the external opinions due to social influence. Our findings suggest that opinions are random variables with varying standard deviations, which are updated via Bayes' rule. Based on these findings, we discuss how to hinder opinion manipulation in the online realm. At the end of my talk, I introduce our project entitled "Current Affairs 2.0: Agenda Setting in the European Union" (http://www.euagendas.org).
Social media data as source for empirical studies have recently come under renewed scrutiny, given the widespread deletion of Russian disinformation pages by Facebook as well as the suspension of Alt Right accounts by Twitter. Missing data is one issue, compounded by the fact that the ‘archives’ (CrowdTangle for Facebook and Gnip for Twitter) are also owned by the companies. Previously questions revolved around the extent to which corporate data collected for one purpose (e.g., advertising) could be employed by social science for another (e.g., political engagement). Social media data also could be said to be far from ‘good data’, since the platforms not only change and introduce new data fields (‘reactions’ on Facebook), but also increasingly narrow what’s available to researchers for privacy reasons. Profound ethical issues were also put on display recently during the Cambridge Analytica scandal, as science became implicated in the subsequent ‘locking down' of social media data by the corporations. How to approach social media data these days? The session is an opportunity to debate the future of research with social media.
In this talk, I will describe the design and the general procedure of the first experimental study, which is carried out within the ODYCCEUS project. The theoretical starting point of this study is the first opinion dynamics model developed and presented by Sven Banisch. Its basic idea is that agents evaluate alternative views on the basis of the social feedback obtained on expressing them. We expect that positive social feedback, that is strong support by the social environment on a persons expressed opinion, will reinforce the value associated to this opinion. In contrast, negative feedback will reduce the likelihood of re-expressing that opinion in that social environment.
To provide first empirical insights in this process, we will conduct a two-staged study in the summer of 2018. During the first stage participants will take part in a standardized online survey in which their attitudes towards the issue of speciesism as well as some basic political attitudes and fundamental personality traits will be collected. In the following second stage of the study, the same participants will be invited to the laboratory for a group discussion on one particular aspect of the general issue they were already questioned about. The discussions will be performed using z-Tree (Zurich Toolbox for Readymade Economic Experiments) on separated computer workstations. Afterwards, the participants will be asked to take part in another online survey, in which their attitudes towards the speciesism issue will be measured again. Our aim is to observe how positive and negative social feedback influences public and private opinions.
Many models of social interaction on a network are based on concepts from Statistical Physics, such as percolation and spin kinetics under pairwise interactions. Sociophysics models thus inherit the hardness of analyzing the dynamics on one specific (quenched) network in exact terms. With frustration in spin interactions, for instance, the computation of ground states is NP-hard, thus unlikely feasible in polynomial time. In practice, the cavity method may provide good approximations by treating networks as tree-like. Here we find, however, that exact computation is possible in many medium-size networks due to existence of a branch decomposition of small width: the edge set of the network may be bipartitioned recursively with only few nodes shared by the two sides of the partition at each level.
Spreading of cultural, social and technological ideas always had a great influence on humanity. Studying such processes and developing new approaches for their modeling, prediction and control has gained a lot of attention in recent years. However, when considering spreading processes that happened in ancient times, we face new challenges: most of the existing archaeological data is sparse and uncertain; lots of information is unknown and there is no procedure to repeat the history and obtain new data.
In this talk, we will present a new framework for modelling of such processes on an example of the wool-bearing sheep spreading in the Near East and Europe, between 6200 and 4200 BC. The introduction of wool had an important influence on the growth of the textile production and had strongly affected the socioeconomic development of past societies. Our new agent-based model combines a data-driven dynamics of human movements with a time-evolving network for possible social interactions. This approach offers an instructive way for studying the qualitative effect of different aspects on the speed and spatial evolution of spreading processes in ancient times, such as: geographic, climatic and cultural aspects.
Spatial metaphors are ubiquitous in political communication. One talks about political positions, political landscapes, and about the distances between opinions or positions of political actors. Thus, it is natural to consider spatial representations of political opinions: Political spaces. The main level of political discourse, that can be observed directly, is the level of political issues. We consider anything on which a political decision can be made as a political issue. The issue space is spanned by the possible attitudes on issues that are the object of political decisions in all fields of politics. This space itself has a multi-level structure, because these issues can be considered at different levels of aggregation. For a single actor, such as a politician, a political party, or a voter the attitude towards these issues is represented as a point (or region) in this issue space. If a population of such actors is considered, the attitudes on the single issues will be normally not statistically independent. Due to these correlations, the attitudes of the population might be approximately represented in a low dimensional space, called “political spaces”. In the present contribution I will discuss, how such political spaces can be inferred from textual data on two examples: Electoral manifestos from the Manifesto project (https://manifesto-project.wzb.eu/) and data from Twitter.
