We propose a five steps methodology to retrieve, reconstruct and analyse conflict related narratives in a standardized and automated way. Our methodology combines AI and network analysis techniques to build a visual representation of key agents and entities involved in a conflict and to characterize their relations. Unlike the majority of existing methods, ours can be applied to any type of conflict, as, through two data downloading phases, it first generates a bird's-eye representation and then a fine-grained map of any conflict. Given the broad applicability of the proposed methodology, we believe that this work moves the first steps towards a better understanding of conflictual narrative mechanics.
Our hypothesis is that conflict narratives are 2.1. Conflict narratives
strategically designed around recurrent story-telling
patterns and frames that assign a set of (asym- Conflicts and their escalation in the physical and
metric and stereotyped) roles to involved parties. narrative space are generally the by-product of
hetWe plan to ultimately utilize the methodology pre- erogeneous beliefs [
When trying to capture the context and view- and objectives, which are often opaque being difipoint of narratives, one of the main dificulties is cult to elicit and model. not to introduce researcher bias to sampled data: Conflict narratives, and more in general debates By collecting, for example, only posts written in about conflicts through which these narratives English, by defining which terms to query for, which evolve and spread, have long been studied in the conflict-related narrative frames are being looked social and political sciences under many diferent up, or which actors in the conflict are of interest approaches and perspectives [17, 18, 19]. For exapart from the main parties, any resulting corpus ample, the study of conflict framing [ 20], factional represents only a certain fraction of the discourse. discourse design [21], and other polarizing commu
Therefore, we define a two-phase approach to nication strategies [22] is key for understanding any investigate conflict narratives based on online con- consensus-building and group-mobilization process versations on Twitter: when there are competing views or interests at stake.
The narrative dimensions of conflicts are gen1. A dataset is collected of tweets that explic- erally more visible in non-authoritarian countries, itly mention the conflict at hand through where partisan narratives [23] are constructed and the names of the main parties or neutral employed in relation to public support-building and conflict-related terms. From this, the enti- policy justification. Conflict narratives operate at ties involved in the conflict, such as actors, all levels of national and international governance events, and locations, either actively or pas- processes. Especially democratic countries aim to sively, are discovered and the most important mitigate and resolve potential or actual conflicts entities and their co-occurrence are identi- through transparent, informed, and participatory ifed. deliberative processes. From local debates about 2. More fine-grained data is then collected by public policies to be implemented in response to searching for all the main entities and cap- a pandemic, like COVID-19, to the renegotiation turing frames used to characterize these enti- of trade agreements between parties that compete ties as well as possible relationships between for the control on strategic natural resources and them. technologies, like conflicts for rare earths or for the control over semi-conductor industry technologies,
So far, we have conducted the first phase of the the emergence of conflict narratives is an ubiquitous analysis on the conflict between Russia and Ukraine, phenomenon in contemporary times. This is particas characterized by a sample of the Twitter conver- ularly evident in decentralized and multi-directional sation in the spring of 2022. In order to explain the online communication mediums such as social media proposed methodology, we present the results of this [24, 25], which have become the default propagation ifrst step, detail the steps planned for the second medium for (popular) narratives, including conflictphase analysis, and discuss which challenges remain related ones. to be solved in order to expand the methodology to a greater scope and other case studies.
2. Related Work shown that conflicts are not necessarily the outcome
of diverging material interests among individuals
The following section presents a brief summary of and groups, as modeled in early game-theoretic
recent literature on conflict-related narratives with works [27]. They can also be the product of: (i)
a particular focus on how these narratives are con- diferently-biased or competing world views used to
structed and difused through online social media. decipher events and to comprehend the intentions
underlying specific actions or communications by
other individuals and groups [28]; or (ii) to the
deterioration of inter-group trust [
Recent works [
The former branches of research open the way to often in this context. We are then able to construct a new AI-augmented research field on all kinds of a network based on their co-occurences. conflicts. Such research could combine the potentials of NLP, network analysis, and computational 3.1.1. Step 1: Corpus construction linguistic methods with the semantic web, serving as interfaces for the real-time observation and un- To construct the first corpus on the current war in derstanding of conflict-related narratives. Ukraine, we collected all tweets
lficts, we developed a methodology consisting of two phases which include two cycles of data collection. The reasons behind this are twofold: Firstly, it allows the discovering of actors and entities of • between the first of January and the first of
May 2022 • which contain the words: “Ukraine”, “Russia”, and any of “conflict(s)”, “tension(s)”, “crises” or “crisis” • which are not retweets.
For this, we have only searched for these terms in same holds for other entities involved, e.g. a person English. Hence, the collected corpus only represents being referred to as a “president” or “dictator”. Eia perspective from English-speaking Twitter users. ther variant will be collected if it appears in a noun However, we want to also collect tweets contain- phrase together with one of the named entities. ing these key terms in other languages, including In this step, we use the open source NLP library Ukrainian and Russian, in order to generate a mul- spaCy for part-of-speech tagging, dependency parstilingual corpus that represents a broader view and ing, and named entity recognition. From this, we contains difering narratives that include those from identify all noun phrases which include a token that both of the opposing parties. was classified as a named entity. Such identified en
An advantage of this two phase approach is that tities would include only geopolitical entities such as
at the beginning, only a small number of terms needs countries, nationalities or religious/political groups,
to be defined, which are compatible with any type organizations, persons, events, and locations.
Howof conflict. We plan to employ the help of native ever, these entities would exclude instances such as
speakers of other languages to translate these terms cardinals.
and validate the results of the first phase. However, Identified noun phrases can then be consolidated
we hope that we will not require the help of experts in entities, by linking them to named entities in a
in the domain which are also native speakers of known knowledge graph such as WikiData
(consolithe additional languages. Apart from the definition dating e.g. “the dictator Putin”, “Vladimir Putin”
of this small set of neutral conflict-related starting and “the russian president”, etc.). For the
prelimterms, the collection of the data (across both phases) inary analysis that follows, the consolidation step
is entirely automatic. We chose to go with only these has yet to be implemented. In future versions of this
three terms (conflict, tension and crisis) firstly in work this will be done through name matching and
order to limit the scope of the twitter query, and name substring matching, without disambiguation,
secondly because we believe that other synonyms or through more advanced distributional semantics
of those terms generally connote either violent or and ML methods, like OpenTapioca [
For this preliminary analysis, we collected a total In the third step, identified entities are mapped of 724400 tweets. In a first pre-processing step, through an undirected network based on their cowe removed special UTF-8 characters, like emojis, occurences in the corpus. Each node represents a emoticons, and URL links. entity, with edge weights denoting how often two entities occur together in one tweet in the dataset. By 3.1.2. Step 2: Entity Recognition so doing we obtain a weighted and undirected network containing 933081 nodes and 10871807 edges.
After generating this dataset, we use this broad- With an average degree of nodes equal to 23.303 coverage overview of the conflict at hand to discover and an average weighted degree of nodes equal to from it which actors or other entities are relevant to 46.162. the conflict. These will constitute the terms which Network nodes are then filtered based on their we will explicitly teach for in the second phase. Of weighted degree centrality metric; alternative ceninterest are trality measures, like pagerank, betweenness or eigen-centrality may also be used for this purpose. • named entities such as known persons, orga- This filtration is done in order to remove entities nizations, countries or peoples, e.g. in this that are less influential (more peripheral) and have case “Putin”, “Ukraine”, “EU”, “Russians”. a marginal role in the entity network for the se• noun phrases which include these same terms lected conflict. As shown in Figure 2, this filtration as signifiers, such as “Ukrainian president”, step removes those entities which are mentioned “russian army”, “Ukraine war”, “russia- rarely in the dataset, like entities weekly related ukraine conflict”, etc. to the conflict and other irrelevant noun phrases.
This might include entities such as sports teams or their fans who might be involved in a metaphorical conflict, misspellings, or other noun phrases that do not play a relevant role in online conflict-related narratives.
lfict and thus characterizing it ourselves – based on the diferent connotations of “war”, “unrest”, “invasion”, “operation”, etc. – we are able to discover which terms are used in online conversations. The
grained data about the conflict on the selected social media, in order to (re-)construct and analyze the narratives surrounding specific relational blocks.
As a starting point for this will serve the key actors/entities and relations (dyads of actors), which were identified in Phase 1 based on the chosen centrality metric, together with the largest/heaviest cliques of order 3+. 3.2.1. Step 4: Second corpus construction At this point, we have identified a number of key entities and relations, as well as diferent reference terms which are used to refer to them. These will then be used to construct a new set of queries from which a second corpus of tweets is collected. In this second phase, we are going to search explicitly for the entities that the first phase identified as being perceived as important actors in the conflict, using the terms that were discovered to be used to refer to them, by twitter users.
While the first corpus included only tweets that reference the conflict itself, thus providing a broader view of the online conversation surrounding it, this 3.2.2. Step 5 - Analysis second corpus will include tweets referring specifically to one or more of the relevant entities and allow for a more fine-grained analysis. By collecting these tweets, we aim to identify how the entities themselves, as well as the relationships between them, are characterized by diferent people online, for example through adjectives and verbs qualifying the relation between two key entities.
Our hypothesis is that the narratives surrounding the actors involved in the conflict are based on recurring phrase fragment patterns and frames that assign specific roles and attributes to the involved parties. In many cases, these roles are a-symmetric and mirror certain stereotypes that are common to (almost) all conflict narratives, like the role of the victim and that of the perpetrator. We plan to use the methodology we describe herein to investigate this hypothesis on the corpus we are currently collecting about the war in Ukraine. Using the tweets from this Phase 2 corpus, we aim to collect and assign a number of diferent frames to the actors and relations, which are essential constituents of the conflict narrative. For example, we expect that we will find • verbs related to asymmetric roles in the conlfict, like: aggression / protection, ofense / defense, attack / counter-attack, ownership claims, “deserviness” claims, resisting/ surrendering, etc. • conflict related nouns and adjectives, like: aggressor / aggressed, invader / invaded, liberator / liberated, oppressor / oppressed, strong / weak, winning / losing, perpetrator / victim, etc. • characterizations of the conflict or its escalation, like: justified / unjustified, legitimate / illegitimate, necessary /unnecessary, explainable / unexplainable, expected / unexpected, hot / cold, violent / non-violent, verbal /physical, ideological, political, economic, financial, military, etc. • equivalently, characterizations of – the peoples or populations involved in
the conflict – the leaders of the factions involved in
the conflict – the countries involved in the conflict – the factions, armies or soldiers involved
in the conflict – etc.
based of RDF-star or a labelled property graph, is an in-depth representation of the narratives describing the relations between the key actors and their characteristics. For this step, we plan to again employ the part-of-speech recognition and dependency tree identified using the spaCy toolkit to reconstruct this network.
Phase 1 dataset shows an interesting bird’s-eye perspective on the online narratives surrounding the selected conflict, that is, the war between Ukraine and Russia. However, this pipeline for automatically reconstructing conflict narratives is still at a very early development stage. Open questions remain concerning the advantages and disadvantages of the centrality measures and AI methods to be used at each stage of the process, for instance, ifltering the actors network. A future aim is to explore and benchmark these alternative procedures and metrics, as well as their impact on the results, This work was funded by the by the FET-Open throughout future works, employing multiple con- Project #951846 “MUHAI – Meaning and Unlfict datasets. Here follows a brief discussion about derstanding for Human-centric AI” by the EU the criticalities that we have identified at the cur- Pathfinder and Horizon 2020 Program and the rent state of the work and that we will address as a “Empowering Digital Media” grant from the Klaus priority in the next development stages. Tschira Foundation.
Firstly, we plan to evaluate the most useful way to construct the co-occurence network. Currently, References the network is based on the counts of co-occurences of identified entities and noun phrases in tweets.
More precisely, the more often two entities appear in the same tweet, the higher the weight of that edge is. An alternative approach is to take the frequency of any two connected nodes into consideration when having to weigh the relation between two entities (as more frequently occurring entities are also more likely to appear together by mere chance). This metric would also allow us to diferentiate between pairs of nodes that appear together more often than random, and pairs of nodes that appear together less often than random. Similar approaches have been used in other fields of studies, for example, for species probabilistic co-occurrence analysis. In connection with this, we are also considering a number of diferent options to calculate the centrality of a node, which is currently based on its degree.
The second part which we will further investigate is the representation of conflict dynamics over time.
The current network is based on the entire set of tweets collected from a defined time frame. During the next iteration, we plan to firstly collect tweets from a larger time interval and secondly to slice the network into smaller time-chunks. This will enable us to visualize and analyze how the network changes over time. We expect an interesting perspective to be added to both the co-occurence network of the first phase as well as to the more fine-grained network.
Thirdly and finally, we plan for the next iteration of this dataset to analyze more of the context data surrounding the tweets themselves: Many tweets include metadata about the country of origin, time zone, location of users, and their language, signifying emojis, flags and hashtags, as well as the metadata connected to the user’s account. Considering users’ biographies or geolocation might make it possible to look at which narratives are more prominent among which groups of users or locations. This is also connected to our previous goal of adding more languages to the corpus. Acknowledgments