Event detection is a key aspect of story development which is composed of multiple narrative layers. Most of the narratives are template-based and follow a narration theory. In this paper, we demonstrate a narrative from events detected in the international relations domain along with classi cation of events using our novel mediation ontology. We also introduce a novel method of classifying events through the mediation ontology. Our methodology involves action classi cation based on the verb categorization of Beth Levin, its arguments determined by universal dependencies and word2vec. The selected feature space is a result of mapping language entities to ontological entities where we obtain substantially good results. The narration also presents interactions of international actors over various topics as well as other visualizations.
Introduction the features used along with machine learning techniques for classi cation are discussed. In subsequent sections, narratives and visualizations are demonstrated in section 6 with a case study. We conclude the paper in section 7 with proposals on the future work sparked by this study. 2
The last few decades have witnessed a considerable escalation in studies which are directed at event coding ontologies in the political domain. This kind of research began during the 1970s with the purpose of forecasting international con ict under the sponsorship of the U.S. Department of Defense Advanced Research Projects Agency (DARPA) [CR78], [AH88]. The kind of research that has been focused is mainly on: 1. the political event data coding ontologies. 2. the generation of the political event data. 3. forecasting of international con ict.
Our focus in this paper is restricted to international relations event coding ontology i.e., Ontology for international relations events or mediation types. Such ontologies include WEIS [Gol92], COPDAB [Aza80], CAMEO [GSYAJ02], IDEA [BBO+03] etc. The WEIS Ontology is made up of 22 top-level categories that encompass actions such as Request or Grant. Each of these 22 top-level categories contain single level children which are further ne-grained. For example, the code 07 is the top-level code for reward with the sub-code 072 representing extended military assistance. The CAMEO ontology is an upgraded version of WEIS with mediation event types added to it. It is more ne-grained with 20 top-level categories that encompass actions such as Make-Statement or Protest. Each of these 20 top-level categories contain ner-grained categories in a hierarchical manner. For example, the code 14 is the top-level code for Protest with the sub-code 141 representing a general demonstration or rally. Under the code 141 is code 1411 which codes demonstrate or rally for leadership change. Thus, as one moves down the hierarchy of CAMEO, it becomes concise. Based on one's need, CAMEO or any event data coding schemes can be evolved using a mix-and-match framework whereby a researcher could adopt most of his or her coding categories from a standard set, and then elaborate on a smaller number of newer categories. Event coding using CAMEO [GSYAJ02] involves event detection and classi cation based on pattern matching from a large set of verb patterns, actors, compound nouns, compound verb phrases, reference to pronouns and deep parsing of sentences in news articles. Prior to CAMEO [GSYAJ02], event encoding was done manually based on the rules mentioned in the corresponding codebooks. Our work presented in this paper carves a similar problem by computing event types and narrative generation of the international events. 3
Bercovitch [Ber97] de nes mediation as \a process of con ict management, related to but distinct from the parties' own negotiations, where those in con ict seek the assistance of, or accept an o er of help from, an outsider (whether an individual, an organization, a group, or a state) to change their perceptions or behaviour, and do so without resorting to physical force or invoking the authority of law." He also mentions, \Mediation may well be the closest thing we have to an e ective technique for dealing with con icts in the twenty- rst century". The main goals of this research are event classi cation and narrative generation so as to help journalists and researchers identify interactions among actors during international con icts.
Our mediation ontology is inspired from social agents grouped under \The ontologies of Persons" mentioned by [Bic12]. He mentions, \Persons, so I argue, are special kinds of agents that arise in and are constituted in interactions with social and cultural processes, including other social persons, and thereby co-constitute the emergence base for those social and cultural realities." He adds that \In this model, agents are constituted in their interactive dynamics; such interactive dynamics is their ontology." Our mediation ontology is inspired from this concept of person and interactive dynamics. He also states that \Agents that develop to become participants, thus constitutive participants, in those environments are themselves, therefore, emergent kinds of agents | social agents:persons." A person's interactive dynamics can be inferred from the pronouncements he makes, the engagements he has with other persons, his response to another person's opinions and the use of force in unhealthy relations. In a similar manner, an actor in international relations interacts with other actors through pronouncements, engagements, responses and force mechanisms. The responses and force mechanisms of an actor determine the pronouncements and engagements made by peers. This is because pronouncements and engagements happen only when some kind of base event has occurred. Hence, force mechanisms and responses are ground event types whereas pronouncements and engagements are lateral event types. As mentioned by [GP12], the relations among the actors is what the eld of international relations is concerned about. Therefore, multiple actors coming together would determine international relations.
The motivation behind a new mediation ontology arises from the 250+ classes in CAMEO [GSYAJ02]. There is an overlap in the mappings from verbs to classes in their verb dictionaries. Also, verb classi cation is an extremely context-sensitive exercise. Hence, we map language entities with ontological entities while proposing a new statistical model of event classi cation system which meets all our requirements. We classi ed an event type into four classes instead of the 20 top-level classes that CAMEO [GSYAJ02] consists of (with nearly 250+ subclasses). Since CAMEO [GSYAJ02] is widely used, we mapped the CAMEO [GSYAJ02] categories as following in order to come-up with the current de nitions of event types.
1. Pronouncements - Declining to comment, making pessimistic and optimistic comment, claiming, denying, empathetic, accord, symbolic act, policy option. - Appeal for material or diplomatic cooperation, aid, political reform, negotiation, settling disputes, accepting mediation. - Expressing intent to cooperate, material or diplomatic cooperation, providing aid, political reform, yield, negotiating, settle disputes, mediation.
- CAMEO Classes - 01, 02, 03. 2. Engage - Consult, discuss, meet, negotiate, mediate. - Engaging in diplomatic, material, economic, military, judicial, intelligence cooperation, endorse, defend verbally, support, recognize, apologize, forgive, formal agreement. 3. Respond
4
Our system listens to 248 media feeds1 for news articles daily. We use the methodology proposed in [KS18] to extract events between September 1, 2017 and September 30, 2017. Since we mapped our categorical information with CAMEO, we used the Petrarch system [CN17] based on CAMEO [GSYAJ02] to generate its event types and map to our categories. We generated a total of 103392 events distributed across all the four classes. Detailed description regarding our dataset is described in table 2.
1http://ceh.iiit.ac.in/international relations/source.txt The methodology is described in gure 2. The sentence in which the event is detected is sent to Stanford Dependency parser[CM14] to identify the action verb and its dependencies. The identi ed action verb, its arguments and universal dependency relations [NdMG+16] are passed through 3 di erent modules which nally unite to form our feature space.
1. In the rst module, we identify the class of the verb with respect to Beth Levin Verb Classes [Lev93] considering verb and its alternations in the sentence. We chose 59 classes which are relevant to our classi cation. Mapping between Beth Levin [Lev93] Verb Classes and our mediation ontology categories is described in table 3. 2. In the second module, the verb and its arguments which are found using the universal dependencies are converted to vectors using Google word2vec pretrained model [TMS]. All the argument vectors are added to the verb vector. 3. In the third module, all the universal dependency relations [NdMG+16] of the verb with its arguments are taken into account. Since, there are 40 universal dependencies as mentioned in [NdMG+16], we consider a 40 dimensional vector which is normalized by the total number of dependencies detected in the sentence.
The results of a few Machine Learning algorithms (viz., Logistic Regression, Random Forest, ensemble of Logistic Regression and Random Forest and Multi-Layer Perceptron) on the feature space obtained from the
Characterize Verbs , Appeal Verbs , Long Verbs , Verbs of Transfer of a Message , Tell Verbs, Verbs of Manner of Speaking , Say Verbs , Complain Verbs , Re exive Verbs of Appearance Pit Verbs , Drive Verbs , Contribute Verbs , Verbs of Future Having , Verbs of Exchange , Build Verbs , Grow Verbs , Create Verbs , Performance Verbs , Dub Verbs , Conjecture Verbs , Admire Verbs , Judgment Verbs , Correspond Verbs , Meet Verbs , Talk Verbs , Chitchat Verbs , Dine Verbs , Gorge Verbs , Verbs of Spatial Con guration , Verbs of Contiguous Location , Verbs of Inherently Directed Motion , Roll Verbs , Verbs that are not Vehicle Names , Accompany Verbs Banish Verbs , Manner Subclass , Verbs of Possessional Deprivation: Cheat Verbs , Get Verbs , Hold Verbs , Verbs of Concealment , Separate Verbs , Split Verbs , Disassemble Verbs , Amuse Verbs , Verbs of Assessment , Search Verbs , Investigate Verbs , Advise Verbs , Break Verbs , Bend Verbs , Other Alternating Verbs of Change of State , Verbs of Lingering
Throw Verbs , Hit Verbs , Swat Verbs , Sight Verbs , Murder Verbs above methodology are described in the table 4. The ensemble technique we used is majority rule voting. All the hyper-parameters of Multi-Layer Perceptron [HM94] are described in table 5. All the metrics (precision, recall and accuracy) are the average of the corresponding class metrics. The optimum result was obtained using Multi-Layer Perceptron [HM94] with precision, recall and accuracy of 80%. The results are favourable using Multi-Layer Perceptron [HM94] because of the backpropagation training algorithm. It is worth to note that all the other Machine Learning algorithms produce nearly same results which gives a strong base for the choice of our feature space. 6
We use the event model presented by [KS18] for extracting events with attributes date-time, location, actors, media-source, event-title, source-url, sentence. Extending this model with action (verb) and action-type (eventtype), our mediation ontology adds two new attributes | action and action-type. These attributes capture individual actions when events are grouped with topics helping in capturing subtler details of the generated narrative. Updated event model is shown in gure 3.
Our system visualizes the event actor interaction using graphical, topical, geographical and temporal features. These visualizations help journalists and researchers in IR domain in understanding the interactions among actors.
Hidden layer sizes Activation function for hidden layer Solver Learning Rate Maximum number of iterations until solver converges Tolerance for optimization
(100,)
Recti ed linear unit function adam (Stochastic gradient-based optimizer proposed by [KB14])
Constant with an initial learning rate of 0.001 200
We consider the de nition of narrative event chain by [CJ08]. [CJ08] states that \A narrative event is a tuple of an event and its participants, represented as typed dependencies. A narrative event chain is a partially ordered set of narrative events that share a common actor (protagonist)." Our narrative event chain is also a partially ordered set of narrative events that shares two common actors.
The narrative starts with the description of two actors followed by various visualizations describing the interaction of the actors.
1. Bar Chart: The bar chart depicts frequency of the events reported involving both the actors with the most reported topics of interaction and unique events. 2. Line Chart: The line chart compares the frequency of the events reported involving individual and combined actors. This provides how important the interaction of the mentioned actors is in comparison to total number of interactions reported with all other actors. 3. Mediation Class Events: The mediation class events gives the number of events reported in each category of the mediation ontology. This visualization provides comparison among the classes of interaction. 4. Graph Visualization: The graphical visualization represents the interaction wherein nodes are the actors and the edges are topics of interaction. This visualization helps place an actor level context to the con ict.
Hovering on the nodes and edges provides with the actor name and the topics of interaction respectively. 5. Topic Cloud: The topical visualization helps situate the gravity of the topics spoken of and thus giving a subjective view of the con ict. It is created based on the frequency of the words used in the topic. 6. Geographic Visualization: The geographical visualization brings to the fore the narrative about actor's stakes in the international con ict and thus add a geopolitical persona to the event. The importance of the actor increases with the increase in the number of sub-actors. Hovering on the circle gives all the actors and sub-actors involved in that area. 7. Timeline Visualization: The timeline visualization helps bring a coherency to the event-actor duo and places the interaction over a span of the dialogue until its closure.
We show an example of a narrative involving Japan and South Korea as common actors surrounded by a narrative event chain. All the narrative visualizations between them are shown in gure 4.
A live prototype of the system is available here: http://ceh.iiit.ac.in/international_politics/ (a) Bar Chart (b) Line Chart
(c) Graph Visualization (d) Mediation Class Counts [AH88]
Edward E. Azar. The con ict and peace data bank (copdab) project. Journal of Con ict Resolution, 24(1):143{152, 1980.
Doug Bond, Joe Bond, Churl Oh, J. Craig Jenkins, and Charles Lewis Taylor. Integrated data for events analysis (idea): An event typology for automated events data development. Journal of Peace Research, 40(6):733{745, 2003.
Jacob Bercovitch. Mediation in international con ict: An overview of theory, a review of practice. Peacemaking in international con ict: Methods and techniques, pages 125{154, 1997.
Mark H. Bickhard. The emergent ontology of persons, page 165180. Cambridge University Press, 2012. [CM14] [CR78] [GP12] [HM94] [Kan09] [KB14] [KS18] [Lev93]
CoRR, VenuMadhav Kattagoni and Navjyoti Singh. Towards an unsupervised learning method to generate international political event data with spatiotemporal annotations. In Andrew U. Frank, Christine Ivanovic, Francesco Mambrini, Marco Passarotti, and Caroline Sporleder, editors, Proceedings of the Second Workshop on Corpus-Based Research in the Humanities CRH-2, volume 1 of Gerastree proceedings, pages 105{112, 2018.
B. Levin. English Verb Classes and Alternations: A Preliminary Investigation. University of Chicago Press, Chicago, IL, 1993.