Skip to content
Back to search results

Event Data Analysis of Conflict

Introduced by
Philip Schrodt (KEDS/TABARI); ICEWS team (Boschee et al.)
Year
1994
Type
Automated extraction of structured political events from news text
Family
Information extraction / NLP for political science
Font size

What does it do?

Event data analysis is the automated extraction of structured records of political interactions — who did what to whom, when, and where — from large volumes of news text, for the quantitative study of conflict and cooperation.

Pioneered for machine coding by Philip Schrodt with the KEDS and TABARI systems and scaled in projects such as ICEWS and GDELT, it turns unstructured reporting into dated actor-action-target triples coded to an ontology like CAMEO, which can then be aggregated into time series of interstate or intrastate hostility.

The intuition behind the method

Newswires report a constant stream of who-did-what-to-whom: a government condemns a neighbor, troops mass at a border, two states sign an accord. Read one by one these are anecdotes; counted and scored systematically they become a measurable pulse of international relations. Event data analysis automates the reading: a parser identifies the source actor, the action, and the target in each sentence, maps the action to a standardized code with a known conflict-cooperation intensity, and tallies the codes over time. The result is a near-real-time conflict barometer for any dyad of actors, built from millions of stories no human team could code by hand.

History

Quantitative event data in international relations began with hand-coded projects like WEIS and COPDAB in the 1960s-70s. Philip Schrodt and colleagues automated the enterprise in the 1990s with KEDS and its successor TABARI, demonstrating that machine coding of news could rival human coders. The CAMEO ontology standardized the action categories. The 2010s brought large-scale near-real-time systems — ICEWS, developed under U.S. defense research funding and released on Harvard Dataverse, and the open GDELT project — alongside open-source coders such as PETRARCH, embedding event data firmly within computational social science and NLP.

When to use it

Use event data analysis when you need fine-grained, high-frequency measures of political interaction across many actors and a long time span — for conflict forecasting, early warning, studying escalation dynamics, or testing theories of cooperation and conflict at daily or weekly resolution. It excels where hand-coding is infeasible because of volume. It is less suitable when the phenomenon is rarely reported or systematically underreported, when fine semantic distinctions exceed the coder's accuracy, or when the research question requires the contextual judgment that structured triples discard; in those cases curated datasets like the Correlates of War militarized disputes may be preferable.

Limitations and considerations

  • Entirely dependent on media coverage: events that go unreported do not exist in the data, and coverage is geographically and politically uneven.
  • Automated parsing makes systematic errors — misattributed actors, miscoded actions, duplicate reports inflating counts.
  • The CAMEO ontology and Goldstein weights impose a fixed schema that may not match a study's theoretical categories.
  • High event volume can create an illusion of precision that masks substantial measurement error and source bias.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Scales to millions of reports, producing near-real-time, high-frequency measures impossible to hand-code.
  • Yields standardized, reproducible codes (CAMEO) and intensity scores (Goldstein) comparable across studies.
  • Covers the full conflict-cooperation spectrum and any pair of actors the news mention, not just wars.
  • Supports forecasting and early-warning models because new events are coded as soon as they are reported.

Applications

  • Conflict early-warning and forecasting systems that predict the onset or escalation of violence.
  • Studying escalation and de-escalation dynamics between rival states at daily or weekly resolution.
  • Measuring cooperation, such as diplomatic engagement or mediation, across the full event spectrum.
  • Constructing dyadic hostility series as covariates in models of war, alliance, or trade.
  • Crisis monitoring and situational awareness for analysts and humanitarian organizations.

How it works

How it works

1

A continuously updated corpus of date-stamped news reports — historically newswire services, now global digital news — is ingested, deduplicated, and segmented into sentences. Coverage, source bias, and duplicate reporting are first-order concerns because the event series can only reflect what the media report.

2

Each sentence is parsed to identify a source actor, an action verb phrase, and a target actor, producing a who-did-what-to-whom triple. Early systems (TABARI) used sparse-parsing pattern dictionaries; modern pipelines (e.g., PETRARCH and neural extractors) use fuller syntactic and semantic parsing to resolve actors and actions.

Formula
\text{sentence} \;\rightarrow\; (\text{source actor},\, \text{event},\, \text{target actor})
3

Actors are resolved to standardized entities (states, organizations, leaders) and actions are mapped to a category in a conflict-cooperation ontology — most commonly CAMEO, which classifies events from verbal cooperation through material conflict. Each coded event e carries a source s, action a, target r, timestamp τ, and geolocation g.

Formula
e_t = \big(s,\, a,\, r,\, \tau,\, g\big),\quad a \in \text{CAMEO}
4

Events are aggregated into directed-dyad time series. Each event's action code is assigned a weight w — for example on the Goldstein scale of conflict-cooperation intensity — and summed over a time window T for each source-target pair, producing a quantitative measure C of how hostile or cooperative actor i was toward actor j.

Formula
C_{ij,T} = \sum_{t \in T} w(a_t)\,\mathbb{1}\{s_t=i,\, r_t=j\}

Common pitfalls

  • Treating event counts as counts of reality rather than counts of reporting, ignoring media-coverage bias.
  • Failing to deduplicate, so a single incident reported by many outlets is counted many times and inflates intensity.
  • Comparing series across periods when the underlying news sources or the coder version changed, confounding real change with pipeline change.
  • Over-trusting actor resolution, so events involving ambiguously named actors are attributed to the wrong state or group.

Frequently asked questions

What is CAMEO and why is it central to conflict event data?

CAMEO (Conflict and Mediation Event Observations) is the standardized ontology that classifies political actions into a hierarchy ranging from verbal cooperation to material conflict, with a parallel scheme for actors. It is central because it makes machine-coded events comparable across projects and time and because the Goldstein conflict-cooperation weights are defined over its categories, turning categorical events into a continuous hostility scale.

How does this differ from datasets like the Correlates of War militarized disputes?

Militarized interstate dispute data are carefully hand-coded, low-frequency records of confrontations involving the threat or use of force, curated for accuracy and theoretical precision. Event data are automatically coded, high-frequency, broad-spectrum records derived from news, trading some accuracy and contextual judgment for scale, timeliness, and coverage. They are complementary: see the related Militarized Interstate Dispute Analysis entry.

Why is event data analysis treated as a natural language processing task?

Because its core technical work is information extraction from text: parsing sentences, resolving named entities (actors), classifying the relation (action) between them, and geolocating and timestamping the result. Modern pipelines use the full NLP toolkit — syntactic parsing, named-entity recognition, coreference resolution, and increasingly neural language models — which is why it sits in the NLP tasks leaf.

Related methods

Sources

  1. Schrodt, P. A., Davis, S. G., & Weddle, J. L. (1994). Political science: KEDS — A program for the machine coding of event data. Social Science Computer Review, 12(4), 561–588. See also Gerner, Schrodt et al. (1994), Machine coding of event data using regional and international sources, International Studies Quarterly, 38(1), 91–119.

    Open DOI reference
    Copied.
  2. Boschee, E., Lautenschlager, J., O'Brien, S., Shellman, S., Starz, J., & Ward, M. (2015). ICEWS Coded Event Data. Harvard Dataverse.

    Open DOI reference
    Copied.