The Jesuits Massacre

See the 1989 murders of Jesuits in El Salvador Wikipedia page for a description of this atrocity committed during the Salvadoran Civil War. Nobody claimed responsibility for the attack - and initial reports contained claim and counter-claim about who was responsible for the murders. The truth emerged over time, ultimately leading to a comprehensive description of events in a UN Truth Commission report.

Contemperaneous reports relating to the murders can be found in the MUC-3 corpus. These can be used to demonstrate developing hypotheses based on testimonial evidence, and in updating hypotheses over time in the light of new evidence.

The Analyst Perspective

  1. Collecting evidence and listing excerpts from reports that make allegations.

  2. Rephrasing these excerpts to to distil out the claims made. Analyst judgements can be added at this stage.

  3. Creating an information list of evidence for use in analysis.

  4. Applying a structured analytic technique, such as ACH.

The Technical Perspective

The example web pages are marked up according to the AIF dialogical model. AIF models the concepts of locution, rephrasing and illocutionary force that link the excerpts, their rephrasing, and the evidence they claim. This means that any conclusions reached by arguments from testimony can be traced back to source, and issues of trust and credibility can be considered in deciding whether to accept those conclusions.

There are number of manual tasks here that gloss over active research topics in natural language processing (NLP): event extraction to identify the event of interest and the documents that report it; argument mining to identify premises and conclusions expressed in those documents; and intertextual correspondence to piece together arguments extracted from various reports to create a cohesive argument map. The idea is that any future machine process conducting these tasks would interact with the linked data model in the same way.

We can decouple the extraneous NLP from that needed specifically for argument mining by making both capable of producing and consuming the same NLP serialization. Baleen XML is suitable for this purpose. See here for the NLP serialization of the documents driving this example.