Annotation Support for Aspects of Textual Entailment

[ghost]

This issues is to discuss the way we can assist the annotation of varies aspects of textual entailment. Please see page four and five of the paper below for examples:

http://www.celct.it/download/Textual-Entailment-LREC2010.pdf

I haven't given this overly much thought yet, but what we really want to avoid is duplicating sentences and still keeping it tidy and comparable between sentences. In particular since there is likely to be several source sentences and only one target sentence which is very different from what we have in brat so far.

To be noted, this is a request from Aizawa-sensei that has a very clear usecase for this.

[spyysalo]

Some early thoughts:

• Could the source text and the related hypothesis sentences be combined into a single document? If not, this might require cross-document references, which is likely to be quite demanding technically (esp. client-side).
• If I understand the aim correctly, the entailed/not entailed decisions would apply on the level of hypothesis sentences. This could be captured either by marking each such sentence as a span and attaching the decisions to the span annotation (inelegant, potentially messy annotation) or by introducing sentence-level annotations. We've had a number of requests for the latter already, so maybe it's time?
• The bits that are related in some way (paraphrases, coreferential, etc.) between sentences would be quite naturally captured by standard spans and relations. The only potentially problematic issue here is that as all relations are cross-sentence, the default layout algorithm might make more of a mess of things than usual, but I suppose we won't know until we try.