Annotation Manual

Given are pairs in natural language, which can be sentences or larger text
segments.

Sentence T is assumed to be true.
Sentence H will be coded according to this assumption.

Each inference will be classified by several attributes, listed below.

value:
  - ALL    H is true in all possible worlds where T is true.
           This is equivalent to TRUE (STRICT).
  - SOME   In all worlds where T is true, H can be true, but is not necessarily
           true.  This is equivalent to TRUE (PLAUSIBLE), UNKNOWN, and FALSE
	   (PLAUSIBLE).  The subjective attribute or context attribute (see
	   below) may give more information as to whether it is more likely to
	   be true than false.
  - NONE   H is false in all worlds where T is true

context:
  This will be coded only in cases of "value=SOME".  In cases where the text T
  introduces a context whose truth conditions are not specified, and the
  assumption of the context will determine the truth value of H, then the type
  of the context will be given in this attribute.
  
  Other contextual-reasoning modules will be able to determine the default
  credibility of certain contexts or even the explicit truth values of them.
  However, just marking them as such is free of any subjectivity.  Multiple
  codings are possible.

  Here is a preliminary list of context types:
  - TEMPORAL	a temporal context is defined (Before dinner, Jim is cranky.)
  - SPATIAL	a spacial context is defined (In Texas, summers are very hot.)
  - BELIEF	a belief/opinion context. (Analysts expect the stocks to rise
                soon.)  Note that we are marking the value of the CONTENT of
		the context, not its source.
  - DOMAIN      a context restriction on the domain. This can be for unstated
                information as well as word/concept senses. (Reggy Miller shot
                the lights out. Domain: NBA vs Domain: Law Suits)
  - CONDITIONAL If-Then type of contexts (If Jim wins the next round, he wins
		$100,000)
  - POSSIBILITY	contexts of probability, likelihood (It might have rained.)
  - PLANNING	contexts of planning. (The train was scheduled to arrive at
                4:00 PM) 

subjective:
  On a scale of 1-5, the Annotator has to present his personal judgement about
  the tendency of the truth value.

  This is important because these judgements may vary. This data might also
  assist in builig weighting mechanisms. However, we do not expect a high
  inter-coder agreement, that is why we chose to introduce this as a side
  metric.  This coding has to be done independently of the context.  Given T:
  
  1: I am absolutely convinced H is false
  2: I am somewhat convinced H is false / I believe H is false
  3: I am neutral on the truth of H
  4: I am somewhat convinced H is true / I believe H is true
  5: I am absolutely convinced H is true

source
  Sources of transformation or inference.  Includes, but is not limited to, the
  following list:
  - LEX	    Lexical information such as lexical chains
  - SR	    Semantic relations
  - LNG	    Linguistic knowledge
  - BWK	    Basic world knowledge
  - XWK	    Extended world knowledge
  

General Interpretation of the Text passages:

  If you cannot constrain word senses, or conceptually equivalent synonyms with
  domains, then assume the reading that would allow H to be true.  Tense and
  time that is not specified in contexts shall not be taken into account
  negatively.  All unspecified times shall be read as being the most
  accurate/current piece of information.  Named entities that are coreferenced
  in a lax fashion should be accepted, as long as they are unambiguous in the
  context of the pair.