The column format we use for entailment corpora encodes the following information
(columns are numbered from 1 to N, 1 being the leftmost):

Column 1: Parse dependency tree information: Phrasetype/Phrasehead index
  -- LEAF entry may be prefixed to indicate leaf node; this is not done 
     if the leaf node constitutes a recognized phrase type.

Column 2: Named Entity information: {B|I|O}-NE1/NE2...
  -- more than one NE may be tagged.  
  -- BIO (Begin/Inside/Outside) format used to indicate multi-word NE boundaries

Column 3: Word index (from zero)

Column 4: Part of Speech

Column 5: Word 

Column 6: Verb lemma 

Columns 7+: Semantic Role Label information.
  -- each column applies to one verb in the sentence (whose lemma appears in column 6)
  -- The format here is similar to the dependency tree for column 1.
     -- ABC/N: ABC is tag, N is sentence constituent ABC relates to;
     -- Arguments of verbs: head is ARG0-5, AM_XYZ; these correspond to subject, direct
        object, adverbial phrases etc. (see PropBank annotation details)
        -- index points to verb of which this word is an argument
        -- other argument constituents are tagged MOD_A0 etc., and their index points to the
           designated head of their argument. (FYI, full parse dependency info is used to
           identify heads of multi-word SRL constituents)