What Makes Discourse Coherent

What Makes Discourse Coherent

Parsing discourse

Centering

Entity grids and self-supervision

Under RST, in the two-EDU passage [The ferry cut its crossing short.] [The fog bank was closing fast.] related by REASON, which span is the nucleus and which is the satellite?

How do annotators in the Penn Discourse Treebank style label a coherence relation between two adjacent sentences that share no explicit connective?

An entity grid has $8$ sentence rows and $4$ entity columns, and a model uses every length-2 column transition as a feature. How many such transitions does the grid contain?

Utterance $U_n$ has backward center the ferry and utterance $U_{n+1}$ evokes a new backward center the pilot, which is also the top-ranked forward-looking center of $U_{n+1}$. Which centering transition is this?

Four sentences have pairwise entity overlap counts, adjacent pair (1,2) shares 1 entity, (2,3) shares 1, (3,4) shares 1, and every other pair shares 0. Compute the entity-overlap coherence score of the ordering $(4, 2, 1, 3)$, where each adjacent pair contributes its shared-entity count.