Ellipses in coordinations

A recent paper in the UD workshop, https://www.aclweb.org/anthology/W18-6012, compares a data-driven approach and rule-based approach for adding enhanced dependencies to existing treebanks in Universal Dependencies. My impression is that ERG can produce already a representation even richer and the Enhanced Dependencies or at least, comparable, right? That was the case for the examples a and b from Figure 1. But for example c from Figure 1 the output of ERG was very strange: “Sam fixed lunch and Kim dinner.”. In the MRS-Derived Semantic Dependencies (DM) output from http://erg.delph-in.net/logon, the output didn’t reveal the expected relation between Kim as subj of the elliptical event ‘fixed’ and its obj dinner. Is my interpretation right? Is it a problem in ERG?

In principle, yes, much of what EUD calls for is available in the ERG’s analyses. It’s phenomenon by phenomenon though and you picked a particularly tricky one (gapping) that’s not in the 1214 release. I know Dan’s been working on it though. Maybe it’ll be in 2018?

1 Like

So @ebender, in principle, the conclusion in this paragraph is not fair:

A more recent approach states that the challenges posed by ellipsis to phrase structure theories of syntax are due to the phrase structure component of the grammar. In other words, the difficulties facing phrase structure theories stem from the theoretical prerequisite that syntactic structure be analyzed in terms of the constituents that are associated with constituency grammars … If the theory departs from phrase structures and acknowledges the dependency structures …, the ability to acknowledge a different sort of syntactic unit as fundamental opens the door to a much more parsimonious theory of ellipsis.

Note the warning box at the top of the Wikipedia article! It looks like something of a personal essay. I think it would be best to read that paragraph as a criticism of movement-based approaches. From the poin of view of that paragraph, Delph-in grammars are something of a hybrid between constituency and dependency approaches, because there’s a context-free backbone but also semantic dependencies.

The examples in the article aren’t analysed in the same way in the ERG, so as @ebender pointed out, it’s something you’d have to look at phenomenon by phenomenon.

1 Like