(Note: this session had no scribe, so I'm summarizing based on memory)

The point of this session was to brainstorm how we can link our resources to non-DELPH-IN resources, e.g., FrameNet or Propbank. There has been a theme of how to make our tools and resources more available, accessible, and understandable to people outside of our group, and this falls in line with that theme. Furthermore, we have our own lexical resources, such as the large and richly informative lexica provided by our grammars, but they are not intended to be used for extragrammatic applications, so we might consider them underutilized. Some non-DELPH-IN lexical resources, such as FrameNet, Propbank, WordNet, etc. are documented and usable for a variety of applications. Perhaps by associating our lexical entries (and possibly other ontological information, like SEM-I properties, parts of speech, etc.), there could be some benefits:

• Our lexica could supplement other resources (in quantity or quality of lexical information)
• Our grammars can be made more understandable to others, and possibly more amenable to integration into projects that already use the external resources
• We could utilize the external resources for our own gain, e.g. WSD for MT, etc.

We brought up how the surface predicates (aka RealPreds) have a sense field that we could exploit. For instance, if it were possible to linked our existing lemma+pos+sense triples to WordNet senses, we could later apply WSD to further specify the sense. For example:

This idea is initially appealing, but the default sense would often be linked very high up (e.g. the sense for "object" or "entity" and not interestingly different across predicates).

I wondered if sense disambiguation, if it could be done during the parsing process, could help to refine the parse selection process. Glenn pointed out (as the sole developer of a parser/generator in the room) that such an application would probably be more trouble than it was worth, and would probably do little more than explode the parse chart with spurious, subtly different trees. So we generally agreed that any such alternations to the sense field should be done as a post-operation, if at all. In some serializations of MRS (e.g. [JSON, or maybe XML) Alternatively, we could maintain the sense linking as a stand-off annotation. E.g., an MRS object would remain as it is now, but it might be accompanied by an additional structure that maps the predicates to the senses. In MRS, intrinsic variables would not be sufficient (as they are not unique per EP), so this could be done by matching from a tuple of (predicate, Lnk span (cfrom/cto)). In DMRS, the nodeid of each predication would be unique and could be used instead.

We did not get far into a discussion about linking to FrameNet, but it would not only involve linking our predicates to FrameNet senses, but also each predicate's core argument inventory (e.g. ARG1, ARG2, etc.) to the roles in each frame (to be complete, maybe also linking non-core roles in the frame to adjunct subgraphs in MRS, perhaps with MRS fingerprints). Such an annotation project would be a significant undertaking, so it might be beneficial to first do some exploratory exercises. For example, to initially link ERG predicates to FrameNet senses, then compare the EP arity with the number of roles in the associated frame. It might also be useful to do SRL using MRS role names and see if the sentence annotations structurally resemble those from FrameNet (if so, it might be feasible to automatically associate predicate+ROLE to frame+ROLE).

(if you attended this session and remember more details, please feel free to edit the page; thanks!)