Frame element hierarchy

O’Hara and Wiebe (2009) provide an inventory of semantic relations derived from the frame elements of FrameNet, along with a mapping identifying what was done for semantic roles occurring at least 50 times. They also discuss a number of other schemes for identifying a set of basic semantic relations or roles. Marquez et al. (2008) also discuss the nature of semantic roles , particularly noting the lack of agreement among linguistic researchers. In all these cases, a common thread is that researcher’s judgment has been used to develop these basic inventories.  While these inventories appear to be reasonable, I would like something more data-driven. I suggest that the frame-to-frame relations in the FrameNet database can be used as the basis for a more principled identification of primitive frame elements. The methods I suggest use a digraph analysis of a dictionary of frame elements constructed out of individual mappings developed in the FrameNet project itself.

As suggested in a previous post, each preposition sense needs to be associated with one or more frame elements. Thus, it is important to consider the relationships among the frame elements. FrameNet uses frame-to-frame relations to map frame elements of one frame into those of a more primitive frame. These can be exploited for developing a frame element hierarchy. Specifically, FrameNet uses the following six frame-to-frame relations: INHERITS, USES, PRECEDES, IS_SUB_OF, PERSP_ON, INCH_OF, and CAUSE_OF. For example, Achieving_first INHERITS Intentionally_create, with the following mapping among frame elements:

  • Instrument Manner Means Place Purpose Time Basis New_idea Cognizer Field
  • Instrument Manner Means Place Purpose Time Components Created_entity Creator Depictive

Notice that the first six frame elements are identical, whereas the last four are different. In Achieving_first, these last four are more specialized versions of the last four in Intentionally_create. We can say that Created_entity is a hypernym of New_idea.

With this basic idea, we created a dictionary of the 1015 distinct frame elements that have been used in FrameNet. In creating this dictionary, we identified all the mappings and used them to create hypernymic relations between the frame elements. In addition, we created a “definition” for each frame element as the number of frames in which the frame element is used; for example, the Agent frame element was used in 125 frames. We then used CL Research’s DIMAP to perform a digraph analysis to create a hierarchy of frame elements. An example of the hierarchy that emerges is Event –> Behavior –> Means –> Transport_means –> Vehicle –> Mode_of_Transportation. (The maximum depth is about 7 or 8 nodes.) The resultant digraph is an image 3.3 MB in size.

There are many problems in this hierarchy, including

  • Circularity
  • Frame elements without hypernyms, thus resulting in a large number of so-called primitive elements
  • Potential weak elements of hypernyms

The image takes quite a while to load (perhaps a few minutes in Firefox), but takes less than a minute to create with DIMAP. As a result, it will be easy to make changes in the entries and create a new digraph.

The next step is to develop a set of procedures for making changes to individual entries to arrive at a consistent set.

You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.

1 Comment »


Leave a Reply

XHTML: You can use these tags: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>