📄 Towards a Discourse-Oriented Representation of Information Structure in HPSG

Summary

The paper looks at how information structure (focus, ground, link, tail) should be represented in HPSG. It argues that the representation should be tied to discourse rather than to syntax. The starting point is the Engdahl & Vallduví (1996) analysis. There, FOCUS and GROUND (LINK, TAIL) take syntactic constituents as values. This handles the contrast between narrow and wide focus neatly, as in "The president hates the Delft china set". It fails when the informational partition does not line up with a single constituent, for example subject-plus-verb focus or complex focus spread over several arguments.

The author then reviews a semantics-oriented alternative from the PLUS dialogue system (Jokinen 1996). PLUS marks NewInfo and topics with semantic indices. This helps with topic tracking and with separating old from new referents. However, a single index often cannot pick out the focused material.

The proposal is to use Minimal Recursion Semantics (MRS). Information-structure features take sets of MRS handles as values, so arbitrary groups of predications can be marked as focus, link or tail. The paper re-encodes the problematic examples this way, including quantifier-scope readings. It frames this as a step toward treating information structure as a separate, discourse-level dimension, though the work is exploratory and has no implementation results.

Key points

• The paper critiques Engdahl & Vallduví (1996). Their FOCUS/LINK/TAIL values are signs, so subject-verb focus and complex focus (e.g. "John introduced Bill to Sue") cannot be expressed cleanly.
• It reviews PLUS (Jokinen 1996), a dialogue system that marks NewInfo and topic via semantic indices. This approach fails when the focus covers several distinct entities, such as an event and an object, which have different indices.
• It adopts MRS (Copestake, Flickinger & Sag). Each predication carries a handle, and FOCUS, LINK and TAIL become sets of handles.
• Handles also cover scoped versus unscoped quantifier readings. The paper gives "every dog chased some cat" in both narrow and wide scope forms with information structure attached.
• Related work includes Teich's treatment of focus scope in Systemic Functional Grammar. The contribution is a representation proposal, with no corpus or evaluation.

Interesting ideas

• I1. Mark focus and topic as sets of pointers to individual predications, not as single phrases or single referents, so discontinuous or non-constituent focus is easy to state.
• I2. Keep information structure as its own layer, independent of both the UD tree and the semantic graph, linked to them only by references.
• I3. Use semantic indices to tell old discourse referents from newly introduced ones, as PLUS does for dialogue topic continuity.
• I4. Treat narrow versus wide focus as alternative assignments over the same semantic content, rather than as different parses.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (frames): prag frames could carry focus and topic as sets of frame-element ids. Questions like "who broke it?" versus "what did the butler do?" would then target the right slots even when the focus is not a subtree.
• I2 (world): the world-of-text could store a per-sentence given/new/topic record pointing into entities and events. This would support explainable topic tracking across a fable without changing the core graph.
• I3 (frames): coreference could tag each mention as given or new. This tag becomes an explainable feature for choosing antecedents and for detecting topic shifts.
• I4 (back): generation could pick word order or emphasis from the focus set over a fixed meaning graph, so the compressor and translator preserve what was new information.