📄 An Underspecified HPSG Representation for Information Structure

Summary

The paper tackles a computational problem with information structure (IS). In German, and similarly in English, a given pitch-accent pattern usually fits several possible focus structures. Existing accounts, including the HPSG treatment by Engdahl/Vallduví (1994), list these readings as disjunctive alternatives. Those alternatives then multiply with every other ambiguity in the sentence.

Kuhn instead proposes a single graph-based underspecified representation:

• Each constituent's meaning is a node.
• Edges lead toward two partitions, focus (FOC) and background (BG).
• Edges are either obligatory or defeasible.

HPSG principles build this graph compositionally from accenting and from F-projection rules adapted from Selkirk (1995). Context then resolves it, using Schwarzschild's (1995) notion of givenness, defined as entailment by prior discourse, together with an F-skeleton. Contextual entailment adds obligatory links that override defeasible ones, and this can collapse the graph into one fully specified IS.

The motivation comes from text-to-speech and speech translation. In these settings even partial IS information can settle choices such as:

• accent placement in the output,
• the translation of the particle noch,
• the lexical sense of Besuch ("visit" vs "visitors").

Key points

• Target applications: text-to-speech and sentence-based speech translation with semantic transfer. Ignoring IS there can produce output that is truth-conditionally correct but contextually misleading.
• Scale of the problem: a short German sentence with two accents already yields nine alternative F-markings under the projection rules. This is the motivation for a compact encoding.
• Graph links:
  • Solid arrows are called h_link (obligatory).
  • Dashed arrows are called s_link (defeasible).
  • A chain of links that reaches FOC is what licenses focus.
• New HPSG features: the CONT type gains O-SEM, F-SKEL, IS-CSTR and MAX-F, and phonology gains PROM (accented/unaccented).
• Principles:
  • Focus Linking
  • Head F-Projection
  • Indirect Head F-Marking
  • Background Linking
  • F-Skeleton Instantiation
  • the F-Skeleton Principle
• Avoid F becomes unnecessary: the design makes Schwarzschild's "Avoid F" filter redundant. The author also questions that filter on pragmatic grounds.

Interesting ideas

• I1. Instead of enumerating every reading, the paper keeps one graph whose edges are marked strict or defeasible, and lets later evidence harden edges until the ambiguity disappears; this avoids a combinatorial explosion of readings.
• I2. A partly resolved representation already supports some decisions, such as choosing the sense of Besuch, while other questions stay open; the paper recommends resolving only when a task actually needs it.
• I3. Givenness is tested by whether the context entails an F-skeleton, the meaning with its focused parts replaced by variables; this explains narrow focus even when a question seems to ask for broad focus.
• I4. The links to FOC and BG come from accenting, while the links between constituents come from grammar alone, regardless of prosody; this cleanly separates evidence sources that are combined in one structure.
• I5. Information structure is used to pick between translation or realization alternatives that are logically equivalent but differ in context fit, as with accenting and the particle noch.

For Mova

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

• I1 (frames): Speech-act or focus ambiguity could be stored as defeasible links on frame nodes and only firmed up when the discourse supplies evidence, rather than forking frames.
• I2 (answer): Readers could query the underspecified structure lazily and resolve only the part relevant to the question being answered, saving work and keeping explanations short.
• I3 (world): The world model could check whether an utterance's skeleton like "photographed a X rose" is already entailed, marking which entities are new versus given for coreference and state tracking.
• I4 (tree): Projection links derived from UD head/complement relations could be computed once per tree and then combined with separate cues, such as emphasis markup or italics, without re-parsing.
• I5 (back): Generation from the meaning graph could consult FOC/BG partitions to choose word order, emphasis or particles so regenerated text fits its context, not just its truth conditions.