📄 Can pitch accent type convey information status in yes-no questions?

Summary

The paper asks whether the tune of a yes-no question in Bari Italian shows if the speaker is requesting new information or seeking confirmation of something already established. Italian polar questions use declarative word order, so intonation carries this distinction. The authors analyse questions from task-oriented dialogues recorded by six speaker pairs using a Bari version of the HCRC Map Task. They label moves with the Map Task conversational-game scheme (QUERY, CHECK, ALIGN) and add OBJECT for echo-like moves that signal a communication breakdown. Contours are transcribed in a modified ToBI.

Move type does not map one-to-one onto pitch accent. A better predictor is how confident the speaker is that the material is already shared:

• Low confidence: questions carry a rising-peak L+H* focal accent, and later words may be reaccented.
• High confidence: questions use an early-peak falling H*+L, followed by deaccenting or a "suppressed" accent.

The result matters because it shows that information status is a graded property of questions, separate from the dialogue act itself.

Key points

• The corpus consists of Map Task–style route-transfer conversations, each held by one of six speaker pairs from Bari whose only contact was by hearing. Landmark names were built mostly from sonorants and controlled for stress placement, so that pitch tracks would be clean.
• QUERIES are realised as L+H* L-L%, where the final boundary may be H%, and an optional initial H* can occur. When focus is non-final, the remainder of the question forms a reduced-range subordinate intermediate phrase that repeats the same accent.
• CHECKS vary with speaker confidence:
  • Tentative CHECKS sound identical to QUERIES.
  • Confident CHECKS use H*+L, with the F0 peak early in the accented syllable, and post-focal material is not reaccented.
• ALIGNS such as "Va bene?" mostly pattern like QUERIES. OBJECT moves share the QUERY tones but differ through breathy voice and/or expanded pitch range, which signal incredulity.
• In Table 1, both CHECK and ALIGN allow either L+H* or H*+L, which rules out a fixed move-to-accent mapping. The authors plan perceptual tests to confirm the confidence hypothesis.

Interesting ideas

• I1. Treat "how sure is the speaker that this is already shared?" as its own graded dimension, separate from the speech-act label, instead of forcing it into the QUERY/CHECK category.
• I2. When the categories fail, the authors pick a different organising variable (information status) rather than adding more categories, a useful pattern whenever a label set shows no clean correlation with the features.
• I3. They add an OBJECT move for echo questions, which both respond within the current game and open a repair sub-game. This models communication breakdown explicitly instead of filing it under questions.
• I4. The authors control the elicitation material (sonorant landmark names, fixed stress) so the signal of interest is not masked by noise, a lightweight design choice for building diagnostic datasets.
• I5. Material that is given in the context but deaccented or "suppressed" is treated as a marker that the speaker presupposes it.

For Mova

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

• I1 (frames): prag could attach a confidence/givenness score to each yes-no question frame, independent of the act type, using textual cues such as "so…?" or declarative-syntax questions.
• I2 (learn): when witness parsers or the teacher disagree on a label, evaluation could test whether a latent variable explains the split better than the existing tag inventory.
• I3 (frames): echo repetitions in dialogue ("Anima Mia?!") could be tagged as repair/objection acts that point back to the echoed utterance, feeding coreference and irony/incredulity detection.
• I4 (answer): test fables or word problems could be built with controlled lexical slots, so a failure can be traced to the targeted phenomenon rather than to unrelated vocabulary.
• I5 (world): in text, analogous givenness signals (definite reference, "so", ellipsis) could mark propositions the speaker assumes are already in the world model, triggering a consistency check rather than adding a new fact.