MultiWhoAudio

MultiWhoAudio is a spoken-language dataset for evaluating how models resolve relative clause (RC) attachment ambiguity from audio input. It was constructed and used by Hu & Issa (2026) to probe prosodic and semantic sensitivity in large audio language models. It is the audio counterpart to the text-based MultiWho dataset (Lee et al., 2025), extending that line of work from written to spoken stimuli across three languages and three experimental conditions.

What the dataset probes

Sentences of the form "The doctor met the son of the gentleman who had a beard" are structurally ambiguous: the relative clause (who had a beard) can attach to either of two candidate nouns —

• DP1 (high attachment): the first, structurally higher noun (son)
• DP2 (low attachment): the second, embedded noun (gentleman)

Each item asks which noun the RC is understood to modify. Attachment preferences are known to vary by language and to be shaped by factors such as constituent length, syntactic position, semantic/pragmatic plausibility, and prosody. The dataset lets you measure a model's attachment preference from listening rather than reading, and compare it against human sentence-processing patterns.

Languages and conditions

The data is organized as a single default subset split into 9 parts — one per language × condition combination.

Language	Code	Baseline	Semantic	Prosodic
English	en	96	192	192
Arabic	ar	96	192	192
Chinese	ch	96	192	192

Total: 1,440 rows (splits: en_baseline, ar_baseline, ch_baseline, en_semantic, ar_semantic, ch_semantic, en_prosodic, ar_prosodic, ch_prosodic).

• Baseline — structurally ambiguous items with no biasing cue.
• Semantic — items where world-knowledge / plausibility favors one attachment.
• Prosodic — items where spoken prosody (phrasing, boundaries) cues an attachment.

Data fields

Field	Type	Description
ID	string	Item identifier (e.g., 1_0).
condition	string	Experimental condition: baseline, semantic, or prosodic.
sentence	string	The ambiguous carrier sentence.
RC	string	The relative clause whose attachment is being tested.
DP1	string	The high-attachment candidate noun (first / structurally higher DP).
DP2	string	The low-attachment candidate noun (second / embedded DP).
yes_no_prompt	string	A Yes/No comprehension prompt targeting one attachment reading.
choice_prompt	string	A two-option (1/2) forced-choice prompt over DP1 vs. DP2.
audio	audio	The spoken rendering of the sentence.

Example usage

from datasets import load_dataset

ds = load_dataset("clap-purdue/MultiWhoAudio", split="en_baseline")
ex = ds[0]

print(ex["sentence"])       # "The doctor met the son of the gentleman who had a beard."
print(ex["DP1"], ex["DP2"]) # "son" "gentleman"
audio = ex["audio"]         # {'array': ..., 'sampling_rate': ...}

Intended uses

• Evaluating audio / speech language models on syntactic ambiguity resolution.
• Studying whether spoken prosody shifts model attachment preferences.
• Cross-linguistic comparison of attachment behavior (English, Arabic, Chinese).
• Comparing model behavior against known human sentence-processing patterns.

Provenance

MultiWhoAudio was constructed and used by Hu & Issa (2026) to probe prosodic and semantic sensitivity in large audio language models. It adapts the relative-clause attachment paradigm of the text-based MultiWho dataset (Lee et al., 2025) to the auditory modality, adding the prosodic condition that is only meaningful in speech.

Citation

If you use this dataset, please cite the paper that constructed it:

@inproceedings{hu-issa-2026-relative,
  title     = "Where Does the Relative Clause Attach? Probing Prosodic and Semantic Sensitivity in Large Audio Language Models",
  author    = "Hu, Jingying and Issa, Elsayed",
  booktitle = "Proceedings of the Conference on Language Modeling (COLM 2026)",
  year      = "2026",
  url       = "https://colmweb.org/AcceptedPapers.html"
}

Please also cite the text-based MultiWho dataset it extends:

@inproceedings{lee-etal-2025-relies,
  title     = "Who Relies More on World Knowledge and Bias for Syntactic Ambiguity Resolution: Humans or {LLM}s?",
  author    = "Lee, So Young and Scheinberg, Russell and Shore, Amber and Agrawal, Ameeta",
  booktitle = "Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)",
  month     = apr,
  year      = "2025",
  address   = "Albuquerque, New Mexico",
  publisher = "Association for Computational Linguistics",
  isbn      = "979-8-89176-189-6",
  url       = "https://aclanthology.org/2025.naacl-long.177/"
}

Contact

Maintained by the Computational Linguistics group at Purdue (clap-purdue).