chnln/grounded-misunderstandings-in-maptask

GMMT: Grounded Misunderstandings in MapTask

The Grounded Misunderstandings in MapTask (GMMT) dataset was produced for the LREC 2026 paper Grounded Misunderstandings in Asymmetric Dialogue: A Perspectivist Annotation Scheme for MapTask by Nan Li, Albert Gatt, and Massimo Poesio.

It provides perspectivist annotations of the HCRC MapTask corpus, capturing both speaker-intended and addressee-interpreted landmarks for every reference expression (RE). The annotations support quantitative study of how understanding emerges, diverges, and repairs in collaborative dialogue with asymmetric map information.

Key numbers: 16 map pairs · 128 dialogues · 13,077 annotated REs · 5 hierarchical attributes + 2 participant interpretation IDs per RE · 3 human-verified dialogues (504 REs).

Related resources

This Hugging Face repository is the loadable Dataset Viewer / Parquet version of GMMT.

• GitHub release: https://github.com/chnln/grounded-misunderstandings-in-maptask
• The GitHub repository contains the raw annotation JSON files, prompt template, output schema, reference-expression pointers, and reconstruction script.
• This Hugging Face repository packages GMMT as loadable dataset configurations, while also keeping the raw JSON, schema, scripts, and metadata assets available as files.

Transcript text and licensing

This release does not redistribute full HCRC MapTask transcript text. It includes our annotations, short reference-expression strings, structural pointers, timing information, landmark metadata, and model-generated reason explanations. The reason field may contain short quoted or paraphrased dialogue evidence, but it does not include full transcript context.

To obtain the dialogue text, including the RE-marked transcripts, run the included reconstruction script against your own copy of the MapTask corpus.

Loading

from datasets import load_dataset

# Default: one row per reference expression (13,077 rows)
ds = load_dataset("chnln/grounded-misunderstandings-in-maptask", "re_level")

# Human-verified gold subset (504 REs from 3 dialogues)
ds = load_dataset("chnln/grounded-misunderstandings-in-maptask", "re_level_verified")

# One row per dialogue: nested RE annotations + text-less context skeleton
ds = load_dataset("chnln/grounded-misunderstandings-in-maptask", "dialogue_level")

There is a single train split per config — this is a research/analysis resource, not a train/test benchmark. The verified subset overlaps the full set (same 3 dialogues), so it is a separate config, not a validation split.

Configurations

Config	Rows	Grain
re_level (default)	13,077	one reference expression
re_level_verified	504	one RE, 3 human-verified dialogues
dialogue_level	128	one dialogue (nested REs + context skeleton)
dialogue_level_verified	3	one dialogue, verified

re_level fields

Annotation labels, short RE surface strings, and pointers into the dialogue structure (no transcript text is stored).

Field	Type	Description
dialogue_id	string	e.g. q1ec2
map_id	string	map pair identifier, e.g. m9
ref_id_unif	string	unique RE ID, e.g. q1ec2.ref.0
annotation_source	string	llm_annotated or human_verified
is_quantificational	bool | null	cascade step 1
is_specified	bool | null	cascade step 2
is_accommodated	bool | null	cascade step 3
is_grounded	bool | null	cascade step 4
is_imagined	bool | null	cascade step 5
giver_interpretation	string	giver's landmark interpretation
follower_interpretation	string	follower's landmark interpretation ("" if not grounded)
status	string	aligned / pending / misunderstood
subtype	string	finer classification
concept_id	string	original MapTask landmark ID used in the raw annotation JSON, e.g. m9_site_of_plane_crash
speaker	string	who produced the RE
addressee	string	who received it
expression	string	the RE surface string (short)
utt_id	int	utterance the RE occurs in
transaction_id	string	the RE's own transaction
context_transaction_ids	list[string]	all transactions from dialogue start through this one — the cumulative context the RE was annotated with
end_utt_id_of_context	int	last utterance in the annotation context (end of the RE's transaction)
timed_unit_ids	list[string]	unified timed-unit IDs the RE spans (used by the reconstruction script)
start_time, end_time	float	RE timing (seconds)
reason	string	model-generated explanation; may include short quoted/paraphrased dialogue evidence
source_file	string	provenance

Nullable boolean fields use null when the cascade has already terminated at an earlier step (e.g. when is_specified = false, the three downstream booleans are null). Use re_level_verified for the human-verified subset; annotation_source identifies whether a row comes from llm_annotated or human_verified annotations.

dialogue_level fields

dialogue_id, map_id, annotation_source, num_reference_expressions, landmark_reference_expressions (the original nested RE annotations), plus a text-less context skeleton:

• utterances: [{ln, utt_id, speaker, transaction_id, ref_ids}] — turn structure and where each RE occurs, without words.
• transactions: [{transaction_id, utt_ids, end_utt_id, move_types}] — the task-subgoal segmentation used to bound the annotation context.

Fill in the words locally with the reconstruction script (below).

Map image correspondence

The original annotation prompts used both participants' maps as images, but this release does not redistribute MapTask map images. Obtain the map images from your own copy of the HCRC MapTask corpus.

The file assets/dialogue_map_images.csv maps each dialogue to the corresponding map pair and expected image filenames:

Field	Description
dialogue_id	dialogue identifier, e.g. q1ec2
map_id	map pair identifier, e.g. m9
giver_map_filename	expected giver map image filename, e.g. map9g.png
follower_map_filename	expected follower map image filename, e.g. map9f.png

These filenames follow the convention used by the MapTask map image files: for map ID m9, the expected giver and follower files are map9g.png and map9f.png.

MapTask background

In the HCRC MapTask, a route giver guides a route follower along a route through landmarks. Both hold maps of the same landscape, but the maps are not identical:

• identical landmarks (on both maps);
• lexical variants — different names for the same location (e.g. "white water" vs "rapids");
• existence discrepancies — a landmark on only one map;
• multiplicity discrepancies — a landmark appearing twice on one map, once on the other.

Participants cannot see each other's maps, which makes MapTask a natural setting for studying how references are understood — and misunderstood. The corpus has 16 map pairs (m0–m15), each used in 8 dialogues (128 total).

Annotation schema

Each RE is annotated with a 5-attribute cascade modeling incremental resolution from both speaker and addressee perspectives. Each step is evaluated only when the previous condition holds; when it terminates early, downstream attributes are null.

Step	Attribute	Perspective	Question	Proceed if
1	is_quantificational	Speaker	Is the RE an existence check?	false
2	is_specified	Addressee	Any uptake of the RE?	true
3	is_accommodated	Addressee	Processed without mishearing/confusion?	true
4	is_grounded	Addressee	Linked to a specific landmark?	true
5	is_imagined	Addressee	Grounded landmark absent from addressee's own map?	(terminal)

Understanding states (status): aligned (both ground to the same/equivalent landmark), pending (not fully grounded), misunderstood (both believe they agree but ground to different landmarks).

Landmark ID format: <map_id>_<concept>#<ordinal>@<side> — e.g. m9_site_of_plane_crash#0@g (#ordinal only for multiplicity landmarks: 0=bottom, 1=top; @g/@f = giver/follower map).

Reconstructing the marked transcripts

Because MapTask transcript text cannot be redistributed, the dialogue text is regenerated locally from your own corpus copy. scripts/reconstruct_transcripts.py reproduces the original RE-marked transcripts byte-for-byte (verified on all 128 dialogues).

1. Download the HCRC Map Task Corpus v2.1 from https://groups.inf.ed.ac.uk/maptask/maptasknxt.html and unpack it. The per-side timed-units live in maptaskv2-1/Data/timed-units/ (files named <dialogue>.g.timed-units.xml / <dialogue>.f.timed-units.xml).

2. Run:

   python scripts/reconstruct_transcripts.py \
     --maptask-tu-dir /path/to/maptaskv2-1/Data/timed-units \
     --out-dir ./transcripts_with_refs
   

   (--re-dir and --map-mapping default to the reference_expressions/ and assets/map_trans_mapping.txt files shipped here.)

Output lines look like:

[giver    ln:19  utt:13 ] pass well it's <<stony desert id:q1ec2.ref.4 lm:m9_stony_desert>> that's what i've got here

The script fills missing utt IDs, merges giver+follower timed-units, re-derives the unified timed_unit_ids, and inserts markers anchored on those IDs — so its output aligns exactly with the timed_unit_ids in this dataset.

Dataset statistics

Cascade attribute distribution across all 13,077 REs:

Attribute	True	False	Null
is_quantificational	1,604 (12.3%)	11,473 (87.7%)	0
is_specified	11,137 (85.2%)	336 (2.6%)	1,604 (12.3%)
is_accommodated	10,977 (83.9%)	160 (1.2%)	1,940 (14.8%)
is_grounded	9,965 (76.2%)	1,012 (7.7%)	2,100 (16.1%)
is_imagined	632 (4.8%)	9,333 (71.4%)	3,112 (23.8%)

Understanding states (after lexical-variant unification):

Status	Count	%
aligned	9,435	72.1%
pending	3,403	26.0%
misunderstood	239	1.8%

Givers produce 8,679 REs (66.4%), followers 4,398 (33.6%).

Files in this repository

data/                        four Parquet configs (the loadable dataset)
raw/dialogues/               128 original per-dialogue annotation JSONs
raw/human_verified_subset/   3 gold-standard annotation JSONs
reference_expressions/       128 RE-extraction JSONs (timed_unit_ids etc.; used by reconstruction)
assets/                      landmark metadata (see below)
schema/                      annotation_output_schema.json, annotation_prompt_template.json
scripts/                     reconstruct_transcripts.py

Asset files (assets/)

• landmarks_configuration.csv — metadata for all 267 landmarks (map presence, discrepancy types). Columns assim/sim/contr/omo are inherited from the original corpus, are not clearly documented, and are unused in our scheme; kept for completeness.
• lexical_variant_landmark_info.json — the 10 lexical-variant landmark pairs.
• multiplicity_landmark_position.json — the 16 multiplicity landmarks with instance positions (ordinal 0=lower, 1=upper).
• map_trans_mapping.txt — map ID numeric suffix (m0–m15) → its 8 dialogue IDs (space-separated).
• dialogue_map_images.csv — dialogue ID → map ID and expected giver/follower map image filenames (e.g., map9g.png, map9f.png).

All asset files were processed from the original HCRC MapTask corpus by the authors.

The original MapTask corpus & licensing

This release contains derived annotations, structure, pointers, short RE strings, landmark metadata, and reconstruction code. The original HCRC MapTask corpus (audio, maps, full transcriptions) is available at https://groups.inf.ed.ac.uk/maptask/ under the corpus's own terms, which do not permit onward redistribution of the transcript text. Accordingly:

• Our annotations, the landmark-metadata assets, prompt template, schema, map-image filename correspondence, and reconstruction code are released under CC-BY-4.0 (see LICENSE).
• Full MapTask transcript text and map images are not included; obtain them from the link above and regenerate the marked transcripts locally with scripts/reconstruct_transcripts.py.

Please also cite and acknowledge the original HCRC MapTask corpus (Anderson et al., 1991).

Citation

If you use the GMMT dataset, please cite:

@inproceedings{li-etal-2026-grounded,
  title     = {Grounded Misunderstandings in Asymmetric Dialogue: A Perspectivist Annotation Scheme for MapTask},
  author    = {Li, Nan and Gatt, Albert and Poesio, Massimo},
  booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference (LREC 2026)},
  month     = {May},
  year      = {2026},
  pages     = {4988--5001},
  address   = {Palma, Mallorca, Spain},
  publisher = {European Language Resources Association (ELRA)},
  editor    = {Piperidis, Stelios and Bel, N{\'u}ria and van den Heuvel, Henk and Ide, Nancy and Krek, Simon and Toral, Antonio},
  doi       = {10.63317/59anbt78wyj7}
}

Preprint: arXiv:2511.03718.

License

CC-BY-4.0. See LICENSE. This license covers the released annotations and derived metadata; MapTask transcript text is not included and remains under the original corpus's terms.