Gaze as Evidence for Common Grounding: A Cross-Corpus Analysis of MapTask and MUNDEX
Nan Li, Albert Gatt, Massimo Poesio
cs.CL · cs.HC
Quick verdict
The paper argues gaze direction consistently tracks grounding states across MapTask and MUNDEX with unusually transparent statistics, but effects are tiny and fragile to participant-level clustering.
Summary
This empirical study asks whether video-coded gaze (partner/task/away) carries evidence of grounding in two asymmetric-information corpora: HCRC MapTask (5,144 reference-expression windows, 46 dialogues) and MUNDEX (807 windows, 26 interactions). Windows around references (MapTask) and understanding judgments (MUNDEX) yield seven feature groups; associations are tested with Mann-Whitney U plus cluster-robust logistic GEE, with Benjamini-Hochberg correction, and a grouped-cross-validation logistic-regression ablation probes predictability. Directional convergence emerges: aligned references and UND judgments show more task gaze, less partner gaze, lower entropy, fewer transitions, strongest for givers and explainers; speaker gaze entropy drops at the resolving mention (dz=-.20, q=.044). Best feature groups beat controls modestly (macro-F1 .532 vs .472 in MapTask; .564 vs .544 in MUNDEX). Key limitations: effects are tiny (pooled |r| <= .058 MapTask), most MapTask associations and some MUNDEX ones vanish when recurring participants are the clustering unit, prediction gains are partition-dependent, and MUNDEX's pooled retrospective labels disagree across perspectives on 32.6% of linked events.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 4
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Across both corpora, aligned reference interpretations (MapTask) and UND judgments (MUNDEX) are associated with more task-directed gaze, less partner-directed gaze, lower gaze entropy, and fewer gaze transitions.
- These associations are clearest for the participant leading the task: giver-produced references in MapTask and explainer judgments in MUNDEX, the latter also co-varying with the explainee's gaze.
- In same-speaker MapTask reference chains, the speaker's gaze entropy is lower at the mention where a previously non-aligned referent becomes aligned (dz=-.20, q=.044), though this depends on the unit of inference.
- The best gaze feature groups improve modestly over controls under grouped cross-validation (macro-F1 .532 vs .472 in MapTask; .564 vs .544 in MUNDEX), with gains that are partition-dependent.
- Because effects are small and several weaken under participant-level clustering, gaze should be modeled as one contributing cue within the interactional state, alongside linguistic and task-context features.
Strengths
Genuine cross-corpus operationalization
The shared partner/task/away vocabulary (Section 3) maps MapTask up/down/off and MUNDEX EX/EE/TABLE/AWAY into one space, making grounding-gaze comparisons across tasks feasible; the authors explicitly claim transferability to other video-coded gaze corpora rather than eye-tracking data.
Unusually rigorous and transparent statistics
Beyond Mann-Whitney tests, the paper reports effect sizes (rank-biserial), BH/FDR correction (Tables 5-10), cluster-robust GEE (Tables 11-12), participant-level re-clustering with Mancl-DeRouen bias-reduced SEs (Appendix E), window-length sensitivity (Tables 3-4), and 30 reshuffled CV partitions (Appendix G) - far above field norms.
Negative and boundary results are reported, not buried
Novel within-speaker reference-chain analysis
Reproducibility support
Weaknesses
Headline findings are unit-of-inference fragile
Very small effect sizes of unclear practical value
Prediction task shows marginal signal
Noisy MUNDEX target construct with double-counted events
Chain-resolution analysis has design confounds
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What does the paper claim about gaze and common ground?
It claims that in two asymmetric dialogue tasks, moments where understanding is achieved (aligned references in MapTask; 'understood' judgments in MUNDEX) coincide with gaze patterns: more looking at the task materials, less looking at the partner, lower gaze variability (entropy), and fewer gaze switches. The effects are consistent in direction across the two corpora but statistically small, so the authors position gaze as one contributing cue rather than a standalone signal of grounding.
How strong is the effect of gaze on grounding state?
Very modest. The largest pooled effect sizes are rank-biserial correlations of .058 (MapTask) and .181 (MUNDEX), and several associations disappear when recurring participants - rather than individual windows or dialogues - are used as the unit of statistical inference. Prediction performance with gaze features exceeds simple control baselines by only about .02-.06 macro-F1.
How did the authors compare the MapTask and MUNDEX corpora?
They mapped each corpus's video-coded gaze categories into one shared three-way vocabulary (partner/task/away) and computed identical feature sets over time windows centered on grounding-relevant events: reference expressions to landmarks in MapTask and retrospectively annotated understanding moments in MUNDEX. They then ran the same association tests, role-stratified analyses, and prediction ablations on both.
Can gaze features be used to predict whether someone understood an instruction?
Only weakly in this setup. Logistic regression on gaze features reaches macro-F1 around .53 (MapTask) and .56 (MUNDEX) under grouped cross-validation, beating control-only models by .06 and .02 respectively - and in MUNDEX the gain is barely above what a role indicator alone provides. The authors themselves call the gains modest and partition-dependent, so this is a signal-detection result, not a deployable predictor.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Review is based on the full extracted text including appendices; Figure 1 and Figure 2 could only be assessed via their captions and referenced numbers.
- External repository contents (github.com/chnln/gaze-as-grounding-evidence) and dataset releases were not inspected; reproducibility rating relies on the paper's availability statements and the operational detail in Appendices A-B.
- The referenced Li et al. 2026a/b, Wang et al. 2026, Murat and Vogel 2026, and Lazarov and Grimminger 2026 works were not independently verified beyond their in-text descriptions.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.